[net-next,v3,4/4] vsock/virtio: MSG_ZEROCOPY flag support

Message ID 20230720214245.457298-5-AVKrasnov@sberdevices.ru
State New
Headers
Series vsock/virtio/vhost: MSG_ZEROCOPY preparations |

Commit Message

Arseniy Krasnov July 20, 2023, 9:42 p.m. UTC
  This adds handling of MSG_ZEROCOPY flag on transmission path: if this
flag is set and zerocopy transmission is possible (enabled in socket
options and transport allows zerocopy), then non-linear skb will be
created and filled with the pages of user's buffer. Pages of user's
buffer are locked in memory by 'get_user_pages()'. Second thing that
this patch does is replace type of skb owning: instead of calling
'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
of socket, so to decrease this field correctly proper skb destructor is
needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.

Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
---
 Changelog:
 v5(big patchset) -> v1:
  * Refactorings of 'if' conditions.
  * Remove extra blank line.
  * Remove 'frag_off' field unneeded init.
  * Add function 'virtio_transport_fill_skb()' which fills both linear
    and non-linear skb with provided data.
 v1 -> v2:
  * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
 v2 -> v3:
  * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
    provided 'iov_iter' with data could be sent in a zerocopy mode.
    If this callback is not set in transport - transport allows to send
    any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
    then zerocopy is allowed. Reason of this callback is that in case of
    G2H transmission we insert whole skb to the tx virtio queue and such
    skb must fit to the size of the virtio queue to be sent in a single
    iteration (may be tx logic in 'virtio_transport.c' could be reworked
    as in vhost to support partial send of current skb). This callback
    will be enabled only for G2H path. For details pls see comment 
    'Check that tx queue...' below.

 include/net/af_vsock.h                  |   3 +
 net/vmw_vsock/virtio_transport.c        |  39 ++++
 net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
 3 files changed, 241 insertions(+), 58 deletions(-)
  

Comments

Arseniy Krasnov July 21, 2023, 5:09 a.m. UTC | #1
On 21.07.2023 00:42, Arseniy Krasnov wrote:
> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
> flag is set and zerocopy transmission is possible (enabled in socket
> options and transport allows zerocopy), then non-linear skb will be
> created and filled with the pages of user's buffer. Pages of user's
> buffer are locked in memory by 'get_user_pages()'. Second thing that
> this patch does is replace type of skb owning: instead of calling
> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
> of socket, so to decrease this field correctly proper skb destructor is
> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
> 
> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
> ---
>  Changelog:
>  v5(big patchset) -> v1:
>   * Refactorings of 'if' conditions.
>   * Remove extra blank line.
>   * Remove 'frag_off' field unneeded init.
>   * Add function 'virtio_transport_fill_skb()' which fills both linear
>     and non-linear skb with provided data.
>  v1 -> v2:
>   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
>  v2 -> v3:
>   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
>     provided 'iov_iter' with data could be sent in a zerocopy mode.
>     If this callback is not set in transport - transport allows to send
>     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
>     then zerocopy is allowed. Reason of this callback is that in case of
>     G2H transmission we insert whole skb to the tx virtio queue and such
>     skb must fit to the size of the virtio queue to be sent in a single
>     iteration (may be tx logic in 'virtio_transport.c' could be reworked
>     as in vhost to support partial send of current skb). This callback
>     will be enabled only for G2H path. For details pls see comment 
>     'Check that tx queue...' below.
> 
>  include/net/af_vsock.h                  |   3 +
>  net/vmw_vsock/virtio_transport.c        |  39 ++++
>  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
>  3 files changed, 241 insertions(+), 58 deletions(-)
> 
> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
> index 0e7504a42925..a6b346eeeb8e 100644
> --- a/include/net/af_vsock.h
> +++ b/include/net/af_vsock.h
> @@ -177,6 +177,9 @@ struct vsock_transport {
>  
>  	/* Read a single skb */
>  	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
> +
> +	/* Zero-copy. */
> +	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
>  };
>  
>  /**** CORE ****/
> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
> index 7bbcc8093e51..23cb8ed638c4 100644
> --- a/net/vmw_vsock/virtio_transport.c
> +++ b/net/vmw_vsock/virtio_transport.c
> @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
>  	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>  }
>  
> +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
> +{
> +	struct virtio_vsock *vsock;
> +	bool res = false;
> +
> +	rcu_read_lock();
> +
> +	vsock = rcu_dereference(the_virtio_vsock);
> +	if (vsock) {
> +		struct virtqueue *vq;
> +		int iov_pages;
> +
> +		vq = vsock->vqs[VSOCK_VQ_TX];
> +
> +		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
> +
> +		/* Check that tx queue is large enough to keep whole
> +		 * data to send. This is needed, because when there is
> +		 * not enough free space in the queue, current skb to
> +		 * send will be reinserted to the head of tx list of
> +		 * the socket to retry transmission later, so if skb
> +		 * is bigger than whole queue, it will be reinserted
> +		 * again and again, thus blocking other skbs to be sent.
> +		 * Each page of the user provided buffer will be added
> +		 * as a single buffer to the tx virtqueue, so compare
> +		 * number of pages against maximum capacity of the queue.
> +		 * +1 means buffer for the packet header.
> +		 */
> +		if (iov_pages + 1 <= vq->num_max)

I think this check is actual only for case one we don't have indirect buffer feature.
With indirect mode whole data to send will be packed into one indirect buffer.

Thanks, Arseniy

> +			res = true;
> +	}
> +
> +	rcu_read_unlock();
> +
> +	return res;
> +}
> +
>  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
>  
>  static struct virtio_transport virtio_transport = {
> @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
>  		.seqpacket_allow          = virtio_transport_seqpacket_allow,
>  		.seqpacket_has_data       = virtio_transport_seqpacket_has_data,
>  
> +		.msgzerocopy_check_iov	  = virtio_transport_msgzerocopy_check_iov,
> +
>  		.notify_poll_in           = virtio_transport_notify_poll_in,
>  		.notify_poll_out          = virtio_transport_notify_poll_out,
>  		.notify_recv_init         = virtio_transport_notify_recv_init,
> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
> index 26a4d10da205..e4e3d541aff4 100644
> --- a/net/vmw_vsock/virtio_transport_common.c
> +++ b/net/vmw_vsock/virtio_transport_common.c
> @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
>  	return container_of(t, struct virtio_transport, transport);
>  }
>  
> -/* Returns a new packet on success, otherwise returns NULL.
> - *
> - * If NULL is returned, errp is set to a negative errno.
> - */
> -static struct sk_buff *
> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
> -			   size_t len,
> -			   u32 src_cid,
> -			   u32 src_port,
> -			   u32 dst_cid,
> -			   u32 dst_port)
> -{
> -	const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
> -	struct virtio_vsock_hdr *hdr;
> -	struct sk_buff *skb;
> -	void *payload;
> -	int err;
> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
> +				       size_t max_to_send)
> +{
> +	const struct vsock_transport *t;
> +	struct iov_iter *iov_iter;
>  
> -	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
> -	if (!skb)
> -		return NULL;
> +	if (!info->msg)
> +		return false;
>  
> -	hdr = virtio_vsock_hdr(skb);
> -	hdr->type	= cpu_to_le16(info->type);
> -	hdr->op		= cpu_to_le16(info->op);
> -	hdr->src_cid	= cpu_to_le64(src_cid);
> -	hdr->dst_cid	= cpu_to_le64(dst_cid);
> -	hdr->src_port	= cpu_to_le32(src_port);
> -	hdr->dst_port	= cpu_to_le32(dst_port);
> -	hdr->flags	= cpu_to_le32(info->flags);
> -	hdr->len	= cpu_to_le32(len);
> +	iov_iter = &info->msg->msg_iter;
>  
> -	if (info->msg && len > 0) {
> -		payload = skb_put(skb, len);
> -		err = memcpy_from_msg(payload, info->msg, len);
> -		if (err)
> -			goto out;
> +	t = vsock_core_get_transport(info->vsk);
>  
> -		if (msg_data_left(info->msg) == 0 &&
> -		    info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
> -			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
> +	if (t->msgzerocopy_check_iov &&
> +	    !t->msgzerocopy_check_iov(iov_iter))
> +		return false;
>  
> -			if (info->msg->msg_flags & MSG_EOR)
> -				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
> -		}
> +	/* Data is simple buffer. */
> +	if (iter_is_ubuf(iov_iter))
> +		return true;
> +
> +	if (!iter_is_iovec(iov_iter))
> +		return false;
> +
> +	if (iov_iter->iov_offset)
> +		return false;
> +
> +	/* We can't send whole iov. */
> +	if (iov_iter->count > max_to_send)
> +		return false;
> +
> +	return true;
> +}
> +
> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
> +					   struct sk_buff *skb,
> +					   struct msghdr *msg,
> +					   bool zerocopy)
> +{
> +	struct ubuf_info *uarg;
> +
> +	if (msg->msg_ubuf) {
> +		uarg = msg->msg_ubuf;
> +		net_zcopy_get(uarg);
> +	} else {
> +		struct iov_iter *iter = &msg->msg_iter;
> +		struct ubuf_info_msgzc *uarg_zc;
> +		int len;
> +
> +		/* Only ITER_IOVEC or ITER_UBUF are allowed and
> +		 * checked before.
> +		 */
> +		if (iter_is_iovec(iter))
> +			len = iov_length(iter->__iov, iter->nr_segs);
> +		else
> +			len = iter->count;
> +
> +		uarg = msg_zerocopy_realloc(sk_vsock(vsk),
> +					    len,
> +					    NULL);
> +		if (!uarg)
> +			return -1;
> +
> +		uarg_zc = uarg_to_msgzc(uarg);
> +		uarg_zc->zerocopy = zerocopy ? 1 : 0;
>  	}
>  
> -	if (info->reply)
> -		virtio_vsock_skb_set_reply(skb);
> +	skb_zcopy_init(skb, uarg);
>  
> -	trace_virtio_transport_alloc_pkt(src_cid, src_port,
> -					 dst_cid, dst_port,
> -					 len,
> -					 info->type,
> -					 info->op,
> -					 info->flags);
> +	return 0;
> +}
>  
> -	if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
> -		WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
> -		goto out;
> +static int virtio_transport_fill_skb(struct sk_buff *skb,
> +				     struct virtio_vsock_pkt_info *info,
> +				     size_t len,
> +				     bool zcopy)
> +{
> +	if (zcopy) {
> +		return __zerocopy_sg_from_iter(info->msg, NULL, skb,
> +					      &info->msg->msg_iter,
> +					      len);
> +	} else {
> +		void *payload;
> +		int err;
> +
> +		payload = skb_put(skb, len);
> +		err = memcpy_from_msg(payload, info->msg, len);
> +		if (err)
> +			return -1;
> +
> +		if (msg_data_left(info->msg))
> +			return 0;
> +
> +		return 0;
>  	}
> +}
>  
> -	return skb;
> +static void virtio_transport_init_hdr(struct sk_buff *skb,
> +				      struct virtio_vsock_pkt_info *info,
> +				      u32 src_cid,
> +				      u32 src_port,
> +				      u32 dst_cid,
> +				      u32 dst_port,
> +				      size_t len)
> +{
> +	struct virtio_vsock_hdr *hdr;
>  
> -out:
> -	kfree_skb(skb);
> -	return NULL;
> +	hdr = virtio_vsock_hdr(skb);
> +	hdr->type	= cpu_to_le16(info->type);
> +	hdr->op		= cpu_to_le16(info->op);
> +	hdr->src_cid	= cpu_to_le64(src_cid);
> +	hdr->dst_cid	= cpu_to_le64(dst_cid);
> +	hdr->src_port	= cpu_to_le32(src_port);
> +	hdr->dst_port	= cpu_to_le32(dst_port);
> +	hdr->flags	= cpu_to_le32(info->flags);
> +	hdr->len	= cpu_to_le32(len);
>  }
>  
>  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
> @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
>  		return VIRTIO_VSOCK_TYPE_SEQPACKET;
>  }
>  
> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
> +						  struct virtio_vsock_pkt_info *info,
> +						  size_t payload_len,
> +						  bool zcopy,
> +						  u32 src_cid,
> +						  u32 src_port,
> +						  u32 dst_cid,
> +						  u32 dst_port)
> +{
> +	struct sk_buff *skb;
> +	size_t skb_len;
> +
> +	skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
> +
> +	if (!zcopy)
> +		skb_len += payload_len;
> +
> +	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
> +	if (!skb)
> +		return NULL;
> +
> +	virtio_transport_init_hdr(skb, info, src_cid, src_port,
> +				  dst_cid, dst_port,
> +				  payload_len);
> +
> +	/* Set owner here, because '__zerocopy_sg_from_iter()' uses
> +	 * owner of skb without check to update 'sk_wmem_alloc'.
> +	 */
> +	if (vsk)
> +		skb_set_owner_w(skb, sk_vsock(vsk));
> +
> +	if (info->msg && payload_len > 0) {
> +		int err;
> +
> +		err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
> +		if (err)
> +			goto out;
> +
> +		if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
> +			struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
> +
> +			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
> +
> +			if (info->msg->msg_flags & MSG_EOR)
> +				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
> +		}
> +	}
> +
> +	if (info->reply)
> +		virtio_vsock_skb_set_reply(skb);
> +
> +	trace_virtio_transport_alloc_pkt(src_cid, src_port,
> +					 dst_cid, dst_port,
> +					 payload_len,
> +					 info->type,
> +					 info->op,
> +					 info->flags);
> +
> +	return skb;
> +out:
> +	kfree_skb(skb);
> +	return NULL;
> +}
> +
>  /* This function can only be used on connecting/connected sockets,
>   * since a socket assigned to a transport is required.
>   *
> @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
>  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>  					  struct virtio_vsock_pkt_info *info)
>  {
> +	u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
>  	u32 src_cid, src_port, dst_cid, dst_port;
>  	const struct virtio_transport *t_ops;
>  	struct virtio_vsock_sock *vvs;
>  	u32 pkt_len = info->pkt_len;
> +	bool can_zcopy = false;
>  	u32 rest_len;
>  	int ret;
>  
> @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>  	if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
>  		return pkt_len;
>  
> +	if (info->msg) {
> +		/* If zerocopy is not enabled by 'setsockopt()', we behave as
> +		 * there is no MSG_ZEROCOPY flag set.
> +		 */
> +		if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
> +			info->msg->msg_flags &= ~MSG_ZEROCOPY;
> +
> +		if (info->msg->msg_flags & MSG_ZEROCOPY)
> +			can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
> +
> +		if (can_zcopy)
> +			max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
> +					    (MAX_SKB_FRAGS * PAGE_SIZE));
> +	}
> +
>  	rest_len = pkt_len;
>  
>  	do {
>  		struct sk_buff *skb;
>  		size_t skb_len;
>  
> -		skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
> +		skb_len = min(max_skb_len, rest_len);
>  
> -		skb = virtio_transport_alloc_skb(info, skb_len,
> +		skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
>  						 src_cid, src_port,
>  						 dst_cid, dst_port);
>  		if (!skb) {
> @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>  			break;
>  		}
>  
> +		/* This is last skb to send this portion of data. */
> +		if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
> +		    skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
> +			if (virtio_transport_init_zcopy_skb(vsk, skb,
> +							    info->msg,
> +							    can_zcopy)) {
> +				ret = -ENOMEM;
> +				break;
> +			}
> +		}
> +
>  		virtio_transport_inc_tx_pkt(vvs, skb);
>  
>  		ret = t_ops->send_pkt(skb);
> @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
>  	if (!t)
>  		return -ENOTCONN;
>  
> -	reply = virtio_transport_alloc_skb(&info, 0,
> +	reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
>  					   le64_to_cpu(hdr->dst_cid),
>  					   le32_to_cpu(hdr->dst_port),
>  					   le64_to_cpu(hdr->src_cid),
  
Michael S. Tsirkin July 25, 2023, 8:25 a.m. UTC | #2
On Fri, Jul 21, 2023 at 12:42:45AM +0300, Arseniy Krasnov wrote:
> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
> flag is set and zerocopy transmission is possible (enabled in socket
> options and transport allows zerocopy), then non-linear skb will be
> created and filled with the pages of user's buffer. Pages of user's
> buffer are locked in memory by 'get_user_pages()'. Second thing that
> this patch does is replace type of skb owning: instead of calling
> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
> of socket, so to decrease this field correctly proper skb destructor is
> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
> 
> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
> ---
>  Changelog:
>  v5(big patchset) -> v1:
>   * Refactorings of 'if' conditions.
>   * Remove extra blank line.
>   * Remove 'frag_off' field unneeded init.
>   * Add function 'virtio_transport_fill_skb()' which fills both linear
>     and non-linear skb with provided data.
>  v1 -> v2:
>   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
>  v2 -> v3:
>   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
>     provided 'iov_iter' with data could be sent in a zerocopy mode.
>     If this callback is not set in transport - transport allows to send
>     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
>     then zerocopy is allowed. Reason of this callback is that in case of
>     G2H transmission we insert whole skb to the tx virtio queue and such
>     skb must fit to the size of the virtio queue to be sent in a single
>     iteration (may be tx logic in 'virtio_transport.c' could be reworked
>     as in vhost to support partial send of current skb). This callback
>     will be enabled only for G2H path. For details pls see comment 
>     'Check that tx queue...' below.
> 
>  include/net/af_vsock.h                  |   3 +
>  net/vmw_vsock/virtio_transport.c        |  39 ++++
>  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
>  3 files changed, 241 insertions(+), 58 deletions(-)
> 
> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
> index 0e7504a42925..a6b346eeeb8e 100644
> --- a/include/net/af_vsock.h
> +++ b/include/net/af_vsock.h
> @@ -177,6 +177,9 @@ struct vsock_transport {
>  
>  	/* Read a single skb */
>  	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
> +
> +	/* Zero-copy. */
> +	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
>  };
>  
>  /**** CORE ****/
> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
> index 7bbcc8093e51..23cb8ed638c4 100644
> --- a/net/vmw_vsock/virtio_transport.c
> +++ b/net/vmw_vsock/virtio_transport.c
> @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
>  	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>  }
>  
> +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
> +{
> +	struct virtio_vsock *vsock;
> +	bool res = false;
> +
> +	rcu_read_lock();
> +
> +	vsock = rcu_dereference(the_virtio_vsock);
> +	if (vsock) {
> +		struct virtqueue *vq;
> +		int iov_pages;
> +
> +		vq = vsock->vqs[VSOCK_VQ_TX];
> +
> +		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
> +
> +		/* Check that tx queue is large enough to keep whole
> +		 * data to send. This is needed, because when there is
> +		 * not enough free space in the queue, current skb to
> +		 * send will be reinserted to the head of tx list of
> +		 * the socket to retry transmission later, so if skb
> +		 * is bigger than whole queue, it will be reinserted
> +		 * again and again, thus blocking other skbs to be sent.
> +		 * Each page of the user provided buffer will be added
> +		 * as a single buffer to the tx virtqueue, so compare
> +		 * number of pages against maximum capacity of the queue.
> +		 * +1 means buffer for the packet header.
> +		 */
> +		if (iov_pages + 1 <= vq->num_max)
> +			res = true;


Yes but can't there already be buffers in the queue?
Then you can't stick num_max there.


> +	}
> +
> +	rcu_read_unlock();
> +
> +	return res;
> +}
> +
>  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
>  
>  static struct virtio_transport virtio_transport = {
> @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
>  		.seqpacket_allow          = virtio_transport_seqpacket_allow,
>  		.seqpacket_has_data       = virtio_transport_seqpacket_has_data,
>  
> +		.msgzerocopy_check_iov	  = virtio_transport_msgzerocopy_check_iov,
> +
>  		.notify_poll_in           = virtio_transport_notify_poll_in,
>  		.notify_poll_out          = virtio_transport_notify_poll_out,
>  		.notify_recv_init         = virtio_transport_notify_recv_init,
> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
> index 26a4d10da205..e4e3d541aff4 100644
> --- a/net/vmw_vsock/virtio_transport_common.c
> +++ b/net/vmw_vsock/virtio_transport_common.c
> @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
>  	return container_of(t, struct virtio_transport, transport);
>  }
>  
> -/* Returns a new packet on success, otherwise returns NULL.
> - *
> - * If NULL is returned, errp is set to a negative errno.
> - */
> -static struct sk_buff *
> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
> -			   size_t len,
> -			   u32 src_cid,
> -			   u32 src_port,
> -			   u32 dst_cid,
> -			   u32 dst_port)
> -{
> -	const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
> -	struct virtio_vsock_hdr *hdr;
> -	struct sk_buff *skb;
> -	void *payload;
> -	int err;
> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
> +				       size_t max_to_send)
> +{
> +	const struct vsock_transport *t;
> +	struct iov_iter *iov_iter;
>  
> -	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
> -	if (!skb)
> -		return NULL;
> +	if (!info->msg)
> +		return false;
>  
> -	hdr = virtio_vsock_hdr(skb);
> -	hdr->type	= cpu_to_le16(info->type);
> -	hdr->op		= cpu_to_le16(info->op);
> -	hdr->src_cid	= cpu_to_le64(src_cid);
> -	hdr->dst_cid	= cpu_to_le64(dst_cid);
> -	hdr->src_port	= cpu_to_le32(src_port);
> -	hdr->dst_port	= cpu_to_le32(dst_port);
> -	hdr->flags	= cpu_to_le32(info->flags);
> -	hdr->len	= cpu_to_le32(len);
> +	iov_iter = &info->msg->msg_iter;
>  
> -	if (info->msg && len > 0) {
> -		payload = skb_put(skb, len);
> -		err = memcpy_from_msg(payload, info->msg, len);
> -		if (err)
> -			goto out;
> +	t = vsock_core_get_transport(info->vsk);
>  
> -		if (msg_data_left(info->msg) == 0 &&
> -		    info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
> -			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
> +	if (t->msgzerocopy_check_iov &&
> +	    !t->msgzerocopy_check_iov(iov_iter))
> +		return false;
>  
> -			if (info->msg->msg_flags & MSG_EOR)
> -				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
> -		}
> +	/* Data is simple buffer. */
> +	if (iter_is_ubuf(iov_iter))
> +		return true;
> +
> +	if (!iter_is_iovec(iov_iter))
> +		return false;
> +
> +	if (iov_iter->iov_offset)
> +		return false;
> +
> +	/* We can't send whole iov. */
> +	if (iov_iter->count > max_to_send)
> +		return false;
> +
> +	return true;
> +}
> +
> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
> +					   struct sk_buff *skb,
> +					   struct msghdr *msg,
> +					   bool zerocopy)
> +{
> +	struct ubuf_info *uarg;
> +
> +	if (msg->msg_ubuf) {
> +		uarg = msg->msg_ubuf;
> +		net_zcopy_get(uarg);
> +	} else {
> +		struct iov_iter *iter = &msg->msg_iter;
> +		struct ubuf_info_msgzc *uarg_zc;
> +		int len;
> +
> +		/* Only ITER_IOVEC or ITER_UBUF are allowed and
> +		 * checked before.
> +		 */
> +		if (iter_is_iovec(iter))
> +			len = iov_length(iter->__iov, iter->nr_segs);
> +		else
> +			len = iter->count;
> +
> +		uarg = msg_zerocopy_realloc(sk_vsock(vsk),
> +					    len,
> +					    NULL);
> +		if (!uarg)
> +			return -1;
> +
> +		uarg_zc = uarg_to_msgzc(uarg);
> +		uarg_zc->zerocopy = zerocopy ? 1 : 0;
>  	}
>  
> -	if (info->reply)
> -		virtio_vsock_skb_set_reply(skb);
> +	skb_zcopy_init(skb, uarg);
>  
> -	trace_virtio_transport_alloc_pkt(src_cid, src_port,
> -					 dst_cid, dst_port,
> -					 len,
> -					 info->type,
> -					 info->op,
> -					 info->flags);
> +	return 0;
> +}
>  
> -	if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
> -		WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
> -		goto out;
> +static int virtio_transport_fill_skb(struct sk_buff *skb,
> +				     struct virtio_vsock_pkt_info *info,
> +				     size_t len,
> +				     bool zcopy)
> +{
> +	if (zcopy) {
> +		return __zerocopy_sg_from_iter(info->msg, NULL, skb,
> +					      &info->msg->msg_iter,
> +					      len);
> +	} else {
> +		void *payload;
> +		int err;
> +
> +		payload = skb_put(skb, len);
> +		err = memcpy_from_msg(payload, info->msg, len);
> +		if (err)
> +			return -1;
> +
> +		if (msg_data_left(info->msg))
> +			return 0;
> +
> +		return 0;
>  	}
> +}
>  
> -	return skb;
> +static void virtio_transport_init_hdr(struct sk_buff *skb,
> +				      struct virtio_vsock_pkt_info *info,
> +				      u32 src_cid,
> +				      u32 src_port,
> +				      u32 dst_cid,
> +				      u32 dst_port,
> +				      size_t len)
> +{
> +	struct virtio_vsock_hdr *hdr;
>  
> -out:
> -	kfree_skb(skb);
> -	return NULL;
> +	hdr = virtio_vsock_hdr(skb);
> +	hdr->type	= cpu_to_le16(info->type);
> +	hdr->op		= cpu_to_le16(info->op);
> +	hdr->src_cid	= cpu_to_le64(src_cid);
> +	hdr->dst_cid	= cpu_to_le64(dst_cid);
> +	hdr->src_port	= cpu_to_le32(src_port);
> +	hdr->dst_port	= cpu_to_le32(dst_port);
> +	hdr->flags	= cpu_to_le32(info->flags);
> +	hdr->len	= cpu_to_le32(len);
>  }
>  
>  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
> @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
>  		return VIRTIO_VSOCK_TYPE_SEQPACKET;
>  }
>  
> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
> +						  struct virtio_vsock_pkt_info *info,
> +						  size_t payload_len,
> +						  bool zcopy,
> +						  u32 src_cid,
> +						  u32 src_port,
> +						  u32 dst_cid,
> +						  u32 dst_port)
> +{
> +	struct sk_buff *skb;
> +	size_t skb_len;
> +
> +	skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
> +
> +	if (!zcopy)
> +		skb_len += payload_len;
> +
> +	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
> +	if (!skb)
> +		return NULL;
> +
> +	virtio_transport_init_hdr(skb, info, src_cid, src_port,
> +				  dst_cid, dst_port,
> +				  payload_len);
> +
> +	/* Set owner here, because '__zerocopy_sg_from_iter()' uses
> +	 * owner of skb without check to update 'sk_wmem_alloc'.
> +	 */
> +	if (vsk)
> +		skb_set_owner_w(skb, sk_vsock(vsk));
> +
> +	if (info->msg && payload_len > 0) {
> +		int err;
> +
> +		err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
> +		if (err)
> +			goto out;
> +
> +		if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
> +			struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
> +
> +			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
> +
> +			if (info->msg->msg_flags & MSG_EOR)
> +				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
> +		}
> +	}
> +
> +	if (info->reply)
> +		virtio_vsock_skb_set_reply(skb);
> +
> +	trace_virtio_transport_alloc_pkt(src_cid, src_port,
> +					 dst_cid, dst_port,
> +					 payload_len,
> +					 info->type,
> +					 info->op,
> +					 info->flags);
> +
> +	return skb;
> +out:
> +	kfree_skb(skb);
> +	return NULL;
> +}
> +
>  /* This function can only be used on connecting/connected sockets,
>   * since a socket assigned to a transport is required.
>   *
> @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
>  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>  					  struct virtio_vsock_pkt_info *info)
>  {
> +	u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
>  	u32 src_cid, src_port, dst_cid, dst_port;
>  	const struct virtio_transport *t_ops;
>  	struct virtio_vsock_sock *vvs;
>  	u32 pkt_len = info->pkt_len;
> +	bool can_zcopy = false;
>  	u32 rest_len;
>  	int ret;
>  
> @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>  	if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
>  		return pkt_len;
>  
> +	if (info->msg) {
> +		/* If zerocopy is not enabled by 'setsockopt()', we behave as
> +		 * there is no MSG_ZEROCOPY flag set.
> +		 */
> +		if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
> +			info->msg->msg_flags &= ~MSG_ZEROCOPY;
> +
> +		if (info->msg->msg_flags & MSG_ZEROCOPY)
> +			can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
> +
> +		if (can_zcopy)
> +			max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
> +					    (MAX_SKB_FRAGS * PAGE_SIZE));
> +	}
> +
>  	rest_len = pkt_len;
>  
>  	do {
>  		struct sk_buff *skb;
>  		size_t skb_len;
>  
> -		skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
> +		skb_len = min(max_skb_len, rest_len);
>  
> -		skb = virtio_transport_alloc_skb(info, skb_len,
> +		skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
>  						 src_cid, src_port,
>  						 dst_cid, dst_port);
>  		if (!skb) {
> @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>  			break;
>  		}
>  
> +		/* This is last skb to send this portion of data. */
> +		if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
> +		    skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
> +			if (virtio_transport_init_zcopy_skb(vsk, skb,
> +							    info->msg,
> +							    can_zcopy)) {
> +				ret = -ENOMEM;
> +				break;
> +			}
> +		}
> +
>  		virtio_transport_inc_tx_pkt(vvs, skb);
>  
>  		ret = t_ops->send_pkt(skb);
> @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
>  	if (!t)
>  		return -ENOTCONN;
>  
> -	reply = virtio_transport_alloc_skb(&info, 0,
> +	reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
>  					   le64_to_cpu(hdr->dst_cid),
>  					   le32_to_cpu(hdr->dst_port),
>  					   le64_to_cpu(hdr->src_cid),
> -- 
> 2.25.1
  
Arseniy Krasnov July 25, 2023, 8:39 a.m. UTC | #3
On 25.07.2023 11:25, Michael S. Tsirkin wrote:
> On Fri, Jul 21, 2023 at 12:42:45AM +0300, Arseniy Krasnov wrote:
>> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
>> flag is set and zerocopy transmission is possible (enabled in socket
>> options and transport allows zerocopy), then non-linear skb will be
>> created and filled with the pages of user's buffer. Pages of user's
>> buffer are locked in memory by 'get_user_pages()'. Second thing that
>> this patch does is replace type of skb owning: instead of calling
>> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
>> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
>> of socket, so to decrease this field correctly proper skb destructor is
>> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
>>
>> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
>> ---
>>  Changelog:
>>  v5(big patchset) -> v1:
>>   * Refactorings of 'if' conditions.
>>   * Remove extra blank line.
>>   * Remove 'frag_off' field unneeded init.
>>   * Add function 'virtio_transport_fill_skb()' which fills both linear
>>     and non-linear skb with provided data.
>>  v1 -> v2:
>>   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
>>  v2 -> v3:
>>   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
>>     provided 'iov_iter' with data could be sent in a zerocopy mode.
>>     If this callback is not set in transport - transport allows to send
>>     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
>>     then zerocopy is allowed. Reason of this callback is that in case of
>>     G2H transmission we insert whole skb to the tx virtio queue and such
>>     skb must fit to the size of the virtio queue to be sent in a single
>>     iteration (may be tx logic in 'virtio_transport.c' could be reworked
>>     as in vhost to support partial send of current skb). This callback
>>     will be enabled only for G2H path. For details pls see comment 
>>     'Check that tx queue...' below.
>>
>>  include/net/af_vsock.h                  |   3 +
>>  net/vmw_vsock/virtio_transport.c        |  39 ++++
>>  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
>>  3 files changed, 241 insertions(+), 58 deletions(-)
>>
>> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
>> index 0e7504a42925..a6b346eeeb8e 100644
>> --- a/include/net/af_vsock.h
>> +++ b/include/net/af_vsock.h
>> @@ -177,6 +177,9 @@ struct vsock_transport {
>>  
>>  	/* Read a single skb */
>>  	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
>> +
>> +	/* Zero-copy. */
>> +	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
>>  };
>>  
>>  /**** CORE ****/
>> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>> index 7bbcc8093e51..23cb8ed638c4 100644
>> --- a/net/vmw_vsock/virtio_transport.c
>> +++ b/net/vmw_vsock/virtio_transport.c
>> @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
>>  	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>>  }
>>  
>> +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
>> +{
>> +	struct virtio_vsock *vsock;
>> +	bool res = false;
>> +
>> +	rcu_read_lock();
>> +
>> +	vsock = rcu_dereference(the_virtio_vsock);
>> +	if (vsock) {
>> +		struct virtqueue *vq;
>> +		int iov_pages;
>> +
>> +		vq = vsock->vqs[VSOCK_VQ_TX];
>> +
>> +		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
>> +
>> +		/* Check that tx queue is large enough to keep whole
>> +		 * data to send. This is needed, because when there is
>> +		 * not enough free space in the queue, current skb to
>> +		 * send will be reinserted to the head of tx list of
>> +		 * the socket to retry transmission later, so if skb
>> +		 * is bigger than whole queue, it will be reinserted
>> +		 * again and again, thus blocking other skbs to be sent.
>> +		 * Each page of the user provided buffer will be added
>> +		 * as a single buffer to the tx virtqueue, so compare
>> +		 * number of pages against maximum capacity of the queue.
>> +		 * +1 means buffer for the packet header.
>> +		 */
>> +		if (iov_pages + 1 <= vq->num_max)
>> +			res = true;
> 
> 
> Yes but can't there already be buffers in the queue?
> Then you can't stick num_max there.

I think, that it is not critical, because vhost part always tries to process all
incoming buffers (yes, 'vhost_exceeds_weight()' breaks at some moment, but it will
reschedule tx kick ('vhost_vsock_handle_tx_kick()') work again), so current "too
big" skb will wait until there will be enough space in queue and as it is requeued
to the head of tx list it will be inserted to tx queue first.

But anyway, I agree that comparing to 'num_free' may be more effective to the whole
system performance...

Thanks, Arseniy

> 
> 
>> +	}
>> +
>> +	rcu_read_unlock();
>> +
>> +	return res;
>> +}
>> +
>>  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
>>  
>>  static struct virtio_transport virtio_transport = {
>> @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
>>  		.seqpacket_allow          = virtio_transport_seqpacket_allow,
>>  		.seqpacket_has_data       = virtio_transport_seqpacket_has_data,
>>  
>> +		.msgzerocopy_check_iov	  = virtio_transport_msgzerocopy_check_iov,
>> +
>>  		.notify_poll_in           = virtio_transport_notify_poll_in,
>>  		.notify_poll_out          = virtio_transport_notify_poll_out,
>>  		.notify_recv_init         = virtio_transport_notify_recv_init,
>> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
>> index 26a4d10da205..e4e3d541aff4 100644
>> --- a/net/vmw_vsock/virtio_transport_common.c
>> +++ b/net/vmw_vsock/virtio_transport_common.c
>> @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
>>  	return container_of(t, struct virtio_transport, transport);
>>  }
>>  
>> -/* Returns a new packet on success, otherwise returns NULL.
>> - *
>> - * If NULL is returned, errp is set to a negative errno.
>> - */
>> -static struct sk_buff *
>> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
>> -			   size_t len,
>> -			   u32 src_cid,
>> -			   u32 src_port,
>> -			   u32 dst_cid,
>> -			   u32 dst_port)
>> -{
>> -	const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
>> -	struct virtio_vsock_hdr *hdr;
>> -	struct sk_buff *skb;
>> -	void *payload;
>> -	int err;
>> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
>> +				       size_t max_to_send)
>> +{
>> +	const struct vsock_transport *t;
>> +	struct iov_iter *iov_iter;
>>  
>> -	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>> -	if (!skb)
>> -		return NULL;
>> +	if (!info->msg)
>> +		return false;
>>  
>> -	hdr = virtio_vsock_hdr(skb);
>> -	hdr->type	= cpu_to_le16(info->type);
>> -	hdr->op		= cpu_to_le16(info->op);
>> -	hdr->src_cid	= cpu_to_le64(src_cid);
>> -	hdr->dst_cid	= cpu_to_le64(dst_cid);
>> -	hdr->src_port	= cpu_to_le32(src_port);
>> -	hdr->dst_port	= cpu_to_le32(dst_port);
>> -	hdr->flags	= cpu_to_le32(info->flags);
>> -	hdr->len	= cpu_to_le32(len);
>> +	iov_iter = &info->msg->msg_iter;
>>  
>> -	if (info->msg && len > 0) {
>> -		payload = skb_put(skb, len);
>> -		err = memcpy_from_msg(payload, info->msg, len);
>> -		if (err)
>> -			goto out;
>> +	t = vsock_core_get_transport(info->vsk);
>>  
>> -		if (msg_data_left(info->msg) == 0 &&
>> -		    info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>> -			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>> +	if (t->msgzerocopy_check_iov &&
>> +	    !t->msgzerocopy_check_iov(iov_iter))
>> +		return false;
>>  
>> -			if (info->msg->msg_flags & MSG_EOR)
>> -				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>> -		}
>> +	/* Data is simple buffer. */
>> +	if (iter_is_ubuf(iov_iter))
>> +		return true;
>> +
>> +	if (!iter_is_iovec(iov_iter))
>> +		return false;
>> +
>> +	if (iov_iter->iov_offset)
>> +		return false;
>> +
>> +	/* We can't send whole iov. */
>> +	if (iov_iter->count > max_to_send)
>> +		return false;
>> +
>> +	return true;
>> +}
>> +
>> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
>> +					   struct sk_buff *skb,
>> +					   struct msghdr *msg,
>> +					   bool zerocopy)
>> +{
>> +	struct ubuf_info *uarg;
>> +
>> +	if (msg->msg_ubuf) {
>> +		uarg = msg->msg_ubuf;
>> +		net_zcopy_get(uarg);
>> +	} else {
>> +		struct iov_iter *iter = &msg->msg_iter;
>> +		struct ubuf_info_msgzc *uarg_zc;
>> +		int len;
>> +
>> +		/* Only ITER_IOVEC or ITER_UBUF are allowed and
>> +		 * checked before.
>> +		 */
>> +		if (iter_is_iovec(iter))
>> +			len = iov_length(iter->__iov, iter->nr_segs);
>> +		else
>> +			len = iter->count;
>> +
>> +		uarg = msg_zerocopy_realloc(sk_vsock(vsk),
>> +					    len,
>> +					    NULL);
>> +		if (!uarg)
>> +			return -1;
>> +
>> +		uarg_zc = uarg_to_msgzc(uarg);
>> +		uarg_zc->zerocopy = zerocopy ? 1 : 0;
>>  	}
>>  
>> -	if (info->reply)
>> -		virtio_vsock_skb_set_reply(skb);
>> +	skb_zcopy_init(skb, uarg);
>>  
>> -	trace_virtio_transport_alloc_pkt(src_cid, src_port,
>> -					 dst_cid, dst_port,
>> -					 len,
>> -					 info->type,
>> -					 info->op,
>> -					 info->flags);
>> +	return 0;
>> +}
>>  
>> -	if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
>> -		WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
>> -		goto out;
>> +static int virtio_transport_fill_skb(struct sk_buff *skb,
>> +				     struct virtio_vsock_pkt_info *info,
>> +				     size_t len,
>> +				     bool zcopy)
>> +{
>> +	if (zcopy) {
>> +		return __zerocopy_sg_from_iter(info->msg, NULL, skb,
>> +					      &info->msg->msg_iter,
>> +					      len);
>> +	} else {
>> +		void *payload;
>> +		int err;
>> +
>> +		payload = skb_put(skb, len);
>> +		err = memcpy_from_msg(payload, info->msg, len);
>> +		if (err)
>> +			return -1;
>> +
>> +		if (msg_data_left(info->msg))
>> +			return 0;
>> +
>> +		return 0;
>>  	}
>> +}
>>  
>> -	return skb;
>> +static void virtio_transport_init_hdr(struct sk_buff *skb,
>> +				      struct virtio_vsock_pkt_info *info,
>> +				      u32 src_cid,
>> +				      u32 src_port,
>> +				      u32 dst_cid,
>> +				      u32 dst_port,
>> +				      size_t len)
>> +{
>> +	struct virtio_vsock_hdr *hdr;
>>  
>> -out:
>> -	kfree_skb(skb);
>> -	return NULL;
>> +	hdr = virtio_vsock_hdr(skb);
>> +	hdr->type	= cpu_to_le16(info->type);
>> +	hdr->op		= cpu_to_le16(info->op);
>> +	hdr->src_cid	= cpu_to_le64(src_cid);
>> +	hdr->dst_cid	= cpu_to_le64(dst_cid);
>> +	hdr->src_port	= cpu_to_le32(src_port);
>> +	hdr->dst_port	= cpu_to_le32(dst_port);
>> +	hdr->flags	= cpu_to_le32(info->flags);
>> +	hdr->len	= cpu_to_le32(len);
>>  }
>>  
>>  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
>> @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>  		return VIRTIO_VSOCK_TYPE_SEQPACKET;
>>  }
>>  
>> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
>> +						  struct virtio_vsock_pkt_info *info,
>> +						  size_t payload_len,
>> +						  bool zcopy,
>> +						  u32 src_cid,
>> +						  u32 src_port,
>> +						  u32 dst_cid,
>> +						  u32 dst_port)
>> +{
>> +	struct sk_buff *skb;
>> +	size_t skb_len;
>> +
>> +	skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
>> +
>> +	if (!zcopy)
>> +		skb_len += payload_len;
>> +
>> +	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>> +	if (!skb)
>> +		return NULL;
>> +
>> +	virtio_transport_init_hdr(skb, info, src_cid, src_port,
>> +				  dst_cid, dst_port,
>> +				  payload_len);
>> +
>> +	/* Set owner here, because '__zerocopy_sg_from_iter()' uses
>> +	 * owner of skb without check to update 'sk_wmem_alloc'.
>> +	 */
>> +	if (vsk)
>> +		skb_set_owner_w(skb, sk_vsock(vsk));
>> +
>> +	if (info->msg && payload_len > 0) {
>> +		int err;
>> +
>> +		err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
>> +		if (err)
>> +			goto out;
>> +
>> +		if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>> +			struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
>> +
>> +			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>> +
>> +			if (info->msg->msg_flags & MSG_EOR)
>> +				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>> +		}
>> +	}
>> +
>> +	if (info->reply)
>> +		virtio_vsock_skb_set_reply(skb);
>> +
>> +	trace_virtio_transport_alloc_pkt(src_cid, src_port,
>> +					 dst_cid, dst_port,
>> +					 payload_len,
>> +					 info->type,
>> +					 info->op,
>> +					 info->flags);
>> +
>> +	return skb;
>> +out:
>> +	kfree_skb(skb);
>> +	return NULL;
>> +}
>> +
>>  /* This function can only be used on connecting/connected sockets,
>>   * since a socket assigned to a transport is required.
>>   *
>> @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>  					  struct virtio_vsock_pkt_info *info)
>>  {
>> +	u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
>>  	u32 src_cid, src_port, dst_cid, dst_port;
>>  	const struct virtio_transport *t_ops;
>>  	struct virtio_vsock_sock *vvs;
>>  	u32 pkt_len = info->pkt_len;
>> +	bool can_zcopy = false;
>>  	u32 rest_len;
>>  	int ret;
>>  
>> @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>  	if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
>>  		return pkt_len;
>>  
>> +	if (info->msg) {
>> +		/* If zerocopy is not enabled by 'setsockopt()', we behave as
>> +		 * there is no MSG_ZEROCOPY flag set.
>> +		 */
>> +		if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
>> +			info->msg->msg_flags &= ~MSG_ZEROCOPY;
>> +
>> +		if (info->msg->msg_flags & MSG_ZEROCOPY)
>> +			can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
>> +
>> +		if (can_zcopy)
>> +			max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
>> +					    (MAX_SKB_FRAGS * PAGE_SIZE));
>> +	}
>> +
>>  	rest_len = pkt_len;
>>  
>>  	do {
>>  		struct sk_buff *skb;
>>  		size_t skb_len;
>>  
>> -		skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
>> +		skb_len = min(max_skb_len, rest_len);
>>  
>> -		skb = virtio_transport_alloc_skb(info, skb_len,
>> +		skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
>>  						 src_cid, src_port,
>>  						 dst_cid, dst_port);
>>  		if (!skb) {
>> @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>  			break;
>>  		}
>>  
>> +		/* This is last skb to send this portion of data. */
>> +		if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
>> +		    skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
>> +			if (virtio_transport_init_zcopy_skb(vsk, skb,
>> +							    info->msg,
>> +							    can_zcopy)) {
>> +				ret = -ENOMEM;
>> +				break;
>> +			}
>> +		}
>> +
>>  		virtio_transport_inc_tx_pkt(vvs, skb);
>>  
>>  		ret = t_ops->send_pkt(skb);
>> @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
>>  	if (!t)
>>  		return -ENOTCONN;
>>  
>> -	reply = virtio_transport_alloc_skb(&info, 0,
>> +	reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
>>  					   le64_to_cpu(hdr->dst_cid),
>>  					   le32_to_cpu(hdr->dst_port),
>>  					   le64_to_cpu(hdr->src_cid),
>> -- 
>> 2.25.1
>
  
Stefano Garzarella July 25, 2023, 8:43 a.m. UTC | #4
On Fri, Jul 21, 2023 at 08:09:03AM +0300, Arseniy Krasnov wrote:
>
>
>On 21.07.2023 00:42, Arseniy Krasnov wrote:
>> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
>> flag is set and zerocopy transmission is possible (enabled in socket
>> options and transport allows zerocopy), then non-linear skb will be
>> created and filled with the pages of user's buffer. Pages of user's
>> buffer are locked in memory by 'get_user_pages()'. Second thing that
>> this patch does is replace type of skb owning: instead of calling
>> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
>> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
>> of socket, so to decrease this field correctly proper skb destructor is
>> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
>>
>> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
>> ---
>>  Changelog:
>>  v5(big patchset) -> v1:
>>   * Refactorings of 'if' conditions.
>>   * Remove extra blank line.
>>   * Remove 'frag_off' field unneeded init.
>>   * Add function 'virtio_transport_fill_skb()' which fills both linear
>>     and non-linear skb with provided data.
>>  v1 -> v2:
>>   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
>>  v2 -> v3:
>>   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
>>     provided 'iov_iter' with data could be sent in a zerocopy mode.
>>     If this callback is not set in transport - transport allows to send
>>     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
>>     then zerocopy is allowed. Reason of this callback is that in case of
>>     G2H transmission we insert whole skb to the tx virtio queue and such
>>     skb must fit to the size of the virtio queue to be sent in a single
>>     iteration (may be tx logic in 'virtio_transport.c' could be reworked
>>     as in vhost to support partial send of current skb). This callback
>>     will be enabled only for G2H path. For details pls see comment
>>     'Check that tx queue...' below.
>>
>>  include/net/af_vsock.h                  |   3 +
>>  net/vmw_vsock/virtio_transport.c        |  39 ++++
>>  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
>>  3 files changed, 241 insertions(+), 58 deletions(-)
>>
>> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
>> index 0e7504a42925..a6b346eeeb8e 100644
>> --- a/include/net/af_vsock.h
>> +++ b/include/net/af_vsock.h
>> @@ -177,6 +177,9 @@ struct vsock_transport {
>>
>>  	/* Read a single skb */
>>  	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
>> +
>> +	/* Zero-copy. */
>> +	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
>>  };
>>
>>  /**** CORE ****/
>> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>> index 7bbcc8093e51..23cb8ed638c4 100644
>> --- a/net/vmw_vsock/virtio_transport.c
>> +++ b/net/vmw_vsock/virtio_transport.c
>> @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
>>  	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>>  }
>>
>> +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
>> +{
>> +	struct virtio_vsock *vsock;
>> +	bool res = false;
>> +
>> +	rcu_read_lock();
>> +
>> +	vsock = rcu_dereference(the_virtio_vsock);
>> +	if (vsock) {
>> +		struct virtqueue *vq;
>> +		int iov_pages;
>> +
>> +		vq = vsock->vqs[VSOCK_VQ_TX];
>> +
>> +		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
>> +
>> +		/* Check that tx queue is large enough to keep whole
>> +		 * data to send. This is needed, because when there is
>> +		 * not enough free space in the queue, current skb to
>> +		 * send will be reinserted to the head of tx list of
>> +		 * the socket to retry transmission later, so if skb
>> +		 * is bigger than whole queue, it will be reinserted
>> +		 * again and again, thus blocking other skbs to be sent.
>> +		 * Each page of the user provided buffer will be added
>> +		 * as a single buffer to the tx virtqueue, so compare
>> +		 * number of pages against maximum capacity of the queue.
>> +		 * +1 means buffer for the packet header.
>> +		 */
>> +		if (iov_pages + 1 <= vq->num_max)
>
>I think this check is actual only for case one we don't have indirect buffer feature.
>With indirect mode whole data to send will be packed into one indirect buffer.

I think so.
So, should we check also that here?

>
>Thanks, Arseniy
>
>> +			res = true;
>> +	}
>> +
>> +	rcu_read_unlock();
>> +
>> +	return res;
>> +}
>> +
>>  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
>>
>>  static struct virtio_transport virtio_transport = {
>> @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
>>  		.seqpacket_allow          = virtio_transport_seqpacket_allow,
>>  		.seqpacket_has_data       = virtio_transport_seqpacket_has_data,
>>
>> +		.msgzerocopy_check_iov	  = virtio_transport_msgzerocopy_check_iov,
>> +
>>  		.notify_poll_in           = virtio_transport_notify_poll_in,
>>  		.notify_poll_out          = virtio_transport_notify_poll_out,
>>  		.notify_recv_init         = virtio_transport_notify_recv_init,
>> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
>> index 26a4d10da205..e4e3d541aff4 100644
>> --- a/net/vmw_vsock/virtio_transport_common.c
>> +++ b/net/vmw_vsock/virtio_transport_common.c
>> @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
>>  	return container_of(t, struct virtio_transport, transport);
>>  }
>>
>> -/* Returns a new packet on success, otherwise returns NULL.
>> - *
>> - * If NULL is returned, errp is set to a negative errno.
>> - */
>> -static struct sk_buff *
>> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
>> -			   size_t len,
>> -			   u32 src_cid,
>> -			   u32 src_port,
>> -			   u32 dst_cid,
>> -			   u32 dst_port)
>> -{
>> -	const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
>> -	struct virtio_vsock_hdr *hdr;
>> -	struct sk_buff *skb;
>> -	void *payload;
>> -	int err;
>> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
>> +				       size_t max_to_send)
>> +{
>> +	const struct vsock_transport *t;
>> +	struct iov_iter *iov_iter;
>>
>> -	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>> -	if (!skb)
>> -		return NULL;
>> +	if (!info->msg)
>> +		return false;
>>
>> -	hdr = virtio_vsock_hdr(skb);
>> -	hdr->type	= cpu_to_le16(info->type);
>> -	hdr->op		= cpu_to_le16(info->op);
>> -	hdr->src_cid	= cpu_to_le64(src_cid);
>> -	hdr->dst_cid	= cpu_to_le64(dst_cid);
>> -	hdr->src_port	= cpu_to_le32(src_port);
>> -	hdr->dst_port	= cpu_to_le32(dst_port);
>> -	hdr->flags	= cpu_to_le32(info->flags);
>> -	hdr->len	= cpu_to_le32(len);
>> +	iov_iter = &info->msg->msg_iter;
>>
>> -	if (info->msg && len > 0) {
>> -		payload = skb_put(skb, len);
>> -		err = memcpy_from_msg(payload, info->msg, len);
>> -		if (err)
>> -			goto out;
>> +	t = vsock_core_get_transport(info->vsk);
>>
>> -		if (msg_data_left(info->msg) == 0 &&
>> -		    info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>> -			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>> +	if (t->msgzerocopy_check_iov &&
>> +	    !t->msgzerocopy_check_iov(iov_iter))
>> +		return false;

I'd avoid adding a new transport callback used only internally in virtio
transports.

Usually the transport callbacks are used in af_vsock.c, if we need a
callback just for virtio transports, maybe better to add it in struct
virtio_vsock_pkt_info or struct virtio_vsock_sock.

Maybe better the last one so we don't have to allocate pointer space
for each packet and you should reach it via info.

Thanks,
Stefano

>>
>> -			if (info->msg->msg_flags & MSG_EOR)
>> -				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>> -		}
>> +	/* Data is simple buffer. */
>> +	if (iter_is_ubuf(iov_iter))
>> +		return true;
>> +
>> +	if (!iter_is_iovec(iov_iter))
>> +		return false;
>> +
>> +	if (iov_iter->iov_offset)
>> +		return false;
>> +
>> +	/* We can't send whole iov. */
>> +	if (iov_iter->count > max_to_send)
>> +		return false;
>> +
>> +	return true;
>> +}
>> +
>> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
>> +					   struct sk_buff *skb,
>> +					   struct msghdr *msg,
>> +					   bool zerocopy)
>> +{
>> +	struct ubuf_info *uarg;
>> +
>> +	if (msg->msg_ubuf) {
>> +		uarg = msg->msg_ubuf;
>> +		net_zcopy_get(uarg);
>> +	} else {
>> +		struct iov_iter *iter = &msg->msg_iter;
>> +		struct ubuf_info_msgzc *uarg_zc;
>> +		int len;
>> +
>> +		/* Only ITER_IOVEC or ITER_UBUF are allowed and
>> +		 * checked before.
>> +		 */
>> +		if (iter_is_iovec(iter))
>> +			len = iov_length(iter->__iov, iter->nr_segs);
>> +		else
>> +			len = iter->count;
>> +
>> +		uarg = msg_zerocopy_realloc(sk_vsock(vsk),
>> +					    len,
>> +					    NULL);
>> +		if (!uarg)
>> +			return -1;
>> +
>> +		uarg_zc = uarg_to_msgzc(uarg);
>> +		uarg_zc->zerocopy = zerocopy ? 1 : 0;
>>  	}
>>
>> -	if (info->reply)
>> -		virtio_vsock_skb_set_reply(skb);
>> +	skb_zcopy_init(skb, uarg);
>>
>> -	trace_virtio_transport_alloc_pkt(src_cid, src_port,
>> -					 dst_cid, dst_port,
>> -					 len,
>> -					 info->type,
>> -					 info->op,
>> -					 info->flags);
>> +	return 0;
>> +}
>>
>> -	if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
>> -		WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
>> -		goto out;
>> +static int virtio_transport_fill_skb(struct sk_buff *skb,
>> +				     struct virtio_vsock_pkt_info *info,
>> +				     size_t len,
>> +				     bool zcopy)
>> +{
>> +	if (zcopy) {
>> +		return __zerocopy_sg_from_iter(info->msg, NULL, skb,
>> +					      &info->msg->msg_iter,
>> +					      len);
>> +	} else {
>> +		void *payload;
>> +		int err;
>> +
>> +		payload = skb_put(skb, len);
>> +		err = memcpy_from_msg(payload, info->msg, len);
>> +		if (err)
>> +			return -1;
>> +
>> +		if (msg_data_left(info->msg))
>> +			return 0;
>> +
>> +		return 0;
>>  	}
>> +}
>>
>> -	return skb;
>> +static void virtio_transport_init_hdr(struct sk_buff *skb,
>> +				      struct virtio_vsock_pkt_info *info,
>> +				      u32 src_cid,
>> +				      u32 src_port,
>> +				      u32 dst_cid,
>> +				      u32 dst_port,
>> +				      size_t len)
>> +{
>> +	struct virtio_vsock_hdr *hdr;
>>
>> -out:
>> -	kfree_skb(skb);
>> -	return NULL;
>> +	hdr = virtio_vsock_hdr(skb);
>> +	hdr->type	= cpu_to_le16(info->type);
>> +	hdr->op		= cpu_to_le16(info->op);
>> +	hdr->src_cid	= cpu_to_le64(src_cid);
>> +	hdr->dst_cid	= cpu_to_le64(dst_cid);
>> +	hdr->src_port	= cpu_to_le32(src_port);
>> +	hdr->dst_port	= cpu_to_le32(dst_port);
>> +	hdr->flags	= cpu_to_le32(info->flags);
>> +	hdr->len	= cpu_to_le32(len);
>>  }
>>
>>  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
>> @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>  		return VIRTIO_VSOCK_TYPE_SEQPACKET;
>>  }
>>
>> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
>> +						  struct virtio_vsock_pkt_info *info,
>> +						  size_t payload_len,
>> +						  bool zcopy,
>> +						  u32 src_cid,
>> +						  u32 src_port,
>> +						  u32 dst_cid,
>> +						  u32 dst_port)
>> +{
>> +	struct sk_buff *skb;
>> +	size_t skb_len;
>> +
>> +	skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
>> +
>> +	if (!zcopy)
>> +		skb_len += payload_len;
>> +
>> +	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>> +	if (!skb)
>> +		return NULL;
>> +
>> +	virtio_transport_init_hdr(skb, info, src_cid, src_port,
>> +				  dst_cid, dst_port,
>> +				  payload_len);
>> +
>> +	/* Set owner here, because '__zerocopy_sg_from_iter()' uses
>> +	 * owner of skb without check to update 'sk_wmem_alloc'.
>> +	 */
>> +	if (vsk)
>> +		skb_set_owner_w(skb, sk_vsock(vsk));
>> +
>> +	if (info->msg && payload_len > 0) {
>> +		int err;
>> +
>> +		err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
>> +		if (err)
>> +			goto out;
>> +
>> +		if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>> +			struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
>> +
>> +			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>> +
>> +			if (info->msg->msg_flags & MSG_EOR)
>> +				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>> +		}
>> +	}
>> +
>> +	if (info->reply)
>> +		virtio_vsock_skb_set_reply(skb);
>> +
>> +	trace_virtio_transport_alloc_pkt(src_cid, src_port,
>> +					 dst_cid, dst_port,
>> +					 payload_len,
>> +					 info->type,
>> +					 info->op,
>> +					 info->flags);
>> +
>> +	return skb;
>> +out:
>> +	kfree_skb(skb);
>> +	return NULL;
>> +}
>> +
>>  /* This function can only be used on connecting/connected sockets,
>>   * since a socket assigned to a transport is required.
>>   *
>> @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>  					  struct virtio_vsock_pkt_info *info)
>>  {
>> +	u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
>>  	u32 src_cid, src_port, dst_cid, dst_port;
>>  	const struct virtio_transport *t_ops;
>>  	struct virtio_vsock_sock *vvs;
>>  	u32 pkt_len = info->pkt_len;
>> +	bool can_zcopy = false;
>>  	u32 rest_len;
>>  	int ret;
>>
>> @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>  	if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
>>  		return pkt_len;
>>
>> +	if (info->msg) {
>> +		/* If zerocopy is not enabled by 'setsockopt()', we behave as
>> +		 * there is no MSG_ZEROCOPY flag set.
>> +		 */
>> +		if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
>> +			info->msg->msg_flags &= ~MSG_ZEROCOPY;
>> +
>> +		if (info->msg->msg_flags & MSG_ZEROCOPY)
>> +			can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
>> +
>> +		if (can_zcopy)
>> +			max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
>> +					    (MAX_SKB_FRAGS * PAGE_SIZE));
>> +	}
>> +
>>  	rest_len = pkt_len;
>>
>>  	do {
>>  		struct sk_buff *skb;
>>  		size_t skb_len;
>>
>> -		skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
>> +		skb_len = min(max_skb_len, rest_len);
>>
>> -		skb = virtio_transport_alloc_skb(info, skb_len,
>> +		skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
>>  						 src_cid, src_port,
>>  						 dst_cid, dst_port);
>>  		if (!skb) {
>> @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>  			break;
>>  		}
>>
>> +		/* This is last skb to send this portion of data. */
>> +		if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
>> +		    skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
>> +			if (virtio_transport_init_zcopy_skb(vsk, skb,
>> +							    info->msg,
>> +							    can_zcopy)) {
>> +				ret = -ENOMEM;
>> +				break;
>> +			}
>> +		}
>> +
>>  		virtio_transport_inc_tx_pkt(vvs, skb);
>>
>>  		ret = t_ops->send_pkt(skb);
>> @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
>>  	if (!t)
>>  		return -ENOTCONN;
>>
>> -	reply = virtio_transport_alloc_skb(&info, 0,
>> +	reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
>>  					   le64_to_cpu(hdr->dst_cid),
>>  					   le32_to_cpu(hdr->dst_port),
>>  					   le64_to_cpu(hdr->src_cid),
>
  
Arseniy Krasnov July 25, 2023, 8:46 a.m. UTC | #5
On 25.07.2023 11:43, Stefano Garzarella wrote:
> On Fri, Jul 21, 2023 at 08:09:03AM +0300, Arseniy Krasnov wrote:
>>
>>
>> On 21.07.2023 00:42, Arseniy Krasnov wrote:
>>> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
>>> flag is set and zerocopy transmission is possible (enabled in socket
>>> options and transport allows zerocopy), then non-linear skb will be
>>> created and filled with the pages of user's buffer. Pages of user's
>>> buffer are locked in memory by 'get_user_pages()'. Second thing that
>>> this patch does is replace type of skb owning: instead of calling
>>> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
>>> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
>>> of socket, so to decrease this field correctly proper skb destructor is
>>> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
>>>
>>> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
>>> ---
>>>  Changelog:
>>>  v5(big patchset) -> v1:
>>>   * Refactorings of 'if' conditions.
>>>   * Remove extra blank line.
>>>   * Remove 'frag_off' field unneeded init.
>>>   * Add function 'virtio_transport_fill_skb()' which fills both linear
>>>     and non-linear skb with provided data.
>>>  v1 -> v2:
>>>   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
>>>  v2 -> v3:
>>>   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
>>>     provided 'iov_iter' with data could be sent in a zerocopy mode.
>>>     If this callback is not set in transport - transport allows to send
>>>     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
>>>     then zerocopy is allowed. Reason of this callback is that in case of
>>>     G2H transmission we insert whole skb to the tx virtio queue and such
>>>     skb must fit to the size of the virtio queue to be sent in a single
>>>     iteration (may be tx logic in 'virtio_transport.c' could be reworked
>>>     as in vhost to support partial send of current skb). This callback
>>>     will be enabled only for G2H path. For details pls see comment
>>>     'Check that tx queue...' below.
>>>
>>>  include/net/af_vsock.h                  |   3 +
>>>  net/vmw_vsock/virtio_transport.c        |  39 ++++
>>>  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
>>>  3 files changed, 241 insertions(+), 58 deletions(-)
>>>
>>> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
>>> index 0e7504a42925..a6b346eeeb8e 100644
>>> --- a/include/net/af_vsock.h
>>> +++ b/include/net/af_vsock.h
>>> @@ -177,6 +177,9 @@ struct vsock_transport {
>>>
>>>      /* Read a single skb */
>>>      int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
>>> +
>>> +    /* Zero-copy. */
>>> +    bool (*msgzerocopy_check_iov)(const struct iov_iter *);
>>>  };
>>>
>>>  /**** CORE ****/
>>> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>>> index 7bbcc8093e51..23cb8ed638c4 100644
>>> --- a/net/vmw_vsock/virtio_transport.c
>>> +++ b/net/vmw_vsock/virtio_transport.c
>>> @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
>>>      queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>>>  }
>>>
>>> +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
>>> +{
>>> +    struct virtio_vsock *vsock;
>>> +    bool res = false;
>>> +
>>> +    rcu_read_lock();
>>> +
>>> +    vsock = rcu_dereference(the_virtio_vsock);
>>> +    if (vsock) {
>>> +        struct virtqueue *vq;
>>> +        int iov_pages;
>>> +
>>> +        vq = vsock->vqs[VSOCK_VQ_TX];
>>> +
>>> +        iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
>>> +
>>> +        /* Check that tx queue is large enough to keep whole
>>> +         * data to send. This is needed, because when there is
>>> +         * not enough free space in the queue, current skb to
>>> +         * send will be reinserted to the head of tx list of
>>> +         * the socket to retry transmission later, so if skb
>>> +         * is bigger than whole queue, it will be reinserted
>>> +         * again and again, thus blocking other skbs to be sent.
>>> +         * Each page of the user provided buffer will be added
>>> +         * as a single buffer to the tx virtqueue, so compare
>>> +         * number of pages against maximum capacity of the queue.
>>> +         * +1 means buffer for the packet header.
>>> +         */
>>> +        if (iov_pages + 1 <= vq->num_max)
>>
>> I think this check is actual only for case one we don't have indirect buffer feature.
>> With indirect mode whole data to send will be packed into one indirect buffer.
> 
> I think so.
> So, should we check also that here?
> 
>>
>> Thanks, Arseniy
>>
>>> +            res = true;
>>> +    }
>>> +
>>> +    rcu_read_unlock();
>>> +
>>> +    return res;
>>> +}
>>> +
>>>  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
>>>
>>>  static struct virtio_transport virtio_transport = {
>>> @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
>>>          .seqpacket_allow          = virtio_transport_seqpacket_allow,
>>>          .seqpacket_has_data       = virtio_transport_seqpacket_has_data,
>>>
>>> +        .msgzerocopy_check_iov      = virtio_transport_msgzerocopy_check_iov,
>>> +
>>>          .notify_poll_in           = virtio_transport_notify_poll_in,
>>>          .notify_poll_out          = virtio_transport_notify_poll_out,
>>>          .notify_recv_init         = virtio_transport_notify_recv_init,
>>> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
>>> index 26a4d10da205..e4e3d541aff4 100644
>>> --- a/net/vmw_vsock/virtio_transport_common.c
>>> +++ b/net/vmw_vsock/virtio_transport_common.c
>>> @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
>>>      return container_of(t, struct virtio_transport, transport);
>>>  }
>>>
>>> -/* Returns a new packet on success, otherwise returns NULL.
>>> - *
>>> - * If NULL is returned, errp is set to a negative errno.
>>> - */
>>> -static struct sk_buff *
>>> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
>>> -               size_t len,
>>> -               u32 src_cid,
>>> -               u32 src_port,
>>> -               u32 dst_cid,
>>> -               u32 dst_port)
>>> -{
>>> -    const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
>>> -    struct virtio_vsock_hdr *hdr;
>>> -    struct sk_buff *skb;
>>> -    void *payload;
>>> -    int err;
>>> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
>>> +                       size_t max_to_send)
>>> +{
>>> +    const struct vsock_transport *t;
>>> +    struct iov_iter *iov_iter;
>>>
>>> -    skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>> -    if (!skb)
>>> -        return NULL;
>>> +    if (!info->msg)
>>> +        return false;
>>>
>>> -    hdr = virtio_vsock_hdr(skb);
>>> -    hdr->type    = cpu_to_le16(info->type);
>>> -    hdr->op        = cpu_to_le16(info->op);
>>> -    hdr->src_cid    = cpu_to_le64(src_cid);
>>> -    hdr->dst_cid    = cpu_to_le64(dst_cid);
>>> -    hdr->src_port    = cpu_to_le32(src_port);
>>> -    hdr->dst_port    = cpu_to_le32(dst_port);
>>> -    hdr->flags    = cpu_to_le32(info->flags);
>>> -    hdr->len    = cpu_to_le32(len);
>>> +    iov_iter = &info->msg->msg_iter;
>>>
>>> -    if (info->msg && len > 0) {
>>> -        payload = skb_put(skb, len);
>>> -        err = memcpy_from_msg(payload, info->msg, len);
>>> -        if (err)
>>> -            goto out;
>>> +    t = vsock_core_get_transport(info->vsk);
>>>
>>> -        if (msg_data_left(info->msg) == 0 &&
>>> -            info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>> -            hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>> +    if (t->msgzerocopy_check_iov &&
>>> +        !t->msgzerocopy_check_iov(iov_iter))
>>> +        return false;
> 
> I'd avoid adding a new transport callback used only internally in virtio
> transports.

Ok, I see.

> 
> Usually the transport callbacks are used in af_vsock.c, if we need a
> callback just for virtio transports, maybe better to add it in struct
> virtio_vsock_pkt_info or struct virtio_vsock_sock.
> 
> Maybe better the last one so we don't have to allocate pointer space
> for each packet and you should reach it via info.

Ok, thanks, I'll try this way

Thanks, Arseniy

> 
> Thanks,
> Stefano
> 
>>>
>>> -            if (info->msg->msg_flags & MSG_EOR)
>>> -                hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>> -        }
>>> +    /* Data is simple buffer. */
>>> +    if (iter_is_ubuf(iov_iter))
>>> +        return true;
>>> +
>>> +    if (!iter_is_iovec(iov_iter))
>>> +        return false;
>>> +
>>> +    if (iov_iter->iov_offset)
>>> +        return false;
>>> +
>>> +    /* We can't send whole iov. */
>>> +    if (iov_iter->count > max_to_send)
>>> +        return false;
>>> +
>>> +    return true;
>>> +}
>>> +
>>> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
>>> +                       struct sk_buff *skb,
>>> +                       struct msghdr *msg,
>>> +                       bool zerocopy)
>>> +{
>>> +    struct ubuf_info *uarg;
>>> +
>>> +    if (msg->msg_ubuf) {
>>> +        uarg = msg->msg_ubuf;
>>> +        net_zcopy_get(uarg);
>>> +    } else {
>>> +        struct iov_iter *iter = &msg->msg_iter;
>>> +        struct ubuf_info_msgzc *uarg_zc;
>>> +        int len;
>>> +
>>> +        /* Only ITER_IOVEC or ITER_UBUF are allowed and
>>> +         * checked before.
>>> +         */
>>> +        if (iter_is_iovec(iter))
>>> +            len = iov_length(iter->__iov, iter->nr_segs);
>>> +        else
>>> +            len = iter->count;
>>> +
>>> +        uarg = msg_zerocopy_realloc(sk_vsock(vsk),
>>> +                        len,
>>> +                        NULL);
>>> +        if (!uarg)
>>> +            return -1;
>>> +
>>> +        uarg_zc = uarg_to_msgzc(uarg);
>>> +        uarg_zc->zerocopy = zerocopy ? 1 : 0;
>>>      }
>>>
>>> -    if (info->reply)
>>> -        virtio_vsock_skb_set_reply(skb);
>>> +    skb_zcopy_init(skb, uarg);
>>>
>>> -    trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>> -                     dst_cid, dst_port,
>>> -                     len,
>>> -                     info->type,
>>> -                     info->op,
>>> -                     info->flags);
>>> +    return 0;
>>> +}
>>>
>>> -    if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
>>> -        WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
>>> -        goto out;
>>> +static int virtio_transport_fill_skb(struct sk_buff *skb,
>>> +                     struct virtio_vsock_pkt_info *info,
>>> +                     size_t len,
>>> +                     bool zcopy)
>>> +{
>>> +    if (zcopy) {
>>> +        return __zerocopy_sg_from_iter(info->msg, NULL, skb,
>>> +                          &info->msg->msg_iter,
>>> +                          len);
>>> +    } else {
>>> +        void *payload;
>>> +        int err;
>>> +
>>> +        payload = skb_put(skb, len);
>>> +        err = memcpy_from_msg(payload, info->msg, len);
>>> +        if (err)
>>> +            return -1;
>>> +
>>> +        if (msg_data_left(info->msg))
>>> +            return 0;
>>> +
>>> +        return 0;
>>>      }
>>> +}
>>>
>>> -    return skb;
>>> +static void virtio_transport_init_hdr(struct sk_buff *skb,
>>> +                      struct virtio_vsock_pkt_info *info,
>>> +                      u32 src_cid,
>>> +                      u32 src_port,
>>> +                      u32 dst_cid,
>>> +                      u32 dst_port,
>>> +                      size_t len)
>>> +{
>>> +    struct virtio_vsock_hdr *hdr;
>>>
>>> -out:
>>> -    kfree_skb(skb);
>>> -    return NULL;
>>> +    hdr = virtio_vsock_hdr(skb);
>>> +    hdr->type    = cpu_to_le16(info->type);
>>> +    hdr->op        = cpu_to_le16(info->op);
>>> +    hdr->src_cid    = cpu_to_le64(src_cid);
>>> +    hdr->dst_cid    = cpu_to_le64(dst_cid);
>>> +    hdr->src_port    = cpu_to_le32(src_port);
>>> +    hdr->dst_port    = cpu_to_le32(dst_port);
>>> +    hdr->flags    = cpu_to_le32(info->flags);
>>> +    hdr->len    = cpu_to_le32(len);
>>>  }
>>>
>>>  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
>>> @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>          return VIRTIO_VSOCK_TYPE_SEQPACKET;
>>>  }
>>>
>>> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
>>> +                          struct virtio_vsock_pkt_info *info,
>>> +                          size_t payload_len,
>>> +                          bool zcopy,
>>> +                          u32 src_cid,
>>> +                          u32 src_port,
>>> +                          u32 dst_cid,
>>> +                          u32 dst_port)
>>> +{
>>> +    struct sk_buff *skb;
>>> +    size_t skb_len;
>>> +
>>> +    skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
>>> +
>>> +    if (!zcopy)
>>> +        skb_len += payload_len;
>>> +
>>> +    skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>> +    if (!skb)
>>> +        return NULL;
>>> +
>>> +    virtio_transport_init_hdr(skb, info, src_cid, src_port,
>>> +                  dst_cid, dst_port,
>>> +                  payload_len);
>>> +
>>> +    /* Set owner here, because '__zerocopy_sg_from_iter()' uses
>>> +     * owner of skb without check to update 'sk_wmem_alloc'.
>>> +     */
>>> +    if (vsk)
>>> +        skb_set_owner_w(skb, sk_vsock(vsk));
>>> +
>>> +    if (info->msg && payload_len > 0) {
>>> +        int err;
>>> +
>>> +        err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
>>> +        if (err)
>>> +            goto out;
>>> +
>>> +        if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>> +            struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
>>> +
>>> +            hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>> +
>>> +            if (info->msg->msg_flags & MSG_EOR)
>>> +                hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>> +        }
>>> +    }
>>> +
>>> +    if (info->reply)
>>> +        virtio_vsock_skb_set_reply(skb);
>>> +
>>> +    trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>> +                     dst_cid, dst_port,
>>> +                     payload_len,
>>> +                     info->type,
>>> +                     info->op,
>>> +                     info->flags);
>>> +
>>> +    return skb;
>>> +out:
>>> +    kfree_skb(skb);
>>> +    return NULL;
>>> +}
>>> +
>>>  /* This function can only be used on connecting/connected sockets,
>>>   * since a socket assigned to a transport is required.
>>>   *
>>> @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>                        struct virtio_vsock_pkt_info *info)
>>>  {
>>> +    u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
>>>      u32 src_cid, src_port, dst_cid, dst_port;
>>>      const struct virtio_transport *t_ops;
>>>      struct virtio_vsock_sock *vvs;
>>>      u32 pkt_len = info->pkt_len;
>>> +    bool can_zcopy = false;
>>>      u32 rest_len;
>>>      int ret;
>>>
>>> @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>      if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
>>>          return pkt_len;
>>>
>>> +    if (info->msg) {
>>> +        /* If zerocopy is not enabled by 'setsockopt()', we behave as
>>> +         * there is no MSG_ZEROCOPY flag set.
>>> +         */
>>> +        if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
>>> +            info->msg->msg_flags &= ~MSG_ZEROCOPY;
>>> +
>>> +        if (info->msg->msg_flags & MSG_ZEROCOPY)
>>> +            can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
>>> +
>>> +        if (can_zcopy)
>>> +            max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
>>> +                        (MAX_SKB_FRAGS * PAGE_SIZE));
>>> +    }
>>> +
>>>      rest_len = pkt_len;
>>>
>>>      do {
>>>          struct sk_buff *skb;
>>>          size_t skb_len;
>>>
>>> -        skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
>>> +        skb_len = min(max_skb_len, rest_len);
>>>
>>> -        skb = virtio_transport_alloc_skb(info, skb_len,
>>> +        skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
>>>                           src_cid, src_port,
>>>                           dst_cid, dst_port);
>>>          if (!skb) {
>>> @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>              break;
>>>          }
>>>
>>> +        /* This is last skb to send this portion of data. */
>>> +        if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
>>> +            skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
>>> +            if (virtio_transport_init_zcopy_skb(vsk, skb,
>>> +                                info->msg,
>>> +                                can_zcopy)) {
>>> +                ret = -ENOMEM;
>>> +                break;
>>> +            }
>>> +        }
>>> +
>>>          virtio_transport_inc_tx_pkt(vvs, skb);
>>>
>>>          ret = t_ops->send_pkt(skb);
>>> @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
>>>      if (!t)
>>>          return -ENOTCONN;
>>>
>>> -    reply = virtio_transport_alloc_skb(&info, 0,
>>> +    reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
>>>                         le64_to_cpu(hdr->dst_cid),
>>>                         le32_to_cpu(hdr->dst_port),
>>>                         le64_to_cpu(hdr->src_cid),
>>
>
  
Arseniy Krasnov July 25, 2023, 9:16 a.m. UTC | #6
On 25.07.2023 11:46, Arseniy Krasnov wrote:
> 
> 
> On 25.07.2023 11:43, Stefano Garzarella wrote:
>> On Fri, Jul 21, 2023 at 08:09:03AM +0300, Arseniy Krasnov wrote:
>>>
>>>
>>> On 21.07.2023 00:42, Arseniy Krasnov wrote:
>>>> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
>>>> flag is set and zerocopy transmission is possible (enabled in socket
>>>> options and transport allows zerocopy), then non-linear skb will be
>>>> created and filled with the pages of user's buffer. Pages of user's
>>>> buffer are locked in memory by 'get_user_pages()'. Second thing that
>>>> this patch does is replace type of skb owning: instead of calling
>>>> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
>>>> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
>>>> of socket, so to decrease this field correctly proper skb destructor is
>>>> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
>>>>
>>>> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
>>>> ---
>>>>  Changelog:
>>>>  v5(big patchset) -> v1:
>>>>   * Refactorings of 'if' conditions.
>>>>   * Remove extra blank line.
>>>>   * Remove 'frag_off' field unneeded init.
>>>>   * Add function 'virtio_transport_fill_skb()' which fills both linear
>>>>     and non-linear skb with provided data.
>>>>  v1 -> v2:
>>>>   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
>>>>  v2 -> v3:
>>>>   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
>>>>     provided 'iov_iter' with data could be sent in a zerocopy mode.
>>>>     If this callback is not set in transport - transport allows to send
>>>>     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
>>>>     then zerocopy is allowed. Reason of this callback is that in case of
>>>>     G2H transmission we insert whole skb to the tx virtio queue and such
>>>>     skb must fit to the size of the virtio queue to be sent in a single
>>>>     iteration (may be tx logic in 'virtio_transport.c' could be reworked
>>>>     as in vhost to support partial send of current skb). This callback
>>>>     will be enabled only for G2H path. For details pls see comment
>>>>     'Check that tx queue...' below.
>>>>
>>>>  include/net/af_vsock.h                  |   3 +
>>>>  net/vmw_vsock/virtio_transport.c        |  39 ++++
>>>>  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
>>>>  3 files changed, 241 insertions(+), 58 deletions(-)
>>>>
>>>> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
>>>> index 0e7504a42925..a6b346eeeb8e 100644
>>>> --- a/include/net/af_vsock.h
>>>> +++ b/include/net/af_vsock.h
>>>> @@ -177,6 +177,9 @@ struct vsock_transport {
>>>>
>>>>      /* Read a single skb */
>>>>      int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
>>>> +
>>>> +    /* Zero-copy. */
>>>> +    bool (*msgzerocopy_check_iov)(const struct iov_iter *);
>>>>  };
>>>>
>>>>  /**** CORE ****/
>>>> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>>>> index 7bbcc8093e51..23cb8ed638c4 100644
>>>> --- a/net/vmw_vsock/virtio_transport.c
>>>> +++ b/net/vmw_vsock/virtio_transport.c
>>>> @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
>>>>      queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>>>>  }
>>>>
>>>> +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
>>>> +{
>>>> +    struct virtio_vsock *vsock;
>>>> +    bool res = false;
>>>> +
>>>> +    rcu_read_lock();
>>>> +
>>>> +    vsock = rcu_dereference(the_virtio_vsock);
>>>> +    if (vsock) {
>>>> +        struct virtqueue *vq;
>>>> +        int iov_pages;
>>>> +
>>>> +        vq = vsock->vqs[VSOCK_VQ_TX];
>>>> +
>>>> +        iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
>>>> +
>>>> +        /* Check that tx queue is large enough to keep whole
>>>> +         * data to send. This is needed, because when there is
>>>> +         * not enough free space in the queue, current skb to
>>>> +         * send will be reinserted to the head of tx list of
>>>> +         * the socket to retry transmission later, so if skb
>>>> +         * is bigger than whole queue, it will be reinserted
>>>> +         * again and again, thus blocking other skbs to be sent.
>>>> +         * Each page of the user provided buffer will be added
>>>> +         * as a single buffer to the tx virtqueue, so compare
>>>> +         * number of pages against maximum capacity of the queue.
>>>> +         * +1 means buffer for the packet header.
>>>> +         */
>>>> +        if (iov_pages + 1 <= vq->num_max)
>>>
>>> I think this check is actual only for case one we don't have indirect buffer feature.
>>> With indirect mode whole data to send will be packed into one indirect buffer.
>>
>> I think so.
>> So, should we check also that here?
>>
>>>
>>> Thanks, Arseniy
>>>
>>>> +            res = true;
>>>> +    }
>>>> +
>>>> +    rcu_read_unlock();
>>>> +
>>>> +    return res;
>>>> +}
>>>> +
>>>>  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
>>>>
>>>>  static struct virtio_transport virtio_transport = {
>>>> @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
>>>>          .seqpacket_allow          = virtio_transport_seqpacket_allow,
>>>>          .seqpacket_has_data       = virtio_transport_seqpacket_has_data,
>>>>
>>>> +        .msgzerocopy_check_iov      = virtio_transport_msgzerocopy_check_iov,
>>>> +
>>>>          .notify_poll_in           = virtio_transport_notify_poll_in,
>>>>          .notify_poll_out          = virtio_transport_notify_poll_out,
>>>>          .notify_recv_init         = virtio_transport_notify_recv_init,
>>>> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
>>>> index 26a4d10da205..e4e3d541aff4 100644
>>>> --- a/net/vmw_vsock/virtio_transport_common.c
>>>> +++ b/net/vmw_vsock/virtio_transport_common.c
>>>> @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
>>>>      return container_of(t, struct virtio_transport, transport);
>>>>  }
>>>>
>>>> -/* Returns a new packet on success, otherwise returns NULL.
>>>> - *
>>>> - * If NULL is returned, errp is set to a negative errno.
>>>> - */
>>>> -static struct sk_buff *
>>>> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
>>>> -               size_t len,
>>>> -               u32 src_cid,
>>>> -               u32 src_port,
>>>> -               u32 dst_cid,
>>>> -               u32 dst_port)
>>>> -{
>>>> -    const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
>>>> -    struct virtio_vsock_hdr *hdr;
>>>> -    struct sk_buff *skb;
>>>> -    void *payload;
>>>> -    int err;
>>>> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
>>>> +                       size_t max_to_send)
>>>> +{
>>>> +    const struct vsock_transport *t;
>>>> +    struct iov_iter *iov_iter;
>>>>
>>>> -    skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>>> -    if (!skb)
>>>> -        return NULL;
>>>> +    if (!info->msg)
>>>> +        return false;
>>>>
>>>> -    hdr = virtio_vsock_hdr(skb);
>>>> -    hdr->type    = cpu_to_le16(info->type);
>>>> -    hdr->op        = cpu_to_le16(info->op);
>>>> -    hdr->src_cid    = cpu_to_le64(src_cid);
>>>> -    hdr->dst_cid    = cpu_to_le64(dst_cid);
>>>> -    hdr->src_port    = cpu_to_le32(src_port);
>>>> -    hdr->dst_port    = cpu_to_le32(dst_port);
>>>> -    hdr->flags    = cpu_to_le32(info->flags);
>>>> -    hdr->len    = cpu_to_le32(len);
>>>> +    iov_iter = &info->msg->msg_iter;
>>>>
>>>> -    if (info->msg && len > 0) {
>>>> -        payload = skb_put(skb, len);
>>>> -        err = memcpy_from_msg(payload, info->msg, len);
>>>> -        if (err)
>>>> -            goto out;
>>>> +    t = vsock_core_get_transport(info->vsk);
>>>>
>>>> -        if (msg_data_left(info->msg) == 0 &&
>>>> -            info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>>> -            hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>>> +    if (t->msgzerocopy_check_iov &&
>>>> +        !t->msgzerocopy_check_iov(iov_iter))
>>>> +        return false;
>>
>> I'd avoid adding a new transport callback used only internally in virtio
>> transports.
> 
> Ok, I see.
> 
>>
>> Usually the transport callbacks are used in af_vsock.c, if we need a
>> callback just for virtio transports, maybe better to add it in struct
>> virtio_vsock_pkt_info or struct virtio_vsock_sock.

Hm, may be I just need to move this callback from 'struct vsock_transport' to parent 'struct virtio_transport',
after 'send_pkt' callback. In this case:
1) AF_VSOCK part is not touched.
2) This callback stays in 'virtio_transport.c' and is set also in this file.
   vhost and loopback are unchanged - only 'send_pkt' still enabled in both
   files for these two transports.

Thanks, Arseniy

>>
>> Maybe better the last one so we don't have to allocate pointer space
>> for each packet and you should reach it via info.
> 
> Ok, thanks, I'll try this way
> 
> Thanks, Arseniy
> 
>>
>> Thanks,
>> Stefano
>>
>>>>
>>>> -            if (info->msg->msg_flags & MSG_EOR)
>>>> -                hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>>> -        }
>>>> +    /* Data is simple buffer. */
>>>> +    if (iter_is_ubuf(iov_iter))
>>>> +        return true;
>>>> +
>>>> +    if (!iter_is_iovec(iov_iter))
>>>> +        return false;
>>>> +
>>>> +    if (iov_iter->iov_offset)
>>>> +        return false;
>>>> +
>>>> +    /* We can't send whole iov. */
>>>> +    if (iov_iter->count > max_to_send)
>>>> +        return false;
>>>> +
>>>> +    return true;
>>>> +}
>>>> +
>>>> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
>>>> +                       struct sk_buff *skb,
>>>> +                       struct msghdr *msg,
>>>> +                       bool zerocopy)
>>>> +{
>>>> +    struct ubuf_info *uarg;
>>>> +
>>>> +    if (msg->msg_ubuf) {
>>>> +        uarg = msg->msg_ubuf;
>>>> +        net_zcopy_get(uarg);
>>>> +    } else {
>>>> +        struct iov_iter *iter = &msg->msg_iter;
>>>> +        struct ubuf_info_msgzc *uarg_zc;
>>>> +        int len;
>>>> +
>>>> +        /* Only ITER_IOVEC or ITER_UBUF are allowed and
>>>> +         * checked before.
>>>> +         */
>>>> +        if (iter_is_iovec(iter))
>>>> +            len = iov_length(iter->__iov, iter->nr_segs);
>>>> +        else
>>>> +            len = iter->count;
>>>> +
>>>> +        uarg = msg_zerocopy_realloc(sk_vsock(vsk),
>>>> +                        len,
>>>> +                        NULL);
>>>> +        if (!uarg)
>>>> +            return -1;
>>>> +
>>>> +        uarg_zc = uarg_to_msgzc(uarg);
>>>> +        uarg_zc->zerocopy = zerocopy ? 1 : 0;
>>>>      }
>>>>
>>>> -    if (info->reply)
>>>> -        virtio_vsock_skb_set_reply(skb);
>>>> +    skb_zcopy_init(skb, uarg);
>>>>
>>>> -    trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>>> -                     dst_cid, dst_port,
>>>> -                     len,
>>>> -                     info->type,
>>>> -                     info->op,
>>>> -                     info->flags);
>>>> +    return 0;
>>>> +}
>>>>
>>>> -    if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
>>>> -        WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
>>>> -        goto out;
>>>> +static int virtio_transport_fill_skb(struct sk_buff *skb,
>>>> +                     struct virtio_vsock_pkt_info *info,
>>>> +                     size_t len,
>>>> +                     bool zcopy)
>>>> +{
>>>> +    if (zcopy) {
>>>> +        return __zerocopy_sg_from_iter(info->msg, NULL, skb,
>>>> +                          &info->msg->msg_iter,
>>>> +                          len);
>>>> +    } else {
>>>> +        void *payload;
>>>> +        int err;
>>>> +
>>>> +        payload = skb_put(skb, len);
>>>> +        err = memcpy_from_msg(payload, info->msg, len);
>>>> +        if (err)
>>>> +            return -1;
>>>> +
>>>> +        if (msg_data_left(info->msg))
>>>> +            return 0;
>>>> +
>>>> +        return 0;
>>>>      }
>>>> +}
>>>>
>>>> -    return skb;
>>>> +static void virtio_transport_init_hdr(struct sk_buff *skb,
>>>> +                      struct virtio_vsock_pkt_info *info,
>>>> +                      u32 src_cid,
>>>> +                      u32 src_port,
>>>> +                      u32 dst_cid,
>>>> +                      u32 dst_port,
>>>> +                      size_t len)
>>>> +{
>>>> +    struct virtio_vsock_hdr *hdr;
>>>>
>>>> -out:
>>>> -    kfree_skb(skb);
>>>> -    return NULL;
>>>> +    hdr = virtio_vsock_hdr(skb);
>>>> +    hdr->type    = cpu_to_le16(info->type);
>>>> +    hdr->op        = cpu_to_le16(info->op);
>>>> +    hdr->src_cid    = cpu_to_le64(src_cid);
>>>> +    hdr->dst_cid    = cpu_to_le64(dst_cid);
>>>> +    hdr->src_port    = cpu_to_le32(src_port);
>>>> +    hdr->dst_port    = cpu_to_le32(dst_port);
>>>> +    hdr->flags    = cpu_to_le32(info->flags);
>>>> +    hdr->len    = cpu_to_le32(len);
>>>>  }
>>>>
>>>>  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
>>>> @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>>          return VIRTIO_VSOCK_TYPE_SEQPACKET;
>>>>  }
>>>>
>>>> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
>>>> +                          struct virtio_vsock_pkt_info *info,
>>>> +                          size_t payload_len,
>>>> +                          bool zcopy,
>>>> +                          u32 src_cid,
>>>> +                          u32 src_port,
>>>> +                          u32 dst_cid,
>>>> +                          u32 dst_port)
>>>> +{
>>>> +    struct sk_buff *skb;
>>>> +    size_t skb_len;
>>>> +
>>>> +    skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
>>>> +
>>>> +    if (!zcopy)
>>>> +        skb_len += payload_len;
>>>> +
>>>> +    skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>>> +    if (!skb)
>>>> +        return NULL;
>>>> +
>>>> +    virtio_transport_init_hdr(skb, info, src_cid, src_port,
>>>> +                  dst_cid, dst_port,
>>>> +                  payload_len);
>>>> +
>>>> +    /* Set owner here, because '__zerocopy_sg_from_iter()' uses
>>>> +     * owner of skb without check to update 'sk_wmem_alloc'.
>>>> +     */
>>>> +    if (vsk)
>>>> +        skb_set_owner_w(skb, sk_vsock(vsk));
>>>> +
>>>> +    if (info->msg && payload_len > 0) {
>>>> +        int err;
>>>> +
>>>> +        err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
>>>> +        if (err)
>>>> +            goto out;
>>>> +
>>>> +        if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>>> +            struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
>>>> +
>>>> +            hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>>> +
>>>> +            if (info->msg->msg_flags & MSG_EOR)
>>>> +                hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>>> +        }
>>>> +    }
>>>> +
>>>> +    if (info->reply)
>>>> +        virtio_vsock_skb_set_reply(skb);
>>>> +
>>>> +    trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>>> +                     dst_cid, dst_port,
>>>> +                     payload_len,
>>>> +                     info->type,
>>>> +                     info->op,
>>>> +                     info->flags);
>>>> +
>>>> +    return skb;
>>>> +out:
>>>> +    kfree_skb(skb);
>>>> +    return NULL;
>>>> +}
>>>> +
>>>>  /* This function can only be used on connecting/connected sockets,
>>>>   * since a socket assigned to a transport is required.
>>>>   *
>>>> @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>>  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>                        struct virtio_vsock_pkt_info *info)
>>>>  {
>>>> +    u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
>>>>      u32 src_cid, src_port, dst_cid, dst_port;
>>>>      const struct virtio_transport *t_ops;
>>>>      struct virtio_vsock_sock *vvs;
>>>>      u32 pkt_len = info->pkt_len;
>>>> +    bool can_zcopy = false;
>>>>      u32 rest_len;
>>>>      int ret;
>>>>
>>>> @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>      if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
>>>>          return pkt_len;
>>>>
>>>> +    if (info->msg) {
>>>> +        /* If zerocopy is not enabled by 'setsockopt()', we behave as
>>>> +         * there is no MSG_ZEROCOPY flag set.
>>>> +         */
>>>> +        if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
>>>> +            info->msg->msg_flags &= ~MSG_ZEROCOPY;
>>>> +
>>>> +        if (info->msg->msg_flags & MSG_ZEROCOPY)
>>>> +            can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
>>>> +
>>>> +        if (can_zcopy)
>>>> +            max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
>>>> +                        (MAX_SKB_FRAGS * PAGE_SIZE));
>>>> +    }
>>>> +
>>>>      rest_len = pkt_len;
>>>>
>>>>      do {
>>>>          struct sk_buff *skb;
>>>>          size_t skb_len;
>>>>
>>>> -        skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
>>>> +        skb_len = min(max_skb_len, rest_len);
>>>>
>>>> -        skb = virtio_transport_alloc_skb(info, skb_len,
>>>> +        skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
>>>>                           src_cid, src_port,
>>>>                           dst_cid, dst_port);
>>>>          if (!skb) {
>>>> @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>              break;
>>>>          }
>>>>
>>>> +        /* This is last skb to send this portion of data. */
>>>> +        if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
>>>> +            skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
>>>> +            if (virtio_transport_init_zcopy_skb(vsk, skb,
>>>> +                                info->msg,
>>>> +                                can_zcopy)) {
>>>> +                ret = -ENOMEM;
>>>> +                break;
>>>> +            }
>>>> +        }
>>>> +
>>>>          virtio_transport_inc_tx_pkt(vvs, skb);
>>>>
>>>>          ret = t_ops->send_pkt(skb);
>>>> @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
>>>>      if (!t)
>>>>          return -ENOTCONN;
>>>>
>>>> -    reply = virtio_transport_alloc_skb(&info, 0,
>>>> +    reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
>>>>                         le64_to_cpu(hdr->dst_cid),
>>>>                         le32_to_cpu(hdr->dst_port),
>>>>                         le64_to_cpu(hdr->src_cid),
>>>
>>
  
Michael S. Tsirkin July 25, 2023, 11:50 a.m. UTC | #7
On Fri, Jul 21, 2023 at 08:09:03AM +0300, Arseniy Krasnov wrote:
> 
> 
> On 21.07.2023 00:42, Arseniy Krasnov wrote:
> > This adds handling of MSG_ZEROCOPY flag on transmission path: if this
> > flag is set and zerocopy transmission is possible (enabled in socket
> > options and transport allows zerocopy), then non-linear skb will be
> > created and filled with the pages of user's buffer. Pages of user's
> > buffer are locked in memory by 'get_user_pages()'. Second thing that
> > this patch does is replace type of skb owning: instead of calling
> > 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
> > change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
> > of socket, so to decrease this field correctly proper skb destructor is
> > needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
> > 
> > Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
> > ---
> >  Changelog:
> >  v5(big patchset) -> v1:
> >   * Refactorings of 'if' conditions.
> >   * Remove extra blank line.
> >   * Remove 'frag_off' field unneeded init.
> >   * Add function 'virtio_transport_fill_skb()' which fills both linear
> >     and non-linear skb with provided data.
> >  v1 -> v2:
> >   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
> >  v2 -> v3:
> >   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
> >     provided 'iov_iter' with data could be sent in a zerocopy mode.
> >     If this callback is not set in transport - transport allows to send
> >     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
> >     then zerocopy is allowed. Reason of this callback is that in case of
> >     G2H transmission we insert whole skb to the tx virtio queue and such
> >     skb must fit to the size of the virtio queue to be sent in a single
> >     iteration (may be tx logic in 'virtio_transport.c' could be reworked
> >     as in vhost to support partial send of current skb). This callback
> >     will be enabled only for G2H path. For details pls see comment 
> >     'Check that tx queue...' below.
> > 
> >  include/net/af_vsock.h                  |   3 +
> >  net/vmw_vsock/virtio_transport.c        |  39 ++++
> >  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
> >  3 files changed, 241 insertions(+), 58 deletions(-)
> > 
> > diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
> > index 0e7504a42925..a6b346eeeb8e 100644
> > --- a/include/net/af_vsock.h
> > +++ b/include/net/af_vsock.h
> > @@ -177,6 +177,9 @@ struct vsock_transport {
> >  
> >  	/* Read a single skb */
> >  	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
> > +
> > +	/* Zero-copy. */
> > +	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
> >  };
> >  
> >  /**** CORE ****/
> > diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
> > index 7bbcc8093e51..23cb8ed638c4 100644
> > --- a/net/vmw_vsock/virtio_transport.c
> > +++ b/net/vmw_vsock/virtio_transport.c
> > @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
> >  	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
> >  }
> >  
> > +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
> > +{
> > +	struct virtio_vsock *vsock;
> > +	bool res = false;
> > +
> > +	rcu_read_lock();
> > +
> > +	vsock = rcu_dereference(the_virtio_vsock);
> > +	if (vsock) {
> > +		struct virtqueue *vq;
> > +		int iov_pages;
> > +
> > +		vq = vsock->vqs[VSOCK_VQ_TX];
> > +
> > +		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
> > +
> > +		/* Check that tx queue is large enough to keep whole
> > +		 * data to send. This is needed, because when there is
> > +		 * not enough free space in the queue, current skb to
> > +		 * send will be reinserted to the head of tx list of
> > +		 * the socket to retry transmission later, so if skb
> > +		 * is bigger than whole queue, it will be reinserted
> > +		 * again and again, thus blocking other skbs to be sent.
> > +		 * Each page of the user provided buffer will be added
> > +		 * as a single buffer to the tx virtqueue, so compare
> > +		 * number of pages against maximum capacity of the queue.
> > +		 * +1 means buffer for the packet header.
> > +		 */
> > +		if (iov_pages + 1 <= vq->num_max)
> 
> I think this check is actual only for case one we don't have indirect buffer feature.
> With indirect mode whole data to send will be packed into one indirect buffer.
> 
> Thanks, Arseniy

Actually the reverse. With indirect you are limited to num_max.
Without you are limited to whatever space is left in the
queue (which you did not check here, so you should).


> > +			res = true;
> > +	}
> > +
> > +	rcu_read_unlock();

Just curious:
is the point of all this RCU dance to allow vsock
to change from under us? then why is it ok to
have it change? the virtio_transport_msgzerocopy_check_iov
will then refer to the old vsock ...


> > +
> > +	return res;
> > +}
> > +
> >  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
> >  
> >  static struct virtio_transport virtio_transport = {
> > @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
> >  		.seqpacket_allow          = virtio_transport_seqpacket_allow,
> >  		.seqpacket_has_data       = virtio_transport_seqpacket_has_data,
> >  
> > +		.msgzerocopy_check_iov	  = virtio_transport_msgzerocopy_check_iov,
> > +
> >  		.notify_poll_in           = virtio_transport_notify_poll_in,
> >  		.notify_poll_out          = virtio_transport_notify_poll_out,
> >  		.notify_recv_init         = virtio_transport_notify_recv_init,
> > diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
> > index 26a4d10da205..e4e3d541aff4 100644
> > --- a/net/vmw_vsock/virtio_transport_common.c
> > +++ b/net/vmw_vsock/virtio_transport_common.c
> > @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
> >  	return container_of(t, struct virtio_transport, transport);
> >  }
> >  
> > -/* Returns a new packet on success, otherwise returns NULL.
> > - *
> > - * If NULL is returned, errp is set to a negative errno.
> > - */
> > -static struct sk_buff *
> > -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
> > -			   size_t len,
> > -			   u32 src_cid,
> > -			   u32 src_port,
> > -			   u32 dst_cid,
> > -			   u32 dst_port)
> > -{
> > -	const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
> > -	struct virtio_vsock_hdr *hdr;
> > -	struct sk_buff *skb;
> > -	void *payload;
> > -	int err;
> > +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
> > +				       size_t max_to_send)
> > +{
> > +	const struct vsock_transport *t;
> > +	struct iov_iter *iov_iter;
> >  
> > -	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
> > -	if (!skb)
> > -		return NULL;
> > +	if (!info->msg)
> > +		return false;
> >  
> > -	hdr = virtio_vsock_hdr(skb);
> > -	hdr->type	= cpu_to_le16(info->type);
> > -	hdr->op		= cpu_to_le16(info->op);
> > -	hdr->src_cid	= cpu_to_le64(src_cid);
> > -	hdr->dst_cid	= cpu_to_le64(dst_cid);
> > -	hdr->src_port	= cpu_to_le32(src_port);
> > -	hdr->dst_port	= cpu_to_le32(dst_port);
> > -	hdr->flags	= cpu_to_le32(info->flags);
> > -	hdr->len	= cpu_to_le32(len);
> > +	iov_iter = &info->msg->msg_iter;
> >  
> > -	if (info->msg && len > 0) {
> > -		payload = skb_put(skb, len);
> > -		err = memcpy_from_msg(payload, info->msg, len);
> > -		if (err)
> > -			goto out;
> > +	t = vsock_core_get_transport(info->vsk);
> >  
> > -		if (msg_data_left(info->msg) == 0 &&
> > -		    info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
> > -			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
> > +	if (t->msgzerocopy_check_iov &&
> > +	    !t->msgzerocopy_check_iov(iov_iter))
> > +		return false;
> >  
> > -			if (info->msg->msg_flags & MSG_EOR)
> > -				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
> > -		}
> > +	/* Data is simple buffer. */
> > +	if (iter_is_ubuf(iov_iter))
> > +		return true;
> > +
> > +	if (!iter_is_iovec(iov_iter))
> > +		return false;
> > +
> > +	if (iov_iter->iov_offset)
> > +		return false;
> > +
> > +	/* We can't send whole iov. */
> > +	if (iov_iter->count > max_to_send)
> > +		return false;
> > +
> > +	return true;
> > +}
> > +
> > +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
> > +					   struct sk_buff *skb,
> > +					   struct msghdr *msg,
> > +					   bool zerocopy)
> > +{
> > +	struct ubuf_info *uarg;
> > +
> > +	if (msg->msg_ubuf) {
> > +		uarg = msg->msg_ubuf;
> > +		net_zcopy_get(uarg);
> > +	} else {
> > +		struct iov_iter *iter = &msg->msg_iter;
> > +		struct ubuf_info_msgzc *uarg_zc;
> > +		int len;
> > +
> > +		/* Only ITER_IOVEC or ITER_UBUF are allowed and
> > +		 * checked before.
> > +		 */
> > +		if (iter_is_iovec(iter))
> > +			len = iov_length(iter->__iov, iter->nr_segs);
> > +		else
> > +			len = iter->count;
> > +
> > +		uarg = msg_zerocopy_realloc(sk_vsock(vsk),
> > +					    len,
> > +					    NULL);
> > +		if (!uarg)
> > +			return -1;
> > +
> > +		uarg_zc = uarg_to_msgzc(uarg);
> > +		uarg_zc->zerocopy = zerocopy ? 1 : 0;
> >  	}
> >  
> > -	if (info->reply)
> > -		virtio_vsock_skb_set_reply(skb);
> > +	skb_zcopy_init(skb, uarg);
> >  
> > -	trace_virtio_transport_alloc_pkt(src_cid, src_port,
> > -					 dst_cid, dst_port,
> > -					 len,
> > -					 info->type,
> > -					 info->op,
> > -					 info->flags);
> > +	return 0;
> > +}
> >  
> > -	if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
> > -		WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
> > -		goto out;
> > +static int virtio_transport_fill_skb(struct sk_buff *skb,
> > +				     struct virtio_vsock_pkt_info *info,
> > +				     size_t len,
> > +				     bool zcopy)
> > +{
> > +	if (zcopy) {
> > +		return __zerocopy_sg_from_iter(info->msg, NULL, skb,
> > +					      &info->msg->msg_iter,
> > +					      len);
> > +	} else {
> > +		void *payload;
> > +		int err;
> > +
> > +		payload = skb_put(skb, len);
> > +		err = memcpy_from_msg(payload, info->msg, len);
> > +		if (err)
> > +			return -1;
> > +
> > +		if (msg_data_left(info->msg))
> > +			return 0;
> > +
> > +		return 0;
> >  	}
> > +}
> >  
> > -	return skb;
> > +static void virtio_transport_init_hdr(struct sk_buff *skb,
> > +				      struct virtio_vsock_pkt_info *info,
> > +				      u32 src_cid,
> > +				      u32 src_port,
> > +				      u32 dst_cid,
> > +				      u32 dst_port,
> > +				      size_t len)
> > +{
> > +	struct virtio_vsock_hdr *hdr;
> >  
> > -out:
> > -	kfree_skb(skb);
> > -	return NULL;
> > +	hdr = virtio_vsock_hdr(skb);
> > +	hdr->type	= cpu_to_le16(info->type);
> > +	hdr->op		= cpu_to_le16(info->op);
> > +	hdr->src_cid	= cpu_to_le64(src_cid);
> > +	hdr->dst_cid	= cpu_to_le64(dst_cid);
> > +	hdr->src_port	= cpu_to_le32(src_port);
> > +	hdr->dst_port	= cpu_to_le32(dst_port);
> > +	hdr->flags	= cpu_to_le32(info->flags);
> > +	hdr->len	= cpu_to_le32(len);
> >  }
> >  
> >  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
> > @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
> >  		return VIRTIO_VSOCK_TYPE_SEQPACKET;
> >  }
> >  
> > +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
> > +						  struct virtio_vsock_pkt_info *info,
> > +						  size_t payload_len,
> > +						  bool zcopy,
> > +						  u32 src_cid,
> > +						  u32 src_port,
> > +						  u32 dst_cid,
> > +						  u32 dst_port)
> > +{
> > +	struct sk_buff *skb;
> > +	size_t skb_len;
> > +
> > +	skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
> > +
> > +	if (!zcopy)
> > +		skb_len += payload_len;
> > +
> > +	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
> > +	if (!skb)
> > +		return NULL;
> > +
> > +	virtio_transport_init_hdr(skb, info, src_cid, src_port,
> > +				  dst_cid, dst_port,
> > +				  payload_len);
> > +
> > +	/* Set owner here, because '__zerocopy_sg_from_iter()' uses
> > +	 * owner of skb without check to update 'sk_wmem_alloc'.
> > +	 */
> > +	if (vsk)
> > +		skb_set_owner_w(skb, sk_vsock(vsk));
> > +
> > +	if (info->msg && payload_len > 0) {
> > +		int err;
> > +
> > +		err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
> > +		if (err)
> > +			goto out;
> > +
> > +		if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
> > +			struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
> > +
> > +			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
> > +
> > +			if (info->msg->msg_flags & MSG_EOR)
> > +				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
> > +		}
> > +	}
> > +
> > +	if (info->reply)
> > +		virtio_vsock_skb_set_reply(skb);
> > +
> > +	trace_virtio_transport_alloc_pkt(src_cid, src_port,
> > +					 dst_cid, dst_port,
> > +					 payload_len,
> > +					 info->type,
> > +					 info->op,
> > +					 info->flags);
> > +
> > +	return skb;
> > +out:
> > +	kfree_skb(skb);
> > +	return NULL;
> > +}
> > +
> >  /* This function can only be used on connecting/connected sockets,
> >   * since a socket assigned to a transport is required.
> >   *
> > @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
> >  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
> >  					  struct virtio_vsock_pkt_info *info)
> >  {
> > +	u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
> >  	u32 src_cid, src_port, dst_cid, dst_port;
> >  	const struct virtio_transport *t_ops;
> >  	struct virtio_vsock_sock *vvs;
> >  	u32 pkt_len = info->pkt_len;
> > +	bool can_zcopy = false;
> >  	u32 rest_len;
> >  	int ret;
> >  
> > @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
> >  	if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
> >  		return pkt_len;
> >  
> > +	if (info->msg) {
> > +		/* If zerocopy is not enabled by 'setsockopt()', we behave as
> > +		 * there is no MSG_ZEROCOPY flag set.
> > +		 */
> > +		if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
> > +			info->msg->msg_flags &= ~MSG_ZEROCOPY;
> > +
> > +		if (info->msg->msg_flags & MSG_ZEROCOPY)
> > +			can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
> > +
> > +		if (can_zcopy)
> > +			max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
> > +					    (MAX_SKB_FRAGS * PAGE_SIZE));
> > +	}
> > +
> >  	rest_len = pkt_len;
> >  
> >  	do {
> >  		struct sk_buff *skb;
> >  		size_t skb_len;
> >  
> > -		skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
> > +		skb_len = min(max_skb_len, rest_len);
> >  
> > -		skb = virtio_transport_alloc_skb(info, skb_len,
> > +		skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
> >  						 src_cid, src_port,
> >  						 dst_cid, dst_port);
> >  		if (!skb) {
> > @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
> >  			break;
> >  		}
> >  
> > +		/* This is last skb to send this portion of data. */
> > +		if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
> > +		    skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
> > +			if (virtio_transport_init_zcopy_skb(vsk, skb,
> > +							    info->msg,
> > +							    can_zcopy)) {
> > +				ret = -ENOMEM;
> > +				break;
> > +			}
> > +		}
> > +
> >  		virtio_transport_inc_tx_pkt(vvs, skb);
> >  
> >  		ret = t_ops->send_pkt(skb);
> > @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
> >  	if (!t)
> >  		return -ENOTCONN;
> >  
> > -	reply = virtio_transport_alloc_skb(&info, 0,
> > +	reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
> >  					   le64_to_cpu(hdr->dst_cid),
> >  					   le32_to_cpu(hdr->dst_port),
> >  					   le64_to_cpu(hdr->src_cid),
  
Michael S. Tsirkin July 25, 2023, 11:59 a.m. UTC | #8
On Tue, Jul 25, 2023 at 11:39:22AM +0300, Arseniy Krasnov wrote:
> 
> 
> On 25.07.2023 11:25, Michael S. Tsirkin wrote:
> > On Fri, Jul 21, 2023 at 12:42:45AM +0300, Arseniy Krasnov wrote:
> >> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
> >> flag is set and zerocopy transmission is possible (enabled in socket
> >> options and transport allows zerocopy), then non-linear skb will be
> >> created and filled with the pages of user's buffer. Pages of user's
> >> buffer are locked in memory by 'get_user_pages()'. Second thing that
> >> this patch does is replace type of skb owning: instead of calling
> >> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
> >> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
> >> of socket, so to decrease this field correctly proper skb destructor is
> >> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
> >>
> >> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
> >> ---
> >>  Changelog:
> >>  v5(big patchset) -> v1:
> >>   * Refactorings of 'if' conditions.
> >>   * Remove extra blank line.
> >>   * Remove 'frag_off' field unneeded init.
> >>   * Add function 'virtio_transport_fill_skb()' which fills both linear
> >>     and non-linear skb with provided data.
> >>  v1 -> v2:
> >>   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
> >>  v2 -> v3:
> >>   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
> >>     provided 'iov_iter' with data could be sent in a zerocopy mode.
> >>     If this callback is not set in transport - transport allows to send
> >>     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
> >>     then zerocopy is allowed. Reason of this callback is that in case of
> >>     G2H transmission we insert whole skb to the tx virtio queue and such
> >>     skb must fit to the size of the virtio queue to be sent in a single
> >>     iteration (may be tx logic in 'virtio_transport.c' could be reworked
> >>     as in vhost to support partial send of current skb). This callback
> >>     will be enabled only for G2H path. For details pls see comment 
> >>     'Check that tx queue...' below.
> >>
> >>  include/net/af_vsock.h                  |   3 +
> >>  net/vmw_vsock/virtio_transport.c        |  39 ++++
> >>  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
> >>  3 files changed, 241 insertions(+), 58 deletions(-)
> >>
> >> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
> >> index 0e7504a42925..a6b346eeeb8e 100644
> >> --- a/include/net/af_vsock.h
> >> +++ b/include/net/af_vsock.h
> >> @@ -177,6 +177,9 @@ struct vsock_transport {
> >>  
> >>  	/* Read a single skb */
> >>  	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
> >> +
> >> +	/* Zero-copy. */
> >> +	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
> >>  };
> >>  
> >>  /**** CORE ****/
> >> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
> >> index 7bbcc8093e51..23cb8ed638c4 100644
> >> --- a/net/vmw_vsock/virtio_transport.c
> >> +++ b/net/vmw_vsock/virtio_transport.c
> >> @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
> >>  	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
> >>  }
> >>  
> >> +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
> >> +{
> >> +	struct virtio_vsock *vsock;
> >> +	bool res = false;
> >> +
> >> +	rcu_read_lock();
> >> +
> >> +	vsock = rcu_dereference(the_virtio_vsock);
> >> +	if (vsock) {
> >> +		struct virtqueue *vq;
> >> +		int iov_pages;
> >> +
> >> +		vq = vsock->vqs[VSOCK_VQ_TX];
> >> +
> >> +		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
> >> +
> >> +		/* Check that tx queue is large enough to keep whole
> >> +		 * data to send. This is needed, because when there is
> >> +		 * not enough free space in the queue, current skb to
> >> +		 * send will be reinserted to the head of tx list of
> >> +		 * the socket to retry transmission later, so if skb
> >> +		 * is bigger than whole queue, it will be reinserted
> >> +		 * again and again, thus blocking other skbs to be sent.
> >> +		 * Each page of the user provided buffer will be added
> >> +		 * as a single buffer to the tx virtqueue, so compare
> >> +		 * number of pages against maximum capacity of the queue.
> >> +		 * +1 means buffer for the packet header.
> >> +		 */
> >> +		if (iov_pages + 1 <= vq->num_max)
> >> +			res = true;
> > 
> > 
> > Yes but can't there already be buffers in the queue?
> > Then you can't stick num_max there.
> 
> I think, that it is not critical, because vhost part always tries to process all
> incoming buffers (yes, 'vhost_exceeds_weight()' breaks at some moment, but it will
> reschedule tx kick ('vhost_vsock_handle_tx_kick()') work again), so current "too
> big" skb will wait until there will be enough space in queue and as it is requeued
> to the head of tx list it will be inserted to tx queue first.
> 
> But anyway, I agree that comparing to 'num_free' may be more effective to the whole
> system performance...
> 
> Thanks, Arseniy

Oh I see. It makes sense then - instead of copying just so we can
stick it in the queue, wait a bit and send later.
Also - for stream transports can't the message be split
and sent chunk by chunk? Better than copying ...


> > 
> > 
> >> +	}
> >> +
> >> +	rcu_read_unlock();
> >> +
> >> +	return res;
> >> +}
> >> +
> >>  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
> >>  
> >>  static struct virtio_transport virtio_transport = {
> >> @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
> >>  		.seqpacket_allow          = virtio_transport_seqpacket_allow,
> >>  		.seqpacket_has_data       = virtio_transport_seqpacket_has_data,
> >>  
> >> +		.msgzerocopy_check_iov	  = virtio_transport_msgzerocopy_check_iov,
> >> +
> >>  		.notify_poll_in           = virtio_transport_notify_poll_in,
> >>  		.notify_poll_out          = virtio_transport_notify_poll_out,
> >>  		.notify_recv_init         = virtio_transport_notify_recv_init,
> >> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
> >> index 26a4d10da205..e4e3d541aff4 100644
> >> --- a/net/vmw_vsock/virtio_transport_common.c
> >> +++ b/net/vmw_vsock/virtio_transport_common.c
> >> @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
> >>  	return container_of(t, struct virtio_transport, transport);
> >>  }
> >>  
> >> -/* Returns a new packet on success, otherwise returns NULL.
> >> - *
> >> - * If NULL is returned, errp is set to a negative errno.
> >> - */
> >> -static struct sk_buff *
> >> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
> >> -			   size_t len,
> >> -			   u32 src_cid,
> >> -			   u32 src_port,
> >> -			   u32 dst_cid,
> >> -			   u32 dst_port)
> >> -{
> >> -	const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
> >> -	struct virtio_vsock_hdr *hdr;
> >> -	struct sk_buff *skb;
> >> -	void *payload;
> >> -	int err;
> >> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
> >> +				       size_t max_to_send)
> >> +{
> >> +	const struct vsock_transport *t;
> >> +	struct iov_iter *iov_iter;
> >>  
> >> -	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
> >> -	if (!skb)
> >> -		return NULL;
> >> +	if (!info->msg)
> >> +		return false;
> >>  
> >> -	hdr = virtio_vsock_hdr(skb);
> >> -	hdr->type	= cpu_to_le16(info->type);
> >> -	hdr->op		= cpu_to_le16(info->op);
> >> -	hdr->src_cid	= cpu_to_le64(src_cid);
> >> -	hdr->dst_cid	= cpu_to_le64(dst_cid);
> >> -	hdr->src_port	= cpu_to_le32(src_port);
> >> -	hdr->dst_port	= cpu_to_le32(dst_port);
> >> -	hdr->flags	= cpu_to_le32(info->flags);
> >> -	hdr->len	= cpu_to_le32(len);
> >> +	iov_iter = &info->msg->msg_iter;
> >>  
> >> -	if (info->msg && len > 0) {
> >> -		payload = skb_put(skb, len);
> >> -		err = memcpy_from_msg(payload, info->msg, len);
> >> -		if (err)
> >> -			goto out;
> >> +	t = vsock_core_get_transport(info->vsk);
> >>  
> >> -		if (msg_data_left(info->msg) == 0 &&
> >> -		    info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
> >> -			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
> >> +	if (t->msgzerocopy_check_iov &&
> >> +	    !t->msgzerocopy_check_iov(iov_iter))
> >> +		return false;
> >>  
> >> -			if (info->msg->msg_flags & MSG_EOR)
> >> -				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
> >> -		}
> >> +	/* Data is simple buffer. */
> >> +	if (iter_is_ubuf(iov_iter))
> >> +		return true;
> >> +
> >> +	if (!iter_is_iovec(iov_iter))
> >> +		return false;
> >> +
> >> +	if (iov_iter->iov_offset)
> >> +		return false;
> >> +
> >> +	/* We can't send whole iov. */
> >> +	if (iov_iter->count > max_to_send)
> >> +		return false;
> >> +
> >> +	return true;
> >> +}
> >> +
> >> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
> >> +					   struct sk_buff *skb,
> >> +					   struct msghdr *msg,
> >> +					   bool zerocopy)
> >> +{
> >> +	struct ubuf_info *uarg;
> >> +
> >> +	if (msg->msg_ubuf) {
> >> +		uarg = msg->msg_ubuf;
> >> +		net_zcopy_get(uarg);
> >> +	} else {
> >> +		struct iov_iter *iter = &msg->msg_iter;
> >> +		struct ubuf_info_msgzc *uarg_zc;
> >> +		int len;
> >> +
> >> +		/* Only ITER_IOVEC or ITER_UBUF are allowed and
> >> +		 * checked before.
> >> +		 */
> >> +		if (iter_is_iovec(iter))
> >> +			len = iov_length(iter->__iov, iter->nr_segs);
> >> +		else
> >> +			len = iter->count;
> >> +
> >> +		uarg = msg_zerocopy_realloc(sk_vsock(vsk),
> >> +					    len,
> >> +					    NULL);
> >> +		if (!uarg)
> >> +			return -1;
> >> +
> >> +		uarg_zc = uarg_to_msgzc(uarg);
> >> +		uarg_zc->zerocopy = zerocopy ? 1 : 0;
> >>  	}
> >>  
> >> -	if (info->reply)
> >> -		virtio_vsock_skb_set_reply(skb);
> >> +	skb_zcopy_init(skb, uarg);
> >>  
> >> -	trace_virtio_transport_alloc_pkt(src_cid, src_port,
> >> -					 dst_cid, dst_port,
> >> -					 len,
> >> -					 info->type,
> >> -					 info->op,
> >> -					 info->flags);
> >> +	return 0;
> >> +}
> >>  
> >> -	if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
> >> -		WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
> >> -		goto out;
> >> +static int virtio_transport_fill_skb(struct sk_buff *skb,
> >> +				     struct virtio_vsock_pkt_info *info,
> >> +				     size_t len,
> >> +				     bool zcopy)
> >> +{
> >> +	if (zcopy) {
> >> +		return __zerocopy_sg_from_iter(info->msg, NULL, skb,
> >> +					      &info->msg->msg_iter,
> >> +					      len);
> >> +	} else {
> >> +		void *payload;
> >> +		int err;
> >> +
> >> +		payload = skb_put(skb, len);
> >> +		err = memcpy_from_msg(payload, info->msg, len);
> >> +		if (err)
> >> +			return -1;
> >> +
> >> +		if (msg_data_left(info->msg))
> >> +			return 0;
> >> +
> >> +		return 0;
> >>  	}
> >> +}
> >>  
> >> -	return skb;
> >> +static void virtio_transport_init_hdr(struct sk_buff *skb,
> >> +				      struct virtio_vsock_pkt_info *info,
> >> +				      u32 src_cid,
> >> +				      u32 src_port,
> >> +				      u32 dst_cid,
> >> +				      u32 dst_port,
> >> +				      size_t len)
> >> +{
> >> +	struct virtio_vsock_hdr *hdr;
> >>  
> >> -out:
> >> -	kfree_skb(skb);
> >> -	return NULL;
> >> +	hdr = virtio_vsock_hdr(skb);
> >> +	hdr->type	= cpu_to_le16(info->type);
> >> +	hdr->op		= cpu_to_le16(info->op);
> >> +	hdr->src_cid	= cpu_to_le64(src_cid);
> >> +	hdr->dst_cid	= cpu_to_le64(dst_cid);
> >> +	hdr->src_port	= cpu_to_le32(src_port);
> >> +	hdr->dst_port	= cpu_to_le32(dst_port);
> >> +	hdr->flags	= cpu_to_le32(info->flags);
> >> +	hdr->len	= cpu_to_le32(len);
> >>  }
> >>  
> >>  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
> >> @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
> >>  		return VIRTIO_VSOCK_TYPE_SEQPACKET;
> >>  }
> >>  
> >> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
> >> +						  struct virtio_vsock_pkt_info *info,
> >> +						  size_t payload_len,
> >> +						  bool zcopy,
> >> +						  u32 src_cid,
> >> +						  u32 src_port,
> >> +						  u32 dst_cid,
> >> +						  u32 dst_port)
> >> +{
> >> +	struct sk_buff *skb;
> >> +	size_t skb_len;
> >> +
> >> +	skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
> >> +
> >> +	if (!zcopy)
> >> +		skb_len += payload_len;
> >> +
> >> +	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
> >> +	if (!skb)
> >> +		return NULL;
> >> +
> >> +	virtio_transport_init_hdr(skb, info, src_cid, src_port,
> >> +				  dst_cid, dst_port,
> >> +				  payload_len);
> >> +
> >> +	/* Set owner here, because '__zerocopy_sg_from_iter()' uses
> >> +	 * owner of skb without check to update 'sk_wmem_alloc'.
> >> +	 */
> >> +	if (vsk)
> >> +		skb_set_owner_w(skb, sk_vsock(vsk));
> >> +
> >> +	if (info->msg && payload_len > 0) {
> >> +		int err;
> >> +
> >> +		err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
> >> +		if (err)
> >> +			goto out;
> >> +
> >> +		if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
> >> +			struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
> >> +
> >> +			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
> >> +
> >> +			if (info->msg->msg_flags & MSG_EOR)
> >> +				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
> >> +		}
> >> +	}
> >> +
> >> +	if (info->reply)
> >> +		virtio_vsock_skb_set_reply(skb);
> >> +
> >> +	trace_virtio_transport_alloc_pkt(src_cid, src_port,
> >> +					 dst_cid, dst_port,
> >> +					 payload_len,
> >> +					 info->type,
> >> +					 info->op,
> >> +					 info->flags);
> >> +
> >> +	return skb;
> >> +out:
> >> +	kfree_skb(skb);
> >> +	return NULL;
> >> +}
> >> +
> >>  /* This function can only be used on connecting/connected sockets,
> >>   * since a socket assigned to a transport is required.
> >>   *
> >> @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
> >>  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
> >>  					  struct virtio_vsock_pkt_info *info)
> >>  {
> >> +	u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
> >>  	u32 src_cid, src_port, dst_cid, dst_port;
> >>  	const struct virtio_transport *t_ops;
> >>  	struct virtio_vsock_sock *vvs;
> >>  	u32 pkt_len = info->pkt_len;
> >> +	bool can_zcopy = false;
> >>  	u32 rest_len;
> >>  	int ret;
> >>  
> >> @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
> >>  	if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
> >>  		return pkt_len;
> >>  
> >> +	if (info->msg) {
> >> +		/* If zerocopy is not enabled by 'setsockopt()', we behave as
> >> +		 * there is no MSG_ZEROCOPY flag set.
> >> +		 */
> >> +		if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
> >> +			info->msg->msg_flags &= ~MSG_ZEROCOPY;
> >> +
> >> +		if (info->msg->msg_flags & MSG_ZEROCOPY)
> >> +			can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
> >> +
> >> +		if (can_zcopy)
> >> +			max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
> >> +					    (MAX_SKB_FRAGS * PAGE_SIZE));
> >> +	}
> >> +
> >>  	rest_len = pkt_len;
> >>  
> >>  	do {
> >>  		struct sk_buff *skb;
> >>  		size_t skb_len;
> >>  
> >> -		skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
> >> +		skb_len = min(max_skb_len, rest_len);
> >>  
> >> -		skb = virtio_transport_alloc_skb(info, skb_len,
> >> +		skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
> >>  						 src_cid, src_port,
> >>  						 dst_cid, dst_port);
> >>  		if (!skb) {
> >> @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
> >>  			break;
> >>  		}
> >>  
> >> +		/* This is last skb to send this portion of data. */
> >> +		if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
> >> +		    skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
> >> +			if (virtio_transport_init_zcopy_skb(vsk, skb,
> >> +							    info->msg,
> >> +							    can_zcopy)) {
> >> +				ret = -ENOMEM;
> >> +				break;
> >> +			}
> >> +		}
> >> +
> >>  		virtio_transport_inc_tx_pkt(vvs, skb);
> >>  
> >>  		ret = t_ops->send_pkt(skb);
> >> @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
> >>  	if (!t)
> >>  		return -ENOTCONN;
> >>  
> >> -	reply = virtio_transport_alloc_skb(&info, 0,
> >> +	reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
> >>  					   le64_to_cpu(hdr->dst_cid),
> >>  					   le32_to_cpu(hdr->dst_port),
> >>  					   le64_to_cpu(hdr->src_cid),
> >> -- 
> >> 2.25.1
> >
  
Stefano Garzarella July 25, 2023, 12:28 p.m. UTC | #9
On Tue, Jul 25, 2023 at 12:16:11PM +0300, Arseniy Krasnov wrote:
>
>
>On 25.07.2023 11:46, Arseniy Krasnov wrote:
>>
>>
>> On 25.07.2023 11:43, Stefano Garzarella wrote:
>>> On Fri, Jul 21, 2023 at 08:09:03AM +0300, Arseniy Krasnov wrote:
>>>>
>>>>
>>>> On 21.07.2023 00:42, Arseniy Krasnov wrote:
>>>>> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
>>>>> flag is set and zerocopy transmission is possible (enabled in socket
>>>>> options and transport allows zerocopy), then non-linear skb will be
>>>>> created and filled with the pages of user's buffer. Pages of user's
>>>>> buffer are locked in memory by 'get_user_pages()'. Second thing that
>>>>> this patch does is replace type of skb owning: instead of calling
>>>>> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
>>>>> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
>>>>> of socket, so to decrease this field correctly proper skb destructor is
>>>>> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
>>>>>
>>>>> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
>>>>> ---
>>>>>  Changelog:
>>>>>  v5(big patchset) -> v1:
>>>>>   * Refactorings of 'if' conditions.
>>>>>   * Remove extra blank line.
>>>>>   * Remove 'frag_off' field unneeded init.
>>>>>   * Add function 'virtio_transport_fill_skb()' which fills both linear
>>>>>     and non-linear skb with provided data.
>>>>>  v1 -> v2:
>>>>>   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
>>>>>  v2 -> v3:
>>>>>   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
>>>>>     provided 'iov_iter' with data could be sent in a zerocopy mode.
>>>>>     If this callback is not set in transport - transport allows to send
>>>>>     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
>>>>>     then zerocopy is allowed. Reason of this callback is that in case of
>>>>>     G2H transmission we insert whole skb to the tx virtio queue and such
>>>>>     skb must fit to the size of the virtio queue to be sent in a single
>>>>>     iteration (may be tx logic in 'virtio_transport.c' could be reworked
>>>>>     as in vhost to support partial send of current skb). This callback
>>>>>     will be enabled only for G2H path. For details pls see comment
>>>>>     'Check that tx queue...' below.
>>>>>
>>>>>  include/net/af_vsock.h                  |   3 +
>>>>>  net/vmw_vsock/virtio_transport.c        |  39 ++++
>>>>>  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
>>>>>  3 files changed, 241 insertions(+), 58 deletions(-)
>>>>>
>>>>> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
>>>>> index 0e7504a42925..a6b346eeeb8e 100644
>>>>> --- a/include/net/af_vsock.h
>>>>> +++ b/include/net/af_vsock.h
>>>>> @@ -177,6 +177,9 @@ struct vsock_transport {
>>>>>
>>>>>      /* Read a single skb */
>>>>>      int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
>>>>> +
>>>>> +    /* Zero-copy. */
>>>>> +    bool (*msgzerocopy_check_iov)(const struct iov_iter *);
>>>>>  };
>>>>>
>>>>>  /**** CORE ****/
>>>>> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>>>>> index 7bbcc8093e51..23cb8ed638c4 100644
>>>>> --- a/net/vmw_vsock/virtio_transport.c
>>>>> +++ b/net/vmw_vsock/virtio_transport.c
>>>>> @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
>>>>>      queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>>>>>  }
>>>>>
>>>>> +static bool virtio_transport_msgzerocopy_check_iov(const struct 
>>>>> iov_iter *iov)
>>>>> +{
>>>>> +    struct virtio_vsock *vsock;
>>>>> +    bool res = false;
>>>>> +
>>>>> +    rcu_read_lock();
>>>>> +
>>>>> +    vsock = rcu_dereference(the_virtio_vsock);
>>>>> +    if (vsock) {

Just noted, what about the following to reduce the indentation?

         if (!vsock) {
             goto out;
         }
             ...
             ...
     out:
         rcu_read_unlock();
         return res;

>>>>> +        struct virtqueue *vq;
>>>>> +        int iov_pages;
>>>>> +
>>>>> +        vq = vsock->vqs[VSOCK_VQ_TX];
>>>>> +
>>>>> +        iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
>>>>> +
>>>>> +        /* Check that tx queue is large enough to keep whole
>>>>> +         * data to send. This is needed, because when there is
>>>>> +         * not enough free space in the queue, current skb to
>>>>> +         * send will be reinserted to the head of tx list of
>>>>> +         * the socket to retry transmission later, so if skb
>>>>> +         * is bigger than whole queue, it will be reinserted
>>>>> +         * again and again, thus blocking other skbs to be sent.
>>>>> +         * Each page of the user provided buffer will be added
>>>>> +         * as a single buffer to the tx virtqueue, so compare
>>>>> +         * number of pages against maximum capacity of the queue.
>>>>> +         * +1 means buffer for the packet header.
>>>>> +         */
>>>>> +        if (iov_pages + 1 <= vq->num_max)
>>>>
>>>> I think this check is actual only for case one we don't have indirect buffer feature.
>>>> With indirect mode whole data to send will be packed into one indirect buffer.
>>>
>>> I think so.
>>> So, should we check also that here?
>>>
>>>>
>>>> Thanks, Arseniy
>>>>
>>>>> +            res = true;
>>>>> +    }
>>>>> +
>>>>> +    rcu_read_unlock();
>>>>> +
>>>>> +    return res;
>>>>> +}
>>>>> +
>>>>>  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
>>>>>
>>>>>  static struct virtio_transport virtio_transport = {
>>>>> @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
>>>>>          .seqpacket_allow          = virtio_transport_seqpacket_allow,
>>>>>          .seqpacket_has_data       = virtio_transport_seqpacket_has_data,
>>>>>
>>>>> +        .msgzerocopy_check_iov      = virtio_transport_msgzerocopy_check_iov,
>>>>> +
>>>>>          .notify_poll_in           = virtio_transport_notify_poll_in,
>>>>>          .notify_poll_out          = virtio_transport_notify_poll_out,
>>>>>          .notify_recv_init         = virtio_transport_notify_recv_init,
>>>>> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
>>>>> index 26a4d10da205..e4e3d541aff4 100644
>>>>> --- a/net/vmw_vsock/virtio_transport_common.c
>>>>> +++ b/net/vmw_vsock/virtio_transport_common.c
>>>>> @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
>>>>>      return container_of(t, struct virtio_transport, transport);
>>>>>  }
>>>>>
>>>>> -/* Returns a new packet on success, otherwise returns NULL.
>>>>> - *
>>>>> - * If NULL is returned, errp is set to a negative errno.
>>>>> - */
>>>>> -static struct sk_buff *
>>>>> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
>>>>> -               size_t len,
>>>>> -               u32 src_cid,
>>>>> -               u32 src_port,
>>>>> -               u32 dst_cid,
>>>>> -               u32 dst_port)
>>>>> -{
>>>>> -    const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
>>>>> -    struct virtio_vsock_hdr *hdr;
>>>>> -    struct sk_buff *skb;
>>>>> -    void *payload;
>>>>> -    int err;
>>>>> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
>>>>> +                       size_t max_to_send)
>>>>> +{
>>>>> +    const struct vsock_transport *t;
>>>>> +    struct iov_iter *iov_iter;
>>>>>
>>>>> -    skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>>>> -    if (!skb)
>>>>> -        return NULL;
>>>>> +    if (!info->msg)
>>>>> +        return false;
>>>>>
>>>>> -    hdr = virtio_vsock_hdr(skb);
>>>>> -    hdr->type    = cpu_to_le16(info->type);
>>>>> -    hdr->op        = cpu_to_le16(info->op);
>>>>> -    hdr->src_cid    = cpu_to_le64(src_cid);
>>>>> -    hdr->dst_cid    = cpu_to_le64(dst_cid);
>>>>> -    hdr->src_port    = cpu_to_le32(src_port);
>>>>> -    hdr->dst_port    = cpu_to_le32(dst_port);
>>>>> -    hdr->flags    = cpu_to_le32(info->flags);
>>>>> -    hdr->len    = cpu_to_le32(len);
>>>>> +    iov_iter = &info->msg->msg_iter;
>>>>>
>>>>> -    if (info->msg && len > 0) {
>>>>> -        payload = skb_put(skb, len);
>>>>> -        err = memcpy_from_msg(payload, info->msg, len);
>>>>> -        if (err)
>>>>> -            goto out;
>>>>> +    t = vsock_core_get_transport(info->vsk);
>>>>>
>>>>> -        if (msg_data_left(info->msg) == 0 &&
>>>>> -            info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>>>> -            hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>>>> +    if (t->msgzerocopy_check_iov &&
>>>>> +        !t->msgzerocopy_check_iov(iov_iter))
>>>>> +        return false;
>>>
>>> I'd avoid adding a new transport callback used only internally in virtio
>>> transports.
>>
>> Ok, I see.
>>
>>>
>>> Usually the transport callbacks are used in af_vsock.c, if we need a
>>> callback just for virtio transports, maybe better to add it in struct
>>> virtio_vsock_pkt_info or struct virtio_vsock_sock.
>
>Hm, may be I just need to move this callback from 'struct vsock_transport' to parent 'struct virtio_transport',
>after 'send_pkt' callback. In this case:
>1) AF_VSOCK part is not touched.
>2) This callback stays in 'virtio_transport.c' and is set also in this file.
>   vhost and loopback are unchanged - only 'send_pkt' still enabled in both
>   files for these two transports.

Yep, this could also work!

Stefano

>
>Thanks, Arseniy
>
>>>
>>> Maybe better the last one so we don't have to allocate pointer space
>>> for each packet and you should reach it via info.
>>
>> Ok, thanks, I'll try this way
>>
>> Thanks, Arseniy
>>
>>>
>>> Thanks,
>>> Stefano
>>>
>>>>>
>>>>> -            if (info->msg->msg_flags & MSG_EOR)
>>>>> -                hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>>>> -        }
>>>>> +    /* Data is simple buffer. */
>>>>> +    if (iter_is_ubuf(iov_iter))
>>>>> +        return true;
>>>>> +
>>>>> +    if (!iter_is_iovec(iov_iter))
>>>>> +        return false;
>>>>> +
>>>>> +    if (iov_iter->iov_offset)
>>>>> +        return false;
>>>>> +
>>>>> +    /* We can't send whole iov. */
>>>>> +    if (iov_iter->count > max_to_send)
>>>>> +        return false;
>>>>> +
>>>>> +    return true;
>>>>> +}
>>>>> +
>>>>> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
>>>>> +                       struct sk_buff *skb,
>>>>> +                       struct msghdr *msg,
>>>>> +                       bool zerocopy)
>>>>> +{
>>>>> +    struct ubuf_info *uarg;
>>>>> +
>>>>> +    if (msg->msg_ubuf) {
>>>>> +        uarg = msg->msg_ubuf;
>>>>> +        net_zcopy_get(uarg);
>>>>> +    } else {
>>>>> +        struct iov_iter *iter = &msg->msg_iter;
>>>>> +        struct ubuf_info_msgzc *uarg_zc;
>>>>> +        int len;
>>>>> +
>>>>> +        /* Only ITER_IOVEC or ITER_UBUF are allowed and
>>>>> +         * checked before.
>>>>> +         */
>>>>> +        if (iter_is_iovec(iter))
>>>>> +            len = iov_length(iter->__iov, iter->nr_segs);
>>>>> +        else
>>>>> +            len = iter->count;
>>>>> +
>>>>> +        uarg = msg_zerocopy_realloc(sk_vsock(vsk),
>>>>> +                        len,
>>>>> +                        NULL);
>>>>> +        if (!uarg)
>>>>> +            return -1;
>>>>> +
>>>>> +        uarg_zc = uarg_to_msgzc(uarg);
>>>>> +        uarg_zc->zerocopy = zerocopy ? 1 : 0;
>>>>>      }
>>>>>
>>>>> -    if (info->reply)
>>>>> -        virtio_vsock_skb_set_reply(skb);
>>>>> +    skb_zcopy_init(skb, uarg);
>>>>>
>>>>> -    trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>>>> -                     dst_cid, dst_port,
>>>>> -                     len,
>>>>> -                     info->type,
>>>>> -                     info->op,
>>>>> -                     info->flags);
>>>>> +    return 0;
>>>>> +}
>>>>>
>>>>> -    if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
>>>>> -        WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
>>>>> -        goto out;
>>>>> +static int virtio_transport_fill_skb(struct sk_buff *skb,
>>>>> +                     struct virtio_vsock_pkt_info *info,
>>>>> +                     size_t len,
>>>>> +                     bool zcopy)
>>>>> +{
>>>>> +    if (zcopy) {
>>>>> +        return __zerocopy_sg_from_iter(info->msg, NULL, skb,
>>>>> +                          &info->msg->msg_iter,
>>>>> +                          len);
>>>>> +    } else {
>>>>> +        void *payload;
>>>>> +        int err;
>>>>> +
>>>>> +        payload = skb_put(skb, len);
>>>>> +        err = memcpy_from_msg(payload, info->msg, len);
>>>>> +        if (err)
>>>>> +            return -1;
>>>>> +
>>>>> +        if (msg_data_left(info->msg))
>>>>> +            return 0;
>>>>> +
>>>>> +        return 0;
>>>>>      }
>>>>> +}
>>>>>
>>>>> -    return skb;
>>>>> +static void virtio_transport_init_hdr(struct sk_buff *skb,
>>>>> +                      struct virtio_vsock_pkt_info *info,
>>>>> +                      u32 src_cid,
>>>>> +                      u32 src_port,
>>>>> +                      u32 dst_cid,
>>>>> +                      u32 dst_port,
>>>>> +                      size_t len)
>>>>> +{
>>>>> +    struct virtio_vsock_hdr *hdr;
>>>>>
>>>>> -out:
>>>>> -    kfree_skb(skb);
>>>>> -    return NULL;
>>>>> +    hdr = virtio_vsock_hdr(skb);
>>>>> +    hdr->type    = cpu_to_le16(info->type);
>>>>> +    hdr->op        = cpu_to_le16(info->op);
>>>>> +    hdr->src_cid    = cpu_to_le64(src_cid);
>>>>> +    hdr->dst_cid    = cpu_to_le64(dst_cid);
>>>>> +    hdr->src_port    = cpu_to_le32(src_port);
>>>>> +    hdr->dst_port    = cpu_to_le32(dst_port);
>>>>> +    hdr->flags    = cpu_to_le32(info->flags);
>>>>> +    hdr->len    = cpu_to_le32(len);
>>>>>  }
>>>>>
>>>>>  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
>>>>> @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>>>          return VIRTIO_VSOCK_TYPE_SEQPACKET;
>>>>>  }
>>>>>
>>>>> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
>>>>> +                          struct virtio_vsock_pkt_info *info,
>>>>> +                          size_t payload_len,
>>>>> +                          bool zcopy,
>>>>> +                          u32 src_cid,
>>>>> +                          u32 src_port,
>>>>> +                          u32 dst_cid,
>>>>> +                          u32 dst_port)
>>>>> +{
>>>>> +    struct sk_buff *skb;
>>>>> +    size_t skb_len;
>>>>> +
>>>>> +    skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
>>>>> +
>>>>> +    if (!zcopy)
>>>>> +        skb_len += payload_len;
>>>>> +
>>>>> +    skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>>>> +    if (!skb)
>>>>> +        return NULL;
>>>>> +
>>>>> +    virtio_transport_init_hdr(skb, info, src_cid, src_port,
>>>>> +                  dst_cid, dst_port,
>>>>> +                  payload_len);
>>>>> +
>>>>> +    /* Set owner here, because '__zerocopy_sg_from_iter()' uses
>>>>> +     * owner of skb without check to update 'sk_wmem_alloc'.
>>>>> +     */
>>>>> +    if (vsk)
>>>>> +        skb_set_owner_w(skb, sk_vsock(vsk));
>>>>> +
>>>>> +    if (info->msg && payload_len > 0) {
>>>>> +        int err;
>>>>> +
>>>>> +        err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
>>>>> +        if (err)
>>>>> +            goto out;
>>>>> +
>>>>> +        if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>>>> +            struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
>>>>> +
>>>>> +            hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>>>> +
>>>>> +            if (info->msg->msg_flags & MSG_EOR)
>>>>> +                hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>>>> +        }
>>>>> +    }
>>>>> +
>>>>> +    if (info->reply)
>>>>> +        virtio_vsock_skb_set_reply(skb);
>>>>> +
>>>>> +    trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>>>> +                     dst_cid, dst_port,
>>>>> +                     payload_len,
>>>>> +                     info->type,
>>>>> +                     info->op,
>>>>> +                     info->flags);
>>>>> +
>>>>> +    return skb;
>>>>> +out:
>>>>> +    kfree_skb(skb);
>>>>> +    return NULL;
>>>>> +}
>>>>> +
>>>>>  /* This function can only be used on connecting/connected sockets,
>>>>>   * since a socket assigned to a transport is required.
>>>>>   *
>>>>> @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>>>  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>>                        struct virtio_vsock_pkt_info *info)
>>>>>  {
>>>>> +    u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
>>>>>      u32 src_cid, src_port, dst_cid, dst_port;
>>>>>      const struct virtio_transport *t_ops;
>>>>>      struct virtio_vsock_sock *vvs;
>>>>>      u32 pkt_len = info->pkt_len;
>>>>> +    bool can_zcopy = false;
>>>>>      u32 rest_len;
>>>>>      int ret;
>>>>>
>>>>> @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>>      if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
>>>>>          return pkt_len;
>>>>>
>>>>> +    if (info->msg) {
>>>>> +        /* If zerocopy is not enabled by 'setsockopt()', we behave as
>>>>> +         * there is no MSG_ZEROCOPY flag set.
>>>>> +         */
>>>>> +        if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
>>>>> +            info->msg->msg_flags &= ~MSG_ZEROCOPY;
>>>>> +
>>>>> +        if (info->msg->msg_flags & MSG_ZEROCOPY)
>>>>> +            can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
>>>>> +
>>>>> +        if (can_zcopy)
>>>>> +            max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
>>>>> +                        (MAX_SKB_FRAGS * PAGE_SIZE));
>>>>> +    }
>>>>> +
>>>>>      rest_len = pkt_len;
>>>>>
>>>>>      do {
>>>>>          struct sk_buff *skb;
>>>>>          size_t skb_len;
>>>>>
>>>>> -        skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
>>>>> +        skb_len = min(max_skb_len, rest_len);
>>>>>
>>>>> -        skb = virtio_transport_alloc_skb(info, skb_len,
>>>>> +        skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
>>>>>                           src_cid, src_port,
>>>>>                           dst_cid, dst_port);
>>>>>          if (!skb) {
>>>>> @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>>              break;
>>>>>          }
>>>>>
>>>>> +        /* This is last skb to send this portion of data. */
>>>>> +        if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
>>>>> +            skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
>>>>> +            if (virtio_transport_init_zcopy_skb(vsk, skb,
>>>>> +                                info->msg,
>>>>> +                                can_zcopy)) {
>>>>> +                ret = -ENOMEM;
>>>>> +                break;
>>>>> +            }
>>>>> +        }
>>>>> +
>>>>>          virtio_transport_inc_tx_pkt(vvs, skb);
>>>>>
>>>>>          ret = t_ops->send_pkt(skb);
>>>>> @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
>>>>>      if (!t)
>>>>>          return -ENOTCONN;
>>>>>
>>>>> -    reply = virtio_transport_alloc_skb(&info, 0,
>>>>> +    reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
>>>>>                         le64_to_cpu(hdr->dst_cid),
>>>>>                         le32_to_cpu(hdr->dst_port),
>>>>>                         le64_to_cpu(hdr->src_cid),
>>>>
>>>
>
  
Michael S. Tsirkin July 25, 2023, 12:39 p.m. UTC | #10
On Tue, Jul 25, 2023 at 02:28:02PM +0200, Stefano Garzarella wrote:
> On Tue, Jul 25, 2023 at 12:16:11PM +0300, Arseniy Krasnov wrote:
> > 
> > 
> > On 25.07.2023 11:46, Arseniy Krasnov wrote:
> > > 
> > > 
> > > On 25.07.2023 11:43, Stefano Garzarella wrote:
> > > > On Fri, Jul 21, 2023 at 08:09:03AM +0300, Arseniy Krasnov wrote:
> > > > > 
> > > > > 
> > > > > On 21.07.2023 00:42, Arseniy Krasnov wrote:
> > > > > > This adds handling of MSG_ZEROCOPY flag on transmission path: if this
> > > > > > flag is set and zerocopy transmission is possible (enabled in socket
> > > > > > options and transport allows zerocopy), then non-linear skb will be
> > > > > > created and filled with the pages of user's buffer. Pages of user's
> > > > > > buffer are locked in memory by 'get_user_pages()'. Second thing that
> > > > > > this patch does is replace type of skb owning: instead of calling
> > > > > > 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
> > > > > > change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
> > > > > > of socket, so to decrease this field correctly proper skb destructor is
> > > > > > needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
> > > > > > 
> > > > > > Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
> > > > > > ---
> > > > > >  Changelog:
> > > > > >  v5(big patchset) -> v1:
> > > > > >   * Refactorings of 'if' conditions.
> > > > > >   * Remove extra blank line.
> > > > > >   * Remove 'frag_off' field unneeded init.
> > > > > >   * Add function 'virtio_transport_fill_skb()' which fills both linear
> > > > > >     and non-linear skb with provided data.
> > > > > >  v1 -> v2:
> > > > > >   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
> > > > > >  v2 -> v3:
> > > > > >   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
> > > > > >     provided 'iov_iter' with data could be sent in a zerocopy mode.
> > > > > >     If this callback is not set in transport - transport allows to send
> > > > > >     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
> > > > > >     then zerocopy is allowed. Reason of this callback is that in case of
> > > > > >     G2H transmission we insert whole skb to the tx virtio queue and such
> > > > > >     skb must fit to the size of the virtio queue to be sent in a single
> > > > > >     iteration (may be tx logic in 'virtio_transport.c' could be reworked
> > > > > >     as in vhost to support partial send of current skb). This callback
> > > > > >     will be enabled only for G2H path. For details pls see comment
> > > > > >     'Check that tx queue...' below.
> > > > > > 
> > > > > >  include/net/af_vsock.h                  |   3 +
> > > > > >  net/vmw_vsock/virtio_transport.c        |  39 ++++
> > > > > >  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
> > > > > >  3 files changed, 241 insertions(+), 58 deletions(-)
> > > > > > 
> > > > > > diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
> > > > > > index 0e7504a42925..a6b346eeeb8e 100644
> > > > > > --- a/include/net/af_vsock.h
> > > > > > +++ b/include/net/af_vsock.h
> > > > > > @@ -177,6 +177,9 @@ struct vsock_transport {
> > > > > > 
> > > > > >      /* Read a single skb */
> > > > > >      int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
> > > > > > +
> > > > > > +    /* Zero-copy. */
> > > > > > +    bool (*msgzerocopy_check_iov)(const struct iov_iter *);
> > > > > >  };
> > > > > > 
> > > > > >  /**** CORE ****/
> > > > > > diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
> > > > > > index 7bbcc8093e51..23cb8ed638c4 100644
> > > > > > --- a/net/vmw_vsock/virtio_transport.c
> > > > > > +++ b/net/vmw_vsock/virtio_transport.c
> > > > > > @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
> > > > > >      queue_work(virtio_vsock_workqueue, &vsock->rx_work);
> > > > > >  }
> > > > > > 
> > > > > > +static bool
> > > > > > virtio_transport_msgzerocopy_check_iov(const struct
> > > > > > iov_iter *iov)
> > > > > > +{
> > > > > > +    struct virtio_vsock *vsock;
> > > > > > +    bool res = false;
> > > > > > +
> > > > > > +    rcu_read_lock();
> > > > > > +
> > > > > > +    vsock = rcu_dereference(the_virtio_vsock);
> > > > > > +    if (vsock) {
> 
> Just noted, what about the following to reduce the indentation?
> 
>         if (!vsock) {
>             goto out;
>         }

no {} pls

>             ...
>             ...
>     out:
>         rcu_read_unlock();
>         return res;

indentation is quite modest here. Not sure goto is worth it.
  
Stefano Garzarella July 25, 2023, 12:45 p.m. UTC | #11
On Tue, Jul 25, 2023 at 08:39:17AM -0400, Michael S. Tsirkin wrote:
>On Tue, Jul 25, 2023 at 02:28:02PM +0200, Stefano Garzarella wrote:
>> On Tue, Jul 25, 2023 at 12:16:11PM +0300, Arseniy Krasnov wrote:
>> >
>> >
>> > On 25.07.2023 11:46, Arseniy Krasnov wrote:
>> > >
>> > >
>> > > On 25.07.2023 11:43, Stefano Garzarella wrote:
>> > > > On Fri, Jul 21, 2023 at 08:09:03AM +0300, Arseniy Krasnov wrote:
>> > > > >
>> > > > >
>> > > > > On 21.07.2023 00:42, Arseniy Krasnov wrote:
>> > > > > > This adds handling of MSG_ZEROCOPY flag on transmission path: if this
>> > > > > > flag is set and zerocopy transmission is possible (enabled in socket
>> > > > > > options and transport allows zerocopy), then non-linear skb will be
>> > > > > > created and filled with the pages of user's buffer. Pages of user's
>> > > > > > buffer are locked in memory by 'get_user_pages()'. Second thing that
>> > > > > > this patch does is replace type of skb owning: instead of calling
>> > > > > > 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
>> > > > > > change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
>> > > > > > of socket, so to decrease this field correctly proper skb destructor is
>> > > > > > needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
>> > > > > >
>> > > > > > Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
>> > > > > > ---
>> > > > > >  Changelog:
>> > > > > >  v5(big patchset) -> v1:
>> > > > > >   * Refactorings of 'if' conditions.
>> > > > > >   * Remove extra blank line.
>> > > > > >   * Remove 'frag_off' field unneeded init.
>> > > > > >   * Add function 'virtio_transport_fill_skb()' which fills both linear
>> > > > > >     and non-linear skb with provided data.
>> > > > > >  v1 -> v2:
>> > > > > >   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
>> > > > > >  v2 -> v3:
>> > > > > >   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
>> > > > > >     provided 'iov_iter' with data could be sent in a zerocopy mode.
>> > > > > >     If this callback is not set in transport - transport allows to send
>> > > > > >     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
>> > > > > >     then zerocopy is allowed. Reason of this callback is that in case of
>> > > > > >     G2H transmission we insert whole skb to the tx virtio queue and such
>> > > > > >     skb must fit to the size of the virtio queue to be sent in a single
>> > > > > >     iteration (may be tx logic in 'virtio_transport.c' could be reworked
>> > > > > >     as in vhost to support partial send of current skb). This callback
>> > > > > >     will be enabled only for G2H path. For details pls see comment
>> > > > > >     'Check that tx queue...' below.
>> > > > > >
>> > > > > >  include/net/af_vsock.h                  |   3 +
>> > > > > >  net/vmw_vsock/virtio_transport.c        |  39 ++++
>> > > > > >  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
>> > > > > >  3 files changed, 241 insertions(+), 58 deletions(-)
>> > > > > >
>> > > > > > diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
>> > > > > > index 0e7504a42925..a6b346eeeb8e 100644
>> > > > > > --- a/include/net/af_vsock.h
>> > > > > > +++ b/include/net/af_vsock.h
>> > > > > > @@ -177,6 +177,9 @@ struct vsock_transport {
>> > > > > >
>> > > > > >      /* Read a single skb */
>> > > > > >      int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
>> > > > > > +
>> > > > > > +    /* Zero-copy. */
>> > > > > > +    bool (*msgzerocopy_check_iov)(const struct iov_iter *);
>> > > > > >  };
>> > > > > >
>> > > > > >  /**** CORE ****/
>> > > > > > diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>> > > > > > index 7bbcc8093e51..23cb8ed638c4 100644
>> > > > > > --- a/net/vmw_vsock/virtio_transport.c
>> > > > > > +++ b/net/vmw_vsock/virtio_transport.c
>> > > > > > @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
>> > > > > >      queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>> > > > > >  }
>> > > > > >
>> > > > > > +static bool
>> > > > > > virtio_transport_msgzerocopy_check_iov(const struct
>> > > > > > iov_iter *iov)
>> > > > > > +{
>> > > > > > +    struct virtio_vsock *vsock;
>> > > > > > +    bool res = false;
>> > > > > > +
>> > > > > > +    rcu_read_lock();
>> > > > > > +
>> > > > > > +    vsock = rcu_dereference(the_virtio_vsock);
>> > > > > > +    if (vsock) {
>>
>> Just noted, what about the following to reduce the indentation?
>>
>>         if (!vsock) {
>>             goto out;
>>         }
>
>no {} pls

ooops, true, too much QEMU code today, but luckily checkpatch would have
spotted it ;-)

>
>>             ...
>>             ...
>>     out:
>>         rcu_read_unlock();
>>         return res;
>
>indentation is quite modest here. Not sure goto is worth it.

It's a pattern we follow a lot in this file, and I find the early
return/goto more readable.
Anyway, I don't have a strong opinion, it's fine the way it is now,
actually we don't have too many indentations for now in this function.

Thanks,
Stefano
  
Stefano Garzarella July 25, 2023, 12:53 p.m. UTC | #12
On Tue, Jul 25, 2023 at 07:50:53AM -0400, Michael S. Tsirkin wrote:
>On Fri, Jul 21, 2023 at 08:09:03AM +0300, Arseniy Krasnov wrote:
>>
>>
>> On 21.07.2023 00:42, Arseniy Krasnov wrote:
>> > This adds handling of MSG_ZEROCOPY flag on transmission path: if this
>> > flag is set and zerocopy transmission is possible (enabled in socket
>> > options and transport allows zerocopy), then non-linear skb will be
>> > created and filled with the pages of user's buffer. Pages of user's
>> > buffer are locked in memory by 'get_user_pages()'. Second thing that
>> > this patch does is replace type of skb owning: instead of calling
>> > 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
>> > change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
>> > of socket, so to decrease this field correctly proper skb destructor is
>> > needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
>> >
>> > Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
>> > ---
>> >  Changelog:
>> >  v5(big patchset) -> v1:
>> >   * Refactorings of 'if' conditions.
>> >   * Remove extra blank line.
>> >   * Remove 'frag_off' field unneeded init.
>> >   * Add function 'virtio_transport_fill_skb()' which fills both linear
>> >     and non-linear skb with provided data.
>> >  v1 -> v2:
>> >   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
>> >  v2 -> v3:
>> >   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
>> >     provided 'iov_iter' with data could be sent in a zerocopy mode.
>> >     If this callback is not set in transport - transport allows to send
>> >     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
>> >     then zerocopy is allowed. Reason of this callback is that in case of
>> >     G2H transmission we insert whole skb to the tx virtio queue and such
>> >     skb must fit to the size of the virtio queue to be sent in a single
>> >     iteration (may be tx logic in 'virtio_transport.c' could be reworked
>> >     as in vhost to support partial send of current skb). This callback
>> >     will be enabled only for G2H path. For details pls see comment
>> >     'Check that tx queue...' below.
>> >
>> >  include/net/af_vsock.h                  |   3 +
>> >  net/vmw_vsock/virtio_transport.c        |  39 ++++
>> >  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
>> >  3 files changed, 241 insertions(+), 58 deletions(-)
>> >
>> > diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
>> > index 0e7504a42925..a6b346eeeb8e 100644
>> > --- a/include/net/af_vsock.h
>> > +++ b/include/net/af_vsock.h
>> > @@ -177,6 +177,9 @@ struct vsock_transport {
>> >
>> >  	/* Read a single skb */
>> >  	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
>> > +
>> > +	/* Zero-copy. */
>> > +	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
>> >  };
>> >
>> >  /**** CORE ****/
>> > diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>> > index 7bbcc8093e51..23cb8ed638c4 100644
>> > --- a/net/vmw_vsock/virtio_transport.c
>> > +++ b/net/vmw_vsock/virtio_transport.c
>> > @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
>> >  	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>> >  }
>> >
>> > +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
>> > +{
>> > +	struct virtio_vsock *vsock;
>> > +	bool res = false;
>> > +
>> > +	rcu_read_lock();
>> > +
>> > +	vsock = rcu_dereference(the_virtio_vsock);
>> > +	if (vsock) {
>> > +		struct virtqueue *vq;
>> > +		int iov_pages;
>> > +
>> > +		vq = vsock->vqs[VSOCK_VQ_TX];
>> > +
>> > +		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
>> > +
>> > +		/* Check that tx queue is large enough to keep whole
>> > +		 * data to send. This is needed, because when there is
>> > +		 * not enough free space in the queue, current skb to
>> > +		 * send will be reinserted to the head of tx list of
>> > +		 * the socket to retry transmission later, so if skb
>> > +		 * is bigger than whole queue, it will be reinserted
>> > +		 * again and again, thus blocking other skbs to be sent.
>> > +		 * Each page of the user provided buffer will be added
>> > +		 * as a single buffer to the tx virtqueue, so compare
>> > +		 * number of pages against maximum capacity of the queue.
>> > +		 * +1 means buffer for the packet header.
>> > +		 */
>> > +		if (iov_pages + 1 <= vq->num_max)
>>
>> I think this check is actual only for case one we don't have indirect buffer feature.
>> With indirect mode whole data to send will be packed into one indirect buffer.
>>
>> Thanks, Arseniy
>
>Actually the reverse. With indirect you are limited to num_max.
>Without you are limited to whatever space is left in the
>queue (which you did not check here, so you should).
>
>
>> > +			res = true;
>> > +	}
>> > +
>> > +	rcu_read_unlock();
>
>Just curious:
>is the point of all this RCU dance to allow vsock
>to change from under us? then why is it ok to
>have it change? the virtio_transport_msgzerocopy_check_iov
>will then refer to the old vsock ...

IIRC we introduced the RCU to handle hot-unplug issues:
commit 0deab087b16a ("vsock/virtio: use RCU to avoid use-after-free on
the_virtio_vsock")

When we remove the device, we flush all the works, etc. so we should
not be in this case (referring the old vsock), except for an irrelevant
transient as the device is disappearing.

Stefano
  
Arseniy Krasnov July 25, 2023, 1:04 p.m. UTC | #13
On 25.07.2023 14:50, Michael S. Tsirkin wrote:
> On Fri, Jul 21, 2023 at 08:09:03AM +0300, Arseniy Krasnov wrote:
>>
>>
>> On 21.07.2023 00:42, Arseniy Krasnov wrote:
>>> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
>>> flag is set and zerocopy transmission is possible (enabled in socket
>>> options and transport allows zerocopy), then non-linear skb will be
>>> created and filled with the pages of user's buffer. Pages of user's
>>> buffer are locked in memory by 'get_user_pages()'. Second thing that
>>> this patch does is replace type of skb owning: instead of calling
>>> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
>>> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
>>> of socket, so to decrease this field correctly proper skb destructor is
>>> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
>>>
>>> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
>>> ---
>>>  Changelog:
>>>  v5(big patchset) -> v1:
>>>   * Refactorings of 'if' conditions.
>>>   * Remove extra blank line.
>>>   * Remove 'frag_off' field unneeded init.
>>>   * Add function 'virtio_transport_fill_skb()' which fills both linear
>>>     and non-linear skb with provided data.
>>>  v1 -> v2:
>>>   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
>>>  v2 -> v3:
>>>   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
>>>     provided 'iov_iter' with data could be sent in a zerocopy mode.
>>>     If this callback is not set in transport - transport allows to send
>>>     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
>>>     then zerocopy is allowed. Reason of this callback is that in case of
>>>     G2H transmission we insert whole skb to the tx virtio queue and such
>>>     skb must fit to the size of the virtio queue to be sent in a single
>>>     iteration (may be tx logic in 'virtio_transport.c' could be reworked
>>>     as in vhost to support partial send of current skb). This callback
>>>     will be enabled only for G2H path. For details pls see comment 
>>>     'Check that tx queue...' below.
>>>
>>>  include/net/af_vsock.h                  |   3 +
>>>  net/vmw_vsock/virtio_transport.c        |  39 ++++
>>>  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
>>>  3 files changed, 241 insertions(+), 58 deletions(-)
>>>
>>> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
>>> index 0e7504a42925..a6b346eeeb8e 100644
>>> --- a/include/net/af_vsock.h
>>> +++ b/include/net/af_vsock.h
>>> @@ -177,6 +177,9 @@ struct vsock_transport {
>>>  
>>>  	/* Read a single skb */
>>>  	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
>>> +
>>> +	/* Zero-copy. */
>>> +	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
>>>  };
>>>  
>>>  /**** CORE ****/
>>> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>>> index 7bbcc8093e51..23cb8ed638c4 100644
>>> --- a/net/vmw_vsock/virtio_transport.c
>>> +++ b/net/vmw_vsock/virtio_transport.c
>>> @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
>>>  	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>>>  }
>>>  
>>> +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
>>> +{
>>> +	struct virtio_vsock *vsock;
>>> +	bool res = false;
>>> +
>>> +	rcu_read_lock();
>>> +
>>> +	vsock = rcu_dereference(the_virtio_vsock);
>>> +	if (vsock) {
>>> +		struct virtqueue *vq;
>>> +		int iov_pages;
>>> +
>>> +		vq = vsock->vqs[VSOCK_VQ_TX];
>>> +
>>> +		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
>>> +
>>> +		/* Check that tx queue is large enough to keep whole
>>> +		 * data to send. This is needed, because when there is
>>> +		 * not enough free space in the queue, current skb to
>>> +		 * send will be reinserted to the head of tx list of
>>> +		 * the socket to retry transmission later, so if skb
>>> +		 * is bigger than whole queue, it will be reinserted
>>> +		 * again and again, thus blocking other skbs to be sent.
>>> +		 * Each page of the user provided buffer will be added
>>> +		 * as a single buffer to the tx virtqueue, so compare
>>> +		 * number of pages against maximum capacity of the queue.
>>> +		 * +1 means buffer for the packet header.
>>> +		 */
>>> +		if (iov_pages + 1 <= vq->num_max)
>>
>> I think this check is actual only for case one we don't have indirect buffer feature.
>> With indirect mode whole data to send will be packed into one indirect buffer.
>>
>> Thanks, Arseniy
> 
> Actually the reverse. With indirect you are limited to num_max.
> Without you are limited to whatever space is left in the
> queue (which you did not check here, so you should).

I mean that with indirect, we only need one buffer, and we can just wait
for enough space - for this single buffer ( as we discussed a little bit before).
But if indirect buffer is not supported - we need that whole packet must fit
to the size of tx queue - otherwise it never be transmitted.

Thanks, Arseniy

> 
> 
>>> +			res = true;
>>> +	}
>>> +
>>> +	rcu_read_unlock();
> 
> Just curious:
> is the point of all this RCU dance to allow vsock
> to change from under us? then why is it ok to
> have it change? the virtio_transport_msgzerocopy_check_iov
> will then refer to the old vsock ...
> 
> 
>>> +
>>> +	return res;
>>> +}
>>> +
>>>  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
>>>  
>>>  static struct virtio_transport virtio_transport = {
>>> @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
>>>  		.seqpacket_allow          = virtio_transport_seqpacket_allow,
>>>  		.seqpacket_has_data       = virtio_transport_seqpacket_has_data,
>>>  
>>> +		.msgzerocopy_check_iov	  = virtio_transport_msgzerocopy_check_iov,
>>> +
>>>  		.notify_poll_in           = virtio_transport_notify_poll_in,
>>>  		.notify_poll_out          = virtio_transport_notify_poll_out,
>>>  		.notify_recv_init         = virtio_transport_notify_recv_init,
>>> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
>>> index 26a4d10da205..e4e3d541aff4 100644
>>> --- a/net/vmw_vsock/virtio_transport_common.c
>>> +++ b/net/vmw_vsock/virtio_transport_common.c
>>> @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
>>>  	return container_of(t, struct virtio_transport, transport);
>>>  }
>>>  
>>> -/* Returns a new packet on success, otherwise returns NULL.
>>> - *
>>> - * If NULL is returned, errp is set to a negative errno.
>>> - */
>>> -static struct sk_buff *
>>> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
>>> -			   size_t len,
>>> -			   u32 src_cid,
>>> -			   u32 src_port,
>>> -			   u32 dst_cid,
>>> -			   u32 dst_port)
>>> -{
>>> -	const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
>>> -	struct virtio_vsock_hdr *hdr;
>>> -	struct sk_buff *skb;
>>> -	void *payload;
>>> -	int err;
>>> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
>>> +				       size_t max_to_send)
>>> +{
>>> +	const struct vsock_transport *t;
>>> +	struct iov_iter *iov_iter;
>>>  
>>> -	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>> -	if (!skb)
>>> -		return NULL;
>>> +	if (!info->msg)
>>> +		return false;
>>>  
>>> -	hdr = virtio_vsock_hdr(skb);
>>> -	hdr->type	= cpu_to_le16(info->type);
>>> -	hdr->op		= cpu_to_le16(info->op);
>>> -	hdr->src_cid	= cpu_to_le64(src_cid);
>>> -	hdr->dst_cid	= cpu_to_le64(dst_cid);
>>> -	hdr->src_port	= cpu_to_le32(src_port);
>>> -	hdr->dst_port	= cpu_to_le32(dst_port);
>>> -	hdr->flags	= cpu_to_le32(info->flags);
>>> -	hdr->len	= cpu_to_le32(len);
>>> +	iov_iter = &info->msg->msg_iter;
>>>  
>>> -	if (info->msg && len > 0) {
>>> -		payload = skb_put(skb, len);
>>> -		err = memcpy_from_msg(payload, info->msg, len);
>>> -		if (err)
>>> -			goto out;
>>> +	t = vsock_core_get_transport(info->vsk);
>>>  
>>> -		if (msg_data_left(info->msg) == 0 &&
>>> -		    info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>> -			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>> +	if (t->msgzerocopy_check_iov &&
>>> +	    !t->msgzerocopy_check_iov(iov_iter))
>>> +		return false;
>>>  
>>> -			if (info->msg->msg_flags & MSG_EOR)
>>> -				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>> -		}
>>> +	/* Data is simple buffer. */
>>> +	if (iter_is_ubuf(iov_iter))
>>> +		return true;
>>> +
>>> +	if (!iter_is_iovec(iov_iter))
>>> +		return false;
>>> +
>>> +	if (iov_iter->iov_offset)
>>> +		return false;
>>> +
>>> +	/* We can't send whole iov. */
>>> +	if (iov_iter->count > max_to_send)
>>> +		return false;
>>> +
>>> +	return true;
>>> +}
>>> +
>>> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
>>> +					   struct sk_buff *skb,
>>> +					   struct msghdr *msg,
>>> +					   bool zerocopy)
>>> +{
>>> +	struct ubuf_info *uarg;
>>> +
>>> +	if (msg->msg_ubuf) {
>>> +		uarg = msg->msg_ubuf;
>>> +		net_zcopy_get(uarg);
>>> +	} else {
>>> +		struct iov_iter *iter = &msg->msg_iter;
>>> +		struct ubuf_info_msgzc *uarg_zc;
>>> +		int len;
>>> +
>>> +		/* Only ITER_IOVEC or ITER_UBUF are allowed and
>>> +		 * checked before.
>>> +		 */
>>> +		if (iter_is_iovec(iter))
>>> +			len = iov_length(iter->__iov, iter->nr_segs);
>>> +		else
>>> +			len = iter->count;
>>> +
>>> +		uarg = msg_zerocopy_realloc(sk_vsock(vsk),
>>> +					    len,
>>> +					    NULL);
>>> +		if (!uarg)
>>> +			return -1;
>>> +
>>> +		uarg_zc = uarg_to_msgzc(uarg);
>>> +		uarg_zc->zerocopy = zerocopy ? 1 : 0;
>>>  	}
>>>  
>>> -	if (info->reply)
>>> -		virtio_vsock_skb_set_reply(skb);
>>> +	skb_zcopy_init(skb, uarg);
>>>  
>>> -	trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>> -					 dst_cid, dst_port,
>>> -					 len,
>>> -					 info->type,
>>> -					 info->op,
>>> -					 info->flags);
>>> +	return 0;
>>> +}
>>>  
>>> -	if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
>>> -		WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
>>> -		goto out;
>>> +static int virtio_transport_fill_skb(struct sk_buff *skb,
>>> +				     struct virtio_vsock_pkt_info *info,
>>> +				     size_t len,
>>> +				     bool zcopy)
>>> +{
>>> +	if (zcopy) {
>>> +		return __zerocopy_sg_from_iter(info->msg, NULL, skb,
>>> +					      &info->msg->msg_iter,
>>> +					      len);
>>> +	} else {
>>> +		void *payload;
>>> +		int err;
>>> +
>>> +		payload = skb_put(skb, len);
>>> +		err = memcpy_from_msg(payload, info->msg, len);
>>> +		if (err)
>>> +			return -1;
>>> +
>>> +		if (msg_data_left(info->msg))
>>> +			return 0;
>>> +
>>> +		return 0;
>>>  	}
>>> +}
>>>  
>>> -	return skb;
>>> +static void virtio_transport_init_hdr(struct sk_buff *skb,
>>> +				      struct virtio_vsock_pkt_info *info,
>>> +				      u32 src_cid,
>>> +				      u32 src_port,
>>> +				      u32 dst_cid,
>>> +				      u32 dst_port,
>>> +				      size_t len)
>>> +{
>>> +	struct virtio_vsock_hdr *hdr;
>>>  
>>> -out:
>>> -	kfree_skb(skb);
>>> -	return NULL;
>>> +	hdr = virtio_vsock_hdr(skb);
>>> +	hdr->type	= cpu_to_le16(info->type);
>>> +	hdr->op		= cpu_to_le16(info->op);
>>> +	hdr->src_cid	= cpu_to_le64(src_cid);
>>> +	hdr->dst_cid	= cpu_to_le64(dst_cid);
>>> +	hdr->src_port	= cpu_to_le32(src_port);
>>> +	hdr->dst_port	= cpu_to_le32(dst_port);
>>> +	hdr->flags	= cpu_to_le32(info->flags);
>>> +	hdr->len	= cpu_to_le32(len);
>>>  }
>>>  
>>>  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
>>> @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>  		return VIRTIO_VSOCK_TYPE_SEQPACKET;
>>>  }
>>>  
>>> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
>>> +						  struct virtio_vsock_pkt_info *info,
>>> +						  size_t payload_len,
>>> +						  bool zcopy,
>>> +						  u32 src_cid,
>>> +						  u32 src_port,
>>> +						  u32 dst_cid,
>>> +						  u32 dst_port)
>>> +{
>>> +	struct sk_buff *skb;
>>> +	size_t skb_len;
>>> +
>>> +	skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
>>> +
>>> +	if (!zcopy)
>>> +		skb_len += payload_len;
>>> +
>>> +	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>> +	if (!skb)
>>> +		return NULL;
>>> +
>>> +	virtio_transport_init_hdr(skb, info, src_cid, src_port,
>>> +				  dst_cid, dst_port,
>>> +				  payload_len);
>>> +
>>> +	/* Set owner here, because '__zerocopy_sg_from_iter()' uses
>>> +	 * owner of skb without check to update 'sk_wmem_alloc'.
>>> +	 */
>>> +	if (vsk)
>>> +		skb_set_owner_w(skb, sk_vsock(vsk));
>>> +
>>> +	if (info->msg && payload_len > 0) {
>>> +		int err;
>>> +
>>> +		err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
>>> +		if (err)
>>> +			goto out;
>>> +
>>> +		if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>> +			struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
>>> +
>>> +			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>> +
>>> +			if (info->msg->msg_flags & MSG_EOR)
>>> +				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>> +		}
>>> +	}
>>> +
>>> +	if (info->reply)
>>> +		virtio_vsock_skb_set_reply(skb);
>>> +
>>> +	trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>> +					 dst_cid, dst_port,
>>> +					 payload_len,
>>> +					 info->type,
>>> +					 info->op,
>>> +					 info->flags);
>>> +
>>> +	return skb;
>>> +out:
>>> +	kfree_skb(skb);
>>> +	return NULL;
>>> +}
>>> +
>>>  /* This function can only be used on connecting/connected sockets,
>>>   * since a socket assigned to a transport is required.
>>>   *
>>> @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>  					  struct virtio_vsock_pkt_info *info)
>>>  {
>>> +	u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
>>>  	u32 src_cid, src_port, dst_cid, dst_port;
>>>  	const struct virtio_transport *t_ops;
>>>  	struct virtio_vsock_sock *vvs;
>>>  	u32 pkt_len = info->pkt_len;
>>> +	bool can_zcopy = false;
>>>  	u32 rest_len;
>>>  	int ret;
>>>  
>>> @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>  	if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
>>>  		return pkt_len;
>>>  
>>> +	if (info->msg) {
>>> +		/* If zerocopy is not enabled by 'setsockopt()', we behave as
>>> +		 * there is no MSG_ZEROCOPY flag set.
>>> +		 */
>>> +		if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
>>> +			info->msg->msg_flags &= ~MSG_ZEROCOPY;
>>> +
>>> +		if (info->msg->msg_flags & MSG_ZEROCOPY)
>>> +			can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
>>> +
>>> +		if (can_zcopy)
>>> +			max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
>>> +					    (MAX_SKB_FRAGS * PAGE_SIZE));
>>> +	}
>>> +
>>>  	rest_len = pkt_len;
>>>  
>>>  	do {
>>>  		struct sk_buff *skb;
>>>  		size_t skb_len;
>>>  
>>> -		skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
>>> +		skb_len = min(max_skb_len, rest_len);
>>>  
>>> -		skb = virtio_transport_alloc_skb(info, skb_len,
>>> +		skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
>>>  						 src_cid, src_port,
>>>  						 dst_cid, dst_port);
>>>  		if (!skb) {
>>> @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>  			break;
>>>  		}
>>>  
>>> +		/* This is last skb to send this portion of data. */
>>> +		if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
>>> +		    skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
>>> +			if (virtio_transport_init_zcopy_skb(vsk, skb,
>>> +							    info->msg,
>>> +							    can_zcopy)) {
>>> +				ret = -ENOMEM;
>>> +				break;
>>> +			}
>>> +		}
>>> +
>>>  		virtio_transport_inc_tx_pkt(vvs, skb);
>>>  
>>>  		ret = t_ops->send_pkt(skb);
>>> @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
>>>  	if (!t)
>>>  		return -ENOTCONN;
>>>  
>>> -	reply = virtio_transport_alloc_skb(&info, 0,
>>> +	reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
>>>  					   le64_to_cpu(hdr->dst_cid),
>>>  					   le32_to_cpu(hdr->dst_port),
>>>  					   le64_to_cpu(hdr->src_cid),
>
  
Michael S. Tsirkin July 25, 2023, 1:06 p.m. UTC | #14
On Tue, Jul 25, 2023 at 02:53:39PM +0200, Stefano Garzarella wrote:
> On Tue, Jul 25, 2023 at 07:50:53AM -0400, Michael S. Tsirkin wrote:
> > On Fri, Jul 21, 2023 at 08:09:03AM +0300, Arseniy Krasnov wrote:
> > > 
> > > 
> > > On 21.07.2023 00:42, Arseniy Krasnov wrote:
> > > > This adds handling of MSG_ZEROCOPY flag on transmission path: if this
> > > > flag is set and zerocopy transmission is possible (enabled in socket
> > > > options and transport allows zerocopy), then non-linear skb will be
> > > > created and filled with the pages of user's buffer. Pages of user's
> > > > buffer are locked in memory by 'get_user_pages()'. Second thing that
> > > > this patch does is replace type of skb owning: instead of calling
> > > > 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
> > > > change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
> > > > of socket, so to decrease this field correctly proper skb destructor is
> > > > needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
> > > >
> > > > Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
> > > > ---
> > > >  Changelog:
> > > >  v5(big patchset) -> v1:
> > > >   * Refactorings of 'if' conditions.
> > > >   * Remove extra blank line.
> > > >   * Remove 'frag_off' field unneeded init.
> > > >   * Add function 'virtio_transport_fill_skb()' which fills both linear
> > > >     and non-linear skb with provided data.
> > > >  v1 -> v2:
> > > >   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
> > > >  v2 -> v3:
> > > >   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
> > > >     provided 'iov_iter' with data could be sent in a zerocopy mode.
> > > >     If this callback is not set in transport - transport allows to send
> > > >     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
> > > >     then zerocopy is allowed. Reason of this callback is that in case of
> > > >     G2H transmission we insert whole skb to the tx virtio queue and such
> > > >     skb must fit to the size of the virtio queue to be sent in a single
> > > >     iteration (may be tx logic in 'virtio_transport.c' could be reworked
> > > >     as in vhost to support partial send of current skb). This callback
> > > >     will be enabled only for G2H path. For details pls see comment
> > > >     'Check that tx queue...' below.
> > > >
> > > >  include/net/af_vsock.h                  |   3 +
> > > >  net/vmw_vsock/virtio_transport.c        |  39 ++++
> > > >  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
> > > >  3 files changed, 241 insertions(+), 58 deletions(-)
> > > >
> > > > diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
> > > > index 0e7504a42925..a6b346eeeb8e 100644
> > > > --- a/include/net/af_vsock.h
> > > > +++ b/include/net/af_vsock.h
> > > > @@ -177,6 +177,9 @@ struct vsock_transport {
> > > >
> > > >  	/* Read a single skb */
> > > >  	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
> > > > +
> > > > +	/* Zero-copy. */
> > > > +	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
> > > >  };
> > > >
> > > >  /**** CORE ****/
> > > > diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
> > > > index 7bbcc8093e51..23cb8ed638c4 100644
> > > > --- a/net/vmw_vsock/virtio_transport.c
> > > > +++ b/net/vmw_vsock/virtio_transport.c
> > > > @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
> > > >  	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
> > > >  }
> > > >
> > > > +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
> > > > +{
> > > > +	struct virtio_vsock *vsock;
> > > > +	bool res = false;
> > > > +
> > > > +	rcu_read_lock();
> > > > +
> > > > +	vsock = rcu_dereference(the_virtio_vsock);
> > > > +	if (vsock) {
> > > > +		struct virtqueue *vq;
> > > > +		int iov_pages;
> > > > +
> > > > +		vq = vsock->vqs[VSOCK_VQ_TX];
> > > > +
> > > > +		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
> > > > +
> > > > +		/* Check that tx queue is large enough to keep whole
> > > > +		 * data to send. This is needed, because when there is
> > > > +		 * not enough free space in the queue, current skb to
> > > > +		 * send will be reinserted to the head of tx list of
> > > > +		 * the socket to retry transmission later, so if skb
> > > > +		 * is bigger than whole queue, it will be reinserted
> > > > +		 * again and again, thus blocking other skbs to be sent.
> > > > +		 * Each page of the user provided buffer will be added
> > > > +		 * as a single buffer to the tx virtqueue, so compare
> > > > +		 * number of pages against maximum capacity of the queue.
> > > > +		 * +1 means buffer for the packet header.
> > > > +		 */
> > > > +		if (iov_pages + 1 <= vq->num_max)
> > > 
> > > I think this check is actual only for case one we don't have indirect buffer feature.
> > > With indirect mode whole data to send will be packed into one indirect buffer.
> > > 
> > > Thanks, Arseniy
> > 
> > Actually the reverse. With indirect you are limited to num_max.
> > Without you are limited to whatever space is left in the
> > queue (which you did not check here, so you should).
> > 
> > 
> > > > +			res = true;
> > > > +	}
> > > > +
> > > > +	rcu_read_unlock();
> > 
> > Just curious:
> > is the point of all this RCU dance to allow vsock
> > to change from under us? then why is it ok to
> > have it change? the virtio_transport_msgzerocopy_check_iov
> > will then refer to the old vsock ...
> 
> IIRC we introduced the RCU to handle hot-unplug issues:
> commit 0deab087b16a ("vsock/virtio: use RCU to avoid use-after-free on
> the_virtio_vsock")
> 
> When we remove the device, we flush all the works, etc. so we should
> not be in this case (referring the old vsock), except for an irrelevant
> transient as the device is disappearing.
> 
> Stefano

what if old device goes away then new one appears?
  
Arseniy Krasnov July 25, 2023, 1:10 p.m. UTC | #15
On 25.07.2023 14:59, Michael S. Tsirkin wrote:
> On Tue, Jul 25, 2023 at 11:39:22AM +0300, Arseniy Krasnov wrote:
>>
>>
>> On 25.07.2023 11:25, Michael S. Tsirkin wrote:
>>> On Fri, Jul 21, 2023 at 12:42:45AM +0300, Arseniy Krasnov wrote:
>>>> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
>>>> flag is set and zerocopy transmission is possible (enabled in socket
>>>> options and transport allows zerocopy), then non-linear skb will be
>>>> created and filled with the pages of user's buffer. Pages of user's
>>>> buffer are locked in memory by 'get_user_pages()'. Second thing that
>>>> this patch does is replace type of skb owning: instead of calling
>>>> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
>>>> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
>>>> of socket, so to decrease this field correctly proper skb destructor is
>>>> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
>>>>
>>>> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
>>>> ---
>>>>  Changelog:
>>>>  v5(big patchset) -> v1:
>>>>   * Refactorings of 'if' conditions.
>>>>   * Remove extra blank line.
>>>>   * Remove 'frag_off' field unneeded init.
>>>>   * Add function 'virtio_transport_fill_skb()' which fills both linear
>>>>     and non-linear skb with provided data.
>>>>  v1 -> v2:
>>>>   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
>>>>  v2 -> v3:
>>>>   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
>>>>     provided 'iov_iter' with data could be sent in a zerocopy mode.
>>>>     If this callback is not set in transport - transport allows to send
>>>>     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
>>>>     then zerocopy is allowed. Reason of this callback is that in case of
>>>>     G2H transmission we insert whole skb to the tx virtio queue and such
>>>>     skb must fit to the size of the virtio queue to be sent in a single
>>>>     iteration (may be tx logic in 'virtio_transport.c' could be reworked
>>>>     as in vhost to support partial send of current skb). This callback
>>>>     will be enabled only for G2H path. For details pls see comment 
>>>>     'Check that tx queue...' below.
>>>>
>>>>  include/net/af_vsock.h                  |   3 +
>>>>  net/vmw_vsock/virtio_transport.c        |  39 ++++
>>>>  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
>>>>  3 files changed, 241 insertions(+), 58 deletions(-)
>>>>
>>>> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
>>>> index 0e7504a42925..a6b346eeeb8e 100644
>>>> --- a/include/net/af_vsock.h
>>>> +++ b/include/net/af_vsock.h
>>>> @@ -177,6 +177,9 @@ struct vsock_transport {
>>>>  
>>>>  	/* Read a single skb */
>>>>  	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
>>>> +
>>>> +	/* Zero-copy. */
>>>> +	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
>>>>  };
>>>>  
>>>>  /**** CORE ****/
>>>> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>>>> index 7bbcc8093e51..23cb8ed638c4 100644
>>>> --- a/net/vmw_vsock/virtio_transport.c
>>>> +++ b/net/vmw_vsock/virtio_transport.c
>>>> @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
>>>>  	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>>>>  }
>>>>  
>>>> +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
>>>> +{
>>>> +	struct virtio_vsock *vsock;
>>>> +	bool res = false;
>>>> +
>>>> +	rcu_read_lock();
>>>> +
>>>> +	vsock = rcu_dereference(the_virtio_vsock);
>>>> +	if (vsock) {
>>>> +		struct virtqueue *vq;
>>>> +		int iov_pages;
>>>> +
>>>> +		vq = vsock->vqs[VSOCK_VQ_TX];
>>>> +
>>>> +		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
>>>> +
>>>> +		/* Check that tx queue is large enough to keep whole
>>>> +		 * data to send. This is needed, because when there is
>>>> +		 * not enough free space in the queue, current skb to
>>>> +		 * send will be reinserted to the head of tx list of
>>>> +		 * the socket to retry transmission later, so if skb
>>>> +		 * is bigger than whole queue, it will be reinserted
>>>> +		 * again and again, thus blocking other skbs to be sent.
>>>> +		 * Each page of the user provided buffer will be added
>>>> +		 * as a single buffer to the tx virtqueue, so compare
>>>> +		 * number of pages against maximum capacity of the queue.
>>>> +		 * +1 means buffer for the packet header.
>>>> +		 */
>>>> +		if (iov_pages + 1 <= vq->num_max)
>>>> +			res = true;
>>>
>>>
>>> Yes but can't there already be buffers in the queue?
>>> Then you can't stick num_max there.
>>
>> I think, that it is not critical, because vhost part always tries to process all
>> incoming buffers (yes, 'vhost_exceeds_weight()' breaks at some moment, but it will
>> reschedule tx kick ('vhost_vsock_handle_tx_kick()') work again), so current "too
>> big" skb will wait until there will be enough space in queue and as it is requeued
>> to the head of tx list it will be inserted to tx queue first.
>>
>> But anyway, I agree that comparing to 'num_free' may be more effective to the whole
>> system performance...
>>
>> Thanks, Arseniy
> 
> Oh I see. It makes sense then - instead of copying just so we can
> stick it in the queue, wait a bit and send later.
> Also - for stream transports can't the message be split
> and sent chunk by chunk? Better than copying ...

Technically yes, also we can split message for non-stream sockets (as vhost
does when it copies data to rx buffers of the guest), but it requires to rework
current implementation by adding buffers one by one to the tx queue. I think
it was not implemented here because until MSG_ZEROCOPY all skbs requires one
(if it is control message) or two (with payload) buffers, so there is no big
sense in processing max two buffers in "one-by-one" mode - we can just wait
for space.

May be, I can add this logic for non-linear skb's here: 

if (skb->len > vq->max_num)
    add buffers "one-by-one", incrementing internal offset in skb,
   if (new skb insertion fails)
       requeue skb, wait for space.

In TX done callback I'll call consume skb only when above mentioned internal
offset == skb->len. I think this approach allows to get rid of new 'check_iov'
callback from this patch.


Stefano, what do You think?

Thanks, Arseniy

> 
> 
>>>
>>>
>>>> +	}
>>>> +
>>>> +	rcu_read_unlock();
>>>> +
>>>> +	return res;
>>>> +}
>>>> +
>>>>  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
>>>>  
>>>>  static struct virtio_transport virtio_transport = {
>>>> @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
>>>>  		.seqpacket_allow          = virtio_transport_seqpacket_allow,
>>>>  		.seqpacket_has_data       = virtio_transport_seqpacket_has_data,
>>>>  
>>>> +		.msgzerocopy_check_iov	  = virtio_transport_msgzerocopy_check_iov,
>>>> +
>>>>  		.notify_poll_in           = virtio_transport_notify_poll_in,
>>>>  		.notify_poll_out          = virtio_transport_notify_poll_out,
>>>>  		.notify_recv_init         = virtio_transport_notify_recv_init,
>>>> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
>>>> index 26a4d10da205..e4e3d541aff4 100644
>>>> --- a/net/vmw_vsock/virtio_transport_common.c
>>>> +++ b/net/vmw_vsock/virtio_transport_common.c
>>>> @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
>>>>  	return container_of(t, struct virtio_transport, transport);
>>>>  }
>>>>  
>>>> -/* Returns a new packet on success, otherwise returns NULL.
>>>> - *
>>>> - * If NULL is returned, errp is set to a negative errno.
>>>> - */
>>>> -static struct sk_buff *
>>>> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
>>>> -			   size_t len,
>>>> -			   u32 src_cid,
>>>> -			   u32 src_port,
>>>> -			   u32 dst_cid,
>>>> -			   u32 dst_port)
>>>> -{
>>>> -	const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
>>>> -	struct virtio_vsock_hdr *hdr;
>>>> -	struct sk_buff *skb;
>>>> -	void *payload;
>>>> -	int err;
>>>> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
>>>> +				       size_t max_to_send)
>>>> +{
>>>> +	const struct vsock_transport *t;
>>>> +	struct iov_iter *iov_iter;
>>>>  
>>>> -	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>>> -	if (!skb)
>>>> -		return NULL;
>>>> +	if (!info->msg)
>>>> +		return false;
>>>>  
>>>> -	hdr = virtio_vsock_hdr(skb);
>>>> -	hdr->type	= cpu_to_le16(info->type);
>>>> -	hdr->op		= cpu_to_le16(info->op);
>>>> -	hdr->src_cid	= cpu_to_le64(src_cid);
>>>> -	hdr->dst_cid	= cpu_to_le64(dst_cid);
>>>> -	hdr->src_port	= cpu_to_le32(src_port);
>>>> -	hdr->dst_port	= cpu_to_le32(dst_port);
>>>> -	hdr->flags	= cpu_to_le32(info->flags);
>>>> -	hdr->len	= cpu_to_le32(len);
>>>> +	iov_iter = &info->msg->msg_iter;
>>>>  
>>>> -	if (info->msg && len > 0) {
>>>> -		payload = skb_put(skb, len);
>>>> -		err = memcpy_from_msg(payload, info->msg, len);
>>>> -		if (err)
>>>> -			goto out;
>>>> +	t = vsock_core_get_transport(info->vsk);
>>>>  
>>>> -		if (msg_data_left(info->msg) == 0 &&
>>>> -		    info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>>> -			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>>> +	if (t->msgzerocopy_check_iov &&
>>>> +	    !t->msgzerocopy_check_iov(iov_iter))
>>>> +		return false;
>>>>  
>>>> -			if (info->msg->msg_flags & MSG_EOR)
>>>> -				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>>> -		}
>>>> +	/* Data is simple buffer. */
>>>> +	if (iter_is_ubuf(iov_iter))
>>>> +		return true;
>>>> +
>>>> +	if (!iter_is_iovec(iov_iter))
>>>> +		return false;
>>>> +
>>>> +	if (iov_iter->iov_offset)
>>>> +		return false;
>>>> +
>>>> +	/* We can't send whole iov. */
>>>> +	if (iov_iter->count > max_to_send)
>>>> +		return false;
>>>> +
>>>> +	return true;
>>>> +}
>>>> +
>>>> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
>>>> +					   struct sk_buff *skb,
>>>> +					   struct msghdr *msg,
>>>> +					   bool zerocopy)
>>>> +{
>>>> +	struct ubuf_info *uarg;
>>>> +
>>>> +	if (msg->msg_ubuf) {
>>>> +		uarg = msg->msg_ubuf;
>>>> +		net_zcopy_get(uarg);
>>>> +	} else {
>>>> +		struct iov_iter *iter = &msg->msg_iter;
>>>> +		struct ubuf_info_msgzc *uarg_zc;
>>>> +		int len;
>>>> +
>>>> +		/* Only ITER_IOVEC or ITER_UBUF are allowed and
>>>> +		 * checked before.
>>>> +		 */
>>>> +		if (iter_is_iovec(iter))
>>>> +			len = iov_length(iter->__iov, iter->nr_segs);
>>>> +		else
>>>> +			len = iter->count;
>>>> +
>>>> +		uarg = msg_zerocopy_realloc(sk_vsock(vsk),
>>>> +					    len,
>>>> +					    NULL);
>>>> +		if (!uarg)
>>>> +			return -1;
>>>> +
>>>> +		uarg_zc = uarg_to_msgzc(uarg);
>>>> +		uarg_zc->zerocopy = zerocopy ? 1 : 0;
>>>>  	}
>>>>  
>>>> -	if (info->reply)
>>>> -		virtio_vsock_skb_set_reply(skb);
>>>> +	skb_zcopy_init(skb, uarg);
>>>>  
>>>> -	trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>>> -					 dst_cid, dst_port,
>>>> -					 len,
>>>> -					 info->type,
>>>> -					 info->op,
>>>> -					 info->flags);
>>>> +	return 0;
>>>> +}
>>>>  
>>>> -	if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
>>>> -		WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
>>>> -		goto out;
>>>> +static int virtio_transport_fill_skb(struct sk_buff *skb,
>>>> +				     struct virtio_vsock_pkt_info *info,
>>>> +				     size_t len,
>>>> +				     bool zcopy)
>>>> +{
>>>> +	if (zcopy) {
>>>> +		return __zerocopy_sg_from_iter(info->msg, NULL, skb,
>>>> +					      &info->msg->msg_iter,
>>>> +					      len);
>>>> +	} else {
>>>> +		void *payload;
>>>> +		int err;
>>>> +
>>>> +		payload = skb_put(skb, len);
>>>> +		err = memcpy_from_msg(payload, info->msg, len);
>>>> +		if (err)
>>>> +			return -1;
>>>> +
>>>> +		if (msg_data_left(info->msg))
>>>> +			return 0;
>>>> +
>>>> +		return 0;
>>>>  	}
>>>> +}
>>>>  
>>>> -	return skb;
>>>> +static void virtio_transport_init_hdr(struct sk_buff *skb,
>>>> +				      struct virtio_vsock_pkt_info *info,
>>>> +				      u32 src_cid,
>>>> +				      u32 src_port,
>>>> +				      u32 dst_cid,
>>>> +				      u32 dst_port,
>>>> +				      size_t len)
>>>> +{
>>>> +	struct virtio_vsock_hdr *hdr;
>>>>  
>>>> -out:
>>>> -	kfree_skb(skb);
>>>> -	return NULL;
>>>> +	hdr = virtio_vsock_hdr(skb);
>>>> +	hdr->type	= cpu_to_le16(info->type);
>>>> +	hdr->op		= cpu_to_le16(info->op);
>>>> +	hdr->src_cid	= cpu_to_le64(src_cid);
>>>> +	hdr->dst_cid	= cpu_to_le64(dst_cid);
>>>> +	hdr->src_port	= cpu_to_le32(src_port);
>>>> +	hdr->dst_port	= cpu_to_le32(dst_port);
>>>> +	hdr->flags	= cpu_to_le32(info->flags);
>>>> +	hdr->len	= cpu_to_le32(len);
>>>>  }
>>>>  
>>>>  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
>>>> @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>>  		return VIRTIO_VSOCK_TYPE_SEQPACKET;
>>>>  }
>>>>  
>>>> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
>>>> +						  struct virtio_vsock_pkt_info *info,
>>>> +						  size_t payload_len,
>>>> +						  bool zcopy,
>>>> +						  u32 src_cid,
>>>> +						  u32 src_port,
>>>> +						  u32 dst_cid,
>>>> +						  u32 dst_port)
>>>> +{
>>>> +	struct sk_buff *skb;
>>>> +	size_t skb_len;
>>>> +
>>>> +	skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
>>>> +
>>>> +	if (!zcopy)
>>>> +		skb_len += payload_len;
>>>> +
>>>> +	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>>> +	if (!skb)
>>>> +		return NULL;
>>>> +
>>>> +	virtio_transport_init_hdr(skb, info, src_cid, src_port,
>>>> +				  dst_cid, dst_port,
>>>> +				  payload_len);
>>>> +
>>>> +	/* Set owner here, because '__zerocopy_sg_from_iter()' uses
>>>> +	 * owner of skb without check to update 'sk_wmem_alloc'.
>>>> +	 */
>>>> +	if (vsk)
>>>> +		skb_set_owner_w(skb, sk_vsock(vsk));
>>>> +
>>>> +	if (info->msg && payload_len > 0) {
>>>> +		int err;
>>>> +
>>>> +		err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
>>>> +		if (err)
>>>> +			goto out;
>>>> +
>>>> +		if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>>> +			struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
>>>> +
>>>> +			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>>> +
>>>> +			if (info->msg->msg_flags & MSG_EOR)
>>>> +				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>>> +		}
>>>> +	}
>>>> +
>>>> +	if (info->reply)
>>>> +		virtio_vsock_skb_set_reply(skb);
>>>> +
>>>> +	trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>>> +					 dst_cid, dst_port,
>>>> +					 payload_len,
>>>> +					 info->type,
>>>> +					 info->op,
>>>> +					 info->flags);
>>>> +
>>>> +	return skb;
>>>> +out:
>>>> +	kfree_skb(skb);
>>>> +	return NULL;
>>>> +}
>>>> +
>>>>  /* This function can only be used on connecting/connected sockets,
>>>>   * since a socket assigned to a transport is required.
>>>>   *
>>>> @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>>  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>  					  struct virtio_vsock_pkt_info *info)
>>>>  {
>>>> +	u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
>>>>  	u32 src_cid, src_port, dst_cid, dst_port;
>>>>  	const struct virtio_transport *t_ops;
>>>>  	struct virtio_vsock_sock *vvs;
>>>>  	u32 pkt_len = info->pkt_len;
>>>> +	bool can_zcopy = false;
>>>>  	u32 rest_len;
>>>>  	int ret;
>>>>  
>>>> @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>  	if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
>>>>  		return pkt_len;
>>>>  
>>>> +	if (info->msg) {
>>>> +		/* If zerocopy is not enabled by 'setsockopt()', we behave as
>>>> +		 * there is no MSG_ZEROCOPY flag set.
>>>> +		 */
>>>> +		if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
>>>> +			info->msg->msg_flags &= ~MSG_ZEROCOPY;
>>>> +
>>>> +		if (info->msg->msg_flags & MSG_ZEROCOPY)
>>>> +			can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
>>>> +
>>>> +		if (can_zcopy)
>>>> +			max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
>>>> +					    (MAX_SKB_FRAGS * PAGE_SIZE));
>>>> +	}
>>>> +
>>>>  	rest_len = pkt_len;
>>>>  
>>>>  	do {
>>>>  		struct sk_buff *skb;
>>>>  		size_t skb_len;
>>>>  
>>>> -		skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
>>>> +		skb_len = min(max_skb_len, rest_len);
>>>>  
>>>> -		skb = virtio_transport_alloc_skb(info, skb_len,
>>>> +		skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
>>>>  						 src_cid, src_port,
>>>>  						 dst_cid, dst_port);
>>>>  		if (!skb) {
>>>> @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>  			break;
>>>>  		}
>>>>  
>>>> +		/* This is last skb to send this portion of data. */
>>>> +		if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
>>>> +		    skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
>>>> +			if (virtio_transport_init_zcopy_skb(vsk, skb,
>>>> +							    info->msg,
>>>> +							    can_zcopy)) {
>>>> +				ret = -ENOMEM;
>>>> +				break;
>>>> +			}
>>>> +		}
>>>> +
>>>>  		virtio_transport_inc_tx_pkt(vvs, skb);
>>>>  
>>>>  		ret = t_ops->send_pkt(skb);
>>>> @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
>>>>  	if (!t)
>>>>  		return -ENOTCONN;
>>>>  
>>>> -	reply = virtio_transport_alloc_skb(&info, 0,
>>>> +	reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
>>>>  					   le64_to_cpu(hdr->dst_cid),
>>>>  					   le32_to_cpu(hdr->dst_port),
>>>>  					   le64_to_cpu(hdr->src_cid),
>>>> -- 
>>>> 2.25.1
>>>
>
  
Stefano Garzarella July 25, 2023, 1:21 p.m. UTC | #16
On Tue, Jul 25, 2023 at 09:06:02AM -0400, Michael S. Tsirkin wrote:
>On Tue, Jul 25, 2023 at 02:53:39PM +0200, Stefano Garzarella wrote:
>> On Tue, Jul 25, 2023 at 07:50:53AM -0400, Michael S. Tsirkin wrote:
>> > On Fri, Jul 21, 2023 at 08:09:03AM +0300, Arseniy Krasnov wrote:
>> > >
>> > >
>> > > On 21.07.2023 00:42, Arseniy Krasnov wrote:
>> > > > This adds handling of MSG_ZEROCOPY flag on transmission path: if this
>> > > > flag is set and zerocopy transmission is possible (enabled in socket
>> > > > options and transport allows zerocopy), then non-linear skb will be
>> > > > created and filled with the pages of user's buffer. Pages of user's
>> > > > buffer are locked in memory by 'get_user_pages()'. Second thing that
>> > > > this patch does is replace type of skb owning: instead of calling
>> > > > 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
>> > > > change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
>> > > > of socket, so to decrease this field correctly proper skb destructor is
>> > > > needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
>> > > >
>> > > > Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
>> > > > ---
>> > > >  Changelog:
>> > > >  v5(big patchset) -> v1:
>> > > >   * Refactorings of 'if' conditions.
>> > > >   * Remove extra blank line.
>> > > >   * Remove 'frag_off' field unneeded init.
>> > > >   * Add function 'virtio_transport_fill_skb()' which fills both linear
>> > > >     and non-linear skb with provided data.
>> > > >  v1 -> v2:
>> > > >   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
>> > > >  v2 -> v3:
>> > > >   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
>> > > >     provided 'iov_iter' with data could be sent in a zerocopy mode.
>> > > >     If this callback is not set in transport - transport allows to send
>> > > >     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
>> > > >     then zerocopy is allowed. Reason of this callback is that in case of
>> > > >     G2H transmission we insert whole skb to the tx virtio queue and such
>> > > >     skb must fit to the size of the virtio queue to be sent in a single
>> > > >     iteration (may be tx logic in 'virtio_transport.c' could be reworked
>> > > >     as in vhost to support partial send of current skb). This callback
>> > > >     will be enabled only for G2H path. For details pls see comment
>> > > >     'Check that tx queue...' below.
>> > > >
>> > > >  include/net/af_vsock.h                  |   3 +
>> > > >  net/vmw_vsock/virtio_transport.c        |  39 ++++
>> > > >  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
>> > > >  3 files changed, 241 insertions(+), 58 deletions(-)
>> > > >
>> > > > diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
>> > > > index 0e7504a42925..a6b346eeeb8e 100644
>> > > > --- a/include/net/af_vsock.h
>> > > > +++ b/include/net/af_vsock.h
>> > > > @@ -177,6 +177,9 @@ struct vsock_transport {
>> > > >
>> > > >  	/* Read a single skb */
>> > > >  	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
>> > > > +
>> > > > +	/* Zero-copy. */
>> > > > +	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
>> > > >  };
>> > > >
>> > > >  /**** CORE ****/
>> > > > diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>> > > > index 7bbcc8093e51..23cb8ed638c4 100644
>> > > > --- a/net/vmw_vsock/virtio_transport.c
>> > > > +++ b/net/vmw_vsock/virtio_transport.c
>> > > > @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
>> > > >  	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>> > > >  }
>> > > >
>> > > > +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
>> > > > +{
>> > > > +	struct virtio_vsock *vsock;
>> > > > +	bool res = false;
>> > > > +
>> > > > +	rcu_read_lock();
>> > > > +
>> > > > +	vsock = rcu_dereference(the_virtio_vsock);
>> > > > +	if (vsock) {
>> > > > +		struct virtqueue *vq;
>> > > > +		int iov_pages;
>> > > > +
>> > > > +		vq = vsock->vqs[VSOCK_VQ_TX];
>> > > > +
>> > > > +		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
>> > > > +
>> > > > +		/* Check that tx queue is large enough to keep whole
>> > > > +		 * data to send. This is needed, because when there is
>> > > > +		 * not enough free space in the queue, current skb to
>> > > > +		 * send will be reinserted to the head of tx list of
>> > > > +		 * the socket to retry transmission later, so if skb
>> > > > +		 * is bigger than whole queue, it will be reinserted
>> > > > +		 * again and again, thus blocking other skbs to be sent.
>> > > > +		 * Each page of the user provided buffer will be added
>> > > > +		 * as a single buffer to the tx virtqueue, so compare
>> > > > +		 * number of pages against maximum capacity of the queue.
>> > > > +		 * +1 means buffer for the packet header.
>> > > > +		 */
>> > > > +		if (iov_pages + 1 <= vq->num_max)
>> > >
>> > > I think this check is actual only for case one we don't have indirect buffer feature.
>> > > With indirect mode whole data to send will be packed into one indirect buffer.
>> > >
>> > > Thanks, Arseniy
>> >
>> > Actually the reverse. With indirect you are limited to num_max.
>> > Without you are limited to whatever space is left in the
>> > queue (which you did not check here, so you should).
>> >
>> >
>> > > > +			res = true;
>> > > > +	}
>> > > > +
>> > > > +	rcu_read_unlock();
>> >
>> > Just curious:
>> > is the point of all this RCU dance to allow vsock
>> > to change from under us? then why is it ok to
>> > have it change? the virtio_transport_msgzerocopy_check_iov
>> > will then refer to the old vsock ...
>>
>> IIRC we introduced the RCU to handle hot-unplug issues:
>> commit 0deab087b16a ("vsock/virtio: use RCU to avoid use-after-free on
>> the_virtio_vsock")
>>
>> When we remove the device, we flush all the works, etc. so we should
>> not be in this case (referring the old vsock), except for an irrelevant
>> transient as the device is disappearing.
>>
>> Stefano
>
>what if old device goes away then new one appears?

In virtio_vsock_remove() (.remove cb) we hold `the_virtio_vsock_mutex`
while flushing all the works/sockets/packets and sync the RCU.

In virtio_vsock_probe (.probe cb) we hold the same lock while adding
the new one and updating the RCU pointer. (only 1 virtio-vsock device
per guest is currently supported)

So when the new one appears, all the previous sockets are closed, all
the queue packets and pending works flushed.

So new packets will see the new vsock device. It looks safe to me.

Stefano
  
Michael S. Tsirkin July 25, 2023, 1:22 p.m. UTC | #17
On Tue, Jul 25, 2023 at 04:04:13PM +0300, Arseniy Krasnov wrote:
> 
> 
> On 25.07.2023 14:50, Michael S. Tsirkin wrote:
> > On Fri, Jul 21, 2023 at 08:09:03AM +0300, Arseniy Krasnov wrote:
> >>
> >>
> >> On 21.07.2023 00:42, Arseniy Krasnov wrote:
> >>> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
> >>> flag is set and zerocopy transmission is possible (enabled in socket
> >>> options and transport allows zerocopy), then non-linear skb will be
> >>> created and filled with the pages of user's buffer. Pages of user's
> >>> buffer are locked in memory by 'get_user_pages()'. Second thing that
> >>> this patch does is replace type of skb owning: instead of calling
> >>> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
> >>> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
> >>> of socket, so to decrease this field correctly proper skb destructor is
> >>> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
> >>>
> >>> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
> >>> ---
> >>>  Changelog:
> >>>  v5(big patchset) -> v1:
> >>>   * Refactorings of 'if' conditions.
> >>>   * Remove extra blank line.
> >>>   * Remove 'frag_off' field unneeded init.
> >>>   * Add function 'virtio_transport_fill_skb()' which fills both linear
> >>>     and non-linear skb with provided data.
> >>>  v1 -> v2:
> >>>   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
> >>>  v2 -> v3:
> >>>   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
> >>>     provided 'iov_iter' with data could be sent in a zerocopy mode.
> >>>     If this callback is not set in transport - transport allows to send
> >>>     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
> >>>     then zerocopy is allowed. Reason of this callback is that in case of
> >>>     G2H transmission we insert whole skb to the tx virtio queue and such
> >>>     skb must fit to the size of the virtio queue to be sent in a single
> >>>     iteration (may be tx logic in 'virtio_transport.c' could be reworked
> >>>     as in vhost to support partial send of current skb). This callback
> >>>     will be enabled only for G2H path. For details pls see comment 
> >>>     'Check that tx queue...' below.
> >>>
> >>>  include/net/af_vsock.h                  |   3 +
> >>>  net/vmw_vsock/virtio_transport.c        |  39 ++++
> >>>  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
> >>>  3 files changed, 241 insertions(+), 58 deletions(-)
> >>>
> >>> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
> >>> index 0e7504a42925..a6b346eeeb8e 100644
> >>> --- a/include/net/af_vsock.h
> >>> +++ b/include/net/af_vsock.h
> >>> @@ -177,6 +177,9 @@ struct vsock_transport {
> >>>  
> >>>  	/* Read a single skb */
> >>>  	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
> >>> +
> >>> +	/* Zero-copy. */
> >>> +	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
> >>>  };
> >>>  
> >>>  /**** CORE ****/
> >>> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
> >>> index 7bbcc8093e51..23cb8ed638c4 100644
> >>> --- a/net/vmw_vsock/virtio_transport.c
> >>> +++ b/net/vmw_vsock/virtio_transport.c
> >>> @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
> >>>  	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
> >>>  }
> >>>  
> >>> +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
> >>> +{
> >>> +	struct virtio_vsock *vsock;
> >>> +	bool res = false;
> >>> +
> >>> +	rcu_read_lock();
> >>> +
> >>> +	vsock = rcu_dereference(the_virtio_vsock);
> >>> +	if (vsock) {
> >>> +		struct virtqueue *vq;
> >>> +		int iov_pages;
> >>> +
> >>> +		vq = vsock->vqs[VSOCK_VQ_TX];
> >>> +
> >>> +		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
> >>> +
> >>> +		/* Check that tx queue is large enough to keep whole
> >>> +		 * data to send. This is needed, because when there is
> >>> +		 * not enough free space in the queue, current skb to
> >>> +		 * send will be reinserted to the head of tx list of
> >>> +		 * the socket to retry transmission later, so if skb
> >>> +		 * is bigger than whole queue, it will be reinserted
> >>> +		 * again and again, thus blocking other skbs to be sent.
> >>> +		 * Each page of the user provided buffer will be added
> >>> +		 * as a single buffer to the tx virtqueue, so compare
> >>> +		 * number of pages against maximum capacity of the queue.
> >>> +		 * +1 means buffer for the packet header.
> >>> +		 */
> >>> +		if (iov_pages + 1 <= vq->num_max)
> >>
> >> I think this check is actual only for case one we don't have indirect buffer feature.
> >> With indirect mode whole data to send will be packed into one indirect buffer.
> >>
> >> Thanks, Arseniy
> > 
> > Actually the reverse. With indirect you are limited to num_max.
> > Without you are limited to whatever space is left in the
> > queue (which you did not check here, so you should).
> 
> I mean that with indirect, we only need one buffer, and we can just wait
> for enough space - for this single buffer ( as we discussed a little bit before).
> But if indirect buffer is not supported - we need that whole packet must fit
> to the size of tx queue - otherwise it never be transmitted.
> 
> Thanks, Arseniy


yes but according to virtio spec it's illegal to add s/g that is bigger
than queue size.

> > 
> > 
> >>> +			res = true;
> >>> +	}
> >>> +
> >>> +	rcu_read_unlock();
> > 
> > Just curious:
> > is the point of all this RCU dance to allow vsock
> > to change from under us? then why is it ok to
> > have it change? the virtio_transport_msgzerocopy_check_iov
> > will then refer to the old vsock ...
> > 
> > 
> >>> +
> >>> +	return res;
> >>> +}
> >>> +
> >>>  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
> >>>  
> >>>  static struct virtio_transport virtio_transport = {
> >>> @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
> >>>  		.seqpacket_allow          = virtio_transport_seqpacket_allow,
> >>>  		.seqpacket_has_data       = virtio_transport_seqpacket_has_data,
> >>>  
> >>> +		.msgzerocopy_check_iov	  = virtio_transport_msgzerocopy_check_iov,
> >>> +
> >>>  		.notify_poll_in           = virtio_transport_notify_poll_in,
> >>>  		.notify_poll_out          = virtio_transport_notify_poll_out,
> >>>  		.notify_recv_init         = virtio_transport_notify_recv_init,
> >>> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
> >>> index 26a4d10da205..e4e3d541aff4 100644
> >>> --- a/net/vmw_vsock/virtio_transport_common.c
> >>> +++ b/net/vmw_vsock/virtio_transport_common.c
> >>> @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
> >>>  	return container_of(t, struct virtio_transport, transport);
> >>>  }
> >>>  
> >>> -/* Returns a new packet on success, otherwise returns NULL.
> >>> - *
> >>> - * If NULL is returned, errp is set to a negative errno.
> >>> - */
> >>> -static struct sk_buff *
> >>> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
> >>> -			   size_t len,
> >>> -			   u32 src_cid,
> >>> -			   u32 src_port,
> >>> -			   u32 dst_cid,
> >>> -			   u32 dst_port)
> >>> -{
> >>> -	const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
> >>> -	struct virtio_vsock_hdr *hdr;
> >>> -	struct sk_buff *skb;
> >>> -	void *payload;
> >>> -	int err;
> >>> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
> >>> +				       size_t max_to_send)
> >>> +{
> >>> +	const struct vsock_transport *t;
> >>> +	struct iov_iter *iov_iter;
> >>>  
> >>> -	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
> >>> -	if (!skb)
> >>> -		return NULL;
> >>> +	if (!info->msg)
> >>> +		return false;
> >>>  
> >>> -	hdr = virtio_vsock_hdr(skb);
> >>> -	hdr->type	= cpu_to_le16(info->type);
> >>> -	hdr->op		= cpu_to_le16(info->op);
> >>> -	hdr->src_cid	= cpu_to_le64(src_cid);
> >>> -	hdr->dst_cid	= cpu_to_le64(dst_cid);
> >>> -	hdr->src_port	= cpu_to_le32(src_port);
> >>> -	hdr->dst_port	= cpu_to_le32(dst_port);
> >>> -	hdr->flags	= cpu_to_le32(info->flags);
> >>> -	hdr->len	= cpu_to_le32(len);
> >>> +	iov_iter = &info->msg->msg_iter;
> >>>  
> >>> -	if (info->msg && len > 0) {
> >>> -		payload = skb_put(skb, len);
> >>> -		err = memcpy_from_msg(payload, info->msg, len);
> >>> -		if (err)
> >>> -			goto out;
> >>> +	t = vsock_core_get_transport(info->vsk);
> >>>  
> >>> -		if (msg_data_left(info->msg) == 0 &&
> >>> -		    info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
> >>> -			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
> >>> +	if (t->msgzerocopy_check_iov &&
> >>> +	    !t->msgzerocopy_check_iov(iov_iter))
> >>> +		return false;
> >>>  
> >>> -			if (info->msg->msg_flags & MSG_EOR)
> >>> -				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
> >>> -		}
> >>> +	/* Data is simple buffer. */
> >>> +	if (iter_is_ubuf(iov_iter))
> >>> +		return true;
> >>> +
> >>> +	if (!iter_is_iovec(iov_iter))
> >>> +		return false;
> >>> +
> >>> +	if (iov_iter->iov_offset)
> >>> +		return false;
> >>> +
> >>> +	/* We can't send whole iov. */
> >>> +	if (iov_iter->count > max_to_send)
> >>> +		return false;
> >>> +
> >>> +	return true;
> >>> +}
> >>> +
> >>> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
> >>> +					   struct sk_buff *skb,
> >>> +					   struct msghdr *msg,
> >>> +					   bool zerocopy)
> >>> +{
> >>> +	struct ubuf_info *uarg;
> >>> +
> >>> +	if (msg->msg_ubuf) {
> >>> +		uarg = msg->msg_ubuf;
> >>> +		net_zcopy_get(uarg);
> >>> +	} else {
> >>> +		struct iov_iter *iter = &msg->msg_iter;
> >>> +		struct ubuf_info_msgzc *uarg_zc;
> >>> +		int len;
> >>> +
> >>> +		/* Only ITER_IOVEC or ITER_UBUF are allowed and
> >>> +		 * checked before.
> >>> +		 */
> >>> +		if (iter_is_iovec(iter))
> >>> +			len = iov_length(iter->__iov, iter->nr_segs);
> >>> +		else
> >>> +			len = iter->count;
> >>> +
> >>> +		uarg = msg_zerocopy_realloc(sk_vsock(vsk),
> >>> +					    len,
> >>> +					    NULL);
> >>> +		if (!uarg)
> >>> +			return -1;
> >>> +
> >>> +		uarg_zc = uarg_to_msgzc(uarg);
> >>> +		uarg_zc->zerocopy = zerocopy ? 1 : 0;
> >>>  	}
> >>>  
> >>> -	if (info->reply)
> >>> -		virtio_vsock_skb_set_reply(skb);
> >>> +	skb_zcopy_init(skb, uarg);
> >>>  
> >>> -	trace_virtio_transport_alloc_pkt(src_cid, src_port,
> >>> -					 dst_cid, dst_port,
> >>> -					 len,
> >>> -					 info->type,
> >>> -					 info->op,
> >>> -					 info->flags);
> >>> +	return 0;
> >>> +}
> >>>  
> >>> -	if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
> >>> -		WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
> >>> -		goto out;
> >>> +static int virtio_transport_fill_skb(struct sk_buff *skb,
> >>> +				     struct virtio_vsock_pkt_info *info,
> >>> +				     size_t len,
> >>> +				     bool zcopy)
> >>> +{
> >>> +	if (zcopy) {
> >>> +		return __zerocopy_sg_from_iter(info->msg, NULL, skb,
> >>> +					      &info->msg->msg_iter,
> >>> +					      len);
> >>> +	} else {
> >>> +		void *payload;
> >>> +		int err;
> >>> +
> >>> +		payload = skb_put(skb, len);
> >>> +		err = memcpy_from_msg(payload, info->msg, len);
> >>> +		if (err)
> >>> +			return -1;
> >>> +
> >>> +		if (msg_data_left(info->msg))
> >>> +			return 0;
> >>> +
> >>> +		return 0;
> >>>  	}
> >>> +}
> >>>  
> >>> -	return skb;
> >>> +static void virtio_transport_init_hdr(struct sk_buff *skb,
> >>> +				      struct virtio_vsock_pkt_info *info,
> >>> +				      u32 src_cid,
> >>> +				      u32 src_port,
> >>> +				      u32 dst_cid,
> >>> +				      u32 dst_port,
> >>> +				      size_t len)
> >>> +{
> >>> +	struct virtio_vsock_hdr *hdr;
> >>>  
> >>> -out:
> >>> -	kfree_skb(skb);
> >>> -	return NULL;
> >>> +	hdr = virtio_vsock_hdr(skb);
> >>> +	hdr->type	= cpu_to_le16(info->type);
> >>> +	hdr->op		= cpu_to_le16(info->op);
> >>> +	hdr->src_cid	= cpu_to_le64(src_cid);
> >>> +	hdr->dst_cid	= cpu_to_le64(dst_cid);
> >>> +	hdr->src_port	= cpu_to_le32(src_port);
> >>> +	hdr->dst_port	= cpu_to_le32(dst_port);
> >>> +	hdr->flags	= cpu_to_le32(info->flags);
> >>> +	hdr->len	= cpu_to_le32(len);
> >>>  }
> >>>  
> >>>  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
> >>> @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
> >>>  		return VIRTIO_VSOCK_TYPE_SEQPACKET;
> >>>  }
> >>>  
> >>> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
> >>> +						  struct virtio_vsock_pkt_info *info,
> >>> +						  size_t payload_len,
> >>> +						  bool zcopy,
> >>> +						  u32 src_cid,
> >>> +						  u32 src_port,
> >>> +						  u32 dst_cid,
> >>> +						  u32 dst_port)
> >>> +{
> >>> +	struct sk_buff *skb;
> >>> +	size_t skb_len;
> >>> +
> >>> +	skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
> >>> +
> >>> +	if (!zcopy)
> >>> +		skb_len += payload_len;
> >>> +
> >>> +	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
> >>> +	if (!skb)
> >>> +		return NULL;
> >>> +
> >>> +	virtio_transport_init_hdr(skb, info, src_cid, src_port,
> >>> +				  dst_cid, dst_port,
> >>> +				  payload_len);
> >>> +
> >>> +	/* Set owner here, because '__zerocopy_sg_from_iter()' uses
> >>> +	 * owner of skb without check to update 'sk_wmem_alloc'.
> >>> +	 */
> >>> +	if (vsk)
> >>> +		skb_set_owner_w(skb, sk_vsock(vsk));
> >>> +
> >>> +	if (info->msg && payload_len > 0) {
> >>> +		int err;
> >>> +
> >>> +		err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
> >>> +		if (err)
> >>> +			goto out;
> >>> +
> >>> +		if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
> >>> +			struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
> >>> +
> >>> +			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
> >>> +
> >>> +			if (info->msg->msg_flags & MSG_EOR)
> >>> +				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
> >>> +		}
> >>> +	}
> >>> +
> >>> +	if (info->reply)
> >>> +		virtio_vsock_skb_set_reply(skb);
> >>> +
> >>> +	trace_virtio_transport_alloc_pkt(src_cid, src_port,
> >>> +					 dst_cid, dst_port,
> >>> +					 payload_len,
> >>> +					 info->type,
> >>> +					 info->op,
> >>> +					 info->flags);
> >>> +
> >>> +	return skb;
> >>> +out:
> >>> +	kfree_skb(skb);
> >>> +	return NULL;
> >>> +}
> >>> +
> >>>  /* This function can only be used on connecting/connected sockets,
> >>>   * since a socket assigned to a transport is required.
> >>>   *
> >>> @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
> >>>  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
> >>>  					  struct virtio_vsock_pkt_info *info)
> >>>  {
> >>> +	u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
> >>>  	u32 src_cid, src_port, dst_cid, dst_port;
> >>>  	const struct virtio_transport *t_ops;
> >>>  	struct virtio_vsock_sock *vvs;
> >>>  	u32 pkt_len = info->pkt_len;
> >>> +	bool can_zcopy = false;
> >>>  	u32 rest_len;
> >>>  	int ret;
> >>>  
> >>> @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
> >>>  	if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
> >>>  		return pkt_len;
> >>>  
> >>> +	if (info->msg) {
> >>> +		/* If zerocopy is not enabled by 'setsockopt()', we behave as
> >>> +		 * there is no MSG_ZEROCOPY flag set.
> >>> +		 */
> >>> +		if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
> >>> +			info->msg->msg_flags &= ~MSG_ZEROCOPY;
> >>> +
> >>> +		if (info->msg->msg_flags & MSG_ZEROCOPY)
> >>> +			can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
> >>> +
> >>> +		if (can_zcopy)
> >>> +			max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
> >>> +					    (MAX_SKB_FRAGS * PAGE_SIZE));
> >>> +	}
> >>> +
> >>>  	rest_len = pkt_len;
> >>>  
> >>>  	do {
> >>>  		struct sk_buff *skb;
> >>>  		size_t skb_len;
> >>>  
> >>> -		skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
> >>> +		skb_len = min(max_skb_len, rest_len);
> >>>  
> >>> -		skb = virtio_transport_alloc_skb(info, skb_len,
> >>> +		skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
> >>>  						 src_cid, src_port,
> >>>  						 dst_cid, dst_port);
> >>>  		if (!skb) {
> >>> @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
> >>>  			break;
> >>>  		}
> >>>  
> >>> +		/* This is last skb to send this portion of data. */
> >>> +		if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
> >>> +		    skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
> >>> +			if (virtio_transport_init_zcopy_skb(vsk, skb,
> >>> +							    info->msg,
> >>> +							    can_zcopy)) {
> >>> +				ret = -ENOMEM;
> >>> +				break;
> >>> +			}
> >>> +		}
> >>> +
> >>>  		virtio_transport_inc_tx_pkt(vvs, skb);
> >>>  
> >>>  		ret = t_ops->send_pkt(skb);
> >>> @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
> >>>  	if (!t)
> >>>  		return -ENOTCONN;
> >>>  
> >>> -	reply = virtio_transport_alloc_skb(&info, 0,
> >>> +	reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
> >>>  					   le64_to_cpu(hdr->dst_cid),
> >>>  					   le32_to_cpu(hdr->dst_port),
> >>>  					   le64_to_cpu(hdr->src_cid),
> >
  
Michael S. Tsirkin July 25, 2023, 1:23 p.m. UTC | #18
On Tue, Jul 25, 2023 at 04:10:40PM +0300, Arseniy Krasnov wrote:
> 
> 
> On 25.07.2023 14:59, Michael S. Tsirkin wrote:
> > On Tue, Jul 25, 2023 at 11:39:22AM +0300, Arseniy Krasnov wrote:
> >>
> >>
> >> On 25.07.2023 11:25, Michael S. Tsirkin wrote:
> >>> On Fri, Jul 21, 2023 at 12:42:45AM +0300, Arseniy Krasnov wrote:
> >>>> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
> >>>> flag is set and zerocopy transmission is possible (enabled in socket
> >>>> options and transport allows zerocopy), then non-linear skb will be
> >>>> created and filled with the pages of user's buffer. Pages of user's
> >>>> buffer are locked in memory by 'get_user_pages()'. Second thing that
> >>>> this patch does is replace type of skb owning: instead of calling
> >>>> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
> >>>> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
> >>>> of socket, so to decrease this field correctly proper skb destructor is
> >>>> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
> >>>>
> >>>> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
> >>>> ---
> >>>>  Changelog:
> >>>>  v5(big patchset) -> v1:
> >>>>   * Refactorings of 'if' conditions.
> >>>>   * Remove extra blank line.
> >>>>   * Remove 'frag_off' field unneeded init.
> >>>>   * Add function 'virtio_transport_fill_skb()' which fills both linear
> >>>>     and non-linear skb with provided data.
> >>>>  v1 -> v2:
> >>>>   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
> >>>>  v2 -> v3:
> >>>>   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
> >>>>     provided 'iov_iter' with data could be sent in a zerocopy mode.
> >>>>     If this callback is not set in transport - transport allows to send
> >>>>     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
> >>>>     then zerocopy is allowed. Reason of this callback is that in case of
> >>>>     G2H transmission we insert whole skb to the tx virtio queue and such
> >>>>     skb must fit to the size of the virtio queue to be sent in a single
> >>>>     iteration (may be tx logic in 'virtio_transport.c' could be reworked
> >>>>     as in vhost to support partial send of current skb). This callback
> >>>>     will be enabled only for G2H path. For details pls see comment 
> >>>>     'Check that tx queue...' below.
> >>>>
> >>>>  include/net/af_vsock.h                  |   3 +
> >>>>  net/vmw_vsock/virtio_transport.c        |  39 ++++
> >>>>  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
> >>>>  3 files changed, 241 insertions(+), 58 deletions(-)
> >>>>
> >>>> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
> >>>> index 0e7504a42925..a6b346eeeb8e 100644
> >>>> --- a/include/net/af_vsock.h
> >>>> +++ b/include/net/af_vsock.h
> >>>> @@ -177,6 +177,9 @@ struct vsock_transport {
> >>>>  
> >>>>  	/* Read a single skb */
> >>>>  	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
> >>>> +
> >>>> +	/* Zero-copy. */
> >>>> +	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
> >>>>  };
> >>>>  
> >>>>  /**** CORE ****/
> >>>> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
> >>>> index 7bbcc8093e51..23cb8ed638c4 100644
> >>>> --- a/net/vmw_vsock/virtio_transport.c
> >>>> +++ b/net/vmw_vsock/virtio_transport.c
> >>>> @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
> >>>>  	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
> >>>>  }
> >>>>  
> >>>> +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
> >>>> +{
> >>>> +	struct virtio_vsock *vsock;
> >>>> +	bool res = false;
> >>>> +
> >>>> +	rcu_read_lock();
> >>>> +
> >>>> +	vsock = rcu_dereference(the_virtio_vsock);
> >>>> +	if (vsock) {
> >>>> +		struct virtqueue *vq;
> >>>> +		int iov_pages;
> >>>> +
> >>>> +		vq = vsock->vqs[VSOCK_VQ_TX];
> >>>> +
> >>>> +		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
> >>>> +
> >>>> +		/* Check that tx queue is large enough to keep whole
> >>>> +		 * data to send. This is needed, because when there is
> >>>> +		 * not enough free space in the queue, current skb to
> >>>> +		 * send will be reinserted to the head of tx list of
> >>>> +		 * the socket to retry transmission later, so if skb
> >>>> +		 * is bigger than whole queue, it will be reinserted
> >>>> +		 * again and again, thus blocking other skbs to be sent.
> >>>> +		 * Each page of the user provided buffer will be added
> >>>> +		 * as a single buffer to the tx virtqueue, so compare
> >>>> +		 * number of pages against maximum capacity of the queue.
> >>>> +		 * +1 means buffer for the packet header.
> >>>> +		 */
> >>>> +		if (iov_pages + 1 <= vq->num_max)
> >>>> +			res = true;
> >>>
> >>>
> >>> Yes but can't there already be buffers in the queue?
> >>> Then you can't stick num_max there.
> >>
> >> I think, that it is not critical, because vhost part always tries to process all
> >> incoming buffers (yes, 'vhost_exceeds_weight()' breaks at some moment, but it will
> >> reschedule tx kick ('vhost_vsock_handle_tx_kick()') work again), so current "too
> >> big" skb will wait until there will be enough space in queue and as it is requeued
> >> to the head of tx list it will be inserted to tx queue first.
> >>
> >> But anyway, I agree that comparing to 'num_free' may be more effective to the whole
> >> system performance...
> >>
> >> Thanks, Arseniy
> > 
> > Oh I see. It makes sense then - instead of copying just so we can
> > stick it in the queue, wait a bit and send later.
> > Also - for stream transports can't the message be split
> > and sent chunk by chunk? Better than copying ...
> 
> Technically yes, also we can split message for non-stream sockets (as vhost
> does when it copies data to rx buffers of the guest),

Won't breaking up messages break applications though?


> but it requires to rework
> current implementation by adding buffers one by one to the tx queue. I think
> it was not implemented here because until MSG_ZEROCOPY all skbs requires one
> (if it is control message) or two (with payload) buffers, so there is no big
> sense in processing max two buffers in "one-by-one" mode - we can just wait
> for space.
> 
> May be, I can add this logic for non-linear skb's here: 
> 
> if (skb->len > vq->max_num)
>     add buffers "one-by-one", incrementing internal offset in skb,
>    if (new skb insertion fails)
>        requeue skb, wait for space.
> 
> In TX done callback I'll call consume skb only when above mentioned internal
> offset == skb->len. I think this approach allows to get rid of new 'check_iov'
> callback from this patch.
> 
> 
> Stefano, what do You think?
> 
> Thanks, Arseniy
> 
> > 
> > 
> >>>
> >>>
> >>>> +	}
> >>>> +
> >>>> +	rcu_read_unlock();
> >>>> +
> >>>> +	return res;
> >>>> +}
> >>>> +
> >>>>  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
> >>>>  
> >>>>  static struct virtio_transport virtio_transport = {
> >>>> @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
> >>>>  		.seqpacket_allow          = virtio_transport_seqpacket_allow,
> >>>>  		.seqpacket_has_data       = virtio_transport_seqpacket_has_data,
> >>>>  
> >>>> +		.msgzerocopy_check_iov	  = virtio_transport_msgzerocopy_check_iov,
> >>>> +
> >>>>  		.notify_poll_in           = virtio_transport_notify_poll_in,
> >>>>  		.notify_poll_out          = virtio_transport_notify_poll_out,
> >>>>  		.notify_recv_init         = virtio_transport_notify_recv_init,
> >>>> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
> >>>> index 26a4d10da205..e4e3d541aff4 100644
> >>>> --- a/net/vmw_vsock/virtio_transport_common.c
> >>>> +++ b/net/vmw_vsock/virtio_transport_common.c
> >>>> @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
> >>>>  	return container_of(t, struct virtio_transport, transport);
> >>>>  }
> >>>>  
> >>>> -/* Returns a new packet on success, otherwise returns NULL.
> >>>> - *
> >>>> - * If NULL is returned, errp is set to a negative errno.
> >>>> - */
> >>>> -static struct sk_buff *
> >>>> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
> >>>> -			   size_t len,
> >>>> -			   u32 src_cid,
> >>>> -			   u32 src_port,
> >>>> -			   u32 dst_cid,
> >>>> -			   u32 dst_port)
> >>>> -{
> >>>> -	const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
> >>>> -	struct virtio_vsock_hdr *hdr;
> >>>> -	struct sk_buff *skb;
> >>>> -	void *payload;
> >>>> -	int err;
> >>>> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
> >>>> +				       size_t max_to_send)
> >>>> +{
> >>>> +	const struct vsock_transport *t;
> >>>> +	struct iov_iter *iov_iter;
> >>>>  
> >>>> -	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
> >>>> -	if (!skb)
> >>>> -		return NULL;
> >>>> +	if (!info->msg)
> >>>> +		return false;
> >>>>  
> >>>> -	hdr = virtio_vsock_hdr(skb);
> >>>> -	hdr->type	= cpu_to_le16(info->type);
> >>>> -	hdr->op		= cpu_to_le16(info->op);
> >>>> -	hdr->src_cid	= cpu_to_le64(src_cid);
> >>>> -	hdr->dst_cid	= cpu_to_le64(dst_cid);
> >>>> -	hdr->src_port	= cpu_to_le32(src_port);
> >>>> -	hdr->dst_port	= cpu_to_le32(dst_port);
> >>>> -	hdr->flags	= cpu_to_le32(info->flags);
> >>>> -	hdr->len	= cpu_to_le32(len);
> >>>> +	iov_iter = &info->msg->msg_iter;
> >>>>  
> >>>> -	if (info->msg && len > 0) {
> >>>> -		payload = skb_put(skb, len);
> >>>> -		err = memcpy_from_msg(payload, info->msg, len);
> >>>> -		if (err)
> >>>> -			goto out;
> >>>> +	t = vsock_core_get_transport(info->vsk);
> >>>>  
> >>>> -		if (msg_data_left(info->msg) == 0 &&
> >>>> -		    info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
> >>>> -			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
> >>>> +	if (t->msgzerocopy_check_iov &&
> >>>> +	    !t->msgzerocopy_check_iov(iov_iter))
> >>>> +		return false;
> >>>>  
> >>>> -			if (info->msg->msg_flags & MSG_EOR)
> >>>> -				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
> >>>> -		}
> >>>> +	/* Data is simple buffer. */
> >>>> +	if (iter_is_ubuf(iov_iter))
> >>>> +		return true;
> >>>> +
> >>>> +	if (!iter_is_iovec(iov_iter))
> >>>> +		return false;
> >>>> +
> >>>> +	if (iov_iter->iov_offset)
> >>>> +		return false;
> >>>> +
> >>>> +	/* We can't send whole iov. */
> >>>> +	if (iov_iter->count > max_to_send)
> >>>> +		return false;
> >>>> +
> >>>> +	return true;
> >>>> +}
> >>>> +
> >>>> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
> >>>> +					   struct sk_buff *skb,
> >>>> +					   struct msghdr *msg,
> >>>> +					   bool zerocopy)
> >>>> +{
> >>>> +	struct ubuf_info *uarg;
> >>>> +
> >>>> +	if (msg->msg_ubuf) {
> >>>> +		uarg = msg->msg_ubuf;
> >>>> +		net_zcopy_get(uarg);
> >>>> +	} else {
> >>>> +		struct iov_iter *iter = &msg->msg_iter;
> >>>> +		struct ubuf_info_msgzc *uarg_zc;
> >>>> +		int len;
> >>>> +
> >>>> +		/* Only ITER_IOVEC or ITER_UBUF are allowed and
> >>>> +		 * checked before.
> >>>> +		 */
> >>>> +		if (iter_is_iovec(iter))
> >>>> +			len = iov_length(iter->__iov, iter->nr_segs);
> >>>> +		else
> >>>> +			len = iter->count;
> >>>> +
> >>>> +		uarg = msg_zerocopy_realloc(sk_vsock(vsk),
> >>>> +					    len,
> >>>> +					    NULL);
> >>>> +		if (!uarg)
> >>>> +			return -1;
> >>>> +
> >>>> +		uarg_zc = uarg_to_msgzc(uarg);
> >>>> +		uarg_zc->zerocopy = zerocopy ? 1 : 0;
> >>>>  	}
> >>>>  
> >>>> -	if (info->reply)
> >>>> -		virtio_vsock_skb_set_reply(skb);
> >>>> +	skb_zcopy_init(skb, uarg);
> >>>>  
> >>>> -	trace_virtio_transport_alloc_pkt(src_cid, src_port,
> >>>> -					 dst_cid, dst_port,
> >>>> -					 len,
> >>>> -					 info->type,
> >>>> -					 info->op,
> >>>> -					 info->flags);
> >>>> +	return 0;
> >>>> +}
> >>>>  
> >>>> -	if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
> >>>> -		WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
> >>>> -		goto out;
> >>>> +static int virtio_transport_fill_skb(struct sk_buff *skb,
> >>>> +				     struct virtio_vsock_pkt_info *info,
> >>>> +				     size_t len,
> >>>> +				     bool zcopy)
> >>>> +{
> >>>> +	if (zcopy) {
> >>>> +		return __zerocopy_sg_from_iter(info->msg, NULL, skb,
> >>>> +					      &info->msg->msg_iter,
> >>>> +					      len);
> >>>> +	} else {
> >>>> +		void *payload;
> >>>> +		int err;
> >>>> +
> >>>> +		payload = skb_put(skb, len);
> >>>> +		err = memcpy_from_msg(payload, info->msg, len);
> >>>> +		if (err)
> >>>> +			return -1;
> >>>> +
> >>>> +		if (msg_data_left(info->msg))
> >>>> +			return 0;
> >>>> +
> >>>> +		return 0;
> >>>>  	}
> >>>> +}
> >>>>  
> >>>> -	return skb;
> >>>> +static void virtio_transport_init_hdr(struct sk_buff *skb,
> >>>> +				      struct virtio_vsock_pkt_info *info,
> >>>> +				      u32 src_cid,
> >>>> +				      u32 src_port,
> >>>> +				      u32 dst_cid,
> >>>> +				      u32 dst_port,
> >>>> +				      size_t len)
> >>>> +{
> >>>> +	struct virtio_vsock_hdr *hdr;
> >>>>  
> >>>> -out:
> >>>> -	kfree_skb(skb);
> >>>> -	return NULL;
> >>>> +	hdr = virtio_vsock_hdr(skb);
> >>>> +	hdr->type	= cpu_to_le16(info->type);
> >>>> +	hdr->op		= cpu_to_le16(info->op);
> >>>> +	hdr->src_cid	= cpu_to_le64(src_cid);
> >>>> +	hdr->dst_cid	= cpu_to_le64(dst_cid);
> >>>> +	hdr->src_port	= cpu_to_le32(src_port);
> >>>> +	hdr->dst_port	= cpu_to_le32(dst_port);
> >>>> +	hdr->flags	= cpu_to_le32(info->flags);
> >>>> +	hdr->len	= cpu_to_le32(len);
> >>>>  }
> >>>>  
> >>>>  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
> >>>> @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
> >>>>  		return VIRTIO_VSOCK_TYPE_SEQPACKET;
> >>>>  }
> >>>>  
> >>>> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
> >>>> +						  struct virtio_vsock_pkt_info *info,
> >>>> +						  size_t payload_len,
> >>>> +						  bool zcopy,
> >>>> +						  u32 src_cid,
> >>>> +						  u32 src_port,
> >>>> +						  u32 dst_cid,
> >>>> +						  u32 dst_port)
> >>>> +{
> >>>> +	struct sk_buff *skb;
> >>>> +	size_t skb_len;
> >>>> +
> >>>> +	skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
> >>>> +
> >>>> +	if (!zcopy)
> >>>> +		skb_len += payload_len;
> >>>> +
> >>>> +	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
> >>>> +	if (!skb)
> >>>> +		return NULL;
> >>>> +
> >>>> +	virtio_transport_init_hdr(skb, info, src_cid, src_port,
> >>>> +				  dst_cid, dst_port,
> >>>> +				  payload_len);
> >>>> +
> >>>> +	/* Set owner here, because '__zerocopy_sg_from_iter()' uses
> >>>> +	 * owner of skb without check to update 'sk_wmem_alloc'.
> >>>> +	 */
> >>>> +	if (vsk)
> >>>> +		skb_set_owner_w(skb, sk_vsock(vsk));
> >>>> +
> >>>> +	if (info->msg && payload_len > 0) {
> >>>> +		int err;
> >>>> +
> >>>> +		err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
> >>>> +		if (err)
> >>>> +			goto out;
> >>>> +
> >>>> +		if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
> >>>> +			struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
> >>>> +
> >>>> +			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
> >>>> +
> >>>> +			if (info->msg->msg_flags & MSG_EOR)
> >>>> +				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
> >>>> +		}
> >>>> +	}
> >>>> +
> >>>> +	if (info->reply)
> >>>> +		virtio_vsock_skb_set_reply(skb);
> >>>> +
> >>>> +	trace_virtio_transport_alloc_pkt(src_cid, src_port,
> >>>> +					 dst_cid, dst_port,
> >>>> +					 payload_len,
> >>>> +					 info->type,
> >>>> +					 info->op,
> >>>> +					 info->flags);
> >>>> +
> >>>> +	return skb;
> >>>> +out:
> >>>> +	kfree_skb(skb);
> >>>> +	return NULL;
> >>>> +}
> >>>> +
> >>>>  /* This function can only be used on connecting/connected sockets,
> >>>>   * since a socket assigned to a transport is required.
> >>>>   *
> >>>> @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
> >>>>  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
> >>>>  					  struct virtio_vsock_pkt_info *info)
> >>>>  {
> >>>> +	u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
> >>>>  	u32 src_cid, src_port, dst_cid, dst_port;
> >>>>  	const struct virtio_transport *t_ops;
> >>>>  	struct virtio_vsock_sock *vvs;
> >>>>  	u32 pkt_len = info->pkt_len;
> >>>> +	bool can_zcopy = false;
> >>>>  	u32 rest_len;
> >>>>  	int ret;
> >>>>  
> >>>> @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
> >>>>  	if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
> >>>>  		return pkt_len;
> >>>>  
> >>>> +	if (info->msg) {
> >>>> +		/* If zerocopy is not enabled by 'setsockopt()', we behave as
> >>>> +		 * there is no MSG_ZEROCOPY flag set.
> >>>> +		 */
> >>>> +		if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
> >>>> +			info->msg->msg_flags &= ~MSG_ZEROCOPY;
> >>>> +
> >>>> +		if (info->msg->msg_flags & MSG_ZEROCOPY)
> >>>> +			can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
> >>>> +
> >>>> +		if (can_zcopy)
> >>>> +			max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
> >>>> +					    (MAX_SKB_FRAGS * PAGE_SIZE));
> >>>> +	}
> >>>> +
> >>>>  	rest_len = pkt_len;
> >>>>  
> >>>>  	do {
> >>>>  		struct sk_buff *skb;
> >>>>  		size_t skb_len;
> >>>>  
> >>>> -		skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
> >>>> +		skb_len = min(max_skb_len, rest_len);
> >>>>  
> >>>> -		skb = virtio_transport_alloc_skb(info, skb_len,
> >>>> +		skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
> >>>>  						 src_cid, src_port,
> >>>>  						 dst_cid, dst_port);
> >>>>  		if (!skb) {
> >>>> @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
> >>>>  			break;
> >>>>  		}
> >>>>  
> >>>> +		/* This is last skb to send this portion of data. */
> >>>> +		if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
> >>>> +		    skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
> >>>> +			if (virtio_transport_init_zcopy_skb(vsk, skb,
> >>>> +							    info->msg,
> >>>> +							    can_zcopy)) {
> >>>> +				ret = -ENOMEM;
> >>>> +				break;
> >>>> +			}
> >>>> +		}
> >>>> +
> >>>>  		virtio_transport_inc_tx_pkt(vvs, skb);
> >>>>  
> >>>>  		ret = t_ops->send_pkt(skb);
> >>>> @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
> >>>>  	if (!t)
> >>>>  		return -ENOTCONN;
> >>>>  
> >>>> -	reply = virtio_transport_alloc_skb(&info, 0,
> >>>> +	reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
> >>>>  					   le64_to_cpu(hdr->dst_cid),
> >>>>  					   le32_to_cpu(hdr->dst_port),
> >>>>  					   le64_to_cpu(hdr->src_cid),
> >>>> -- 
> >>>> 2.25.1
> >>>
> >
  
Arseniy Krasnov July 25, 2023, 1:28 p.m. UTC | #19
On 25.07.2023 16:22, Michael S. Tsirkin wrote:
> On Tue, Jul 25, 2023 at 04:04:13PM +0300, Arseniy Krasnov wrote:
>>
>>
>> On 25.07.2023 14:50, Michael S. Tsirkin wrote:
>>> On Fri, Jul 21, 2023 at 08:09:03AM +0300, Arseniy Krasnov wrote:
>>>>
>>>>
>>>> On 21.07.2023 00:42, Arseniy Krasnov wrote:
>>>>> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
>>>>> flag is set and zerocopy transmission is possible (enabled in socket
>>>>> options and transport allows zerocopy), then non-linear skb will be
>>>>> created and filled with the pages of user's buffer. Pages of user's
>>>>> buffer are locked in memory by 'get_user_pages()'. Second thing that
>>>>> this patch does is replace type of skb owning: instead of calling
>>>>> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
>>>>> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
>>>>> of socket, so to decrease this field correctly proper skb destructor is
>>>>> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
>>>>>
>>>>> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
>>>>> ---
>>>>>  Changelog:
>>>>>  v5(big patchset) -> v1:
>>>>>   * Refactorings of 'if' conditions.
>>>>>   * Remove extra blank line.
>>>>>   * Remove 'frag_off' field unneeded init.
>>>>>   * Add function 'virtio_transport_fill_skb()' which fills both linear
>>>>>     and non-linear skb with provided data.
>>>>>  v1 -> v2:
>>>>>   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
>>>>>  v2 -> v3:
>>>>>   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
>>>>>     provided 'iov_iter' with data could be sent in a zerocopy mode.
>>>>>     If this callback is not set in transport - transport allows to send
>>>>>     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
>>>>>     then zerocopy is allowed. Reason of this callback is that in case of
>>>>>     G2H transmission we insert whole skb to the tx virtio queue and such
>>>>>     skb must fit to the size of the virtio queue to be sent in a single
>>>>>     iteration (may be tx logic in 'virtio_transport.c' could be reworked
>>>>>     as in vhost to support partial send of current skb). This callback
>>>>>     will be enabled only for G2H path. For details pls see comment 
>>>>>     'Check that tx queue...' below.
>>>>>
>>>>>  include/net/af_vsock.h                  |   3 +
>>>>>  net/vmw_vsock/virtio_transport.c        |  39 ++++
>>>>>  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
>>>>>  3 files changed, 241 insertions(+), 58 deletions(-)
>>>>>
>>>>> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
>>>>> index 0e7504a42925..a6b346eeeb8e 100644
>>>>> --- a/include/net/af_vsock.h
>>>>> +++ b/include/net/af_vsock.h
>>>>> @@ -177,6 +177,9 @@ struct vsock_transport {
>>>>>  
>>>>>  	/* Read a single skb */
>>>>>  	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
>>>>> +
>>>>> +	/* Zero-copy. */
>>>>> +	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
>>>>>  };
>>>>>  
>>>>>  /**** CORE ****/
>>>>> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>>>>> index 7bbcc8093e51..23cb8ed638c4 100644
>>>>> --- a/net/vmw_vsock/virtio_transport.c
>>>>> +++ b/net/vmw_vsock/virtio_transport.c
>>>>> @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
>>>>>  	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>>>>>  }
>>>>>  
>>>>> +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
>>>>> +{
>>>>> +	struct virtio_vsock *vsock;
>>>>> +	bool res = false;
>>>>> +
>>>>> +	rcu_read_lock();
>>>>> +
>>>>> +	vsock = rcu_dereference(the_virtio_vsock);
>>>>> +	if (vsock) {
>>>>> +		struct virtqueue *vq;
>>>>> +		int iov_pages;
>>>>> +
>>>>> +		vq = vsock->vqs[VSOCK_VQ_TX];
>>>>> +
>>>>> +		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
>>>>> +
>>>>> +		/* Check that tx queue is large enough to keep whole
>>>>> +		 * data to send. This is needed, because when there is
>>>>> +		 * not enough free space in the queue, current skb to
>>>>> +		 * send will be reinserted to the head of tx list of
>>>>> +		 * the socket to retry transmission later, so if skb
>>>>> +		 * is bigger than whole queue, it will be reinserted
>>>>> +		 * again and again, thus blocking other skbs to be sent.
>>>>> +		 * Each page of the user provided buffer will be added
>>>>> +		 * as a single buffer to the tx virtqueue, so compare
>>>>> +		 * number of pages against maximum capacity of the queue.
>>>>> +		 * +1 means buffer for the packet header.
>>>>> +		 */
>>>>> +		if (iov_pages + 1 <= vq->num_max)
>>>>
>>>> I think this check is actual only for case one we don't have indirect buffer feature.
>>>> With indirect mode whole data to send will be packed into one indirect buffer.
>>>>
>>>> Thanks, Arseniy
>>>
>>> Actually the reverse. With indirect you are limited to num_max.
>>> Without you are limited to whatever space is left in the
>>> queue (which you did not check here, so you should).
>>
>> I mean that with indirect, we only need one buffer, and we can just wait
>> for enough space - for this single buffer ( as we discussed a little bit before).
>> But if indirect buffer is not supported - we need that whole packet must fit
>> to the size of tx queue - otherwise it never be transmitted.
>>
>> Thanks, Arseniy
> 
> 
> yes but according to virtio spec it's illegal to add s/g that is bigger
> than queue size.

Aah, so even in case of indirect buffers feature, buffer descriptors stored in memory
pointed by indirect buffer must be accounted against queue size ?

Thanks, Arseniy

> 
>>>
>>>
>>>>> +			res = true;
>>>>> +	}
>>>>> +
>>>>> +	rcu_read_unlock();
>>>
>>> Just curious:
>>> is the point of all this RCU dance to allow vsock
>>> to change from under us? then why is it ok to
>>> have it change? the virtio_transport_msgzerocopy_check_iov
>>> will then refer to the old vsock ...
>>>
>>>
>>>>> +
>>>>> +	return res;
>>>>> +}
>>>>> +
>>>>>  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
>>>>>  
>>>>>  static struct virtio_transport virtio_transport = {
>>>>> @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
>>>>>  		.seqpacket_allow          = virtio_transport_seqpacket_allow,
>>>>>  		.seqpacket_has_data       = virtio_transport_seqpacket_has_data,
>>>>>  
>>>>> +		.msgzerocopy_check_iov	  = virtio_transport_msgzerocopy_check_iov,
>>>>> +
>>>>>  		.notify_poll_in           = virtio_transport_notify_poll_in,
>>>>>  		.notify_poll_out          = virtio_transport_notify_poll_out,
>>>>>  		.notify_recv_init         = virtio_transport_notify_recv_init,
>>>>> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
>>>>> index 26a4d10da205..e4e3d541aff4 100644
>>>>> --- a/net/vmw_vsock/virtio_transport_common.c
>>>>> +++ b/net/vmw_vsock/virtio_transport_common.c
>>>>> @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
>>>>>  	return container_of(t, struct virtio_transport, transport);
>>>>>  }
>>>>>  
>>>>> -/* Returns a new packet on success, otherwise returns NULL.
>>>>> - *
>>>>> - * If NULL is returned, errp is set to a negative errno.
>>>>> - */
>>>>> -static struct sk_buff *
>>>>> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
>>>>> -			   size_t len,
>>>>> -			   u32 src_cid,
>>>>> -			   u32 src_port,
>>>>> -			   u32 dst_cid,
>>>>> -			   u32 dst_port)
>>>>> -{
>>>>> -	const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
>>>>> -	struct virtio_vsock_hdr *hdr;
>>>>> -	struct sk_buff *skb;
>>>>> -	void *payload;
>>>>> -	int err;
>>>>> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
>>>>> +				       size_t max_to_send)
>>>>> +{
>>>>> +	const struct vsock_transport *t;
>>>>> +	struct iov_iter *iov_iter;
>>>>>  
>>>>> -	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>>>> -	if (!skb)
>>>>> -		return NULL;
>>>>> +	if (!info->msg)
>>>>> +		return false;
>>>>>  
>>>>> -	hdr = virtio_vsock_hdr(skb);
>>>>> -	hdr->type	= cpu_to_le16(info->type);
>>>>> -	hdr->op		= cpu_to_le16(info->op);
>>>>> -	hdr->src_cid	= cpu_to_le64(src_cid);
>>>>> -	hdr->dst_cid	= cpu_to_le64(dst_cid);
>>>>> -	hdr->src_port	= cpu_to_le32(src_port);
>>>>> -	hdr->dst_port	= cpu_to_le32(dst_port);
>>>>> -	hdr->flags	= cpu_to_le32(info->flags);
>>>>> -	hdr->len	= cpu_to_le32(len);
>>>>> +	iov_iter = &info->msg->msg_iter;
>>>>>  
>>>>> -	if (info->msg && len > 0) {
>>>>> -		payload = skb_put(skb, len);
>>>>> -		err = memcpy_from_msg(payload, info->msg, len);
>>>>> -		if (err)
>>>>> -			goto out;
>>>>> +	t = vsock_core_get_transport(info->vsk);
>>>>>  
>>>>> -		if (msg_data_left(info->msg) == 0 &&
>>>>> -		    info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>>>> -			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>>>> +	if (t->msgzerocopy_check_iov &&
>>>>> +	    !t->msgzerocopy_check_iov(iov_iter))
>>>>> +		return false;
>>>>>  
>>>>> -			if (info->msg->msg_flags & MSG_EOR)
>>>>> -				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>>>> -		}
>>>>> +	/* Data is simple buffer. */
>>>>> +	if (iter_is_ubuf(iov_iter))
>>>>> +		return true;
>>>>> +
>>>>> +	if (!iter_is_iovec(iov_iter))
>>>>> +		return false;
>>>>> +
>>>>> +	if (iov_iter->iov_offset)
>>>>> +		return false;
>>>>> +
>>>>> +	/* We can't send whole iov. */
>>>>> +	if (iov_iter->count > max_to_send)
>>>>> +		return false;
>>>>> +
>>>>> +	return true;
>>>>> +}
>>>>> +
>>>>> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
>>>>> +					   struct sk_buff *skb,
>>>>> +					   struct msghdr *msg,
>>>>> +					   bool zerocopy)
>>>>> +{
>>>>> +	struct ubuf_info *uarg;
>>>>> +
>>>>> +	if (msg->msg_ubuf) {
>>>>> +		uarg = msg->msg_ubuf;
>>>>> +		net_zcopy_get(uarg);
>>>>> +	} else {
>>>>> +		struct iov_iter *iter = &msg->msg_iter;
>>>>> +		struct ubuf_info_msgzc *uarg_zc;
>>>>> +		int len;
>>>>> +
>>>>> +		/* Only ITER_IOVEC or ITER_UBUF are allowed and
>>>>> +		 * checked before.
>>>>> +		 */
>>>>> +		if (iter_is_iovec(iter))
>>>>> +			len = iov_length(iter->__iov, iter->nr_segs);
>>>>> +		else
>>>>> +			len = iter->count;
>>>>> +
>>>>> +		uarg = msg_zerocopy_realloc(sk_vsock(vsk),
>>>>> +					    len,
>>>>> +					    NULL);
>>>>> +		if (!uarg)
>>>>> +			return -1;
>>>>> +
>>>>> +		uarg_zc = uarg_to_msgzc(uarg);
>>>>> +		uarg_zc->zerocopy = zerocopy ? 1 : 0;
>>>>>  	}
>>>>>  
>>>>> -	if (info->reply)
>>>>> -		virtio_vsock_skb_set_reply(skb);
>>>>> +	skb_zcopy_init(skb, uarg);
>>>>>  
>>>>> -	trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>>>> -					 dst_cid, dst_port,
>>>>> -					 len,
>>>>> -					 info->type,
>>>>> -					 info->op,
>>>>> -					 info->flags);
>>>>> +	return 0;
>>>>> +}
>>>>>  
>>>>> -	if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
>>>>> -		WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
>>>>> -		goto out;
>>>>> +static int virtio_transport_fill_skb(struct sk_buff *skb,
>>>>> +				     struct virtio_vsock_pkt_info *info,
>>>>> +				     size_t len,
>>>>> +				     bool zcopy)
>>>>> +{
>>>>> +	if (zcopy) {
>>>>> +		return __zerocopy_sg_from_iter(info->msg, NULL, skb,
>>>>> +					      &info->msg->msg_iter,
>>>>> +					      len);
>>>>> +	} else {
>>>>> +		void *payload;
>>>>> +		int err;
>>>>> +
>>>>> +		payload = skb_put(skb, len);
>>>>> +		err = memcpy_from_msg(payload, info->msg, len);
>>>>> +		if (err)
>>>>> +			return -1;
>>>>> +
>>>>> +		if (msg_data_left(info->msg))
>>>>> +			return 0;
>>>>> +
>>>>> +		return 0;
>>>>>  	}
>>>>> +}
>>>>>  
>>>>> -	return skb;
>>>>> +static void virtio_transport_init_hdr(struct sk_buff *skb,
>>>>> +				      struct virtio_vsock_pkt_info *info,
>>>>> +				      u32 src_cid,
>>>>> +				      u32 src_port,
>>>>> +				      u32 dst_cid,
>>>>> +				      u32 dst_port,
>>>>> +				      size_t len)
>>>>> +{
>>>>> +	struct virtio_vsock_hdr *hdr;
>>>>>  
>>>>> -out:
>>>>> -	kfree_skb(skb);
>>>>> -	return NULL;
>>>>> +	hdr = virtio_vsock_hdr(skb);
>>>>> +	hdr->type	= cpu_to_le16(info->type);
>>>>> +	hdr->op		= cpu_to_le16(info->op);
>>>>> +	hdr->src_cid	= cpu_to_le64(src_cid);
>>>>> +	hdr->dst_cid	= cpu_to_le64(dst_cid);
>>>>> +	hdr->src_port	= cpu_to_le32(src_port);
>>>>> +	hdr->dst_port	= cpu_to_le32(dst_port);
>>>>> +	hdr->flags	= cpu_to_le32(info->flags);
>>>>> +	hdr->len	= cpu_to_le32(len);
>>>>>  }
>>>>>  
>>>>>  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
>>>>> @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>>>  		return VIRTIO_VSOCK_TYPE_SEQPACKET;
>>>>>  }
>>>>>  
>>>>> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
>>>>> +						  struct virtio_vsock_pkt_info *info,
>>>>> +						  size_t payload_len,
>>>>> +						  bool zcopy,
>>>>> +						  u32 src_cid,
>>>>> +						  u32 src_port,
>>>>> +						  u32 dst_cid,
>>>>> +						  u32 dst_port)
>>>>> +{
>>>>> +	struct sk_buff *skb;
>>>>> +	size_t skb_len;
>>>>> +
>>>>> +	skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
>>>>> +
>>>>> +	if (!zcopy)
>>>>> +		skb_len += payload_len;
>>>>> +
>>>>> +	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>>>> +	if (!skb)
>>>>> +		return NULL;
>>>>> +
>>>>> +	virtio_transport_init_hdr(skb, info, src_cid, src_port,
>>>>> +				  dst_cid, dst_port,
>>>>> +				  payload_len);
>>>>> +
>>>>> +	/* Set owner here, because '__zerocopy_sg_from_iter()' uses
>>>>> +	 * owner of skb without check to update 'sk_wmem_alloc'.
>>>>> +	 */
>>>>> +	if (vsk)
>>>>> +		skb_set_owner_w(skb, sk_vsock(vsk));
>>>>> +
>>>>> +	if (info->msg && payload_len > 0) {
>>>>> +		int err;
>>>>> +
>>>>> +		err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
>>>>> +		if (err)
>>>>> +			goto out;
>>>>> +
>>>>> +		if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>>>> +			struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
>>>>> +
>>>>> +			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>>>> +
>>>>> +			if (info->msg->msg_flags & MSG_EOR)
>>>>> +				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>>>> +		}
>>>>> +	}
>>>>> +
>>>>> +	if (info->reply)
>>>>> +		virtio_vsock_skb_set_reply(skb);
>>>>> +
>>>>> +	trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>>>> +					 dst_cid, dst_port,
>>>>> +					 payload_len,
>>>>> +					 info->type,
>>>>> +					 info->op,
>>>>> +					 info->flags);
>>>>> +
>>>>> +	return skb;
>>>>> +out:
>>>>> +	kfree_skb(skb);
>>>>> +	return NULL;
>>>>> +}
>>>>> +
>>>>>  /* This function can only be used on connecting/connected sockets,
>>>>>   * since a socket assigned to a transport is required.
>>>>>   *
>>>>> @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>>>  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>>  					  struct virtio_vsock_pkt_info *info)
>>>>>  {
>>>>> +	u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
>>>>>  	u32 src_cid, src_port, dst_cid, dst_port;
>>>>>  	const struct virtio_transport *t_ops;
>>>>>  	struct virtio_vsock_sock *vvs;
>>>>>  	u32 pkt_len = info->pkt_len;
>>>>> +	bool can_zcopy = false;
>>>>>  	u32 rest_len;
>>>>>  	int ret;
>>>>>  
>>>>> @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>>  	if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
>>>>>  		return pkt_len;
>>>>>  
>>>>> +	if (info->msg) {
>>>>> +		/* If zerocopy is not enabled by 'setsockopt()', we behave as
>>>>> +		 * there is no MSG_ZEROCOPY flag set.
>>>>> +		 */
>>>>> +		if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
>>>>> +			info->msg->msg_flags &= ~MSG_ZEROCOPY;
>>>>> +
>>>>> +		if (info->msg->msg_flags & MSG_ZEROCOPY)
>>>>> +			can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
>>>>> +
>>>>> +		if (can_zcopy)
>>>>> +			max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
>>>>> +					    (MAX_SKB_FRAGS * PAGE_SIZE));
>>>>> +	}
>>>>> +
>>>>>  	rest_len = pkt_len;
>>>>>  
>>>>>  	do {
>>>>>  		struct sk_buff *skb;
>>>>>  		size_t skb_len;
>>>>>  
>>>>> -		skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
>>>>> +		skb_len = min(max_skb_len, rest_len);
>>>>>  
>>>>> -		skb = virtio_transport_alloc_skb(info, skb_len,
>>>>> +		skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
>>>>>  						 src_cid, src_port,
>>>>>  						 dst_cid, dst_port);
>>>>>  		if (!skb) {
>>>>> @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>>  			break;
>>>>>  		}
>>>>>  
>>>>> +		/* This is last skb to send this portion of data. */
>>>>> +		if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
>>>>> +		    skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
>>>>> +			if (virtio_transport_init_zcopy_skb(vsk, skb,
>>>>> +							    info->msg,
>>>>> +							    can_zcopy)) {
>>>>> +				ret = -ENOMEM;
>>>>> +				break;
>>>>> +			}
>>>>> +		}
>>>>> +
>>>>>  		virtio_transport_inc_tx_pkt(vvs, skb);
>>>>>  
>>>>>  		ret = t_ops->send_pkt(skb);
>>>>> @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
>>>>>  	if (!t)
>>>>>  		return -ENOTCONN;
>>>>>  
>>>>> -	reply = virtio_transport_alloc_skb(&info, 0,
>>>>> +	reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
>>>>>  					   le64_to_cpu(hdr->dst_cid),
>>>>>  					   le32_to_cpu(hdr->dst_port),
>>>>>  					   le64_to_cpu(hdr->src_cid),
>>>
>
  
Arseniy Krasnov July 25, 2023, 1:30 p.m. UTC | #20
On 25.07.2023 16:23, Michael S. Tsirkin wrote:
> On Tue, Jul 25, 2023 at 04:10:40PM +0300, Arseniy Krasnov wrote:
>>
>>
>> On 25.07.2023 14:59, Michael S. Tsirkin wrote:
>>> On Tue, Jul 25, 2023 at 11:39:22AM +0300, Arseniy Krasnov wrote:
>>>>
>>>>
>>>> On 25.07.2023 11:25, Michael S. Tsirkin wrote:
>>>>> On Fri, Jul 21, 2023 at 12:42:45AM +0300, Arseniy Krasnov wrote:
>>>>>> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
>>>>>> flag is set and zerocopy transmission is possible (enabled in socket
>>>>>> options and transport allows zerocopy), then non-linear skb will be
>>>>>> created and filled with the pages of user's buffer. Pages of user's
>>>>>> buffer are locked in memory by 'get_user_pages()'. Second thing that
>>>>>> this patch does is replace type of skb owning: instead of calling
>>>>>> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
>>>>>> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
>>>>>> of socket, so to decrease this field correctly proper skb destructor is
>>>>>> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
>>>>>>
>>>>>> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
>>>>>> ---
>>>>>>  Changelog:
>>>>>>  v5(big patchset) -> v1:
>>>>>>   * Refactorings of 'if' conditions.
>>>>>>   * Remove extra blank line.
>>>>>>   * Remove 'frag_off' field unneeded init.
>>>>>>   * Add function 'virtio_transport_fill_skb()' which fills both linear
>>>>>>     and non-linear skb with provided data.
>>>>>>  v1 -> v2:
>>>>>>   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
>>>>>>  v2 -> v3:
>>>>>>   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
>>>>>>     provided 'iov_iter' with data could be sent in a zerocopy mode.
>>>>>>     If this callback is not set in transport - transport allows to send
>>>>>>     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
>>>>>>     then zerocopy is allowed. Reason of this callback is that in case of
>>>>>>     G2H transmission we insert whole skb to the tx virtio queue and such
>>>>>>     skb must fit to the size of the virtio queue to be sent in a single
>>>>>>     iteration (may be tx logic in 'virtio_transport.c' could be reworked
>>>>>>     as in vhost to support partial send of current skb). This callback
>>>>>>     will be enabled only for G2H path. For details pls see comment 
>>>>>>     'Check that tx queue...' below.
>>>>>>
>>>>>>  include/net/af_vsock.h                  |   3 +
>>>>>>  net/vmw_vsock/virtio_transport.c        |  39 ++++
>>>>>>  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
>>>>>>  3 files changed, 241 insertions(+), 58 deletions(-)
>>>>>>
>>>>>> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
>>>>>> index 0e7504a42925..a6b346eeeb8e 100644
>>>>>> --- a/include/net/af_vsock.h
>>>>>> +++ b/include/net/af_vsock.h
>>>>>> @@ -177,6 +177,9 @@ struct vsock_transport {
>>>>>>  
>>>>>>  	/* Read a single skb */
>>>>>>  	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
>>>>>> +
>>>>>> +	/* Zero-copy. */
>>>>>> +	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
>>>>>>  };
>>>>>>  
>>>>>>  /**** CORE ****/
>>>>>> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>>>>>> index 7bbcc8093e51..23cb8ed638c4 100644
>>>>>> --- a/net/vmw_vsock/virtio_transport.c
>>>>>> +++ b/net/vmw_vsock/virtio_transport.c
>>>>>> @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
>>>>>>  	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>>>>>>  }
>>>>>>  
>>>>>> +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
>>>>>> +{
>>>>>> +	struct virtio_vsock *vsock;
>>>>>> +	bool res = false;
>>>>>> +
>>>>>> +	rcu_read_lock();
>>>>>> +
>>>>>> +	vsock = rcu_dereference(the_virtio_vsock);
>>>>>> +	if (vsock) {
>>>>>> +		struct virtqueue *vq;
>>>>>> +		int iov_pages;
>>>>>> +
>>>>>> +		vq = vsock->vqs[VSOCK_VQ_TX];
>>>>>> +
>>>>>> +		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
>>>>>> +
>>>>>> +		/* Check that tx queue is large enough to keep whole
>>>>>> +		 * data to send. This is needed, because when there is
>>>>>> +		 * not enough free space in the queue, current skb to
>>>>>> +		 * send will be reinserted to the head of tx list of
>>>>>> +		 * the socket to retry transmission later, so if skb
>>>>>> +		 * is bigger than whole queue, it will be reinserted
>>>>>> +		 * again and again, thus blocking other skbs to be sent.
>>>>>> +		 * Each page of the user provided buffer will be added
>>>>>> +		 * as a single buffer to the tx virtqueue, so compare
>>>>>> +		 * number of pages against maximum capacity of the queue.
>>>>>> +		 * +1 means buffer for the packet header.
>>>>>> +		 */
>>>>>> +		if (iov_pages + 1 <= vq->num_max)
>>>>>> +			res = true;
>>>>>
>>>>>
>>>>> Yes but can't there already be buffers in the queue?
>>>>> Then you can't stick num_max there.
>>>>
>>>> I think, that it is not critical, because vhost part always tries to process all
>>>> incoming buffers (yes, 'vhost_exceeds_weight()' breaks at some moment, but it will
>>>> reschedule tx kick ('vhost_vsock_handle_tx_kick()') work again), so current "too
>>>> big" skb will wait until there will be enough space in queue and as it is requeued
>>>> to the head of tx list it will be inserted to tx queue first.
>>>>
>>>> But anyway, I agree that comparing to 'num_free' may be more effective to the whole
>>>> system performance...
>>>>
>>>> Thanks, Arseniy
>>>
>>> Oh I see. It makes sense then - instead of copying just so we can
>>> stick it in the queue, wait a bit and send later.
>>> Also - for stream transports can't the message be split
>>> and sent chunk by chunk? Better than copying ...
>>
>> Technically yes, also we can split message for non-stream sockets (as vhost
>> does when it copies data to rx buffers of the guest),
> 
> Won't breaking up messages break applications though?

No, for seqpacket we have special marker + port in each message, so we
can restore original message at the receiver and pass it to the socket.

Thanks, Arseniy

> 
> 
>> but it requires to rework
>> current implementation by adding buffers one by one to the tx queue. I think
>> it was not implemented here because until MSG_ZEROCOPY all skbs requires one
>> (if it is control message) or two (with payload) buffers, so there is no big
>> sense in processing max two buffers in "one-by-one" mode - we can just wait
>> for space.
>>
>> May be, I can add this logic for non-linear skb's here: 
>>
>> if (skb->len > vq->max_num)
>>     add buffers "one-by-one", incrementing internal offset in skb,
>>    if (new skb insertion fails)
>>        requeue skb, wait for space.
>>
>> In TX done callback I'll call consume skb only when above mentioned internal
>> offset == skb->len. I think this approach allows to get rid of new 'check_iov'
>> callback from this patch.
>>
>>
>> Stefano, what do You think?
>>
>> Thanks, Arseniy
>>
>>>
>>>
>>>>>
>>>>>
>>>>>> +	}
>>>>>> +
>>>>>> +	rcu_read_unlock();
>>>>>> +
>>>>>> +	return res;
>>>>>> +}
>>>>>> +
>>>>>>  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
>>>>>>  
>>>>>>  static struct virtio_transport virtio_transport = {
>>>>>> @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
>>>>>>  		.seqpacket_allow          = virtio_transport_seqpacket_allow,
>>>>>>  		.seqpacket_has_data       = virtio_transport_seqpacket_has_data,
>>>>>>  
>>>>>> +		.msgzerocopy_check_iov	  = virtio_transport_msgzerocopy_check_iov,
>>>>>> +
>>>>>>  		.notify_poll_in           = virtio_transport_notify_poll_in,
>>>>>>  		.notify_poll_out          = virtio_transport_notify_poll_out,
>>>>>>  		.notify_recv_init         = virtio_transport_notify_recv_init,
>>>>>> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
>>>>>> index 26a4d10da205..e4e3d541aff4 100644
>>>>>> --- a/net/vmw_vsock/virtio_transport_common.c
>>>>>> +++ b/net/vmw_vsock/virtio_transport_common.c
>>>>>> @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
>>>>>>  	return container_of(t, struct virtio_transport, transport);
>>>>>>  }
>>>>>>  
>>>>>> -/* Returns a new packet on success, otherwise returns NULL.
>>>>>> - *
>>>>>> - * If NULL is returned, errp is set to a negative errno.
>>>>>> - */
>>>>>> -static struct sk_buff *
>>>>>> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
>>>>>> -			   size_t len,
>>>>>> -			   u32 src_cid,
>>>>>> -			   u32 src_port,
>>>>>> -			   u32 dst_cid,
>>>>>> -			   u32 dst_port)
>>>>>> -{
>>>>>> -	const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
>>>>>> -	struct virtio_vsock_hdr *hdr;
>>>>>> -	struct sk_buff *skb;
>>>>>> -	void *payload;
>>>>>> -	int err;
>>>>>> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
>>>>>> +				       size_t max_to_send)
>>>>>> +{
>>>>>> +	const struct vsock_transport *t;
>>>>>> +	struct iov_iter *iov_iter;
>>>>>>  
>>>>>> -	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>>>>> -	if (!skb)
>>>>>> -		return NULL;
>>>>>> +	if (!info->msg)
>>>>>> +		return false;
>>>>>>  
>>>>>> -	hdr = virtio_vsock_hdr(skb);
>>>>>> -	hdr->type	= cpu_to_le16(info->type);
>>>>>> -	hdr->op		= cpu_to_le16(info->op);
>>>>>> -	hdr->src_cid	= cpu_to_le64(src_cid);
>>>>>> -	hdr->dst_cid	= cpu_to_le64(dst_cid);
>>>>>> -	hdr->src_port	= cpu_to_le32(src_port);
>>>>>> -	hdr->dst_port	= cpu_to_le32(dst_port);
>>>>>> -	hdr->flags	= cpu_to_le32(info->flags);
>>>>>> -	hdr->len	= cpu_to_le32(len);
>>>>>> +	iov_iter = &info->msg->msg_iter;
>>>>>>  
>>>>>> -	if (info->msg && len > 0) {
>>>>>> -		payload = skb_put(skb, len);
>>>>>> -		err = memcpy_from_msg(payload, info->msg, len);
>>>>>> -		if (err)
>>>>>> -			goto out;
>>>>>> +	t = vsock_core_get_transport(info->vsk);
>>>>>>  
>>>>>> -		if (msg_data_left(info->msg) == 0 &&
>>>>>> -		    info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>>>>> -			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>>>>> +	if (t->msgzerocopy_check_iov &&
>>>>>> +	    !t->msgzerocopy_check_iov(iov_iter))
>>>>>> +		return false;
>>>>>>  
>>>>>> -			if (info->msg->msg_flags & MSG_EOR)
>>>>>> -				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>>>>> -		}
>>>>>> +	/* Data is simple buffer. */
>>>>>> +	if (iter_is_ubuf(iov_iter))
>>>>>> +		return true;
>>>>>> +
>>>>>> +	if (!iter_is_iovec(iov_iter))
>>>>>> +		return false;
>>>>>> +
>>>>>> +	if (iov_iter->iov_offset)
>>>>>> +		return false;
>>>>>> +
>>>>>> +	/* We can't send whole iov. */
>>>>>> +	if (iov_iter->count > max_to_send)
>>>>>> +		return false;
>>>>>> +
>>>>>> +	return true;
>>>>>> +}
>>>>>> +
>>>>>> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
>>>>>> +					   struct sk_buff *skb,
>>>>>> +					   struct msghdr *msg,
>>>>>> +					   bool zerocopy)
>>>>>> +{
>>>>>> +	struct ubuf_info *uarg;
>>>>>> +
>>>>>> +	if (msg->msg_ubuf) {
>>>>>> +		uarg = msg->msg_ubuf;
>>>>>> +		net_zcopy_get(uarg);
>>>>>> +	} else {
>>>>>> +		struct iov_iter *iter = &msg->msg_iter;
>>>>>> +		struct ubuf_info_msgzc *uarg_zc;
>>>>>> +		int len;
>>>>>> +
>>>>>> +		/* Only ITER_IOVEC or ITER_UBUF are allowed and
>>>>>> +		 * checked before.
>>>>>> +		 */
>>>>>> +		if (iter_is_iovec(iter))
>>>>>> +			len = iov_length(iter->__iov, iter->nr_segs);
>>>>>> +		else
>>>>>> +			len = iter->count;
>>>>>> +
>>>>>> +		uarg = msg_zerocopy_realloc(sk_vsock(vsk),
>>>>>> +					    len,
>>>>>> +					    NULL);
>>>>>> +		if (!uarg)
>>>>>> +			return -1;
>>>>>> +
>>>>>> +		uarg_zc = uarg_to_msgzc(uarg);
>>>>>> +		uarg_zc->zerocopy = zerocopy ? 1 : 0;
>>>>>>  	}
>>>>>>  
>>>>>> -	if (info->reply)
>>>>>> -		virtio_vsock_skb_set_reply(skb);
>>>>>> +	skb_zcopy_init(skb, uarg);
>>>>>>  
>>>>>> -	trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>>>>> -					 dst_cid, dst_port,
>>>>>> -					 len,
>>>>>> -					 info->type,
>>>>>> -					 info->op,
>>>>>> -					 info->flags);
>>>>>> +	return 0;
>>>>>> +}
>>>>>>  
>>>>>> -	if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
>>>>>> -		WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
>>>>>> -		goto out;
>>>>>> +static int virtio_transport_fill_skb(struct sk_buff *skb,
>>>>>> +				     struct virtio_vsock_pkt_info *info,
>>>>>> +				     size_t len,
>>>>>> +				     bool zcopy)
>>>>>> +{
>>>>>> +	if (zcopy) {
>>>>>> +		return __zerocopy_sg_from_iter(info->msg, NULL, skb,
>>>>>> +					      &info->msg->msg_iter,
>>>>>> +					      len);
>>>>>> +	} else {
>>>>>> +		void *payload;
>>>>>> +		int err;
>>>>>> +
>>>>>> +		payload = skb_put(skb, len);
>>>>>> +		err = memcpy_from_msg(payload, info->msg, len);
>>>>>> +		if (err)
>>>>>> +			return -1;
>>>>>> +
>>>>>> +		if (msg_data_left(info->msg))
>>>>>> +			return 0;
>>>>>> +
>>>>>> +		return 0;
>>>>>>  	}
>>>>>> +}
>>>>>>  
>>>>>> -	return skb;
>>>>>> +static void virtio_transport_init_hdr(struct sk_buff *skb,
>>>>>> +				      struct virtio_vsock_pkt_info *info,
>>>>>> +				      u32 src_cid,
>>>>>> +				      u32 src_port,
>>>>>> +				      u32 dst_cid,
>>>>>> +				      u32 dst_port,
>>>>>> +				      size_t len)
>>>>>> +{
>>>>>> +	struct virtio_vsock_hdr *hdr;
>>>>>>  
>>>>>> -out:
>>>>>> -	kfree_skb(skb);
>>>>>> -	return NULL;
>>>>>> +	hdr = virtio_vsock_hdr(skb);
>>>>>> +	hdr->type	= cpu_to_le16(info->type);
>>>>>> +	hdr->op		= cpu_to_le16(info->op);
>>>>>> +	hdr->src_cid	= cpu_to_le64(src_cid);
>>>>>> +	hdr->dst_cid	= cpu_to_le64(dst_cid);
>>>>>> +	hdr->src_port	= cpu_to_le32(src_port);
>>>>>> +	hdr->dst_port	= cpu_to_le32(dst_port);
>>>>>> +	hdr->flags	= cpu_to_le32(info->flags);
>>>>>> +	hdr->len	= cpu_to_le32(len);
>>>>>>  }
>>>>>>  
>>>>>>  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
>>>>>> @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>>>>  		return VIRTIO_VSOCK_TYPE_SEQPACKET;
>>>>>>  }
>>>>>>  
>>>>>> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
>>>>>> +						  struct virtio_vsock_pkt_info *info,
>>>>>> +						  size_t payload_len,
>>>>>> +						  bool zcopy,
>>>>>> +						  u32 src_cid,
>>>>>> +						  u32 src_port,
>>>>>> +						  u32 dst_cid,
>>>>>> +						  u32 dst_port)
>>>>>> +{
>>>>>> +	struct sk_buff *skb;
>>>>>> +	size_t skb_len;
>>>>>> +
>>>>>> +	skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
>>>>>> +
>>>>>> +	if (!zcopy)
>>>>>> +		skb_len += payload_len;
>>>>>> +
>>>>>> +	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>>>>> +	if (!skb)
>>>>>> +		return NULL;
>>>>>> +
>>>>>> +	virtio_transport_init_hdr(skb, info, src_cid, src_port,
>>>>>> +				  dst_cid, dst_port,
>>>>>> +				  payload_len);
>>>>>> +
>>>>>> +	/* Set owner here, because '__zerocopy_sg_from_iter()' uses
>>>>>> +	 * owner of skb without check to update 'sk_wmem_alloc'.
>>>>>> +	 */
>>>>>> +	if (vsk)
>>>>>> +		skb_set_owner_w(skb, sk_vsock(vsk));
>>>>>> +
>>>>>> +	if (info->msg && payload_len > 0) {
>>>>>> +		int err;
>>>>>> +
>>>>>> +		err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
>>>>>> +		if (err)
>>>>>> +			goto out;
>>>>>> +
>>>>>> +		if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>>>>> +			struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
>>>>>> +
>>>>>> +			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>>>>> +
>>>>>> +			if (info->msg->msg_flags & MSG_EOR)
>>>>>> +				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>>>>> +		}
>>>>>> +	}
>>>>>> +
>>>>>> +	if (info->reply)
>>>>>> +		virtio_vsock_skb_set_reply(skb);
>>>>>> +
>>>>>> +	trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>>>>> +					 dst_cid, dst_port,
>>>>>> +					 payload_len,
>>>>>> +					 info->type,
>>>>>> +					 info->op,
>>>>>> +					 info->flags);
>>>>>> +
>>>>>> +	return skb;
>>>>>> +out:
>>>>>> +	kfree_skb(skb);
>>>>>> +	return NULL;
>>>>>> +}
>>>>>> +
>>>>>>  /* This function can only be used on connecting/connected sockets,
>>>>>>   * since a socket assigned to a transport is required.
>>>>>>   *
>>>>>> @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>>>>  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>>>  					  struct virtio_vsock_pkt_info *info)
>>>>>>  {
>>>>>> +	u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
>>>>>>  	u32 src_cid, src_port, dst_cid, dst_port;
>>>>>>  	const struct virtio_transport *t_ops;
>>>>>>  	struct virtio_vsock_sock *vvs;
>>>>>>  	u32 pkt_len = info->pkt_len;
>>>>>> +	bool can_zcopy = false;
>>>>>>  	u32 rest_len;
>>>>>>  	int ret;
>>>>>>  
>>>>>> @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>>>  	if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
>>>>>>  		return pkt_len;
>>>>>>  
>>>>>> +	if (info->msg) {
>>>>>> +		/* If zerocopy is not enabled by 'setsockopt()', we behave as
>>>>>> +		 * there is no MSG_ZEROCOPY flag set.
>>>>>> +		 */
>>>>>> +		if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
>>>>>> +			info->msg->msg_flags &= ~MSG_ZEROCOPY;
>>>>>> +
>>>>>> +		if (info->msg->msg_flags & MSG_ZEROCOPY)
>>>>>> +			can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
>>>>>> +
>>>>>> +		if (can_zcopy)
>>>>>> +			max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
>>>>>> +					    (MAX_SKB_FRAGS * PAGE_SIZE));
>>>>>> +	}
>>>>>> +
>>>>>>  	rest_len = pkt_len;
>>>>>>  
>>>>>>  	do {
>>>>>>  		struct sk_buff *skb;
>>>>>>  		size_t skb_len;
>>>>>>  
>>>>>> -		skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
>>>>>> +		skb_len = min(max_skb_len, rest_len);
>>>>>>  
>>>>>> -		skb = virtio_transport_alloc_skb(info, skb_len,
>>>>>> +		skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
>>>>>>  						 src_cid, src_port,
>>>>>>  						 dst_cid, dst_port);
>>>>>>  		if (!skb) {
>>>>>> @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>>>  			break;
>>>>>>  		}
>>>>>>  
>>>>>> +		/* This is last skb to send this portion of data. */
>>>>>> +		if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
>>>>>> +		    skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
>>>>>> +			if (virtio_transport_init_zcopy_skb(vsk, skb,
>>>>>> +							    info->msg,
>>>>>> +							    can_zcopy)) {
>>>>>> +				ret = -ENOMEM;
>>>>>> +				break;
>>>>>> +			}
>>>>>> +		}
>>>>>> +
>>>>>>  		virtio_transport_inc_tx_pkt(vvs, skb);
>>>>>>  
>>>>>>  		ret = t_ops->send_pkt(skb);
>>>>>> @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
>>>>>>  	if (!t)
>>>>>>  		return -ENOTCONN;
>>>>>>  
>>>>>> -	reply = virtio_transport_alloc_skb(&info, 0,
>>>>>> +	reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
>>>>>>  					   le64_to_cpu(hdr->dst_cid),
>>>>>>  					   le32_to_cpu(hdr->dst_port),
>>>>>>  					   le64_to_cpu(hdr->src_cid),
>>>>>> -- 
>>>>>> 2.25.1
>>>>>
>>>
>
  
Arseniy Krasnov July 25, 2023, 1:35 p.m. UTC | #21
On 25.07.2023 16:36, Michael S. Tsirkin wrote:
> On Tue, Jul 25, 2023 at 04:28:14PM +0300, Arseniy Krasnov wrote:
>>
>>
>> On 25.07.2023 16:22, Michael S. Tsirkin wrote:
>>> On Tue, Jul 25, 2023 at 04:04:13PM +0300, Arseniy Krasnov wrote:
>>>>
>>>>
>>>> On 25.07.2023 14:50, Michael S. Tsirkin wrote:
>>>>> On Fri, Jul 21, 2023 at 08:09:03AM +0300, Arseniy Krasnov wrote:
>>>>>>
>>>>>>
>>>>>> On 21.07.2023 00:42, Arseniy Krasnov wrote:
>>>>>>> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
>>>>>>> flag is set and zerocopy transmission is possible (enabled in socket
>>>>>>> options and transport allows zerocopy), then non-linear skb will be
>>>>>>> created and filled with the pages of user's buffer. Pages of user's
>>>>>>> buffer are locked in memory by 'get_user_pages()'. Second thing that
>>>>>>> this patch does is replace type of skb owning: instead of calling
>>>>>>> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
>>>>>>> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
>>>>>>> of socket, so to decrease this field correctly proper skb destructor is
>>>>>>> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
>>>>>>>
>>>>>>> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
>>>>>>> ---
>>>>>>>  Changelog:
>>>>>>>  v5(big patchset) -> v1:
>>>>>>>   * Refactorings of 'if' conditions.
>>>>>>>   * Remove extra blank line.
>>>>>>>   * Remove 'frag_off' field unneeded init.
>>>>>>>   * Add function 'virtio_transport_fill_skb()' which fills both linear
>>>>>>>     and non-linear skb with provided data.
>>>>>>>  v1 -> v2:
>>>>>>>   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
>>>>>>>  v2 -> v3:
>>>>>>>   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
>>>>>>>     provided 'iov_iter' with data could be sent in a zerocopy mode.
>>>>>>>     If this callback is not set in transport - transport allows to send
>>>>>>>     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
>>>>>>>     then zerocopy is allowed. Reason of this callback is that in case of
>>>>>>>     G2H transmission we insert whole skb to the tx virtio queue and such
>>>>>>>     skb must fit to the size of the virtio queue to be sent in a single
>>>>>>>     iteration (may be tx logic in 'virtio_transport.c' could be reworked
>>>>>>>     as in vhost to support partial send of current skb). This callback
>>>>>>>     will be enabled only for G2H path. For details pls see comment 
>>>>>>>     'Check that tx queue...' below.
>>>>>>>
>>>>>>>  include/net/af_vsock.h                  |   3 +
>>>>>>>  net/vmw_vsock/virtio_transport.c        |  39 ++++
>>>>>>>  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
>>>>>>>  3 files changed, 241 insertions(+), 58 deletions(-)
>>>>>>>
>>>>>>> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
>>>>>>> index 0e7504a42925..a6b346eeeb8e 100644
>>>>>>> --- a/include/net/af_vsock.h
>>>>>>> +++ b/include/net/af_vsock.h
>>>>>>> @@ -177,6 +177,9 @@ struct vsock_transport {
>>>>>>>  
>>>>>>>  	/* Read a single skb */
>>>>>>>  	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
>>>>>>> +
>>>>>>> +	/* Zero-copy. */
>>>>>>> +	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
>>>>>>>  };
>>>>>>>  
>>>>>>>  /**** CORE ****/
>>>>>>> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>>>>>>> index 7bbcc8093e51..23cb8ed638c4 100644
>>>>>>> --- a/net/vmw_vsock/virtio_transport.c
>>>>>>> +++ b/net/vmw_vsock/virtio_transport.c
>>>>>>> @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
>>>>>>>  	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>>>>>>>  }
>>>>>>>  
>>>>>>> +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
>>>>>>> +{
>>>>>>> +	struct virtio_vsock *vsock;
>>>>>>> +	bool res = false;
>>>>>>> +
>>>>>>> +	rcu_read_lock();
>>>>>>> +
>>>>>>> +	vsock = rcu_dereference(the_virtio_vsock);
>>>>>>> +	if (vsock) {
>>>>>>> +		struct virtqueue *vq;
>>>>>>> +		int iov_pages;
>>>>>>> +
>>>>>>> +		vq = vsock->vqs[VSOCK_VQ_TX];
>>>>>>> +
>>>>>>> +		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
>>>>>>> +
>>>>>>> +		/* Check that tx queue is large enough to keep whole
>>>>>>> +		 * data to send. This is needed, because when there is
>>>>>>> +		 * not enough free space in the queue, current skb to
>>>>>>> +		 * send will be reinserted to the head of tx list of
>>>>>>> +		 * the socket to retry transmission later, so if skb
>>>>>>> +		 * is bigger than whole queue, it will be reinserted
>>>>>>> +		 * again and again, thus blocking other skbs to be sent.
>>>>>>> +		 * Each page of the user provided buffer will be added
>>>>>>> +		 * as a single buffer to the tx virtqueue, so compare
>>>>>>> +		 * number of pages against maximum capacity of the queue.
>>>>>>> +		 * +1 means buffer for the packet header.
>>>>>>> +		 */
>>>>>>> +		if (iov_pages + 1 <= vq->num_max)
>>>>>>
>>>>>> I think this check is actual only for case one we don't have indirect buffer feature.
>>>>>> With indirect mode whole data to send will be packed into one indirect buffer.
>>>>>>
>>>>>> Thanks, Arseniy
>>>>>
>>>>> Actually the reverse. With indirect you are limited to num_max.
>>>>> Without you are limited to whatever space is left in the
>>>>> queue (which you did not check here, so you should).
>>>>
>>>> I mean that with indirect, we only need one buffer, and we can just wait
>>>> for enough space - for this single buffer ( as we discussed a little bit before).
>>>> But if indirect buffer is not supported - we need that whole packet must fit
>>>> to the size of tx queue - otherwise it never be transmitted.
>>>>
>>>> Thanks, Arseniy
>>>
>>>
>>> yes but according to virtio spec it's illegal to add s/g that is bigger
>>> than queue size.
>>
>> Aah, so even in case of indirect buffers feature, buffer descriptors stored in memory
>> pointed by indirect buffer must be accounted against queue size ?
>>
>> Thanks, Arseniy
> 
> a single indirect buffer can't exceed vq size.

I see, so i guess right way is to compare length of data "in buffers"
again 'num_max', as it is implemented. And we should not refer to whether
vq support indirect buffering or not.

Thanks, Arseniy

> 
> 
>>>
>>>>>
>>>>>
>>>>>>> +			res = true;
>>>>>>> +	}
>>>>>>> +
>>>>>>> +	rcu_read_unlock();
>>>>>
>>>>> Just curious:
>>>>> is the point of all this RCU dance to allow vsock
>>>>> to change from under us? then why is it ok to
>>>>> have it change? the virtio_transport_msgzerocopy_check_iov
>>>>> will then refer to the old vsock ...
>>>>>
>>>>>
>>>>>>> +
>>>>>>> +	return res;
>>>>>>> +}
>>>>>>> +
>>>>>>>  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
>>>>>>>  
>>>>>>>  static struct virtio_transport virtio_transport = {
>>>>>>> @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
>>>>>>>  		.seqpacket_allow          = virtio_transport_seqpacket_allow,
>>>>>>>  		.seqpacket_has_data       = virtio_transport_seqpacket_has_data,
>>>>>>>  
>>>>>>> +		.msgzerocopy_check_iov	  = virtio_transport_msgzerocopy_check_iov,
>>>>>>> +
>>>>>>>  		.notify_poll_in           = virtio_transport_notify_poll_in,
>>>>>>>  		.notify_poll_out          = virtio_transport_notify_poll_out,
>>>>>>>  		.notify_recv_init         = virtio_transport_notify_recv_init,
>>>>>>> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
>>>>>>> index 26a4d10da205..e4e3d541aff4 100644
>>>>>>> --- a/net/vmw_vsock/virtio_transport_common.c
>>>>>>> +++ b/net/vmw_vsock/virtio_transport_common.c
>>>>>>> @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
>>>>>>>  	return container_of(t, struct virtio_transport, transport);
>>>>>>>  }
>>>>>>>  
>>>>>>> -/* Returns a new packet on success, otherwise returns NULL.
>>>>>>> - *
>>>>>>> - * If NULL is returned, errp is set to a negative errno.
>>>>>>> - */
>>>>>>> -static struct sk_buff *
>>>>>>> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
>>>>>>> -			   size_t len,
>>>>>>> -			   u32 src_cid,
>>>>>>> -			   u32 src_port,
>>>>>>> -			   u32 dst_cid,
>>>>>>> -			   u32 dst_port)
>>>>>>> -{
>>>>>>> -	const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
>>>>>>> -	struct virtio_vsock_hdr *hdr;
>>>>>>> -	struct sk_buff *skb;
>>>>>>> -	void *payload;
>>>>>>> -	int err;
>>>>>>> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
>>>>>>> +				       size_t max_to_send)
>>>>>>> +{
>>>>>>> +	const struct vsock_transport *t;
>>>>>>> +	struct iov_iter *iov_iter;
>>>>>>>  
>>>>>>> -	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>>>>>> -	if (!skb)
>>>>>>> -		return NULL;
>>>>>>> +	if (!info->msg)
>>>>>>> +		return false;
>>>>>>>  
>>>>>>> -	hdr = virtio_vsock_hdr(skb);
>>>>>>> -	hdr->type	= cpu_to_le16(info->type);
>>>>>>> -	hdr->op		= cpu_to_le16(info->op);
>>>>>>> -	hdr->src_cid	= cpu_to_le64(src_cid);
>>>>>>> -	hdr->dst_cid	= cpu_to_le64(dst_cid);
>>>>>>> -	hdr->src_port	= cpu_to_le32(src_port);
>>>>>>> -	hdr->dst_port	= cpu_to_le32(dst_port);
>>>>>>> -	hdr->flags	= cpu_to_le32(info->flags);
>>>>>>> -	hdr->len	= cpu_to_le32(len);
>>>>>>> +	iov_iter = &info->msg->msg_iter;
>>>>>>>  
>>>>>>> -	if (info->msg && len > 0) {
>>>>>>> -		payload = skb_put(skb, len);
>>>>>>> -		err = memcpy_from_msg(payload, info->msg, len);
>>>>>>> -		if (err)
>>>>>>> -			goto out;
>>>>>>> +	t = vsock_core_get_transport(info->vsk);
>>>>>>>  
>>>>>>> -		if (msg_data_left(info->msg) == 0 &&
>>>>>>> -		    info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>>>>>> -			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>>>>>> +	if (t->msgzerocopy_check_iov &&
>>>>>>> +	    !t->msgzerocopy_check_iov(iov_iter))
>>>>>>> +		return false;
>>>>>>>  
>>>>>>> -			if (info->msg->msg_flags & MSG_EOR)
>>>>>>> -				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>>>>>> -		}
>>>>>>> +	/* Data is simple buffer. */
>>>>>>> +	if (iter_is_ubuf(iov_iter))
>>>>>>> +		return true;
>>>>>>> +
>>>>>>> +	if (!iter_is_iovec(iov_iter))
>>>>>>> +		return false;
>>>>>>> +
>>>>>>> +	if (iov_iter->iov_offset)
>>>>>>> +		return false;
>>>>>>> +
>>>>>>> +	/* We can't send whole iov. */
>>>>>>> +	if (iov_iter->count > max_to_send)
>>>>>>> +		return false;
>>>>>>> +
>>>>>>> +	return true;
>>>>>>> +}
>>>>>>> +
>>>>>>> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
>>>>>>> +					   struct sk_buff *skb,
>>>>>>> +					   struct msghdr *msg,
>>>>>>> +					   bool zerocopy)
>>>>>>> +{
>>>>>>> +	struct ubuf_info *uarg;
>>>>>>> +
>>>>>>> +	if (msg->msg_ubuf) {
>>>>>>> +		uarg = msg->msg_ubuf;
>>>>>>> +		net_zcopy_get(uarg);
>>>>>>> +	} else {
>>>>>>> +		struct iov_iter *iter = &msg->msg_iter;
>>>>>>> +		struct ubuf_info_msgzc *uarg_zc;
>>>>>>> +		int len;
>>>>>>> +
>>>>>>> +		/* Only ITER_IOVEC or ITER_UBUF are allowed and
>>>>>>> +		 * checked before.
>>>>>>> +		 */
>>>>>>> +		if (iter_is_iovec(iter))
>>>>>>> +			len = iov_length(iter->__iov, iter->nr_segs);
>>>>>>> +		else
>>>>>>> +			len = iter->count;
>>>>>>> +
>>>>>>> +		uarg = msg_zerocopy_realloc(sk_vsock(vsk),
>>>>>>> +					    len,
>>>>>>> +					    NULL);
>>>>>>> +		if (!uarg)
>>>>>>> +			return -1;
>>>>>>> +
>>>>>>> +		uarg_zc = uarg_to_msgzc(uarg);
>>>>>>> +		uarg_zc->zerocopy = zerocopy ? 1 : 0;
>>>>>>>  	}
>>>>>>>  
>>>>>>> -	if (info->reply)
>>>>>>> -		virtio_vsock_skb_set_reply(skb);
>>>>>>> +	skb_zcopy_init(skb, uarg);
>>>>>>>  
>>>>>>> -	trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>>>>>> -					 dst_cid, dst_port,
>>>>>>> -					 len,
>>>>>>> -					 info->type,
>>>>>>> -					 info->op,
>>>>>>> -					 info->flags);
>>>>>>> +	return 0;
>>>>>>> +}
>>>>>>>  
>>>>>>> -	if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
>>>>>>> -		WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
>>>>>>> -		goto out;
>>>>>>> +static int virtio_transport_fill_skb(struct sk_buff *skb,
>>>>>>> +				     struct virtio_vsock_pkt_info *info,
>>>>>>> +				     size_t len,
>>>>>>> +				     bool zcopy)
>>>>>>> +{
>>>>>>> +	if (zcopy) {
>>>>>>> +		return __zerocopy_sg_from_iter(info->msg, NULL, skb,
>>>>>>> +					      &info->msg->msg_iter,
>>>>>>> +					      len);
>>>>>>> +	} else {
>>>>>>> +		void *payload;
>>>>>>> +		int err;
>>>>>>> +
>>>>>>> +		payload = skb_put(skb, len);
>>>>>>> +		err = memcpy_from_msg(payload, info->msg, len);
>>>>>>> +		if (err)
>>>>>>> +			return -1;
>>>>>>> +
>>>>>>> +		if (msg_data_left(info->msg))
>>>>>>> +			return 0;
>>>>>>> +
>>>>>>> +		return 0;
>>>>>>>  	}
>>>>>>> +}
>>>>>>>  
>>>>>>> -	return skb;
>>>>>>> +static void virtio_transport_init_hdr(struct sk_buff *skb,
>>>>>>> +				      struct virtio_vsock_pkt_info *info,
>>>>>>> +				      u32 src_cid,
>>>>>>> +				      u32 src_port,
>>>>>>> +				      u32 dst_cid,
>>>>>>> +				      u32 dst_port,
>>>>>>> +				      size_t len)
>>>>>>> +{
>>>>>>> +	struct virtio_vsock_hdr *hdr;
>>>>>>>  
>>>>>>> -out:
>>>>>>> -	kfree_skb(skb);
>>>>>>> -	return NULL;
>>>>>>> +	hdr = virtio_vsock_hdr(skb);
>>>>>>> +	hdr->type	= cpu_to_le16(info->type);
>>>>>>> +	hdr->op		= cpu_to_le16(info->op);
>>>>>>> +	hdr->src_cid	= cpu_to_le64(src_cid);
>>>>>>> +	hdr->dst_cid	= cpu_to_le64(dst_cid);
>>>>>>> +	hdr->src_port	= cpu_to_le32(src_port);
>>>>>>> +	hdr->dst_port	= cpu_to_le32(dst_port);
>>>>>>> +	hdr->flags	= cpu_to_le32(info->flags);
>>>>>>> +	hdr->len	= cpu_to_le32(len);
>>>>>>>  }
>>>>>>>  
>>>>>>>  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
>>>>>>> @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>>>>>  		return VIRTIO_VSOCK_TYPE_SEQPACKET;
>>>>>>>  }
>>>>>>>  
>>>>>>> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
>>>>>>> +						  struct virtio_vsock_pkt_info *info,
>>>>>>> +						  size_t payload_len,
>>>>>>> +						  bool zcopy,
>>>>>>> +						  u32 src_cid,
>>>>>>> +						  u32 src_port,
>>>>>>> +						  u32 dst_cid,
>>>>>>> +						  u32 dst_port)
>>>>>>> +{
>>>>>>> +	struct sk_buff *skb;
>>>>>>> +	size_t skb_len;
>>>>>>> +
>>>>>>> +	skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
>>>>>>> +
>>>>>>> +	if (!zcopy)
>>>>>>> +		skb_len += payload_len;
>>>>>>> +
>>>>>>> +	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>>>>>> +	if (!skb)
>>>>>>> +		return NULL;
>>>>>>> +
>>>>>>> +	virtio_transport_init_hdr(skb, info, src_cid, src_port,
>>>>>>> +				  dst_cid, dst_port,
>>>>>>> +				  payload_len);
>>>>>>> +
>>>>>>> +	/* Set owner here, because '__zerocopy_sg_from_iter()' uses
>>>>>>> +	 * owner of skb without check to update 'sk_wmem_alloc'.
>>>>>>> +	 */
>>>>>>> +	if (vsk)
>>>>>>> +		skb_set_owner_w(skb, sk_vsock(vsk));
>>>>>>> +
>>>>>>> +	if (info->msg && payload_len > 0) {
>>>>>>> +		int err;
>>>>>>> +
>>>>>>> +		err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
>>>>>>> +		if (err)
>>>>>>> +			goto out;
>>>>>>> +
>>>>>>> +		if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>>>>>> +			struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
>>>>>>> +
>>>>>>> +			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>>>>>> +
>>>>>>> +			if (info->msg->msg_flags & MSG_EOR)
>>>>>>> +				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>>>>>> +		}
>>>>>>> +	}
>>>>>>> +
>>>>>>> +	if (info->reply)
>>>>>>> +		virtio_vsock_skb_set_reply(skb);
>>>>>>> +
>>>>>>> +	trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>>>>>> +					 dst_cid, dst_port,
>>>>>>> +					 payload_len,
>>>>>>> +					 info->type,
>>>>>>> +					 info->op,
>>>>>>> +					 info->flags);
>>>>>>> +
>>>>>>> +	return skb;
>>>>>>> +out:
>>>>>>> +	kfree_skb(skb);
>>>>>>> +	return NULL;
>>>>>>> +}
>>>>>>> +
>>>>>>>  /* This function can only be used on connecting/connected sockets,
>>>>>>>   * since a socket assigned to a transport is required.
>>>>>>>   *
>>>>>>> @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>>>>>  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>>>>  					  struct virtio_vsock_pkt_info *info)
>>>>>>>  {
>>>>>>> +	u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
>>>>>>>  	u32 src_cid, src_port, dst_cid, dst_port;
>>>>>>>  	const struct virtio_transport *t_ops;
>>>>>>>  	struct virtio_vsock_sock *vvs;
>>>>>>>  	u32 pkt_len = info->pkt_len;
>>>>>>> +	bool can_zcopy = false;
>>>>>>>  	u32 rest_len;
>>>>>>>  	int ret;
>>>>>>>  
>>>>>>> @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>>>>  	if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
>>>>>>>  		return pkt_len;
>>>>>>>  
>>>>>>> +	if (info->msg) {
>>>>>>> +		/* If zerocopy is not enabled by 'setsockopt()', we behave as
>>>>>>> +		 * there is no MSG_ZEROCOPY flag set.
>>>>>>> +		 */
>>>>>>> +		if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
>>>>>>> +			info->msg->msg_flags &= ~MSG_ZEROCOPY;
>>>>>>> +
>>>>>>> +		if (info->msg->msg_flags & MSG_ZEROCOPY)
>>>>>>> +			can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
>>>>>>> +
>>>>>>> +		if (can_zcopy)
>>>>>>> +			max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
>>>>>>> +					    (MAX_SKB_FRAGS * PAGE_SIZE));
>>>>>>> +	}
>>>>>>> +
>>>>>>>  	rest_len = pkt_len;
>>>>>>>  
>>>>>>>  	do {
>>>>>>>  		struct sk_buff *skb;
>>>>>>>  		size_t skb_len;
>>>>>>>  
>>>>>>> -		skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
>>>>>>> +		skb_len = min(max_skb_len, rest_len);
>>>>>>>  
>>>>>>> -		skb = virtio_transport_alloc_skb(info, skb_len,
>>>>>>> +		skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
>>>>>>>  						 src_cid, src_port,
>>>>>>>  						 dst_cid, dst_port);
>>>>>>>  		if (!skb) {
>>>>>>> @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>>>>  			break;
>>>>>>>  		}
>>>>>>>  
>>>>>>> +		/* This is last skb to send this portion of data. */
>>>>>>> +		if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
>>>>>>> +		    skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
>>>>>>> +			if (virtio_transport_init_zcopy_skb(vsk, skb,
>>>>>>> +							    info->msg,
>>>>>>> +							    can_zcopy)) {
>>>>>>> +				ret = -ENOMEM;
>>>>>>> +				break;
>>>>>>> +			}
>>>>>>> +		}
>>>>>>> +
>>>>>>>  		virtio_transport_inc_tx_pkt(vvs, skb);
>>>>>>>  
>>>>>>>  		ret = t_ops->send_pkt(skb);
>>>>>>> @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
>>>>>>>  	if (!t)
>>>>>>>  		return -ENOTCONN;
>>>>>>>  
>>>>>>> -	reply = virtio_transport_alloc_skb(&info, 0,
>>>>>>> +	reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
>>>>>>>  					   le64_to_cpu(hdr->dst_cid),
>>>>>>>  					   le32_to_cpu(hdr->dst_port),
>>>>>>>  					   le64_to_cpu(hdr->src_cid),
>>>>>
>>>
>
  
Michael S. Tsirkin July 25, 2023, 1:36 p.m. UTC | #22
On Tue, Jul 25, 2023 at 04:28:14PM +0300, Arseniy Krasnov wrote:
> 
> 
> On 25.07.2023 16:22, Michael S. Tsirkin wrote:
> > On Tue, Jul 25, 2023 at 04:04:13PM +0300, Arseniy Krasnov wrote:
> >>
> >>
> >> On 25.07.2023 14:50, Michael S. Tsirkin wrote:
> >>> On Fri, Jul 21, 2023 at 08:09:03AM +0300, Arseniy Krasnov wrote:
> >>>>
> >>>>
> >>>> On 21.07.2023 00:42, Arseniy Krasnov wrote:
> >>>>> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
> >>>>> flag is set and zerocopy transmission is possible (enabled in socket
> >>>>> options and transport allows zerocopy), then non-linear skb will be
> >>>>> created and filled with the pages of user's buffer. Pages of user's
> >>>>> buffer are locked in memory by 'get_user_pages()'. Second thing that
> >>>>> this patch does is replace type of skb owning: instead of calling
> >>>>> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
> >>>>> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
> >>>>> of socket, so to decrease this field correctly proper skb destructor is
> >>>>> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
> >>>>>
> >>>>> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
> >>>>> ---
> >>>>>  Changelog:
> >>>>>  v5(big patchset) -> v1:
> >>>>>   * Refactorings of 'if' conditions.
> >>>>>   * Remove extra blank line.
> >>>>>   * Remove 'frag_off' field unneeded init.
> >>>>>   * Add function 'virtio_transport_fill_skb()' which fills both linear
> >>>>>     and non-linear skb with provided data.
> >>>>>  v1 -> v2:
> >>>>>   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
> >>>>>  v2 -> v3:
> >>>>>   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
> >>>>>     provided 'iov_iter' with data could be sent in a zerocopy mode.
> >>>>>     If this callback is not set in transport - transport allows to send
> >>>>>     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
> >>>>>     then zerocopy is allowed. Reason of this callback is that in case of
> >>>>>     G2H transmission we insert whole skb to the tx virtio queue and such
> >>>>>     skb must fit to the size of the virtio queue to be sent in a single
> >>>>>     iteration (may be tx logic in 'virtio_transport.c' could be reworked
> >>>>>     as in vhost to support partial send of current skb). This callback
> >>>>>     will be enabled only for G2H path. For details pls see comment 
> >>>>>     'Check that tx queue...' below.
> >>>>>
> >>>>>  include/net/af_vsock.h                  |   3 +
> >>>>>  net/vmw_vsock/virtio_transport.c        |  39 ++++
> >>>>>  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
> >>>>>  3 files changed, 241 insertions(+), 58 deletions(-)
> >>>>>
> >>>>> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
> >>>>> index 0e7504a42925..a6b346eeeb8e 100644
> >>>>> --- a/include/net/af_vsock.h
> >>>>> +++ b/include/net/af_vsock.h
> >>>>> @@ -177,6 +177,9 @@ struct vsock_transport {
> >>>>>  
> >>>>>  	/* Read a single skb */
> >>>>>  	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
> >>>>> +
> >>>>> +	/* Zero-copy. */
> >>>>> +	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
> >>>>>  };
> >>>>>  
> >>>>>  /**** CORE ****/
> >>>>> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
> >>>>> index 7bbcc8093e51..23cb8ed638c4 100644
> >>>>> --- a/net/vmw_vsock/virtio_transport.c
> >>>>> +++ b/net/vmw_vsock/virtio_transport.c
> >>>>> @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
> >>>>>  	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
> >>>>>  }
> >>>>>  
> >>>>> +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
> >>>>> +{
> >>>>> +	struct virtio_vsock *vsock;
> >>>>> +	bool res = false;
> >>>>> +
> >>>>> +	rcu_read_lock();
> >>>>> +
> >>>>> +	vsock = rcu_dereference(the_virtio_vsock);
> >>>>> +	if (vsock) {
> >>>>> +		struct virtqueue *vq;
> >>>>> +		int iov_pages;
> >>>>> +
> >>>>> +		vq = vsock->vqs[VSOCK_VQ_TX];
> >>>>> +
> >>>>> +		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
> >>>>> +
> >>>>> +		/* Check that tx queue is large enough to keep whole
> >>>>> +		 * data to send. This is needed, because when there is
> >>>>> +		 * not enough free space in the queue, current skb to
> >>>>> +		 * send will be reinserted to the head of tx list of
> >>>>> +		 * the socket to retry transmission later, so if skb
> >>>>> +		 * is bigger than whole queue, it will be reinserted
> >>>>> +		 * again and again, thus blocking other skbs to be sent.
> >>>>> +		 * Each page of the user provided buffer will be added
> >>>>> +		 * as a single buffer to the tx virtqueue, so compare
> >>>>> +		 * number of pages against maximum capacity of the queue.
> >>>>> +		 * +1 means buffer for the packet header.
> >>>>> +		 */
> >>>>> +		if (iov_pages + 1 <= vq->num_max)
> >>>>
> >>>> I think this check is actual only for case one we don't have indirect buffer feature.
> >>>> With indirect mode whole data to send will be packed into one indirect buffer.
> >>>>
> >>>> Thanks, Arseniy
> >>>
> >>> Actually the reverse. With indirect you are limited to num_max.
> >>> Without you are limited to whatever space is left in the
> >>> queue (which you did not check here, so you should).
> >>
> >> I mean that with indirect, we only need one buffer, and we can just wait
> >> for enough space - for this single buffer ( as we discussed a little bit before).
> >> But if indirect buffer is not supported - we need that whole packet must fit
> >> to the size of tx queue - otherwise it never be transmitted.
> >>
> >> Thanks, Arseniy
> > 
> > 
> > yes but according to virtio spec it's illegal to add s/g that is bigger
> > than queue size.
> 
> Aah, so even in case of indirect buffers feature, buffer descriptors stored in memory
> pointed by indirect buffer must be accounted against queue size ?
> 
> Thanks, Arseniy

a single indirect buffer can't exceed vq size.


> > 
> >>>
> >>>
> >>>>> +			res = true;
> >>>>> +	}
> >>>>> +
> >>>>> +	rcu_read_unlock();
> >>>
> >>> Just curious:
> >>> is the point of all this RCU dance to allow vsock
> >>> to change from under us? then why is it ok to
> >>> have it change? the virtio_transport_msgzerocopy_check_iov
> >>> will then refer to the old vsock ...
> >>>
> >>>
> >>>>> +
> >>>>> +	return res;
> >>>>> +}
> >>>>> +
> >>>>>  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
> >>>>>  
> >>>>>  static struct virtio_transport virtio_transport = {
> >>>>> @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
> >>>>>  		.seqpacket_allow          = virtio_transport_seqpacket_allow,
> >>>>>  		.seqpacket_has_data       = virtio_transport_seqpacket_has_data,
> >>>>>  
> >>>>> +		.msgzerocopy_check_iov	  = virtio_transport_msgzerocopy_check_iov,
> >>>>> +
> >>>>>  		.notify_poll_in           = virtio_transport_notify_poll_in,
> >>>>>  		.notify_poll_out          = virtio_transport_notify_poll_out,
> >>>>>  		.notify_recv_init         = virtio_transport_notify_recv_init,
> >>>>> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
> >>>>> index 26a4d10da205..e4e3d541aff4 100644
> >>>>> --- a/net/vmw_vsock/virtio_transport_common.c
> >>>>> +++ b/net/vmw_vsock/virtio_transport_common.c
> >>>>> @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
> >>>>>  	return container_of(t, struct virtio_transport, transport);
> >>>>>  }
> >>>>>  
> >>>>> -/* Returns a new packet on success, otherwise returns NULL.
> >>>>> - *
> >>>>> - * If NULL is returned, errp is set to a negative errno.
> >>>>> - */
> >>>>> -static struct sk_buff *
> >>>>> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
> >>>>> -			   size_t len,
> >>>>> -			   u32 src_cid,
> >>>>> -			   u32 src_port,
> >>>>> -			   u32 dst_cid,
> >>>>> -			   u32 dst_port)
> >>>>> -{
> >>>>> -	const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
> >>>>> -	struct virtio_vsock_hdr *hdr;
> >>>>> -	struct sk_buff *skb;
> >>>>> -	void *payload;
> >>>>> -	int err;
> >>>>> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
> >>>>> +				       size_t max_to_send)
> >>>>> +{
> >>>>> +	const struct vsock_transport *t;
> >>>>> +	struct iov_iter *iov_iter;
> >>>>>  
> >>>>> -	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
> >>>>> -	if (!skb)
> >>>>> -		return NULL;
> >>>>> +	if (!info->msg)
> >>>>> +		return false;
> >>>>>  
> >>>>> -	hdr = virtio_vsock_hdr(skb);
> >>>>> -	hdr->type	= cpu_to_le16(info->type);
> >>>>> -	hdr->op		= cpu_to_le16(info->op);
> >>>>> -	hdr->src_cid	= cpu_to_le64(src_cid);
> >>>>> -	hdr->dst_cid	= cpu_to_le64(dst_cid);
> >>>>> -	hdr->src_port	= cpu_to_le32(src_port);
> >>>>> -	hdr->dst_port	= cpu_to_le32(dst_port);
> >>>>> -	hdr->flags	= cpu_to_le32(info->flags);
> >>>>> -	hdr->len	= cpu_to_le32(len);
> >>>>> +	iov_iter = &info->msg->msg_iter;
> >>>>>  
> >>>>> -	if (info->msg && len > 0) {
> >>>>> -		payload = skb_put(skb, len);
> >>>>> -		err = memcpy_from_msg(payload, info->msg, len);
> >>>>> -		if (err)
> >>>>> -			goto out;
> >>>>> +	t = vsock_core_get_transport(info->vsk);
> >>>>>  
> >>>>> -		if (msg_data_left(info->msg) == 0 &&
> >>>>> -		    info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
> >>>>> -			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
> >>>>> +	if (t->msgzerocopy_check_iov &&
> >>>>> +	    !t->msgzerocopy_check_iov(iov_iter))
> >>>>> +		return false;
> >>>>>  
> >>>>> -			if (info->msg->msg_flags & MSG_EOR)
> >>>>> -				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
> >>>>> -		}
> >>>>> +	/* Data is simple buffer. */
> >>>>> +	if (iter_is_ubuf(iov_iter))
> >>>>> +		return true;
> >>>>> +
> >>>>> +	if (!iter_is_iovec(iov_iter))
> >>>>> +		return false;
> >>>>> +
> >>>>> +	if (iov_iter->iov_offset)
> >>>>> +		return false;
> >>>>> +
> >>>>> +	/* We can't send whole iov. */
> >>>>> +	if (iov_iter->count > max_to_send)
> >>>>> +		return false;
> >>>>> +
> >>>>> +	return true;
> >>>>> +}
> >>>>> +
> >>>>> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
> >>>>> +					   struct sk_buff *skb,
> >>>>> +					   struct msghdr *msg,
> >>>>> +					   bool zerocopy)
> >>>>> +{
> >>>>> +	struct ubuf_info *uarg;
> >>>>> +
> >>>>> +	if (msg->msg_ubuf) {
> >>>>> +		uarg = msg->msg_ubuf;
> >>>>> +		net_zcopy_get(uarg);
> >>>>> +	} else {
> >>>>> +		struct iov_iter *iter = &msg->msg_iter;
> >>>>> +		struct ubuf_info_msgzc *uarg_zc;
> >>>>> +		int len;
> >>>>> +
> >>>>> +		/* Only ITER_IOVEC or ITER_UBUF are allowed and
> >>>>> +		 * checked before.
> >>>>> +		 */
> >>>>> +		if (iter_is_iovec(iter))
> >>>>> +			len = iov_length(iter->__iov, iter->nr_segs);
> >>>>> +		else
> >>>>> +			len = iter->count;
> >>>>> +
> >>>>> +		uarg = msg_zerocopy_realloc(sk_vsock(vsk),
> >>>>> +					    len,
> >>>>> +					    NULL);
> >>>>> +		if (!uarg)
> >>>>> +			return -1;
> >>>>> +
> >>>>> +		uarg_zc = uarg_to_msgzc(uarg);
> >>>>> +		uarg_zc->zerocopy = zerocopy ? 1 : 0;
> >>>>>  	}
> >>>>>  
> >>>>> -	if (info->reply)
> >>>>> -		virtio_vsock_skb_set_reply(skb);
> >>>>> +	skb_zcopy_init(skb, uarg);
> >>>>>  
> >>>>> -	trace_virtio_transport_alloc_pkt(src_cid, src_port,
> >>>>> -					 dst_cid, dst_port,
> >>>>> -					 len,
> >>>>> -					 info->type,
> >>>>> -					 info->op,
> >>>>> -					 info->flags);
> >>>>> +	return 0;
> >>>>> +}
> >>>>>  
> >>>>> -	if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
> >>>>> -		WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
> >>>>> -		goto out;
> >>>>> +static int virtio_transport_fill_skb(struct sk_buff *skb,
> >>>>> +				     struct virtio_vsock_pkt_info *info,
> >>>>> +				     size_t len,
> >>>>> +				     bool zcopy)
> >>>>> +{
> >>>>> +	if (zcopy) {
> >>>>> +		return __zerocopy_sg_from_iter(info->msg, NULL, skb,
> >>>>> +					      &info->msg->msg_iter,
> >>>>> +					      len);
> >>>>> +	} else {
> >>>>> +		void *payload;
> >>>>> +		int err;
> >>>>> +
> >>>>> +		payload = skb_put(skb, len);
> >>>>> +		err = memcpy_from_msg(payload, info->msg, len);
> >>>>> +		if (err)
> >>>>> +			return -1;
> >>>>> +
> >>>>> +		if (msg_data_left(info->msg))
> >>>>> +			return 0;
> >>>>> +
> >>>>> +		return 0;
> >>>>>  	}
> >>>>> +}
> >>>>>  
> >>>>> -	return skb;
> >>>>> +static void virtio_transport_init_hdr(struct sk_buff *skb,
> >>>>> +				      struct virtio_vsock_pkt_info *info,
> >>>>> +				      u32 src_cid,
> >>>>> +				      u32 src_port,
> >>>>> +				      u32 dst_cid,
> >>>>> +				      u32 dst_port,
> >>>>> +				      size_t len)
> >>>>> +{
> >>>>> +	struct virtio_vsock_hdr *hdr;
> >>>>>  
> >>>>> -out:
> >>>>> -	kfree_skb(skb);
> >>>>> -	return NULL;
> >>>>> +	hdr = virtio_vsock_hdr(skb);
> >>>>> +	hdr->type	= cpu_to_le16(info->type);
> >>>>> +	hdr->op		= cpu_to_le16(info->op);
> >>>>> +	hdr->src_cid	= cpu_to_le64(src_cid);
> >>>>> +	hdr->dst_cid	= cpu_to_le64(dst_cid);
> >>>>> +	hdr->src_port	= cpu_to_le32(src_port);
> >>>>> +	hdr->dst_port	= cpu_to_le32(dst_port);
> >>>>> +	hdr->flags	= cpu_to_le32(info->flags);
> >>>>> +	hdr->len	= cpu_to_le32(len);
> >>>>>  }
> >>>>>  
> >>>>>  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
> >>>>> @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
> >>>>>  		return VIRTIO_VSOCK_TYPE_SEQPACKET;
> >>>>>  }
> >>>>>  
> >>>>> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
> >>>>> +						  struct virtio_vsock_pkt_info *info,
> >>>>> +						  size_t payload_len,
> >>>>> +						  bool zcopy,
> >>>>> +						  u32 src_cid,
> >>>>> +						  u32 src_port,
> >>>>> +						  u32 dst_cid,
> >>>>> +						  u32 dst_port)
> >>>>> +{
> >>>>> +	struct sk_buff *skb;
> >>>>> +	size_t skb_len;
> >>>>> +
> >>>>> +	skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
> >>>>> +
> >>>>> +	if (!zcopy)
> >>>>> +		skb_len += payload_len;
> >>>>> +
> >>>>> +	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
> >>>>> +	if (!skb)
> >>>>> +		return NULL;
> >>>>> +
> >>>>> +	virtio_transport_init_hdr(skb, info, src_cid, src_port,
> >>>>> +				  dst_cid, dst_port,
> >>>>> +				  payload_len);
> >>>>> +
> >>>>> +	/* Set owner here, because '__zerocopy_sg_from_iter()' uses
> >>>>> +	 * owner of skb without check to update 'sk_wmem_alloc'.
> >>>>> +	 */
> >>>>> +	if (vsk)
> >>>>> +		skb_set_owner_w(skb, sk_vsock(vsk));
> >>>>> +
> >>>>> +	if (info->msg && payload_len > 0) {
> >>>>> +		int err;
> >>>>> +
> >>>>> +		err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
> >>>>> +		if (err)
> >>>>> +			goto out;
> >>>>> +
> >>>>> +		if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
> >>>>> +			struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
> >>>>> +
> >>>>> +			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
> >>>>> +
> >>>>> +			if (info->msg->msg_flags & MSG_EOR)
> >>>>> +				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
> >>>>> +		}
> >>>>> +	}
> >>>>> +
> >>>>> +	if (info->reply)
> >>>>> +		virtio_vsock_skb_set_reply(skb);
> >>>>> +
> >>>>> +	trace_virtio_transport_alloc_pkt(src_cid, src_port,
> >>>>> +					 dst_cid, dst_port,
> >>>>> +					 payload_len,
> >>>>> +					 info->type,
> >>>>> +					 info->op,
> >>>>> +					 info->flags);
> >>>>> +
> >>>>> +	return skb;
> >>>>> +out:
> >>>>> +	kfree_skb(skb);
> >>>>> +	return NULL;
> >>>>> +}
> >>>>> +
> >>>>>  /* This function can only be used on connecting/connected sockets,
> >>>>>   * since a socket assigned to a transport is required.
> >>>>>   *
> >>>>> @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
> >>>>>  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
> >>>>>  					  struct virtio_vsock_pkt_info *info)
> >>>>>  {
> >>>>> +	u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
> >>>>>  	u32 src_cid, src_port, dst_cid, dst_port;
> >>>>>  	const struct virtio_transport *t_ops;
> >>>>>  	struct virtio_vsock_sock *vvs;
> >>>>>  	u32 pkt_len = info->pkt_len;
> >>>>> +	bool can_zcopy = false;
> >>>>>  	u32 rest_len;
> >>>>>  	int ret;
> >>>>>  
> >>>>> @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
> >>>>>  	if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
> >>>>>  		return pkt_len;
> >>>>>  
> >>>>> +	if (info->msg) {
> >>>>> +		/* If zerocopy is not enabled by 'setsockopt()', we behave as
> >>>>> +		 * there is no MSG_ZEROCOPY flag set.
> >>>>> +		 */
> >>>>> +		if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
> >>>>> +			info->msg->msg_flags &= ~MSG_ZEROCOPY;
> >>>>> +
> >>>>> +		if (info->msg->msg_flags & MSG_ZEROCOPY)
> >>>>> +			can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
> >>>>> +
> >>>>> +		if (can_zcopy)
> >>>>> +			max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
> >>>>> +					    (MAX_SKB_FRAGS * PAGE_SIZE));
> >>>>> +	}
> >>>>> +
> >>>>>  	rest_len = pkt_len;
> >>>>>  
> >>>>>  	do {
> >>>>>  		struct sk_buff *skb;
> >>>>>  		size_t skb_len;
> >>>>>  
> >>>>> -		skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
> >>>>> +		skb_len = min(max_skb_len, rest_len);
> >>>>>  
> >>>>> -		skb = virtio_transport_alloc_skb(info, skb_len,
> >>>>> +		skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
> >>>>>  						 src_cid, src_port,
> >>>>>  						 dst_cid, dst_port);
> >>>>>  		if (!skb) {
> >>>>> @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
> >>>>>  			break;
> >>>>>  		}
> >>>>>  
> >>>>> +		/* This is last skb to send this portion of data. */
> >>>>> +		if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
> >>>>> +		    skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
> >>>>> +			if (virtio_transport_init_zcopy_skb(vsk, skb,
> >>>>> +							    info->msg,
> >>>>> +							    can_zcopy)) {
> >>>>> +				ret = -ENOMEM;
> >>>>> +				break;
> >>>>> +			}
> >>>>> +		}
> >>>>> +
> >>>>>  		virtio_transport_inc_tx_pkt(vvs, skb);
> >>>>>  
> >>>>>  		ret = t_ops->send_pkt(skb);
> >>>>> @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
> >>>>>  	if (!t)
> >>>>>  		return -ENOTCONN;
> >>>>>  
> >>>>> -	reply = virtio_transport_alloc_skb(&info, 0,
> >>>>> +	reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
> >>>>>  					   le64_to_cpu(hdr->dst_cid),
> >>>>>  					   le32_to_cpu(hdr->dst_port),
> >>>>>  					   le64_to_cpu(hdr->src_cid),
> >>>
> >
  
Arseniy Krasnov July 27, 2023, 8:32 a.m. UTC | #23
On 25.07.2023 15:28, Stefano Garzarella wrote:
> On Tue, Jul 25, 2023 at 12:16:11PM +0300, Arseniy Krasnov wrote:
>>
>>
>> On 25.07.2023 11:46, Arseniy Krasnov wrote:
>>>
>>>
>>> On 25.07.2023 11:43, Stefano Garzarella wrote:
>>>> On Fri, Jul 21, 2023 at 08:09:03AM +0300, Arseniy Krasnov wrote:
>>>>>
>>>>>
>>>>> On 21.07.2023 00:42, Arseniy Krasnov wrote:
>>>>>> This adds handling of MSG_ZEROCOPY flag on transmission path: if this
>>>>>> flag is set and zerocopy transmission is possible (enabled in socket
>>>>>> options and transport allows zerocopy), then non-linear skb will be
>>>>>> created and filled with the pages of user's buffer. Pages of user's
>>>>>> buffer are locked in memory by 'get_user_pages()'. Second thing that
>>>>>> this patch does is replace type of skb owning: instead of calling
>>>>>> 'skb_set_owner_sk_safe()' it calls 'skb_set_owner_w()'. Reason of this
>>>>>> change is that '__zerocopy_sg_from_iter()' increments 'sk_wmem_alloc'
>>>>>> of socket, so to decrease this field correctly proper skb destructor is
>>>>>> needed: 'sock_wfree()'. This destructor is set by 'skb_set_owner_w()'.
>>>>>>
>>>>>> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
>>>>>> ---
>>>>>>  Changelog:
>>>>>>  v5(big patchset) -> v1:
>>>>>>   * Refactorings of 'if' conditions.
>>>>>>   * Remove extra blank line.
>>>>>>   * Remove 'frag_off' field unneeded init.
>>>>>>   * Add function 'virtio_transport_fill_skb()' which fills both linear
>>>>>>     and non-linear skb with provided data.
>>>>>>  v1 -> v2:
>>>>>>   * Use original order of last four arguments in 'virtio_transport_alloc_skb()'.
>>>>>>  v2 -> v3:
>>>>>>   * Add new transport callback: 'msgzerocopy_check_iov'. It checks that
>>>>>>     provided 'iov_iter' with data could be sent in a zerocopy mode.
>>>>>>     If this callback is not set in transport - transport allows to send
>>>>>>     any 'iov_iter' in zerocopy mode. Otherwise - if callback returns 'true'
>>>>>>     then zerocopy is allowed. Reason of this callback is that in case of
>>>>>>     G2H transmission we insert whole skb to the tx virtio queue and such
>>>>>>     skb must fit to the size of the virtio queue to be sent in a single
>>>>>>     iteration (may be tx logic in 'virtio_transport.c' could be reworked
>>>>>>     as in vhost to support partial send of current skb). This callback
>>>>>>     will be enabled only for G2H path. For details pls see comment
>>>>>>     'Check that tx queue...' below.
>>>>>>
>>>>>>  include/net/af_vsock.h                  |   3 +
>>>>>>  net/vmw_vsock/virtio_transport.c        |  39 ++++
>>>>>>  net/vmw_vsock/virtio_transport_common.c | 257 ++++++++++++++++++------
>>>>>>  3 files changed, 241 insertions(+), 58 deletions(-)
>>>>>>
>>>>>> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
>>>>>> index 0e7504a42925..a6b346eeeb8e 100644
>>>>>> --- a/include/net/af_vsock.h
>>>>>> +++ b/include/net/af_vsock.h
>>>>>> @@ -177,6 +177,9 @@ struct vsock_transport {
>>>>>>
>>>>>>      /* Read a single skb */
>>>>>>      int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
>>>>>> +
>>>>>> +    /* Zero-copy. */
>>>>>> +    bool (*msgzerocopy_check_iov)(const struct iov_iter *);
>>>>>>  };
>>>>>>
>>>>>>  /**** CORE ****/
>>>>>> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>>>>>> index 7bbcc8093e51..23cb8ed638c4 100644
>>>>>> --- a/net/vmw_vsock/virtio_transport.c
>>>>>> +++ b/net/vmw_vsock/virtio_transport.c
>>>>>> @@ -442,6 +442,43 @@ static void virtio_vsock_rx_done(struct virtqueue *vq)
>>>>>>      queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>>>>>>  }
>>>>>>
>>>>>> +static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
>>>>>> +{
>>>>>> +    struct virtio_vsock *vsock;
>>>>>> +    bool res = false;
>>>>>> +
>>>>>> +    rcu_read_lock();
>>>>>> +
>>>>>> +    vsock = rcu_dereference(the_virtio_vsock);
>>>>>> +    if (vsock) {
> 
> Just noted, what about the following to reduce the indentation?
> 
>         if (!vsock) {
>             goto out;
>         }
>             ...
>             ...
>     out:
>         rcu_read_unlock();
>         return res;
> 
>>>>>> +        struct virtqueue *vq;
>>>>>> +        int iov_pages;
>>>>>> +
>>>>>> +        vq = vsock->vqs[VSOCK_VQ_TX];
>>>>>> +
>>>>>> +        iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
>>>>>> +
>>>>>> +        /* Check that tx queue is large enough to keep whole
>>>>>> +         * data to send. This is needed, because when there is
>>>>>> +         * not enough free space in the queue, current skb to
>>>>>> +         * send will be reinserted to the head of tx list of
>>>>>> +         * the socket to retry transmission later, so if skb
>>>>>> +         * is bigger than whole queue, it will be reinserted
>>>>>> +         * again and again, thus blocking other skbs to be sent.
>>>>>> +         * Each page of the user provided buffer will be added
>>>>>> +         * as a single buffer to the tx virtqueue, so compare
>>>>>> +         * number of pages against maximum capacity of the queue.
>>>>>> +         * +1 means buffer for the packet header.
>>>>>> +         */
>>>>>> +        if (iov_pages + 1 <= vq->num_max)
>>>>>
>>>>> I think this check is actual only for case one we don't have indirect buffer feature.
>>>>> With indirect mode whole data to send will be packed into one indirect buffer.
>>>>
>>>> I think so.
>>>> So, should we check also that here?
>>>>
>>>>>
>>>>> Thanks, Arseniy
>>>>>
>>>>>> +            res = true;
>>>>>> +    }
>>>>>> +
>>>>>> +    rcu_read_unlock();
>>>>>> +
>>>>>> +    return res;
>>>>>> +}
>>>>>> +
>>>>>>  static bool virtio_transport_seqpacket_allow(u32 remote_cid);
>>>>>>
>>>>>>  static struct virtio_transport virtio_transport = {
>>>>>> @@ -475,6 +512,8 @@ static struct virtio_transport virtio_transport = {
>>>>>>          .seqpacket_allow          = virtio_transport_seqpacket_allow,
>>>>>>          .seqpacket_has_data       = virtio_transport_seqpacket_has_data,
>>>>>>
>>>>>> +        .msgzerocopy_check_iov      = virtio_transport_msgzerocopy_check_iov,
>>>>>> +
>>>>>>          .notify_poll_in           = virtio_transport_notify_poll_in,
>>>>>>          .notify_poll_out          = virtio_transport_notify_poll_out,
>>>>>>          .notify_recv_init         = virtio_transport_notify_recv_init,
>>>>>> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
>>>>>> index 26a4d10da205..e4e3d541aff4 100644
>>>>>> --- a/net/vmw_vsock/virtio_transport_common.c
>>>>>> +++ b/net/vmw_vsock/virtio_transport_common.c
>>>>>> @@ -37,73 +37,122 @@ virtio_transport_get_ops(struct vsock_sock *vsk)
>>>>>>      return container_of(t, struct virtio_transport, transport);
>>>>>>  }
>>>>>>
>>>>>> -/* Returns a new packet on success, otherwise returns NULL.
>>>>>> - *
>>>>>> - * If NULL is returned, errp is set to a negative errno.
>>>>>> - */
>>>>>> -static struct sk_buff *
>>>>>> -virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
>>>>>> -               size_t len,
>>>>>> -               u32 src_cid,
>>>>>> -               u32 src_port,
>>>>>> -               u32 dst_cid,
>>>>>> -               u32 dst_port)
>>>>>> -{
>>>>>> -    const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
>>>>>> -    struct virtio_vsock_hdr *hdr;
>>>>>> -    struct sk_buff *skb;
>>>>>> -    void *payload;
>>>>>> -    int err;
>>>>>> +static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
>>>>>> +                       size_t max_to_send)
>>>>>> +{
>>>>>> +    const struct vsock_transport *t;
>>>>>> +    struct iov_iter *iov_iter;
>>>>>>
>>>>>> -    skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>>>>> -    if (!skb)
>>>>>> -        return NULL;
>>>>>> +    if (!info->msg)
>>>>>> +        return false;
>>>>>>
>>>>>> -    hdr = virtio_vsock_hdr(skb);
>>>>>> -    hdr->type    = cpu_to_le16(info->type);
>>>>>> -    hdr->op        = cpu_to_le16(info->op);
>>>>>> -    hdr->src_cid    = cpu_to_le64(src_cid);
>>>>>> -    hdr->dst_cid    = cpu_to_le64(dst_cid);
>>>>>> -    hdr->src_port    = cpu_to_le32(src_port);
>>>>>> -    hdr->dst_port    = cpu_to_le32(dst_port);
>>>>>> -    hdr->flags    = cpu_to_le32(info->flags);
>>>>>> -    hdr->len    = cpu_to_le32(len);
>>>>>> +    iov_iter = &info->msg->msg_iter;
>>>>>>
>>>>>> -    if (info->msg && len > 0) {
>>>>>> -        payload = skb_put(skb, len);
>>>>>> -        err = memcpy_from_msg(payload, info->msg, len);
>>>>>> -        if (err)
>>>>>> -            goto out;
>>>>>> +    t = vsock_core_get_transport(info->vsk);
>>>>>>
>>>>>> -        if (msg_data_left(info->msg) == 0 &&
>>>>>> -            info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>>>>> -            hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>>>>> +    if (t->msgzerocopy_check_iov &&
>>>>>> +        !t->msgzerocopy_check_iov(iov_iter))
>>>>>> +        return false;
>>>>
>>>> I'd avoid adding a new transport callback used only internally in virtio
>>>> transports.
>>>
>>> Ok, I see.
>>>
>>>>
>>>> Usually the transport callbacks are used in af_vsock.c, if we need a
>>>> callback just for virtio transports, maybe better to add it in struct
>>>> virtio_vsock_pkt_info or struct virtio_vsock_sock.
>>
>> Hm, may be I just need to move this callback from 'struct vsock_transport' to parent 'struct virtio_transport',
>> after 'send_pkt' callback. In this case:
>> 1) AF_VSOCK part is not touched.
>> 2) This callback stays in 'virtio_transport.c' and is set also in this file.
>>   vhost and loopback are unchanged - only 'send_pkt' still enabled in both
>>   files for these two transports.
> 
> Yep, this could also work!
> 
> Stefano

Great! I'll send this implementation when this patchset for MSG_PEEK will be merged
to net-next as both conflicts with each other.

https://lore.kernel.org/netdev/20230726060150-mutt-send-email-mst@kernel.org/T/#m56f3b850361a412735616145162d2d9df25f6350

Thanks, Arseniy

> 
>>
>> Thanks, Arseniy
>>
>>>>
>>>> Maybe better the last one so we don't have to allocate pointer space
>>>> for each packet and you should reach it via info.
>>>
>>> Ok, thanks, I'll try this way
>>>
>>> Thanks, Arseniy
>>>
>>>>
>>>> Thanks,
>>>> Stefano
>>>>
>>>>>>
>>>>>> -            if (info->msg->msg_flags & MSG_EOR)
>>>>>> -                hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>>>>> -        }
>>>>>> +    /* Data is simple buffer. */
>>>>>> +    if (iter_is_ubuf(iov_iter))
>>>>>> +        return true;
>>>>>> +
>>>>>> +    if (!iter_is_iovec(iov_iter))
>>>>>> +        return false;
>>>>>> +
>>>>>> +    if (iov_iter->iov_offset)
>>>>>> +        return false;
>>>>>> +
>>>>>> +    /* We can't send whole iov. */
>>>>>> +    if (iov_iter->count > max_to_send)
>>>>>> +        return false;
>>>>>> +
>>>>>> +    return true;
>>>>>> +}
>>>>>> +
>>>>>> +static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
>>>>>> +                       struct sk_buff *skb,
>>>>>> +                       struct msghdr *msg,
>>>>>> +                       bool zerocopy)
>>>>>> +{
>>>>>> +    struct ubuf_info *uarg;
>>>>>> +
>>>>>> +    if (msg->msg_ubuf) {
>>>>>> +        uarg = msg->msg_ubuf;
>>>>>> +        net_zcopy_get(uarg);
>>>>>> +    } else {
>>>>>> +        struct iov_iter *iter = &msg->msg_iter;
>>>>>> +        struct ubuf_info_msgzc *uarg_zc;
>>>>>> +        int len;
>>>>>> +
>>>>>> +        /* Only ITER_IOVEC or ITER_UBUF are allowed and
>>>>>> +         * checked before.
>>>>>> +         */
>>>>>> +        if (iter_is_iovec(iter))
>>>>>> +            len = iov_length(iter->__iov, iter->nr_segs);
>>>>>> +        else
>>>>>> +            len = iter->count;
>>>>>> +
>>>>>> +        uarg = msg_zerocopy_realloc(sk_vsock(vsk),
>>>>>> +                        len,
>>>>>> +                        NULL);
>>>>>> +        if (!uarg)
>>>>>> +            return -1;
>>>>>> +
>>>>>> +        uarg_zc = uarg_to_msgzc(uarg);
>>>>>> +        uarg_zc->zerocopy = zerocopy ? 1 : 0;
>>>>>>      }
>>>>>>
>>>>>> -    if (info->reply)
>>>>>> -        virtio_vsock_skb_set_reply(skb);
>>>>>> +    skb_zcopy_init(skb, uarg);
>>>>>>
>>>>>> -    trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>>>>> -                     dst_cid, dst_port,
>>>>>> -                     len,
>>>>>> -                     info->type,
>>>>>> -                     info->op,
>>>>>> -                     info->flags);
>>>>>> +    return 0;
>>>>>> +}
>>>>>>
>>>>>> -    if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
>>>>>> -        WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
>>>>>> -        goto out;
>>>>>> +static int virtio_transport_fill_skb(struct sk_buff *skb,
>>>>>> +                     struct virtio_vsock_pkt_info *info,
>>>>>> +                     size_t len,
>>>>>> +                     bool zcopy)
>>>>>> +{
>>>>>> +    if (zcopy) {
>>>>>> +        return __zerocopy_sg_from_iter(info->msg, NULL, skb,
>>>>>> +                          &info->msg->msg_iter,
>>>>>> +                          len);
>>>>>> +    } else {
>>>>>> +        void *payload;
>>>>>> +        int err;
>>>>>> +
>>>>>> +        payload = skb_put(skb, len);
>>>>>> +        err = memcpy_from_msg(payload, info->msg, len);
>>>>>> +        if (err)
>>>>>> +            return -1;
>>>>>> +
>>>>>> +        if (msg_data_left(info->msg))
>>>>>> +            return 0;
>>>>>> +
>>>>>> +        return 0;
>>>>>>      }
>>>>>> +}
>>>>>>
>>>>>> -    return skb;
>>>>>> +static void virtio_transport_init_hdr(struct sk_buff *skb,
>>>>>> +                      struct virtio_vsock_pkt_info *info,
>>>>>> +                      u32 src_cid,
>>>>>> +                      u32 src_port,
>>>>>> +                      u32 dst_cid,
>>>>>> +                      u32 dst_port,
>>>>>> +                      size_t len)
>>>>>> +{
>>>>>> +    struct virtio_vsock_hdr *hdr;
>>>>>>
>>>>>> -out:
>>>>>> -    kfree_skb(skb);
>>>>>> -    return NULL;
>>>>>> +    hdr = virtio_vsock_hdr(skb);
>>>>>> +    hdr->type    = cpu_to_le16(info->type);
>>>>>> +    hdr->op        = cpu_to_le16(info->op);
>>>>>> +    hdr->src_cid    = cpu_to_le64(src_cid);
>>>>>> +    hdr->dst_cid    = cpu_to_le64(dst_cid);
>>>>>> +    hdr->src_port    = cpu_to_le32(src_port);
>>>>>> +    hdr->dst_port    = cpu_to_le32(dst_port);
>>>>>> +    hdr->flags    = cpu_to_le32(info->flags);
>>>>>> +    hdr->len    = cpu_to_le32(len);
>>>>>>  }
>>>>>>
>>>>>>  static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
>>>>>> @@ -214,6 +263,70 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>>>>          return VIRTIO_VSOCK_TYPE_SEQPACKET;
>>>>>>  }
>>>>>>
>>>>>> +static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
>>>>>> +                          struct virtio_vsock_pkt_info *info,
>>>>>> +                          size_t payload_len,
>>>>>> +                          bool zcopy,
>>>>>> +                          u32 src_cid,
>>>>>> +                          u32 src_port,
>>>>>> +                          u32 dst_cid,
>>>>>> +                          u32 dst_port)
>>>>>> +{
>>>>>> +    struct sk_buff *skb;
>>>>>> +    size_t skb_len;
>>>>>> +
>>>>>> +    skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
>>>>>> +
>>>>>> +    if (!zcopy)
>>>>>> +        skb_len += payload_len;
>>>>>> +
>>>>>> +    skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
>>>>>> +    if (!skb)
>>>>>> +        return NULL;
>>>>>> +
>>>>>> +    virtio_transport_init_hdr(skb, info, src_cid, src_port,
>>>>>> +                  dst_cid, dst_port,
>>>>>> +                  payload_len);
>>>>>> +
>>>>>> +    /* Set owner here, because '__zerocopy_sg_from_iter()' uses
>>>>>> +     * owner of skb without check to update 'sk_wmem_alloc'.
>>>>>> +     */
>>>>>> +    if (vsk)
>>>>>> +        skb_set_owner_w(skb, sk_vsock(vsk));
>>>>>> +
>>>>>> +    if (info->msg && payload_len > 0) {
>>>>>> +        int err;
>>>>>> +
>>>>>> +        err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
>>>>>> +        if (err)
>>>>>> +            goto out;
>>>>>> +
>>>>>> +        if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>>>>> +            struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
>>>>>> +
>>>>>> +            hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>>>>> +
>>>>>> +            if (info->msg->msg_flags & MSG_EOR)
>>>>>> +                hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
>>>>>> +        }
>>>>>> +    }
>>>>>> +
>>>>>> +    if (info->reply)
>>>>>> +        virtio_vsock_skb_set_reply(skb);
>>>>>> +
>>>>>> +    trace_virtio_transport_alloc_pkt(src_cid, src_port,
>>>>>> +                     dst_cid, dst_port,
>>>>>> +                     payload_len,
>>>>>> +                     info->type,
>>>>>> +                     info->op,
>>>>>> +                     info->flags);
>>>>>> +
>>>>>> +    return skb;
>>>>>> +out:
>>>>>> +    kfree_skb(skb);
>>>>>> +    return NULL;
>>>>>> +}
>>>>>> +
>>>>>>  /* This function can only be used on connecting/connected sockets,
>>>>>>   * since a socket assigned to a transport is required.
>>>>>>   *
>>>>>> @@ -222,10 +335,12 @@ static u16 virtio_transport_get_type(struct sock *sk)
>>>>>>  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>>>                        struct virtio_vsock_pkt_info *info)
>>>>>>  {
>>>>>> +    u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
>>>>>>      u32 src_cid, src_port, dst_cid, dst_port;
>>>>>>      const struct virtio_transport *t_ops;
>>>>>>      struct virtio_vsock_sock *vvs;
>>>>>>      u32 pkt_len = info->pkt_len;
>>>>>> +    bool can_zcopy = false;
>>>>>>      u32 rest_len;
>>>>>>      int ret;
>>>>>>
>>>>>> @@ -254,15 +369,30 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>>>      if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
>>>>>>          return pkt_len;
>>>>>>
>>>>>> +    if (info->msg) {
>>>>>> +        /* If zerocopy is not enabled by 'setsockopt()', we behave as
>>>>>> +         * there is no MSG_ZEROCOPY flag set.
>>>>>> +         */
>>>>>> +        if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
>>>>>> +            info->msg->msg_flags &= ~MSG_ZEROCOPY;
>>>>>> +
>>>>>> +        if (info->msg->msg_flags & MSG_ZEROCOPY)
>>>>>> +            can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
>>>>>> +
>>>>>> +        if (can_zcopy)
>>>>>> +            max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
>>>>>> +                        (MAX_SKB_FRAGS * PAGE_SIZE));
>>>>>> +    }
>>>>>> +
>>>>>>      rest_len = pkt_len;
>>>>>>
>>>>>>      do {
>>>>>>          struct sk_buff *skb;
>>>>>>          size_t skb_len;
>>>>>>
>>>>>> -        skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
>>>>>> +        skb_len = min(max_skb_len, rest_len);
>>>>>>
>>>>>> -        skb = virtio_transport_alloc_skb(info, skb_len,
>>>>>> +        skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
>>>>>>                           src_cid, src_port,
>>>>>>                           dst_cid, dst_port);
>>>>>>          if (!skb) {
>>>>>> @@ -270,6 +400,17 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
>>>>>>              break;
>>>>>>          }
>>>>>>
>>>>>> +        /* This is last skb to send this portion of data. */
>>>>>> +        if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
>>>>>> +            skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
>>>>>> +            if (virtio_transport_init_zcopy_skb(vsk, skb,
>>>>>> +                                info->msg,
>>>>>> +                                can_zcopy)) {
>>>>>> +                ret = -ENOMEM;
>>>>>> +                break;
>>>>>> +            }
>>>>>> +        }
>>>>>> +
>>>>>>          virtio_transport_inc_tx_pkt(vvs, skb);
>>>>>>
>>>>>>          ret = t_ops->send_pkt(skb);
>>>>>> @@ -934,7 +1075,7 @@ static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
>>>>>>      if (!t)
>>>>>>          return -ENOTCONN;
>>>>>>
>>>>>> -    reply = virtio_transport_alloc_skb(&info, 0,
>>>>>> +    reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
>>>>>>                         le64_to_cpu(hdr->dst_cid),
>>>>>>                         le32_to_cpu(hdr->dst_port),
>>>>>>                         le64_to_cpu(hdr->src_cid),
>>>>>
>>>>
>>
>
  
Stefano Garzarella July 27, 2023, 8:54 a.m. UTC | #24
On Thu, Jul 27, 2023 at 11:32:00AM +0300, Arseniy Krasnov wrote:
>On 25.07.2023 15:28, Stefano Garzarella wrote:
>> On Tue, Jul 25, 2023 at 12:16:11PM +0300, Arseniy Krasnov wrote:
>>> On 25.07.2023 11:46, Arseniy Krasnov wrote:
>>>> On 25.07.2023 11:43, Stefano Garzarella wrote:
>>>>> On Fri, Jul 21, 2023 at 08:09:03AM +0300, Arseniy Krasnov wrote:

[...]

>>>>>>> +    t = vsock_core_get_transport(info->vsk);
>>>>>>>
>>>>>>> -        if (msg_data_left(info->msg) == 0 &&
>>>>>>> -            info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
>>>>>>> -            hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
>>>>>>> +    if (t->msgzerocopy_check_iov &&
>>>>>>> +        !t->msgzerocopy_check_iov(iov_iter))
>>>>>>> +        return false;
>>>>>
>>>>> I'd avoid adding a new transport callback used only internally in virtio
>>>>> transports.
>>>>
>>>> Ok, I see.
>>>>
>>>>>
>>>>> Usually the transport callbacks are used in af_vsock.c, if we need a
>>>>> callback just for virtio transports, maybe better to add it in struct
>>>>> virtio_vsock_pkt_info or struct virtio_vsock_sock.
>>>
>>> Hm, may be I just need to move this callback from 'struct vsock_transport' to parent 'struct virtio_transport',
>>> after 'send_pkt' callback. In this case:
>>> 1) AF_VSOCK part is not touched.
>>> 2) This callback stays in 'virtio_transport.c' and is set also in this file.
>>>   vhost and loopback are unchanged - only 'send_pkt' still enabled in both
>>>   files for these two transports.
>>
>> Yep, this could also work!
>>
>> Stefano
>
>Great! I'll send this implementation when this patchset for MSG_PEEK will be merged
>to net-next as both conflicts with each other.
>
>https://lore.kernel.org/netdev/20230726060150-mutt-send-email-mst@kernel.org/T/#m56f3b850361a412735616145162d2d9df25f6350

Ack!

Thanks,
Stefano
  

Patch

diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
index 0e7504a42925..a6b346eeeb8e 100644
--- a/include/net/af_vsock.h
+++ b/include/net/af_vsock.h
@@ -177,6 +177,9 @@  struct vsock_transport {
 
 	/* Read a single skb */
 	int (*read_skb)(struct vsock_sock *, skb_read_actor_t);
+
+	/* Zero-copy. */
+	bool (*msgzerocopy_check_iov)(const struct iov_iter *);
 };
 
 /**** CORE ****/
diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
index 7bbcc8093e51..23cb8ed638c4 100644
--- a/net/vmw_vsock/virtio_transport.c
+++ b/net/vmw_vsock/virtio_transport.c
@@ -442,6 +442,43 @@  static void virtio_vsock_rx_done(struct virtqueue *vq)
 	queue_work(virtio_vsock_workqueue, &vsock->rx_work);
 }
 
+static bool virtio_transport_msgzerocopy_check_iov(const struct iov_iter *iov)
+{
+	struct virtio_vsock *vsock;
+	bool res = false;
+
+	rcu_read_lock();
+
+	vsock = rcu_dereference(the_virtio_vsock);
+	if (vsock) {
+		struct virtqueue *vq;
+		int iov_pages;
+
+		vq = vsock->vqs[VSOCK_VQ_TX];
+
+		iov_pages = round_up(iov->count, PAGE_SIZE) / PAGE_SIZE;
+
+		/* Check that tx queue is large enough to keep whole
+		 * data to send. This is needed, because when there is
+		 * not enough free space in the queue, current skb to
+		 * send will be reinserted to the head of tx list of
+		 * the socket to retry transmission later, so if skb
+		 * is bigger than whole queue, it will be reinserted
+		 * again and again, thus blocking other skbs to be sent.
+		 * Each page of the user provided buffer will be added
+		 * as a single buffer to the tx virtqueue, so compare
+		 * number of pages against maximum capacity of the queue.
+		 * +1 means buffer for the packet header.
+		 */
+		if (iov_pages + 1 <= vq->num_max)
+			res = true;
+	}
+
+	rcu_read_unlock();
+
+	return res;
+}
+
 static bool virtio_transport_seqpacket_allow(u32 remote_cid);
 
 static struct virtio_transport virtio_transport = {
@@ -475,6 +512,8 @@  static struct virtio_transport virtio_transport = {
 		.seqpacket_allow          = virtio_transport_seqpacket_allow,
 		.seqpacket_has_data       = virtio_transport_seqpacket_has_data,
 
+		.msgzerocopy_check_iov	  = virtio_transport_msgzerocopy_check_iov,
+
 		.notify_poll_in           = virtio_transport_notify_poll_in,
 		.notify_poll_out          = virtio_transport_notify_poll_out,
 		.notify_recv_init         = virtio_transport_notify_recv_init,
diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
index 26a4d10da205..e4e3d541aff4 100644
--- a/net/vmw_vsock/virtio_transport_common.c
+++ b/net/vmw_vsock/virtio_transport_common.c
@@ -37,73 +37,122 @@  virtio_transport_get_ops(struct vsock_sock *vsk)
 	return container_of(t, struct virtio_transport, transport);
 }
 
-/* Returns a new packet on success, otherwise returns NULL.
- *
- * If NULL is returned, errp is set to a negative errno.
- */
-static struct sk_buff *
-virtio_transport_alloc_skb(struct virtio_vsock_pkt_info *info,
-			   size_t len,
-			   u32 src_cid,
-			   u32 src_port,
-			   u32 dst_cid,
-			   u32 dst_port)
-{
-	const size_t skb_len = VIRTIO_VSOCK_SKB_HEADROOM + len;
-	struct virtio_vsock_hdr *hdr;
-	struct sk_buff *skb;
-	void *payload;
-	int err;
+static bool virtio_transport_can_zcopy(struct virtio_vsock_pkt_info *info,
+				       size_t max_to_send)
+{
+	const struct vsock_transport *t;
+	struct iov_iter *iov_iter;
 
-	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
-	if (!skb)
-		return NULL;
+	if (!info->msg)
+		return false;
 
-	hdr = virtio_vsock_hdr(skb);
-	hdr->type	= cpu_to_le16(info->type);
-	hdr->op		= cpu_to_le16(info->op);
-	hdr->src_cid	= cpu_to_le64(src_cid);
-	hdr->dst_cid	= cpu_to_le64(dst_cid);
-	hdr->src_port	= cpu_to_le32(src_port);
-	hdr->dst_port	= cpu_to_le32(dst_port);
-	hdr->flags	= cpu_to_le32(info->flags);
-	hdr->len	= cpu_to_le32(len);
+	iov_iter = &info->msg->msg_iter;
 
-	if (info->msg && len > 0) {
-		payload = skb_put(skb, len);
-		err = memcpy_from_msg(payload, info->msg, len);
-		if (err)
-			goto out;
+	t = vsock_core_get_transport(info->vsk);
 
-		if (msg_data_left(info->msg) == 0 &&
-		    info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
-			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
+	if (t->msgzerocopy_check_iov &&
+	    !t->msgzerocopy_check_iov(iov_iter))
+		return false;
 
-			if (info->msg->msg_flags & MSG_EOR)
-				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
-		}
+	/* Data is simple buffer. */
+	if (iter_is_ubuf(iov_iter))
+		return true;
+
+	if (!iter_is_iovec(iov_iter))
+		return false;
+
+	if (iov_iter->iov_offset)
+		return false;
+
+	/* We can't send whole iov. */
+	if (iov_iter->count > max_to_send)
+		return false;
+
+	return true;
+}
+
+static int virtio_transport_init_zcopy_skb(struct vsock_sock *vsk,
+					   struct sk_buff *skb,
+					   struct msghdr *msg,
+					   bool zerocopy)
+{
+	struct ubuf_info *uarg;
+
+	if (msg->msg_ubuf) {
+		uarg = msg->msg_ubuf;
+		net_zcopy_get(uarg);
+	} else {
+		struct iov_iter *iter = &msg->msg_iter;
+		struct ubuf_info_msgzc *uarg_zc;
+		int len;
+
+		/* Only ITER_IOVEC or ITER_UBUF are allowed and
+		 * checked before.
+		 */
+		if (iter_is_iovec(iter))
+			len = iov_length(iter->__iov, iter->nr_segs);
+		else
+			len = iter->count;
+
+		uarg = msg_zerocopy_realloc(sk_vsock(vsk),
+					    len,
+					    NULL);
+		if (!uarg)
+			return -1;
+
+		uarg_zc = uarg_to_msgzc(uarg);
+		uarg_zc->zerocopy = zerocopy ? 1 : 0;
 	}
 
-	if (info->reply)
-		virtio_vsock_skb_set_reply(skb);
+	skb_zcopy_init(skb, uarg);
 
-	trace_virtio_transport_alloc_pkt(src_cid, src_port,
-					 dst_cid, dst_port,
-					 len,
-					 info->type,
-					 info->op,
-					 info->flags);
+	return 0;
+}
 
-	if (info->vsk && !skb_set_owner_sk_safe(skb, sk_vsock(info->vsk))) {
-		WARN_ONCE(1, "failed to allocate skb on vsock socket with sk_refcnt == 0\n");
-		goto out;
+static int virtio_transport_fill_skb(struct sk_buff *skb,
+				     struct virtio_vsock_pkt_info *info,
+				     size_t len,
+				     bool zcopy)
+{
+	if (zcopy) {
+		return __zerocopy_sg_from_iter(info->msg, NULL, skb,
+					      &info->msg->msg_iter,
+					      len);
+	} else {
+		void *payload;
+		int err;
+
+		payload = skb_put(skb, len);
+		err = memcpy_from_msg(payload, info->msg, len);
+		if (err)
+			return -1;
+
+		if (msg_data_left(info->msg))
+			return 0;
+
+		return 0;
 	}
+}
 
-	return skb;
+static void virtio_transport_init_hdr(struct sk_buff *skb,
+				      struct virtio_vsock_pkt_info *info,
+				      u32 src_cid,
+				      u32 src_port,
+				      u32 dst_cid,
+				      u32 dst_port,
+				      size_t len)
+{
+	struct virtio_vsock_hdr *hdr;
 
-out:
-	kfree_skb(skb);
-	return NULL;
+	hdr = virtio_vsock_hdr(skb);
+	hdr->type	= cpu_to_le16(info->type);
+	hdr->op		= cpu_to_le16(info->op);
+	hdr->src_cid	= cpu_to_le64(src_cid);
+	hdr->dst_cid	= cpu_to_le64(dst_cid);
+	hdr->src_port	= cpu_to_le32(src_port);
+	hdr->dst_port	= cpu_to_le32(dst_port);
+	hdr->flags	= cpu_to_le32(info->flags);
+	hdr->len	= cpu_to_le32(len);
 }
 
 static void virtio_transport_copy_nonlinear_skb(const struct sk_buff *skb,
@@ -214,6 +263,70 @@  static u16 virtio_transport_get_type(struct sock *sk)
 		return VIRTIO_VSOCK_TYPE_SEQPACKET;
 }
 
+static struct sk_buff *virtio_transport_alloc_skb(struct vsock_sock *vsk,
+						  struct virtio_vsock_pkt_info *info,
+						  size_t payload_len,
+						  bool zcopy,
+						  u32 src_cid,
+						  u32 src_port,
+						  u32 dst_cid,
+						  u32 dst_port)
+{
+	struct sk_buff *skb;
+	size_t skb_len;
+
+	skb_len = VIRTIO_VSOCK_SKB_HEADROOM;
+
+	if (!zcopy)
+		skb_len += payload_len;
+
+	skb = virtio_vsock_alloc_skb(skb_len, GFP_KERNEL);
+	if (!skb)
+		return NULL;
+
+	virtio_transport_init_hdr(skb, info, src_cid, src_port,
+				  dst_cid, dst_port,
+				  payload_len);
+
+	/* Set owner here, because '__zerocopy_sg_from_iter()' uses
+	 * owner of skb without check to update 'sk_wmem_alloc'.
+	 */
+	if (vsk)
+		skb_set_owner_w(skb, sk_vsock(vsk));
+
+	if (info->msg && payload_len > 0) {
+		int err;
+
+		err = virtio_transport_fill_skb(skb, info, payload_len, zcopy);
+		if (err)
+			goto out;
+
+		if (info->type == VIRTIO_VSOCK_TYPE_SEQPACKET) {
+			struct virtio_vsock_hdr *hdr = virtio_vsock_hdr(skb);
+
+			hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOM);
+
+			if (info->msg->msg_flags & MSG_EOR)
+				hdr->flags |= cpu_to_le32(VIRTIO_VSOCK_SEQ_EOR);
+		}
+	}
+
+	if (info->reply)
+		virtio_vsock_skb_set_reply(skb);
+
+	trace_virtio_transport_alloc_pkt(src_cid, src_port,
+					 dst_cid, dst_port,
+					 payload_len,
+					 info->type,
+					 info->op,
+					 info->flags);
+
+	return skb;
+out:
+	kfree_skb(skb);
+	return NULL;
+}
+
 /* This function can only be used on connecting/connected sockets,
  * since a socket assigned to a transport is required.
  *
@@ -222,10 +335,12 @@  static u16 virtio_transport_get_type(struct sock *sk)
 static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
 					  struct virtio_vsock_pkt_info *info)
 {
+	u32 max_skb_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
 	u32 src_cid, src_port, dst_cid, dst_port;
 	const struct virtio_transport *t_ops;
 	struct virtio_vsock_sock *vvs;
 	u32 pkt_len = info->pkt_len;
+	bool can_zcopy = false;
 	u32 rest_len;
 	int ret;
 
@@ -254,15 +369,30 @@  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
 	if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW)
 		return pkt_len;
 
+	if (info->msg) {
+		/* If zerocopy is not enabled by 'setsockopt()', we behave as
+		 * there is no MSG_ZEROCOPY flag set.
+		 */
+		if (!sock_flag(sk_vsock(vsk), SOCK_ZEROCOPY))
+			info->msg->msg_flags &= ~MSG_ZEROCOPY;
+
+		if (info->msg->msg_flags & MSG_ZEROCOPY)
+			can_zcopy = virtio_transport_can_zcopy(info, pkt_len);
+
+		if (can_zcopy)
+			max_skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE,
+					    (MAX_SKB_FRAGS * PAGE_SIZE));
+	}
+
 	rest_len = pkt_len;
 
 	do {
 		struct sk_buff *skb;
 		size_t skb_len;
 
-		skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len);
+		skb_len = min(max_skb_len, rest_len);
 
-		skb = virtio_transport_alloc_skb(info, skb_len,
+		skb = virtio_transport_alloc_skb(vsk, info, skb_len, can_zcopy,
 						 src_cid, src_port,
 						 dst_cid, dst_port);
 		if (!skb) {
@@ -270,6 +400,17 @@  static int virtio_transport_send_pkt_info(struct vsock_sock *vsk,
 			break;
 		}
 
+		/* This is last skb to send this portion of data. */
+		if (info->msg && info->msg->msg_flags & MSG_ZEROCOPY &&
+		    skb_len == rest_len && info->op == VIRTIO_VSOCK_OP_RW) {
+			if (virtio_transport_init_zcopy_skb(vsk, skb,
+							    info->msg,
+							    can_zcopy)) {
+				ret = -ENOMEM;
+				break;
+			}
+		}
+
 		virtio_transport_inc_tx_pkt(vvs, skb);
 
 		ret = t_ops->send_pkt(skb);
@@ -934,7 +1075,7 @@  static int virtio_transport_reset_no_sock(const struct virtio_transport *t,
 	if (!t)
 		return -ENOTCONN;
 
-	reply = virtio_transport_alloc_skb(&info, 0,
+	reply = virtio_transport_alloc_skb(NULL, &info, 0, false,
 					   le64_to_cpu(hdr->dst_cid),
 					   le32_to_cpu(hdr->dst_port),
 					   le64_to_cpu(hdr->src_cid),