[v4,01/21] net/tcp: Prepare tcp_md5sig_pool for TCP-AO

Message ID 20230215183335.800122-2-dima@arista.com
State New
Headers
Series net/tcp: Add TCP-AO support |

Commit Message

Dmitry Safonov Feb. 15, 2023, 6:33 p.m. UTC
  TCP-AO similarly to TCP-MD5 needs to allocate tfms on a slow-path, which
is setsockopt() and use crypto ahash requests on fast paths, which are
RX/TX softirqs. It as well needs a temporary/scratch buffer for
preparing the hashing request.

Extend tcp_md5sig_pool to support other hashing algorithms than MD5.
Move it in a separate file.

This patch was previously submitted as more generic crypto_pool [1],
but Herbert nacked making it generic crypto API. His view is that crypto
requests should be atomically allocated on fast-paths. So, in this
version I don't move this pool anywhere outside TCP, only extending it
for TCP-AO use-case. It can be converted once there will be per-request
hashing crypto keys.

[1]: https://lore.kernel.org/all/20230118214111.394416-1-dima@arista.com/T/#u
Signed-off-by: Dmitry Safonov <dima@arista.com>
---
 include/net/tcp.h        |  48 ++++--
 net/ipv4/Kconfig         |   4 +
 net/ipv4/Makefile        |   1 +
 net/ipv4/tcp.c           | 103 +++---------
 net/ipv4/tcp_ipv4.c      |  97 +++++++-----
 net/ipv4/tcp_minisocks.c |  21 ++-
 net/ipv4/tcp_sigpool.c   | 333 +++++++++++++++++++++++++++++++++++++++
 net/ipv6/tcp_ipv6.c      |  58 +++----
 8 files changed, 493 insertions(+), 172 deletions(-)
 create mode 100644 net/ipv4/tcp_sigpool.c
  

Comments

Herbert Xu Feb. 20, 2023, 9:41 a.m. UTC | #1
On Wed, Feb 15, 2023 at 06:33:15PM +0000, Dmitry Safonov wrote:
> TCP-AO similarly to TCP-MD5 needs to allocate tfms on a slow-path, which
> is setsockopt() and use crypto ahash requests on fast paths, which are
> RX/TX softirqs. It as well needs a temporary/scratch buffer for
> preparing the hashing request.
> 
> Extend tcp_md5sig_pool to support other hashing algorithms than MD5.
> Move it in a separate file.
> 
> This patch was previously submitted as more generic crypto_pool [1],
> but Herbert nacked making it generic crypto API. His view is that crypto
> requests should be atomically allocated on fast-paths. So, in this
> version I don't move this pool anywhere outside TCP, only extending it
> for TCP-AO use-case. It can be converted once there will be per-request
> hashing crypto keys.
> 
> [1]: https://lore.kernel.org/all/20230118214111.394416-1-dima@arista.com/T/#u
> Signed-off-by: Dmitry Safonov <dima@arista.com>
> ---
>  include/net/tcp.h        |  48 ++++--
>  net/ipv4/Kconfig         |   4 +
>  net/ipv4/Makefile        |   1 +
>  net/ipv4/tcp.c           | 103 +++---------
>  net/ipv4/tcp_ipv4.c      |  97 +++++++-----
>  net/ipv4/tcp_minisocks.c |  21 ++-
>  net/ipv4/tcp_sigpool.c   | 333 +++++++++++++++++++++++++++++++++++++++
>  net/ipv6/tcp_ipv6.c      |  58 +++----
>  8 files changed, 493 insertions(+), 172 deletions(-)
>  create mode 100644 net/ipv4/tcp_sigpool.c

Please wait for my per-request hash work before you resubmit this.
Once that's in place all you need is a single tfm for the whole
system.

As to request pools what exactly is the point of that? Just kmalloc
them on demand.

Cheers,
  
Dmitry Safonov Feb. 20, 2023, 4:57 p.m. UTC | #2
Hi Herbert,

On 2/20/23 09:41, Herbert Xu wrote:
> On Wed, Feb 15, 2023 at 06:33:15PM +0000, Dmitry Safonov wrote:
>> TCP-AO similarly to TCP-MD5 needs to allocate tfms on a slow-path, which
>> is setsockopt() and use crypto ahash requests on fast paths, which are
>> RX/TX softirqs. It as well needs a temporary/scratch buffer for
>> preparing the hashing request.
>>
>> Extend tcp_md5sig_pool to support other hashing algorithms than MD5.
>> Move it in a separate file.
>>
>> This patch was previously submitted as more generic crypto_pool [1],
>> but Herbert nacked making it generic crypto API. His view is that crypto
>> requests should be atomically allocated on fast-paths. So, in this
>> version I don't move this pool anywhere outside TCP, only extending it
>> for TCP-AO use-case. It can be converted once there will be per-request
>> hashing crypto keys.
>>
>> [1]: https://lore.kernel.org/all/20230118214111.394416-1-dima@arista.com/T/#u
>> Signed-off-by: Dmitry Safonov <dima@arista.com>
>> ---
>>  include/net/tcp.h        |  48 ++++--
>>  net/ipv4/Kconfig         |   4 +
>>  net/ipv4/Makefile        |   1 +
>>  net/ipv4/tcp.c           | 103 +++---------
>>  net/ipv4/tcp_ipv4.c      |  97 +++++++-----
>>  net/ipv4/tcp_minisocks.c |  21 ++-
>>  net/ipv4/tcp_sigpool.c   | 333 +++++++++++++++++++++++++++++++++++++++
>>  net/ipv6/tcp_ipv6.c      |  58 +++----
>>  8 files changed, 493 insertions(+), 172 deletions(-)
>>  create mode 100644 net/ipv4/tcp_sigpool.c
> 
> Please wait for my per-request hash work before you resubmit this.

Do you have a timeline for that work?
And if you don't mind I keep re-iterating, as I'm trying to address TCP
reviews and missed functionality/selftests.

> Once that's in place all you need is a single tfm for the whole
> system.

Unfortunately, not really: RFC5926 prescribes the mandatory-to-implement
MAC algorithms for TCP-AO: HMAC-SHA-1-96 and AES-128-CMAC-96. But since
the RFC was written sha1 is now more eligible for attacks as well as
RFC5925 has:
> The option should support algorithms other than the default, to
> allow agility over time.
> TCP-AO allows any desired algorithm, subject to TCP option
> space limitations, as noted in Section 2.2. The use of a set
> of MKTs allows separate connections to use different
> algorithms, both for the MAC and the KDF.

As well as from a customer's request we need to support more than two
required algorithms. So, this implementation let the user choose the
algorithm that is supported by crypto/ layer (more or less like xfrm does).

Which means, that it still has to support multiple tfms. I guess that
pool of tfms can be converted to use per-request keys quite easily.

> As to request pools what exactly is the point of that? Just kmalloc
> them on demand.

1) before your per-request key patches - it's not possible.
2) after your patches - my question would be: "is it better to
kmalloc(GFP_ATOMIC) in RX/TX for every signed TCP segment, rather than
pre-allocate it?"

The price of (2) may just well be negligible, but worth measuring before
switching.

Thanks,
          Dmitry
  
Herbert Xu Feb. 21, 2023, 2:43 a.m. UTC | #3
On Mon, Feb 20, 2023 at 04:57:20PM +0000, Dmitry Safonov wrote:
. 
> Do you have a timeline for that work?
> And if you don't mind I keep re-iterating, as I'm trying to address TCP
> reviews and missed functionality/selftests.

I'm hoping to get it ready for the next merge window.
 
> 1) before your per-request key patches - it's not possible.
> 2) after your patches - my question would be: "is it better to
> kmalloc(GFP_ATOMIC) in RX/TX for every signed TCP segment, rather than
> pre-allocate it?"
> 
> The price of (2) may just well be negligible, but worth measuring before
> switching.

Please keep in mind that you're already performing crypto which
is usually a lot slower than a kmalloc.  In any case, if there is
any optimisation to be done to make the kmalloc faster by using
pools, then that optimisation should go into mm.

Cheers,
  
Dmitry Safonov Feb. 21, 2023, 2:52 p.m. UTC | #4
On 2/21/23 02:43, Herbert Xu wrote:
> On Mon, Feb 20, 2023 at 04:57:20PM +0000, Dmitry Safonov wrote:
> . 
>> Do you have a timeline for that work?
>> And if you don't mind I keep re-iterating, as I'm trying to address TCP
>> reviews and missed functionality/selftests.
> 
> I'm hoping to get it ready for the next merge window.

Nice! I'll mark this 1/21 patch as [draft], mentioning your work as it
will need to be re-made using per-request keys.
Still, I will keep iterating TCP-AO patches set during 6.3 RCs in order
to get more reviews/suggestions related to TCP changes.

>> 1) before your per-request key patches - it's not possible.
>> 2) after your patches - my question would be: "is it better to
>> kmalloc(GFP_ATOMIC) in RX/TX for every signed TCP segment, rather than
>> pre-allocate it?"
>>
>> The price of (2) may just well be negligible, but worth measuring before
>> switching.
> 
> Please keep in mind that you're already performing crypto which
> is usually a lot slower than a kmalloc.  In any case, if there is
> any optimisation to be done to make the kmalloc faster by using
> pools, then that optimisation should go into mm.

Fair point. Probably, kmalloc() is negligible. I'll measure as I have a
patch for iperf for TCP-MD5/TCP-AO measurements.

Thanks,
          Dmitry
  

Patch

diff --git a/include/net/tcp.h b/include/net/tcp.h
index db9f828e9d1e..e77080003800 100644
--- a/include/net/tcp.h
+++ b/include/net/tcp.h
@@ -1664,12 +1664,35 @@  union tcp_md5sum_block {
 #endif
 };
 
-/* - pool: digest algorithm, hash description and scratch buffer */
-struct tcp_md5sig_pool {
-	struct ahash_request	*md5_req;
-	void			*scratch;
+/*
+ * struct tcp_sigpool - per-CPU pool of ahash_requests
+ * @scratch: per-CPU temporary area, that can be used between
+ *	     tcp_sigpool_start() and tcp_sigpool_end() to perform
+ *	     crypto request
+ * @req: pre-allocated ahash request
+ */
+struct tcp_sigpool {
+	void *scratch;
+	struct ahash_request *req;
 };
+int tcp_sigpool_alloc_ahash(const char *alg, size_t scratch_size);
+void tcp_sigpool_get(unsigned int id);
+void tcp_sigpool_release(unsigned int id);
 
+/**
+ * tcp_sigpool_start - disable bh and start using tcp_sigpool_ahash
+ * @id: tcp_sigpool that was previously allocated by tcp_sigpool_alloc_ahash()
+ * @c: returned tcp_sigpool for usage (uninitialized on failure)
+ */
+int tcp_sigpool_start(unsigned int id, struct tcp_sigpool *c);
+/**
+ * tcp_sigpool_end - enable bh and stop using tcp_sigpool
+ */
+static inline void tcp_sigpool_end(void)
+{
+	rcu_read_unlock_bh();
+}
+size_t tcp_sigpool_algo(unsigned int id, char *buf, size_t buf_len);
 /* - functions */
 int tcp_v4_md5_hash_skb(char *md5_hash, const struct tcp_md5sig_key *key,
 			const struct sock *sk, const struct sk_buff *skb);
@@ -1725,17 +1748,14 @@  tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
 #define tcp_twsk_md5_key(twsk)	NULL
 #endif
 
-bool tcp_alloc_md5sig_pool(void);
-
-struct tcp_md5sig_pool *tcp_get_md5sig_pool(void);
-static inline void tcp_put_md5sig_pool(void)
-{
-	local_bh_enable();
-}
+int tcp_md5_alloc_sigpool(void);
+void tcp_md5_release_sigpool(void);
+void tcp_md5_add_sigpool(void);
+extern int tcp_md5_sigpool_id;
 
-int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *, const struct sk_buff *,
-			  unsigned int header_len);
-int tcp_md5_hash_key(struct tcp_md5sig_pool *hp,
+int tcp_md5_hash_skb_data(struct tcp_sigpool *hp,
+			  const struct sk_buff *skb, unsigned int header_len);
+int tcp_md5_hash_key(struct tcp_sigpool *hp,
 		     const struct tcp_md5sig_key *key);
 
 /* From tcp_fastopen.c */
diff --git a/net/ipv4/Kconfig b/net/ipv4/Kconfig
index 2dfb12230f08..89e2ab023272 100644
--- a/net/ipv4/Kconfig
+++ b/net/ipv4/Kconfig
@@ -741,10 +741,14 @@  config DEFAULT_TCP_CONG
 	default "bbr" if DEFAULT_BBR
 	default "cubic"
 
+config TCP_SIGPOOL
+	tristate
+
 config TCP_MD5SIG
 	bool "TCP: MD5 Signature Option support (RFC2385)"
 	select CRYPTO
 	select CRYPTO_MD5
+	select TCP_SIGPOOL
 	help
 	  RFC2385 specifies a method of giving MD5 protection to TCP sessions.
 	  Its main (only?) use is to protect BGP sessions between core routers
diff --git a/net/ipv4/Makefile b/net/ipv4/Makefile
index af7d2cf490fb..de8ba8cdc056 100644
--- a/net/ipv4/Makefile
+++ b/net/ipv4/Makefile
@@ -61,6 +61,7 @@  obj-$(CONFIG_TCP_CONG_SCALABLE) += tcp_scalable.o
 obj-$(CONFIG_TCP_CONG_LP) += tcp_lp.o
 obj-$(CONFIG_TCP_CONG_YEAH) += tcp_yeah.o
 obj-$(CONFIG_TCP_CONG_ILLINOIS) += tcp_illinois.o
+obj-$(CONFIG_TCP_SIGPOOL) += tcp_sigpool.o
 obj-$(CONFIG_NET_SOCK_MSG) += tcp_bpf.o
 obj-$(CONFIG_BPF_SYSCALL) += udp_bpf.o
 obj-$(CONFIG_NETLABEL) += cipso_ipv4.o
diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 33f559f491c8..5522c5ac24bc 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -4413,98 +4413,42 @@  int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
 EXPORT_SYMBOL(tcp_getsockopt);
 
 #ifdef CONFIG_TCP_MD5SIG
-static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
-static DEFINE_MUTEX(tcp_md5sig_mutex);
-static bool tcp_md5sig_pool_populated = false;
+int tcp_md5_sigpool_id = -1;
+EXPORT_SYMBOL(tcp_md5_sigpool_id);
 
-static void __tcp_alloc_md5sig_pool(void)
+int tcp_md5_alloc_sigpool(void)
 {
-	struct crypto_ahash *hash;
-	int cpu;
-
-	hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
-	if (IS_ERR(hash))
-		return;
-
-	for_each_possible_cpu(cpu) {
-		void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
-		struct ahash_request *req;
-
-		if (!scratch) {
-			scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
-					       sizeof(struct tcphdr),
-					       GFP_KERNEL,
-					       cpu_to_node(cpu));
-			if (!scratch)
-				return;
-			per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
-		}
-		if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
-			continue;
-
-		req = ahash_request_alloc(hash, GFP_KERNEL);
-		if (!req)
-			return;
-
-		ahash_request_set_callback(req, 0, NULL, NULL);
+	size_t scratch_size;
+	int ret;
 
-		per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
+	scratch_size = sizeof(union tcp_md5sum_block) + sizeof(struct tcphdr);
+	ret = tcp_sigpool_alloc_ahash("md5", scratch_size);
+	if (ret >= 0) {
+		tcp_md5_sigpool_id = ret;
+		return 0;
 	}
-	/* before setting tcp_md5sig_pool_populated, we must commit all writes
-	 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
-	 */
-	smp_wmb();
-	/* Paired with READ_ONCE() from tcp_alloc_md5sig_pool()
-	 * and tcp_get_md5sig_pool().
-	*/
-	WRITE_ONCE(tcp_md5sig_pool_populated, true);
+	return ret;
 }
+EXPORT_SYMBOL(tcp_md5_alloc_sigpool);
 
-bool tcp_alloc_md5sig_pool(void)
+void tcp_md5_release_sigpool(void)
 {
-	/* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
-	if (unlikely(!READ_ONCE(tcp_md5sig_pool_populated))) {
-		mutex_lock(&tcp_md5sig_mutex);
-
-		if (!tcp_md5sig_pool_populated)
-			__tcp_alloc_md5sig_pool();
-
-		mutex_unlock(&tcp_md5sig_mutex);
-	}
-	/* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
-	return READ_ONCE(tcp_md5sig_pool_populated);
+	tcp_sigpool_release(tcp_md5_sigpool_id);
 }
-EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
+EXPORT_SYMBOL(tcp_md5_release_sigpool);
 
-
-/**
- *	tcp_get_md5sig_pool - get md5sig_pool for this user
- *
- *	We use percpu structure, so if we succeed, we exit with preemption
- *	and BH disabled, to make sure another thread or softirq handling
- *	wont try to get same context.
- */
-struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
+void tcp_md5_add_sigpool(void)
 {
-	local_bh_disable();
-
-	/* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
-	if (READ_ONCE(tcp_md5sig_pool_populated)) {
-		/* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
-		smp_rmb();
-		return this_cpu_ptr(&tcp_md5sig_pool);
-	}
-	local_bh_enable();
-	return NULL;
+	tcp_sigpool_get(tcp_md5_sigpool_id);
 }
-EXPORT_SYMBOL(tcp_get_md5sig_pool);
+EXPORT_SYMBOL(tcp_md5_add_sigpool);
 
-int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
+int tcp_md5_hash_skb_data(struct tcp_sigpool *hp,
 			  const struct sk_buff *skb, unsigned int header_len)
 {
 	struct scatterlist sg;
 	const struct tcphdr *tp = tcp_hdr(skb);
-	struct ahash_request *req = hp->md5_req;
+	struct ahash_request *req = hp->req;
 	unsigned int i;
 	const unsigned int head_data_len = skb_headlen(skb) > header_len ?
 					   skb_headlen(skb) - header_len : 0;
@@ -4538,16 +4482,17 @@  int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
 }
 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
 
-int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
+int tcp_md5_hash_key(struct tcp_sigpool *hp,
+		     const struct tcp_md5sig_key *key)
 {
 	u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
 	struct scatterlist sg;
 
 	sg_init_one(&sg, key->key, keylen);
-	ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
+	ahash_request_set_crypt(hp->req, &sg, NULL, keylen);
 
 	/* We use data_race() because tcp_md5_do_add() might change key->key under us */
-	return data_race(crypto_ahash_update(hp->md5_req));
+	return data_race(crypto_ahash_update(hp->req));
 }
 EXPORT_SYMBOL(tcp_md5_hash_key);
 
diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
index 8320d0ecb13a..6701af5922cb 100644
--- a/net/ipv4/tcp_ipv4.c
+++ b/net/ipv4/tcp_ipv4.c
@@ -1212,10 +1212,6 @@  static int __tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
 	key = sock_kmalloc(sk, sizeof(*key), gfp | __GFP_ZERO);
 	if (!key)
 		return -ENOMEM;
-	if (!tcp_alloc_md5sig_pool()) {
-		sock_kfree_s(sk, key, sizeof(*key));
-		return -ENOMEM;
-	}
 
 	memcpy(key->key, newkey, newkeylen);
 	key->keylen = newkeylen;
@@ -1237,8 +1233,13 @@  int tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
 	struct tcp_sock *tp = tcp_sk(sk);
 
 	if (!rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk))) {
-		if (tcp_md5sig_info_add(sk, GFP_KERNEL))
+		if (tcp_md5_alloc_sigpool())
+			return -ENOMEM;
+
+		if (tcp_md5sig_info_add(sk, GFP_KERNEL)) {
+			tcp_md5_release_sigpool();
 			return -ENOMEM;
+		}
 
 		if (!static_branch_inc(&tcp_md5_needed.key)) {
 			struct tcp_md5sig_info *md5sig;
@@ -1246,6 +1247,7 @@  int tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
 			md5sig = rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk));
 			rcu_assign_pointer(tp->md5sig_info, NULL);
 			kfree_rcu(md5sig, rcu);
+			tcp_md5_release_sigpool();
 			return -EUSERS;
 		}
 	}
@@ -1262,8 +1264,12 @@  int tcp_md5_key_copy(struct sock *sk, const union tcp_md5_addr *addr,
 	struct tcp_sock *tp = tcp_sk(sk);
 
 	if (!rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk))) {
-		if (tcp_md5sig_info_add(sk, sk_gfp_mask(sk, GFP_ATOMIC)))
+		tcp_md5_add_sigpool();
+
+		if (tcp_md5sig_info_add(sk, sk_gfp_mask(sk, GFP_ATOMIC))) {
+			tcp_md5_release_sigpool();
 			return -ENOMEM;
+		}
 
 		if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key)) {
 			struct tcp_md5sig_info *md5sig;
@@ -1272,6 +1278,7 @@  int tcp_md5_key_copy(struct sock *sk, const union tcp_md5_addr *addr,
 			net_warn_ratelimited("Too many TCP-MD5 keys in the system\n");
 			rcu_assign_pointer(tp->md5sig_info, NULL);
 			kfree_rcu(md5sig, rcu);
+			tcp_md5_release_sigpool();
 			return -EUSERS;
 		}
 	}
@@ -1371,7 +1378,7 @@  static int tcp_v4_parse_md5_keys(struct sock *sk, int optname,
 			      cmd.tcpm_key, cmd.tcpm_keylen);
 }
 
-static int tcp_v4_md5_hash_headers(struct tcp_md5sig_pool *hp,
+static int tcp_v4_md5_hash_headers(struct tcp_sigpool *hp,
 				   __be32 daddr, __be32 saddr,
 				   const struct tcphdr *th, int nbytes)
 {
@@ -1391,38 +1398,35 @@  static int tcp_v4_md5_hash_headers(struct tcp_md5sig_pool *hp,
 	_th->check = 0;
 
 	sg_init_one(&sg, bp, sizeof(*bp) + sizeof(*th));
-	ahash_request_set_crypt(hp->md5_req, &sg, NULL,
+	ahash_request_set_crypt(hp->req, &sg, NULL,
 				sizeof(*bp) + sizeof(*th));
-	return crypto_ahash_update(hp->md5_req);
+	return crypto_ahash_update(hp->req);
 }
 
 static int tcp_v4_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
 			       __be32 daddr, __be32 saddr, const struct tcphdr *th)
 {
-	struct tcp_md5sig_pool *hp;
-	struct ahash_request *req;
+	struct tcp_sigpool hp;
 
-	hp = tcp_get_md5sig_pool();
-	if (!hp)
-		goto clear_hash_noput;
-	req = hp->md5_req;
+	if (tcp_sigpool_start(tcp_md5_sigpool_id, &hp))
+		goto clear_hash_nostart;
 
-	if (crypto_ahash_init(req))
+	if (crypto_ahash_init(hp.req))
 		goto clear_hash;
-	if (tcp_v4_md5_hash_headers(hp, daddr, saddr, th, th->doff << 2))
+	if (tcp_v4_md5_hash_headers(&hp, daddr, saddr, th, th->doff << 2))
 		goto clear_hash;
-	if (tcp_md5_hash_key(hp, key))
+	if (tcp_md5_hash_key(&hp, key))
 		goto clear_hash;
-	ahash_request_set_crypt(req, NULL, md5_hash, 0);
-	if (crypto_ahash_final(req))
+	ahash_request_set_crypt(hp.req, NULL, md5_hash, 0);
+	if (crypto_ahash_final(hp.req))
 		goto clear_hash;
 
-	tcp_put_md5sig_pool();
+	tcp_sigpool_end();
 	return 0;
 
 clear_hash:
-	tcp_put_md5sig_pool();
-clear_hash_noput:
+	tcp_sigpool_end();
+clear_hash_nostart:
 	memset(md5_hash, 0, 16);
 	return 1;
 }
@@ -1431,8 +1435,7 @@  int tcp_v4_md5_hash_skb(char *md5_hash, const struct tcp_md5sig_key *key,
 			const struct sock *sk,
 			const struct sk_buff *skb)
 {
-	struct tcp_md5sig_pool *hp;
-	struct ahash_request *req;
+	struct tcp_sigpool hp;
 	const struct tcphdr *th = tcp_hdr(skb);
 	__be32 saddr, daddr;
 
@@ -1445,30 +1448,28 @@  int tcp_v4_md5_hash_skb(char *md5_hash, const struct tcp_md5sig_key *key,
 		daddr = iph->daddr;
 	}
 
-	hp = tcp_get_md5sig_pool();
-	if (!hp)
-		goto clear_hash_noput;
-	req = hp->md5_req;
+	if (tcp_sigpool_start(tcp_md5_sigpool_id, &hp))
+		goto clear_hash_nostart;
 
-	if (crypto_ahash_init(req))
+	if (crypto_ahash_init(hp.req))
 		goto clear_hash;
 
-	if (tcp_v4_md5_hash_headers(hp, daddr, saddr, th, skb->len))
+	if (tcp_v4_md5_hash_headers(&hp, daddr, saddr, th, skb->len))
 		goto clear_hash;
-	if (tcp_md5_hash_skb_data(hp, skb, th->doff << 2))
+	if (tcp_md5_hash_skb_data(&hp, skb, th->doff << 2))
 		goto clear_hash;
-	if (tcp_md5_hash_key(hp, key))
+	if (tcp_md5_hash_key(&hp, key))
 		goto clear_hash;
-	ahash_request_set_crypt(req, NULL, md5_hash, 0);
-	if (crypto_ahash_final(req))
+	ahash_request_set_crypt(hp.req, NULL, md5_hash, 0);
+	if (crypto_ahash_final(hp.req))
 		goto clear_hash;
 
-	tcp_put_md5sig_pool();
+	tcp_sigpool_end();
 	return 0;
 
 clear_hash:
-	tcp_put_md5sig_pool();
-clear_hash_noput:
+	tcp_sigpool_end();
+clear_hash_nostart:
 	memset(md5_hash, 0, 16);
 	return 1;
 }
@@ -2285,6 +2286,18 @@  static int tcp_v4_init_sock(struct sock *sk)
 	return 0;
 }
 
+#ifdef CONFIG_TCP_MD5SIG
+static void tcp_md5sig_info_free_rcu(struct rcu_head *head)
+{
+	struct tcp_md5sig_info *md5sig;
+
+	md5sig = container_of(head, struct tcp_md5sig_info, rcu);
+	kfree(md5sig);
+	static_branch_slow_dec_deferred(&tcp_md5_needed);
+	tcp_md5_release_sigpool();
+}
+#endif
+
 void tcp_v4_destroy_sock(struct sock *sk)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
@@ -2309,10 +2322,12 @@  void tcp_v4_destroy_sock(struct sock *sk)
 #ifdef CONFIG_TCP_MD5SIG
 	/* Clean up the MD5 key list, if any */
 	if (tp->md5sig_info) {
+		struct tcp_md5sig_info *md5sig;
+
+		md5sig = rcu_dereference_protected(tp->md5sig_info, 1);
 		tcp_clear_md5_list(sk);
-		kfree_rcu(rcu_dereference_protected(tp->md5sig_info, 1), rcu);
-		tp->md5sig_info = NULL;
-		static_branch_slow_dec_deferred(&tcp_md5_needed);
+		call_rcu(&md5sig->rcu, tcp_md5sig_info_free_rcu);
+		rcu_assign_pointer(tp->md5sig_info, NULL);
 	}
 #endif
 
diff --git a/net/ipv4/tcp_minisocks.c b/net/ipv4/tcp_minisocks.c
index e002f2e1d4f2..0219c0e5e2df 100644
--- a/net/ipv4/tcp_minisocks.c
+++ b/net/ipv4/tcp_minisocks.c
@@ -261,10 +261,9 @@  static void tcp_time_wait_init(struct sock *sk, struct tcp_timewait_sock *tcptw)
 		tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
 		if (!tcptw->tw_md5_key)
 			return;
-		if (!tcp_alloc_md5sig_pool())
-			goto out_free;
 		if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key))
 			goto out_free;
+		tcp_md5_add_sigpool();
 	}
 	return;
 out_free:
@@ -349,16 +348,26 @@  void tcp_time_wait(struct sock *sk, int state, int timeo)
 }
 EXPORT_SYMBOL(tcp_time_wait);
 
+#ifdef CONFIG_TCP_MD5SIG
+static void tcp_md5_twsk_free_rcu(struct rcu_head *head)
+{
+	struct tcp_md5sig_key *key;
+
+	key = container_of(head, struct tcp_md5sig_key, rcu);
+	kfree(key);
+	static_branch_slow_dec_deferred(&tcp_md5_needed);
+	tcp_md5_release_sigpool();
+}
+#endif
+
 void tcp_twsk_destructor(struct sock *sk)
 {
 #ifdef CONFIG_TCP_MD5SIG
 	if (static_branch_unlikely(&tcp_md5_needed.key)) {
 		struct tcp_timewait_sock *twsk = tcp_twsk(sk);
 
-		if (twsk->tw_md5_key) {
-			kfree_rcu(twsk->tw_md5_key, rcu);
-			static_branch_slow_dec_deferred(&tcp_md5_needed);
-		}
+		if (twsk->tw_md5_key)
+			call_rcu(&twsk->tw_md5_key->rcu, tcp_md5_twsk_free_rcu);
 	}
 #endif
 }
diff --git a/net/ipv4/tcp_sigpool.c b/net/ipv4/tcp_sigpool.c
new file mode 100644
index 000000000000..2fbd6795ced6
--- /dev/null
+++ b/net/ipv4/tcp_sigpool.c
@@ -0,0 +1,333 @@ 
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include <crypto/hash.h>
+#include <linux/cpu.h>
+#include <linux/kref.h>
+#include <linux/module.h>
+#include <linux/mutex.h>
+#include <linux/percpu.h>
+#include <linux/workqueue.h>
+#include <net/tcp.h>
+
+static unsigned long __scratch_size;
+static DEFINE_PER_CPU(void __rcu *, sigpool_scratch);
+
+struct sigpool_entry {
+	struct ahash_request * __percpu *req;
+	const char			*alg;
+	struct kref			kref;
+	bool				needs_key;
+};
+
+#define CPOOL_SIZE (PAGE_SIZE / sizeof(struct sigpool_entry))
+static struct sigpool_entry cpool[CPOOL_SIZE];
+static unsigned int cpool_populated;
+static DEFINE_MUTEX(cpool_mutex);
+
+/* Slow-path */
+struct scratches_to_free {
+	struct rcu_head rcu;
+	unsigned int cnt;
+	void *scratches[];
+};
+
+static void free_old_scratches(struct rcu_head *head)
+{
+	struct scratches_to_free *stf;
+
+	stf = container_of(head, struct scratches_to_free, rcu);
+	while (stf->cnt--)
+		kfree(stf->scratches[stf->cnt]);
+	kfree(stf);
+}
+
+/*
+ * sigpool_reserve_scratch - re-allocates scratch buffer, slow-path
+ * @size: request size for the scratch/temp buffer
+ */
+static int sigpool_reserve_scratch(size_t size)
+{
+	struct scratches_to_free *stf;
+	size_t stf_sz = struct_size(stf, scratches, num_possible_cpus());
+	int cpu, err = 0;
+
+	lockdep_assert_held(&cpool_mutex);
+	if (__scratch_size >= size)
+		return 0;
+
+	stf = kmalloc(stf_sz, GFP_KERNEL);
+	if (!stf)
+		return -ENOMEM;
+	stf->cnt = 0;
+
+	size = max(size, __scratch_size);
+	cpus_read_lock();
+	for_each_possible_cpu(cpu) {
+		void *scratch, *old_scratch;
+
+		scratch = kmalloc_node(size, GFP_KERNEL, cpu_to_node(cpu));
+		if (!scratch) {
+			err = -ENOMEM;
+			break;
+		}
+
+		old_scratch = rcu_replace_pointer(per_cpu(sigpool_scratch, cpu), scratch, lockdep_is_held(&cpool_mutex));
+		if (!cpu_online(cpu) || !old_scratch) {
+			kfree(old_scratch);
+			continue;
+		}
+		stf->scratches[stf->cnt++] = old_scratch;
+	}
+	cpus_read_unlock();
+	if (!err)
+		__scratch_size = size;
+
+	call_rcu(&stf->rcu, free_old_scratches);
+	return err;
+}
+
+static void sigpool_scratch_free(void)
+{
+	int cpu;
+
+	for_each_possible_cpu(cpu)
+		kfree(rcu_replace_pointer(per_cpu(sigpool_scratch, cpu),
+					  NULL, lockdep_is_held(&cpool_mutex)));
+	__scratch_size = 0;
+}
+
+static int __cpool_alloc_ahash(struct sigpool_entry *e, const char *alg)
+{
+	struct crypto_ahash *hash, *cpu0_hash;
+	int cpu, ret = -ENOMEM;
+
+	e->alg = kstrdup(alg, GFP_KERNEL);
+	if (!e->alg)
+		return -ENOMEM;
+
+	e->req = alloc_percpu(struct ahash_request *);
+	if (!e->req)
+		goto out_free_alg;
+
+	cpu0_hash = crypto_alloc_ahash(alg, 0, CRYPTO_ALG_ASYNC);
+	if (IS_ERR(cpu0_hash)) {
+		ret = PTR_ERR(cpu0_hash);
+		goto out_free_req;
+	}
+
+	/* If hash has .setkey(), allocate ahash per-CPU, not only request */
+	e->needs_key = crypto_ahash_get_flags(cpu0_hash) & CRYPTO_TFM_NEED_KEY;
+
+	hash = cpu0_hash;
+	for_each_possible_cpu(cpu) {
+		struct ahash_request *req;
+
+		/* If ahash has a key - it has to be allocated per-CPU.
+		 * In such case re-use for CPU0 hash that just have been
+		 * allocated above.
+		 */
+		if (!hash)
+			hash = crypto_alloc_ahash(alg, 0, CRYPTO_ALG_ASYNC);
+		if (IS_ERR(hash))
+			goto out_free_per_cpu;
+
+		req = ahash_request_alloc(hash, GFP_KERNEL);
+		if (!req)
+			goto out_free_hash;
+
+		ahash_request_set_callback(req, 0, NULL, NULL);
+
+		*per_cpu_ptr(e->req, cpu) = req;
+
+		if (e->needs_key)
+			hash = NULL;
+	}
+	kref_init(&e->kref);
+	return 0;
+
+out_free_hash:
+	if (hash != cpu0_hash)
+		crypto_free_ahash(hash);
+
+out_free_per_cpu:
+	for_each_possible_cpu(cpu) {
+		struct ahash_request *req = *per_cpu_ptr(e->req, cpu);
+		struct crypto_ahash *pcpu_hash;
+
+		if (!req)
+			break;
+		pcpu_hash = crypto_ahash_reqtfm(req);
+		ahash_request_free(req);
+		/* hash per-CPU, e->needs_key == true */
+		if (pcpu_hash != cpu0_hash)
+			crypto_free_ahash(pcpu_hash);
+	}
+
+	crypto_free_ahash(cpu0_hash);
+out_free_req:
+	free_percpu(e->req);
+out_free_alg:
+	kfree(e->alg);
+	e->alg = NULL;
+	return ret;
+}
+
+/**
+ * tcp_sigpool_alloc_ahash - allocates pool for ahash requests
+ * @alg: name of async hash algorithm
+ * @scratch_size: reserve a tcp_sigpool::scratch buffer of this size
+ */
+int tcp_sigpool_alloc_ahash(const char *alg, size_t scratch_size)
+{
+	int i, ret;
+
+	/* slow-path */
+	mutex_lock(&cpool_mutex);
+	ret = sigpool_reserve_scratch(scratch_size);
+	if (ret)
+		goto out;
+	for (i = 0; i < cpool_populated; i++) {
+		if (cpool[i].alg && !strcmp(cpool[i].alg, alg)) {
+			if (kref_read(&cpool[i].kref) > 0)
+				kref_get(&cpool[i].kref);
+			else
+				kref_init(&cpool[i].kref);
+			ret = i;
+			goto out;
+		}
+	}
+
+	for (i = 0; i < cpool_populated; i++) {
+		if (!cpool[i].alg)
+			break;
+	}
+	if (i >= CPOOL_SIZE) {
+		ret = -ENOSPC;
+		goto out;
+	}
+
+	ret = __cpool_alloc_ahash(&cpool[i], alg);
+	if (!ret) {
+		ret = i;
+		if (i == cpool_populated)
+			cpool_populated++;
+	}
+out:
+	mutex_unlock(&cpool_mutex);
+	return ret;
+}
+EXPORT_SYMBOL_GPL(tcp_sigpool_alloc_ahash);
+
+static void __cpool_free_entry(struct sigpool_entry *e)
+{
+	struct crypto_ahash *hash = NULL;
+	int cpu;
+
+	for_each_possible_cpu(cpu) {
+		if (*per_cpu_ptr(e->req, cpu) == NULL)
+			continue;
+
+		hash = crypto_ahash_reqtfm(*per_cpu_ptr(e->req, cpu));
+		ahash_request_free(*per_cpu_ptr(e->req, cpu));
+		if (e->needs_key) {
+			crypto_free_ahash(hash);
+			hash = NULL;
+		}
+	}
+	if (hash)
+		crypto_free_ahash(hash);
+	free_percpu(e->req);
+	kfree(e->alg);
+	memset(e, 0, sizeof(*e));
+}
+
+static void cpool_cleanup_work_cb(struct work_struct *work)
+{
+	unsigned int i;
+	bool free_scratch = true;
+
+	mutex_lock(&cpool_mutex);
+	for (i = 0; i < cpool_populated; i++) {
+		if (kref_read(&cpool[i].kref) > 0) {
+			free_scratch = false;
+			continue;
+		}
+		if (!cpool[i].alg)
+			continue;
+		__cpool_free_entry(&cpool[i]);
+	}
+	if (free_scratch)
+		sigpool_scratch_free();
+	mutex_unlock(&cpool_mutex);
+}
+
+static DECLARE_WORK(cpool_cleanup_work, cpool_cleanup_work_cb);
+static void cpool_schedule_cleanup(struct kref *kref)
+{
+	schedule_work(&cpool_cleanup_work);
+}
+
+/**
+ * tcp_sigpool_release - decreases number of users for a pool. If it was
+ * the last user of the pool, releases any memory that was consumed.
+ * @id: tcp_sigpool that was previously allocated by tcp_sigpool_alloc_ahash()
+ */
+void tcp_sigpool_release(unsigned int id)
+{
+	if (WARN_ON_ONCE(id > cpool_populated || !cpool[id].alg))
+		return;
+
+	/* slow-path */
+	kref_put(&cpool[id].kref, cpool_schedule_cleanup);
+}
+EXPORT_SYMBOL_GPL(tcp_sigpool_release);
+
+/**
+ * tcp_sigpool_get - increases number of users (refcounter) for a pool
+ * @id: tcp_sigpool that was previously allocated by tcp_sigpool_alloc_ahash()
+ */
+void tcp_sigpool_get(unsigned int id)
+{
+	if (WARN_ON_ONCE(id > cpool_populated || !cpool[id].alg))
+		return;
+	kref_get(&cpool[id].kref);
+}
+EXPORT_SYMBOL_GPL(tcp_sigpool_get);
+
+int tcp_sigpool_start(unsigned int id, struct tcp_sigpool *c)
+{
+	rcu_read_lock_bh();
+	if (WARN_ON_ONCE(id > cpool_populated || !cpool[id].alg)) {
+		rcu_read_unlock_bh();
+		return -EINVAL;
+	}
+	c->req = *this_cpu_ptr(cpool[id].req);
+	/* Pairs with tcp_sigpool_reserve_scratch(), scratch area is
+	 * valid (allocated) until tcp_sigpool_end().
+	 */
+	c->scratch = rcu_dereference_bh(*this_cpu_ptr(&sigpool_scratch));
+	return 0;
+}
+EXPORT_SYMBOL_GPL(tcp_sigpool_start);
+
+/**
+ * tcp_sigpool_algo - return algorithm of tcp_sigpool
+ * @id: tcp_sigpool that was previously allocated by tcp_sigpool_alloc_ahash()
+ * @buf: buffer to return name of algorithm
+ * @buf_len: size of @buf
+ */
+size_t tcp_sigpool_algo(unsigned int id, char *buf, size_t buf_len)
+{
+	size_t ret = 0;
+
+	/* slow-path */
+	mutex_lock(&cpool_mutex);
+	if (cpool[id].alg)
+		ret = strscpy(buf, cpool[id].alg, buf_len);
+	mutex_unlock(&cpool_mutex);
+	return ret;
+}
+EXPORT_SYMBOL_GPL(tcp_sigpool_algo);
+
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("Per-CPU pool of crypto requests");
diff --git a/net/ipv6/tcp_ipv6.c b/net/ipv6/tcp_ipv6.c
index 11b736a76bd7..116319d56ab2 100644
--- a/net/ipv6/tcp_ipv6.c
+++ b/net/ipv6/tcp_ipv6.c
@@ -672,7 +672,7 @@  static int tcp_v6_parse_md5_keys(struct sock *sk, int optname,
 			      cmd.tcpm_key, cmd.tcpm_keylen);
 }
 
-static int tcp_v6_md5_hash_headers(struct tcp_md5sig_pool *hp,
+static int tcp_v6_md5_hash_headers(struct tcp_sigpool *hp,
 				   const struct in6_addr *daddr,
 				   const struct in6_addr *saddr,
 				   const struct tcphdr *th, int nbytes)
@@ -693,39 +693,36 @@  static int tcp_v6_md5_hash_headers(struct tcp_md5sig_pool *hp,
 	_th->check = 0;
 
 	sg_init_one(&sg, bp, sizeof(*bp) + sizeof(*th));
-	ahash_request_set_crypt(hp->md5_req, &sg, NULL,
+	ahash_request_set_crypt(hp->req, &sg, NULL,
 				sizeof(*bp) + sizeof(*th));
-	return crypto_ahash_update(hp->md5_req);
+	return crypto_ahash_update(hp->req);
 }
 
 static int tcp_v6_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
 			       const struct in6_addr *daddr, struct in6_addr *saddr,
 			       const struct tcphdr *th)
 {
-	struct tcp_md5sig_pool *hp;
-	struct ahash_request *req;
+	struct tcp_sigpool hp;
 
-	hp = tcp_get_md5sig_pool();
-	if (!hp)
-		goto clear_hash_noput;
-	req = hp->md5_req;
+	if (tcp_sigpool_start(tcp_md5_sigpool_id, &hp))
+		goto clear_hash_nostart;
 
-	if (crypto_ahash_init(req))
+	if (crypto_ahash_init(hp.req))
 		goto clear_hash;
-	if (tcp_v6_md5_hash_headers(hp, daddr, saddr, th, th->doff << 2))
+	if (tcp_v6_md5_hash_headers(&hp, daddr, saddr, th, th->doff << 2))
 		goto clear_hash;
-	if (tcp_md5_hash_key(hp, key))
+	if (tcp_md5_hash_key(&hp, key))
 		goto clear_hash;
-	ahash_request_set_crypt(req, NULL, md5_hash, 0);
-	if (crypto_ahash_final(req))
+	ahash_request_set_crypt(hp.req, NULL, md5_hash, 0);
+	if (crypto_ahash_final(hp.req))
 		goto clear_hash;
 
-	tcp_put_md5sig_pool();
+	tcp_sigpool_end();
 	return 0;
 
 clear_hash:
-	tcp_put_md5sig_pool();
-clear_hash_noput:
+	tcp_sigpool_end();
+clear_hash_nostart:
 	memset(md5_hash, 0, 16);
 	return 1;
 }
@@ -736,8 +733,7 @@  static int tcp_v6_md5_hash_skb(char *md5_hash,
 			       const struct sk_buff *skb)
 {
 	const struct in6_addr *saddr, *daddr;
-	struct tcp_md5sig_pool *hp;
-	struct ahash_request *req;
+	struct tcp_sigpool hp;
 	const struct tcphdr *th = tcp_hdr(skb);
 
 	if (sk) { /* valid for establish/request sockets */
@@ -749,30 +745,28 @@  static int tcp_v6_md5_hash_skb(char *md5_hash,
 		daddr = &ip6h->daddr;
 	}
 
-	hp = tcp_get_md5sig_pool();
-	if (!hp)
-		goto clear_hash_noput;
-	req = hp->md5_req;
+	if (tcp_sigpool_start(tcp_md5_sigpool_id, &hp))
+		goto clear_hash_nostart;
 
-	if (crypto_ahash_init(req))
+	if (crypto_ahash_init(hp.req))
 		goto clear_hash;
 
-	if (tcp_v6_md5_hash_headers(hp, daddr, saddr, th, skb->len))
+	if (tcp_v6_md5_hash_headers(&hp, daddr, saddr, th, skb->len))
 		goto clear_hash;
-	if (tcp_md5_hash_skb_data(hp, skb, th->doff << 2))
+	if (tcp_md5_hash_skb_data(&hp, skb, th->doff << 2))
 		goto clear_hash;
-	if (tcp_md5_hash_key(hp, key))
+	if (tcp_md5_hash_key(&hp, key))
 		goto clear_hash;
-	ahash_request_set_crypt(req, NULL, md5_hash, 0);
-	if (crypto_ahash_final(req))
+	ahash_request_set_crypt(hp.req, NULL, md5_hash, 0);
+	if (crypto_ahash_final(hp.req))
 		goto clear_hash;
 
-	tcp_put_md5sig_pool();
+	tcp_sigpool_end();
 	return 0;
 
 clear_hash:
-	tcp_put_md5sig_pool();
-clear_hash_noput:
+	tcp_sigpool_end();
+clear_hash_nostart:
 	memset(md5_hash, 0, 16);
 	return 1;
 }