[net,v1] octeontx2-pf: Fix SQE threshold checking
Commit Message
Current way of checking available SQE count which is based on
HW updated SQB count could result in driver submitting an SQE
even before CQE for the previously transmitted SQE at the same
index is processed in NAPI resulting losing SKB pointers,
hence a leak. Fix this by checking a consumer index which
is updated once CQE is processed.
Fixes: 3ca6c4c882a7 ("octeontx2-pf: Add packet transmission support")
Signed-off-by: Ratheesh Kannoth <rkannoth@marvell.com>
Reviewed-by: Sunil Kovvuri Goutham <sgoutham@marvell.com>
---
.../marvell/octeontx2/nic/otx2_common.c | 1 +
.../marvell/octeontx2/nic/otx2_txrx.c | 32 +++++++++++--------
.../marvell/octeontx2/nic/otx2_txrx.h | 1 +
3 files changed, 21 insertions(+), 13 deletions(-)
Comments
Hello:
This patch was applied to netdev/net.git (master)
by Paolo Abeni <pabeni@redhat.com>:
On Mon, 7 Nov 2022 09:05:05 +0530 you wrote:
> Current way of checking available SQE count which is based on
> HW updated SQB count could result in driver submitting an SQE
> even before CQE for the previously transmitted SQE at the same
> index is processed in NAPI resulting losing SKB pointers,
> hence a leak. Fix this by checking a consumer index which
> is updated once CQE is processed.
>
> [...]
Here is the summary with links:
- [net,v1] octeontx2-pf: Fix SQE threshold checking
https://git.kernel.org/netdev/net/c/f0dfc4c88ef3
You are awesome, thank you!
@@ -898,6 +898,7 @@ static int otx2_sq_init(struct otx2_nic *pfvf, u16 qidx, u16 sqb_aura)
}
sq->head = 0;
+ sq->cons_head = 0;
sq->sqe_per_sqb = (pfvf->hw.sqb_size / sq->sqe_size) - 1;
sq->num_sqbs = (qset->sqe_cnt + sq->sqe_per_sqb) / sq->sqe_per_sqb;
/* Set SQE threshold to 10% of total SQEs */
@@ -441,6 +441,7 @@ static int otx2_tx_napi_handler(struct otx2_nic *pfvf,
struct otx2_cq_queue *cq, int budget)
{
int tx_pkts = 0, tx_bytes = 0, qidx;
+ struct otx2_snd_queue *sq;
struct nix_cqe_tx_s *cqe;
int processed_cqe = 0;
@@ -451,6 +452,9 @@ static int otx2_tx_napi_handler(struct otx2_nic *pfvf,
return 0;
process_cqe:
+ qidx = cq->cq_idx - pfvf->hw.rx_queues;
+ sq = &pfvf->qset.sq[qidx];
+
while (likely(processed_cqe < budget) && cq->pend_cqe) {
cqe = (struct nix_cqe_tx_s *)otx2_get_next_cqe(cq);
if (unlikely(!cqe)) {
@@ -458,18 +462,20 @@ static int otx2_tx_napi_handler(struct otx2_nic *pfvf,
return 0;
break;
}
+
if (cq->cq_type == CQ_XDP) {
- qidx = cq->cq_idx - pfvf->hw.rx_queues;
- otx2_xdp_snd_pkt_handler(pfvf, &pfvf->qset.sq[qidx],
- cqe);
+ otx2_xdp_snd_pkt_handler(pfvf, sq, cqe);
} else {
- otx2_snd_pkt_handler(pfvf, cq,
- &pfvf->qset.sq[cq->cint_idx],
- cqe, budget, &tx_pkts, &tx_bytes);
+ otx2_snd_pkt_handler(pfvf, cq, sq, cqe, budget,
+ &tx_pkts, &tx_bytes);
}
+
cqe->hdr.cqe_type = NIX_XQE_TYPE_INVALID;
processed_cqe++;
cq->pend_cqe--;
+
+ sq->cons_head++;
+ sq->cons_head &= (sq->sqe_cnt - 1);
}
/* Free CQEs to HW */
@@ -1072,17 +1078,17 @@ bool otx2_sq_append_skb(struct net_device *netdev, struct otx2_snd_queue *sq,
{
struct netdev_queue *txq = netdev_get_tx_queue(netdev, qidx);
struct otx2_nic *pfvf = netdev_priv(netdev);
- int offset, num_segs, free_sqe;
+ int offset, num_segs, free_desc;
struct nix_sqe_hdr_s *sqe_hdr;
- /* Check if there is room for new SQE.
- * 'Num of SQBs freed to SQ's pool - SQ's Aura count'
- * will give free SQE count.
+ /* Check if there is enough room between producer
+ * and consumer index.
*/
- free_sqe = (sq->num_sqbs - *sq->aura_fc_addr) * sq->sqe_per_sqb;
+ free_desc = (sq->cons_head - sq->head - 1 + sq->sqe_cnt) & (sq->sqe_cnt - 1);
+ if (free_desc < sq->sqe_thresh)
+ return false;
- if (free_sqe < sq->sqe_thresh ||
- free_sqe < otx2_get_sqe_count(pfvf, skb))
+ if (free_desc < otx2_get_sqe_count(pfvf, skb))
return false;
num_segs = skb_shinfo(skb)->nr_frags + 1;
@@ -79,6 +79,7 @@ struct sg_list {
struct otx2_snd_queue {
u8 aura_id;
u16 head;
+ u16 cons_head;
u16 sqe_size;
u32 sqe_cnt;
u16 num_sqbs;