Skip to contentSkip to conversation

net/mlx5: fixes for rx queue count calculation

PATCH Superseded Maxime Leroy 2 participants 5 messages Nov 12, 2020 – Nov 13, 2020 Network drivers

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

v2 cover net/mlx5: fixes for rx queue count calculation

This patchset provides several bug fixes for rx queue count calculation for mlx5 driver.

---
V2:
* squash first patch and second patch
* fix wrong init of used for compressed cqes

Didier Pallard (1):
net/mlx5: fix Rx descriptor status returned value

Maxime Leroy (1):
net/mlx5: fix Rx queue count calculation

 drivers/net/mlx5/mlx5_rxtx.c | 19 ++++++++++++-------
 1 file changed, 12 insertions(+), 7 deletions(-)
Signature · 1 line
2.27.0
#1 lore raw

PATCH v2 1/2 net/mlx5: fix Rx queue count calculation

The commit d2d57605522d ("net/mlx5: fix Rx queue count calculation") is uncorrect because the count calculation is wrong for the next cqe:

Example:

Compressed Set of packets 1 | Compressed Set of packets 2
C | a | e0 | e1 | e2 | e3 | e4 | e5 | C | a | e0

There are 2 compressed set of packets in the first queue. For the first set, n is computed correctly.

But for the second, n is not computed properly. Because the zip context is for the first set. The second set is not yet decompressed, so there are no context.

To fix the issue, we should only use the zip context for the first CQEs serie.

Fixes d2d57605522d ("net/mlx5: fix Rx queue count calculation")
---
 drivers/net/mlx5/mlx5_rxtx.c | 16 ++++++++++------
 1 file changed, 10 insertions(+), 6 deletions(-)
drivers/net/mlx5/mlx5_rxtx.c +10 -6
index 844a1c63..2733dcd3 100644
@@ -462,11 +462,18 @@ rx_queue_count(struct mlx5_rxq_data *rxq)
 {
 	struct rxq_zip *zip = &rxq->zip;
 	volatile struct mlx5_cqe *cqe;
-	unsigned int cq_ci = rxq->cq_ci;
 	const unsigned int cqe_n = (1 << rxq->cqe_n);
 	const unsigned int cqe_cnt = cqe_n - 1;
-	unsigned int used = 0;
+	unsigned int cq_ci, used;
 
+	/* if we are processing a compressed cqe */
+	if (zip->ai) {
+		used = zip->cqe_cnt - zip->ai;
+		cq_ci = zip->cq_ci;
+	} else {
+		used = 0;
+		cq_ci = rxq->cq_ci;
+	}
 	cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
 	while (check_cqe(cqe, cqe_n, cq_ci) != MLX5_CQE_STATUS_HW_OWN) {
 		int8_t op_own;
@@ -474,10 +481,7 @@ rx_queue_count(struct mlx5_rxq_data *rxq)
 
 		op_own = cqe->op_own;
 		if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED)
-			if (unlikely(zip->ai))
-				n = zip->cqe_cnt - zip->ai;
-			else
-				n = rte_be_to_cpu_32(cqe->byte_cnt);
+			n = rte_be_to_cpu_32(cqe->byte_cnt);
 		else
 			n = 1;
 		cq_ci += n;
Signature · 1 line
2.27.0
Acked-by on v2 1/2
#2 lore raw

Re: [dpdk-dev] [PATCH v2 1/2] net/mlx5: fix Rx queue count calculation

Quoted 21 lines

-----Original Message-----
From: Maxime Leroy <[email protected]>
Sent: Thursday, November 12, 2020 17:39
To: Matan Azrad <[email protected]>; Shahaf Shuler <[email protected]>; Slava Ovsiienko <[email protected]>; Alexander Kozyrev <[email protected]>
Cc: [email protected]; NBU-Contact-N?lio Laranjeiro <[email protected]>
Subject: [PATCH v2 1/2] net/mlx5: fix Rx queue count calculation

The commit d2d57605522d ("net/mlx5: fix Rx queue count calculation") is uncorrect because the count calculation is wrong for the next cqe:

Example:

Compressed Set of packets 1 | Compressed Set of packets 2
C | a | e0 | e1 | e2 | e3 | e4 | e5 | C | a | e0

There are 2 compressed set of packets in the first queue. For the first set, n is computed correctly.

Quoted 8 lines

But for the second, n is not computed properly. Because the zip context is for the first set. The second set is not yet decompressed, so there are no context.

To fix the issue, we should only use the zip context for the first CQEs serie.

Fixes d2d57605522d ("net/mlx5: fix Rx queue count calculation")

Thank you for the fix. The second patch is still on review - I have some doubts about final RTE_MIN(), checking.

#3 lore raw

PATCH v2 2/2 net/mlx5: fix Rx descriptor status returned value

In reply to Maxime Leroy

From: Didier Pallard <[email protected]>

One entry may contain several segments, so 'used' must be multiplied by number of segments per entry to properly reflect the queue usage.

Fixes 8788fec1f269 ("net/mlx5: implement descriptor status API")
---
 drivers/net/mlx5/mlx5_rxtx.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/net/mlx5/mlx5_rxtx.c +2 -1
index 2733dcd3..f390dd66 100644
@@ -463,6 +463,7 @@ rx_queue_count(struct mlx5_rxq_data *rxq)
 	struct rxq_zip *zip = &rxq->zip;
 	volatile struct mlx5_cqe *cqe;
 	const unsigned int cqe_n = (1 << rxq->cqe_n);
+	const unsigned int sges_n = (1 << rxq->sges_n);
 	const unsigned int cqe_cnt = cqe_n - 1;
 	unsigned int cq_ci, used;
 
@@ -488,7 +489,7 @@ rx_queue_count(struct mlx5_rxq_data *rxq)
 		used += n;
 		cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
 	}
-	used = RTE_MIN(used, cqe_n);
+	used = RTE_MIN(used * sges_n, cqe_n);
 	return used;
 }
 
Signature · 1 line
2.27.0
#4 lore raw

Re: [dpdk-dev] [PATCH v2 2/2] net/mlx5: fix Rx descriptor status returned value

Hi, Maxime

Quoted 34 lines

-----Original Message-----
From: Maxime Leroy <[email protected]>
Sent: Thursday, November 12, 2020 17:39
To: Matan Azrad <[email protected]>; Shahaf Shuler <[email protected]>; Slava Ovsiienko <[email protected]>; Olivier Matz <[email protected]>
Cc: [email protected]; Didier Pallard <[email protected]>
Subject: [PATCH v2 2/2] net/mlx5: fix Rx descriptor status returned value

From: Didier Pallard <[email protected]>

One entry may contain several segments, so 'used' must be multiplied by number of segments per entry to properly reflect the queue usage.

Fixes 8788fec1f269 ("net/mlx5: implement descriptor status API")
---
 drivers/net/mlx5/mlx5_rxtx.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/net/mlx5/mlx5_rxtx.c index renamed from drivers/net/mlx5/mlx5_rxtx.c +0 -0

2733dcd3..f390dd66 100644

drivers/net/mlx5/mlx5_rxtx.c +1 -0
@@ -463,6 +463,7 @@ rx_queue_count(struct mlx5_rxq_data *rxq)
 	struct rxq_zip *zip = &rxq->zip;
 	volatile struct mlx5_cqe *cqe;
 	const unsigned int cqe_n = (1 << rxq->cqe_n);
+	const unsigned int sges_n = (1 << rxq->sges_n);
 	const unsigned int cqe_cnt = cqe_n - 1;
 	unsigned int cq_ci, used;
 

drivers/net/mlx5/mlx5_rxtx.c rx_queue_count(struct mlx5_rxq_data *rxq) · line 492

Quoted 5 lines
drivers/net/mlx5/mlx5_rxtx.c +1 -1
@@ -488,7 +489,7 @@ rx_queue_count(struct mlx5_rxq_data *rxq)
 		used += n;
 		cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
 	}
-	used = RTE_MIN(used, cqe_n);
+	used = RTE_MIN(used * sges_n, cqe_n);

cqe_n reflects the number of the data descriptors in the RxQ, there might be convergency. I suppose the clamping should be:
- for non-MPRQ rx_burst (regular and vectorized) - (1 <<rxq->elts_n)
- for MPRQ (1 <<rxq->elts_n) * (1 << rxq->strd_n)

For non-MPRQ the rxq->strd_n is zero, hence, it could look like:

used = RTE_MIN(used * sges_n, (1 <<rxq->elts_n) * (1 << rxq->strd_n));

With best regards,
Slava