patches for DPDK stable branches
 help / color / mirror / Atom feed
* [dpdk-stable] [PATCH 19.11] net/mlx5: fix counter query
@ 2020-08-11  8:59 Suanming Mou
  2020-08-12 12:42 ` Luca Boccassi
  0 siblings, 1 reply; 2+ messages in thread
From: Suanming Mou @ 2020-08-11  8:59 UTC (permalink / raw)
  To: luca.boccassi; +Cc: stable, rasland

[ upstream commit e1293b10dea9bbae2cb2ebc47b5066df73275e2f ]

Currently, the counter query requires the counter ID should start
with 4 aligned. In none-batch mode, the counter pool might have the
chance to get the counter ID not 4 aligned. In this case, the counter
should be skipped, or the query will be failed.

Skip the counter with ID not 4 aligned as the first counter in the
none-batch count pool to avoid invalid counter query. Once having
new min_dcs ID in the poll less than the skipped counters, the
skipped counters will be returned to the pool free list to use.

Fixes: 5382d28c2110 ("net/mlx5: accelerate DV flow counter transactions")
Signed-off-by: Suanming Mou <suanmingm@mellanox.com>
---
 drivers/net/mlx5/mlx5.h         |  4 +-
 drivers/net/mlx5/mlx5_flow.c    |  6 +++
 drivers/net/mlx5/mlx5_flow_dv.c | 94 ++++++++++++++++++++++++++++++++-
 drivers/net/mlx5/mlx5_prm.h     |  3 ++
 4 files changed, 104 insertions(+), 3 deletions(-)

diff --git a/drivers/net/mlx5/mlx5.h b/drivers/net/mlx5/mlx5.h
index 51f067b4b..07be509f6 100644
--- a/drivers/net/mlx5/mlx5.h
+++ b/drivers/net/mlx5/mlx5.h
@@ -486,7 +486,8 @@ struct mlx5_flow_counter {
 	uint32_t shared:1; /**< Share counter ID with other flow rules. */
 	uint32_t batch: 1;
 	/**< Whether the counter was allocated by batch command. */
-	uint32_t ref_cnt:30; /**< Reference counter. */
+	uint32_t ref_cnt:29; /**< Reference counter. */
+	uint32_t skipped:1; /* This counter is skipped or not. */
 	uint32_t id; /**< Counter ID. */
 	union {  /**< Holds the counters for the rule. */
 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
@@ -518,6 +519,7 @@ struct mlx5_flow_counter_pool {
 	/* The devx object of the minimum counter ID. */
 	rte_atomic64_t query_gen;
 	uint32_t n_counters: 16; /* Number of devx allocated counters. */
+	uint32_t skip_cnt:1; /* Pool contains skipped counter. */
 	rte_spinlock_t sl; /* The pool lock. */
 	struct mlx5_counter_stats_raw *raw;
 	struct mlx5_counter_stats_raw *raw_hw; /* The raw on HW working. */
diff --git a/drivers/net/mlx5/mlx5_flow.c b/drivers/net/mlx5/mlx5_flow.c
index a4bbb74fc..3868b1cf5 100644
--- a/drivers/net/mlx5/mlx5_flow.c
+++ b/drivers/net/mlx5/mlx5_flow.c
@@ -5594,6 +5594,11 @@ mlx5_flow_query_alarm(void *arg)
 		goto set_alarm;
 	dcs = (struct mlx5_devx_obj *)(uintptr_t)rte_atomic64_read
 							      (&pool->a64_dcs);
+	if (dcs->id & (MLX5_CNT_BATCH_QUERY_ID_ALIGNMENT - 1)) {
+		/* Pool without valid counter. */
+		pool->raw_hw = NULL;
+		goto next_pool;
+	}
 	offset = batch ? 0 : dcs->id % MLX5_COUNTERS_PER_POOL;
 	ret = mlx5_devx_cmd_flow_counter_query(dcs, 0, MLX5_COUNTERS_PER_POOL -
 					       offset, NULL, NULL,
@@ -5611,6 +5616,7 @@ mlx5_flow_query_alarm(void *arg)
 	pool->raw_hw->min_dcs_id = dcs->id;
 	LIST_REMOVE(pool->raw_hw, next);
 	sh->cmng.pending_queries++;
+next_pool:
 	pool_index++;
 	if (pool_index >= rte_atomic16_read(&cont->n_valid)) {
 		batch ^= 0x1;
diff --git a/drivers/net/mlx5/mlx5_flow_dv.c b/drivers/net/mlx5/mlx5_flow_dv.c
index 387036d0b..60d6af04e 100644
--- a/drivers/net/mlx5/mlx5_flow_dv.c
+++ b/drivers/net/mlx5/mlx5_flow_dv.c
@@ -4101,6 +4101,63 @@ flow_dv_pool_create(struct rte_eth_dev *dev, struct mlx5_devx_obj *dcs,
 	rte_atomic16_add(&cont->n_valid, 1);
 	return cont;
 }
+/**
+ * Restore skipped counters in the pool.
+ *
+ * As counter pool query requires the first counter dcs
+ * ID start with 4 alinged, if the pool counters with
+ * min_dcs ID are not aligned with 4, the counters will
+ * be skipped.
+ * Once other min_dcs ID less than these skipped counter
+ * dcs ID appears, the skipped counters will be safe to
+ * use.
+ * Should be called when min_dcs is updated.
+ *
+ * @param[in] pool
+ *   Current counter pool.
+ * @param[in] last_min_dcs
+ *   Last min_dcs.
+ */
+static void
+flow_dv_counter_restore(struct mlx5_flow_counter_pool *pool,
+			struct mlx5_devx_obj *last_min_dcs)
+{
+	struct mlx5_flow_counter *cnt;
+	uint32_t offset, new_offset;
+	uint32_t skip_cnt = 0;
+	uint32_t i;
+
+	if (!pool->skip_cnt)
+		return;
+	/*
+	 * If last min_dcs is not valid. The skipped counter may even after
+	 * last min_dcs, set the offset to the whole pool.
+	 */
+	if (last_min_dcs->id & (MLX5_CNT_BATCH_QUERY_ID_ALIGNMENT - 1))
+		offset = MLX5_COUNTERS_PER_POOL;
+	else
+		offset = last_min_dcs->id % MLX5_COUNTERS_PER_POOL;
+	new_offset = pool->min_dcs->id % MLX5_COUNTERS_PER_POOL;
+	/*
+	 * Check the counters from 1 to the last_min_dcs range. Counters
+	 * before new min_dcs indicates pool still has skipped counters.
+	 * Counters be skipped after new min_dcs will be ready to use.
+	 * Offset 0 counter must be empty or min_dcs, start from 1.
+	 */
+	for (i = 1; i < offset; i++) {
+		cnt = &pool->counters_raw[i];
+		if (cnt->skipped) {
+			if (i > new_offset) {
+				cnt->skipped = 0;
+				TAILQ_INSERT_TAIL(&pool->counters, cnt, next);
+			} else {
+				skip_cnt++;
+			}
+		}
+	}
+	if (!skip_cnt)
+		pool->skip_cnt = 0;
+}
 
 /**
  * Prepare a new counter and/or a new counter pool.
@@ -4125,11 +4182,13 @@ flow_dv_counter_pool_prepare(struct rte_eth_dev *dev,
 	struct mlx5_pools_container *cont;
 	struct mlx5_flow_counter_pool *pool;
 	struct mlx5_devx_obj *dcs = NULL;
+	struct mlx5_devx_obj *last_min_dcs;
 	struct mlx5_flow_counter *cnt;
 	uint32_t i;
 
 	cont = MLX5_CNT_CONTAINER(priv->sh, batch, 0);
 	if (!batch) {
+retry:
 		/* bulk_bitmap must be 0 for single counter allocation. */
 		dcs = mlx5_devx_cmd_flow_counter_alloc(priv->sh->ctx, 0);
 		if (!dcs)
@@ -4142,13 +4201,44 @@ flow_dv_counter_pool_prepare(struct rte_eth_dev *dev,
 				return NULL;
 			}
 			pool = TAILQ_FIRST(&cont->pool_list);
-		} else if (dcs->id < pool->min_dcs->id) {
+		} else if (((dcs->id < pool->min_dcs->id) ||
+			   pool->min_dcs->id &
+			   (MLX5_CNT_BATCH_QUERY_ID_ALIGNMENT - 1)) &&
+			   !(dcs->id &
+			   (MLX5_CNT_BATCH_QUERY_ID_ALIGNMENT - 1))) {
+			/*
+			 * Update the pool min_dcs only if current dcs is
+			 * valid and exist min_dcs is not valid or greater
+			 * than new dcs.
+			 */
+			last_min_dcs = pool->min_dcs;
 			rte_atomic64_set(&pool->a64_dcs,
 					 (int64_t)(uintptr_t)dcs);
+			/*
+			 * Restore any skipped counters if the new min_dcs
+			 * ID is smaller or min_dcs is not valid.
+			 */
+			if (dcs->id < last_min_dcs->id ||
+			    last_min_dcs->id &
+			    (MLX5_CNT_BATCH_QUERY_ID_ALIGNMENT - 1))
+				flow_dv_counter_restore(pool, last_min_dcs);
 		}
 		cnt = &pool->counters_raw[dcs->id % MLX5_COUNTERS_PER_POOL];
-		TAILQ_INSERT_HEAD(&pool->counters, cnt, next);
 		cnt->dcs = dcs;
+		/*
+		 * If min_dcs is not valid, it means the new allocated dcs
+		 * also fail to become the valid min_dcs, just skip it.
+		 * Or if min_dcs is valid, and new dcs ID is smaller than
+		 * min_dcs, but not become the min_dcs, also skip it.
+		 */
+		if (pool->min_dcs->id &
+		    (MLX5_CNT_BATCH_QUERY_ID_ALIGNMENT - 1) ||
+		    dcs->id < pool->min_dcs->id) {
+			cnt->skipped = 1;
+			pool->skip_cnt = 1;
+			goto retry;
+		}
+		TAILQ_INSERT_HEAD(&pool->counters, cnt, next);
 		*cnt_free = cnt;
 		return cont;
 	}
diff --git a/drivers/net/mlx5/mlx5_prm.h b/drivers/net/mlx5/mlx5_prm.h
index 4c8671976..c82a854bd 100644
--- a/drivers/net/mlx5/mlx5_prm.h
+++ b/drivers/net/mlx5/mlx5_prm.h
@@ -725,6 +725,9 @@ enum {
 	MLX5_MKC_ACCESS_MODE_MTT   = 0x1,
 };
 
+/* The counter batch query requires ID align with 4. */
+#define MLX5_CNT_BATCH_QUERY_ID_ALIGNMENT 4
+
 /* Flow counters. */
 struct mlx5_ifc_alloc_flow_counter_out_bits {
 	u8         status[0x8];
-- 
2.25.1


^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2020-08-12 12:42 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2020-08-11  8:59 [dpdk-stable] [PATCH 19.11] net/mlx5: fix counter query Suanming Mou
2020-08-12 12:42 ` Luca Boccassi

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).