common/mlx5: fix shared mempool subscription

PATCH In review Gregory Etelson 2 participants 2 messages Nov 3, 2022 – Nov 6, 2022

Share & tools

Gregory Etelson

author

2022-11-03 10:45 UTC #0 lore raw

PATCH 1/1 common/mlx5: fix shared mempool subscription

MLX5 PMD counted each mempool subscribe invocation. The PMD expected that the mempool subscription will be deleted after the mempool counter dropped to 0. However, current PMD design unsubscribes mempool callbacks only once.
As the result, the PMD destroyed mlx5_common_device but kept shared RX subscription callback. EAL tried to activate that callback and crashed.

The patch removes mempool subscriptions counter.
The PMD registers mempool subscription once only. An attempt to register existing subscription returns EEXIST.
Also, the PMD expects to remove subscription when mempool unsubscribe was activated.

Fixes 8ad97e4b3215 ("common/mlx5: fix multi-process mempool registration")

Cc [email protected]

Signed-off-by Gregory Etelson <[email protected]>

Acked-by Matan Azrad <[email protected]>

---
 drivers/common/mlx5/mlx5_common.c    | 22 +++++++++++-----------
 drivers/common/mlx5/mlx5_common_mr.c |  1 -
 drivers/common/mlx5/mlx5_common_mr.h |  1 -
 3 files changed, 11 insertions(+), 13 deletions(-)

drivers/common/mlx5/mlx5_common.c +11 ~~-11~~

index bf22c0694d..0ad14a48c7 100644
@@ -577,6 +577,11 @@ mlx5_dev_mempool_event_cb(enum rte_mempool_event event, struct rte_mempool *mp,
 	}
 }
 
+/**
+ * Primary and secondary processes share the `cdev` pointer.
+ * Callbacks addresses are local in each process.
+ * Therefore, each process can register private callbacks.
+ */
 int
 mlx5_dev_mempool_subscribe(struct mlx5_common_device *cdev)
 {

@@ -588,14 +593,13 @@ mlx5_dev_mempool_subscribe(struct mlx5_common_device *cdev)
 	/* Callback for this device may be already registered. */
 	ret = rte_mempool_event_callback_register(mlx5_dev_mempool_event_cb,
 						  cdev);
-	if (ret != 0 && rte_errno != EEXIST)
-		goto exit;
-	__atomic_add_fetch(&cdev->mr_scache.mempool_cb_reg_n, 1,
-			   __ATOMIC_ACQUIRE);
 	/* Register mempools only once for this device. */
-	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
+	if (ret == 0 && rte_eal_process_type() == RTE_PROC_PRIMARY) {
 		rte_mempool_walk(mlx5_dev_mempool_register_cb, cdev);
-	ret = 0;
+		goto exit;
+	}
+	if (ret != 0 && rte_errno == EEXIST)
+		ret = 0;
 exit:
 	rte_rwlock_write_unlock(&cdev->mr_scache.mprwlock);
 	return ret;

@@ -604,15 +608,11 @@ mlx5_dev_mempool_subscribe(struct mlx5_common_device *cdev)
 static void
 mlx5_dev_mempool_unsubscribe(struct mlx5_common_device *cdev)
 {
-	uint32_t mempool_cb_reg_n;
 	int ret;
 
+	MLX5_ASSERT(cdev->dev != NULL);
 	if (!cdev->config.mr_mempool_reg_en)
 		return;
-	mempool_cb_reg_n = __atomic_sub_fetch(&cdev->mr_scache.mempool_cb_reg_n,
-					      1, __ATOMIC_RELEASE);
-	if (mempool_cb_reg_n > 0)
-		return;
 	/* Stop watching for mempool events and unregister all mempools. */
 	ret = rte_mempool_event_callback_unregister(mlx5_dev_mempool_event_cb,
 						    cdev);

drivers/common/mlx5/mlx5_common_mr.c +0 -1

index 1d54102b54..0e1d2434ab 100644
@@ -1138,7 +1138,6 @@ mlx5_mr_create_cache(struct mlx5_mr_share_cache *share_cache, int socket)
 			      &share_cache->dereg_mr_cb);
 	rte_rwlock_init(&share_cache->rwlock);
 	rte_rwlock_init(&share_cache->mprwlock);
-	share_cache->mempool_cb_reg_n = 0;
 	/* Initialize B-tree and allocate memory for global MR cache table. */
 	return mlx5_mr_btree_init(&share_cache->cache,
 				  MLX5_MR_BTREE_CACHE_N * 2, socket);

drivers/common/mlx5/mlx5_common_mr.h +0 -1

index f774ccbf33..13eb350980 100644
@@ -81,7 +81,6 @@ struct mlx5_mr_share_cache {
 	uint32_t dev_gen; /* Generation number to flush local caches. */
 	rte_rwlock_t rwlock; /* MR cache Lock. */
 	rte_rwlock_t mprwlock; /* Mempool Registration Lock. */
-	uint32_t mempool_cb_reg_n; /* Mempool event callback registrants. */
 	struct mlx5_mr_btree cache; /* Global MR cache table. */
 	struct mlx5_mr_list mr_list; /* Registered MR list. */
 	struct mlx5_mr_list mr_free_list; /* Freed MR list. */

Signature · 1 line

2.34.1

Raslan Darawsheh

2022-11-06 11:08 UTC #1 lore raw

Hi,

Quoted 24 lines

-----Original Message-----
From: Gregory Etelson <[email protected]>
Sent: Thursday, November 3, 2022 12:44 PM
To: [email protected]
Cc: Gregory Etelson <[email protected]>; Matan Azrad <[email protected]>; Raslan Darawsheh <[email protected]>; [email protected]; Slava Ovsiienko <[email protected]>; Anatoly Burakov <[email protected]>; Dmitry Kozlyuk <[email protected]>
Subject: [PATCH] common/mlx5: fix shared mempool subscription
MLX5 PMD counted each mempool subscribe invocation. The PMD expected that the mempool subscription will be deleted after the mempool counter dropped to 0. However, current PMD design unsubscribes mempool callbacks only once.
As the result, the PMD destroyed mlx5_common_device but kept shared RX subscription callback. EAL tried to activate that callback and crashed.
The patch removes mempool subscriptions counter.
The PMD registers mempool subscription once only. An attempt to register existing subscription returns EEXIST.
Also, the PMD expects to remove subscription when mempool unsubscribe was activated.

Quoted 7 lines

Fixes 8ad97e4b3215 ("common/mlx5: fix multi-process mempool
registration")
Cc [email protected]
Signed-off-by Gregory Etelson <[email protected]>
Acked-by Matan Azrad <[email protected]>

Patch applied to next-net-mlx,

Kindest regards,
Raslan Darawsheh