patch 'net/mlx5: fix flow workspace destruction' has been queued to stable release 21.11.5

[Kevin Traynor]

Hi,

FYI, your patch has been queued to stable release 21.11.5

Note it hasn't been pushed to http://dpdk.org/browse/dpdk-stable yet.
It will be pushed if I get no objections before 07/25/23. So please
shout if anyone has objections.

Also note that after the patch there's a diff of the upstream commit vs the
patch applied to the branch. This will indicate if there was any rebasing
needed to apply to the stable branch. If there were code changes for rebasing
(ie: not only metadata diffs), please double check that the rebase was
correctly done.

Queued patches are on a temporary branch at:
https://github.com/kevintraynor/dpdk-stable

This queued commit can be viewed at:
https://github.com/kevintraynor/dpdk-stable/commit/cd2b1999eb8c504f0d669e2022518b93c71085a0

Thanks.

Kevin

---
>From cd2b1999eb8c504f0d669e2022518b93c71085a0 Mon Sep 17 00:00:00 2001
From: Gregory Etelson <getelson at nvidia.com>
Date: Mon, 3 Jul 2023 12:50:52 +0300
Subject: [PATCH] net/mlx5: fix flow workspace destruction

[ upstream commit dc7c5e0aa905b675c56a66b2014b01b7f5ae8a1d ]

PMD uses pthread key to allocate and access per thread flow
workspace memory buffers.

PMD registered a key destructor function to clean up flow workspace
buffers. However, the key destructor was not called by the pthread
library.

The patch keeps track of per-thread flow workspaces in PMD.
Flow workspaces memory release is activated from PMD destructor.

In the meanwhile, workspace buffer and RSS queues array are allocated
in a single memory chunk with this patch. The maximal number of
queues RTE_ETH_RSS_RETA_SIZE_512 is chosen. Then the workspace
adjustment can be removed to reduce the software hiccup:
  1. realloc and content copy
  2. spinlock acquire and release

Bugzilla ID: 1255
Fixes: 5d55a494f4e6 ("net/mlx5: split multi-thread flow handling per OS")

Reported-by: David Marchand <david.marchand at redhat.com>
Signed-off-by: Gregory Etelson <getelson at nvidia.com>
Signed-off-by: Bing Zhao <bingz at nvidia.com>
Acked-by: Matan Azrad <matan at nvidia.com>
---
 drivers/net/mlx5/linux/mlx5_flow_os.c |  2 +-
 drivers/net/mlx5/mlx5.c               |  1 +
 drivers/net/mlx5/mlx5_flow.c          | 77 +++++++++++----------------
 drivers/net/mlx5/mlx5_flow.h          |  4 +-
 4 files changed, 36 insertions(+), 48 deletions(-)

diff --git a/drivers/net/mlx5/linux/mlx5_flow_os.c b/drivers/net/mlx5/linux/mlx5_flow_os.c
index a5956c255a..08337a9b24 100644
--- a/drivers/net/mlx5/linux/mlx5_flow_os.c
+++ b/drivers/net/mlx5/linux/mlx5_flow_os.c
@@ -13,5 +13,5 @@ int
 mlx5_flow_os_init_workspace_once(void)
 {
-	if (rte_thread_key_create(&key_workspace, flow_release_workspace)) {
+	if (rte_thread_key_create(&key_workspace, NULL)) {
 		DRV_LOG(ERR, "Can't create flow workspace data thread key.");
 		rte_errno = ENOMEM;
diff --git a/drivers/net/mlx5/mlx5.c b/drivers/net/mlx5/mlx5.c
index c5f2db1e4a..5645e8656c 100644
--- a/drivers/net/mlx5/mlx5.c
+++ b/drivers/net/mlx5/mlx5.c
@@ -1328,4 +1328,5 @@ mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
 		mlx5_os_net_cleanup();
 		mlx5_flow_os_release_workspace();
+		mlx5_flow_workspace_gc_release();
 	}
 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
diff --git a/drivers/net/mlx5/mlx5_flow.c b/drivers/net/mlx5/mlx5_flow.c
index 192e2d7718..133dca5dcf 100644
--- a/drivers/net/mlx5/mlx5_flow.c
+++ b/drivers/net/mlx5/mlx5_flow.c
@@ -6527,34 +6527,4 @@ flow_tunnel_from_rule(const struct mlx5_flow *flow)
 }
 
-/**
- * Adjust flow RSS workspace if needed.
- *
- * @param wks
- *   Pointer to thread flow work space.
- * @param rss_desc
- *   Pointer to RSS descriptor.
- * @param[in] nrssq_num
- *   New RSS queue number.
- *
- * @return
- *   0 on success, -1 otherwise and rte_errno is set.
- */
-static int
-flow_rss_workspace_adjust(struct mlx5_flow_workspace *wks,
-			  struct mlx5_flow_rss_desc *rss_desc,
-			  uint32_t nrssq_num)
-{
-	if (likely(nrssq_num <= wks->rssq_num))
-		return 0;
-	rss_desc->queue = realloc(rss_desc->queue,
-			  sizeof(*rss_desc->queue) * RTE_ALIGN(nrssq_num, 2));
-	if (!rss_desc->queue) {
-		rte_errno = ENOMEM;
-		return -1;
-	}
-	wks->rssq_num = RTE_ALIGN(nrssq_num, 2);
-	return 0;
-}
-
 /**
  * Create a flow and add it to @p list.
@@ -6674,6 +6644,5 @@ flow_list_create(struct rte_eth_dev *dev, enum mlx5_flow_type type,
 		rss = flow_get_rss_action(dev, p_actions_rx);
 	if (rss) {
-		if (flow_rss_workspace_adjust(wks, rss_desc, rss->queue_num))
-			return 0;
+		MLX5_ASSERT(rss->queue_num <= RTE_ETH_RSS_RETA_SIZE_512);
 		/*
 		 * The following information is required by
@@ -7108,5 +7077,4 @@ flow_release_workspace(void *data)
 	while (wks) {
 		next = wks->next;
-		free(wks->rss_desc.queue);
 		free(wks);
 		wks = next;
@@ -7114,4 +7082,27 @@ flow_release_workspace(void *data)
 }
 
+static struct mlx5_flow_workspace *gc_head;
+static rte_spinlock_t mlx5_flow_workspace_lock = RTE_SPINLOCK_INITIALIZER;
+
+static void
+mlx5_flow_workspace_gc_add(struct mlx5_flow_workspace *ws)
+{
+	rte_spinlock_lock(&mlx5_flow_workspace_lock);
+	ws->gc = gc_head;
+	gc_head = ws;
+	rte_spinlock_unlock(&mlx5_flow_workspace_lock);
+}
+
+void
+mlx5_flow_workspace_gc_release(void)
+{
+	while (gc_head) {
+		struct mlx5_flow_workspace *wks = gc_head;
+
+		gc_head = wks->gc;
+		flow_release_workspace(wks);
+	}
+}
+
 /**
  * Get thread specific current flow workspace.
@@ -7139,22 +7130,15 @@ static struct mlx5_flow_workspace*
 flow_alloc_thread_workspace(void)
 {
-	struct mlx5_flow_workspace *data = calloc(1, sizeof(*data));
+	size_t data_size = RTE_ALIGN(sizeof(struct mlx5_flow_workspace), sizeof(long));
+	size_t rss_queue_array_size = sizeof(uint16_t) * RTE_ETH_RSS_RETA_SIZE_512;
+	struct mlx5_flow_workspace *data = calloc(1, data_size +
+						     rss_queue_array_size);
 
 	if (!data) {
-		DRV_LOG(ERR, "Failed to allocate flow workspace "
-			"memory.");
+		DRV_LOG(ERR, "Failed to allocate flow workspace memory.");
 		return NULL;
 	}
-	data->rss_desc.queue = calloc(1,
-			sizeof(uint16_t) * MLX5_RSSQ_DEFAULT_NUM);
-	if (!data->rss_desc.queue)
-		goto err;
-	data->rssq_num = MLX5_RSSQ_DEFAULT_NUM;
+	data->rss_desc.queue = RTE_PTR_ADD(data, data_size);
 	return data;
-err:
-	if (data->rss_desc.queue)
-		free(data->rss_desc.queue);
-	free(data);
-	return NULL;
 }
 
@@ -7177,4 +7161,5 @@ mlx5_flow_push_thread_workspace(void)
 		if (!data)
 			return NULL;
+		mlx5_flow_workspace_gc_add(data);
 	} else if (!curr->inuse) {
 		data = curr;
diff --git a/drivers/net/mlx5/mlx5_flow.h b/drivers/net/mlx5/mlx5_flow.h
index 8ad8e63205..8a571f977e 100644
--- a/drivers/net/mlx5/mlx5_flow.h
+++ b/drivers/net/mlx5/mlx5_flow.h
@@ -1099,8 +1099,8 @@ struct mlx5_flow_workspace {
 	struct mlx5_flow_workspace *prev;
 	struct mlx5_flow_workspace *next;
+	struct mlx5_flow_workspace *gc;
 	uint32_t inuse; /* can't create new flow with current. */
 	struct mlx5_flow flows[MLX5_NUM_MAX_DEV_FLOWS];
 	struct mlx5_flow_rss_desc rss_desc;
-	uint32_t rssq_num; /* Allocated queue num in rss_desc. */
 	uint32_t flow_idx; /* Intermediate device flow index. */
 	struct mlx5_flow_meter_info *fm; /* Pointer to the meter in flow. */
@@ -1305,4 +1305,6 @@ struct mlx5_flow_workspace *mlx5_flow_push_thread_workspace(void);
 void mlx5_flow_pop_thread_workspace(void);
 struct mlx5_flow_workspace *mlx5_flow_get_thread_workspace(void);
+void mlx5_flow_workspace_gc_release(void);
+
 __extension__
 struct flow_grp_info {
-- 
2.41.0

---
  Diff of the applied patch vs upstream commit (please double-check if non-empty:
---
--- -	2023-07-20 16:18:10.154319379 +0100
+++ 0126-net-mlx5-fix-flow-workspace-destruction.patch	2023-07-20 16:17:55.178752619 +0100
@@ -1 +1 @@
-From dc7c5e0aa905b675c56a66b2014b01b7f5ae8a1d Mon Sep 17 00:00:00 2001
+From cd2b1999eb8c504f0d669e2022518b93c71085a0 Mon Sep 17 00:00:00 2001
@@ -5,0 +6,2 @@
+[ upstream commit dc7c5e0aa905b675c56a66b2014b01b7f5ae8a1d ]
+
@@ -25 +26,0 @@
-Cc: stable at dpdk.org
@@ -34 +35 @@
- drivers/net/mlx5/mlx5_flow.c          | 76 +++++++++++----------------
+ drivers/net/mlx5/mlx5_flow.c          | 77 +++++++++++----------------
@@ -36 +37 @@
- 4 files changed, 36 insertions(+), 47 deletions(-)
+ 4 files changed, 36 insertions(+), 48 deletions(-)
@@ -39 +40 @@
-index 3c9a823edf..b139bb75b9 100644
+index a5956c255a..08337a9b24 100644
@@ -42 +43 @@
-@@ -52,5 +52,5 @@ int
+@@ -13,5 +13,5 @@ int
@@ -50 +51 @@
-index 5f0aa296ba..fd9b76027d 100644
+index c5f2db1e4a..5645e8656c 100644
@@ -53 +54 @@
-@@ -1839,4 +1839,5 @@ mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
+@@ -1328,4 +1328,5 @@ mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
@@ -60 +61 @@
-index abb86241fc..1071ef0c3e 100644
+index 192e2d7718..133dca5dcf 100644
@@ -63 +64 @@
-@@ -7156,34 +7156,4 @@ flow_tunnel_from_rule(const struct mlx5_flow *flow)
+@@ -6527,34 +6527,4 @@ flow_tunnel_from_rule(const struct mlx5_flow *flow)
@@ -98 +99 @@
-@@ -7304,6 +7274,5 @@ flow_list_create(struct rte_eth_dev *dev, enum mlx5_flow_type type,
+@@ -6674,6 +6644,5 @@ flow_list_create(struct rte_eth_dev *dev, enum mlx5_flow_type type,
@@ -106 +107 @@
-@@ -8073,5 +8042,4 @@ flow_release_workspace(void *data)
+@@ -7108,5 +7077,4 @@ flow_release_workspace(void *data)
@@ -112 +113 @@
-@@ -8079,4 +8047,27 @@ flow_release_workspace(void *data)
+@@ -7114,4 +7082,27 @@ flow_release_workspace(void *data)
@@ -140 +141 @@
-@@ -8104,21 +8095,15 @@ static struct mlx5_flow_workspace*
+@@ -7139,22 +7130,15 @@ static struct mlx5_flow_workspace*
@@ -163 +164,2 @@
--	free(data->rss_desc.queue);
+-	if (data->rss_desc.queue)
+-		free(data->rss_desc.queue);
@@ -168 +170 @@
-@@ -8141,4 +8126,5 @@ mlx5_flow_push_thread_workspace(void)
+@@ -7177,4 +7161,5 @@ mlx5_flow_push_thread_workspace(void)
@@ -175 +177 @@
-index 003e7da3a6..62789853ab 100644
+index 8ad8e63205..8a571f977e 100644
@@ -178 +180 @@
-@@ -1497,8 +1497,8 @@ struct mlx5_flow_workspace {
+@@ -1099,8 +1099,8 @@ struct mlx5_flow_workspace {
@@ -188 +190 @@
-@@ -2023,4 +2023,6 @@ struct mlx5_flow_workspace *mlx5_flow_push_thread_workspace(void);
+@@ -1305,4 +1305,6 @@ struct mlx5_flow_workspace *mlx5_flow_push_thread_workspace(void);