DPDK patches and discussions
 help / color / mirror / Atom feed
From: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
To: <dev@dpdk.org>
Cc: <orika@nvidia.com>, <rasland@nvidia.com>, <matan@nvidia.com>,
	<stable@dpdk.org>
Subject: [PATCH 1/1] net/mlx5: fix device removal event handling
Date: Tue, 30 May 2023 18:13:28 +0300	[thread overview]
Message-ID: <20230530151328.29738-1-viacheslavo@nvidia.com> (raw)

On the device removal kernel notifies user space application
with queueing the IBV_DEVICE_FATAL_EVENT and triggering appropriate
file descriptor. Mellanox kernel driver stack emits this event
twice from different layers (mlx5 and uverbs). The IB port index
is not applicable in the event structure and should be ignored
for IBV_DEVICE_FATAL_EVENT events.

Also, on the older kernels (at least from OFED 4.9) there might be
race conditions causing the event queue close before application
fetches the IBV_DEVICE_FATAL_EVENT message with ibv_get_async_event()
API.

To provide the reliable device removal event detection the patch:

  - ignores the IB port index for the IBV_DEVICE_FATAL_EVENT
  - introduces the flag to notify PMD about removal only once
  - acks event with ibv_ack_async_event after actual handling
  - checks for EIO error, making sure queue is not closed yet

Fixes: 40d9f906f4e2 ("net/mlx5: fix device removal handler for multiport")
Cc: stable@dpdk.org

Signed-off-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
---
 drivers/net/mlx5/linux/mlx5_ethdev_os.c | 34 +++++++++++++++++--------
 drivers/net/mlx5/mlx5.h                 |  1 +
 2 files changed, 25 insertions(+), 10 deletions(-)

diff --git a/drivers/net/mlx5/linux/mlx5_ethdev_os.c b/drivers/net/mlx5/linux/mlx5_ethdev_os.c
index 55801534d1..639e629fe4 100644
--- a/drivers/net/mlx5/linux/mlx5_ethdev_os.c
+++ b/drivers/net/mlx5/linux/mlx5_ethdev_os.c
@@ -746,6 +746,7 @@ mlx5_dev_interrupt_device_fatal(struct mlx5_dev_ctx_shared *sh)
 
 	for (i = 0; i < sh->max_port; ++i) {
 		struct rte_eth_dev *dev;
+		struct mlx5_priv *priv;
 
 		if (sh->port[i].ih_port_id >= RTE_MAX_ETHPORTS) {
 			/*
@@ -756,9 +757,14 @@ mlx5_dev_interrupt_device_fatal(struct mlx5_dev_ctx_shared *sh)
 		}
 		dev = &rte_eth_devices[sh->port[i].ih_port_id];
 		MLX5_ASSERT(dev);
-		if (dev->data->dev_conf.intr_conf.rmv)
+		priv = dev->data->dev_private;
+		MLX5_ASSERT(priv);
+		if (!priv->rmv_notified && dev->data->dev_conf.intr_conf.rmv) {
+			/* Notify driver about removal only once. */
+			priv->rmv_notified = 1;
 			rte_eth_dev_callback_process
 				(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
+		}
 	}
 }
 
@@ -830,21 +836,29 @@ mlx5_dev_interrupt_handler(void *cb_arg)
 		struct rte_eth_dev *dev;
 		uint32_t tmp;
 
-		if (mlx5_glue->get_async_event(sh->cdev->ctx, &event))
+		if (mlx5_glue->get_async_event(sh->cdev->ctx, &event)) {
+			if (errno == EIO) {
+				DRV_LOG(DEBUG,
+					"IBV async event queue closed on: %s",
+					sh->ibdev_name);
+				mlx5_dev_interrupt_device_fatal(sh);
+			}
 			break;
-		/* Retrieve and check IB port index. */
-		tmp = (uint32_t)event.element.port_num;
-		if (!tmp && event.event_type == IBV_EVENT_DEVICE_FATAL) {
+		}
+		if (event.event_type == IBV_EVENT_DEVICE_FATAL) {
 			/*
-			 * The DEVICE_FATAL event is called once for
-			 * entire device without port specifying.
-			 * We should notify all existing ports.
+			 * The DEVICE_FATAL event can be called by kernel
+			 * twice - from mlx5 and uverbs layers, and port
+			 * index is not applicable. We should notify all
+			 * existing ports.
 			 */
-			mlx5_glue->ack_async_event(&event);
 			mlx5_dev_interrupt_device_fatal(sh);
+			mlx5_glue->ack_async_event(&event);
 			continue;
 		}
-		MLX5_ASSERT(tmp && (tmp <= sh->max_port));
+		/* Retrieve and check IB port index. */
+		tmp = (uint32_t)event.element.port_num;
+		MLX5_ASSERT(tmp <= sh->max_port);
 		if (!tmp) {
 			/* Unsupported device level event. */
 			mlx5_glue->ack_async_event(&event);
diff --git a/drivers/net/mlx5/mlx5.h b/drivers/net/mlx5/mlx5.h
index 021049ad2b..6aae8fe3f4 100644
--- a/drivers/net/mlx5/mlx5.h
+++ b/drivers/net/mlx5/mlx5.h
@@ -1743,6 +1743,7 @@ struct mlx5_priv {
 	unsigned int mtr_en:1; /* Whether support meter. */
 	unsigned int mtr_reg_share:1; /* Whether support meter REG_C share. */
 	unsigned int lb_used:1; /* Loopback queue is referred to. */
+	unsigned int rmv_notified:1; /* Notified about removal event */
 	uint32_t mark_enabled:1; /* If mark action is enabled on rxqs. */
 	uint32_t num_lag_ports:4; /* Number of ports can be bonded. */
 	uint16_t domain_id; /* Switch domain identifier. */
-- 
2.18.1


             reply	other threads:[~2023-05-30 15:14 UTC|newest]

Thread overview: 2+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2023-05-30 15:13 Viacheslav Ovsiienko [this message]
2023-06-19 12:14 ` Raslan Darawsheh

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20230530151328.29738-1-viacheslavo@nvidia.com \
    --to=viacheslavo@nvidia.com \
    --cc=dev@dpdk.org \
    --cc=matan@nvidia.com \
    --cc=orika@nvidia.com \
    --cc=rasland@nvidia.com \
    --cc=stable@dpdk.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
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).