[dpdk-dev] [PATCH] net/mlx5: fix event handler uninstall

[Viacheslav Ovsiienko <viacheslavo@mellanox.com>]

When device is being closed and tries to unregister interrupt callback,
there is a chance the handler is still active (called in context of
eal_intr_thread_main thread). If so the rte_intr_callback_unregister
returns -EAGAIN and keeps the handler registered, causing crash when
underlaying resourse is gone away.

This race condition may happen if event handling in application takes
a long time. We should check the return code of unregistering routine
and try again to unregister the handler. The diagnostic messages are
shown once a second, while trying to unregister.

Fixes: 028b2a28c3cb ("net/mlx5: update event handler for multiport IB devices")

Signed-off-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
Acked-by: Yongseok Koh <yskoh@mellanox.com>
---
 drivers/net/mlx5/mlx5.c        |  2 +-
 drivers/net/mlx5/mlx5.h        |  2 ++
 drivers/net/mlx5/mlx5_ethdev.c | 79 +++++++++++++++++++++++++++++++++++++++---
 3 files changed, 78 insertions(+), 5 deletions(-)

diff --git a/drivers/net/mlx5/mlx5.c b/drivers/net/mlx5/mlx5.c
index 9f5ec97..2344cb4 100644
--- a/drivers/net/mlx5/mlx5.c
+++ b/drivers/net/mlx5/mlx5.c
@@ -313,7 +313,7 @@ struct mlx5_dev_spawn_data {
 	 **/
 	assert(!sh->intr_cnt);
 	if (sh->intr_cnt)
-		rte_intr_callback_unregister
+		mlx5_intr_callback_unregister
 			(&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
 	pthread_mutex_destroy(&sh->intr_mutex);
 	if (sh->pd)
diff --git a/drivers/net/mlx5/mlx5.h b/drivers/net/mlx5/mlx5.h
index 3eaaafd..5b5b93d 100644
--- a/drivers/net/mlx5/mlx5.h
+++ b/drivers/net/mlx5/mlx5.h
@@ -437,6 +437,8 @@ void mlx5_nl_check_switch_info(bool nun_vf_set,
 			       struct mlx5_switch_info *switch_info);
 void mlx5_translate_port_name(const char *port_name_in,
 			      struct mlx5_switch_info *port_info_out);
+void mlx5_intr_callback_unregister(const struct rte_intr_handle *handle,
+				   rte_intr_callback_fn cb_fn, void *cb_arg);
 
 /* mlx5_mac.c */
 
diff --git a/drivers/net/mlx5/mlx5_ethdev.c b/drivers/net/mlx5/mlx5_ethdev.c
index a8a7ece..f47297c 100644
--- a/drivers/net/mlx5/mlx5_ethdev.c
+++ b/drivers/net/mlx5/mlx5_ethdev.c
@@ -1226,9 +1226,80 @@ int mlx5_fw_version_get(struct rte_eth_dev *dev, char *fw_ver, size_t fw_size)
 	}
 }
 
+/*
+ * Unregister callback handler safely. The handler may be active
+ * while we are trying to unregister it, in this case code -EAGAIN
+ * is returned by rte_intr_callback_unregister(). This routine checks
+ * the return code and tries to unregister handler again.
+ *
+ * @param handle
+ *   interrupt handle
+ * @param cb_fn
+ *   pointer to callback routine
+ * @cb_arg
+ *   opaque callback parameter
+ */
+void
+mlx5_intr_callback_unregister(const struct rte_intr_handle *handle,
+			      rte_intr_callback_fn cb_fn, void *cb_arg)
+{
+	/*
+	 * Try to reduce timeout management overhead by not calling
+	 * the timer related routines on the first iteration. If the
+	 * unregistering succeeds on first call there will be no
+	 * timer calls at all.
+	 */
+	uint64_t twait = 0;
+	uint64_t start = 0;
+
+	do {
+		int ret;
+
+		ret = rte_intr_callback_unregister(handle, cb_fn, cb_arg);
+		if (ret >= 0)
+			return;
+		if (ret != -EAGAIN) {
+			DRV_LOG(INFO, "failed to unregister interrupt"
+				      " handler (error: %d)", ret);
+			assert(false);
+			return;
+		}
+		if (twait) {
+			struct timespec onems;
+
+			/* Wait one millisecond and try again. */
+			onems.tv_sec = 0;
+			onems.tv_nsec = NS_PER_S / MS_PER_S;
+			nanosleep(&onems, 0);
+			/* Check whether one second elapsed. */
+			if ((rte_get_timer_cycles() - start) <= twait)
+				continue;
+		} else {
+			/*
+			 * We get the amount of timer ticks for one second.
+			 * If this amount elapsed it means we spent one
+			 * second in waiting. This branch is executed once
+			 * on first iteration.
+			 */
+			twait = rte_get_timer_hz();
+			assert(twait);
+		}
+		/*
+		 * Timeout elapsed, show message (once a second) and retry.
+		 * We have no other acceptable option here, if we ignore
+		 * the unregistering return code the handler will not
+		 * be unregistered, fd will be closed and we may get the
+		 * crush. Hanging and messaging in the loop seems not to be
+		 * the worst choice.
+		 */
+		DRV_LOG(INFO, "Retrying to unregister interrupt handler");
+		start = rte_get_timer_cycles();
+	} while (true);
+}
+
 /**
  * Uninstall shared asynchronous device events handler.
- * This function is implemeted to support event sharing
+ * This function is implemented to support event sharing
  * between multiple ports of single IB device.
  *
  * @param dev
@@ -1254,7 +1325,7 @@ int mlx5_fw_version_get(struct rte_eth_dev *dev, char *fw_ver, size_t fw_size)
 	sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
 	if (!sh->intr_cnt || --sh->intr_cnt)
 		goto exit;
-	rte_intr_callback_unregister(&sh->intr_handle,
+	mlx5_intr_callback_unregister(&sh->intr_handle,
 				     mlx5_dev_interrupt_handler, sh);
 	sh->intr_handle.fd = 0;
 	sh->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
@@ -1263,8 +1334,8 @@ int mlx5_fw_version_get(struct rte_eth_dev *dev, char *fw_ver, size_t fw_size)
 }
 
 /**
- * Install shared asyncronous device events handler.
- * This function is implemeted to support event sharing
+ * Install shared asynchronous device events handler.
+ * This function is implemented to support event sharing
  * between multiple ports of single IB device.
  *
  * @param dev
-- 
1.8.3.1