net/mlx5: fix event handler uninstall
authorViacheslav Ovsiienko <viacheslavo@mellanox.com>
Mon, 27 May 2019 04:58:32 +0000 (04:58 +0000)
committerFerruh Yigit <ferruh.yigit@intel.com>
Thu, 6 Jun 2019 11:21:20 +0000 (20:21 +0900)
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")
Cc: stable@dpdk.org
Signed-off-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
Acked-by: Yongseok Koh <yskoh@mellanox.com>
drivers/net/mlx5/mlx5.c
drivers/net/mlx5/mlx5.h
drivers/net/mlx5/mlx5_ethdev.c

index aee9f97..3ef33c9 100644 (file)
@@ -313,7 +313,7 @@ mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
         **/
        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)
index caabdfc..01cfd65 100644 (file)
@@ -438,6 +438,8 @@ 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 */
 
index 7785341..ac0500a 100644 (file)
@@ -33,6 +33,7 @@
 #include <rte_malloc.h>
 #include <rte_string_fns.h>
 #include <rte_rwlock.h>
+#include <rte_cycles.h>
 
 #include "mlx5.h"
 #include "mlx5_glue.h"
@@ -1256,9 +1257,80 @@ mlx5_dev_interrupt_handler(void *cb_arg)
        }
 }
 
+/*
+ * 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
@@ -1284,7 +1356,7 @@ mlx5_dev_shared_handler_uninstall(struct rte_eth_dev *dev)
        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;
@@ -1293,8 +1365,8 @@ exit:
 }
 
 /**
- * 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