net/failsafe: fix hotplug races
authorMatan Azrad <matan@mellanox.com>
Mon, 12 Feb 2018 20:51:42 +0000 (20:51 +0000)
committerThomas Monjalon <thomas@monjalon.net>
Tue, 13 Feb 2018 17:17:30 +0000 (18:17 +0100)
Fail-safe uses a periodic alarm mechanism, running from the host
thread, to manage the hot-plug events of its sub-devices. This
management requires a lot of sub-devices PMDs operations
(stop, close, start, configure, etc.).

While the hot-plug alarm runs in the host thread, the application may
call fail-safe operations, which directly trigger the sub-devices PMDs
operations as well. This call may occur from any thread decided by the
application (probably the master thread).

Thus, more than one operation can be executed to a sub-device at the
same time. This can initiate a lot of races in the sub-PMDs.

Moreover, some control operations update the fail-safe internal
databases, which can be used by the alarm mechanism at the same time.
This can also initiate races and crashes.

Fail-safe is the owner of its sub-devices and must synchronize their
use according to the ETHDEV ownership rules.

Synchronize hot-plug management by a new lock mechanism uses a mutex to
atomically defend each critical section in the fail-safe hot-plug
mechanism and control operations to prevent any races between them.

Fixes: a46f8d5 ("net/failsafe: add fail-safe PMD")
Cc: stable@dpdk.org
Signed-off-by: Matan Azrad <matan@mellanox.com>
Acked-by: Gaetan Rivet <gaetan.rivet@6wind.com>
drivers/net/failsafe/Makefile
drivers/net/failsafe/failsafe.c
drivers/net/failsafe/failsafe_ether.c
drivers/net/failsafe/failsafe_flow.c
drivers/net/failsafe/failsafe_ops.c
drivers/net/failsafe/failsafe_private.h

index d1ae899..bd2f019 100644 (file)
@@ -68,5 +68,6 @@ CFLAGS += -pedantic
 LDLIBS += -lrte_eal -lrte_mbuf -lrte_mempool -lrte_ring
 LDLIBS += -lrte_ethdev -lrte_net -lrte_kvargs
 LDLIBS += -lrte_bus_vdev
+LDLIBS += -lpthread
 
 include $(RTE_SDK)/mk/rte.lib.mk
index 7b2cdbb..c499bfb 100644 (file)
@@ -113,16 +113,45 @@ fs_hotplug_alarm(void *arg)
                        break;
        /* if we have non-probed device */
        if (i != PRIV(dev)->subs_tail) {
+               if (fs_lock(dev, 1) != 0)
+                       goto reinstall;
                ret = failsafe_eth_dev_state_sync(dev);
+               fs_unlock(dev, 1);
                if (ret)
                        ERROR("Unable to synchronize sub_device state");
        }
        failsafe_dev_remove(dev);
+reinstall:
        ret = failsafe_hotplug_alarm_install(dev);
        if (ret)
                ERROR("Unable to set up next alarm");
 }
 
+static int
+fs_mutex_init(struct fs_priv *priv)
+{
+       int ret;
+       pthread_mutexattr_t attr;
+
+       ret = pthread_mutexattr_init(&attr);
+       if (ret) {
+               ERROR("Cannot initiate mutex attributes - %s", strerror(ret));
+               return ret;
+       }
+       /* Allow mutex relocks for the thread holding the mutex. */
+       ret = pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE);
+       if (ret) {
+               ERROR("Cannot set mutex type - %s", strerror(ret));
+               return ret;
+       }
+       ret = pthread_mutex_init(&priv->hotplug_mutex, &attr);
+       if (ret) {
+               ERROR("Cannot initiate mutex - %s", strerror(ret));
+               return ret;
+       }
+       return 0;
+}
+
 static int
 fs_eth_dev_create(struct rte_vdev_device *vdev)
 {
@@ -174,6 +203,9 @@ fs_eth_dev_create(struct rte_vdev_device *vdev)
        snprintf(priv->my_owner.name, sizeof(priv->my_owner.name),
                 FAILSAFE_OWNER_NAME);
        ret = failsafe_eal_init(dev);
+       if (ret)
+               goto free_args;
+       ret = fs_mutex_init(priv);
        if (ret)
                goto free_args;
        ret = failsafe_hotplug_alarm_install(dev);
@@ -250,6 +282,9 @@ fs_rte_eth_free(const char *name)
                ERROR("Error while uninitializing sub-EAL");
        failsafe_args_free(dev);
        fs_sub_device_free(dev);
+       ret = pthread_mutex_destroy(&PRIV(dev)->hotplug_mutex);
+       if (ret)
+               ERROR("Error while destroying hotplug mutex");
        rte_free(PRIV(dev));
        rte_eth_dev_release_port(dev);
        return ret;
index d820faf..2c0bf93 100644 (file)
@@ -328,8 +328,11 @@ failsafe_dev_remove(struct rte_eth_dev *dev)
 
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE)
                if (sdev->remove && fs_rxtx_clean(sdev)) {
+                       if (fs_lock(dev, 1) != 0)
+                               return;
                        fs_dev_stats_save(sdev);
                        fs_dev_remove(sdev);
+                       fs_unlock(dev, 1);
                }
 }
 
@@ -429,6 +432,7 @@ failsafe_eth_rmv_event_callback(uint16_t port_id __rte_unused,
 {
        struct sub_device *sdev = cb_arg;
 
+       fs_lock(sdev->fs_dev, 0);
        /* Switch as soon as possible tx_dev. */
        fs_switch_dev(sdev->fs_dev, sdev);
        /* Use safe bursts in any case. */
@@ -438,6 +442,7 @@ failsafe_eth_rmv_event_callback(uint16_t port_id __rte_unused,
         * the callback at the source of the current thread context.
         */
        sdev->remove = 1;
+       fs_unlock(sdev->fs_dev, 0);
        return 0;
 }
 
index 4d18e8e..ec8c909 100644 (file)
@@ -55,6 +55,7 @@ fs_flow_validate(struct rte_eth_dev *dev,
        uint8_t i;
        int ret;
 
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
                DEBUG("Calling rte_flow_validate on sub_device %d", i);
                ret = rte_flow_validate(PORT_ID(sdev),
@@ -62,9 +63,11 @@ fs_flow_validate(struct rte_eth_dev *dev,
                if ((ret = fs_err(sdev, ret))) {
                        ERROR("Operation rte_flow_validate failed for sub_device %d"
                              " with error %d", i, ret);
+                       fs_unlock(dev, 0);
                        return ret;
                }
        }
+       fs_unlock(dev, 0);
        return 0;
 }
 
@@ -79,6 +82,7 @@ fs_flow_create(struct rte_eth_dev *dev,
        struct rte_flow *flow;
        uint8_t i;
 
+       fs_lock(dev, 0);
        flow = fs_flow_allocate(attr, patterns, actions);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
                flow->flows[i] = rte_flow_create(PORT_ID(sdev),
@@ -90,6 +94,7 @@ fs_flow_create(struct rte_eth_dev *dev,
                }
        }
        TAILQ_INSERT_TAIL(&PRIV(dev)->flow_list, flow, next);
+       fs_unlock(dev, 0);
        return flow;
 err:
        FOREACH_SUBDEV(sdev, i, dev) {
@@ -98,6 +103,7 @@ err:
                                flow->flows[i], error);
        }
        fs_flow_release(&flow);
+       fs_unlock(dev, 0);
        return NULL;
 }
 
@@ -115,6 +121,7 @@ fs_flow_destroy(struct rte_eth_dev *dev,
                return -EINVAL;
        }
        ret = 0;
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
                int local_ret;
 
@@ -131,6 +138,7 @@ fs_flow_destroy(struct rte_eth_dev *dev,
        }
        TAILQ_REMOVE(&PRIV(dev)->flow_list, flow, next);
        fs_flow_release(&flow);
+       fs_unlock(dev, 0);
        return ret;
 }
 
@@ -144,12 +152,14 @@ fs_flow_flush(struct rte_eth_dev *dev,
        uint8_t i;
        int ret;
 
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
                DEBUG("Calling rte_flow_flush on sub_device %d", i);
                ret = rte_flow_flush(PORT_ID(sdev), error);
                if ((ret = fs_err(sdev, ret))) {
                        ERROR("Operation rte_flow_flush failed for sub_device %d"
                              " with error %d", i, ret);
+                       fs_unlock(dev, 0);
                        return ret;
                }
        }
@@ -157,6 +167,7 @@ fs_flow_flush(struct rte_eth_dev *dev,
                TAILQ_REMOVE(&PRIV(dev)->flow_list, flow, next);
                fs_flow_release(&flow);
        }
+       fs_unlock(dev, 0);
        return 0;
 }
 
@@ -169,15 +180,19 @@ fs_flow_query(struct rte_eth_dev *dev,
 {
        struct sub_device *sdev;
 
+       fs_lock(dev, 0);
        sdev = TX_SUBDEV(dev);
        if (sdev != NULL) {
                int ret = rte_flow_query(PORT_ID(sdev),
                                         flow->flows[SUB_ID(sdev)],
                                         type, arg, error);
 
-               if ((ret = fs_err(sdev, ret)))
+               if ((ret = fs_err(sdev, ret))) {
+                       fs_unlock(dev, 0);
                        return ret;
+               }
        }
+       fs_unlock(dev, 0);
        WARN("No active sub_device to query about its flow");
        return -1;
 }
@@ -191,6 +206,7 @@ fs_flow_isolate(struct rte_eth_dev *dev,
        uint8_t i;
        int ret;
 
+       fs_lock(dev, 0);
        FOREACH_SUBDEV(sdev, i, dev) {
                if (sdev->state < DEV_PROBED)
                        continue;
@@ -202,11 +218,13 @@ fs_flow_isolate(struct rte_eth_dev *dev,
                if ((ret = fs_err(sdev, ret))) {
                        ERROR("Operation rte_flow_isolate failed for sub_device %d"
                              " with error %d", i, ret);
+                       fs_unlock(dev, 0);
                        return ret;
                }
                sdev->flow_isolated = set;
        }
        PRIV(dev)->flow_isolated = set;
+       fs_unlock(dev, 0);
        return 0;
 }
 
index f0e48c1..fe64c68 100644 (file)
@@ -94,6 +94,7 @@ fs_dev_configure(struct rte_eth_dev *dev)
        uint8_t i;
        int ret;
 
+       fs_lock(dev, 0);
        supp_tx_offloads = PRIV(dev)->infos.tx_offload_capa;
        tx_offloads = dev->data->dev_conf.txmode.offloads;
        if ((tx_offloads & supp_tx_offloads) != tx_offloads) {
@@ -101,6 +102,7 @@ fs_dev_configure(struct rte_eth_dev *dev)
                ERROR("Some Tx offloads are not supported, "
                      "requested 0x%" PRIx64 " supported 0x%" PRIx64,
                      tx_offloads, supp_tx_offloads);
+               fs_unlock(dev, 0);
                return -rte_errno;
        }
        FOREACH_SUBDEV(sdev, i, dev) {
@@ -139,6 +141,7 @@ fs_dev_configure(struct rte_eth_dev *dev)
                        if (!fs_err(sdev, ret))
                                continue;
                        ERROR("Could not configure sub_device %d", i);
+                       fs_unlock(dev, 0);
                        return ret;
                }
                if (rmv_interrupt) {
@@ -165,6 +168,7 @@ fs_dev_configure(struct rte_eth_dev *dev)
        }
        if (PRIV(dev)->state < DEV_ACTIVE)
                PRIV(dev)->state = DEV_ACTIVE;
+       fs_unlock(dev, 0);
        return 0;
 }
 
@@ -175,9 +179,12 @@ fs_dev_start(struct rte_eth_dev *dev)
        uint8_t i;
        int ret;
 
+       fs_lock(dev, 0);
        ret = failsafe_rx_intr_install(dev);
-       if (ret)
+       if (ret) {
+               fs_unlock(dev, 0);
                return ret;
+       }
        FOREACH_SUBDEV(sdev, i, dev) {
                if (sdev->state != DEV_ACTIVE)
                        continue;
@@ -186,6 +193,7 @@ fs_dev_start(struct rte_eth_dev *dev)
                if (ret) {
                        if (!fs_err(sdev, ret))
                                continue;
+                       fs_unlock(dev, 0);
                        return ret;
                }
                ret = failsafe_rx_intr_install_subdevice(sdev);
@@ -193,6 +201,7 @@ fs_dev_start(struct rte_eth_dev *dev)
                        if (!fs_err(sdev, ret))
                                continue;
                        rte_eth_dev_stop(PORT_ID(sdev));
+                       fs_unlock(dev, 0);
                        return ret;
                }
                sdev->state = DEV_STARTED;
@@ -200,6 +209,7 @@ fs_dev_start(struct rte_eth_dev *dev)
        if (PRIV(dev)->state < DEV_STARTED)
                PRIV(dev)->state = DEV_STARTED;
        fs_switch_dev(dev, NULL);
+       fs_unlock(dev, 0);
        return 0;
 }
 
@@ -209,6 +219,7 @@ fs_dev_stop(struct rte_eth_dev *dev)
        struct sub_device *sdev;
        uint8_t i;
 
+       fs_lock(dev, 0);
        PRIV(dev)->state = DEV_STARTED - 1;
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_STARTED) {
                rte_eth_dev_stop(PORT_ID(sdev));
@@ -216,6 +227,7 @@ fs_dev_stop(struct rte_eth_dev *dev)
                sdev->state = DEV_STARTED - 1;
        }
        failsafe_rx_intr_uninstall(dev);
+       fs_unlock(dev, 0);
 }
 
 static int
@@ -225,15 +237,18 @@ fs_dev_set_link_up(struct rte_eth_dev *dev)
        uint8_t i;
        int ret;
 
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
                DEBUG("Calling rte_eth_dev_set_link_up on sub_device %d", i);
                ret = rte_eth_dev_set_link_up(PORT_ID(sdev));
                if ((ret = fs_err(sdev, ret))) {
                        ERROR("Operation rte_eth_dev_set_link_up failed for sub_device %d"
                              " with error %d", i, ret);
+                       fs_unlock(dev, 0);
                        return ret;
                }
        }
+       fs_unlock(dev, 0);
        return 0;
 }
 
@@ -244,15 +259,18 @@ fs_dev_set_link_down(struct rte_eth_dev *dev)
        uint8_t i;
        int ret;
 
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
                DEBUG("Calling rte_eth_dev_set_link_down on sub_device %d", i);
                ret = rte_eth_dev_set_link_down(PORT_ID(sdev));
                if ((ret = fs_err(sdev, ret))) {
                        ERROR("Operation rte_eth_dev_set_link_down failed for sub_device %d"
                              " with error %d", i, ret);
+                       fs_unlock(dev, 0);
                        return ret;
                }
        }
+       fs_unlock(dev, 0);
        return 0;
 }
 
@@ -263,6 +281,7 @@ fs_dev_close(struct rte_eth_dev *dev)
        struct sub_device *sdev;
        uint8_t i;
 
+       fs_lock(dev, 0);
        failsafe_hotplug_alarm_cancel(dev);
        if (PRIV(dev)->state == DEV_STARTED)
                dev->dev_ops->dev_stop(dev);
@@ -273,6 +292,7 @@ fs_dev_close(struct rte_eth_dev *dev)
                sdev->state = DEV_ACTIVE - 1;
        }
        fs_dev_free_queues(dev);
+       fs_unlock(dev, 0);
 }
 
 static bool
@@ -305,14 +325,16 @@ fs_rx_queue_release(void *queue)
        if (queue == NULL)
                return;
        rxq = queue;
+       dev = rxq->priv->dev;
+       fs_lock(dev, 0);
        if (rxq->event_fd > 0)
                close(rxq->event_fd);
-       dev = rxq->priv->dev;
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE)
                SUBOPS(sdev, rx_queue_release)
                        (ETH(sdev)->data->rx_queues[rxq->qid]);
        dev->data->rx_queues[rxq->qid] = NULL;
        rte_free(rxq);
+       fs_unlock(dev, 0);
 }
 
 static int
@@ -338,6 +360,7 @@ fs_rx_queue_setup(struct rte_eth_dev *dev,
        uint8_t i;
        int ret;
 
+       fs_lock(dev, 0);
        rxq = dev->data->rx_queues[rx_queue_id];
        if (rxq != NULL) {
                fs_rx_queue_release(rxq);
@@ -353,14 +376,17 @@ fs_rx_queue_setup(struct rte_eth_dev *dev,
                      dev->data->dev_conf.rxmode.offloads,
                      PRIV(dev)->infos.rx_offload_capa |
                      PRIV(dev)->infos.rx_queue_offload_capa);
+               fs_unlock(dev, 0);
                return -rte_errno;
        }
        rxq = rte_zmalloc(NULL,
                          sizeof(*rxq) +
                          sizeof(rte_atomic64_t) * PRIV(dev)->subs_tail,
                          RTE_CACHE_LINE_SIZE);
-       if (rxq == NULL)
+       if (rxq == NULL) {
+               fs_unlock(dev, 0);
                return -ENOMEM;
+       }
        FOREACH_SUBDEV(sdev, i, dev)
                rte_atomic64_init(&rxq->refcnt[i]);
        rxq->qid = rx_queue_id;
@@ -371,8 +397,10 @@ fs_rx_queue_setup(struct rte_eth_dev *dev,
        rxq->priv = PRIV(dev);
        rxq->sdev = PRIV(dev)->subs;
        ret = rte_intr_efd_enable(&intr_handle, 1);
-       if (ret < 0)
+       if (ret < 0) {
+               fs_unlock(dev, 0);
                return ret;
+       }
        rxq->event_fd = intr_handle.efds[0];
        dev->data->rx_queues[rx_queue_id] = rxq;
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
@@ -385,9 +413,11 @@ fs_rx_queue_setup(struct rte_eth_dev *dev,
                        goto free_rxq;
                }
        }
+       fs_unlock(dev, 0);
        return 0;
 free_rxq:
        fs_rx_queue_release(rxq);
+       fs_unlock(dev, 0);
        return ret;
 }
 
@@ -400,20 +430,21 @@ fs_rx_intr_enable(struct rte_eth_dev *dev, uint16_t idx)
        int ret;
        int rc = 0;
 
+       fs_lock(dev, 0);
        if (idx >= dev->data->nb_rx_queues) {
-               rte_errno = EINVAL;
-               return -rte_errno;
+               rc = -EINVAL;
+               goto unlock;
        }
        rxq = dev->data->rx_queues[idx];
        if (rxq == NULL || rxq->event_fd <= 0) {
-               rte_errno = EINVAL;
-               return -rte_errno;
+               rc = -EINVAL;
+               goto unlock;
        }
        /* Fail if proxy service is nor running. */
        if (PRIV(dev)->rxp.sstate != SS_RUNNING) {
                ERROR("failsafe interrupt services are not running");
-               rte_errno = EAGAIN;
-               return -rte_errno;
+               rc = -EAGAIN;
+               goto unlock;
        }
        rxq->enable_events = 1;
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
@@ -422,6 +453,8 @@ fs_rx_intr_enable(struct rte_eth_dev *dev, uint16_t idx)
                if (ret)
                        rc = ret;
        }
+unlock:
+       fs_unlock(dev, 0);
        if (rc)
                rte_errno = -rc;
        return rc;
@@ -437,14 +470,15 @@ fs_rx_intr_disable(struct rte_eth_dev *dev, uint16_t idx)
        int rc = 0;
        int ret;
 
+       fs_lock(dev, 0);
        if (idx >= dev->data->nb_rx_queues) {
-               rte_errno = EINVAL;
-               return -rte_errno;
+               rc = -EINVAL;
+               goto unlock;
        }
        rxq = dev->data->rx_queues[idx];
        if (rxq == NULL || rxq->event_fd <= 0) {
-               rte_errno = EINVAL;
-               return -rte_errno;
+               rc = -EINVAL;
+               goto unlock;
        }
        rxq->enable_events = 0;
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
@@ -456,6 +490,8 @@ fs_rx_intr_disable(struct rte_eth_dev *dev, uint16_t idx)
        /* Clear pending events */
        while (read(rxq->event_fd, &u64, sizeof(uint64_t)) >  0)
                ;
+unlock:
+       fs_unlock(dev, 0);
        if (rc)
                rte_errno = -rc;
        return rc;
@@ -492,11 +528,13 @@ fs_tx_queue_release(void *queue)
                return;
        txq = queue;
        dev = txq->priv->dev;
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE)
                SUBOPS(sdev, tx_queue_release)
                        (ETH(sdev)->data->tx_queues[txq->qid]);
        dev->data->tx_queues[txq->qid] = NULL;
        rte_free(txq);
+       fs_unlock(dev, 0);
 }
 
 static int
@@ -511,6 +549,7 @@ fs_tx_queue_setup(struct rte_eth_dev *dev,
        uint8_t i;
        int ret;
 
+       fs_lock(dev, 0);
        txq = dev->data->tx_queues[tx_queue_id];
        if (txq != NULL) {
                fs_tx_queue_release(txq);
@@ -531,14 +570,17 @@ fs_tx_queue_setup(struct rte_eth_dev *dev,
                      dev->data->dev_conf.txmode.offloads,
                      PRIV(dev)->infos.tx_offload_capa |
                      PRIV(dev)->infos.tx_queue_offload_capa);
+               fs_unlock(dev, 0);
                return -rte_errno;
        }
        txq = rte_zmalloc("ethdev TX queue",
                          sizeof(*txq) +
                          sizeof(rte_atomic64_t) * PRIV(dev)->subs_tail,
                          RTE_CACHE_LINE_SIZE);
-       if (txq == NULL)
+       if (txq == NULL) {
+               fs_unlock(dev, 0);
                return -ENOMEM;
+       }
        FOREACH_SUBDEV(sdev, i, dev)
                rte_atomic64_init(&txq->refcnt[i]);
        txq->qid = tx_queue_id;
@@ -557,9 +599,11 @@ fs_tx_queue_setup(struct rte_eth_dev *dev,
                        goto free_txq;
                }
        }
+       fs_unlock(dev, 0);
        return 0;
 free_txq:
        fs_tx_queue_release(txq);
+       fs_unlock(dev, 0);
        return ret;
 }
 
@@ -586,8 +630,10 @@ fs_promiscuous_enable(struct rte_eth_dev *dev)
        struct sub_device *sdev;
        uint8_t i;
 
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE)
                rte_eth_promiscuous_enable(PORT_ID(sdev));
+       fs_unlock(dev, 0);
 }
 
 static void
@@ -596,8 +642,10 @@ fs_promiscuous_disable(struct rte_eth_dev *dev)
        struct sub_device *sdev;
        uint8_t i;
 
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE)
                rte_eth_promiscuous_disable(PORT_ID(sdev));
+       fs_unlock(dev, 0);
 }
 
 static void
@@ -606,8 +654,10 @@ fs_allmulticast_enable(struct rte_eth_dev *dev)
        struct sub_device *sdev;
        uint8_t i;
 
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE)
                rte_eth_allmulticast_enable(PORT_ID(sdev));
+       fs_unlock(dev, 0);
 }
 
 static void
@@ -616,8 +666,10 @@ fs_allmulticast_disable(struct rte_eth_dev *dev)
        struct sub_device *sdev;
        uint8_t i;
 
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE)
                rte_eth_allmulticast_disable(PORT_ID(sdev));
+       fs_unlock(dev, 0);
 }
 
 static int
@@ -628,6 +680,7 @@ fs_link_update(struct rte_eth_dev *dev,
        uint8_t i;
        int ret;
 
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
                DEBUG("Calling link_update on sub_device %d", i);
                ret = (SUBOPS(sdev, link_update))(ETH(sdev), wait_to_complete);
@@ -635,6 +688,7 @@ fs_link_update(struct rte_eth_dev *dev,
                    rte_eth_dev_is_removed(PORT_ID(sdev)) == 0) {
                        ERROR("Link update failed for sub_device %d with error %d",
                              i, ret);
+                       fs_unlock(dev, 0);
                        return ret;
                }
        }
@@ -646,9 +700,11 @@ fs_link_update(struct rte_eth_dev *dev,
                l2 = &ETH(TX_SUBDEV(dev))->data->dev_link;
                if (memcmp(l1, l2, sizeof(*l1))) {
                        *l1 = *l2;
+                       fs_unlock(dev, 0);
                        return 0;
                }
        }
+       fs_unlock(dev, 0);
        return -1;
 }
 
@@ -661,6 +717,7 @@ fs_stats_get(struct rte_eth_dev *dev,
        uint8_t i;
        int ret;
 
+       fs_lock(dev, 0);
        rte_memcpy(stats, &PRIV(dev)->stats_accumulator, sizeof(*stats));
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
                struct rte_eth_stats *snapshot = &sdev->stats_snapshot.stats;
@@ -676,12 +733,14 @@ fs_stats_get(struct rte_eth_dev *dev,
                        ERROR("Operation rte_eth_stats_get failed for sub_device %d with error %d",
                                  i, ret);
                        *timestamp = 0;
+                       fs_unlock(dev, 0);
                        return ret;
                }
                *timestamp = rte_rdtsc();
 inc:
                failsafe_stats_increment(stats, snapshot);
        }
+       fs_unlock(dev, 0);
        return 0;
 }
 
@@ -691,11 +750,13 @@ fs_stats_reset(struct rte_eth_dev *dev)
        struct sub_device *sdev;
        uint8_t i;
 
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
                rte_eth_stats_reset(PORT_ID(sdev));
                memset(&sdev->stats_snapshot, 0, sizeof(struct rte_eth_stats));
        }
        memset(&PRIV(dev)->stats_accumulator, 0, sizeof(struct rte_eth_stats));
+       fs_unlock(dev, 0);
 }
 
 /**
@@ -771,14 +832,20 @@ fs_dev_supported_ptypes_get(struct rte_eth_dev *dev)
 {
        struct sub_device *sdev;
        struct rte_eth_dev *edev;
+       const uint32_t *ret;
 
+       fs_lock(dev, 0);
        sdev = TX_SUBDEV(dev);
-       if (sdev == NULL)
-               return NULL;
+       if (sdev == NULL) {
+               ret = NULL;
+               goto unlock;
+       }
        edev = ETH(sdev);
        /* ENOTSUP: counts as no supported ptypes */
-       if (SUBOPS(sdev, dev_supported_ptypes_get) == NULL)
-               return NULL;
+       if (SUBOPS(sdev, dev_supported_ptypes_get) == NULL) {
+               ret = NULL;
+               goto unlock;
+       }
        /*
         * The API does not permit to do a clean AND of all ptypes,
         * It is also incomplete by design and we do not really care
@@ -786,7 +853,10 @@ fs_dev_supported_ptypes_get(struct rte_eth_dev *dev)
         * We just return the ptypes of the device of highest
         * priority, usually the PREFERRED device.
         */
-       return SUBOPS(sdev, dev_supported_ptypes_get)(edev);
+       ret = SUBOPS(sdev, dev_supported_ptypes_get)(edev);
+unlock:
+       fs_unlock(dev, 0);
+       return ret;
 }
 
 static int
@@ -796,15 +866,18 @@ fs_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
        uint8_t i;
        int ret;
 
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
                DEBUG("Calling rte_eth_dev_set_mtu on sub_device %d", i);
                ret = rte_eth_dev_set_mtu(PORT_ID(sdev), mtu);
                if ((ret = fs_err(sdev, ret))) {
                        ERROR("Operation rte_eth_dev_set_mtu failed for sub_device %d with error %d",
                              i, ret);
+                       fs_unlock(dev, 0);
                        return ret;
                }
        }
+       fs_unlock(dev, 0);
        return 0;
 }
 
@@ -815,15 +888,18 @@ fs_vlan_filter_set(struct rte_eth_dev *dev, uint16_t vlan_id, int on)
        uint8_t i;
        int ret;
 
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
                DEBUG("Calling rte_eth_dev_vlan_filter on sub_device %d", i);
                ret = rte_eth_dev_vlan_filter(PORT_ID(sdev), vlan_id, on);
                if ((ret = fs_err(sdev, ret))) {
                        ERROR("Operation rte_eth_dev_vlan_filter failed for sub_device %d"
                              " with error %d", i, ret);
+                       fs_unlock(dev, 0);
                        return ret;
                }
        }
+       fs_unlock(dev, 0);
        return 0;
 }
 
@@ -832,13 +908,22 @@ fs_flow_ctrl_get(struct rte_eth_dev *dev,
                struct rte_eth_fc_conf *fc_conf)
 {
        struct sub_device *sdev;
+       int ret;
 
+       fs_lock(dev, 0);
        sdev = TX_SUBDEV(dev);
-       if (sdev == NULL)
-               return 0;
-       if (SUBOPS(sdev, flow_ctrl_get) == NULL)
-               return -ENOTSUP;
-       return SUBOPS(sdev, flow_ctrl_get)(ETH(sdev), fc_conf);
+       if (sdev == NULL) {
+               ret = 0;
+               goto unlock;
+       }
+       if (SUBOPS(sdev, flow_ctrl_get) == NULL) {
+               ret = -ENOTSUP;
+               goto unlock;
+       }
+       ret = SUBOPS(sdev, flow_ctrl_get)(ETH(sdev), fc_conf);
+unlock:
+       fs_unlock(dev, 0);
+       return ret;
 }
 
 static int
@@ -849,15 +934,18 @@ fs_flow_ctrl_set(struct rte_eth_dev *dev,
        uint8_t i;
        int ret;
 
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
                DEBUG("Calling rte_eth_dev_flow_ctrl_set on sub_device %d", i);
                ret = rte_eth_dev_flow_ctrl_set(PORT_ID(sdev), fc_conf);
                if ((ret = fs_err(sdev, ret))) {
                        ERROR("Operation rte_eth_dev_flow_ctrl_set failed for sub_device %d"
                              " with error %d", i, ret);
+                       fs_unlock(dev, 0);
                        return ret;
                }
        }
+       fs_unlock(dev, 0);
        return 0;
 }
 
@@ -867,6 +955,7 @@ fs_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
        struct sub_device *sdev;
        uint8_t i;
 
+       fs_lock(dev, 0);
        /* No check: already done within the rte_eth_dev_mac_addr_remove
         * call for the fail-safe device.
         */
@@ -874,6 +963,7 @@ fs_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
                rte_eth_dev_mac_addr_remove(PORT_ID(sdev),
                                &dev->data->mac_addrs[index]);
        PRIV(dev)->mac_addr_pool[index] = 0;
+       fs_unlock(dev, 0);
 }
 
 static int
@@ -887,11 +977,13 @@ fs_mac_addr_add(struct rte_eth_dev *dev,
        uint8_t i;
 
        RTE_ASSERT(index < FAILSAFE_MAX_ETHADDR);
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
                ret = rte_eth_dev_mac_addr_add(PORT_ID(sdev), mac_addr, vmdq);
                if ((ret = fs_err(sdev, ret))) {
                        ERROR("Operation rte_eth_dev_mac_addr_add failed for sub_device %"
                              PRIu8 " with error %d", i, ret);
+                       fs_unlock(dev, 0);
                        return ret;
                }
        }
@@ -900,6 +992,7 @@ fs_mac_addr_add(struct rte_eth_dev *dev,
                PRIV(dev)->nb_mac_addr = index;
        }
        PRIV(dev)->mac_addr_pool[index] = vmdq;
+       fs_unlock(dev, 0);
        return 0;
 }
 
@@ -909,8 +1002,10 @@ fs_mac_addr_set(struct rte_eth_dev *dev, struct ether_addr *mac_addr)
        struct sub_device *sdev;
        uint8_t i;
 
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE)
                rte_eth_dev_default_mac_addr_set(PORT_ID(sdev), mac_addr);
+       fs_unlock(dev, 0);
 }
 
 static int
@@ -928,15 +1023,18 @@ fs_filter_ctrl(struct rte_eth_dev *dev,
                *(const void **)arg = &fs_flow_ops;
                return 0;
        }
+       fs_lock(dev, 0);
        FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
                DEBUG("Calling rte_eth_dev_filter_ctrl on sub_device %d", i);
                ret = rte_eth_dev_filter_ctrl(PORT_ID(sdev), type, op, arg);
                if ((ret = fs_err(sdev, ret))) {
                        ERROR("Operation rte_eth_dev_filter_ctrl failed for sub_device %d"
                              " with error %d", i, ret);
+                       fs_unlock(dev, 0);
                        return ret;
                }
        }
+       fs_unlock(dev, 0);
        return 0;
 }
 
index f3be152..5b84db9 100644 (file)
@@ -7,6 +7,7 @@
 #define _RTE_ETH_FAILSAFE_PRIVATE_H_
 
 #include <sys/queue.h>
+#include <pthread.h>
 
 #include <rte_atomic.h>
 #include <rte_dev.h>
@@ -161,6 +162,9 @@ struct fs_priv {
         * appropriate failsafe Rx queue.
         */
        struct rx_proxy rxp;
+       pthread_mutex_t hotplug_mutex;
+       /* Hot-plug mutex is locked by the alarm mechanism. */
+       volatile unsigned int alarm_lock:1;
        unsigned int pending_alarm:1; /* An alarm is pending */
        /* flow isolation state */
        int flow_isolated:1;
@@ -346,6 +350,58 @@ fs_find_next(struct rte_eth_dev *dev,
        return &subs[sid];
 }
 
+/*
+ * Lock hot-plug mutex.
+ * is_alarm means that the caller is, for sure, the hot-plug alarm mechanism.
+ */
+static inline int
+fs_lock(struct rte_eth_dev *dev, unsigned int is_alarm)
+{
+       int ret;
+
+       if (is_alarm) {
+               ret = pthread_mutex_trylock(&PRIV(dev)->hotplug_mutex);
+               if (ret) {
+                       DEBUG("Hot-plug mutex lock trying failed(%s), will try"
+                             " again later...", strerror(ret));
+                       return ret;
+               }
+               PRIV(dev)->alarm_lock = 1;
+       } else {
+               ret = pthread_mutex_lock(&PRIV(dev)->hotplug_mutex);
+               if (ret) {
+                       ERROR("Cannot lock mutex(%s)", strerror(ret));
+                       return ret;
+               }
+       }
+       DEBUG("Hot-plug mutex was locked by thread %lu%s", pthread_self(),
+             PRIV(dev)->alarm_lock ? " by the hot-plug alarm" : "");
+       return ret;
+}
+
+/*
+ * Unlock hot-plug mutex.
+ * is_alarm means that the caller is, for sure, the hot-plug alarm mechanism.
+ */
+static inline void
+fs_unlock(struct rte_eth_dev *dev, unsigned int is_alarm)
+{
+       int ret;
+       unsigned int prev_alarm_lock = PRIV(dev)->alarm_lock;
+
+       if (is_alarm) {
+               RTE_ASSERT(PRIV(dev)->alarm_lock == 1);
+               PRIV(dev)->alarm_lock = 0;
+       }
+       ret = pthread_mutex_unlock(&PRIV(dev)->hotplug_mutex);
+       if (ret)
+               ERROR("Cannot unlock hot-plug mutex(%s)", strerror(ret));
+       else
+               DEBUG("Hot-plug mutex was unlocked by thread %lu%s",
+                     pthread_self(),
+                     prev_alarm_lock ? " by the hot-plug alarm" : "");
+}
+
 /*
  * Switch emitting device.
  * If banned is set, banned must not be considered for