net/mlx5: support Rx queue delay drop
[dpdk.git] / drivers / common / mlx5 / mlx5_common.c
index 17a54ac..e6ff045 100644 (file)
@@ -13,6 +13,7 @@
 
 #include "mlx5_common.h"
 #include "mlx5_common_os.h"
+#include "mlx5_common_mp.h"
 #include "mlx5_common_log.h"
 #include "mlx5_common_defs.h"
 #include "mlx5_common_private.h"
@@ -258,12 +259,6 @@ is_valid_class_combination(uint32_t user_classes)
        return 0;
 }
 
-static bool
-device_class_enabled(const struct mlx5_common_device *device, uint32_t class)
-{
-       return (device->classes_loaded & class) > 0;
-}
-
 static bool
 mlx5_bus_match(const struct mlx5_class_driver *drv,
               const struct rte_device *dev)
@@ -308,6 +303,187 @@ mlx5_dev_to_pci_str(const struct rte_device *dev, char *addr, size_t size)
 #endif
 }
 
+/**
+ * Register the mempool for the protection domain.
+ *
+ * @param cdev
+ *   Pointer to the mlx5 common device.
+ * @param mp
+ *   Mempool being registered.
+ *
+ * @return
+ *   0 on success, (-1) on failure and rte_errno is set.
+ */
+static int
+mlx5_dev_mempool_register(struct mlx5_common_device *cdev,
+                         struct rte_mempool *mp)
+{
+       struct mlx5_mp_id mp_id;
+
+       mlx5_mp_id_init(&mp_id, 0);
+       return mlx5_mr_mempool_register(&cdev->mr_scache, cdev->pd, mp, &mp_id);
+}
+
+/**
+ * Unregister the mempool from the protection domain.
+ *
+ * @param cdev
+ *   Pointer to the mlx5 common device.
+ * @param mp
+ *   Mempool being unregistered.
+ */
+void
+mlx5_dev_mempool_unregister(struct mlx5_common_device *cdev,
+                           struct rte_mempool *mp)
+{
+       struct mlx5_mp_id mp_id;
+
+       mlx5_mp_id_init(&mp_id, 0);
+       if (mlx5_mr_mempool_unregister(&cdev->mr_scache, mp, &mp_id) < 0)
+               DRV_LOG(WARNING, "Failed to unregister mempool %s for PD %p: %s",
+                       mp->name, cdev->pd, rte_strerror(rte_errno));
+}
+
+/**
+ * rte_mempool_walk() callback to register mempools for the protection domain.
+ *
+ * @param mp
+ *   The mempool being walked.
+ * @param arg
+ *   Pointer to the device shared context.
+ */
+static void
+mlx5_dev_mempool_register_cb(struct rte_mempool *mp, void *arg)
+{
+       struct mlx5_common_device *cdev = arg;
+       int ret;
+
+       ret = mlx5_dev_mempool_register(cdev, mp);
+       if (ret < 0 && rte_errno != EEXIST)
+               DRV_LOG(ERR,
+                       "Failed to register existing mempool %s for PD %p: %s",
+                       mp->name, cdev->pd, rte_strerror(rte_errno));
+}
+
+/**
+ * rte_mempool_walk() callback to unregister mempools
+ * from the protection domain.
+ *
+ * @param mp
+ *   The mempool being walked.
+ * @param arg
+ *   Pointer to the device shared context.
+ */
+static void
+mlx5_dev_mempool_unregister_cb(struct rte_mempool *mp, void *arg)
+{
+       mlx5_dev_mempool_unregister((struct mlx5_common_device *)arg, mp);
+}
+
+/**
+ * Mempool life cycle callback for mlx5 common devices.
+ *
+ * @param event
+ *   Mempool life cycle event.
+ * @param mp
+ *   Associated mempool.
+ * @param arg
+ *   Pointer to a device shared context.
+ */
+static void
+mlx5_dev_mempool_event_cb(enum rte_mempool_event event, struct rte_mempool *mp,
+                         void *arg)
+{
+       struct mlx5_common_device *cdev = arg;
+
+       switch (event) {
+       case RTE_MEMPOOL_EVENT_READY:
+               if (mlx5_dev_mempool_register(cdev, mp) < 0)
+                       DRV_LOG(ERR,
+                               "Failed to register new mempool %s for PD %p: %s",
+                               mp->name, cdev->pd, rte_strerror(rte_errno));
+               break;
+       case RTE_MEMPOOL_EVENT_DESTROY:
+               mlx5_dev_mempool_unregister(cdev, mp);
+               break;
+       }
+}
+
+int
+mlx5_dev_mempool_subscribe(struct mlx5_common_device *cdev)
+{
+       int ret = 0;
+
+       if (!cdev->config.mr_mempool_reg_en)
+               return 0;
+       rte_rwlock_write_lock(&cdev->mr_scache.mprwlock);
+       if (cdev->mr_scache.mp_cb_registered)
+               goto exit;
+       /* Callback for this device may be already registered. */
+       ret = rte_mempool_event_callback_register(mlx5_dev_mempool_event_cb,
+                                                 cdev);
+       if (ret != 0 && rte_errno != EEXIST)
+               goto exit;
+       /* Register mempools only once for this device. */
+       if (ret == 0)
+               rte_mempool_walk(mlx5_dev_mempool_register_cb, cdev);
+       ret = 0;
+       cdev->mr_scache.mp_cb_registered = 1;
+exit:
+       rte_rwlock_write_unlock(&cdev->mr_scache.mprwlock);
+       return ret;
+}
+
+static void
+mlx5_dev_mempool_unsubscribe(struct mlx5_common_device *cdev)
+{
+       int ret;
+
+       if (!cdev->mr_scache.mp_cb_registered ||
+           !cdev->config.mr_mempool_reg_en)
+               return;
+       /* Stop watching for mempool events and unregister all mempools. */
+       ret = rte_mempool_event_callback_unregister(mlx5_dev_mempool_event_cb,
+                                                   cdev);
+       if (ret == 0)
+               rte_mempool_walk(mlx5_dev_mempool_unregister_cb, cdev);
+}
+
+/**
+ * Callback for memory event.
+ *
+ * @param event_type
+ *   Memory event type.
+ * @param addr
+ *   Address of memory.
+ * @param len
+ *   Size of memory.
+ */
+static void
+mlx5_mr_mem_event_cb(enum rte_mem_event event_type, const void *addr,
+                    size_t len, void *arg __rte_unused)
+{
+       struct mlx5_common_device *cdev;
+
+       /* Must be called from the primary process. */
+       MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
+       switch (event_type) {
+       case RTE_MEM_EVENT_FREE:
+               pthread_mutex_lock(&devices_list_lock);
+               /* Iterate all the existing mlx5 devices. */
+               TAILQ_FOREACH(cdev, &devices_list, next)
+                       mlx5_free_mr_by_addr(&cdev->mr_scache,
+                                            mlx5_os_get_ctx_device_name
+                                                                   (cdev->ctx),
+                                            addr, len);
+               pthread_mutex_unlock(&devices_list_lock);
+               break;
+       case RTE_MEM_EVENT_ALLOC:
+       default:
+               break;
+       }
+}
+
 /**
  * Uninitialize all HW global of device context.
  *
@@ -376,8 +552,14 @@ mlx5_common_dev_release(struct mlx5_common_device *cdev)
        pthread_mutex_lock(&devices_list_lock);
        TAILQ_REMOVE(&devices_list, cdev, next);
        pthread_mutex_unlock(&devices_list_lock);
-       if (rte_eal_process_type() == RTE_PROC_PRIMARY)
+       if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
+               if (TAILQ_EMPTY(&devices_list))
+                       rte_mem_event_callback_unregister("MLX5_MEM_EVENT_CB",
+                                                         NULL);
+               mlx5_dev_mempool_unsubscribe(cdev);
+               mlx5_mr_release_cache(&cdev->mr_scache);
                mlx5_dev_hw_global_release(cdev);
+       }
        rte_free(cdev);
 }
 
@@ -412,6 +594,18 @@ mlx5_common_dev_create(struct rte_device *eal_dev, uint32_t classes)
                rte_free(cdev);
                return NULL;
        }
+       /* Initialize global MR cache resources and update its functions. */
+       ret = mlx5_mr_create_cache(&cdev->mr_scache, eal_dev->numa_node);
+       if (ret) {
+               DRV_LOG(ERR, "Failed to initialize global MR share cache.");
+               mlx5_dev_hw_global_release(cdev);
+               rte_free(cdev);
+               return NULL;
+       }
+       /* Register callback function for global shared MR cache management. */
+       if (TAILQ_EMPTY(&devices_list))
+               rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
+                                               mlx5_mr_mem_event_cb, NULL);
 exit:
        pthread_mutex_lock(&devices_list_lock);
        TAILQ_INSERT_HEAD(&devices_list, cdev, next);
@@ -545,62 +739,106 @@ mlx5_common_dev_remove(struct rte_device *eal_dev)
        return ret;
 }
 
+/**
+ * Callback to DMA map external memory to a device.
+ *
+ * @param rte_dev
+ *   Pointer to the generic device.
+ * @param addr
+ *   Starting virtual address of memory to be mapped.
+ * @param iova
+ *   Starting IOVA address of memory to be mapped.
+ * @param len
+ *   Length of memory segment being mapped.
+ *
+ * @return
+ *   0 on success, negative value on error.
+ */
 int
-mlx5_common_dev_dma_map(struct rte_device *dev, void *addr, uint64_t iova,
-                       size_t len)
+mlx5_common_dev_dma_map(struct rte_device *rte_dev, void *addr,
+                       uint64_t iova __rte_unused, size_t len)
 {
-       struct mlx5_class_driver *driver = NULL;
-       struct mlx5_class_driver *temp;
-       struct mlx5_common_device *mdev;
-       int ret = -EINVAL;
+       struct mlx5_common_device *dev;
+       struct mlx5_mr *mr;
 
-       mdev = to_mlx5_device(dev);
-       if (!mdev)
-               return -ENODEV;
-       TAILQ_FOREACH(driver, &drivers_list, next) {
-               if (!device_class_enabled(mdev, driver->drv_class) ||
-                   driver->dma_map == NULL)
-                       continue;
-               ret = driver->dma_map(dev, addr, iova, len);
-               if (ret)
-                       goto map_err;
+       dev = to_mlx5_device(rte_dev);
+       if (!dev) {
+               DRV_LOG(WARNING,
+                       "Unable to find matching mlx5 device to device %s",
+                       rte_dev->name);
+               rte_errno = ENODEV;
+               return -1;
        }
-       return ret;
-map_err:
-       TAILQ_FOREACH(temp, &drivers_list, next) {
-               if (temp == driver)
-                       break;
-               if (device_class_enabled(mdev, temp->drv_class) &&
-                   temp->dma_map && temp->dma_unmap)
-                       temp->dma_unmap(dev, addr, iova, len);
+       mr = mlx5_create_mr_ext(dev->pd, (uintptr_t)addr, len,
+                               SOCKET_ID_ANY, dev->mr_scache.reg_mr_cb);
+       if (!mr) {
+               DRV_LOG(WARNING, "Device %s unable to DMA map", rte_dev->name);
+               rte_errno = EINVAL;
+               return -1;
        }
-       return ret;
+       rte_rwlock_write_lock(&dev->mr_scache.rwlock);
+       LIST_INSERT_HEAD(&dev->mr_scache.mr_list, mr, mr);
+       /* Insert to the global cache table. */
+       mlx5_mr_insert_cache(&dev->mr_scache, mr);
+       rte_rwlock_write_unlock(&dev->mr_scache.rwlock);
+       return 0;
 }
 
+/**
+ * Callback to DMA unmap external memory to a device.
+ *
+ * @param rte_dev
+ *   Pointer to the generic device.
+ * @param addr
+ *   Starting virtual address of memory to be unmapped.
+ * @param iova
+ *   Starting IOVA address of memory to be unmapped.
+ * @param len
+ *   Length of memory segment being unmapped.
+ *
+ * @return
+ *   0 on success, negative value on error.
+ */
 int
-mlx5_common_dev_dma_unmap(struct rte_device *dev, void *addr, uint64_t iova,
-                         size_t len)
+mlx5_common_dev_dma_unmap(struct rte_device *rte_dev, void *addr,
+                         uint64_t iova __rte_unused, size_t len __rte_unused)
 {
-       struct mlx5_class_driver *driver;
-       struct mlx5_common_device *mdev;
-       int local_ret = -EINVAL;
-       int ret = 0;
+       struct mlx5_common_device *dev;
+       struct mr_cache_entry entry;
+       struct mlx5_mr *mr;
 
-       mdev = to_mlx5_device(dev);
-       if (!mdev)
-               return -ENODEV;
-       /* There is no unmap error recovery in current implementation. */
-       TAILQ_FOREACH_REVERSE(driver, &drivers_list, mlx5_drivers, next) {
-               if (!device_class_enabled(mdev, driver->drv_class) ||
-                   driver->dma_unmap == NULL)
-                       continue;
-               local_ret = driver->dma_unmap(dev, addr, iova, len);
-               if (local_ret && (ret == 0))
-                       ret = local_ret;
+       dev = to_mlx5_device(rte_dev);
+       if (!dev) {
+               DRV_LOG(WARNING,
+                       "Unable to find matching mlx5 device to device %s.",
+                       rte_dev->name);
+               rte_errno = ENODEV;
+               return -1;
        }
-       if (local_ret)
-               ret = local_ret;
-       return ret;
+       rte_rwlock_read_lock(&dev->mr_scache.rwlock);
+       mr = mlx5_mr_lookup_list(&dev->mr_scache, &entry, (uintptr_t)addr);
+       if (!mr) {
+               rte_rwlock_read_unlock(&dev->mr_scache.rwlock);
+               DRV_LOG(WARNING,
+                       "Address 0x%" PRIxPTR " wasn't registered to device %s",
+                       (uintptr_t)addr, rte_dev->name);
+               rte_errno = EINVAL;
+               return -1;
+       }
+       LIST_REMOVE(mr, mr);
+       DRV_LOG(DEBUG, "MR(%p) is removed from list.", (void *)mr);
+       mlx5_mr_free(mr, dev->mr_scache.dereg_mr_cb);
+       mlx5_mr_rebuild_cache(&dev->mr_scache);
+       /*
+        * No explicit wmb is needed after updating dev_gen due to
+        * store-release ordering in unlock that provides the
+        * implicit barrier at the software visible level.
+        */
+       ++dev->mr_scache.dev_gen;
+       DRV_LOG(DEBUG, "Broadcasting local cache flush, gen=%d.",
+               dev->mr_scache.dev_gen);
+       rte_rwlock_read_unlock(&dev->mr_scache.rwlock);
+       return 0;
 }
 
 void