ethdev: make flow API thread safe
authorSuanming Mou <suanmingm@nvidia.com>
Thu, 15 Oct 2020 01:07:47 +0000 (09:07 +0800)
committerThomas Monjalon <thomas@monjalon.net>
Thu, 15 Oct 2020 22:44:58 +0000 (00:44 +0200)
Currently, the rte_flow functions are not defined as thread safe.
DPDK applications either call the functions in single thread or
protect any concurrent calling for the rte_flow operations using
a lock.

For PMDs support the flow operations thread safe natively, the
redundant protection in application hurts the performance of the
rte_flow operation functions.

And the restriction of thread safe is not guaranteed for the
rte_flow functions also limits the applications' expectation.

This feature is going to change the rte_flow functions to be thread
safe. As different PMDs have different flow operations, some may
support thread safe already and others may not. For PMDs don't
support flow thread safe operation, a new lock is defined in ethdev
in order to protects thread unsafe PMDs from rte_flow level.

A new RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE device flag is added to
determine whether the PMD supports thread safe flow operation or not.
For PMDs support thread safe flow operations, set the
RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE flag, rte_flow level functions will
skip the thread safe helper lock for these PMDs. Again the rte_flow
level thread safe lock only works when PMD operation functions are
not thread safe.

For the PMDs which don't want the default mutex lock, just set the
flag in the PMD, and add the prefer type of lock in the PMD. Then
the default mutex lock is easily replaced by the PMD level lock.

The change has no effect on the current DPDK applications. No change
is required for the current DPDK applications. For the standard posix
pthread_mutex, if no lock contention with the added rte_flow level
mutex, the mutex only does the atomic increasing in
pthread_mutex_lock() and decreasing in
pthread_mutex_unlock(). No futex() syscall will be involved.

Signed-off-by: Suanming Mou <suanmingm@nvidia.com>
Acked-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
Acked-by: Ori Kam <orika@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
Acked-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>
doc/guides/prog_guide/rte_flow.rst
doc/guides/rel_notes/release_20_11.rst
lib/librte_ethdev/rte_ethdev.c
lib/librte_ethdev/rte_ethdev.h
lib/librte_ethdev/rte_ethdev_core.h
lib/librte_ethdev/rte_flow.c

index 119b128..07f44bb 100644 (file)
@@ -3046,10 +3046,6 @@ Caveats
 - API operations are synchronous and blocking (``EAGAIN`` cannot be
   returned).
 
-- There is no provision for re-entrancy/multi-thread safety, although nothing
-  should prevent different devices from being configured at the same
-  time. PMDs may protect their control path functions accordingly.
-
 - Stopping the data path (TX/RX) should not be necessary when managing flow
   rules. If this cannot be achieved naturally or with workarounds (such as
   temporarily replacing the burst function pointers), an appropriate error
@@ -3101,6 +3097,14 @@ This interface additionally defines the following helper function:
 - ``rte_flow_ops_get()``: get generic flow operations structure from a
   port.
 
+If PMD interfaces don't support re-entrancy/multi-thread safety,
+the rte_flow API functions will protect threads by mutex per port.
+The application can check whether ``RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE``
+is set in ``dev_flags``, meaning the PMD is thread-safe regarding rte_flow,
+so the API level protection is disabled.
+Please note that this API-level mutex protects only rte_flow functions,
+other control path functions are not in scope.
+
 More will be added over time.
 
 Device compatibility
index 708ebb0..1ddbdce 100644 (file)
@@ -79,6 +79,12 @@ New Features
   Added the FEC API which provides functions for query FEC capabilities and
   current FEC mode from device. Also, API for configuring FEC mode is also provided.
 
+* **Added thread safety to rte_flow functions.**
+
+  Added ``RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE`` device flag to indicate
+  whether PMD supports thread safe operations. If PMD doesn't set the flag,
+  rte_flow API level functions will protect the flow operations with mutex.
+
 * **Updated Broadcom bnxt driver.**
 
   Updated the Broadcom bnxt driver with new features and improvements, including:
index 59beb8a..c8ed890 100644 (file)
@@ -500,6 +500,7 @@ rte_eth_dev_allocate(const char *name)
        strlcpy(eth_dev->data->name, name, sizeof(eth_dev->data->name));
        eth_dev->data->port_id = port_id;
        eth_dev->data->mtu = RTE_ETHER_MTU;
+       pthread_mutex_init(&eth_dev->data->flow_ops_mutex, NULL);
 
 unlock:
        rte_spinlock_unlock(&rte_eth_dev_shared_data->ownership_lock);
@@ -564,6 +565,7 @@ rte_eth_dev_release_port(struct rte_eth_dev *eth_dev)
                rte_free(eth_dev->data->mac_addrs);
                rte_free(eth_dev->data->hash_mac_addrs);
                rte_free(eth_dev->data->dev_private);
+               pthread_mutex_destroy(&eth_dev->data->flow_ops_mutex);
                memset(eth_dev->data, 0, sizeof(struct rte_eth_dev_data));
        }
 
index 3a31f94..500b2ff 100644 (file)
@@ -1677,6 +1677,8 @@ struct rte_eth_dev_owner {
        char name[RTE_ETH_MAX_OWNER_NAME_LEN]; /**< The owner name. */
 };
 
+/** PMD supports thread-safe flow operations */
+#define RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE  0x0001
 /** Device supports link state interrupt */
 #define RTE_ETH_DEV_INTR_LSC     0x0002
 /** Device is a bonded slave */
index fd3bf92..918a34e 100644 (file)
@@ -5,6 +5,9 @@
 #ifndef _RTE_ETHDEV_CORE_H_
 #define _RTE_ETHDEV_CORE_H_
 
+#include <pthread.h>
+#include <sys/types.h>
+
 /**
  * @file
  *
@@ -180,6 +183,7 @@ struct rte_eth_dev_data {
                         *   Valid if RTE_ETH_DEV_REPRESENTOR in dev_flags.
                         */
 
+       pthread_mutex_t flow_ops_mutex; /**< rte_flow ops mutex. */
        uint64_t reserved_64s[4]; /**< Reserved for future fields */
        void *reserved_ptrs[4];   /**< Reserved for future fields */
 } __rte_cache_aligned;
index 8d1b279..4101b27 100644 (file)
@@ -207,6 +207,20 @@ error:
        return -rte_errno;
 }
 
+static inline void
+fts_enter(struct rte_eth_dev *dev)
+{
+       if (!(dev->data->dev_flags & RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE))
+               pthread_mutex_lock(&dev->data->flow_ops_mutex);
+}
+
+static inline void
+fts_exit(struct rte_eth_dev *dev)
+{
+       if (!(dev->data->dev_flags & RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE))
+               pthread_mutex_unlock(&dev->data->flow_ops_mutex);
+}
+
 static int
 flow_err(uint16_t port_id, int ret, struct rte_flow_error *error)
 {
@@ -253,12 +267,16 @@ rte_flow_validate(uint16_t port_id,
 {
        const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
        struct rte_eth_dev *dev = &rte_eth_devices[port_id];
+       int ret;
 
        if (unlikely(!ops))
                return -rte_errno;
-       if (likely(!!ops->validate))
-               return flow_err(port_id, ops->validate(dev, attr, pattern,
-                                                      actions, error), error);
+       if (likely(!!ops->validate)) {
+               fts_enter(dev);
+               ret = ops->validate(dev, attr, pattern, actions, error);
+               fts_exit(dev);
+               return flow_err(port_id, ret, error);
+       }
        return rte_flow_error_set(error, ENOSYS,
                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                  NULL, rte_strerror(ENOSYS));
@@ -279,7 +297,9 @@ rte_flow_create(uint16_t port_id,
        if (unlikely(!ops))
                return NULL;
        if (likely(!!ops->create)) {
+               fts_enter(dev);
                flow = ops->create(dev, attr, pattern, actions, error);
+               fts_exit(dev);
                if (flow == NULL)
                        flow_err(port_id, -rte_errno, error);
                return flow;
@@ -297,12 +317,16 @@ rte_flow_destroy(uint16_t port_id,
 {
        struct rte_eth_dev *dev = &rte_eth_devices[port_id];
        const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
+       int ret;
 
        if (unlikely(!ops))
                return -rte_errno;
-       if (likely(!!ops->destroy))
-               return flow_err(port_id, ops->destroy(dev, flow, error),
-                               error);
+       if (likely(!!ops->destroy)) {
+               fts_enter(dev);
+               ret = ops->destroy(dev, flow, error);
+               fts_exit(dev);
+               return flow_err(port_id, ret, error);
+       }
        return rte_flow_error_set(error, ENOSYS,
                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                  NULL, rte_strerror(ENOSYS));
@@ -315,11 +339,16 @@ rte_flow_flush(uint16_t port_id,
 {
        struct rte_eth_dev *dev = &rte_eth_devices[port_id];
        const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
+       int ret;
 
        if (unlikely(!ops))
                return -rte_errno;
-       if (likely(!!ops->flush))
-               return flow_err(port_id, ops->flush(dev, error), error);
+       if (likely(!!ops->flush)) {
+               fts_enter(dev);
+               ret = ops->flush(dev, error);
+               fts_exit(dev);
+               return flow_err(port_id, ret, error);
+       }
        return rte_flow_error_set(error, ENOSYS,
                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                  NULL, rte_strerror(ENOSYS));
@@ -335,12 +364,16 @@ rte_flow_query(uint16_t port_id,
 {
        struct rte_eth_dev *dev = &rte_eth_devices[port_id];
        const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
+       int ret;
 
        if (!ops)
                return -rte_errno;
-       if (likely(!!ops->query))
-               return flow_err(port_id, ops->query(dev, flow, action, data,
-                                                   error), error);
+       if (likely(!!ops->query)) {
+               fts_enter(dev);
+               ret = ops->query(dev, flow, action, data, error);
+               fts_exit(dev);
+               return flow_err(port_id, ret, error);
+       }
        return rte_flow_error_set(error, ENOSYS,
                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                  NULL, rte_strerror(ENOSYS));
@@ -354,11 +387,16 @@ rte_flow_isolate(uint16_t port_id,
 {
        struct rte_eth_dev *dev = &rte_eth_devices[port_id];
        const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
+       int ret;
 
        if (!ops)
                return -rte_errno;
-       if (likely(!!ops->isolate))
-               return flow_err(port_id, ops->isolate(dev, set, error), error);
+       if (likely(!!ops->isolate)) {
+               fts_enter(dev);
+               ret = ops->isolate(dev, set, error);
+               fts_exit(dev);
+               return flow_err(port_id, ret, error);
+       }
        return rte_flow_error_set(error, ENOSYS,
                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                  NULL, rte_strerror(ENOSYS));
@@ -962,12 +1000,16 @@ rte_flow_dev_dump(uint16_t port_id, FILE *file, struct rte_flow_error *error)
 {
        struct rte_eth_dev *dev = &rte_eth_devices[port_id];
        const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
+       int ret;
 
        if (unlikely(!ops))
                return -rte_errno;
-       if (likely(!!ops->dev_dump))
-               return flow_err(port_id, ops->dev_dump(dev, file, error),
-                               error);
+       if (likely(!!ops->dev_dump)) {
+               fts_enter(dev);
+               ret = ops->dev_dump(dev, file, error);
+               fts_exit(dev);
+               return flow_err(port_id, ret, error);
+       }
        return rte_flow_error_set(error, ENOSYS,
                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                  NULL, rte_strerror(ENOSYS));
@@ -979,12 +1021,16 @@ rte_flow_get_aged_flows(uint16_t port_id, void **contexts,
 {
        struct rte_eth_dev *dev = &rte_eth_devices[port_id];
        const struct rte_flow_ops *ops = rte_flow_ops_get(port_id, error);
+       int ret;
 
        if (unlikely(!ops))
                return -rte_errno;
-       if (likely(!!ops->get_aged_flows))
-               return flow_err(port_id, ops->get_aged_flows(dev, contexts,
-                               nb_contexts, error), error);
+       if (likely(!!ops->get_aged_flows)) {
+               fts_enter(dev);
+               ret = ops->get_aged_flows(dev, contexts, nb_contexts, error);
+               fts_exit(dev);
+               return flow_err(port_id, ret, error);
+       }
        return rte_flow_error_set(error, ENOTSUP,
                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                  NULL, rte_strerror(ENOTSUP));