[ovs-dev] [PATCH v3 04/14] netdev-dpdk: Serialise non-pmds mbufs' alloc/free.

[Tiago Lam]

A new mutex, 'nonpmd_mp_mutex', has been introduced to serialise
allocation and free operations by non-pmd threads on a given mempool.

free_dpdk_buf() has been modified to make use of the introduced mutex.

Signed-off-by: Tiago Lam <tiago.lam at intel.com>
Acked-by: Eelco Chaudron <echaudro at redhat.com>
---
 lib/netdev-dpdk.c | 30 +++++++++++++++++++++++++++---
 1 file changed, 27 insertions(+), 3 deletions(-)

diff --git a/lib/netdev-dpdk.c b/lib/netdev-dpdk.c
index 675dd48..6c9f0e7 100644
--- a/lib/netdev-dpdk.c
+++ b/lib/netdev-dpdk.c
@@ -294,6 +294,16 @@ static struct ovs_mutex dpdk_mp_mutex OVS_ACQ_AFTER(dpdk_mutex)
 static struct ovs_list dpdk_mp_free_list OVS_GUARDED_BY(dpdk_mp_mutex)
     = OVS_LIST_INITIALIZER(&dpdk_mp_free_list);
 
+/* This mutex must be used by non pmd threads when allocating or freeing
+ * mbufs through mempools, when outside of the `non_pmd_mutex` mutex, in struct
+ * dp_netdev.
+ * The reason, as pointed out in the "Known Issues" section in DPDK's EAL docs,
+ * is that the implementation on which mempool is based off is non-preemptable.
+ * Since non-pmds may end up not being pinned this could lead to the preemption
+ * between non-pmds performing operations on the same mempool, which could lead
+ * to memory corruption. */
+static struct ovs_mutex nonpmd_mp_mutex = OVS_MUTEX_INITIALIZER;
+
 /* Wrapper for a mempool released but not yet freed. */
 struct dpdk_mp {
      struct rte_mempool *mp;
@@ -462,6 +472,8 @@ struct netdev_rxq_dpdk {
     dpdk_port_t port_id;
 };
 
+static bool dpdk_thread_is_pmd(void);
+
 static void netdev_dpdk_destruct(struct netdev *netdev);
 static void netdev_dpdk_vhost_destruct(struct netdev *netdev);
 
@@ -495,6 +507,12 @@ dpdk_buf_size(int mtu)
                      NETDEV_DPDK_MBUF_ALIGN);
 }
 
+static bool
+dpdk_thread_is_pmd(void)
+{
+     return rte_lcore_id() != NON_PMD_CORE_ID;
+}
+
 /* Allocates an area of 'sz' bytes from DPDK.  The memory is zero'ed.
  *
  * Unlike xmalloc(), this function can return NULL on failure. */
@@ -505,11 +523,17 @@ dpdk_rte_mzalloc(size_t sz)
 }
 
 void
-free_dpdk_buf(struct dp_packet *p)
+free_dpdk_buf(struct dp_packet *packet)
 {
-    struct rte_mbuf *pkt = (struct rte_mbuf *) p;
+    if (!dpdk_thread_is_pmd()) {
+        ovs_mutex_lock(&nonpmd_mp_mutex);
+    }
 
-    rte_pktmbuf_free(pkt);
+    rte_pktmbuf_free(&packet->mbuf);
+
+    if (!dpdk_thread_is_pmd()) {
+        ovs_mutex_unlock(&nonpmd_mp_mutex);
+    }
 }
 
 static void
-- 
2.7.4