compress/mlx5: fix memory region unregistration
authorMichael Baum <michaelba@nvidia.com>
Mon, 12 Jul 2021 07:19:34 +0000 (10:19 +0300)
committerAkhil Goyal <gakhil@marvell.com>
Tue, 20 Jul 2021 08:32:05 +0000 (10:32 +0200)
The issue can cause illegal physical address access while a huge-page A
is released and huge-page B is allocated on the same virtual address.
The old MR can be matched using the virtual address of huge-page B but
the HW will access the physical address of huge-page A which is no more
part of the DPDK process.

Register a driver callback for memory event in order to free out all the
MRs of memory that is going to be freed from the dpdk process.

Fixes: f8c97babc9f4 ("compress/mlx5: add data-path functions")
Cc: stable@dpdk.org
Signed-off-by: Michael Baum <michaelba@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
drivers/compress/mlx5/mlx5_compress.c

index 80c564f..f5f51c0 100644 (file)
@@ -258,6 +258,8 @@ mlx5_compress_qp_setup(struct rte_compressdev *dev, uint16_t qp_id,
                DRV_LOG(ERR, "Can't change SQ state to ready.");
                goto err;
        }
+       /* Save pointer of global generation number to check memory event. */
+       qp->mr_ctrl.dev_gen_ptr = &priv->mr_scache.dev_gen;
        DRV_LOG(INFO, "QP %u: SQN=0x%X CQN=0x%X entries num = %u",
                (uint32_t)qp_id, qp->sq.sq->id, qp->cq.cq->id, qp->entries_n);
        return 0;
@@ -428,6 +430,40 @@ static struct rte_compressdev_ops mlx5_compress_ops = {
        .stream_free            = NULL,
 };
 
+/**
+ * Query LKey from a packet buffer for QP. If not found, add the mempool.
+ *
+ * @param priv
+ *   Pointer to the priv object.
+ * @param addr
+ *   Search key.
+ * @param mr_ctrl
+ *   Pointer to per-queue MR control structure.
+ * @param ol_flags
+ *   Mbuf offload features.
+ *
+ * @return
+ *   Searched LKey on success, UINT32_MAX on no match.
+ */
+static __rte_always_inline uint32_t
+mlx5_compress_addr2mr(struct mlx5_compress_priv *priv, uintptr_t addr,
+                     struct mlx5_mr_ctrl *mr_ctrl, uint64_t ol_flags)
+{
+       uint32_t lkey;
+
+       /* Check generation bit to see if there's any change on existing MRs. */
+       if (unlikely(*mr_ctrl->dev_gen_ptr != mr_ctrl->cur_gen))
+               mlx5_mr_flush_local_cache(mr_ctrl);
+       /* Linear search on MR cache array. */
+       lkey = mlx5_mr_lookup_lkey(mr_ctrl->cache, &mr_ctrl->mru,
+                                  MLX5_MR_CACHE_N, addr);
+       if (likely(lkey != UINT32_MAX))
+               return lkey;
+       /* Take slower bottom-half on miss. */
+       return mlx5_mr_addr2mr_bh(priv->pd, 0, &priv->mr_scache, mr_ctrl, addr,
+                                 !!(ol_flags & EXT_ATTACHED_MBUF));
+}
+
 static __rte_always_inline uint32_t
 mlx5_compress_dseg_set(struct mlx5_compress_qp *qp,
                       volatile struct mlx5_wqe_dseg *restrict dseg,
@@ -437,9 +473,8 @@ mlx5_compress_dseg_set(struct mlx5_compress_qp *qp,
        uintptr_t addr = rte_pktmbuf_mtod_offset(mbuf, uintptr_t, offset);
 
        dseg->bcount = rte_cpu_to_be_32(len);
-       dseg->lkey = mlx5_mr_addr2mr_bh(qp->priv->pd, 0, &qp->priv->mr_scache,
-                                       &qp->mr_ctrl, addr,
-                                       !!(mbuf->ol_flags & EXT_ATTACHED_MBUF));
+       dseg->lkey = mlx5_compress_addr2mr(qp->priv, addr, &qp->mr_ctrl,
+                                          mbuf->ol_flags);
        dseg->pbuf = rte_cpu_to_be_64(addr);
        return dseg->lkey;
 }
@@ -711,6 +746,40 @@ mlx5_compress_hw_global_prepare(struct mlx5_compress_priv *priv)
        return 0;
 }
 
+/**
+ * Callback for memory event.
+ *
+ * @param event_type
+ *   Memory event type.
+ * @param addr
+ *   Address of memory.
+ * @param len
+ *   Size of memory.
+ */
+static void
+mlx5_compress_mr_mem_event_cb(enum rte_mem_event event_type, const void *addr,
+                             size_t len, void *arg __rte_unused)
+{
+       struct mlx5_compress_priv *priv;
+
+       /* 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(&priv_list_lock);
+               /* Iterate all the existing mlx5 devices. */
+               TAILQ_FOREACH(priv, &mlx5_compress_priv_list, next)
+                       mlx5_free_mr_by_addr(&priv->mr_scache,
+                                            priv->ctx->device->name,
+                                            addr, len);
+               pthread_mutex_unlock(&priv_list_lock);
+               break;
+       case RTE_MEM_EVENT_ALLOC:
+       default:
+               break;
+       }
+}
+
 /**
  * DPDK callback to register a PCI device.
  *
@@ -804,6 +873,11 @@ mlx5_compress_pci_probe(struct rte_pci_driver *pci_drv,
        }
        priv->mr_scache.reg_mr_cb = mlx5_common_verbs_reg_mr;
        priv->mr_scache.dereg_mr_cb = mlx5_common_verbs_dereg_mr;
+       /* Register callback function for global shared MR cache management. */
+       if (TAILQ_EMPTY(&mlx5_compress_priv_list))
+               rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
+                                               mlx5_compress_mr_mem_event_cb,
+                                               NULL);
        pthread_mutex_lock(&priv_list_lock);
        TAILQ_INSERT_TAIL(&mlx5_compress_priv_list, priv, next);
        pthread_mutex_unlock(&priv_list_lock);
@@ -834,6 +908,9 @@ mlx5_compress_pci_remove(struct rte_pci_device *pdev)
                TAILQ_REMOVE(&mlx5_compress_priv_list, priv, next);
        pthread_mutex_unlock(&priv_list_lock);
        if (priv) {
+               if (TAILQ_EMPTY(&mlx5_compress_priv_list))
+                       rte_mem_event_callback_unregister("MLX5_MEM_EVENT_CB",
+                                                         NULL);
                mlx5_mr_release_cache(&priv->mr_scache);
                mlx5_compress_hw_global_release(priv);
                rte_compressdev_pmd_destroy(priv->cdev);