RDMA/mlx5: Consolidate MR destruction to mlx5_ib_dereg_mr()

Now that the SRCU stuff has been removed the entire MR destroy logic can
be made a lot simpler. Currently there are many different ways to destroy a
MR and it makes it really hard to do this task correctly. Route all
destruction through mlx5_ib_dereg_mr() and make it work for all
situations.

Since it turns out all the different MR types do basically the same thing
this removes a lot of knowledge of MR internals from ODP and leaves ODP
just exporting an operation to clean up children.

This fixes a few weird corner cases bugs and firmly uses the correct
ordering of the MR destruction:
 - Stop parallel access to the mkey via the ODP xarray
 - Stop DMA
 - Release the umem
 - Clean up ODP children
 - Free/Recycle the MR

Link: https://lore.kernel.org/r/20210304120745.1090751-4-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>

1 files changed, 62 insertions, 71 deletions

diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c
index 54fd38b01a7e..6304ba54a42d 100644
--- a/drivers/infiniband/hw/mlx5/mr.c
+++ b/drivers/infiniband/hw/mlx5/mr.c
@@ -119,8 +119,6 @@ mlx5_ib_create_mkey_cb(struct mlx5_ib_dev *dev,
 				create_mkey_callback, context);
 }
 
-static void clean_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr);
-static void dereg_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr);
 static int mr_cache_max_order(struct mlx5_ib_dev *dev);
 static void queue_adjust_cache_locked(struct mlx5_cache_ent *ent);
 
@@ -627,30 +625,10 @@ static struct mlx5_ib_mr *get_cache_mr(struct mlx5_cache_ent *req_ent)
 	return NULL;
 }
 
-static void detach_mr_from_cache(struct mlx5_ib_mr *mr)
+static void mlx5_mr_cache_free(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
 {
 	struct mlx5_cache_ent *ent = mr->cache_ent;
 
-	mr->cache_ent = NULL;
-	spin_lock_irq(&ent->lock);
-	ent->total_mrs--;
-	spin_unlock_irq(&ent->lock);
-}
-
-void mlx5_mr_cache_free(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
-{
-	struct mlx5_cache_ent *ent = mr->cache_ent;
-
-	if (!ent)
-		return;
-
-	if (mlx5_mr_cache_invalidate(mr)) {
-		detach_mr_from_cache(mr);
-		destroy_mkey(dev, mr);
-		kfree(mr);
-		return;
-	}
-
 	spin_lock_irq(&ent->lock);
 	list_add_tail(&mr->list, &ent->head);
 	ent->available_mrs++;
@@ -1503,7 +1481,7 @@ static struct ib_mr *create_real_mr(struct ib_pd *pd, struct ib_umem *umem,
 		 */
 		err = mlx5_ib_update_mr_pas(mr, MLX5_IB_UPD_XLT_ENABLE);
 		if (err) {
-			dereg_mr(dev, mr);
+			mlx5_ib_dereg_mr(&mr->ibmr, NULL);
 			return ERR_PTR(err);
 		}
 	}
@@ -1560,7 +1538,7 @@ static struct ib_mr *create_user_odp_mr(struct ib_pd *pd, u64 start, u64 length,
 	return &mr->ibmr;
 
 err_dereg_mr:
-	dereg_mr(dev, mr);
+	mlx5_ib_dereg_mr(&mr->ibmr, NULL);
 	return ERR_PTR(err);
 }
 
@@ -1657,7 +1635,7 @@ struct ib_mr *mlx5_ib_reg_user_mr_dmabuf(struct ib_pd *pd, u64 offset,
 	return &mr->ibmr;
 
 err_dereg_mr:
-	dereg_mr(dev, mr);
+	mlx5_ib_dereg_mr(&mr->ibmr, NULL);
 	return ERR_PTR(err);
 }
 
@@ -1669,7 +1647,7 @@ err_dereg_mr:
  * and any DMA inprogress will be completed. Failure of this function
  * indicates the HW has failed catastrophically.
  */
-int mlx5_mr_cache_invalidate(struct mlx5_ib_mr *mr)
+static int mlx5_mr_cache_invalidate(struct mlx5_ib_mr *mr)
 {
 	struct mlx5_umr_wr umrwr = {};
 
@@ -1941,69 +1919,82 @@ mlx5_free_priv_descs(struct mlx5_ib_mr *mr)
 	}
 }
 
-static void clean_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
+int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
 {
-	if (mr->ibmr.type == IB_MR_TYPE_INTEGRITY) {
+	struct mlx5_ib_mr *mr = to_mmr(ibmr);
+	struct mlx5_ib_dev *dev = to_mdev(ibmr->device);
+	int rc;
+
+	/*
+	 * Any async use of the mr must hold the refcount, once the refcount
+	 * goes to zero no other thread, such as ODP page faults, prefetch, any
+	 * UMR activity, etc can touch the mkey. Thus it is safe to destroy it.
+	 */
+	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING) &&
+	    refcount_read(&mr->mmkey.usecount) != 0 &&
+	    xa_erase(&mr_to_mdev(mr)->odp_mkeys, mlx5_base_mkey(mr->mmkey.key)))
+		mlx5r_deref_wait_odp_mkey(&mr->mmkey);
+
+	if (ibmr->type == IB_MR_TYPE_INTEGRITY) {
+		xa_cmpxchg(&dev->sig_mrs, mlx5_base_mkey(mr->mmkey.key), ibmr,
+			   NULL, GFP_KERNEL);
+
+		if (mr->mtt_mr) {
+			rc = mlx5_ib_dereg_mr(&mr->mtt_mr->ibmr, NULL);
+			if (rc)
+				return rc;
+			mr->mtt_mr = NULL;
+		}
+		if (mr->klm_mr) {
+			mlx5_ib_dereg_mr(&mr->klm_mr->ibmr, NULL);
+			if (rc)
+				return rc;
+			mr->klm_mr = NULL;
+		}
+
 		if (mlx5_core_destroy_psv(dev->mdev,
 					  mr->sig->psv_memory.psv_idx))
 			mlx5_ib_warn(dev, "failed to destroy mem psv %d\n",
 				     mr->sig->psv_memory.psv_idx);
-		if (mlx5_core_destroy_psv(dev->mdev,
-					  mr->sig->psv_wire.psv_idx))
+		if (mlx5_core_destroy_psv(dev->mdev, mr->sig->psv_wire.psv_idx))
 			mlx5_ib_warn(dev, "failed to destroy wire psv %d\n",
 				     mr->sig->psv_wire.psv_idx);
-		xa_erase(&dev->sig_mrs, mlx5_base_mkey(mr->mmkey.key));
 		kfree(mr->sig);
 		mr->sig = NULL;
 	}
 
+	/* Stop DMA */
+	if (mr->cache_ent) {
+		if (mlx5_mr_cache_invalidate(mr)) {
+			spin_lock_irq(&mr->cache_ent->lock);
+			mr->cache_ent->total_mrs--;
+			spin_unlock_irq(&mr->cache_ent->lock);
+			mr->cache_ent = NULL;
+		}
+	}
 	if (!mr->cache_ent) {
-		destroy_mkey(dev, mr);
-		mlx5_free_priv_descs(mr);
+		rc = destroy_mkey(to_mdev(mr->ibmr.device), mr);
+		if (rc)
+			return rc;
 	}
-}
-
-static void dereg_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
-{
-	struct ib_umem *umem = mr->umem;
 
-	/* Stop all DMA */
-	if (is_odp_mr(mr))
-		mlx5_ib_fence_odp_mr(mr);
-	else if (is_dmabuf_mr(mr))
-		mlx5_ib_fence_dmabuf_mr(mr);
-	else
-		clean_mr(dev, mr);
+	if (mr->umem) {
+		bool is_odp = is_odp_mr(mr);
 
-	if (umem) {
-		if (!is_odp_mr(mr))
-			atomic_sub(ib_umem_num_pages(umem),
+		if (!is_odp)
+			atomic_sub(ib_umem_num_pages(mr->umem),
 				   &dev->mdev->priv.reg_pages);
-		ib_umem_release(umem);
+		ib_umem_release(mr->umem);
+		if (is_odp)
+			mlx5_ib_free_odp_mr(mr);
 	}
 
-	if (mr->cache_ent)
+	if (mr->cache_ent) {
 		mlx5_mr_cache_free(dev, mr);
-	else
+	} else {
+		mlx5_free_priv_descs(mr);
 		kfree(mr);
-}
-
-int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
-{
-	struct mlx5_ib_mr *mmr = to_mmr(ibmr);
-
-	if (ibmr->type == IB_MR_TYPE_INTEGRITY) {
-		dereg_mr(to_mdev(mmr->mtt_mr->ibmr.device), mmr->mtt_mr);
-		dereg_mr(to_mdev(mmr->klm_mr->ibmr.device), mmr->klm_mr);
-	}
-
-	if (is_odp_mr(mmr) && to_ib_umem_odp(mmr->umem)->is_implicit_odp) {
-		mlx5_ib_free_implicit_mr(mmr);
-		return 0;
 	}
-
-	dereg_mr(to_mdev(ibmr->device), mmr);
-
 	return 0;
 }
 
@@ -2175,10 +2166,10 @@ err_free_descs:
 	destroy_mkey(dev, mr);
 	mlx5_free_priv_descs(mr);
 err_free_mtt_mr:
-	dereg_mr(to_mdev(mr->mtt_mr->ibmr.device), mr->mtt_mr);
+	mlx5_ib_dereg_mr(&mr->mtt_mr->ibmr, NULL);
 	mr->mtt_mr = NULL;
 err_free_klm_mr:
-	dereg_mr(to_mdev(mr->klm_mr->ibmr.device), mr->klm_mr);
+	mlx5_ib_dereg_mr(&mr->klm_mr->ibmr, NULL);
 	mr->klm_mr = NULL;
 err_destroy_psv:
 	if (mlx5_core_destroy_psv(dev->mdev, mr->sig->psv_memory.psv_idx))