RDMA/mlx5: Do not store implicit children in the odp_mkeys xarray

These mkeys are entirely internal and are never used by the HW for
page fault. They should also never be used by userspace for prefetch.
Simplify & optimize things by not including them in the xarray.

Since the prefetch path can now never see a child mkey there is no need
for the second synchronize_srcu() during imr destroy.

Link: https://lore.kernel.org/r/20191009160934.3143-14-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>

diff --git a/drivers/infiniband/hw/mlx5/odp.c b/drivers/infiniband/hw/mlx5/odp.c
index 5cf93d8129a9cab312673bdc603be24761245117..06e24d5e7609584b31b1263579f327426a93a2b5 100644 (file)
--- a/drivers/infiniband/hw/mlx5/odp.c
+++ b/drivers/infiniband/hw/mlx5/odp.c
@@ -146,9 +146,9 @@ void mlx5_odp_populate_klm(struct mlx5_klm *pklm, size_t idx, size_t nentries,
 
 /*
  * This must be called after the mr has been removed from implicit_children
- * and odp_mkeys and the SRCU synchronized.  NOTE: The MR does not necessarily
- * have to be empty here, parallel page faults could have raced with the free
- * process and added pages to it.
+ * and the SRCU synchronized.  NOTE: The MR does not necessarily have to be
+ * empty here, parallel page faults could have raced with the free process and
+ * added pages to it.
  */
 static void free_implicit_child_mr(struct mlx5_ib_mr *mr, bool need_imr_xlt)
 {
@@ -210,7 +210,6 @@ static void destroy_unused_implicit_child_mr(struct mlx5_ib_mr *mr)
            mr)
                goto out_unlock;
 
-       __xa_erase(&mr->dev->odp_mkeys, mlx5_base_mkey(mr->mmkey.key));
        atomic_inc(&imr->num_deferred_work);
        call_srcu(&mr->dev->odp_srcu, &mr->odp_destroy.rcu,
                  free_implicit_child_mr_rcu);
@@ -401,13 +400,6 @@ static struct mlx5_ib_mr *implicit_get_child_mr(struct mlx5_ib_mr *imr,
        if (IS_ERR(mr))
                goto out_umem;
 
-       err = xa_reserve(&imr->dev->odp_mkeys, mlx5_base_mkey(mr->mmkey.key),
-                        GFP_KERNEL);
-       if (err) {
-               ret = ERR_PTR(err);
-               goto out_mr;
-       }
-
        mr->ibmr.pd = imr->ibmr.pd;
        mr->access_flags = imr->access_flags;
        mr->umem = &odp->umem;
@@ -424,7 +416,7 @@ static struct mlx5_ib_mr *implicit_get_child_mr(struct mlx5_ib_mr *imr,
                                 MLX5_IB_UPD_XLT_ENABLE);
        if (err) {
                ret = ERR_PTR(err);
-               goto out_release;
+               goto out_mr;
        }
 
        /*
@@ -433,26 +425,21 @@ static struct mlx5_ib_mr *implicit_get_child_mr(struct mlx5_ib_mr *imr,
         */
        ret = xa_cmpxchg(&imr->implicit_children, idx, NULL, mr,
                         GFP_KERNEL);
-       if (likely(!ret))
-               xa_store(&imr->dev->odp_mkeys, mlx5_base_mkey(mr->mmkey.key),
-                        &mr->mmkey, GFP_ATOMIC);
        if (unlikely(ret)) {
                if (xa_is_err(ret)) {
                        ret = ERR_PTR(xa_err(ret));
-                       goto out_release;
+                       goto out_mr;
                }
                /*
                 * Another thread beat us to creating the child mr, use
                 * theirs.
                 */
-               goto out_release;
+               goto out_mr;
        }
 
        mlx5_ib_dbg(imr->dev, "key %x mr %p\n", mr->mmkey.key, mr);
        return mr;
 
-out_release:
-       xa_release(&imr->dev->odp_mkeys, mlx5_base_mkey(mr->mmkey.key));
 out_mr:
        mlx5_mr_cache_free(imr->dev, mr);
 out_umem:
@@ -535,14 +522,10 @@ void mlx5_ib_free_implicit_mr(struct mlx5_ib_mr *imr)
        xa_lock(&imr->implicit_children);
        xa_for_each (&imr->implicit_children, idx, mtt) {
                __xa_erase(&imr->implicit_children, idx);
-               __xa_erase(&dev->odp_mkeys, mlx5_base_mkey(mtt->mmkey.key));
                list_add(&mtt->odp_destroy.elm, &destroy_list);
        }
        xa_unlock(&imr->implicit_children);
 
-       /* Fence access to the child pointers via the pagefault thread */
-       synchronize_srcu(&dev->odp_srcu);
-
        /*
         * num_deferred_work can only be incremented inside the odp_srcu, or
         * under xa_lock while the child is in the xarray. Thus at this point
@@ -1655,13 +1638,6 @@ get_prefetchable_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
        if (mr->ibmr.pd != pd)
                return NULL;
 
-       /*
-        * Implicit child MRs are internal and userspace should not refer to
-        * them.
-        */
-       if (mr->parent)
-               return NULL;
-
        odp = to_ib_umem_odp(mr->umem);
 
        /* prefetch with write-access must be supported by the MR */