RDMA/mlx5: Add missing synchronize_srcu() for MW cases
authorJason Gunthorpe <jgg@mellanox.com>
Tue, 1 Oct 2019 15:38:21 +0000 (12:38 -0300)
committerJason Gunthorpe <jgg@mellanox.com>
Fri, 4 Oct 2019 18:54:22 +0000 (15:54 -0300)
While MR uses live as the SRCU 'update', the MW case uses the xarray
directly, xa_erase() causes the MW to become inaccessible to the pagefault
thread.

Thus whenever a MW is removed from the xarray we must synchronize_srcu()
before freeing it.

This must be done before freeing the mkey as re-use of the mkey while the
pagefault thread is using the stale mkey is undesirable.

Add the missing synchronizes to MW and DEVX indirect mkey and delete the
bogus protection against double destroy in mlx5_core_destroy_mkey()

Fixes: 534fd7aac56a ("IB/mlx5: Manage indirection mkey upon DEVX flow for ODP")
Fixes: 6aec21f6a832 ("IB/mlx5: Page faults handling infrastructure")
Link: https://lore.kernel.org/r/20191001153821.23621-7-jgg@ziepe.ca
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
drivers/infiniband/hw/mlx5/devx.c
drivers/infiniband/hw/mlx5/mlx5_ib.h
drivers/infiniband/hw/mlx5/mr.c
drivers/net/ethernet/mellanox/mlx5/core/mr.c

index 59022b7441448f0b5eca7e5f406196e4cfe777d3..d609f4659afb7a93d134528e974467e961847749 100644 (file)
@@ -1298,29 +1298,6 @@ static int devx_handle_mkey_create(struct mlx5_ib_dev *dev,
        return 0;
 }
 
-static void devx_free_indirect_mkey(struct rcu_head *rcu)
-{
-       kfree(container_of(rcu, struct devx_obj, devx_mr.rcu));
-}
-
-/* This function to delete from the radix tree needs to be called before
- * destroying the underlying mkey. Otherwise a race might occur in case that
- * other thread will get the same mkey before this one will be deleted,
- * in that case it will fail via inserting to the tree its own data.
- *
- * Note:
- * An error in the destroy is not expected unless there is some other indirect
- * mkey which points to this one. In a kernel cleanup flow it will be just
- * destroyed in the iterative destruction call. In a user flow, in case
- * the application didn't close in the expected order it's its own problem,
- * the mkey won't be part of the tree, in both cases the kernel is safe.
- */
-static void devx_cleanup_mkey(struct devx_obj *obj)
-{
-       xa_erase(&obj->ib_dev->mdev->priv.mkey_table,
-                mlx5_base_mkey(obj->devx_mr.mmkey.key));
-}
-
 static void devx_cleanup_subscription(struct mlx5_ib_dev *dev,
                                      struct devx_event_subscription *sub)
 {
@@ -1362,8 +1339,16 @@ static int devx_obj_cleanup(struct ib_uobject *uobject,
        int ret;
 
        dev = mlx5_udata_to_mdev(&attrs->driver_udata);
-       if (obj->flags & DEVX_OBJ_FLAGS_INDIRECT_MKEY)
-               devx_cleanup_mkey(obj);
+       if (obj->flags & DEVX_OBJ_FLAGS_INDIRECT_MKEY) {
+               /*
+                * The pagefault_single_data_segment() does commands against
+                * the mmkey, we must wait for that to stop before freeing the
+                * mkey, as another allocation could get the same mkey #.
+                */
+               xa_erase(&obj->ib_dev->mdev->priv.mkey_table,
+                        mlx5_base_mkey(obj->devx_mr.mmkey.key));
+               synchronize_srcu(&dev->mr_srcu);
+       }
 
        if (obj->flags & DEVX_OBJ_FLAGS_DCT)
                ret = mlx5_core_destroy_dct(obj->ib_dev->mdev, &obj->core_dct);
@@ -1382,12 +1367,6 @@ static int devx_obj_cleanup(struct ib_uobject *uobject,
                devx_cleanup_subscription(dev, sub_entry);
        mutex_unlock(&devx_event_table->event_xa_lock);
 
-       if (obj->flags & DEVX_OBJ_FLAGS_INDIRECT_MKEY) {
-               call_srcu(&dev->mr_srcu, &obj->devx_mr.rcu,
-                         devx_free_indirect_mkey);
-               return ret;
-       }
-
        kfree(obj);
        return ret;
 }
@@ -1491,26 +1470,21 @@ static int UVERBS_HANDLER(MLX5_IB_METHOD_DEVX_OBJ_CREATE)(
                                   &obj_id);
        WARN_ON(obj->dinlen > MLX5_MAX_DESTROY_INBOX_SIZE_DW * sizeof(u32));
 
-       if (obj->flags & DEVX_OBJ_FLAGS_INDIRECT_MKEY) {
-               err = devx_handle_mkey_indirect(obj, dev, cmd_in, cmd_out);
-               if (err)
-                       goto obj_destroy;
-       }
-
        err = uverbs_copy_to(attrs, MLX5_IB_ATTR_DEVX_OBJ_CREATE_CMD_OUT, cmd_out, cmd_out_len);
        if (err)
-               goto err_copy;
+               goto obj_destroy;
 
        if (opcode == MLX5_CMD_OP_CREATE_GENERAL_OBJECT)
                obj_type = MLX5_GET(general_obj_in_cmd_hdr, cmd_in, obj_type);
-
        obj->obj_id = get_enc_obj_id(opcode | obj_type << 16, obj_id);
 
+       if (obj->flags & DEVX_OBJ_FLAGS_INDIRECT_MKEY) {
+               err = devx_handle_mkey_indirect(obj, dev, cmd_in, cmd_out);
+               if (err)
+                       goto obj_destroy;
+       }
        return 0;
 
-err_copy:
-       if (obj->flags & DEVX_OBJ_FLAGS_INDIRECT_MKEY)
-               devx_cleanup_mkey(obj);
 obj_destroy:
        if (obj->flags & DEVX_OBJ_FLAGS_DCT)
                mlx5_core_destroy_dct(obj->ib_dev->mdev, &obj->core_dct);
index 15e42825cc976e1beba0423f453194f251a9eecb..1a98ee2e01c4b991b10b3c83f8395cf79de94524 100644 (file)
@@ -639,7 +639,6 @@ struct mlx5_ib_mw {
 struct mlx5_ib_devx_mr {
        struct mlx5_core_mkey   mmkey;
        int                     ndescs;
-       struct rcu_head         rcu;
 };
 
 struct mlx5_ib_umr_context {
index 3a27bddfcf31f57de9707c3160719b597c74cc76..630599311586ec23491025dac5603da7d848ead7 100644 (file)
@@ -1962,14 +1962,25 @@ free:
 
 int mlx5_ib_dealloc_mw(struct ib_mw *mw)
 {
+       struct mlx5_ib_dev *dev = to_mdev(mw->device);
        struct mlx5_ib_mw *mmw = to_mmw(mw);
        int err;
 
-       err =  mlx5_core_destroy_mkey((to_mdev(mw->device))->mdev,
-                                     &mmw->mmkey);
-       if (!err)
-               kfree(mmw);
-       return err;
+       if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
+               xa_erase(&dev->mdev->priv.mkey_table,
+                        mlx5_base_mkey(mmw->mmkey.key));
+               /*
+                * pagefault_single_data_segment() may be accessing mmw under
+                * SRCU if the user bound an ODP MR to this MW.
+                */
+               synchronize_srcu(&dev->mr_srcu);
+       }
+
+       err = mlx5_core_destroy_mkey(dev->mdev, &mmw->mmkey);
+       if (err)
+               return err;
+       kfree(mmw);
+       return 0;
 }
 
 int mlx5_ib_check_mr_status(struct ib_mr *ibmr, u32 check_mask,
index 9231b39d18b21c11f971773fa65376f442d010ae..c501bf2a0252104e4460c33191605f4f9a329dff 100644 (file)
@@ -112,17 +112,11 @@ int mlx5_core_destroy_mkey(struct mlx5_core_dev *dev,
        u32 out[MLX5_ST_SZ_DW(destroy_mkey_out)] = {0};
        u32 in[MLX5_ST_SZ_DW(destroy_mkey_in)]   = {0};
        struct xarray *mkeys = &dev->priv.mkey_table;
-       struct mlx5_core_mkey *deleted_mkey;
        unsigned long flags;
 
        xa_lock_irqsave(mkeys, flags);
-       deleted_mkey = __xa_erase(mkeys, mlx5_base_mkey(mkey->key));
+       __xa_erase(mkeys, mlx5_base_mkey(mkey->key));
        xa_unlock_irqrestore(mkeys, flags);
-       if (!deleted_mkey) {
-               mlx5_core_dbg(dev, "failed xarray delete of mkey 0x%x\n",
-                             mlx5_base_mkey(mkey->key));
-               return -ENOENT;
-       }
 
        MLX5_SET(destroy_mkey_in, in, opcode, MLX5_CMD_OP_DESTROY_MKEY);
        MLX5_SET(destroy_mkey_in, in, mkey_index, mlx5_mkey_to_idx(mkey->key));