nvme: drop scan_lock and always kick requeue list when removing namespaces
authorHannes Reinecke <hare@suse.de>
Wed, 20 Oct 2021 05:59:10 +0000 (07:59 +0200)
committerChristoph Hellwig <hch@lst.de>
Wed, 20 Oct 2021 17:23:30 +0000 (19:23 +0200)
When reading the partition table on initial scan hits an I/O error the
I/O will hang with the scan_mutex held:

[<0>] do_read_cache_page+0x49b/0x790
[<0>] read_part_sector+0x39/0xe0
[<0>] read_lba+0xf9/0x1d0
[<0>] efi_partition+0xf1/0x7f0
[<0>] bdev_disk_changed+0x1ee/0x550
[<0>] blkdev_get_whole+0x81/0x90
[<0>] blkdev_get_by_dev+0x128/0x2e0
[<0>] device_add_disk+0x377/0x3c0
[<0>] nvme_mpath_set_live+0x130/0x1b0 [nvme_core]
[<0>] nvme_mpath_add_disk+0x150/0x160 [nvme_core]
[<0>] nvme_alloc_ns+0x417/0x950 [nvme_core]
[<0>] nvme_validate_or_alloc_ns+0xe9/0x1e0 [nvme_core]
[<0>] nvme_scan_work+0x168/0x310 [nvme_core]
[<0>] process_one_work+0x231/0x420

and trying to delete the controller will deadlock as it tries to grab
the scan mutex:

[<0>] nvme_mpath_clear_ctrl_paths+0x25/0x80 [nvme_core]
[<0>] nvme_remove_namespaces+0x31/0xf0 [nvme_core]
[<0>] nvme_do_delete_ctrl+0x4b/0x80 [nvme_core]

As we're now properly ordering the namespace list there is no need to
hold the scan_mutex in nvme_mpath_clear_ctrl_paths() anymore.
And we always need to kick the requeue list as the path will be marked
as unusable and I/O will be requeued _without_ a current path.

Signed-off-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
drivers/nvme/host/multipath.c

index bc1e7d8..954e84d 100644 (file)
@@ -146,13 +146,12 @@ void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl)
 {
        struct nvme_ns *ns;
 
-       mutex_lock(&ctrl->scan_lock);
        down_read(&ctrl->namespaces_rwsem);
-       list_for_each_entry(ns, &ctrl->namespaces, list)
-               if (nvme_mpath_clear_current_path(ns))
-                       kblockd_schedule_work(&ns->head->requeue_work);
+       list_for_each_entry(ns, &ctrl->namespaces, list) {
+               nvme_mpath_clear_current_path(ns);
+               kblockd_schedule_work(&ns->head->requeue_work);
+       }
        up_read(&ctrl->namespaces_rwsem);
-       mutex_unlock(&ctrl->scan_lock);
 }
 
 void nvme_mpath_revalidate_paths(struct nvme_ns *ns)