[PATCH] scsi: target: release SAM task attribute on the abort path

[PATCH] scsi: target: release SAM task attribute on the abort path

[Yu Zhang]

A command that has started execution and is then aborted is completed
through target_handle_abort() (target_cmd_interrupted() ->
target_abort_work()), which bypasses target_complete_ok_work().

target_handle_abort() never calls transport_complete_task_attr(), the
only function that releases the SAM task-attribute accounting a command
takes in target_handle_task_attr() at execute time: for a SIMPLE command
the dev->non_ordered reference, and for an ORDERED command the
dev->ordered_sync_in_progress state (cleared, and non_ordered
resurrected, by transport_complete_ordered_sync()).

The normal completion path (target_complete_ok_work()) and the failure
path (transport_generic_request_failure()) both call
transport_complete_task_attr() before handling the command; the abort
path does not, so aborting an in-flight command leaks the accounting:

 - Aborting an in-flight SIMPLE command leaks a dev->non_ordered
   reference. A later ORDERED command then waits forever for non_ordered
   to drain and every subsequent command piles up on
   dev->delayed_cmd_list, wedging all I/O to the device.

 - Aborting an in-flight ORDERED command leaves
   dev->ordered_sync_in_progress set forever, so target_do_delayed_work()
   never makes progress and all subsequent commands to the device hang.

Call transport_complete_task_attr() from target_handle_abort() as the
completion and failure paths do. It is a no-op for commands that never
set SCF_TASK_ATTR_SET (e.g. task management requests) and is idempotent
via that flag, so it is also safe on the failure path that already
called it before dispatching to target_abort_work().

Reproduced over an iSCSI session to a demo-mode LIO target backed by a
dm-delay device: an ORDERED WRITE held in flight at the backend is
aborted with ABORT_TASK, after which a follow-up command never completes
(device wedged); with this patch the follow-up command completes
normally.

Fixes: ed1227e08099 ("scsi: target: Fix ordered tag handling")
Cc: stable@vger.kernel.org
Signed-off-by: Yu Zhang <yuz08559@gmail.com>
---
 drivers/target/target_core_transport.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/drivers/target/target_core_transport.c b/drivers/target/target_core_transport.c
index 1111111..2222222 100644
--- a/drivers/target/target_core_transport.c
+++ b/drivers/target/target_core_transport.c
@@ -870,6 +870,8 @@ static void target_handle_abort(struct se_cmd *cmd)

 	WARN_ON_ONCE(kref_read(&cmd->cmd_kref) == 0);

+	transport_complete_task_attr(cmd);
+
 	transport_lun_remove_cmd(cmd);

 	transport_cmd_check_stop_to_fabric(cmd);
--
2.43.0