Hi Bart,
On Thu, Apr 27, 2017 at 08:54:32AM -0700, Bart Van Assche wrote:
> blk_mq_quiesce_queue() callers, e.g. elevator_switch_mq(), assume
> that no .queue_rq() calls occur while switching to another I/O
I think we should call blk_mq_freeze_queue_wait() instead
of blk_quiesce_queue() in elevator_switch_mq(), otherwise
it is easy to cause use-after-free.
> scheduler. This patch fixes the following kernel crash if another
> I/O scheduler than "none" is the default scheduler:
>
> general protection fault: 0000 [#1] SMP
> RIP: 0010:__lock_acquire+0xfe/0x1280
> Call Trace:
> lock_acquire+0xd5/0x1c0
> _raw_spin_lock+0x2a/0x40
> dd_dispatch_request+0x29/0x1e0
> blk_mq_sched_dispatch_requests+0x139/0x190
> __blk_mq_run_hw_queue+0x12d/0x1c0
> blk_mq_run_work_fn+0xd/0x10
> process_one_work+0x206/0x6a0
> worker_thread+0x49/0x4a0
> kthread+0x107/0x140
> ret_from_fork+0x2e/0x40
>
> Signed-off-by: Bart Van Assche <bart.vanassche@xxxxxxxxxxx>
> Cc: Hannes Reinecke <hare@xxxxxxxx>
> Cc: Omar Sandoval <osandov@xxxxxx>
> Cc: Ming Lei <ming.lei@xxxxxxxxxx>
> Cc: <stable@xxxxxxxxxxxxxxx>
> ---
> block/blk-mq.c | 5 +++--
> 1 file changed, 3 insertions(+), 2 deletions(-)
>
> diff --git a/block/blk-mq.c b/block/blk-mq.c
> index b75ef2392db7..3b3420f76b5a 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -1224,8 +1224,9 @@ EXPORT_SYMBOL(blk_mq_queue_stopped);
>
> void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
> {
> - cancel_work(&hctx->run_work);
> - cancel_delayed_work(&hctx->delay_work);
> + cancel_work_sync(&hctx->run_work);
> + cancel_delayed_work_sync(&hctx->delay_work);
Could you explain it a bit why we need the sync version?
> + cancel_delayed_work_sync(&hctx->delayed_run_work);
More introduced, more bugs may come, :-)
So I suggest to unity both .run_work and .dealyed_run_work
into one work, just as what Jens did in the following link:
http://marc.info/?t=149183989800010&r=1&w=2
Thanks,
Ming