On Thu, 2018-10-25 at 21:51 +0200, Johannes Berg wrote:
> [ ... ]
> diff --git a/fs/direct-io.c b/fs/direct-io.c
> index 093fb54cd316..9ef33d6cba56 100644
> --- a/fs/direct-io.c
> +++ b/fs/direct-io.c
> @@ -629,9 +629,16 @@ int sb_init_dio_done_wq(struct super_block *sb)
> * This has to be atomic as more DIOs can race to create the workqueue
> */
> old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
> - /* Someone created workqueue before us? Free ours... */
> + /*
> + * Someone created workqueue before us? Free ours...
> + * Note the _nested(), that pushes down to the (in this case actually
> + * pointless) flush_workqueue() happening inside, since this function
> + * might be called in contexts that hold the same locks that an fs may
> + * take while being called from dio_aio_complete_work() from another
> + * instance of the workqueue we allocate here.
> + */
> if (old)
> - destroy_workqueue(wq);
> + destroy_workqueue_nested(wq, SINGLE_DEPTH_NESTING);
> return 0;
> }
> [ ... ]
> /**
> - * flush_workqueue - ensure that any scheduled work has run to completion.
> + * flush_workqueue_nested - ensure that any scheduled work has run to
> completion.
> * @wq: workqueue to flush
> + * @subclass: subclass for lockdep
> *
> * This function sleeps until all work items which were queued on entry
> * have finished execution, but it is not livelocked by new incoming ones.
> */
> -void flush_workqueue(struct workqueue_struct *wq)
> +void flush_workqueue_nested(struct workqueue_struct *wq, int subclass)
> {
> struct wq_flusher this_flusher = {
> .list = LIST_HEAD_INIT(this_flusher.list),
> @@ -2652,7 +2653,7 @@ void flush_workqueue(struct workqueue_struct *wq)
> if (WARN_ON(!wq_online))
> return;
>
> - lock_map_acquire(&wq->lockdep_map);
> + lock_acquire_exclusive(&wq->lockdep_map, subclass, 0, NULL, _THIS_IP_);
> lock_map_release(&wq->lockdep_map);
>
> mutex_lock(&wq->mutex);
> [ ... ]
I don't like this approach because it doesn't match how other kernel code uses
lockdep annotations. All other kernel code I know of only annotates lock depmaps
as nested if the same kernel thread calls lock_acquire() twice for the same
depmap
without intervening lock_release(). My understanding is that with your patch
applied flush_workqueue_nested(wq, 1) calls lock_acquire() only once and with
the
subclass argument set to one. I think this will confuse other people who will
read
the workqueue implementation and who have not followed this conversation.
I like Tejuns proposal much better than the above proposal.
Bart.
