Hi,

在 2023/09/20 16:46, Xiao Ni 写道:

On Mon, Aug 28, 2023 at 10:04 AM Yu Kuai [off-list ref] wrote:

quoted

From: Yu Kuai <redacted>

'pers->prepare_suspend' is introduced to prevent a deadlock for raid456,
this change prepares to clean this up in later patches while refactoring
mddev_suspend(). Specifically allow reshape to make progress while
waiting for 'active_io' to be 0.

Hi Kuai

quoted

From my side, I can't understand the comments. The change has

relationship with pers->prepare_suspend? And why this change can
affect reshape? If this change indeed can affect these two things, can
you explain more?

First of all, 'prepare_suspend' is used to fix a deadlock in raid456:

1) suspend is waiting for normal io to be done.

mddev_suspend
  mddev->suspended++ -> new sync_thread can't start
  percpu_ref_kill(active_io)
  wait_event(percpu_ref_is_zero(active_io))

2) normal io is waiting for reshape to make progress.
3) reshape is waiting for suspended array to be resumed.

md_check_recovery
  if (is_md_suspended(mddev))
   return

Then prepare_suspend will failed the io that is waiting for reshape to
make progress:

mddev_suspend
  mddev->suspended++
  percpu_ref_kill(active_io)
   -> new io will be stuck in md_handle_request
  pers->prepare_suspend() -> raid5_prepare_suspend
   -> wake_up(wait_for_overlap)
		// woke up
		raid5_make_request
		 make_stripe_request
		  !reshape_inprogress(mddev) && reshape_disabled(mddev)
		   // return io error for the io that is waiting for
		   // reshape to make progress

  wait_event(percpu_ref_is_zero(active_io))

With this patch and the new api to suspend array:

mddev_suspend
  percpu_ref_kill(active_io)
  wait_event(percpu_ref_is_zero(active_io))
  -> while waiting for normal io to be done, new sync_thread can still
     start, and reshape can still make progress.
  mddev->suspended++

quoted

Signed-off-by: Yu Kuai <redacted>
---
  drivers/md/md.c | 2 +-
  drivers/md/md.h | 2 +-
  2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/md/md.c b/drivers/md/md.c
index 9d8dff9d923c..7fa311a14317 100644
--- a/drivers/md/md.c
+++ b/drivers/md/md.c

@@ -355,7 +355,7 @@ static DEFINE_SPINLOCK(all_mddevs_lock);
   */
  static bool is_suspended(struct mddev *mddev, struct bio *bio)
  {
-       if (is_md_suspended(mddev))
+       if (is_md_suspended(mddev) || percpu_ref_is_dying(&mddev->active_io))

If we use mddev->suspended to judge if the raid is suspended, it
should be enough? Because mddev->suspended must be true when active_io
is dying.

In the new api, active_io is killed before increasing suspended, and the
difference is that the timing that array is suspended will be delayed
from the start of mddev_suspend() to when all dispatched io is done.

I think this is OK because this doesn't change behaviour when
mddev_suspend() returns.

Thanks,
Kuai
`

Best Regards
Xiao

quoted

                 return true;
         if (bio_data_dir(bio) != WRITE)
                 return false;

diff --git a/drivers/md/md.h b/drivers/md/md.h
index b628c292506e..fb3b123f16dd 100644
--- a/drivers/md/md.h
+++ b/drivers/md/md.h

@@ -584,7 +584,7 @@ static inline bool md_is_rdwr(struct mddev *mddev)

  static inline bool is_md_suspended(struct mddev *mddev)
  {
-       return percpu_ref_is_dying(&mddev->active_io);
+       return READ_ONCE(mddev->suspended);
  }

  static inline int __must_check mddev_lock(struct mddev *mddev)
--

2.39.2

.