On Wed, Jun 8, 2022 at 11:21 AM Logan Gunthorpe [off-list ref] wrote:

On 2022-06-08 11:59, Song Liu wrote:

quoted

On Wed, Jun 8, 2022 at 9:28 AM Logan Gunthorpe [off-list ref] wrote:

quoted

The raid5-cache code relies on there being no IO in flight when
log_exit() is called. There are two places where this is not
guaranteed so add mddev_suspend() and mddev_resume() calls to these
sites.

The site in raid5_remove_disk() has a comment saying that it is
called in raid5d and thus cannot wait for pending writes; however that
does not appear to be correct anymore (if it ever was) as
raid5_remove_disk() is called from hot_remove_disk() which only
appears to be called in the md_ioctl(). Thus, the comment is removed,
as well as the racy check and replaced with calls to suspend/resume.

The site in raid5_change_consistency_policy() is in the error path,
and another similar call site already has suspend/resume calls just
below it; so it should be equally safe to make that change here.

Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
---
 drivers/md/raid5.c | 18 ++++++------------
 1 file changed, 6 insertions(+), 12 deletions(-)

diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c
index 5d09256d7f81..3ad37dd4c5cd 100644
--- a/drivers/md/raid5.c
+++ b/drivers/md/raid5.c

@@ -7938,18 +7938,9 @@ static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)

        print_raid5_conf(conf);
        if (test_bit(Journal, &rdev->flags) && conf->log) {
-               /*
-                * we can't wait pending write here, as this is called in
-                * raid5d, wait will deadlock.
-                * neilb: there is no locking about new writes here,
-                * so this cannot be safe.
-                */
-               if (atomic_read(&conf->active_stripes) ||
-                   atomic_read(&conf->r5c_cached_full_stripes) ||
-                   atomic_read(&conf->r5c_cached_partial_stripes)) {
-                       return -EBUSY;
-               }
+               mddev_suspend(mddev);

Unfortunately, the comment about deadlock is still true, and we cannot call
mddev_suspend here. To trigger it:

Ah, yes. What a tangle. I think we can just drop this patch. Now that we
are removing RCU it isn't actually necessary to fix the bug I was
seeing. It's still probably broken as the comment notes though.

How about we keep the suspend/resume in raid5_change_consistency_policy()?
Like the one I just pushed to md-next:

https://git.kernel.org/pub/scm/linux/kernel/git/song/md.git/commit/?h=md-next&id=ac1506992459fe45a085c1f93df74d51c381887b

Thanks,
Song