Hi,
在 2025/01/27 17:00, Doug V Johnson 写道:
While adding an additional drive to a raid6 array, the reshape stalled
at about 13% complete and any I/O operations on the array hung,
creating an effective soft lock. The kernel reported a hung task in
mdXX_reshape thread and I had to use magic sysrq to recover as systemd
hung as well.
I first suspected an issue with one of the underlying block devices and
as precaution I recovered the data in read only mode to a new array, but
it turned out to be in the RAID layer as I was able to recreate the
issue from a superblock dump in sparse files.
After poking around some I discovered that I had somehow propagated the
bad block list to several devices in the array such that a few blocks
were unreable. The bad read reported correctly in userspace during
recovery, but it wasn't obvious that it was from a bad block list
metadata at the time and instead confirmed my bias suspecting hardware
issues
I was able to reproduce the issue with a minimal test case using small
loopback devices. I put a script for this in a github repository:
https://github.com/dougvj/md_badblock_reshape_stall_test
This patch handles bad reads during a reshape by introducing a
handle_failed_reshape function in a similar manner to
handle_failed_resync. The function aborts the current stripe by
unmarking STRIPE_EXPANDING and STRIP_EXPAND_READY, sets the
MD_RECOVERY_FROZEN bit, reverts the head of the reshape to the safe
position, and reports the situation in dmesg.
Signed-off-by: Doug V Johnson <dougvj@xxxxxxxxxx>
---
drivers/md/raid5.c | 23 +++++++++++++++++++++++
1 file changed, 23 insertions(+)
diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c
index 5c79429acc64..bc0b0c2540f0 100644
--- a/drivers/md/raid5.c
+++ b/drivers/md/raid5.c
@@ -3738,6 +3738,27 @@ handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), !abort);
}
+static void
+handle_failed_reshape(struct r5conf *conf, struct stripe_head *sh,
+ struct stripe_head_state *s)
+{
+ // Abort the current stripe
+ clear_bit(STRIPE_EXPANDING, &sh->state);
+ clear_bit(STRIPE_EXPAND_READY, &sh->state);
+ pr_warn_ratelimited("md/raid:%s: read error during reshape at %lu, cannot progress",
+ mdname(conf->mddev),
+ (unsigned long)sh->sector);
pr_err() is fine here.
+ // Freeze the reshape
+ set_bit(MD_RECOVERY_FROZEN, &conf->mddev->recovery);
+ // Revert progress to safe position
+ spin_lock_irq(&conf->device_lock);
+ conf->reshape_progress = conf->reshape_safe;
+ spin_unlock_irq(&conf->device_lock);
+ // report failed md sync
+ md_done_sync(conf->mddev, 0, 0);
+ wake_up(&conf->wait_for_reshape);
Just wonder, this will leave the array in the state that reshape is
still in progress, and forbid any other sync action(details in
md_choose_sync_action()). It's better to avoid that, not quite
sure yet how. :(
In theory, if user provide a new disk, this werid state can be resolved
by doing a recovery concurrent with the reshape. However, this is too
complicated and doesn't seem worth it.
Perhaps, check badblocks before starting reshape in the first place can
solve most problems in this case? We can keep this patch just in case
new badblocks are set while performing reshape.
Thanks,
Kuai
+}
+
static int want_replace(struct stripe_head *sh, int disk_idx)
{
struct md_rdev *rdev;
@@ -4987,6 +5008,8 @@ static void handle_stripe(struct stripe_head *sh)
handle_failed_stripe(conf, sh, &s, disks);
if (s.syncing + s.replacing)
handle_failed_sync(conf, sh, &s);
+ if (test_bit(STRIPE_EXPANDING, &sh->state))
+ handle_failed_reshape(conf, sh, &s);
}
/* Now we check to see if any write operations have recently
