From: Li Nan <[email protected]>
Recovery will go to giveup and let chunks_skipped++ in
raid10_sync_request() if there are some bad_blocks, and it will return
max_sector when chunks_skipped >= geo.raid_disks. Now, recovery fail and
data is inconsistent but user think recovery is done, it is wrong.
Fix it by set mirror's recovery_disabled, spare device will not be added
to here. The same issue alos exists on resync, it will be fixd in future.
Signed-off-by: Li Nan <[email protected]>
---
drivers/md/raid10.c | 18 +++++++++++++++++-
1 file changed, 17 insertions(+), 1 deletion(-)
diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c
index d93d8cb2b620..3ba1516ea160 100644
--- a/drivers/md/raid10.c
+++ b/drivers/md/raid10.c
@@ -3303,6 +3303,7 @@ static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
int chunks_skipped = 0;
sector_t chunk_mask = conf->geo.chunk_mask;
int page_idx = 0;
+ int error_disk = -1;
/*
* Allow skipping a full rebuild for incremental assembly
@@ -3386,7 +3387,20 @@ static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
return reshape_request(mddev, sector_nr, skipped);
if (chunks_skipped >= conf->geo.raid_disks) {
- /* if there has been nothing to do on any drive,
+ pr_err("md/raid10:%s: %s fail\n", mdname(mddev),
+ test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ? "resync" : "recovery");
+ if (error_disk >= 0 &&
+ !test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
+ /*
+ * recovery fail, set mirrors.recovory_disabled,
+ * device shouldn't be added to there.
+ */
+ conf->mirrors[error_disk].recovery_disabled =
+ mddev->recovery_disabled;
+ return 0;
+ }
+ /*
+ * if there has been nothing to do on any drive,
* then there is nothing to do at all..
*/
*skipped = 1;
@@ -3638,6 +3652,8 @@ static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
mdname(mddev));
mirror->recovery_disabled
= mddev->recovery_disabled;
+ } else {
+ error_disk = i;
}
put_buf(r10_bio);
if (rb2)
--
2.31.1
On Sat, May 27, 2023 at 12:24 AM <[email protected]> wrote:
>
> From: Li Nan <[email protected]>
>
> Recovery will go to giveup and let chunks_skipped++ in
> raid10_sync_request() if there are some bad_blocks, and it will return
> max_sector when chunks_skipped >= geo.raid_disks. Now, recovery fail and
> data is inconsistent but user think recovery is done, it is wrong.
>
> Fix it by set mirror's recovery_disabled, spare device will not be added
> to here. The same issue alos exists on resync, it will be fixd in future.
>
> Signed-off-by: Li Nan <[email protected]>
I applied 1/4 and 2/4 of the set to md-next.
For 3/4 and 4/4, please improve the commit log (rephrase confusing statements,
fix typo's, etc.). Please also add a mdadm test for 3/4.
Thanks,
Song
> ---
> drivers/md/raid10.c | 18 +++++++++++++++++-
> 1 file changed, 17 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c
> index d93d8cb2b620..3ba1516ea160 100644
> --- a/drivers/md/raid10.c
> +++ b/drivers/md/raid10.c
> @@ -3303,6 +3303,7 @@ static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
> int chunks_skipped = 0;
> sector_t chunk_mask = conf->geo.chunk_mask;
> int page_idx = 0;
> + int error_disk = -1;
>
> /*
> * Allow skipping a full rebuild for incremental assembly
> @@ -3386,7 +3387,20 @@ static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
> return reshape_request(mddev, sector_nr, skipped);
>
> if (chunks_skipped >= conf->geo.raid_disks) {
> - /* if there has been nothing to do on any drive,
> + pr_err("md/raid10:%s: %s fail\n", mdname(mddev),
> + test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ? "resync" : "recovery");
> + if (error_disk >= 0 &&
> + !test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
> + /*
> + * recovery fail, set mirrors.recovory_disabled,
> + * device shouldn't be added to there.
> + */
> + conf->mirrors[error_disk].recovery_disabled =
> + mddev->recovery_disabled;
> + return 0;
> + }
> + /*
> + * if there has been nothing to do on any drive,
> * then there is nothing to do at all..
> */
> *skipped = 1;
> @@ -3638,6 +3652,8 @@ static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
> mdname(mddev));
> mirror->recovery_disabled
> = mddev->recovery_disabled;
> + } else {
> + error_disk = i;
> }
> put_buf(r10_bio);
> if (rb2)
> --
> 2.31.1
>
在 2023/5/31 5:55, Song Liu 写道:
> On Sat, May 27, 2023 at 12:24 AM <[email protected]> wrote:
>>
>> From: Li Nan <[email protected]>
>>
>> Recovery will go to giveup and let chunks_skipped++ in
>> raid10_sync_request() if there are some bad_blocks, and it will return
>> max_sector when chunks_skipped >= geo.raid_disks. Now, recovery fail and
>> data is inconsistent but user think recovery is done, it is wrong.
>>
>> Fix it by set mirror's recovery_disabled, spare device will not be added
>> to here. The same issue alos exists on resync, it will be fixd in future.
>>
>> Signed-off-by: Li Nan <[email protected]>
>
> I applied 1/4 and 2/4 of the set to md-next.
>
> For 3/4 and 4/4, please improve the commit log (rephrase confusing statements,
> fix typo's, etc.). Please also add a mdadm test for 3/4.
>
> Thanks,
> Song
>
I will add a test later. Thanks for review.
--
Thanks,
Nan