Re: [PATCH v2 3/3] md/raid10: skip clean regions in bulk during recovery

[yu kuai]

Hi,

在 2026/8/4 11:55, Yunye Zhao 写道:
> Hi Kuai,
>
> On Thu, 31 Jul 2026 at 17:46, Yu Kuai wrote:
>> I'm confused, recovery means replace a new disk ... the whole new disk
>> have to rebuild ... there is nothing to skip
>>
>> So are you talking about the case that the array is already broken?
> The recovery target here is a *re-added existing member*, not a fresh
> spare.

By re-add a disk, this is really a resync, and bitmap can be used in this case
to skip clean bits. In this case, I still think the correct solution, for old bitmap,
is to let bitmap_start_sync() return much bigger skipped blocks.

>
> Right, for a fresh spare: saved_raid_disk != slot sets conf->fullsync=1,
> this skip is bypassed, and the whole disk rebuilds. This patch does not
> touch that.
This is not re-add and it's a real recovery.
>
> The case I fix is a member transiently removed and re-added:
>
>    - a member drops out; its mirror partner stays online, so the array is
>      still redundant and readable (not broken);
>    - nothing is written while it is gone, so the write-intent bitmap
>      stays clean;
>    - the same disk is re-added: valid superblock, saved_raid_disk == old
>      slot, so conf->fullsync stays 0 and md does a bitmap-based recovery
>      (MD_RECOVERY_RECOVER; /proc/mdstat shows "recovery").
>
> The bitmap is fully clean, every chunk returns must_sync == false,
> biolist stays NULL, and we hit the existing "*skipped=1; return
> max_sync;" path. Only its step size is wrong.
>
> Reproducer (near=2, 4 disks, nothing written while degraded, each pair
> keeps a live copy):
>
>    mdadm --create /dev/md0 -l10 -n4 -p n2 --bitmap=internal \
>          --bitmap-chunk=512M --assume-clean --run <4 x 8T>
>    mdadm /dev/md0 --fail /dev/loop1 --remove /dev/loop1
>    mdadm /dev/md0 --fail /dev/loop3 --remove /dev/loop3
>    # no writes here
>    mdadm /dev/md0 --re-add /dev/loop1
>    mdadm /dev/md0 --re-add /dev/loop3
>
> Thanks,
> Yunye

-- 
Thanks,
Kuai