Hi,
在 2024/01/25 8:50, Xiao Ni 写道:
On Wed, Jan 24, 2024 at 8:19 PM Xiao Ni <xni@xxxxxxxxxx> wrote:
On Wed, Jan 24, 2024 at 5:18 PM Yu Kuai <yukuai3@xxxxxxxxxx> wrote:
First regression related to stop sync thread:
The lifetime of sync_thread is designed as following:
1) Decide want to start sync_thread, set MD_RECOVERY_NEEDED, and wake up
daemon thread;
2) Daemon thread detect that MD_RECOVERY_NEEDED is set, then set
MD_RECOVERY_RUNNING and register sync_thread;
3) Execute md_do_sync() for the actual work, if it's done or
interrupted, it will set MD_RECOVERY_DONE and wake up daemone thread;
4) Daemon thread detect that MD_RECOVERY_DONE is set, then clear
MD_RECOVERY_RUNNING and unregister sync_thread;
In v6.7, we fix md/raid to follow this design by commit f52f5c71f3d4
("md: fix stopping sync thread"), however, dm-raid is not considered at
that time, and following test will hang:
Hi Kuai
Thanks very much for the patch set. I reported the dm raid deadlock
when stopping dm raid and we had the patch set "[PATCH v5 md-fixes
0/3] md: fix stopping sync thread" which has patch f52f5c71f3d4. So we
indeed considered dm-raid that time. Because we want to resolve the
deadlock problem. I re-look patch f52f5c71f3d4. It has two major
changes. One is to use a common function stop_sync_thread for stopping
sync thread. This can fix the deadlock problem. The second change
changes the way to reap sync thread. mdraid and dmraid reap sync
thread in __md_stop_writes. So the patch looks overweight.
Before f52f5c71f3d4 do_md_stop release reconfig_mutex before waiting
sync_thread to finish. So there should not be the deadlock problem
which has been fixed in 130443d60b1b ("md: refactor
idle/frozen_sync_thread() to fix deadlock"). So we only need to change
__md_stop_writes to stop sync thread like do_md_stop and reap sync
thread directly.
Maybe this can avoid deadlock? I'll try this way and give the test result.
Please ignore my last comment. There is something wrong. Only dmraid
calls reap_sync_thread directly in __md_stop_writes before.
130443d60b1b ("md: refactor idle/frozen_sync_thread() to fix
deadlock") fixes a deadlock problem. sync io is running and user io
comes. sync io needs to wait user io. user io needs to update
suerblock and it needs mddev->reconfig_mutex. But user action happens
with this lock to stop sync thread. So this is the deadlock. For
dmraid, it doesn't update superblock like md. I'm not sure if dmraid
has such deadlock problem. If not, dmraid can call md_reap_sync_thread
directly, right?
Yes, the deadlock problem is because holding the lock to call
md_reap_sync_thread() directly will block daemon thread to handle IO.
However, for dm-raid superblock, I'm confused here, the code looks like
md superblock is still there, for example:
rs_update_sbs
set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
md_update_sb(mddev, 1);
And the code in raid1/10/5 to update md superblock doesn't have any
special handling for dm-raid. Or am I missing something here?
Thanks,
Kuai
shell/integrity-caching.sh
shell/lvconvert-raid-reshape.sh
This patch set fix the broken test by patch 1-4;
- patch 1 fix that step 4) is broken by suspended array;
- patch 2 fix that step 4) is broken by read-only array;
- patch 3 fix that step 3) is broken that md_do_sync() doesn't set
MD_RECOVERY_DONE; Noted that this patch will introdece new problem that
data will be corrupted, which will be fixed in later patches.
- patch 4 fix that setp 1) is broken that sync_thread is register and
MD_RECOVERY_RUNNING is set directly;
With patch 1-4, the above test won't hang anymore, however, the test
will still fail and complain that ext4 is corrupted;
For patch3, as I mentioned today, the root cause is
dm-raid->rs_start_reshape sets MD_RECOVERY_WAIT. So md_do_sync returns
when MD_RECOVERY_WAIT is set. It's the reason why dm-raid can't stop
sync thread when start a new reshape. . The way in patch3 looks like a
workaround. We need to figure out if dm raid really needs to set
MD_RECOVERY_WAIT. Because now we stop sync thread in an asynchronous
way. So the deadlock problem which was fixed in 644e2537f (dm raid:
fix stripe adding reshape deadlock) may disappear. Maybe we can revert
the patch.
In fact, the flag MD_RECOVERY_WAIT looks like a workaround to prevent
new sync thread to start to me. I actually frozen the sync_thread during
suspend, and prevent user to unfrozen it from raid_message() in patch 6.
I think this way is better and probably MD_RECOVERY_WAIT can be removed.
After talking with Heinz, he mentioned dmraid needs this bit to avoid
md sync thread to start during reshape. So patch3 looks good.
Best Regards
Xiao
Best Regards
Xiao
Second regression related to frozen sync thread:
Noted that for raid456, if reshape is interrupted, then call
"pers->start_reshape" will corrupt data. This is because dm-raid rely on
md_do_sync() doesn't set MD_RECOVERY_DONE so that new sync_thread won't
be registered, and patch 3 just break this.
- Patch 5-6 fix this problem by interrupting reshape and frozen
sync_thread in dm_suspend(), then unfrozen and continue reshape in
dm_resume(). It's verified that dm-raid tests won't complain that
ext4 is corrupted anymore.
- Patch 7 fix the problem that raid_message() call
md_reap_sync_thread() directly, without holding 'reconfig_mutex'.
Last regression related to dm-raid456 IO concurrent with reshape:
For raid456, if reshape is still in progress, then IO across reshape
position will wait for reshape to make progress. However, for dm-raid,
in following cases reshape will never make progress hence IO will hang:
1) the array is read-only;
2) MD_RECOVERY_WAIT is set;
3) MD_RECOVERY_FROZEN is set;
After commit c467e97f079f ("md/raid6: use valid sector values to determine
if an I/O should wait on the reshape") fix the problem that IO across
reshape position doesn't wait for reshape, the dm-raid test
shell/lvconvert-raid-reshape.sh start to hang at raid5_make_request().
For md/raid, the problem doesn't exist because:
1) If array is read-only, it can switch to read-write by ioctl/sysfs;
2) md/raid never set MD_RECOVERY_WAIT;
3) If MD_RECOVERY_FROZEN is set, mddev_suspend() doesn't hold
'reconfig_mutex' anymore, it can be cleared and reshape can continue by
sysfs api 'sync_action'.
However, I'm not sure yet how to avoid the problem in dm-raid yet.
- patch 9-11 fix this problem by detecting the above 3 cases in
dm_suspend(), and fail those IO directly.
If user really meet the IO error, then it means they're reading the wrong
data before c467e97f079f. And it's safe to read/write the array after
reshape make progress successfully.
Tests:
I already run the following two tests many times and verified that they
won't fail anymore:
shell/integrity-caching.sh
shell/lvconvert-raid-reshape.sh
For other tests, I'm still running. However, I'm sending this patchset
in case people think the fixes is not appropriate. Running the full
tests will cost lots of time in my VM, and I'll update full test results
soon.
Yu Kuai (11):
md: don't ignore suspended array in md_check_recovery()
md: don't ignore read-only array in md_check_recovery()
md: make sure md_do_sync() will set MD_RECOVERY_DONE
md: don't register sync_thread for reshape directly
md: export helpers to stop sync_thread
dm-raid: really frozen sync_thread during suspend
md/dm-raid: don't call md_reap_sync_thread() directly
dm-raid: remove mddev_suspend/resume()
dm-raid: add a new helper prepare_suspend() in md_personality
md: export helper md_is_rdwr()
md/raid456: fix a deadlock for dm-raid456 while io concurrent with
reshape
drivers/md/dm-raid.c | 76 +++++++++++++++++++++----------
drivers/md/md.c | 104 ++++++++++++++++++++++++++++---------------
drivers/md/md.h | 16 +++++++
drivers/md/raid10.c | 16 +------
drivers/md/raid5.c | 61 +++++++++++++------------
5 files changed, 171 insertions(+), 102 deletions(-)
--
2.39.2
.
