lkml.org 
[lkml]   [2024]   [Mar]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [REGRESSION] 6.7.1: md: raid5 hang and unresponsive system; successfully bisected
From
Date
Hi,

在 2024/03/13 6:56, junxiao.bi@oracle.com 写道:
> On 3/10/24 6:50 PM, Yu Kuai wrote:
>
>> Hi,
>>
>> 在 2024/03/09 7:49, junxiao.bi@oracle.com 写道:
>>> Here is the root cause for this issue:
>>>
>>> Commit 5e2cf333b7bd ("md/raid5: Wait for MD_SB_CHANGE_PENDING in
>>> raid5d") introduced a regression, it got reverted through commit
>>> bed9e27baf52 ("Revert "md/raid5: Wait for MD_SB_CHANGE_PENDING in
>>> raid5d"). To fix the original issue commit 5e2cf333b7bd was fixing,
>>> commit d6e035aad6c0 ("md: bypass block throttle for superblock
>>> update") was created, it avoids md superblock write getting throttled
>>> by block layer which is good, but md superblock write could be stuck
>>> in block layer due to block flush as well, and that is what was
>>> happening in this regression report.
>>>
>>> Process "md0_reclaim" got stuck while waiting IO for md superblock
>>> write done, that IO was marked with REQ_PREFLUSH | REQ_FUA flags,
>>> these 3 steps ( PREFLUSH, DATA and POSTFLUSH ) will be executed
>>> before done, the hung of this process is because the last step
>>> "POSTFLUSH" never done. And that was because of  process "md0_raid5"
>>> submitted another IO with REQ_FUA flag marked just before that step
>>> started. To handle that IO, blk_insert_flush() will be invoked and
>>> hit "REQ_FSEQ_DATA | REQ_FSEQ_POSTFLUSH" case where
>>> "fq->flush_data_in_flight" will be increased. When the IO for md
>>> superblock write was to issue "POSTFLUSH" step through
>>> blk_kick_flush(), it found that "fq->flush_data_in_flight" was not
>>> zero, so it will skip that step, that is expected, because flush will
>>> be triggered when "fq->flush_data_in_flight" dropped to zero.
>>>
>>> Unfortunately here that inflight data IO from "md0_raid5" will never
>>> done, because it was added into the blk_plug list of that process,
>>> but "md0_raid5" run into infinite loop due to "MD_SB_CHANGE_PENDING"
>>> which made it never had a chance to finish the blk plug until
>>> "MD_SB_CHANGE_PENDING" was cleared. Process "md0_reclaim" was
>>> supposed to clear that flag but it was stuck by "md0_raid5", so this
>>> is a deadlock.
>>>
>>> Looks like the approach in the RFC patch trying to resolve the
>>> regression of commit 5e2cf333b7bd can help this issue. Once
>>> "md0_raid5" starts looping due to "MD_SB_CHANGE_PENDING", it should
>>> release all its staging IO requests to avoid blocking others. Also a
>>> cond_reschedule() will avoid it run into lockup.
>>
>> The analysis sounds good, however, it seems to me that the behaviour
>> raid5d() pings the cpu to wait for 'MD_SB_CHANGE_PENDING' to be cleared
>> is not reasonable, because md_check_recovery() must hold
>> 'reconfig_mutex' to clear the flag.
>
> That's the behavior before commit 5e2cf333b7bd which was added into Sep
> 2022, so this behavior has been with raid5 for many years.
>

Yes, it exists for a long time doesn't mean it's good. It is really
weird to hold spinlock to wait for a mutex.
>
>>
>> Look at raid1/raid10, there are two different behaviour that seems can
>> avoid this problem as well:
>>
>> 1) blk_start_plug() is delayed until all failed IO is handled. This look
>> reasonable because in order to get better performance, IO should be
>> handled by submitted thread as much as possible, and meanwhile, the
>> deadlock can be triggered here.
>> 2) if 'MD_SB_CHANGE_PENDING' is not cleared by md_check_recovery(), skip
>> the handling of failed IO, and when mddev_unlock() is called, daemon
>> thread will be woken up again to handle failed IO.
>>
>> How about the following patch?
>>
>> Thanks,
>> Kuai
>>
>> diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c
>> index 3ad5f3c7f91e..0b2e6060f2c9 100644
>> --- a/drivers/md/raid5.c
>> +++ b/drivers/md/raid5.c
>> @@ -6720,7 +6720,6 @@ static void raid5d(struct md_thread *thread)
>>
>>         md_check_recovery(mddev);
>>
>> -       blk_start_plug(&plug);
>>         handled = 0;
>>         spin_lock_irq(&conf->device_lock);
>>         while (1) {
>> @@ -6728,6 +6727,14 @@ static void raid5d(struct md_thread *thread)
>>                 int batch_size, released;
>>                 unsigned int offset;
>>
>> +               /*
>> +                * md_check_recovery() can't clear sb_flags, usually
>> because of
>> +                * 'reconfig_mutex' can't be grabbed, wait for
>> mddev_unlock() to
>> +                * wake up raid5d().
>> +                */
>> +               if (test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
>> +                       goto skip;
>> +
>>                 released = release_stripe_list(conf,
>> conf->temp_inactive_list);
>>                 if (released)
>>                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
>> @@ -6766,8 +6773,8 @@ static void raid5d(struct md_thread *thread)
>>                         spin_lock_irq(&conf->device_lock);
>>                 }
>>         }
>> +skip:
>>         pr_debug("%d stripes handled\n", handled);
>> -
>>         spin_unlock_irq(&conf->device_lock);
>>         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
>>             mutex_trylock(&conf->cache_size_mutex)) {
>> @@ -6779,6 +6786,7 @@ static void raid5d(struct md_thread *thread)
>>                 mutex_unlock(&conf->cache_size_mutex);
>>         }
>>
>> +       blk_start_plug(&plug);
>>         flush_deferred_bios(conf);
>>
>>         r5l_flush_stripe_to_raid(conf->log);
>
> This patch eliminated the benefit of blk_plug, i think it will not be
> good for IO performance perspective?

There is only one daemon thread, so IO should not be handled here as
much as possible. The IO should be handled by the thread that is
submitting the IO, and let daemon to hanldle the case that IO failed or
can't be submitted at that time.

Thanks,
Kuai

>
>
> Thanks,
>
> Junxiao.
>
>>
>>>
>>> https://www.spinics.net/lists/raid/msg75338.html
>>>
>>> Dan, can you try the following patch?
>>>
>>> diff --git a/block/blk-core.c b/block/blk-core.c
>>> index de771093b526..474462abfbdc 100644
>>> --- a/block/blk-core.c
>>> +++ b/block/blk-core.c
>>> @@ -1183,6 +1183,7 @@ void __blk_flush_plug(struct blk_plug *plug,
>>> bool from_schedule)
>>>          if (unlikely(!rq_list_empty(plug->cached_rq)))
>>>                  blk_mq_free_plug_rqs(plug);
>>>   }
>>> +EXPORT_SYMBOL(__blk_flush_plug);
>>>
>>>   /**
>>>    * blk_finish_plug - mark the end of a batch of submitted I/O
>>> diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c
>>> index 8497880135ee..26e09cdf46a3 100644
>>> --- a/drivers/md/raid5.c
>>> +++ b/drivers/md/raid5.c
>>> @@ -6773,6 +6773,11 @@ static void raid5d(struct md_thread *thread)
>>> spin_unlock_irq(&conf->device_lock);
>>>                          md_check_recovery(mddev);
>>> spin_lock_irq(&conf->device_lock);
>>> +               } else {
>>> + spin_unlock_irq(&conf->device_lock);
>>> +                       blk_flush_plug(&plug, false);
>>> +                       cond_resched();
>>> + spin_lock_irq(&conf->device_lock);
>>>                  }
>>>          }
>>>          pr_debug("%d stripes handled\n", handled);
>>>
>>> Thanks,
>>>
>>> Junxiao.
>>>
>>> On 3/1/24 12:26 PM, junxiao.bi@oracle.com wrote:
>>>> Hi Dan & Song,
>>>>
>>>> I have not root cause this yet, but would like share some findings
>>>> from the vmcore Dan shared. From what i can see, this doesn't look
>>>> like a md issue, but something wrong with block layer or below.
>>>>
>>>> 1. There were multiple process hung by IO over 15mins.
>>>>
>>>> crash> ps -m | grep UN
>>>> [0 00:15:50.424] [UN]  PID: 957      TASK: ffff88810baa0ec0 CPU: 1
>>>> COMMAND: "jbd2/dm-3-8"
>>>> [0 00:15:56.151] [UN]  PID: 1835     TASK: ffff888108a28ec0 CPU: 2
>>>> COMMAND: "dd"
>>>> [0 00:15:56.187] [UN]  PID: 876      TASK: ffff888108bebb00 CPU: 3
>>>> COMMAND: "md0_reclaim"
>>>> [0 00:15:56.185] [UN]  PID: 1914     TASK: ffff8881015e6740 CPU: 1
>>>> COMMAND: "kworker/1:2"
>>>> [0 00:15:56.255] [UN]  PID: 403      TASK: ffff888101351d80 CPU: 7
>>>> COMMAND: "kworker/u21:1"
>>>>
>>>> 2. Let pick md0_reclaim to take a look, it is waiting done
>>>> super_block update. We can see there were two pending superblock
>>>> write and other pending io for the underling physical disk, which
>>>> caused these process hung.
>>>>
>>>> crash> bt 876
>>>> PID: 876      TASK: ffff888108bebb00  CPU: 3    COMMAND: "md0_reclaim"
>>>>  #0 [ffffc900008c3d10] __schedule at ffffffff81ac18ac
>>>>  #1 [ffffc900008c3d70] schedule at ffffffff81ac1d82
>>>>  #2 [ffffc900008c3d88] md_super_wait at ffffffff817df27a
>>>>  #3 [ffffc900008c3dd0] md_update_sb at ffffffff817df609
>>>>  #4 [ffffc900008c3e20] r5l_do_reclaim at ffffffff817d1cf4
>>>>  #5 [ffffc900008c3e98] md_thread at ffffffff817db1ef
>>>>  #6 [ffffc900008c3ef8] kthread at ffffffff8114f8ee
>>>>  #7 [ffffc900008c3f30] ret_from_fork at ffffffff8108bb98
>>>>  #8 [ffffc900008c3f50] ret_from_fork_asm at ffffffff81000da1
>>>>
>>>> crash> mddev.pending_writes,disks 0xffff888108335800
>>>>   pending_writes = {
>>>>     counter = 2  <<<<<<<<<< 2 active super block write
>>>>   },
>>>>   disks = {
>>>>     next = 0xffff88810ce85a00,
>>>>     prev = 0xffff88810ce84c00
>>>>   },
>>>> crash> list -l md_rdev.same_set -s md_rdev.kobj.name,nr_pending
>>>> 0xffff88810ce85a00
>>>> ffff88810ce85a00
>>>>   kobj.name = 0xffff8881067c1a00 "dev-dm-1",
>>>>   nr_pending = {
>>>>     counter = 0
>>>>   },
>>>> ffff8881083ace00
>>>>   kobj.name = 0xffff888100a93280 "dev-sde",
>>>>   nr_pending = {
>>>>     counter = 10 <<<<
>>>>   },
>>>> ffff8881010ad200
>>>>   kobj.name = 0xffff8881012721c8 "dev-sdc",
>>>>   nr_pending = {
>>>>     counter = 8 <<<<<
>>>>   },
>>>> ffff88810ce84c00
>>>>   kobj.name = 0xffff888100325f08 "dev-sdd",
>>>>   nr_pending = {
>>>>     counter = 2 <<<<<
>>>>   },
>>>>
>>>> 3. From block layer, i can find the inflight IO for md superblock
>>>> write which has been pending 955s which matches with the hung time
>>>> of "md0_reclaim"
>>>>
>>>> crash>
>>>> request.q,mq_hctx,cmd_flags,rq_flags,start_time_ns,bio,biotail,state,__data_len,flush,end_io
>>>> ffff888103b4c300
>>>>   q = 0xffff888103a00d80,
>>>>   mq_hctx = 0xffff888103c5d200,
>>>>   cmd_flags = 38913,
>>>>   rq_flags = 139408,
>>>>   start_time_ns = 1504179024146,
>>>>   bio = 0x0,
>>>>   biotail = 0xffff888120758e40,
>>>>   state = MQ_RQ_COMPLETE,
>>>>   __data_len = 0,
>>>>   flush = {
>>>>     seq = 3, <<<< REQ_FSEQ_PREFLUSH | REQ_FSEQ_DATA
>>>>     saved_end_io = 0x0
>>>>   },
>>>>   end_io = 0xffffffff815186e0 <mq_flush_data_end_io>,
>>>>
>>>> crash> p tk_core.timekeeper.tkr_mono.base
>>>> $1 = 2459916243002
>>>> crash> eval 2459916243002-1504179024146
>>>> hexadecimal: de86609f28
>>>>     decimal: 955737218856  <<<<<<< IO pending time is 955s
>>>>       octal: 15720630117450
>>>>      binary:
>>>> 0000000000000000000000001101111010000110011000001001111100101000
>>>>
>>>> crash> bio.bi_iter,bi_end_io 0xffff888120758e40
>>>>   bi_iter = {
>>>>     bi_sector = 8, <<<< super block offset
>>>>     bi_size = 0,
>>>>     bi_idx = 0,
>>>>     bi_bvec_done = 0
>>>>   },
>>>>   bi_end_io = 0xffffffff817dca50 <super_written>,
>>>> crash> dev -d | grep ffff888103a00d80
>>>>     8 ffff8881003ab000   sdd        ffff888103a00d80       0 0 0
>>>>
>>>> 4. Check above request, even its state is "MQ_RQ_COMPLETE", but it
>>>> is still pending. That's because each md superblock write was marked
>>>> with REQ_PREFLUSH | REQ_FUA, so it will be handled in 3 steps:
>>>> pre_flush, data, and post_flush. Once each step complete, it will be
>>>> marked in "request.flush.seq", here the value is 3, which is
>>>> REQ_FSEQ_PREFLUSH |  REQ_FSEQ_DATA, so the last step "post_flush"
>>>> has not be done. Another wired thing is that
>>>> blk_flush_queue.flush_data_in_flight is still 1 even "data" step
>>>> already done.
>>>>
>>>> crash> blk_mq_hw_ctx.fq 0xffff888103c5d200
>>>>   fq = 0xffff88810332e240,
>>>> crash> blk_flush_queue 0xffff88810332e240
>>>> struct blk_flush_queue {
>>>>   mq_flush_lock = {
>>>>     {
>>>>       rlock = {
>>>>         raw_lock = {
>>>>           {
>>>>             val = {
>>>>               counter = 0
>>>>             },
>>>>             {
>>>>               locked = 0 '\000',
>>>>               pending = 0 '\000'
>>>>             },
>>>>             {
>>>>               locked_pending = 0,
>>>>               tail = 0
>>>>             }
>>>>           }
>>>>         }
>>>>       }
>>>>     }
>>>>   },
>>>>   flush_pending_idx = 1,
>>>>   flush_running_idx = 1,
>>>>   rq_status = 0 '\000',
>>>>   flush_pending_since = 4296171408,
>>>>   flush_queue = {{
>>>>       next = 0xffff88810332e250,
>>>>       prev = 0xffff88810332e250
>>>>     }, {
>>>>       next = 0xffff888103b4c348, <<<< the request is in this list
>>>>       prev = 0xffff888103b4c348
>>>>     }},
>>>>   flush_data_in_flight = 1,  >>>>>> still 1
>>>>   flush_rq = 0xffff888103c2e000
>>>> }
>>>>
>>>> crash> list 0xffff888103b4c348
>>>> ffff888103b4c348
>>>> ffff88810332e260
>>>>
>>>> crash> request.tag,state,ref 0xffff888103c2e000 >>>> flush_rq of hw
>>>> queue
>>>>   tag = -1,
>>>>   state = MQ_RQ_IDLE,
>>>>   ref = {
>>>>     counter = 0
>>>>   },
>>>>
>>>> 5. Looks like the block layer or underlying(scsi/virtio-scsi) may
>>>> have some issue which leading to the io request from md layer stayed
>>>> in a partial complete statue. I can't see how this can be related
>>>> with the commit bed9e27baf52 ("Revert "md/raid5: Wait for
>>>> MD_SB_CHANGE_PENDING in raid5d"")
>>>>
>>>>
>>>> Dan,
>>>>
>>>> Are you able to reproduce using some regular scsi disk, would like
>>>> to rule out whether this is related with virtio-scsi?
>>>>
>>>> And I see the kernel version is 6.8.0-rc5 from vmcore, is this the
>>>> official mainline v6.8-rc5 without any other patches?
>>>>
>>>>
>>>> Thanks,
>>>>
>>>> Junxiao.
>>>>
>>>> On 2/23/24 6:13 PM, Song Liu wrote:
>>>>> Hi,
>>>>>
>>>>> On Fri, Feb 23, 2024 at 12:07 AM Linux regression tracking (Thorsten
>>>>> Leemhuis) <regressions@leemhuis.info> wrote:
>>>>>> On 21.02.24 00:06, Dan Moulding wrote:
>>>>>>> Just a friendly reminder that this regression still exists on the
>>>>>>> mainline. It has been reverted in 6.7 stable. But I upgraded a
>>>>>>> development system to 6.8-rc5 today and immediately hit this issue
>>>>>>> again. Then I saw that it hasn't yet been reverted in Linus' tree.
>>>>>> Song Liu, what's the status here? I aware that you fixed with quite a
>>>>>> few regressions recently, but it seems like resolving this one is
>>>>>> stalled. Or were you able to reproduce the issue or make some
>>>>>> progress
>>>>>> and I just missed it?
>>>>> Sorry for the delay with this issue. I have been occupied with some
>>>>> other stuff this week.
>>>>>
>>>>> I haven't got luck to reproduce this issue. I will spend more time
>>>>> looking
>>>>> into it next week.
>>>>>
>>>>>> And if not, what's the way forward here wrt to the release of 6.8?
>>>>>> Revert the culprit and try again later? Or is that not an option
>>>>>> for one
>>>>>> reason or another?
>>>>> If we don't make progress with it in the next week, we will do the
>>>>> revert,
>>>>> same as we did with stable kernels.
>>>>>
>>>>>> Or do we assume that this is not a real issue? That it's caused by
>>>>>> some
>>>>>> oddity (bit-flip in the metadata or something like that?) only to be
>>>>>> found in Dan's setup?
>>>>> I don't think this is because of oddities. Hopefully we can get more
>>>>> information about this soon.
>>>>>
>>>>> Thanks,
>>>>> Song
>>>>>
>>>>>> Ciao, Thorsten (wearing his 'the Linux kernel's regression
>>>>>> tracker' hat)
>>>>>> --
>>>>>> Everything you wanna know about Linux kernel regression tracking:
>>>>>> https://linux-regtracking.leemhuis.info/about/#tldr
>>>>>> If I did something stupid, please tell me, as explained on that page.
>>>>>>
>>>>>> #regzbot poke
>>>>>>
>>>
>>> .
>>>
>>
> .
>


\
 
 \ /
  Last update: 2024-03-13 02:20    [W:0.125 / U:0.176 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site