Re: [PATCH] Bcache: don't return BLK_STS_IOERR during cache detach

[Coly Li <colyli@suse.de>]

On 3/14/22 8:04 PM, Zhang Zhen wrote:
>
>
> On 3/10/22 5:10 PM, Coly Li wrote:
>> On 3/10/22 10:50 AM, Zhang Zhen wrote:
>>> Before this patch, if cache device missing, cached_dev_submit_bio 
>>> return io err
>>> to fs during cache detach, randomly lead to xfs do force shutdown.
>>>
>>> This patch delay the cache io submit in cached_dev_submit_bio
>>> and wait for cache set detach finish.
>>> So if the cache device become missing, bcache detach cache set 
>>> automatically,
>>> and the io will sumbit normally.
>>>
>>> Feb  2 20:59:23  kernel: bcache: bch_count_io_errors() nvme0n1p56: 
>>> IO error on writing btree.
>>> Feb  2 20:59:23  kernel: bcache: bch_count_io_errors() nvme0n1p57: 
>>> IO error on writing btree.
>>> Feb  2 20:59:23  kernel: bcache: bch_count_io_errors() nvme0n1p56: 
>>> IO error on writing btree.
>>> Feb  2 20:59:23  kernel: bcache: bch_btree_insert() error -5
>>> Feb  2 20:59:23  kernel: XFS (bcache43): metadata I/O error in 
>>> "xfs_buf_iodone_callback_error" at daddr 0x80034658 len 32 error 12
>>> Feb  2 20:59:23  kernel: bcache: bch_btree_insert() error -5
>>> Feb  2 20:59:23  kernel: bcache: bch_btree_insert() error -5
>>> Feb  2 20:59:23  kernel: bcache: bch_btree_insert() error -5
>>> Feb  2 20:59:23  kernel: bcache: bch_btree_insert() error -5
>>> Feb  2 20:59:23  kernel: bcache: bch_cache_set_error() bcache: error 
>>> on 004f8aa7-561a-4ba7-bf7b-292e461d3f18:
>>> Feb  2 20:59:23  kernel: journal io error
>>> Feb  2 20:59:23  kernel: bcache: bch_cache_set_error() , disabling 
>>> caching
>>> Feb  2 20:59:23  kernel: bcache: bch_btree_insert() error -5
>>> Feb  2 20:59:23  kernel: bcache: conditional_stop_bcache_device() 
>>> stop_when_cache_set_failed of bcache43 is "auto" and cache is clean, 
>>> keep it alive.
>>> Feb  2 20:59:23  kernel: XFS (bcache43): metadata I/O error in 
>>> "xlog_iodone" at daddr 0x400123e60 len 64 error 12
>>> Feb  2 20:59:23  kernel: XFS (bcache43): xfs_do_force_shutdown(0x2) 
>>> called from line 1298 of file fs/xfs/xfs_log.c. Return address = 
>>> 00000000c1c8077f
>>> Feb  2 20:59:23  kernel: XFS (bcache43): Log I/O Error Detected. 
>>> Shutting down filesystem
>>> Feb  2 20:59:23  kernel: XFS (bcache43): Please unmount the 
>>> filesystem and rectify the problem(s)
>>>
>>> Signed-off-by: Zhen Zhang <zhangzhen.email@gmail.com>
>>> ---
>>>  drivers/md/bcache/bcache.h  | 5 -----
>>>  drivers/md/bcache/request.c | 8 ++++----
>>>  drivers/md/bcache/super.c   | 3 ++-
>>>  3 files changed, 6 insertions(+), 10 deletions(-)
>>>
>>> diff --git a/drivers/md/bcache/bcache.h b/drivers/md/bcache/bcache.h
>>> index 9ed9c955add7..e5227dd08e3a 100644
>>> --- a/drivers/md/bcache/bcache.h
>>> +++ b/drivers/md/bcache/bcache.h
>>> @@ -928,11 +928,6 @@ static inline void closure_bio_submit(struct 
>>> cache_set *c,
>>>                        struct closure *cl)
>>>  {
>>>      closure_get(cl);
>>> -    if (unlikely(test_bit(CACHE_SET_IO_DISABLE, &c->flags))) {
>>> -        bio->bi_status = BLK_STS_IOERR;
>>> -        bio_endio(bio);
>>> -        return;
>>> -    }
>>>      submit_bio_noacct(bio);
>>>  }
>>
>>
>> Comparing to make bcache device living as it looks like, avoiding 
>> data corruption or stale is much more critical. Therefore once there 
>> is critical device failure detected, the following I/O requests must 
>> be stopped (especially write request) to avoid further data 
>> corruption. Without the above checking for CACHE_SET_IO_DISABLE, the 
>> cache has to be attached until there is no I/O. For a busy system it 
>> should be quite long time. Then users may encounter silent data 
>> corruption or inconsistency after a long time since hardware failed.
>>
>>
>> Again, with the above change, you may introduce other issue which 
>> more hard to detect.
>>
>>
>>>  diff --git a/drivers/md/bcache/request.c b/drivers/md/bcache/request.c
>>> index d15aae6c51c1..36f0ee95b51f 100644
>>> --- a/drivers/md/bcache/request.c
>>> +++ b/drivers/md/bcache/request.c
>>> @@ -13,6 +13,7 @@
>>>  #include "request.h"
>>>  #include "writeback.h"
>>>  +#include <linux/delay.h>
>>>  #include <linux/module.h>
>>>  #include <linux/hash.h>
>>>  #include <linux/random.h>
>>> @@ -1172,11 +1173,10 @@ void cached_dev_submit_bio(struct bio *bio)
>>>      unsigned long start_time;
>>>      int rw = bio_data_dir(bio);
>>>  -    if (unlikely((d->c && test_bit(CACHE_SET_IO_DISABLE, 
>>> &d->c->flags)) ||
>>> +    while (unlikely((d->c && test_bit(CACHE_SET_IO_DISABLE, 
>>> &d->c->flags)) ||
>>>               dc->io_disable)) {
>>> -        bio->bi_status = BLK_STS_IOERR;
>>> -        bio_endio(bio);
>>> -        return;
>>> +        /* wait for detach finish and d->c == NULL. */
>>> +        msleep(2);
>>>      }
>>
>> This is unacceptible, neither the infinite loop nor the msleep. You 
>> cannot solve the target issue by an infinite retry, such method will 
>> introduce more issue from other place.
>>
>>
>>>       if (likely(d->c)) {
>>> diff --git a/drivers/md/bcache/super.c b/drivers/md/bcache/super.c
>>> index 140f35dc0c45..8d9a5e937bc8 100644
>>> --- a/drivers/md/bcache/super.c
>>> +++ b/drivers/md/bcache/super.c
>>> @@ -661,7 +661,8 @@ int bch_prio_write(struct cache *ca, bool wait)
>>>          p->csum        = bch_crc64(&p->magic, 
>>> meta_bucket_bytes(&ca->sb) - 8);
>>>           bucket = bch_bucket_alloc(ca, RESERVE_PRIO, wait);
>>> -        BUG_ON(bucket == -1);
>>> +        if (bucket == -1)
>>> +            return -1;
>>
>> This change is wrong. bucket == -1 indicates the bucket allocator 
>> doesn't work properly, if it happens something really critical 
>> happening. This is why BUG_ON is used here.
>>
>> With the above change, you will encounter other strange issue sooner 
>> or later and hard to tell the root cause.
>>
>>
>>> mutex_unlock(&ca->set->bucket_lock);
>>>          prio_io(ca, bucket, REQ_OP_WRITE, 0);
>>
>>
>> Currently I don't have clear idea on how to avoid the IO error return 
>> value during cache set detaching for a failed cache device. But it 
>> cannot be such simple change by the this patch.
> Hi Coly,
>
> Thanks for your review，
> It seems that this is a difficult problem.
>
> Maybe we can try another method.
> If critical device failure detected, we just set IO_DISABLE flag and 
> detach cache device. But don't auto recover.


What do you mean on "auto recover", could you describe it with more 
details ?


>
> This will at least not confuse users，some disk is normal and some disk 
> is error.
> Let user recover it manually.

When you see panic, it is from XFS meta data I/O error, which is 
critical to XFS and trigger its panic. When the I/O error happens for 
normal file system data blocks, the user space receives I/O error return 
values that's why you don't see the panic.

In your case, the cache device is always clean, so cache device failure 
will detach the cache from backing device. It will be better if we can 
re-read data from backing device if no dirty data on cache, but we need 
to handle the potential race window between cache-seen-clean and 
read-from-backing-device, because the cache mode can change on-fly and 
writing may come at any time. This is not simple and should be careful, 
but it can be improved.


Coly Li