Re: [PATCH] md/raid1,raid10: fix IO handle for REQ_NOWAIT

[Paul Menzel <pmenzel@molgen.mpg.de>]

Dear Zheng,


Am 14.06.25 um 08:50 schrieb Zheng Qixing:

> Please disregard the previous reply email, as it contained garbled text.

Thank you for noticing this, and resending.

> 在 2025/6/13 16:02, Paul Menzel 写道:

>> Am 12.06.25 um 15:21 schrieb Zheng Qixing:
>>> From: Zheng Qixing <zhengqixing@huawei.com>
>>>
>>> IO with REQ_NOWAIT should not set R1BIO_Uptodate when it fails,
>>> and bad blocks should also be cleared when REQ_NOWAIT IO succeeds.
>>
>> It’d be great if you could add an explanation for the *should*. Why 
>> should it not be done?
>>
>> Do you have a reproducer for this?
> 
> If we set R1BIO_Uptodate when IO with REQ_NOWAIT fails, the request will
> return a success.

Understood. So no command to check for this automatically on a test system.

For the explanation, I guess my problem is, that I was not familiar with 
REQ_NOWAIT, which means that it fails for blocked IO. (If I am correct.)

> But actually it should return BLK_STS_IOERR or BLK_STS_AGAIN, right?

Sorry, I do not know. Hopefully the maintainers can answer this.

>>> Fixes: 9f346f7d4ea7 ("md/raid1,raid10: don't handle IO error for REQ_RAHEAD and REQ_NOWAIT")
>>> Signed-off-by: Zheng Qixing <zhengqixing@huawei.com>
>>> ---
>>>   drivers/md/raid1.c  | 11 ++++++-----
>>>   drivers/md/raid10.c |  9 +++++----
>>>   2 files changed, 11 insertions(+), 9 deletions(-)
>>>
>>> diff --git a/drivers/md/raid1.c b/drivers/md/raid1.c
>>> index 19c5a0ce5a40..a1cddd24b178 100644
>>> --- a/drivers/md/raid1.c
>>> +++ b/drivers/md/raid1.c
>>> @@ -455,13 +455,13 @@ static void raid1_end_write_request(struct bio *bio)
>>>       struct md_rdev *rdev = conf->mirrors[mirror].rdev;
>>>       sector_t lo = r1_bio->sector;
>>>       sector_t hi = r1_bio->sector + r1_bio->sectors;
>>> -    bool ignore_error = !raid1_should_handle_error(bio) ||
>>> -        (bio->bi_status && bio_op(bio) == REQ_OP_DISCARD);
>>> +    bool discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
>>
>> Excuse my ignorance. What is the difference between ignore and discard?
> 
> REQ_OP_DISCARD is a operation type while REQ_NOWAIT is just a request flag.
> 
> These two can be combined together. IO with REQ_NOWAIT can fail early, even
> though the storage medium is fine. So, we better handle this type of
> error specially.
> 
> I hope this clarifies your doubts.

Sorry about being not clear enough. My question was more about changing 
the naming of the variable.

>>>       /*
>>>        * 'one mirror IO has finished' event handler:
>>>        */
>>> -    if (bio->bi_status && !ignore_error) {
>>> +    if (bio->bi_status && !discard_error &&
>>> +        raid1_should_handle_error(bio)) {
>>>           set_bit(WriteErrorSeen,    &rdev->flags);
>>>           if (!test_and_set_bit(WantReplacement, &rdev->flags))
>>>               set_bit(MD_RECOVERY_NEEDED, &
>>> @@ -507,12 +507,13 @@ static void raid1_end_write_request(struct bio *bio)
>>>            * check this here.
>>>            */
>>>           if (test_bit(In_sync, &rdev->flags) &&
>>> -            !test_bit(Faulty, &rdev->flags))
>>> +            !test_bit(Faulty, &rdev->flags) &&
>>> +            (!bio->bi_status || discard_error))
>>>               set_bit(R1BIO_Uptodate, &r1_bio->state);
>>>             /* Maybe we can clear some bad blocks. */
>>>           if (rdev_has_badblock(rdev, r1_bio->sector, r1_bio->sectors) &&
>>> -            !ignore_error) {
>>> +            !bio->bi_status) {
>>>               r1_bio->bios[mirror] = IO_MADE_GOOD;
>>>               set_bit(R1BIO_MadeGood, &r1_bio->state);
>>>           }
>>> diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c
>>> index b74780af4c22..1848947b0a6d 100644
>>> --- a/drivers/md/raid10.c
>>> +++ b/drivers/md/raid10.c
>>> @@ -458,8 +458,8 @@ static void raid10_end_write_request(struct bio *bio)
>>>       int slot, repl;
>>>       struct md_rdev *rdev = NULL;
>>>       struct bio *to_put = NULL;
>>> -    bool ignore_error = !raid1_should_handle_error(bio) ||
>>> -        (bio->bi_status && bio_op(bio) == REQ_OP_DISCARD);
>>> +    bool discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
>>> +    bool ignore_error = !raid1_should_handle_error(bio) || discard_error;
>>>         dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
>>>   @@ -522,13 +522,14 @@ static void raid10_end_write_request(struct bio *bio)
>>>            * check this here.
>>>            */
>>>           if (test_bit(In_sync, &rdev->flags) &&
>>> -            !test_bit(Faulty, &rdev->flags))
>>> +            !test_bit(Faulty, &rdev->flags) &&
>>> +            (!bio->bi_status || discard_error))
>>>               set_bit(R10BIO_Uptodate, &r10_bio->state);
>>>             /* Maybe we can clear some bad blocks. */
>>>           if (rdev_has_badblock(rdev, r10_bio->devs[slot].addr,
>>>                         r10_bio->sectors) &&
>>> -            !ignore_error) {
>>> +            !bio->bi_status) {
>>>               bio_put(bio);
>>>               if (repl)
>>>                   r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;


Kind regards,

Paul


PS: As it’s two hunks only connected through REQ_NOWAIT, maybe make it 
two commits: one for raid1 and one for raid10? Feel free to ignore.