On 9/1/25 11:40 AM, Yu Kuai wrote:
> Hi,
>
> 在 2025/08/30 12:28, Yu Kuai 写道:
>>>> @@ -745,12 +745,16 @@ void submit_bio_noacct_nocheck(struct bio *bio)
>>>> * to collect a list of requests submited by a ->submit_bio method while
>>>> * it is active, and then process them after it returned.
>>>> */
>>>> - if (current->bio_list)
>>>> - bio_list_add(¤t->bio_list[0], bio);
>>>> - else if (!bdev_test_flag(bio->bi_bdev, BD_HAS_SUBMIT_BIO))
>>>> + if (current->bio_list) {
>>>> + if (split)
>>>> + bio_list_add_head(¤t->bio_list[0], bio);
>>>> + else
>>>> + bio_list_add(¤t->bio_list[0], bio);
>>> This really needs a comment clarifying why we do an add at tail instead of
>>> keeping the original order with a add at head. I am also scared that this may
>>> break sequential write ordering for zoned devices.
>>
>> I think add at head is exactly what we do here to keep the orginal order for
>> the case bio split. Other than split, if caller do generate multiple sequential
>> bios, we should keep the order by add at tail.
>>
>> Not sure about zoned devices for now, I'll have a look in details.
>
> For zoned devices, can we somehow trigger this recursive split? I
> suspect bio disordered will apear in this case but I don't know for
> now and I can't find a way to reporduce it.
dm-linear can be stacked on e.g. dm-crypt, or the reverse. So recursive
splitting may be possible. Though since for DM everything is zone aligned, it
may be hard to find a reproducer. Though dm-crypt will always split BIOs to
BIO_MAX_VECS << PAGE_SECTORS_SHIFT sectors, so it may be possible with very
large BIOs. Would need to try, but really overloaded with other things right now.
>
> Perhaps I can bypass zoned devices for now, and if we really met the
> recursive split case and there is a problem, we can fix it later:
>
> if (split && !bdev_is_zoned(bio->bi_bdev))
> bio_list_add_head()
>
> Thanks,
> Kuai
>
--
Damien Le Moal
Western Digital Research
