Re: [PATCH RFC v3 14/15] block: fix disordered IO in the case recursive split

[Yu Kuai <yukuai1@xxxxxxxxxxxxxxx>]

Hi,

在 2025/09/03 9:12, Bart Van Assche 写道:
> On 9/2/25 6:00 PM, Yu Kuai wrote:
> > Hi,
> >
> > 在 2025/09/03 1:20, Bart Van Assche 写道:
> > > On 8/31/25 8:32 PM, Yu Kuai wrote:
> > > > -void submit_bio_noacct_nocheck(struct bio *bio)
> > > > +void submit_bio_noacct_nocheck(struct bio *bio, bool split)
> > > >   {
> > > >       blk_cgroup_bio_start(bio);
> > > >       blkcg_bio_issue_init(bio);
> > > > @@ -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(&current->bio_list[0], bio);
> > > > -    else if (!bdev_test_flag(bio->bi_bdev, BD_HAS_SUBMIT_BIO))
> > > > +    if (current->bio_list) {
> > > > +        if (split && !bdev_is_zoned(bio->bi_bdev))
> > > > +            bio_list_add_head(&current->bio_list[0], bio);
> > > > +        else
> > > > +            bio_list_add(&current->bio_list[0], bio);
> > >
> > > The above change will cause write errors for zoned block devices. As I
> > > have shown before, also for zoned block devices, if a bio is split
> > > insertion must happen at the head of the list. See e.g.
> > > "Re: [PATCH 1/2] block: Make __submit_bio_noacct() preserve the bio 
> > > submission order"
> > > (https://lore.kernel.org/linux-block/a0c89df8-4b33-409c-ba43- 
> > > f9543fb1b091@xxxxxxx/)
> >
> > Do you mean we should remove the bdev_is_zoned() checking? I added this
> > checking because I'm not quite sure about details in zone device, and
> > this checking is aimed at prevent functional changes in zone device.
>
> Yes, the bdev_is_zoned() check should be removed. I spent a significant
> amount of time on root-causing and proposing fixes for write errors
> caused by recursive bio splitting for zoned devices. What I learned by
> analyzing these write errors is the basis for this feedback.
>
> > So I don't think this change will cause write errors, the write errors
> > should already exist before this set, right?
>
> Agreed. Although I haven't checked this yet, if the bdev_is_zoned()
> check is removed from this patch, this patch should fix the write errors
> triggered by stacking a dm driver on top of a zoned block device if
> inline encryption is enabled.

Ok, I'll remove the checking and mention this problem in the next formal
version.

Thanks,
Kuai

> Thanks,
>
> Bart.
> .