[f2fs-dev] [PATCH V3 6/6] xfs: ignore discard return value

[Yongpeng Yang]

On 11/26/25 16:07, Chaitanya Kulkarni via Linux-f2fs-devel wrote:
> On 11/25/25 18:37, Yongpeng Yang wrote:
>>> diff --git a/fs/xfs/xfs_discard.c b/fs/xfs/xfs_discard.c
>>> index 6917de832191..b6ffe4807a11 100644
>>> --- a/fs/xfs/xfs_discard.c
>>> +++ b/fs/xfs/xfs_discard.c
>>> @@ -108,7 +108,7 @@ xfs_discard_endio(
>>>     * list. We plug and chain the bios so that we only need a single
>>> completion
>>>     * call to clear all the busy extents once the discards are complete.
>>>     */
>>> -int
>>> +void
>>>    xfs_discard_extents(
>>>        struct xfs_mount    *mp,
>>>        struct xfs_busy_extents    *extents)
>>> @@ -116,7 +116,6 @@ xfs_discard_extents(
>>>        struct xfs_extent_busy    *busyp;
>>>        struct bio        *bio = NULL;
>>>        struct blk_plug        plug;
>>> -    int            error = 0;
>>>          blk_start_plug(&plug);
>>>        list_for_each_entry(busyp, &extents->extent_list, list) {
>>> @@ -126,18 +125,10 @@ xfs_discard_extents(
>>>              trace_xfs_discard_extent(xg, busyp->bno, busyp->length);
>>>    -        error = __blkdev_issue_discard(btp->bt_bdev,
>>> +        __blkdev_issue_discard(btp->bt_bdev,
>>>                    xfs_gbno_to_daddr(xg, busyp->bno),
>>>                    XFS_FSB_TO_BB(mp, busyp->length),
>>>                    GFP_KERNEL, &bio);
>>
>> If blk_alloc_discard_bio() fails to allocate a bio inside
>> __blkdev_issue_discard(), this may lead to an invalid loop in
>> list_for_each_entry{}. Instead of using __blkdev_issue_discard(), how
>> about allocate and submit the discard bios explicitly in
>> list_for_each_entry{}?
>>
>> Yongpeng,
> 
> 
> Calling __blkdev_issue_discard() keeps managing all the bio with the
> appropriate GFP mask, so the semantics stay the same. You are just
> moving memory allocation to the caller and potentially looking at
> implementing retry on bio allocation failure.
> 
> The retry for discard bio memory allocation is not desired I think,
> since it's only a hint to the controller.

Agreed. I'm not trying to retry bio allocation inside the
list_for_each_entry{} loop. Instead, since blk_alloc_discard_bio()
returning NULL cannot reliably indicate whether the failure is due to
bio allocation failure, it could also be caused by 'bio_sects == 0', I'd
like to allocate the bio explicitly.

> 
> This patch is simply cleaning up dead error-handling branches at the
> callers no behavioral changes intended.
> 
> What maybe useful is to stop iterating once we fail to allocate the
> bio [1].
> 
> -ck
> 
> [1] Potential addition on the top of this to exit early in discard loop
>       on bio allocation failure.
> 
> diff --git a/fs/xfs/xfs_discard.c b/fs/xfs/xfs_discard.c
> index b6ffe4807a11..1519f708bb79 100644
> --- a/fs/xfs/xfs_discard.c
> +++ b/fs/xfs/xfs_discard.c
> @@ -129,6 +129,13 @@ xfs_discard_extents(
>                                   xfs_gbno_to_daddr(xg, busyp->bno),
>                                   XFS_FSB_TO_BB(mp, busyp->length),
>                                   GFP_KERNEL, &bio);
> +               /*
> +                * We failed to allocate bio instead of continuing the loop
> +                * so it will lead to inconsistent discards to the disk
> +                * exit early and jump into xfs_discard_busy_clear().
> +                */
> +               if (!bio)
> +                       break;

I noticed that as long as XFS_FSB_TO_BB(mp, busyp->length) is greater
than 0 and there is no bio allocation failure, __blkdev_issue_discard()
will never return NULL. I'm not familiar with this part of the xfs, so
I'm not sure whether there are cases where 'XFS_FSB_TO_BB(mp,
busyp->length)' could be 0. If such cases do not exist, then
checking whether the bio is NULL should be sufficient.

Yongpeng,

>           }
>    
>           if (bio) {
> > If we keep looping after the first bio == NULL, the rest of the range is
> guaranteed to be inconsistent anyways, because every subsequent iteration
> will fall into one of three cases:
> 
> - The allocator keeps returning NULL, so none of the remaining LBAs receive
>     discard.
> - Rest of the allocator succeeds, but we’ve already skipped a chunk, leaving
>     a hole in the discard range.
> - We get intermittent successes, which produces alternating chunks of
>     discarded and undiscarded blocks.
> 
> In each of those scenarios, the disk ends up with a partially discarded
> range, so the correct fix is to break out of the loop immediately and
> proceed to xfs_discard_busy_clear() once the very first allocation fails.