Re: [bug report] fstests generic/774 hang

[John Garry <john.g.garry@oracle.com>]

On 09/11/2025 11:58, Ojaswin Mujoo wrote:
>> aw_bsize for  me is 1M, and threads is 32
>>
>> aw_bsize is large as XFS supports software-based atomics, which is generally
>> going to be huge compared to anything which HW can support.
>>
>> When I tried to run this test, it was not completing in a sane amount of
>> time - it was taking many minutes before I gave up.
> Hi John, Shinichiro, Darrick.
> 
> Thanks for looking into this. Sorry, I'm on vacation so a bit slow in
> responding.
> 
> Anyways, the logic behind the filesize calculation is that we want each
> thread to do 100 atomic writes in their own isolated ranges in the file.
> But seems like it is being especially slow when we have high CPUs.

It's not just the number of CPUs which is the problem. The test does the 
awu max size writes - for XFS, this size can be many MBs, and not like 
typically < 100 KBs for any FS which relies only on HW-based atomic 
writes, i.e. ext4. Please also consider limiting the awu max size.

> 
> I think in that sense, it'll be better to limit the threads itself
> rather than filesize. Since its a stress test we dont want it to be too
> less. Maybe:
> 
> diff --git a/tests/generic/774 b/tests/generic/774
> index 7a4d7016..c68fb4b7 100755
> --- a/tests/generic/774
> +++ b/tests/generic/774
> @@ -28,7 +28,7 @@ awu_max_write=$(_get_atomic_write_unit_max "$SCRATCH_MNT/f1")
>   aw_bsize=$(_max "$awu_min_write" "$((awu_max_write/4))")
>   fsbsize=$(_get_block_size $SCRATCH_MNT)
> 
> -threads=$(_min "$(($(nproc) * 2 * LOAD_FACTOR))" "100")
> +threads=$(_min "$(($(nproc) * 2 * LOAD_FACTOR))" "16")
>   filesize=$((aw_bsize * threads * 100))
>   depth=$threads
>   aw_io_size=$((filesize / threads))
> 
> Can you check if this helps?