Re: [5.10.y] Regression in EXT3/4 with v5.10.227, LTP preadv03 is failing

Re: [5.10.y] Regression in EXT3/4 with v5.10.227, LTP preadv03 is failing

[Greg KH]

On Mon, Oct 21, 2024 at 05:43:54PM +0200, Mahmoud Adam wrote:
> 
> Hello,
> 
>   In the latest 5.10.227 we see LTP preadv03 failing on EXT3/4, when
> bisected it was found that the commit dde4c1e1663b6 ("ext4: properly
> sync file size update after O_SYNC direct IO") is causing it
> 
> This seems similar to what occurred before[1] and if I understand
> correctly it was because of missing dependency backport 936e114a245b6
> ("iomap: update ki_pos a little later in iomap_dio_complete") which was
> then backported to 5.15 & 6.1 but not to 5.10.
> 
> *This is not failing on 5.15 nor 6.1, and it does fail consistently in x86-64 & ARM
> 
> [1]: https://lore.kernel.org/all/20231205122122.dfhhoaswsfscuhc3@quack3/
> 


What do you suggest be done?

thanks,

greg k-h

Re: [5.10.y] Regression in EXT3/4 with v5.10.227, LTP preadv03 is failing

[Mahmoud Adam]

Greg KH <gregkh@linuxfoundation.org> writes:

> On Mon, Oct 21, 2024 at 05:43:54PM +0200, Mahmoud Adam wrote:
>> 
>> Hello,
>> 
>>   In the latest 5.10.227 we see LTP preadv03 failing on EXT3/4, when
>> bisected it was found that the commit dde4c1e1663b6 ("ext4: properly
>> sync file size update after O_SYNC direct IO") is causing it
>> 
>> This seems similar to what occurred before[1] and if I understand
>> correctly it was because of missing dependency backport 936e114a245b6
>> ("iomap: update ki_pos a little later in iomap_dio_complete") which was
>> then backported to 5.15 & 6.1 but not to 5.10.
>> 
>> *This is not failing on 5.15 nor 6.1, and it does fail consistently in x86-64 & ARM
>> 
>> [1]: https://lore.kernel.org/all/20231205122122.dfhhoaswsfscuhc3@quack3/
>> 
>
>
> What do you suggest be done?
>

I think as an urgent fix I would suggest reverting the mentioned commit
and commits that depend on it from 5.10.y..

4911610c7a1fe ("ext4: fix warning in ext4_dio_write_end_io()")
f8a7c342326f6 ("ext4: dax: fix overflowing extents beyond inode size when partially writing")
dde4c1e1663b6 ("ext4: properly sync file size update after O_SYNC direct
IO")

maybe someone from the ext4 side confirm that as well?

thanks,
MNAdam

Re: [5.10.y] Regression in EXT3/4 with v5.10.227, LTP preadv03 is failing

[Greg KH]

On Mon, Oct 21, 2024 at 07:15:58PM +0200, Mahmoud Adam wrote:
> Greg KH <gregkh@linuxfoundation.org> writes:
> 
> > On Mon, Oct 21, 2024 at 05:43:54PM +0200, Mahmoud Adam wrote:
> >> 
> >> Hello,
> >> 
> >>   In the latest 5.10.227 we see LTP preadv03 failing on EXT3/4, when
> >> bisected it was found that the commit dde4c1e1663b6 ("ext4: properly
> >> sync file size update after O_SYNC direct IO") is causing it
> >> 
> >> This seems similar to what occurred before[1] and if I understand
> >> correctly it was because of missing dependency backport 936e114a245b6
> >> ("iomap: update ki_pos a little later in iomap_dio_complete") which was
> >> then backported to 5.15 & 6.1 but not to 5.10.
> >> 
> >> *This is not failing on 5.15 nor 6.1, and it does fail consistently in x86-64 & ARM
> >> 
> >> [1]: https://lore.kernel.org/all/20231205122122.dfhhoaswsfscuhc3@quack3/
> >> 
> >
> >
> > What do you suggest be done?
> >
> 
> I think as an urgent fix I would suggest reverting the mentioned commit
> and commits that depend on it from 5.10.y..
> 
> 4911610c7a1fe ("ext4: fix warning in ext4_dio_write_end_io()")
> f8a7c342326f6 ("ext4: dax: fix overflowing extents beyond inode size when partially writing")
> dde4c1e1663b6 ("ext4: properly sync file size update after O_SYNC direct
> IO")

Great, can you submit a series of these reverted that you have tested to
fix the issue?

thanks,

greg k-h

Re: [5.10.y] Regression in EXT3/4 with v5.10.227, LTP preadv03 is failing

[Jan Kara]

On Mon 21-10-24 19:15:58, Mahmoud Adam wrote:
> Greg KH <gregkh@linuxfoundation.org> writes:
> 
> > On Mon, Oct 21, 2024 at 05:43:54PM +0200, Mahmoud Adam wrote:
> >> 
> >> Hello,
> >> 
> >>   In the latest 5.10.227 we see LTP preadv03 failing on EXT3/4, when
> >> bisected it was found that the commit dde4c1e1663b6 ("ext4: properly
> >> sync file size update after O_SYNC direct IO") is causing it
> >> 
> >> This seems similar to what occurred before[1] and if I understand
> >> correctly it was because of missing dependency backport 936e114a245b6
> >> ("iomap: update ki_pos a little later in iomap_dio_complete") which was
> >> then backported to 5.15 & 6.1 but not to 5.10.
> >> 
> >> *This is not failing on 5.15 nor 6.1, and it does fail consistently in x86-64 & ARM
> >> 
> >> [1]: https://lore.kernel.org/all/20231205122122.dfhhoaswsfscuhc3@quack3/
> >> 
> >
> >
> > What do you suggest be done?
> >
> 
> I think as an urgent fix I would suggest reverting the mentioned commit
> and commits that depend on it from 5.10.y..
> 
> 4911610c7a1fe ("ext4: fix warning in ext4_dio_write_end_io()")
> f8a7c342326f6 ("ext4: dax: fix overflowing extents beyond inode size when partially writing")
> dde4c1e1663b6 ("ext4: properly sync file size update after O_SYNC direct
> IO")

Looks sensible to me. Another possibility would be to backport commit
936e114a245b6e3 ("iomap: update ki_pos a little later in
iomap_dio_complete") to 5.10-stable. We've done this for other stable
branches which had this issue and so far I didn't hear about any fallout
from that.

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR

Re: [5.10.y] Regression in EXT3/4 with v5.10.227, LTP preadv03 is failing

[Mahmoud Adam]

Jan Kara <jack@suse.cz> writes:

> On Mon 21-10-24 19:15:58, Mahmoud Adam wrote:
>> Greg KH <gregkh@linuxfoundation.org> writes:
>> 
>> > On Mon, Oct 21, 2024 at 05:43:54PM +0200, Mahmoud Adam wrote:
>> >> 
>> >> Hello,
>> >> 
>> >>   In the latest 5.10.227 we see LTP preadv03 failing on EXT3/4, when
>> >> bisected it was found that the commit dde4c1e1663b6 ("ext4: properly
>> >> sync file size update after O_SYNC direct IO") is causing it
>> >> 
>> >> This seems similar to what occurred before[1] and if I understand
>> >> correctly it was because of missing dependency backport 936e114a245b6
>> >> ("iomap: update ki_pos a little later in iomap_dio_complete") which was
>> >> then backported to 5.15 & 6.1 but not to 5.10.
>> >> 
>> >> *This is not failing on 5.15 nor 6.1, and it does fail consistently in x86-64 & ARM
>> >> 
>> >> [1]: https://lore.kernel.org/all/20231205122122.dfhhoaswsfscuhc3@quack3/
>> >> 
>> >
>> >
>> > What do you suggest be done?
>> >
>> 
>> I think as an urgent fix I would suggest reverting the mentioned commit
>> and commits that depend on it from 5.10.y..
>> 
>> 4911610c7a1fe ("ext4: fix warning in ext4_dio_write_end_io()")
>> f8a7c342326f6 ("ext4: dax: fix overflowing extents beyond inode size when partially writing")
>> dde4c1e1663b6 ("ext4: properly sync file size update after O_SYNC direct
>> IO")
>
> Looks sensible to me. Another possibility would be to backport commit
> 936e114a245b6e3 ("iomap: update ki_pos a little later in
> iomap_dio_complete") to 5.10-stable. We've done this for other stable
> branches which had this issue and so far I didn't hear about any fallout
> from that.

Sure, I can backport that today, if any blockers will send the reverts
instead.

- MNAdam