generic/456 regression on 5.7-rc1, 1k test case

generic/456 regression on 5.7-rc1, 1k test case

[Eric Whitney]

I'm seeing consistent failures for generic/456 while running kvm-xfstests' 1k
test case on 5.7-rc1.  This is with an x86-64 test appliance root file system
image dated 23 March 2020.

The test fails when e2fsck reports "inconsistent fs: inode 12, i_size is
147456, should be 163840".

Bisecting 5.7-rc1 identified the following patch as the cause:
ext4: don't set dioread_nolock by default for blocksize < pagesize
(626b035b816b).  Reverting the patch in 5.7-rc1 reliably eliminates the test
failure.

Eric

Re: generic/456 regression on 5.7-rc1, 1k test case

[Ritesh Harjani]

Hello Eric,

On 4/14/20 1:42 AM, Eric Whitney wrote:
> I'm seeing consistent failures for generic/456 while running kvm-xfstests' 1k
> test case on 5.7-rc1.  This is with an x86-64 test appliance root file system
> image dated 23 March 2020.
> 
> The test fails when e2fsck reports "inconsistent fs: inode 12, i_size is
> 147456, should be 163840".
> 
> Bisecting 5.7-rc1 identified the following patch as the cause:
> ext4: don't set dioread_nolock by default for blocksize < pagesize
> (626b035b816b).  Reverting the patch in 5.7-rc1 reliably eliminates the test
> failure.
> 

Since you could reliably reproduce it. Could you please try with this
patch and see if this fixes it for you?

https://patchwork.ozlabs.org/project/linux-ext4/patch/20200331105016.8674-1-jack@suse.cz/

-ritesh

Re: generic/456 regression on 5.7-rc1, 1k test case

[Ritesh Harjani]

Hello Eric,

On 4/14/20 8:22 AM, Ritesh Harjani wrote:
> Hello Eric,
> 
> On 4/14/20 1:42 AM, Eric Whitney wrote:
>> I'm seeing consistent failures for generic/456 while running 
>> kvm-xfstests' 1k
>> test case on 5.7-rc1.  This is with an x86-64 test appliance root file 
>> system
>> image dated 23 March 2020.
>>
>> The test fails when e2fsck reports "inconsistent fs: inode 12, i_size is
>> 147456, should be 163840".
>>
>> Bisecting 5.7-rc1 identified the following patch as the cause:
>> ext4: don't set dioread_nolock by default for blocksize < pagesize
>> (626b035b816b).  Reverting the patch in 5.7-rc1 reliably eliminates 
>> the test
>> failure.
>>
> 
> Since you could reliably reproduce it. Could you please try with this
> patch and see if this fixes it for you?
> 
> https://patchwork.ozlabs.org/project/linux-ext4/patch/20200331105016.8674-1-jack@suse.cz/ 


Ok, so after updating the xfstests to latest, I could reliably reproduce
generic/456 failing on x86 with 1K blocksize on my setup too.
Although, with my limited testing, I couldn't see this issue on Power 
(where blocksize == 4K and PAGESIZE=64K).

But either ways, after applying above patch the tests always passes for
me (tested on x86). So this should indeed fix your reported problem.
Saw an email too that Ted has now picked up this patch.


-ritesh

Re: generic/456 regression on 5.7-rc1, 1k test case

[Eric Whitney]

* Ritesh Harjani <riteshh@linux.ibm.com>:
> Hello Eric,
> 
> On 4/14/20 8:22 AM, Ritesh Harjani wrote:
> > Hello Eric,
> > 
> > On 4/14/20 1:42 AM, Eric Whitney wrote:
> > > I'm seeing consistent failures for generic/456 while running
> > > kvm-xfstests' 1k
> > > test case on 5.7-rc1.  This is with an x86-64 test appliance root
> > > file system
> > > image dated 23 March 2020.
> > > 
> > > The test fails when e2fsck reports "inconsistent fs: inode 12, i_size is
> > > 147456, should be 163840".
> > > 
> > > Bisecting 5.7-rc1 identified the following patch as the cause:
> > > ext4: don't set dioread_nolock by default for blocksize < pagesize
> > > (626b035b816b).  Reverting the patch in 5.7-rc1 reliably eliminates
> > > the test
> > > failure.
> > > 
> > 
> > Since you could reliably reproduce it. Could you please try with this
> > patch and see if this fixes it for you?
> > 
> > https://patchwork.ozlabs.org/project/linux-ext4/patch/20200331105016.8674-1-jack@suse.cz/
> 
> 
> Ok, so after updating the xfstests to latest, I could reliably reproduce
> generic/456 failing on x86 with 1K blocksize on my setup too.
> Although, with my limited testing, I couldn't see this issue on Power (where
> blocksize == 4K and PAGESIZE=64K).
> 
> But either ways, after applying above patch the tests always passes for
> me (tested on x86). So this should indeed fix your reported problem.
> Saw an email too that Ted has now picked up this patch.
> 
> 
> -ritesh
>

Hi, Ritesh:

Thanks for pointing out that patch - I'd not noticed it.  Out of general
thoroughness, I applied it to 5.7-rc1 and ran the entire 1k test case
without a generic/456 failure or any other regressions.  So, that should
resolve this issue.  (5.7-rc1 otherwise looks good to me generally after
regression against 5.6 on x86_64).

Thanks very much for your help!

Eric