[0/3] selftests: cachestat: fix build and run on older kernels

Message

Andre Przywara Aug. 15, 2023, 3:56 p.m. UTC

I ran all kernel selftests on some test machine, and stumbled upon
cachestat failing (among others).
Those patches fix the cachestat test compilation and run on older
kernels.

Also I found that the but-last test (on a normal file) fails when run on
a tmpfs mounted directory, as it happens on an initramfs-only system, or
when the current directory happens to be /dev/shm or /tmp:
# Create/open tmpfilecachestat
# Cachestat call returned 0
# Using cachestat: Cached: 4, Dirty: 4, Writeback: 0, Evicted: 0, Recently Evicted: 0
# Cachestat call (after fsync) returned 0
# Using cachestat: Cached: 4, Dirty: 4, Writeback: 0, Evicted: 0, Recently Evicted: 0
# Number of dirty should be zero after fsync.
not ok 6 cachestat fails with normal file

That same test binary succeeds on the same machine right afterwards if
the current directory is changed to an ext4 filesystem.

I don't really know if this is expected, and whether we should try to
figure out if the test file lives on a tmpfs filesystem, or whether the
test itself is not strict enough, and requires more "flushing"
(drop_caches?) to cover tmpfs directories as well.

Any ideas how to fix this would be appreciated.

Cheers,
Andre

Andre Przywara (3):
  selftests: cachestat: properly link in librt
  selftests: cachestat: use proper syscall number macro
  selftests: cachestat: test for cachestat availability

 tools/testing/selftests/cachestat/Makefile    |  2 +-
 .../selftests/cachestat/test_cachestat.c      | 29 +++++++++++++++----
 2 files changed, 25 insertions(+), 6 deletions(-)

Comments

Nhat Pham Aug. 15, 2023, 6:07 p.m. UTC | #1

On Tue, Aug 15, 2023 at 8:56 AM Andre Przywara <[e-mail]> wrote:
>
> I ran all kernel selftests on some test machine, and stumbled upon
> cachestat failing (among others).
> Those patches fix the cachestat test compilation and run on older
> kernels.
>
> Also I found that the but-last test (on a normal file) fails when run on
> a tmpfs mounted directory, as it happens on an initramfs-only system, or
> when the current directory happens to be /dev/shm or /tmp:
> # Create/open tmpfilecachestat
> # Cachestat call returned 0
> # Using cachestat: Cached: 4, Dirty: 4, Writeback: 0, Evicted: 0, Recently Evicted: 0
> # Cachestat call (after fsync) returned 0
> # Using cachestat: Cached: 4, Dirty: 4, Writeback: 0, Evicted: 0, Recently Evicted: 0
> # Number of dirty should be zero after fsync.
> not ok 6 cachestat fails with normal file
>
> That same test binary succeeds on the same machine right afterwards if
> the current directory is changed to an ext4 filesystem.

Ah, if I recall correctly, these kinds of fs have no-op fsync, correct?
Something along the line of:
https://github.com/torvalds/linux/blob/91aa6c4/mm/shmem.c#L4111
The fsync logic would fail indeed. Thanks for pointing that out!

>
> I don't really know if this is expected, and whether we should try to
> figure out if the test file lives on a tmpfs filesystem, or whether the

This would be nice. I think there's a userspace method to check
this, right? There's a TMPFS_MAGIC here - not sure if relevant:

https://man7.org/linux/man-pages/man2/statfs.2.html

> test itself is not strict enough, and requires more "flushing"
> (drop_caches?) to cover tmpfs directories as well.
>
> Any ideas how to fix this would be appreciated.
>
> Cheers,
> Andre
>
> Andre Przywara (3):
>   selftests: cachestat: properly link in librt
>   selftests: cachestat: use proper syscall number macro
>   selftests: cachestat: test for cachestat availability
>
>  tools/testing/selftests/cachestat/Makefile    |  2 +-
>  .../selftests/cachestat/test_cachestat.c      | 29 +++++++++++++++----
>  2 files changed, 25 insertions(+), 6 deletions(-)
>
> --
> 2.25.1
>