Skip to contentSkip to conversation

generic/732: fix mount option munging

PATCH In review Jeff Layton 2 participants 5 messages Nov 1, 2023 – Nov 5, 2023

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 generic/732: fix mount option munging

With NFS in particular, we are usually testing with some mount options. Ensure that we preserve those and just add "nosharecache" onto the end of the string.

Cc Yongcheng Yang <[email protected]>
---
 tests/generic/732 | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tests/generic/732 +1 -1
index 785aac58f361..ae49152e42dc 100755
@@ -40,7 +40,7 @@ mkdir -p $testdir1 $testdir2
 # Don't share the data and attribute caches among mount points for NFS.
 # This caching behavior is necessary to reproduce this issue as we're
 # checking the alignment of each mount point's own unique cache.
-[ "$FSTYP" = "nfs" ] && MOUNT_OPTIONS="-o nosharecache"
+[ "$FSTYP" = "nfs" ] && MOUNT_OPTIONS="$MOUNT_OPTIONS -o nosharecache"
 
 SCRATCH_MNT=$testdir1 _scratch_mount
 SCRATCH_MNT=$testdir2 _scratch_mount

---
base-commit: 59299b65ac8f15935ab45e7920cbfda8a6beffd1
change-id: 20231101-master-328ff30ce66f

Best regards,

Signature · 1 line
Jeff Layton <[email protected]>
Reviewed-by on v1 1/1
#1 lore raw
Quoted 15 lines · Jeff Layton

With NFS in particular, we are usually testing with some mount options. Ensure that we preserve those and just add "nosharecache" onto the end of the string.

Cc Yongcheng Yang <[email protected]>
---
 tests/generic/732 | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tests/generic/732 +0 -0
index 785aac58f361..ae49152e42dc 100755

tests/generic/732 mkdir -p $testdir1 $testdir2 · line 43

Quoted 5 lines
tests/generic/732 +1 -1
@@ -40,7 +40,7 @@ mkdir -p $testdir1 $testdir2
 # Don't share the data and attribute caches among mount points for NFS.
 # This caching behavior is necessary to reproduce this issue as we're
 # checking the alignment of each mount point's own unique cache.
-[ "$FSTYP" = "nfs" ] && MOUNT_OPTIONS="-o nosharecache"
+[ "$FSTYP" = "nfs" ] && MOUNT_OPTIONS="$MOUNT_OPTIONS -o nosharecache"

Good to me, and looks like the later option replaces the former one, if there're same options (e.g. -o sharecache -o nosharecache).

Quoted 13 lines
tests/generic/732 +0 -1
 SCRATCH_MNT=$testdir1 _scratch_mount
 SCRATCH_MNT=$testdir2 _scratch_mount
 
---

base-commit: 59299b65ac8f15935ab45e7920cbfda8a6beffd1
change-id: 20231101-master-328ff30ce66f

Best regards,

Signature · 1 line
Jeff Layton <[email protected]>
#2 lore raw
Quoted 25 lines · Zorro Lang
Quoted 15 lines · Jeff Layton

With NFS in particular, we are usually testing with some mount options. Ensure that we preserve those and just add "nosharecache" onto the end of the string.

Cc Yongcheng Yang <[email protected]>
---
 tests/generic/732 | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tests/generic/732 +0 -0
index 785aac58f361..ae49152e42dc 100755

tests/generic/732 mkdir -p $testdir1 $testdir2 · line 43

Quoted 5 lines
tests/generic/732 +1 -1
@@ -40,7 +40,7 @@ mkdir -p $testdir1 $testdir2
 # Don't share the data and attribute caches among mount points for NFS.
 # This caching behavior is necessary to reproduce this issue as we're
 # checking the alignment of each mount point's own unique cache.
-[ "$FSTYP" = "nfs" ] && MOUNT_OPTIONS="-o nosharecache"
+[ "$FSTYP" = "nfs" ] && MOUNT_OPTIONS="$MOUNT_OPTIONS -o nosharecache"

Good to me, and looks like the later option replaces the former one, if there're same options (e.g. -o sharecache -o nosharecache).

Quoted 3 lines

Thanks, Zorro.

Now that I look more closely, I'm not sure this test will ever pass properly on NFS. It's basically doing this:

client1: A/f -> B/f
client2: B/f -> A/f
client1: A/f -> B/f

...and the two clients aren't aware of the changes the other is making (because they were mounted using -o nosharecache). The 3rd rename ends up thinking that the B/f dentry is still positive, and the rename syscall fails with EEXIST.

The really confusing bit is that this test passes against servers running older kernels, because the rename response had the wrong change info in it and that tricks the client into invalidating the directory caches when it shouldn't need to do that.

We fixed that in fdd2630a7398 (nfsd: fix change_info in NFSv4 RENAME replies), and now this test pretty reliably fails when testing against modern nfsd.

We have some longer term plans to add support for directory delegations eventually, which may make it easier to keep the caches more coherent, in this situation, but until then we might want to skip this test.

Signature · 1 line
Jeff Layton <[email protected]>
#3 lore raw
Quoted 44 lines · Jeff Layton
Quoted 25 lines · Zorro Lang
Quoted 15 lines · Jeff Layton

With NFS in particular, we are usually testing with some mount options. Ensure that we preserve those and just add "nosharecache" onto the end of the string.

Cc Yongcheng Yang <[email protected]>
---
 tests/generic/732 | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tests/generic/732 +0 -0
index 785aac58f361..ae49152e42dc 100755

tests/generic/732 mkdir -p $testdir1 $testdir2 · line 43

Quoted 5 lines
tests/generic/732 +1 -1
@@ -40,7 +40,7 @@ mkdir -p $testdir1 $testdir2
 # Don't share the data and attribute caches among mount points for NFS.
 # This caching behavior is necessary to reproduce this issue as we're
 # checking the alignment of each mount point's own unique cache.
-[ "$FSTYP" = "nfs" ] && MOUNT_OPTIONS="-o nosharecache"
+[ "$FSTYP" = "nfs" ] && MOUNT_OPTIONS="$MOUNT_OPTIONS -o nosharecache"

Good to me, and looks like the later option replaces the former one, if there're same options (e.g. -o sharecache -o nosharecache).

Quoted 3 lines

Thanks, Zorro.

Now that I look more closely, I'm not sure this test will ever pass properly on NFS. It's basically doing this:

client1: A/f -> B/f
client2: B/f -> A/f
client1: A/f -> B/f

...and the two clients aren't aware of the changes the other is making (because they were mounted using -o nosharecache). The 3rd rename ends up thinking that the B/f dentry is still positive, and the rename syscall fails with EEXIST.

Quoted 4 lines

The really confusing bit is that this test passes against servers running older kernels, because the rename response had the wrong change info in it and that tricks the client into invalidating the directory caches when it shouldn't need to do that.

Hi Jeff,

Thanks for this information.

Do you mean this case tests passed on nfs due to a bug, and that bug will be fixed soon?

I remember yoyang@ wrote this case for a known nfs issue. And this case tests passed on most filesystems. If it won't be suitable for newer nfsd, can we change it for new nfs? Or _notrun it by someone condition?

Quoted 8 lines

We fixed that in fdd2630a7398 (nfsd: fix change_info in NFSv4 RENAME replies), and now this test pretty reliably fails when testing against modern nfsd.

We have some longer term plans to add support for directory delegations eventually, which may make it easier to keep the caches more coherent, in this situation, but until then we might want to skip this test.

Is there any chance (change it a bit) to make this case still works for nfs?

Thanks,
Zorro

Quoted 4 lines
Signature · 2 lines
-- 
Jeff Layton <[email protected]>
#4 lore raw
Quoted 50 lines · Zorro Lang
Quoted 44 lines · Jeff Layton
Quoted 25 lines · Zorro Lang
Quoted 15 lines · Jeff Layton

With NFS in particular, we are usually testing with some mount options. Ensure that we preserve those and just add "nosharecache" onto the end of the string.

Cc Yongcheng Yang <[email protected]>
---
 tests/generic/732 | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tests/generic/732 +0 -0
index 785aac58f361..ae49152e42dc 100755

tests/generic/732 mkdir -p $testdir1 $testdir2 · line 43

Quoted 5 lines
tests/generic/732 +1 -1
@@ -40,7 +40,7 @@ mkdir -p $testdir1 $testdir2
 # Don't share the data and attribute caches among mount points for NFS.
 # This caching behavior is necessary to reproduce this issue as we're
 # checking the alignment of each mount point's own unique cache.
-[ "$FSTYP" = "nfs" ] && MOUNT_OPTIONS="-o nosharecache"
+[ "$FSTYP" = "nfs" ] && MOUNT_OPTIONS="$MOUNT_OPTIONS -o nosharecache"

Good to me, and looks like the later option replaces the former one, if there're same options (e.g. -o sharecache -o nosharecache).

Quoted 3 lines

Thanks, Zorro.

Now that I look more closely, I'm not sure this test will ever pass properly on NFS. It's basically doing this:

client1: A/f -> B/f
client2: B/f -> A/f
client1: A/f -> B/f

...and the two clients aren't aware of the changes the other is making (because they were mounted using -o nosharecache). The 3rd rename ends up thinking that the B/f dentry is still positive, and the rename syscall fails with EEXIST.

Quoted 4 lines

The really confusing bit is that this test passes against servers running older kernels, because the rename response had the wrong change info in it and that tricks the client into invalidating the directory caches when it shouldn't need to do that.

Quoted 7 lines

Hi Jeff,

Thanks for this information.

Do you mean this case tests passed on nfs due to a bug, and that bug will be fixed soon?

It has already been fixed (fdd2630a7398).

This is a bit of a confusing bug. It turns out that in one specific case, that patch actually helps a test somewhat similar to this to pass, but in most cases, having that bug fixed will make this test fail on NFS (and that behavior is expected, if not ideal).

The real problem is that this is trying to test the behavior of the NFSv4 change attribute, but that value isn't directly observable from userland today. All you can do is try to infer its effect on the cache, which is also impossible to observe directly.

Quoted 9 lines

I remember yoyang@ wrote this case for a known nfs issue. And this case tests passed on most filesystems. If it won't be suitable for newer nfsd, can we change it for new nfs? Or _notrun it by someone condition?

Quoted 5 lines

We fixed that in fdd2630a7398 (nfsd: fix change_info in NFSv4 RENAME replies), and now this test pretty reliably fails when testing against modern nfsd.

Quoted 6 lines
Quoted 3 lines

We have some longer term plans to add support for directory delegations eventually, which may make it easier to keep the caches more coherent, in this situation, but until then we might want to skip this test.

Is there any chance (change it a bit) to make this case still works for nfs?

Sure: You could wait out the dircache timeout (60s) between rename calls. See acdirmin/acdirmax mount options too, which could be used to lower those timeouts to make the test run more quickly. But...if you're using those mount options, you wouldn't hit the original bug that this is was trying to test in the first place.

It's probably better to just skip this on NFS. This would probably be a better test done via pynfs rather than xfstests, since that would allow you to directly observe that the change attribute changes when you expect it to.

Signature · 1 line
Jeff Layton <[email protected]>