Linux 6.14 compat: Check for migratepage VFS

[tonyhutter]

Motivation and Context

Support 6.14 kernel.

Description

The 6.14 kernel doesn't expect a 'migratepage' VFS op. Check for migratepage.

I'm not sure how this slipped though the cracks when we bumped META to 6.14. Maybe because I was testing against the Fedora 6.14 kernel and that's different from kernel.org one?

How Has This Been Tested?

Built against kernel.org source with ./configure --with-linux=<path to source>.

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Performance enhancement (non-breaking change which improves efficiency)
• Code cleanup (non-breaking change which makes code smaller or more readable)
• Breaking change (fix or feature that would cause existing functionality to change)
• Library ABI change (libzfs, libzfs_core, libnvpair, libuutil and libzfsbootenv)
• Documentation (a change to man pages or other documentation)

Checklist:

• My code follows the OpenZFS code style requirements.
• I have updated the documentation accordingly.
• I have read the contributing document.
• I have added tests to cover my changes.
• I have run the ZFS Test Suite with this change applied.
• All commit messages are properly formatted and contain Signed-off-by.

[robn]

The kernel.org 6.14.0 has indeed removed migratepage, but has migrate_folio, so never gets into the #else. I've just built against 6.14.0 again and it worked fine (HAVE_VFS_MIGRATE_FOLIO is set), so I wonder if you've got something else weird going on?

I don't mind this PR as such, because I rather prefer tests that ask for the thing they need, rather than rely on the absence of some other thing, so I'll go test it on some older kernels too. But I'd still like to know what's happening so you needed it?

(FWIW, I tend to hold Fedora kernels at arms length when thinking specific kernel versions. They're not as weird as full Redhat kernels can be, but they're still a bit weird. Another reason to test for the specific things we want!)

[robn]

This should be .migratepage: no one has .migrate_page :)

[robn]

I'd prefer #elif defined(HAVE_VFS_MIGRATEPAGE) etc but it hardly matters.

[robn]

FYI, with the test changed to .migratepage it compiles with the correct results for migrate_folio vs migratepage on kernels: 4.19.325 5.4.291 5.10.235 5.15.179 6.1.131 6.2.16 6.3.13 6.4.16 6.5.13 6.6.84 6.7.12 6.8.12 6.9.12 6.10.14 6.11.11 6.12.20 6.13.8 6.14.0

(I did not try using it though)

[tonyhutter]

I've just built against 6.14.0 again and it worked fine (HAVE_VFS_MIGRATE_FOLIO is set), so I wonder if you've got something else weird going on?

Here's the specific failure I'm seeing:

  CC [M]  os/linux/zfs/zvol_os.o
  CC [M]  icp/algs/aes/aes_impl_aesni.o
  CC [M]  icp/algs/aes/aes_impl_x86-64.o
os/linux/zfs/zpl_file.c:1062:10: error: ‘const struct address_space_operations’ has no member named ‘migratepage’; did you mean ‘writepage’?
 1062 |         .migratepage    = migrate_page,
      |          ^~~~~~~~~~~
      |          writepage
os/linux/zfs/zpl_file.c:1062:27: error: ‘migrate_page’ undeclared here (not in a function); did you mean ‘migrate_pages’?
 1062 |         .migratepage    = migrate_page,
      |                           ^~~~~~~~~~~~
      |                           migrate_pages
make[6]: *** [/mnt/kernel/linux-6.14.1/scripts/Makefile.build:207: os/linux/zfs/zpl_file.o] Error 1
make[6]: *** Waiting for unfinished jobs....
make[5]: *** [/mnt/kernel/linux-6.14.1/Makefile:1994: .] Error 2

I'm building against the kernel.org 6.14.1 kernel (https://cdn.kernel.org/pub/linux/kernel/v6.x/linux-6.14.1.tar.xz) which I built using a make tinyconfig configuration.

Now I'm thinking we can just get rid of .migratepage entirely since we don't use it:

diff --git a/module/os/linux/zfs/zpl_file.c b/module/os/linux/zfs/zpl_file.c
index 787d3cb31..e6ed55157 100644
--- a/module/os/linux/zfs/zpl_file.c
+++ b/module/os/linux/zfs/zpl_file.c
@@ -1058,8 +1058,6 @@ const struct address_space_operations zpl_address_space_operations = {
 #endif
 #ifdef HAVE_VFS_MIGRATE_FOLIO
        .migrate_folio  = migrate_folio,
-#else
-       .migratepage    = migrate_page,
 #endif
 };
 

[robn]

Ahh, I think I see what's happening. The test doesn't just test that .migrate_folio exists, but also that the default function can be assigned to (rather than one we provide). But, the default migrate_folio function doesn't exist when CONFIG_MIGRATION is not set, which it is not for tinyconfig (and same for migrate_page pre-6.0). So the assignment fails in the test and we don't set HAVE_VFS_MIGRATE_FOLIO.

I wrote 74ae760 and was just about to offer it, but then realised that's basically what you'd already done! So if you fix up your migratepage test to set .migratepage, we'll have basically the same thing, and I'll be happy to +1 it.

Glad to understand it :)

[robn]

6.0.

[robn]

Re getting rid of it, I think we have to keep it - kernel looks like it it doesn't always check for it to be non-NULL when calling it, so we at least have to set it to the fallback.

The couple of filesystems that do set it put it behind a #ifdef CONFIG_MIGRATION gate, but we don't need to because the autoconf test wont set our define anyways.

[tonyhutter]

@robn thanks for detective work on what was going on. My latest push should include all the changes you were looking for.

[robn]

Great 👍