VFS: busy inodes on changed media or resized disk pool1/ext4

[tobert]

My test system is a Dell 1950 being loaned to me from another team with 16GB RAM, 24 1TB SAS drives (external JBOD), and two quad-core L5420's. The OS is Ubuntu 9.10 64-bit with the 2.6.31-14-server kernel (not what I'm targeting in production). The team loaning me the server loaded Ubuntu over Solaris 10. SPL is 0.5.1 and I'm switching between ZFS 0.5.1 and git head.

My benchmark ran out of space so I thought I'd try resizing the ext4 filesystem I have on one of the zvols and discovered that even though the "zfs set volsize=2T pool1/foobar" succeeded, the kernel did not pick up the change in size and therefore resize2fs will not resize the filesystem. This works fine when using LVM2, so I would assume there's no reason zvols shouldn't work.

dmesg shows this issue's title, "VFS: busy inodes on changed media or resized disk pool1/foobar".

I also tried running sfdisk -R and blockdev --getsize64 $device to no avail.

Going off clues on the ChangeLog mentioning the switch to check_disk_change() for compatibility, I went into modules/zfs/zvol.c and changed the call to check_disk_change() in zvol_update_volsize() to use revalidate_disk(bdev->bd_disk), rebuilt/reloaded, and now resizing seems to work. I tested resizing the zvol with heavy write I/O going to the filesystem (both xfs and ext4) and verified that the resize is recognized by the kernel and that xfs_grow and resize2fs behave as expected after the change to revalidate_disk.

Now the kernel shows "pool1/foobar: detected capacity change from 536870912000 to 1099511627776".

It may be best to put an ifdef or configure check in place to use revalidate_disk() where it's available.

    if (bdev)
        return EIO;

/* this isn't the correct way, but sufficient for testing the concept */
#ifdef revalidate_disk
    error = revalidate_disk(bdev->bd_disk);
#else
    error = check_disk_change(bdev);
#endif
    ASSERT3U(error, !=, 0);

[behlendorf]

Interesting. Thanks for the bug report and digging in to the issue. I'll get a proper fix merged in to the latest code.

[behlendorf]

I've dug in to your fix more carefully and I think it's a viable workaround for now but probably not the right long term fix. The check_disk_change() function should do what we want although we need to make sure we get the media_changed and revalidate_disk callbacks right. This fix is going to miss the next tag but we'll get it fixed.