On Tue, Dec 06, 2016 at 05:00:47PM +0800, Hou Tao wrote:
> During the 2nd stage of log recovery, if the filesystem is firstly mounted
> as read-only, the unlink inodes will not be destroyed and the unlinked list
> in AGI will not be cleared. Even after a read-write remount or umount,
> the unlinked inodes will still be valid and be kept on disk, and the
> available freespace will be incorrect.
>
> To fix the problem, we need to force xfs_inactive() to destroy the
> unlinked inode when the filesystem is mounted as read-only.
> So clear the XFS_MOUNT_RDONLY flag temporarily before the recovery
> of unlinked inodes and restore the flag after the recovery has done.
>
> The problem can be reproduced by the following steps:
> 1. mount a xfs fs on a KVM VM
> 2. on the VM launch an application which does the following things:
> open(xfs_file); unlink(xfs_file);
> while(1) { write(xfs_file, 2MB); sleep(1); }
> 3. wait 5 seconds, sync the xfs fs, and wait 5 seconds
> 4. terminate the VM
> 5. start the VM and mount the xfs as read-only
> 6. remount the xfs as read-write or umount
> 7. check the unlinked list and the available freespace
This is only papering over the larger problem.
I was talking to Eric about this larger "recovery on read-only
mount" problem last week on IRC - I can't find it my logs right now,
but IIRC I'd suggested that we should always run xfs_mountfs()
in read-write mount if the underlying device can be written to,
and then once that is complete do a rw->ro transition exactly as we
do for a remount,ro operation.
That way we remove all the special "write on read only mount" hacks
we currently have throughout the code to enable log recovery to run
on read-only mounts.
Essentially is requires moving the device read only check from the
log code to xfs_fs_fill_super() and handling the no-recovery flag
there before we call xfs_mountfs() and adding the rw->ro state
transition after we return. This will be much simpler and much more
reliable than trying to turn off "read only" state around certain
operations...
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
--
To unsubscribe from this list: send the line "unsubscribe stable" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html