Re: xfs_repair: add '-F' option to ignore writable mount checking

[Brian Foster <bfoster@redhat.com>]

On Mon, Feb 26, 2018 at 10:59:07AM +0800, Yang Joseph wrote:
> In our case, there is a mountpoint of ceph-fuse type and this mountpoint is
> abnormal.
> I execute 'xfs_repair -n /dev/nbd4' cmd. Then xfs_repair is blocked in
> stat()
> systemcall. '/dev/nbd4' has no relationship with the ceph-fuse mountpoint.
> 
> [root@compute5 ~]# ps aux | grep xfs_repair
> root     16469  0.0  0.0 114744   564 ?        D    10:50   0:00 xfs_repair
> -n /dev/nbd4
> 
> [root@compute5 ~]# cat /proc/16469/stack
> [<ffffffffa04b953d>] __fuse_request_send+0x13d/0x2c0 [fuse]
> [<ffffffffa04b96d2>] fuse_request_send+0x12/0x20 [fuse]
> [<ffffffffa04be67a>] fuse_do_getattr+0x11a/0x2e0 [fuse]
> [<ffffffffa04bfba5>] fuse_update_attributes+0x75/0x80 [fuse]
> [<ffffffffa04bfbf3>] fuse_getattr+0x43/0x50 [fuse]
> [<ffffffff81203976>] vfs_getattr+0x46/0x80
> [<ffffffff81203aa5>] vfs_fstatat+0x75/0xc0
> [<ffffffff81203ffe>] SYSC_newstat+0x2e/0x60
> [<ffffffff812042de>] SyS_newstat+0xe/0x10
> [<ffffffff81697809>] system_call_fastpath+0x16/0x1b
> [<ffffffffffffffff>] 0xffffffffffffffff
> 
> The stat() is from the following code:
> 
> // libxfs/linux.c:platform_check_mount()
>   while ((mnt = getmntent(f)) != NULL) {
>     if (stat64(mnt->mnt_fsname, &mst) < 0) <---------<<<< unconditionally
> stat all mountpoints
>       continue;
> 
> xfs_repair have to check all mountpoints of the system to make sure there is
> no writable mount point of user specified device. If there is one abnormal
> mountpoint, event it not related to user specified device, xfs_repair will
> be blocked.
> 
> I can make sure there is no writable mountpoint of /dev/nbd4, so xfs_repair
> don't need to check all mountpoints of the system. This is why I want to add
> this '-F' option.
> 
> Because there are lots of other services on this node, I can't reboot the
> machine.
> 

Suffice it to say that I agree with the other comments that this
probably isn't something we want to "fix" in xfs_repair... but given
your particular circumstances, would a lazy unmount of the borked mount
allow the repair to proceed?

Brian

> thx
> 
> Yang Honggang
> 
> On 02/25/2018 06:04 AM, Dave Chinner wrote:
> > On Sat, Feb 24, 2018 at 11:56:44AM -0600, Eric Sandeen wrote:
> > > On 2/24/18 5:23 AM, Yang Joseph wrote:
> > > > hello,
> > > > 
> > > > Before the repair process, xfs_repair will check if user specified device already
> > > > has a writable mountpoint. And it will stat all the mountpoints of the system. If there
> > > > is a dead mountpoint, this checking will be blocked and xfs_repair will enter 'D' state.
> > So why is the mount point dead?
> > 
> > That kinda means that the filesystem is still mounted, but something
> > has hung somewhere and the filesystem may still have active
> > references to the underlying device and be doing stuff that is
> > modifying the filesystem....
> > 
> > And if the device is still busy, then you aren't going to be able to
> > mount the repaired device, anyway, because the block device is still
> > busy...
> > 
> > > That sounds like a bug worth fixing, but I am much
> > > less excited about adding options which could do serious damage
> > > to a filesystem.
> > TO me it sounds like something that should be fixed by a reboot, not
> > by adding dangerous options to xfs_repair...
> > 
> > Cheers,
> > 
> > Dave.
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html