Re: [PATCH RESEND v4 6/8] vfs: Add get_vfsmount_sb() function to get vfsmount from a given sb.

[Al Viro <viro@ZenIV.linux.org.uk>]

On Fri, Jan 30, 2015 at 12:14:24PM +0800, Miao Xie wrote:
> On Fri, 30 Jan 2015 02:14:45 +0000, Al Viro wrote:
> > On Fri, Jan 30, 2015 at 09:44:03AM +0800, Qu Wenruo wrote:
> > 
> >> This shouldn't happen. If someone is ro, the whole fs should be ro, right?
> > 
> > Wrong.  Individual vfsmounts over an r/w superblock might very well be r/o.
> > As for that trylock...  What for?  It invites transient failures for no
> > good reason.  Removal of sysfs entry will block while write(2) to that sucker
> > is in progress, so btrfs shutdown will block at that point in ctree_close().
> > It won't go away under you.
> 
> could you explain the race condition? I think the deadlock won't happen, during
> the btrfs shutdown, we hold s_umount, the write operation will fail to lock it,
> and quit quickly, and then umount will continue.

	First of all, ->s_umount is not a mutex; it's rwsem.  So you mean
down_read_trylock().  As for the transient failures - grep for down_write
on it...  E.g. have somebody call mount() from the same device.  We call
sget(), which finds existing superblock and calls grab_super().  Sure,
that ->s_umount will be released shortly, but in the meanwhile your trylock
will fail...

> I think sb_want_write() is similar to trylock(s_umount), the difference is that
> sb_want_write() is more complex.
> 
> > 
> > Now, you might want to move those sysfs entry removals to the very beginning
> > of btrfs_kill_super(), but that's a different story - you need only to make
> > sure that they are removed not later than the destruction of the data
> > structures they need (IOW, the current location might very well be OK - I
> > hadn't checked the details).
> 
> Yes, we need move those sysfs entry removals, but needn't move to the very
> beginning of btrfs_kill_super(), just at the beginning of close_ctree();

So move them...  It's a matter of moving one function call around a bit.