Re: VFS deadlock ?

[Dave Jones <davej@redhat.com>]

On Thu, Mar 21, 2013 at 01:32:45PM -0700, Linus Torvalds wrote:
 > On Thu, Mar 21, 2013 at 1:26 PM, Dave Jones <davej@redhat.com> wrote:
 > >
 > > those are qstr's, so I used d_name.name, right ?
 > 
 > Yup. And if you want to, you could do p1->d_parent->d_name.name too,
 > just to make things obvious. It's technically racy, but by the time
 > the bug happens, who cares?

I'll add that next time around if the current run doesn't turn up anything interesting.
 
 > > I'd be surprised actually, I've got sysfs excluded from its list of victim files,
 > > due to unrelated issues still unresolved.  So unless it followed a symlink into
 > > sys from somewhere in /proc or /dev...
 > >
 > > It took a few hours to reproduce last time, I'll increase the number of child
 > > processes to see if I can trigger it faster now that I have the debug stuff in there.
 > 
 > Hmm, ok. Do you have any network mounts or fuse or other "odd"
 > filesystems etc? The whole "aliased inodes" thing might come from
 > something like that.

at some point during the fuzz run, this happened..

Mar 20 15:20:41 kernel: [ 7578.784674] fuse init (API version 7.21)
Mar 20 15:20:41 systemd[1]: Mounting FUSE Control File System...
Mar 20 15:20:41 systemd[1]: Mounted FUSE Control File System.

I guess something wandered into /dev/fuse and did something. Not sure why 
systemd reacted though...

	Dave