On Fri, Mar 29, 2013 at 2:12 PM, Linus Torvalds <[email protected]> wrote: > > I dunno. I'm still not sure this is triggerable, but it looks bad. But > both the semaphore case and the msg cases seem to be solvable by > moving the unlock down, and shm seem to have no getref/putref users to > race with, so this (whitespace-damaged) patch *may* be sufficient:
Well, the patch doesn't seem to cause any problems, at least neither
lockdep nor spinlock sleep debugging complains. I have no idea whether
it actually fixes any problems, though.
I do wonder if this might explain the problem Emmanuel saw. A double
free of a RCU-freeable object would possibly result in exactly the
kind of mess that Emmanuel reported with the semaphore scalability
patches.
Emmanuel, can you try the attached patch? I think it applies cleanly
on top of the scalability series too without any changes, but I didn't
check if the patches perhaps changed some of the naming or something.
Linus
patch.diff
Description: Binary data
