On Tue, 17 Feb 2015 17:21:02 -0500 "J. Bruce Fields" <bfields@xxxxxxxxxxxx> wrote: > On Tue, Feb 17, 2015 at 02:11:36PM -0500, J. Bruce Fields wrote: > > On Tue, Feb 17, 2015 at 12:56:49PM -0500, Jeff Layton wrote: > > > On Tue, 17 Feb 2015 12:10:17 -0500 > > > "J. Bruce Fields" <bfields@xxxxxxxxxxxx> wrote: > > > > > > > On Tue, Feb 17, 2015 at 07:46:28AM -0500, Jeff Layton wrote: > > > > > As Linus pointed out: > > > > > > > > > > Say we have an existing flock, and now do a new one that conflicts. I > > > > > see what looks like three separate bugs. > > > > > > > > > > - We go through the first loop, find a lock of another type, and > > > > > delete it in preparation for replacing it > > > > > > > > > > - we *drop* the lock context spinlock. > > > > > > > > > > - BUG #1? So now there is no lock at all, and somebody can come in > > > > > and see that unlocked state. Is that really valid? > > > > > > > > > > - another thread comes in while the first thread dropped the lock > > > > > context lock, and wants to add its own lock. It doesn't see the > > > > > deleted or pending locks, so it just adds it > > > > > > > > > > - the first thread gets the context spinlock again, and adds the lock > > > > > that replaced the original > > > > > > > > > > - BUG #2? So now there are *two* locks on the thing, and the next > > > > > time you do an unlock (or when you close the file), it will only > > > > > remove/replace the first one. > > > > > > > > > > ...remove the "drop the spinlock" code in the middle of this function as > > > > > it has always been suspicious. > > > > > > > > Looking back through a historical git repo, purely out of curiosity--the > > > > cond_resched was previously a yield, and then I *think* bug #2 was > > > > introduced in 2002 by a "[PATCH] fs/locks.c: more cleanup". Before that > > > > it retried the first loop after the yield. > > > > > > > > Before that the yield was in locks_wake_up_blocks, removed by: > > > > > > > > commit 7962ad19e6300531784722c16849837864304d84 > > > > Author: Matthew Wilcox <willy@xxxxxxxxxx> > > > > Date: Sat Jun 8 02:09:24 2002 -0700 > > > > > > > > [PATCH] fs/locks.c: Only yield once for flocks > > > > > > > > This patch removes the annoying and confusing `wait' argument > > > > from many places. The only change in behaviour is that we now > > > > yield once when unblocking other BSD-style flocks instead of > > > > once for each lock. > > > > > > > > This slightly improves the semantics for userspace. Before, > > > > when we had two tasks waiting on a lock, the first one would > > > > receive the lock. Now, the one with the highest priority > > > > receives the lock. > > > > > > > > So this really was intended to guarantee other waiters the lock before > > > > allowing an upgrade. Could that actually have worked? > > > > > > > > The non-atomic behavior is documented in flock(2), which says it's > > > > "original BSD behavior". > > > > > > > > A comment at the top of locks.c says this is to avoid deadlock. Hm, so, > > > > are we introducing a deadlock?: > > > > > > > > 1. Two processes both get shared lock on different filps. > > > > 2. Both request exclusive lock. > > > > > > > > Now they're stuck, whereas previously they might have recovered? > > > > > > > > --b. > > > > > > > > > > > > > Yes, I see that now. It might be best to preserve the existing behavior > > > for that reason then. I'd rather not introduce any behavioral changes in this > > > set if we can help it, particularly if there are userland apps that > > > might rely on it. > > > > > > It may be best then to keep this patch, but drop patch #3. > > > > Unfortunately it's this patch that I'm worried about. > > (Urp, never mind, you and Linus were right here.) > > > Also these patches are introducing some failure in my locking tests > > (probably unrelated, I doubt I ever wrote a test for this case). I'll > > investigate. > > I didn't try to figure out, but after dropping patch 3 everything does > pass. > > --b. > Ok, thanks. I've also figured out the problem that Al reported too. The "if (left == right)" case in __posix_lock_file was inserting "left" into the wrong spot. It needs to be inserted just before the new lock, instead of just before "fl". I'll include the fix for that in the pull request once I send it. In the meantime, I plan to sit down this evening with the LTP tests and make sure there is no other breakage there. -- Jeff Layton <jeff.layton@xxxxxxxxxxxxxxx> -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
