[lkml]   [2020]   [Mar]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [locks] 6d390e4b5d: will-it-scale.per_process_ops -96.6% regression
On Wed, 2020-03-11 at 09:57 +0800, yangerkun wrote:


> On 2020/3/11 5:01, NeilBrown wrote:
> >
> > I think this patch contains an assumption which is not justified. It
> > assumes that if a wait_event completes without error, then the wake_up()
> > must have happened. I don't think that is correct.
> >
> > In the patch that caused the recent regression, the race described
> > involved a signal arriving just as __locks_wake_up_blocks() was being
> > called on another thread.
> > So the waiting process was woken by a signal *after* ->fl_blocker was set
> > to NULL, and *before* the wake_up(). If wait_event_interruptible()
> > finds that the condition is true, it will report success whether there
> > was a signal or not.
> Neil and Jeff, Hi,
> But after this, like in flock_lock_inode_wait, we will go another
> flock_lock_inode. And the flock_lock_inode it may return
> - 0: If there is a try lock, it means that we have call
> locks_move_blocks, and fl->fl_blocked_requests will be NULL, no need to
> wake up at all. If there is a unlock, no one call wait for me, no need
> to wake up too.
> - ENOENT: means we are doing unlock, no one will wait for me, no need to
> wake up.
> - ENOMEM: since last time we go through flock_lock_inode someone may
> wait for me, so for this error, we need to wake up them.
> - EAGAIN: since we has go through flock_lock_inode before, these may
> never happen because FL_SLEEP will not lose.
> So the assumption may be ok and for some error case we need to wake up
> someone may wait for me before(the reason for the patch "cifs: call
> locks_delete_block for all error case in cifs_posix_lock_set"). If I am
> wrong, please point out!

That's the basic dilemma. We need to know whether we'll need to delete
the block before taking the blocked_lock_lock.

Your most recent patch used the return code from the wait to determine
this, but that's not 100% reliable (as Neil pointed out). Could we try
to do this by doing the delete only when we get certain error codes?
Maybe, but that's a bit fragile-sounding.

Neil's most recent patch used presence on the fl_blocked_requests list
to determine whether to take the lock, but that relied on some very
subtle memory ordering. We could of course do that, but that's a bit
brittle too.

That's the main reason I'm leaning toward the patch Neil sent
originally and that uses the fl_wait.lock. The existing alternate lock
managers (nfsd and lockd) don't use fl_wait at all, so I don't think
doing that will cause any issues.

Jeff Layton <>

 \ /
  Last update: 2020-03-11 13:52    [W:0.109 / U:3.572 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site