Re: [PATCH] lglock: make lg_lock_global() actually lock globally

[Nick Piggin <npiggin@kernel.dk>]

On Wed, Aug 25, 2010 at 02:16:44PM -0600, Jonathan Corbet wrote:
> On Wed, 25 Aug 2010 13:00:59 -0700
> Linus Torvalds <torvalds@linux-foundation.org> wrote:
> 
> > Grrr. Same disease as Nick and others. Why do you repeat the subject
> > line in the body? Don't do that. We don't want the summary line twice
> > in the commit message, and we don't want it twice in the email.
> > 
> > We simply don't want it twice. Full stop.
> 
> Sorry, I just pasted in the "git format-patch" output.  Will never ever
> ever do it again I promise cross my heart.
> 
> > > lg_lock_global() currently only acquires spinlocks for online CPUs, but
> > > it's meant to lock all possible CPUs.  At Nick's suggestion, change
> > > for_each_online_cpu() to for_each_possible_cpu() to get the expected
> > > behavior.  
> > 
> > Can you say what this actually matters for? Don't we do stop-machine
> > for CPU hotplug anyway? And if we don't, shouldn't we? Exactly because
> > otherwise "for_each_online_cpu()" is always racy (and that has nothing
> > to do with the lglock).
> 
> As I understand it from Nick (after I asked him why the two lock
> primitives were identical): the files_lock scalability work puts a
> per-CPU list of open files into each superblock.  A CPU can be offlined
> while there are open files in "its" lists, and nothing is done to shift
> those files to a still-online CPU's list.  So there will still be
> cross-CPU accesses to those lists as those files are closed; that means
> we need to be sure to acquire locks associated with offline CPUs if we
> want to avoid races.
> 
> lg_global_lock_online() is used (only) in the brlock implementation,
> instead.  In this case, there's no leftover data if a CPU goes
> offline, so no need to take locks associated with offline CPUs.

Yep, thanks Jon, I owe a bit more documentation in that file, coming up.