Re: [PATCH 00/11] V4: rwsem changes + down_read_critical() proposal

[Peter Zijlstra <peterz@infradead.org>]

On Thu, 2010-05-27 at 03:59 -0700, Michel Lespinasse wrote:
> On Tue, May 25, 2010 at 11:27:55AM +0200, Peter Zijlstra wrote:
> > On Tue, 2010-05-25 at 02:12 -0700, Michel Lespinasse wrote:
> > > Yes, we do have patches trying to release the mmap_sem when a page
> > > fault for a file backed VMA blocks on accessing the corresponding
> > > file. We have not given up on these, and we intend to try submitting
> > > them again. However, these patches do *not* address the case of a page
> > > fault blocking while trying to get a free page (i.e. when you get
> > > under high memory pressure).
> > 
> > But I guess they could, right? Simply make the allocation under mmap_sem
> > be __GFP_HARDWALL|__GFP_HIGHMEM|__GFP_MOVABLE__GFP_NOWARN or
> > (GFP_HUGHUSER_MOVABLE & ~(__GFP_WAIT|__GFP_IO|__GFP_FS))|__GFP_NOWARN
> > 
> > and drop the mmap_sem when that fails.
> 
> It's not clear to me if this can lead to a clean uncontroversial solution.
> Doing this for file backed VMAs does not sound any harder in principle,
> but we could not get it past linus's NACK last time. I think it's worth
> exploring again, but I don't expect it to be so easy :)

Right, I was planning to have a look again, but my todo list is getting
out of hand again.

> > > > I really don't like people tinkering with the lock implementations like
> > > > this. Nor do I like the naming, stats are in no way _critical_.
> > > 
> > > Critical here refers to the fact that you're not allowed to block
> > > while holding the unfairly acquired rwsem.
> > 
> > We usually call that atomic, your 0/n patch didn't explain any of that.
> 
> Would replacing the 'critical' name with 'atomic' address your concern
> though, or would you remain fundamentally opposed to anything that involves
> an unfair acquire path ?

Yeah, I don't think its a good idea to add unfairness to the lock, such
things tend to backfire and generate starvation cases. For instance, a
tight loop around your /proc/$pid/maps mod will starve $pid from doing
anything requiring mmap_sem for writing.

> What about patches 1-7 which don't deal with the critical/atomic API;
> can we agree to get these in before we we figure out what to do with
> the the last 4 ?

patch 6 looks like it will break fairness, either that or the changelog
got me totally confused (already had to read it twice).