linux-kernel - Re: [PATCH 1/4] locking/ww_mutex: add ww_mutex_is_owned

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20180220235621.GD22199@phenom.ffwll.local>
Date:   Wed, 21 Feb 2018 00:56:21 +0100
From:   Daniel Vetter <daniel@...ll.ch>
To:     Peter Zijlstra <peterz@...radead.org>
Cc:     christian.koenig@....com, dev@...ankhorst.nl,
        amd-gfx@...ts.freedesktop.org, dri-devel@...ts.freedesktop.org,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH 1/4] locking/ww_mutex: add ww_mutex_is_owned_by function
 v3

On Tue, Feb 20, 2018 at 04:21:58PM +0100, Peter Zijlstra wrote:
> On Tue, Feb 20, 2018 at 04:05:49PM +0100, Christian König wrote:
> > Am 20.02.2018 um 15:54 schrieb Peter Zijlstra:
> > > On Tue, Feb 20, 2018 at 03:34:07PM +0100, Christian König wrote:
> > > > > OK, but neither case would in fact need the !ctx case right? That's just
> > > > > there for completeness sake?
> > > > Unfortunately not. TTM uses trylock to lock BOs which are about to be
> > > > evicted to make room for all the BOs locked with a ctx.
> > > > 
> > > > I need to be able to distinct between the BOs which are trylocked and those
> > > > which are locked with a ctx.
> > > > 
> > > > Writing this I actually noticed the current version is buggy, cause even
> > > > when we check the mutex owner we still need to make sure that the ctx in the
> > > > lock is NULL.
> > > Hurm... I can't remember why trylocks behave like that, and it seems
> > > rather unfortunate / inconsistent.
> > 
> > Actually for me that is rather fortunate, cause I need to distinct between
> > the locks acquired through trylock and lock.
> 
> I suppose that would always be possible using:
> ww_mutex_trylock(.ctx=NULL), and it could be that there simply weren't
> any immediate uses for a !NULL trylock and it was thus not implemented.
> 
> But that is all very long ago..

I think we simple never had a use-case for interleaving ww_mutex_lock(ctx)
and ww_mutex_trylock(ctx). Nesting multiple trylocks in ctx-locks happens
plenty, but not further:

The common use-case for that is locking a bunch of buffers you need (for
command submission or whatever), and then trylocking other buffers to make
space for the buffers you need to move into VRAM. I guess if only
trylocking buffers doesn't succeed in freeing up enough VRAM then we could
go into blocking ww_mutex_locks which need the ctx (and which would need
all the trylock-acquired buffers to be annotated with the ctx too). TTM
currently tries to be far enough away from that corner case (using a
defensive "never use more than 50% of all memory for gfx" approach) that
it doesn't seem to need that.

Once we get there it should indeed be simply to add a ctx parameter to
ww_mutex_trylock to fix this case. The TTM side rework is definitely going
to be the much bigger issue here ...
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch