linux-kernel - Re: [Intel-gfx] [rfc PATCH] drm/i915: Simplify shrinker

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] [day] [month] [year] [list]

Date:	Tue, 19 Jul 2016 09:03:17 +0200
From:	Daniel Vetter <daniel@...ll.ch>
To:	Chris Wilson <chris@...is-wilson.co.uk>,
	Davidlohr Bueso <dave@...olabs.net>, daniel.vetter@...el.com,
	jani.nikula@...ux.intel.com, intel-gfx@...ts.freedesktop.org,
	linux-kernel@...r.kernel.org
Subject: Re: [Intel-gfx] [rfc PATCH] drm/i915: Simplify shrinker_lock

On Sun, Jul 17, 2016 at 10:54:51PM +0100, Chris Wilson wrote:
> On Sun, Jul 17, 2016 at 11:45:44AM -0700, Davidlohr Bueso wrote:
> > In addition, we can simplify the overall function wrt (2), by first
> > checking if we are the lock owner, then address the trylock and
> > deal with (2) if locked/contended by a traditional mutex_lock().
> > This should be safe considering that if current is the lock owner,
> > then we are guaranteed not to race with the counter->owner updates
> > (the counter is updated first which sets the mutex to be visibly locked).
> 
> However, that is then subject to an indirect ABBA deadlock, between the
> shrinker lock and the struct mutex (or at least that used to be the case
> where the kswapd reclaim would be blocked on the mutex and an alloc
> blocked on kswapd).
> 
> Unravelling the gross locking is an ongoing task, with one of the chief
> goals being able to reclaim memory whenever required. It is not pretty
> and often fails under pressure.

Yeah, what we need is to split up the dev->struct_mutex Big Driver Lock to
separate concerns. What's propably needed is a low-level mm lock (under
which we never ever allocate anything to avoid the deadlock with reclaim).
Plus probably per-object locks (using ww_mutex) to be able to protect
buffer against both from the shrinker (which would trylock, considering
locked objects busy) against threads and each another. We also might need
per-submission context locks to avoid havoc there, but not sure.

The reason this is taking forever to get done is that compared to the
existing locking, this new scheme is even more complex ;-)
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch