From mboxrd@z Thu Jan  1 00:00:00 1970
From: Joerg Roedel <joro@8bytes.org>
Subject: mmu_notifier and i915_gem_userptr.c
Date: Thu, 19 Jun 2014 17:36:55 +0200
Message-ID: <20140619153655.GG31771@8bytes.org>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Return-path: <linux-kernel-owner@vger.kernel.org>
Content-Disposition: inline
Sender: linux-kernel-owner@vger.kernel.org
To: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>, Jani Nikula <jani.nikula@linux.intel.com>, David Airlie <airlied@linux.ie>, intel-gfx@lists.freedesktop.org, linux-kernel@vger.kernel.org
List-Id: intel-gfx@lists.freedesktop.org

Hey Chris,

recently I had a look at i915_gem_userptr.c in order to extend the
mmu_notifier call-backs implemented there. My goal is to implement the
change_pte call-back where it is necessary to get rid of it being
wrapped mn_invalidate_range_start/end() calls (for the reason see
commit 6bdb913f).

For most users of mmu_notifiers this is easy, except the i915 driver :)
The invalidate_range_start notifier implemented there can sleep, so it
can't be reused for a change_pte implementation (because change_pte is
called under ptl spin_lock and is not allowed to sleep). On the other
hand you also didn't implement the invalidate_page notifier, so I am not
sure whether the code actually cares about the somewhat similar
change_pte events?

Here is where change_pte is called from:

	* In KSM code when pages are merged (shouldn't be relevant
	  because KSM doesn't merged pages returned by get_user_pages())
	* In uprobes code when a user-page is replaced by a kernel page
	  (should only handle .text sections, so probaly not relevant
	   here)
	* When someone writes to a COW page in mm/memory.c (this looks
	  relevant looking at forked processes, on the other hand,
	  this is currently handled by unbinding the vma from the
	  object list in the i915 driver)

I am not familiar with the i915 hardware and the driver implementation
details, so I wanted to ask whether the driver

	1) Cares about the change_pte event?
	2) If it cares, what is the best way to implement it? What the
	   invalidate_range_start() notifier does seems a bit overkill,
	   since for the change_pte event nothing is unmapped (but maybe
	   remapped)

So any insight you could provide here would be useful :)


Thanks,

	Joerg