Re: [PATCH] drm/i915: fixup i915_gem_object_get_page inline helper

[Chris Wilson <chris@chris-wilson.co.uk>]

On Wed, 10 Oct 2012 11:21:44 +0200, Daniel Vetter <daniel@ffwll.ch> wrote:
> On Wed, Oct 10, 2012 at 10:01:47AM +0100, Chris Wilson wrote:
> > On Tue, 09 Oct 2012 23:16:02 +0100, Chris Wilson <chris@chris-wilson.co.uk> wrote:
> > > On Tue,  9 Oct 2012 22:50:48 +0200, Daniel Vetter <daniel.vetter@ffwll.ch> wrote:
> > > > The obj->pages to obj->pages->sgl rework introduced this helper, but
> > > > it doesn't actually work for n >= SG_MAX_SINGLE_ALLOC.
> > > > 
> > > > For simplicity (and since right now I seem to be too stupid to see
> > > > the bug), let's just grab the right page with a for_each_sg loop.
> > > > 
> > > > This is exercised by the improved hangman tests and the gem_exec_big
> > > > test in i-g-t.
> > > > 
> > > > v2: Compared to v1, don't try to be clever since I seemingly only
> > > > manage to prove that I'm not clever.
> > > 
> > > Only I expect that loop to show up on profiles even higher than the
> > > sg_next() from pwrite. :|
> > > 
> > > I expect it to have a measureable impact upon relocation throughput,
> > > so I should measure it...
> > > -Chris
> > > 
> > > -- 
> > > Chris Wilson, Intel Open Source Technology Centre
> > > _______________________________________________
> > > Intel-gfx mailing list
> > > Intel-gfx@lists.freedesktop.org
> > > http://lists.freedesktop.org/mailman/listinfo/intel-gfx
> > From: Chris Wilson <chris@chris-wilson.co.uk>
> > Subject: Re: [PATCH] drm/i915: fixup i915_gem_object_get_page inline helper
> > To: Daniel Vetter <daniel.vetter@ffwll.ch>, Intel Graphics Development <intel-gfx@lists.freedesktop.org>
> > Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
> > In-Reply-To: <1349815848-1824-1-git-send-email-daniel.vetter@ffwll.ch>
> > References: <84c8a8$624sk4@orsmga001.jf.intel.com> <1349815848-1824-1-git-send-email-daniel.vetter@ffwll.ch>
> > 
> > On Tue,  9 Oct 2012 22:50:48 +0200, Daniel Vetter <daniel.vetter@ffwll.ch> wrote:
> > > The obj->pages to obj->pages->sgl rework introduced this helper, but
> > > it doesn't actually work for n >= SG_MAX_SINGLE_ALLOC.
> > > 
> > > For simplicity (and since right now I seem to be too stupid to see
> > > the bug), let's just grab the right page with a for_each_sg loop.
> > > 
> > > This is exercised by the improved hangman tests and the gem_exec_big
> > > test in i-g-t.
> > > 
> > > v2: Compared to v1, don't try to be clever since I seemingly only
> > > manage to prove that I'm not clever.
> > > 
> > > Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
> > 
> > Looks like my worries are baseless. It can always be attacked latter if
> > need be. I'd still like to know what the mistake was...
> > 
> > Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
> 
> Merged to -fixes, with the missing regression-sha1 citation added.
> -Daniel
> -- 
> Daniel Vetter
> Software Engineer, Intel Corporation
> +41 (0) 79 365 57 48 - http://blog.ffwll.ch
From: Chris Wilson <chris@chris-wilson.co.uk>
Subject: Re: [Intel-gfx] [PATCH] drm/i915: fixup i915_gem_object_get_page inline helper
To: Daniel Vetter <daniel@ffwll.ch>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>, Intel Graphics Development <intel-gfx@lists.freedesktop.org>
In-Reply-To: <20121010091538.GB5533@phenom.ffwll.local>
References: <84c8a8$624sk4@orsmga001.jf.intel.com> <1349815848-1824-1-git-send-email-daniel.vetter@ffwll.ch> <6c3329$6m595h@orsmga002.jf.intel.com> <453bf0$61a4mf@azsmga001.ch.intel.com> <20121010091538.GB5533@phenom.ffwll.local>

On Wed, 10 Oct 2012 11:15:38 +0200, Daniel Vetter <daniel@ffwll.ch> wrote:
> On Wed, Oct 10, 2012 at 10:01:47AM +0100, Chris Wilson wrote:
> > On Tue, 09 Oct 2012 23:16:02 +0100, Chris Wilson <chris@chris-wilson.co.uk> wrote:
> > > On Tue,  9 Oct 2012 22:50:48 +0200, Daniel Vetter <daniel.vetter@ffwll.ch> wrote:
> > > > The obj->pages to obj->pages->sgl rework introduced this helper, but
> > > > it doesn't actually work for n >= SG_MAX_SINGLE_ALLOC.
> > > > 
> > > > For simplicity (and since right now I seem to be too stupid to see
> > > > the bug), let's just grab the right page with a for_each_sg loop.
> > > > 
> > > > This is exercised by the improved hangman tests and the gem_exec_big
> > > > test in i-g-t.
> > > > 
> > > > v2: Compared to v1, don't try to be clever since I seemingly only
> > > > manage to prove that I'm not clever.
> > > 
> > > Only I expect that loop to show up on profiles even higher than the
> > > sg_next() from pwrite. :|
> > > 
> > > I expect it to have a measureable impact upon relocation throughput,
> > > so I should measure it...
> > > -Chris
> > > 
> > > -- 
> > > Chris Wilson, Intel Open Source Technology Centre
> > > _______________________________________________
> > > Intel-gfx mailing list
> > > Intel-gfx@lists.freedesktop.org
> > > http://lists.freedesktop.org/mailman/listinfo/intel-gfx
> > From: Chris Wilson <chris@chris-wilson.co.uk>
> > Subject: Re: [PATCH] drm/i915: fixup i915_gem_object_get_page inline helper
> > To: Daniel Vetter <daniel.vetter@ffwll.ch>, Intel Graphics Development <intel-gfx@lists.freedesktop.org>
> > Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
> > In-Reply-To: <1349815848-1824-1-git-send-email-daniel.vetter@ffwll.ch>
> > References: <84c8a8$624sk4@orsmga001.jf.intel.com> <1349815848-1824-1-git-send-email-daniel.vetter@ffwll.ch>
> > 
> > On Tue,  9 Oct 2012 22:50:48 +0200, Daniel Vetter <daniel.vetter@ffwll.ch> wrote:
> > > The obj->pages to obj->pages->sgl rework introduced this helper, but
> > > it doesn't actually work for n >= SG_MAX_SINGLE_ALLOC.
> > > 
> > > For simplicity (and since right now I seem to be too stupid to see
> > > the bug), let's just grab the right page with a for_each_sg loop.
> > > 
> > > This is exercised by the improved hangman tests and the gem_exec_big
> > > test in i-g-t.
> > > 
> > > v2: Compared to v1, don't try to be clever since I seemingly only
> > > manage to prove that I'm not clever.
> > > 
> > > Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
> > 
> > Looks like my worries are baseless. It can always be attacked latter if
> > need be. I'd still like to know what the mistake was...
> 
> I think it was two mistakes:
> - One was the off-by-one fixed in v1.
> - Second seemed to be the special case that if the table fits exactly,
>   sg_alloc doesn't set a chain ptr with another sg table with just one
>   entry.

That oddity is the reason why the loop was structured so.

The mistake is that just because we have n == MAX, does not mean that
there are only n elements left in the sg! *hides*

diff --git a/drivers/gpu/drm/i915/i915_drv.h
b/drivers/gpu/drm/i915/i915_drv.h
index d2dda78..e6707e7 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -1383,9 +1383,14 @@ int __must_check i915_gem_object_get_pages(struct drm_i91
 static inline struct page *i915_gem_object_get_page(struct drm_i915_gem_object 
 {
        struct scatterlist *sg = obj->pages->sgl;
-       while (n >= SG_MAX_SINGLE_ALLOC) {
+       int nents = obj->pages->orig;
+       while (nents > SG_MAX_SINGLE_ALLOC) {
+               if (n < SG_MAX_SINGLE_ALLOC - 1)
+                       break;
+
                sg = sg_chain_ptr(sg + SG_MAX_SINGLE_ALLOC - 1);
                n -= SG_MAX_SINGLE_ALLOC - 1;
+               nents -= SG_MAX_SINGLE_ALLOC - 1;
        }
        return sg_page(sg+n);
 }

-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre