[Intel-gfx] [RFC] drm/i915: Handle E2BIG error in i915_gem_fault

[Chris Wilson]

On Wed, Sep 09, 2015 at 05:53:57PM +0100, Michel Thierry wrote:
> On 9/9/2015 11:33 AM, Chris Wilson wrote:
> >On Wed, Sep 09, 2015 at 11:29:52AM +0100, Michel Thierry wrote:
> >>i915_gem_object_bind_to_vm returns -E2BIG when the user tries to bind an
> >>object larger than the aperture, but i915_gem_fault does not handle this
> >>return code:
> >>
> >>[501906.530985] gem_mmap_gtt: starting subtest big-bo-tiledX
> >>[501906.541992] gem_mmap_gtt (22362): drop_caches: 3
> >>[501906.610774] WARNING: CPU: 2 PID: 22362 at
> >>drivers/gpu/drm/i915/i915_gem.c:1880 i915_gem_fault+0x24f/0x470 [i915]()
> >>[501906.623568] unhandled error in i915_gem_fault: -7
> >>[501906.825115] Call Trace:
> >>[501906.830322]  [<ffffffff8178ffcc>] dump_stack+0x45/0x57
> >>[501906.838589]  [<ffffffff810759aa>] warn_slowpath_common+0x8a/0xc0
> >>[501906.847846]  [<ffffffff81075a26>] warn_slowpath_fmt+0x46/0x50
> >>[501906.856776]  [<ffffffffc079591f>] i915_gem_fault+0x24f/0x470 [i915]
> >>[501906.866276]  [<ffffffff8119e11d>] __do_fault+0x3d/0xa0
> >>[501906.874464]  [<ffffffff81068cc0>] ? pte_alloc_one+0x30/0x50
> >>[501906.883169]  [<ffffffff811a26a7>] handle_mm_fault+0xe27/0x1810
> >>[501906.892202]  [<ffffffff81306e8a>] ? security_mmap_file+0xca/0xe0
> >>[501906.900389]  [<ffffffff811fb6ad>] ? do_vfs_ioctl+0x2cd/0x4b0
> >>[501906.908143]  [<ffffffff81063ada>] __do_page_fault+0x19a/0x430
> >>[501906.916024]  [<ffffffff81063d92>] do_page_fault+0x22/0x30
> >>[501906.923532]  [<ffffffff81799248>] page_fault+0x28/0x30
> >>
> >>RFC about the error code that should be returned by i915_gem_fault.
> >
> >There are two fixes here, change E2BIG to ENOSPC. The differentiate is
> >painful to all consumers (and missing in userspace).
> 
> I'll change the return code in gem_object_bind_to_vm instead.
> 
> >
> >The second is that gem_fault was supposed to be fixed to handle it and
> >sigbus is the legimate error for that case (not enomem/sigsegv).
> 
> These subtests use objs tiling x/y, so the driver won't use partial views.
> 
> Should it be better if i915_gem_mmap_gtt checks for this (obj size
> bigger than apperture and some tiling mode), and return an error at
> that point? Instead of waiting for the sigbus error in gem_fault.

No, the driver is meant to use a partial fenced view in this case. The
feature was merged very early and is still very, very incomplete.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre