[Intel-gfx] [BrownBag] drm/i915/gtt: Neuter the deferred unbind callback from gen6_ppgtt_cleanup

[Tvrtko Ursulin]

On 24/05/2019 09:51, Tvrtko Ursulin wrote:
> 
> On 24/05/2019 09:36, Chris Wilson wrote:
>> Quoting Tvrtko Ursulin (2019-05-24 09:31:45)
>>>
>>> On 24/05/2019 09:29, Chris Wilson wrote:
>>>> Quoting Tvrtko Ursulin (2019-05-24 09:23:40)
>>>>>
>>>>> On 24/05/2019 09:17, Chris Wilson wrote:
>>>>>> Quoting Tvrtko Ursulin (2019-05-24 09:13:14)
>>>>>>>
>>>>>>> On 24/05/2019 07:45, Chris Wilson wrote:
>>>>>>>> Having deferred the vma destruction to a worker where we can 
>>>>>>>> acquire the
>>>>>>>> struct_mutex, we have to avoid chasing back into the now destroyed
>>>>>>>> ppgtt. The pd_vma is special in having a custom unbind function 
>>>>>>>> to scan
>>>>>>>> for unused pages despite the VMA itself being notionally part of 
>>>>>>>> the
>>>>>>>> GGTT. As such, we need to disable that callback to avoid a
>>>>>>>> use-after-free.
>>>>>>>>
>>>>>>>> This unfortunately blew up so early during boot that CI declared 
>>>>>>>> the
>>>>>>>> machine unreachable as opposed to being the major failure it 
>>>>>>>> was. Oops.
>>>>>>>>
>>>>>>>> Fixes: d3622099c76f ("drm/i915/gtt: Always acquire struct_mutex 
>>>>>>>> for gen6_ppgtt_cleanup")
>>>>>>>> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
>>>>>>>> Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>>>>>> Cc: Tomi Sarvela <tomi.p.sarvela at intel.com>
>>>>>>>> ---
>>>>>>>>      drivers/gpu/drm/i915/i915_gem_gtt.c | 28 
>>>>>>>> ++++++++++++++++++++++++++++
>>>>>>>>      1 file changed, 28 insertions(+)
>>>>>>>>
>>>>>>>> diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c 
>>>>>>>> b/drivers/gpu/drm/i915/i915_gem_gtt.c
>>>>>>>> index 8d8a4b0ad4d9..266baa11df64 100644
>>>>>>>> --- a/drivers/gpu/drm/i915/i915_gem_gtt.c
>>>>>>>> +++ b/drivers/gpu/drm/i915/i915_gem_gtt.c
>>>>>>>> @@ -1847,6 +1847,33 @@ static void 
>>>>>>>> gen6_ppgtt_cleanup_work(struct work_struct *wrk)
>>>>>>>>          kfree(work);
>>>>>>>>      }
>>>>>>>> +static int nop_set_pages(struct i915_vma *vma)
>>>>>>>> +{
>>>>>>>> +     return -ENODEV;
>>>>>>>> +}
>>>>>>>> +
>>>>>>>> +static void nop_clear_pages(struct i915_vma *vma)
>>>>>>>> +{
>>>>>>>> +}
>>>>>>>> +
>>>>>>>> +static int nop_bind(struct i915_vma *vma,
>>>>>>>> +                 enum i915_cache_level cache_level,
>>>>>>>> +                 u32 unused)
>>>>>>>> +{
>>>>>>>> +     return -ENODEV;
>>>>>>>> +}
>>>>>>>> +
>>>>>>>> +static void nop_unbind(struct i915_vma *vma)
>>>>>>>> +{
>>>>>>>> +}
>>>>>>>> +
>>>>>>>> +static const struct i915_vma_ops nop_vma_ops = {
>>>>>>>> +     .set_pages = nop_set_pages,
>>>>>>>> +     .clear_pages = nop_clear_pages,
>>>>>>>> +     .bind_vma = nop_bind,
>>>>>>>> +     .unbind_vma = nop_unbind,
>>>>>>>> +};
>>>>>>>> +
>>>>>>>>      static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
>>>>>>>>      {
>>>>>>>>          struct gen6_hw_ppgtt *ppgtt = 
>>>>>>>> to_gen6_ppgtt(i915_vm_to_ppgtt(vm));
>>>>>>>> @@ -1855,6 +1882,7 @@ static void gen6_ppgtt_cleanup(struct 
>>>>>>>> i915_address_space *vm)
>>>>>>>>          /* FIXME remove the struct_mutex to bring the locking 
>>>>>>>> under control */
>>>>>>>>          INIT_WORK(&work->base, gen6_ppgtt_cleanup_work);
>>>>>>>>          work->vma = ppgtt->vma;
>>>>>>>> +     work->vma->ops = &nop_vma_ops;
>>>>>>>
>>>>>>> Could we use some asserts before overriding the vma ops? Like
>>>>>>> GEM_BUG_ON(vma->pages)? And something for still bound?
>>>>>>
>>>>>> It technically still is bound as it is in the GGTT but currently
>>>>>> unpinned -- that will be checked on destroy, it's just we also get an
>>>>>> unbind callback. vma->pages doesn't exist for this (set to ERR_PTR).
>>>>>
>>>>> If we are getting the unbind callback and we nop-ed it, who will
>>>>> actually do it's job?
>>>>
>>>> The callback is just a hook for us to prune within the ppgtt.
>>>> It still is removed from GGTT by i915_vma_unbind().
>>>
>>> So it needs GEM_BUG_ON(ppgtt->scan_for_unused_pt) before overriding the
>>> unbind?
>>
>> No. They get freed by the cleanup itself. The scan is just an
>> opportunistic prune if either the context/mm is evicted but still alive.
> 
> Then the same assert in gen6_ppgtt_cleanup_work? :)

Okay ppgtt is gone so can't do it.. annoying.. Cleanup seems to support 
your claims but I think we need a BFC (big fat comment) above the vma 
ops override to explains this. With that:

Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>

Regards,

Tvrtko