Re: [Intel-gfx] [PATCH v2 11/16] drm/i915: Call i915_gem_evict_vm in vm_fault_gtt to prevent new ENOSPC errors

[Maarten Lankhorst <maarten.lankhorst@linux.intel.com>]

On 09-12-2021 13:17, Matthew Auld wrote:
> On Mon, 29 Nov 2021 at 13:58, Maarten Lankhorst
> <maarten.lankhorst@linux.intel.com> wrote:
>> Now that we cannot unbind kill the currently locked object directly
> Can this be reworded slightly? Not sure what is meant by "unbind kill" here.
Oops, the word 'kill' doesn't belong here.
>> because we're removing short term pinning, we may have to unbind the
>> object from gtt manually, using a i915_gem_evict_vm() call.
>>
>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
>> ---
>>  drivers/gpu/drm/i915/gem/i915_gem_mman.c | 18 ++++++++++++++++--
>>  1 file changed, 16 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_mman.c b/drivers/gpu/drm/i915/gem/i915_gem_mman.c
>> index 65fc6ff5f59d..6d557bb9926f 100644
>> --- a/drivers/gpu/drm/i915/gem/i915_gem_mman.c
>> +++ b/drivers/gpu/drm/i915/gem/i915_gem_mman.c
>> @@ -357,8 +357,22 @@ static vm_fault_t vm_fault_gtt(struct vm_fault *vmf)
>>                         vma = i915_gem_object_ggtt_pin_ww(obj, &ww, &view, 0, 0, flags);
>>                 }
>>
>> -               /* The entire mappable GGTT is pinned? Unexpected! */
>> -               GEM_BUG_ON(vma == ERR_PTR(-ENOSPC));
>> +               /*
>> +                * The entire mappable GGTT is pinned? Unexpected!
>> +                * Try to evict the object we locked too, as normally we skip it
>> +                * due to lack of short term pinning inside execbuf.
>> +                */
>> +               if (vma == ERR_PTR(-ENOSPC)) {
>> +                       ret = mutex_lock_interruptible(&ggtt->vm.mutex);
>> +                       if (!ret) {
>> +                               ret = i915_gem_evict_vm(&ggtt->vm, &ww);
> Would it make sense to pass an extra flag for the above ggtt_pin(maybe
> PIN_EVICT_SHARED)? Such that evict_for_something can handle the
> already locked object and then also any vma sharing the same dma-resv
> object here? Or at least trying to nuke the entire vm, just for the
> mappable portion seems maybe overkill? Or perhaps we never expect to
> hit this in the real world?
>
> Reviewed-by: Matthew Auld <matthew.auld@intel.com>

Yeah, effect would be the same though. When fully reworking eviction and vm locks, it might be better to do so though.

>
>> +                               mutex_unlock(&ggtt->vm.mutex);
>> +                       }
>> +                       if (ret)
>> +                               goto err_reset;
>> +                       vma = i915_gem_object_ggtt_pin_ww(obj, &ww, &view, 0, 0, flags);
>> +               }
>> +               GEM_WARN_ON(vma == ERR_PTR(-ENOSPC));
>>         }
>>         if (IS_ERR(vma)) {
>>                 ret = PTR_ERR(vma);
>> --
>> 2.34.0
>>