[PATCH v3 08/17] drm/i915: Call i915_gem_evict_vm in vm_fault_gtt to prevent new ENOSPC errors

[Matthew Auld]

On Thu, 16 Dec 2021 at 14:28, Maarten Lankhorst
<maarten.lankhorst at linux.intel.com> wrote:
>
> Now that we cannot unbind kill the currently locked object directly

"unbind kill"

> 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 at linux.intel.com>

Maybe mention that this only in preparation for some future patches,
once the actual eviction is trylock and evict_for_vm can also handle
shared dma-resv? At this point in the series we shouldn't expect to
hit -ENOSPC, right?

> ---
>  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 af81d6c3332a..00cd9642669a 100644
> --- a/drivers/gpu/drm/i915/gem/i915_gem_mman.c
> +++ b/drivers/gpu/drm/i915/gem/i915_gem_mman.c
> @@ -358,8 +358,22 @@ 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);
> +                               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));

Looks like this is being triggered in CI, I assume because the trylock
could easily fail, due to contention? Is this expected for now? Do we
keep the WARN and track it as a known issue?

>         }
>         if (IS_ERR(vma)) {
>                 ret = PTR_ERR(vma);
> --
> 2.34.1
>