[Intel-gfx] [PATCH] drm/i915: vma/ppgtt lifetime rules

[Ben Widawsky]

On Tue, Jul 29, 2014 at 11:08:05AM +0100, Michel Thierry wrote:
> VMAs should take a reference of the address space they use.
> 
> Now, when the fd is closed, it will release the ref that the context was
> holding, but it will still be referenced by any vmas that are still
> active.
> 
> ppgtt_release() should then only be called when the last thing referencing
> it releases the ref, and it can just call the base cleanup and free the
> ppgtt.
> 
> Signed-off-by: Michel Thierry <michel.thierry at intel.com>
> ---
>  drivers/gpu/drm/i915/i915_drv.h         |  2 ++
>  drivers/gpu/drm/i915/i915_gem.c         |  8 ++++++++
>  drivers/gpu/drm/i915/i915_gem_context.c | 23 +++--------------------
>  drivers/gpu/drm/i915/i915_gem_gtt.c     |  5 +++++
>  4 files changed, 18 insertions(+), 20 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> index 2acc03f..a879a93 100644
> --- a/drivers/gpu/drm/i915/i915_drv.h
> +++ b/drivers/gpu/drm/i915/i915_drv.h
> @@ -2495,7 +2495,9 @@ void i915_gem_object_ggtt_unpin(struct drm_i915_gem_object *obj);
>  
>  /* i915_gem_context.c */
>  #define ctx_to_ppgtt(ctx) container_of((ctx)->vm, struct i915_hw_ppgtt, base)
> +#define vm_to_ppgtt(vm) container_of(vm, struct i915_hw_ppgtt, base)
>  int __must_check i915_gem_context_init(struct drm_device *dev);
> +void ppgtt_release(struct kref *kref);
>  void i915_gem_context_fini(struct drm_device *dev);
>  void i915_gem_context_reset(struct drm_device *dev);
>  int i915_gem_context_open(struct drm_device *dev, struct drm_file *file);
> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> index dcd8d7b..25a32b9 100644
> --- a/drivers/gpu/drm/i915/i915_gem.c
> +++ b/drivers/gpu/drm/i915/i915_gem.c
> @@ -4499,12 +4499,20 @@ struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj,
>  
>  void i915_gem_vma_destroy(struct i915_vma *vma)
>  {
> +	struct i915_address_space *vm = NULL;
> +	struct i915_hw_ppgtt *ppgtt = NULL;
>  	WARN_ON(vma->node.allocated);
>  
>  	/* Keep the vma as a placeholder in the execbuffer reservation lists */
>  	if (!list_empty(&vma->exec_list))
>  		return;
>  
> +	vm = vma->vm;
> +	ppgtt = vm_to_ppgtt(vm);
> +
> +	if (ppgtt)
> +		kref_put(&ppgtt->ref, ppgtt_release);
> +
>  	list_del(&vma->vma_link);
>  
>  	kfree(vma);
> diff --git a/drivers/gpu/drm/i915/i915_gem_context.c b/drivers/gpu/drm/i915/i915_gem_context.c
> index 5b5af6c..59272f9 100644
> --- a/drivers/gpu/drm/i915/i915_gem_context.c
> +++ b/drivers/gpu/drm/i915/i915_gem_context.c
> @@ -108,30 +108,13 @@ static void do_ppgtt_cleanup(struct i915_hw_ppgtt *ppgtt)
>  		return;
>  	}
>  
> -	/*
> -	 * Make sure vmas are unbound before we take down the drm_mm
> -	 *
> -	 * FIXME: Proper refcounting should take care of this, this shouldn't be
> -	 * needed at all.
> -	 */
> -	if (!list_empty(&vm->active_list)) {
> -		struct i915_vma *vma;
> -
> -		list_for_each_entry(vma, &vm->active_list, mm_list)
> -			if (WARN_ON(list_empty(&vma->vma_link) ||
> -				    list_is_singular(&vma->vma_link)))
> -				break;
> -
> -		i915_gem_evict_vm(&ppgtt->base, true);
> -	} else {
> -		i915_gem_retire_requests(dev);
> -		i915_gem_evict_vm(&ppgtt->base, false);
> -	}
> +	/* vmas should already be unbound */
> +	WARN_ON(!list_empty(&vm->active_list));
>  
>  	ppgtt->base.cleanup(&ppgtt->base);
>  }
>  
> -static void ppgtt_release(struct kref *kref)
> +void ppgtt_release(struct kref *kref)
>  {
>  	struct i915_hw_ppgtt *ppgtt =
>  		container_of(kref, struct i915_hw_ppgtt, ref);
> diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c b/drivers/gpu/drm/i915/i915_gem_gtt.c
> index 1411613..90c3d0f 100644
> --- a/drivers/gpu/drm/i915/i915_gem_gtt.c
> +++ b/drivers/gpu/drm/i915/i915_gem_gtt.c
> @@ -2159,10 +2159,15 @@ i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
>  				  struct i915_address_space *vm)
>  {
>  	struct i915_vma *vma;
> +	struct i915_hw_ppgtt *ppgtt = NULL;
>  
>  	vma = i915_gem_obj_to_vma(obj, vm);
>  	if (!vma)
>  		vma = __i915_gem_vma_create(obj, vm);
>  
> +	ppgtt = vm_to_ppgtt(vm);
> +	if (ppgtt)
> +		kref_get(&ppgtt->ref);
> +
>  	return vma;
>  }

Seems like a reasonable fix to me, and it's simpler than Chris' request
to ref when active. Logically his makes more sense, but it's not quite
as simple to review. Actually active isn't the right thing to track IMO,
bound is what you want to track. I think if anything, do that one as a
patch on top, it does address the problem mentioned below. I also wonder
given the recent reset lolz how this behaves there; and I'd make sure to
thoroughly test the various reset/hang conditions.

Also, this puts a hard limit on the maximum number of VMAs in an address
space based on sizeof(atomic_t). (unless you do what I said above) I
doubt it's a problem in the near future, but it should be documented
somewhere and I hope kref will explode loudly if you overflow it.

For future safety, I'd recommend dropping vm->ctx at this point. It
shouldn't be hard at all to get rid of (quick grep shows only debugfs).
-- 
Ben Widawsky, Intel Open Source Technology Center