[Intel-gfx] [PATCH] drm/i915: Release i915_gem_context from a worker

[Maarten Lankhorst]

Op 14-08-2021 om 12:43 schreef Daniel Vetter:
> The only reason for this really is the i915_gem_engines->fence
> callback engines_notify(), which exists purely as a fairly funky
> reference counting scheme for that. Otherwise all other callers are
> from process context, and generally fairly benign locking context.
>
> Unfortunately untangling that requires some major surgery, and we have
> a few i915_gem_context reference counting bugs that need fixing, and
> they blow in the current hardirq calling context, so we need a
> stop-gap measure.
>
> Put a FIXME comment in when this should be removable again.
>
> v2: Fix mock_context(), noticed by intel-gfx-ci.
>
> Signed-off-by: Daniel Vetter <daniel.vetter at intel.com>
> Cc: Jon Bloomfield <jon.bloomfield at intel.com>
> Cc: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: Maarten Lankhorst <maarten.lankhorst at linux.intel.com>
> Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
> Cc: Daniel Vetter <daniel.vetter at ffwll.ch>
> Cc: "Thomas Hellström" <thomas.hellstrom at linux.intel.com>
> Cc: Matthew Auld <matthew.auld at intel.com>
> Cc: Lionel Landwerlin <lionel.g.landwerlin at intel.com>
> Cc: Dave Airlie <airlied at redhat.com>
> Cc: Jason Ekstrand <jason at jlekstrand.net>
> ---
>  drivers/gpu/drm/i915/gem/i915_gem_context.c       | 13 +++++++++++--
>  drivers/gpu/drm/i915/gem/i915_gem_context_types.h | 12 ++++++++++++
>  drivers/gpu/drm/i915/gem/selftests/mock_context.c |  1 +
>  3 files changed, 24 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context.c b/drivers/gpu/drm/i915/gem/i915_gem_context.c
> index fd169cf2f75a..051bc357ff65 100644
> --- a/drivers/gpu/drm/i915/gem/i915_gem_context.c
> +++ b/drivers/gpu/drm/i915/gem/i915_gem_context.c
> @@ -986,9 +986,10 @@ static struct i915_gem_engines *user_engines(struct i915_gem_context *ctx,
>  	return err;
>  }
>  
> -void i915_gem_context_release(struct kref *ref)
> +static void i915_gem_context_release_work(struct work_struct *work)
>  {
> -	struct i915_gem_context *ctx = container_of(ref, typeof(*ctx), ref);
> +	struct i915_gem_context *ctx = container_of(work, typeof(*ctx),
> +						    release_work);
>  
>  	trace_i915_context_free(ctx);
>  	GEM_BUG_ON(!i915_gem_context_is_closed(ctx));
> @@ -1002,6 +1003,13 @@ void i915_gem_context_release(struct kref *ref)
>  	kfree_rcu(ctx, rcu);
>  }
>  
> +void i915_gem_context_release(struct kref *ref)
> +{
> +	struct i915_gem_context *ctx = container_of(ref, typeof(*ctx), ref);
> +
> +	queue_work(ctx->i915->wq, &ctx->release_work);
> +}
> +
>  static inline struct i915_gem_engines *
>  __context_engines_static(const struct i915_gem_context *ctx)
>  {
> @@ -1303,6 +1311,7 @@ i915_gem_create_context(struct drm_i915_private *i915,
>  	ctx->sched = pc->sched;
>  	mutex_init(&ctx->mutex);
>  	INIT_LIST_HEAD(&ctx->link);
> +	INIT_WORK(&ctx->release_work, i915_gem_context_release_work);
>  
>  	spin_lock_init(&ctx->stale.lock);
>  	INIT_LIST_HEAD(&ctx->stale.engines);
> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context_types.h b/drivers/gpu/drm/i915/gem/i915_gem_context_types.h
> index 94c03a97cb77..0c38789bd4a8 100644
> --- a/drivers/gpu/drm/i915/gem/i915_gem_context_types.h
> +++ b/drivers/gpu/drm/i915/gem/i915_gem_context_types.h
> @@ -288,6 +288,18 @@ struct i915_gem_context {
>  	 */
>  	struct kref ref;
>  
> +	/**
> +	 * @release_work:
> +	 *
> +	 * Work item for deferred cleanup, since i915_gem_context_put() tends to
> +	 * be called from hardirq context.
> +	 *
> +	 * FIXME: The only real reason for this is &i915_gem_engines.fence, all
> +	 * other callers are from process context and need at most some mild
> +	 * shuffling to pull the i915_gem_context_put() call out of a spinlock.
> +	 */
> +	struct work_struct release_work;
> +
>  	/**
>  	 * @rcu: rcu_head for deferred freeing.
>  	 */
> diff --git a/drivers/gpu/drm/i915/gem/selftests/mock_context.c b/drivers/gpu/drm/i915/gem/selftests/mock_context.c
> index fee070df1c97..067d68a6fe4c 100644
> --- a/drivers/gpu/drm/i915/gem/selftests/mock_context.c
> +++ b/drivers/gpu/drm/i915/gem/selftests/mock_context.c
> @@ -23,6 +23,7 @@ mock_context(struct drm_i915_private *i915,
>  	kref_init(&ctx->ref);
>  	INIT_LIST_HEAD(&ctx->link);
>  	ctx->i915 = i915;
> +	INIT_WORK(&ctx->release_work, i915_gem_context_release_work);
>  
>  	mutex_init(&ctx->mutex);
>  

----
Is the workqueue really needed? I'm not sure you could still race in drm_syncobj_free when refcount is zero, so in that case removing locking from _release would work as well as a workqueue.

Something like below would keep the drm_sync_obj_put hardirq safe.

I assume when freeing, the  cb list is supposed to be empty, so I added a WARN_ON just to be sure, otherwise we should just tear down the list without locking too.

This should be a better alternative for patch 1.
----8<-------
diff --git a/drivers/gpu/drm/drm_syncobj.c b/drivers/gpu/drm/drm_syncobj.c
index c9a9d74f338c..9d561decd97e 100644
--- a/drivers/gpu/drm/drm_syncobj.c
+++ b/drivers/gpu/drm/drm_syncobj.c
@@ -462,7 +462,13 @@ void drm_syncobj_free(struct kref *kref)
 	struct drm_syncobj *syncobj = container_of(kref,
 						   struct drm_syncobj,
 						   refcount);
-	drm_syncobj_replace_fence(syncobj, NULL);
+	struct dma_fence *old_fence;
+
+	old_fence = rcu_dereference_protected(syncobj->fence, !kref_read(&syncobj->refcount));
+	dma_fence_put(old_fence);
+
+	WARN_ON(!list_empty(&syncobj->cb_list));
+
 	kfree(syncobj);
 }
 EXPORT_SYMBOL(drm_syncobj_free);