Re: [PATCH] drm/i915: Release i915_gem_context from a worker

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Tue, Aug 31, 2021 at 11:38:27AM +0200, Maarten Lankhorst wrote:
> 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@xxxxxxxxx>
> > Cc: Jon Bloomfield <jon.bloomfield@xxxxxxxxx>
> > Cc: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> > Cc: Maarten Lankhorst <maarten.lankhorst@xxxxxxxxxxxxxxx>
> > Cc: Joonas Lahtinen <joonas.lahtinen@xxxxxxxxxxxxxxx>
> > Cc: Daniel Vetter <daniel.vetter@xxxxxxxx>
> > Cc: "Thomas Hellström" <thomas.hellstrom@xxxxxxxxxxxxxxx>
> > Cc: Matthew Auld <matthew.auld@xxxxxxxxx>
> > Cc: Lionel Landwerlin <lionel.g.landwerlin@xxxxxxxxx>
> > Cc: Dave Airlie <airlied@xxxxxxxxxx>
> > Cc: Jason Ekstrand <jason@xxxxxxxxxxxxxx>
> > ---
> >  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.

This isn't enough, because the problem isn't just the syncobj put. It's
also the i915_vm_put, and if we dercuify the intel_context stuff too, then
there will be more intel_context_put on top.

So we really need the worker here I think. Trying to make every _unpin() and
_put() work from hardirq context with clever locking tricks is why the
current code is so incomprehensible.

Also vms are rare enough that we really don't care about some
overhead/delay here.
-Daniel

> ----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);
> 
> 

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch



[Index of Archives]     [AMD Graphics]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux