[Intel-gfx] [PATCH v3 6/6] freezer, sched: Rewrite core freezer logic

[Peter Zijlstra]

On Tue, Oct 25, 2022 at 07:52:07AM +0300, Ville Syrjälä wrote:
> On Fri, Oct 21, 2022 at 08:22:41PM +0300, Ville Syrjälä wrote:
> > On Mon, Aug 22, 2022 at 01:18:22PM +0200, Peter Zijlstra wrote:
> > > +#ifdef CONFIG_LOCKDEP
> > > +	/*
> > > +	 * It's dangerous to freeze with locks held; there be dragons there.
> > > +	 */
> > > +	if (!(state & __TASK_FREEZABLE_UNSAFE))
> > > +		WARN_ON_ONCE(debug_locks && p->lockdep_depth);
> > > +#endif
> > 
> > We now seem to be hitting this sporadically in the intel gfx CI.
> > 
> > I've spotted it on two machines so far:
> > https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_12270/shard-tglb7/igt@gem_ctx_isolation@preservation-s3@vcs0.html
> > https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_109950v1/shard-snb5/igt@kms_flip@flip-vs-suspend-interruptible@a-vga1.html
> 
> Sadly no luck in reproducing this locally so far. In the meantime
> I added the following patch into our topic/core-for-CI branch in
> the hopes of CI stumbling on it again and dumping a bit more data:
> 
> --- a/kernel/freezer.c
> +++ b/kernel/freezer.c
> @@ -125,8 +125,16 @@ static int __set_task_frozen(struct task_struct *p, void *arg)
>  	/*
>  	 * It's dangerous to freeze with locks held; there be dragons there.
>  	 */
> -	if (!(state & __TASK_FREEZABLE_UNSAFE))
> -		WARN_ON_ONCE(debug_locks && p->lockdep_depth);
> +	if (!(state & __TASK_FREEZABLE_UNSAFE)) {
> +		static bool warned = false;
> +
> +		if (!warned && debug_locks && p->lockdep_depth) {
> +			debug_show_held_locks(p);
> +			WARN(1, "%s/%d holding locks while freezing\n",
> +			     p->comm, task_pid_nr(p));
> +			warned = true;
> +		}
> +	}
>  #endif
>  
>  	WRITE_ONCE(p->__state, TASK_FROZEN);

That seems reasonable. But note that this constraint isn't new; the
previous freezer had much the same constraint but perhaps it wasn't
triggered for mysterious raisins. see the previous
try_to_freeze_unsafe() function.