linux-kernel - Re: [PATCH drm-misc-next v3] drm/sched: implement dynamic job-flow control

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [day] [month] [year] [list]

Message-ID: <20231027102237.0cdb85af@collabora.com>
Date:   Fri, 27 Oct 2023 10:22:37 +0200
From:   Boris Brezillon <boris.brezillon@...labora.com>
To:     Christian König <christian.koenig@....com>
Cc:     Danilo Krummrich <dakr@...hat.com>, airlied@...il.com,
        daniel@...ll.ch, matthew.brost@...el.com, faith@...strand.net,
        luben.tuikov@....com, dri-devel@...ts.freedesktop.org,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH drm-misc-next v3] drm/sched: implement dynamic job-flow
 control

On Fri, 27 Oct 2023 09:44:13 +0200
Christian König <christian.koenig@....com> wrote:

> Am 27.10.23 um 09:39 schrieb Boris Brezillon:
> > On Fri, 27 Oct 2023 09:35:01 +0200
> > Christian König<christian.koenig@....com>  wrote:
> >  
> >> Am 27.10.23 um 09:32 schrieb Boris Brezillon:  
> >>> On Fri, 27 Oct 2023 09:22:12 +0200
> >>> Christian König<christian.koenig@....com>  wrote:
> >>>     
> >>>>> +
> >>>>> +	/**
> >>>>> +	 * @update_job_credits: Called once the scheduler is considering this
> >>>>> +	 * job for execution.
> >>>>> +	 *
> >>>>> +	 * Drivers may use this to update the job's submission credits, which is
> >>>>> +	 * useful to e.g. deduct the number of native fences which have been
> >>>>> +	 * signaled meanwhile.
> >>>>> +	 *
> >>>>> +	 * The callback must either return the new number of submission credits
> >>>>> +	 * for the given job, or zero if no update is required.
> >>>>> +	 *
> >>>>> +	 * This callback is optional.
> >>>>> +	 */
> >>>>> +	u32 (*update_job_credits)(struct drm_sched_job *sched_job);  
> >>>> Why do we need an extra callback for this?
> >>>>
> >>>> Just document that prepare_job() is allowed to reduce the number of
> >>>> credits the job might need.
> >>> ->prepare_job() is called only once if the returned fence is NULL, but  
> >>> we need this credit-update to happen every time a job is considered for
> >>> execution by the scheduler.  
> >> But the job is only considered for execution once. How do you see that
> >> this is called multiple times?  
> > Nope, it's not. If drm_sched_can_queue() returns false, the scheduler
> > will go look for another entity that has a job ready for execution, and
> > get back to this entity later, and test drm_sched_can_queue() again.
> > Basically, any time drm_sched_can_queue() is called, the job credits
> > update should happen, so we have an accurate view of how many credits
> > this job needs.  
> 
> Well, that is the handling which I already rejected because it creates 
> unfairness between processes. When you consider the credits needed 
> *before* scheduling jobs with a lower credit count are always preferred 
> over jobs with a higher credit count.

My bad, it doesn't pick another entity when an entity with a
ready job that doesn't fit the queue is found, it just bails out from
drm_sched_rq_select_entity_rr() and returns NULL (AKA: no ready entity
found). But we still want to update the job credits before checking if
the job fits or not (next time this entity is tested).

> What you can do is to look at the credits of a job *after* it was picked 
> up for scheduling so that you can scheduler more jobs.

Sure, but then you might further delay your job if something made it
smaller (ie. native fences got signaled) between ->prepare_job() and
drm_sched_can_queue(). And any new drm_sched_can_queue() test would
just see the old credits value.

Out of curiosity, what are you worried about with this optional
->update_job_credits() call in the drm_sched_can_queue() path? Is the
if (sched->update_job_credits) overhead considered too high for drivers
that don't need it?