[Intel-gfx] [RFC PATCH 36/97] drm/i915/guc: Add non blocking CTB send function
Daniel Vetter
daniel at ffwll.ch
Thu Jun 10 15:27:48 UTC 2021
On Wed, Jun 09, 2021 at 04:10:23PM -0700, Matthew Brost wrote:
> On Tue, Jun 08, 2021 at 10:46:15AM +0200, Daniel Vetter wrote:
> > On Tue, Jun 8, 2021 at 10:39 AM Tvrtko Ursulin
> > <tvrtko.ursulin at linux.intel.com> wrote:
> > >
> > >
> > > On 07/06/2021 18:31, Matthew Brost wrote:
> > > > On Thu, May 27, 2021 at 04:11:50PM +0100, Tvrtko Ursulin wrote:
> > > >>
> > > >> On 27/05/2021 15:35, Matthew Brost wrote:
> > > >>> On Thu, May 27, 2021 at 11:02:24AM +0100, Tvrtko Ursulin wrote:
> > > >>>>
> > > >>>> On 26/05/2021 19:10, Matthew Brost wrote:
> > > >>>>
> > > >>>> [snip]
> > > >>>>
> > > >>>>>>>>> +static int ct_send_nb(struct intel_guc_ct *ct,
> > > >>>>>>>>> + const u32 *action,
> > > >>>>>>>>> + u32 len,
> > > >>>>>>>>> + u32 flags)
> > > >>>>>>>>> +{
> > > >>>>>>>>> + struct intel_guc_ct_buffer *ctb = &ct->ctbs.send;
> > > >>>>>>>>> + unsigned long spin_flags;
> > > >>>>>>>>> + u32 fence;
> > > >>>>>>>>> + int ret;
> > > >>>>>>>>> +
> > > >>>>>>>>> + spin_lock_irqsave(&ctb->lock, spin_flags);
> > > >>>>>>>>> +
> > > >>>>>>>>> + ret = ctb_has_room(ctb, len + 1);
> > > >>>>>>>>> + if (unlikely(ret))
> > > >>>>>>>>> + goto out;
> > > >>>>>>>>> +
> > > >>>>>>>>> + fence = ct_get_next_fence(ct);
> > > >>>>>>>>> + ret = ct_write(ct, action, len, fence, flags);
> > > >>>>>>>>> + if (unlikely(ret))
> > > >>>>>>>>> + goto out;
> > > >>>>>>>>> +
> > > >>>>>>>>> + intel_guc_notify(ct_to_guc(ct));
> > > >>>>>>>>> +
> > > >>>>>>>>> +out:
> > > >>>>>>>>> + spin_unlock_irqrestore(&ctb->lock, spin_flags);
> > > >>>>>>>>> +
> > > >>>>>>>>> + return ret;
> > > >>>>>>>>> +}
> > > >>>>>>>>> +
> > > >>>>>>>>> static int ct_send(struct intel_guc_ct *ct,
> > > >>>>>>>>> const u32 *action,
> > > >>>>>>>>> u32 len,
> > > >>>>>>>>> @@ -473,6 +541,7 @@ static int ct_send(struct intel_guc_ct *ct,
> > > >>>>>>>>> u32 response_buf_size,
> > > >>>>>>>>> u32 *status)
> > > >>>>>>>>> {
> > > >>>>>>>>> + struct intel_guc_ct_buffer *ctb = &ct->ctbs.send;
> > > >>>>>>>>> struct ct_request request;
> > > >>>>>>>>> unsigned long flags;
> > > >>>>>>>>> u32 fence;
> > > >>>>>>>>> @@ -482,8 +551,20 @@ static int ct_send(struct intel_guc_ct *ct,
> > > >>>>>>>>> GEM_BUG_ON(!len);
> > > >>>>>>>>> GEM_BUG_ON(len & ~GUC_CT_MSG_LEN_MASK);
> > > >>>>>>>>> GEM_BUG_ON(!response_buf && response_buf_size);
> > > >>>>>>>>> + might_sleep();
> > > >>>>>>>>
> > > >>>>>>>> Sleep is just cond_resched below or there is more?
> > > >>>>>>>>
> > > >>>>>>>
> > > >>>>>>> Yes, the cond_resched.
> > > >>>>>>>
> > > >>>>>>>>> + /*
> > > >>>>>>>>> + * We use a lazy spin wait loop here as we believe that if the CT
> > > >>>>>>>>> + * buffers are sized correctly the flow control condition should be
> > > >>>>>>>>> + * rare.
> > > >>>>>>>>> + */
> > > >>>>>>>>> +retry:
> > > >>>>>>>>> spin_lock_irqsave(&ct->ctbs.send.lock, flags);
> > > >>>>>>>>> + if (unlikely(!ctb_has_room(ctb, len + 1))) {
> > > >>>>>>>>> + spin_unlock_irqrestore(&ct->ctbs.send.lock, flags);
> > > >>>>>>>>> + cond_resched();
> > > >>>>>>>>> + goto retry;
> > > >>>>>>>>> + }
> > > >>>>>>>>
> > > >>>>>>>> If this patch is about adding a non-blocking send function, and below we can
> > > >>>>>>>> see that it creates a fork:
> > > >>>>>>>>
> > > >>>>>>>> intel_guc_ct_send:
> > > >>>>>>>> ...
> > > >>>>>>>> if (flags & INTEL_GUC_SEND_NB)
> > > >>>>>>>> return ct_send_nb(ct, action, len, flags);
> > > >>>>>>>>
> > > >>>>>>>> ret = ct_send(ct, action, len, response_buf, response_buf_size, &status);
> > > >>>>>>>>
> > > >>>>>>>> Then why is there a change in ct_send here, which is not the new
> > > >>>>>>>> non-blocking path?
> > > >>>>>>>>
> > > >>>>>>>
> > > >>>>>>> There is not a change to ct_send(), just to intel_guc_ct_send.
> > > >>>>>>
> > > >>>>>> I was doing by the diff which says:
> > > >>>>>>
> > > >>>>>> static int ct_send(struct intel_guc_ct *ct,
> > > >>>>>> const u32 *action,
> > > >>>>>> u32 len,
> > > >>>>>> @@ -473,6 +541,7 @@ static int ct_send(struct intel_guc_ct *ct,
> > > >>>>>> u32 response_buf_size,
> > > >>>>>> u32 *status)
> > > >>>>>> {
> > > >>>>>> + struct intel_guc_ct_buffer *ctb = &ct->ctbs.send;
> > > >>>>>> struct ct_request request;
> > > >>>>>> unsigned long flags;
> > > >>>>>> u32 fence;
> > > >>>>>> @@ -482,8 +551,20 @@ static int ct_send(struct intel_guc_ct *ct,
> > > >>>>>> GEM_BUG_ON(!len);
> > > >>>>>> GEM_BUG_ON(len & ~GUC_CT_MSG_LEN_MASK);
> > > >>>>>> GEM_BUG_ON(!response_buf && response_buf_size);
> > > >>>>>> + might_sleep();
> > > >>>>>> + /*
> > > >>>>>> + * We use a lazy spin wait loop here as we believe that if the CT
> > > >>>>>> + * buffers are sized correctly the flow control condition should be
> > > >>>>>> + * rare.
> > > >>>>>> + */
> > > >>>>>> +retry:
> > > >>>>>> spin_lock_irqsave(&ct->ctbs.send.lock, flags);
> > > >>>>>> + if (unlikely(!ctb_has_room(ctb, len + 1))) {
> > > >>>>>> + spin_unlock_irqrestore(&ct->ctbs.send.lock, flags);
> > > >>>>>> + cond_resched();
> > > >>>>>> + goto retry;
> > > >>>>>> + }
> > > >>>>>>
> > > >>>>>> So it looks like a change to ct_send to me. Is that wrong?
> > > >>>>
> > > >>>> What about this part - is the patch changing the blocking ct_send or not,
> > > >>>> and if it is why?
> > > >>>>
> > > >>>
> > > >>> Yes, ct_send() changes. Sorry for the confusion.
> > > >>>
> > > >>> This function needs to be updated to account for the H2G space and
> > > >>> backoff if no space is available.
> > > >>
> > > >> Since this one is the sleeping path, it probably can and needs to be smarter
> > > >> than having a cond_resched busy loop added. Like sleep and get woken up when
> > > >> there is space. Otherwise it can degenerate to busy looping via contention
> > > >> with the non-blocking path.
> > > >>
> > > >
> > > > That screams over enginerring a simple problem to me. If the CT channel
> > > > is full we are really in trouble anyways - i.e. the performance is going
> > > > to terrible as we overwhelmed the GuC with traffic. That being said,
> > >
> > > Performance of what would be terrible? Something relating to submitting
> > > new jobs to the GPU I guess. Or something SRIOV related as you hint below.
> > >
> > > But there is no real reason why CPU cycles/power should suffer if GuC is
> > > busy.
> > >
> > > Okay, if it can't happen in real world then it's possibly passable as a
> > > design of a communication interface. But to me it leaves a bad taste and
> > > a doubt that there is this other aspect of the real world. And that is
> > > when the unexpected happens. Even the most trivial things like a bug in
> > > GuC firmware causes the driver to busy spin in there. So not much
> > > happening on the machine but CPU cores pinned burning cycles in this
> > > code. It's just lazy and not robust design. "Bug #nnnnn - High CPU usage
> > > and GUI blocked - Solution: Upgrade GuC firmware and _reboot_ the
> > > machine". Oh well..
> > >
> > > At least I think the commit message should spell out clearly that a busy
> > > looping path is being added to the sleeping send as a downside of
> > > implementation choices. Still, for the record, I object to the design.
> > >
> > > > IGTs can do this but that really isn't a real world use case. For the
> > > > real world, this buffer is large enough that it won't ever be full hence
> > > > the comment + lazy spin loop.
> > > >
> > > > Next, it isn't like we get an interrupt or something when space
> > > > becomes available so how would we wake this thread? Could we come up
> > > > with a convoluted scheme where we insert ops that generated an interrupt
> > > > at regular intervals, probably? Would it be super complicated, totally
> > > > unnecessary, and gain use nothing - absolutely.
> > > >
> > > > Lastly, blocking CTBs really shouldn't ever be used. Certainly the
> > > > submission code doesn't use these. I think SRIOV might, but those can
> > > > probably be reworked too to use non-blocking. At some point we might
> > > > want to scrub the driver and just delete the blocking path.
> >
> > I'd do an s/cond_resched()/msleep(1)/ and comment explaining why we
> > just don't care about this. That checks of the cpu wasting in this
> > case (GuC is overloaded, it wont come back anytime soon anyway) and
> > explains why we really don't want to make this any more clever or
> > complex code (because comment can explain why we wont hit this in
> > actual real world usage except when something else is on fire already
> > anyway).
> >
>
> Sounds good.
>
> > If you want to go absolutely overkill and it's not too much work, make
> > the msleep interruptible or check for signals, and bail out. That way
> > the process can be made unstuck with ^C at least.
>
> This loop is already bound by a timer and if no forward progress is made
> we pop out of this loop. It is assumed if this happens the GuC / GPU is
> dead a and full GPU reset will have to be issued. A following patch
> adds the timer, a bit later in submission section of the series a patch
> is added to trigger the reset.
Yeah timeout bail-out works too, and if you then switch it from timeout to
also interruptible it shouldn't be much more code. It's just nice to not
have any uninterruptible sleep.
-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
More information about the Intel-gfx
mailing list