Re: [PATCH] iommu/arm-smmu: fix "hang" when games exit

From: Rob Clark
Date: Tue Sep 10 2019 - 11:45:42 EST


On Tue, Sep 10, 2019 at 8:01 AM Robin Murphy <robin.murphy@xxxxxxx> wrote:
>
> On 07/09/2019 18:50, Rob Clark wrote:
> > From: Rob Clark <robdclark@xxxxxxxxxxxx>
> >
> > When games, browser, or anything using a lot of GPU buffers exits, there
> > can be many hundreds or thousands of buffers to unmap and free. If the
> > GPU is otherwise suspended, this can cause arm-smmu to resume/suspend
> > for each buffer, resulting 5-10 seconds worth of reprogramming the
> > context bank (arm_smmu_write_context_bank()/arm_smmu_write_s2cr()/etc).
> > To the user it would appear that the system is locked up.
> >
> > A simple solution is to use pm_runtime_put_autosuspend() instead, so we
> > don't immediately suspend the SMMU device.
> >
> > Signed-off-by: Rob Clark <robdclark@xxxxxxxxxxxx>
> > ---
> > Note: I've tied the autosuspend enable/delay to the consumer device,
> > based on the reasoning that if the consumer device benefits from using
> > an autosuspend delay, then it's corresponding SMMU probably does too.
> > Maybe that is overkill and we should just unconditionally enable
> > autosuspend.
>
> I'm not sure there's really any reason to expect that a supplier's usage
> model when doing things for itself bears any relation to that of its
> consumer(s), so I'd certainly lean towards the "unconditional" argument
> myself.

Sounds good, I'll respin w/ unconditional autosuspend

> Of course ideally we'd skip resuming altogether in the map/unmap paths
> (since resume implies a full TLB reset anyway), but IIRC that approach
> started to get messy in the context of the initial RPM patchset. I'm
> planning to fiddle around a bit more to clean up the implementation of
> the new iommu_flush_ops stuff, so I've made a note to myself to revisit
> RPM to see if there's a sufficiently clean way to do better. In the
> meantime, though, I don't have any real objection to using some
> reasonable autosuspend delay on the principle that if we've been woken
> up to map/unmap one page, there's a high likelihood that more will
> follow in short order (and in the configuration slow-paths it won't have
> much impact either way).

It does sort of remind me about something I was chatting with Jordan
the other day.. about how we could possibly skip the TLB inv for
unmaps from non-current pagetables once we have per-context
pagetables.

The challenge is, since the GPU's command parser is the one switching
pagetables, we don't have any race-free way to know which pagetables
are current. But we do know which contexts have work queued up for
the GPU, so we can know either that a given context definitely isn't
current, or that it might be current. And in the "definitely not
current" case we could skip TLB inv.

BR,
-R

>
> Robin.
>
> > drivers/iommu/arm-smmu.c | 11 ++++++++++-
> > 1 file changed, 10 insertions(+), 1 deletion(-)
> >
> > diff --git a/drivers/iommu/arm-smmu.c b/drivers/iommu/arm-smmu.c
> > index c2733b447d9c..73a0dd53c8a3 100644
> > --- a/drivers/iommu/arm-smmu.c
> > +++ b/drivers/iommu/arm-smmu.c
> > @@ -289,7 +289,7 @@ static inline int arm_smmu_rpm_get(struct arm_smmu_device *smmu)
> > static inline void arm_smmu_rpm_put(struct arm_smmu_device *smmu)
> > {
> > if (pm_runtime_enabled(smmu->dev))
> > - pm_runtime_put(smmu->dev);
> > + pm_runtime_put_autosuspend(smmu->dev);
> > }
> >
> > static struct arm_smmu_domain *to_smmu_domain(struct iommu_domain *dom)
> > @@ -1445,6 +1445,15 @@ static int arm_smmu_attach_dev(struct iommu_domain *domain, struct device *dev)
> > /* Looks ok, so add the device to the domain */
> > ret = arm_smmu_domain_add_master(smmu_domain, fwspec);
> >
> > +#ifdef CONFIG_PM
> > + /* TODO maybe device_link_add() should do this for us? */
> > + if (dev->power.use_autosuspend) {
> > + pm_runtime_set_autosuspend_delay(smmu->dev,
> > + dev->power.autosuspend_delay);
> > + pm_runtime_use_autosuspend(smmu->dev);
> > + }
> > +#endif
> > +
> > rpm_put:
> > arm_smmu_rpm_put(smmu);
> > return ret;
> >