Re: [PATCH 7/7] iommu/ipmmu-vmsa: Add suspend/resume support
From: Geert Uytterhoeven
Date: Wed Feb 20 2019 - 14:47:21 EST
Hi Laurent,
On Wed, Feb 20, 2019 at 5:11 PM Laurent Pinchart
<laurent.pinchart@xxxxxxxxxxxxxxxx> wrote:
> On Wed, Feb 20, 2019 at 05:05:49PM +0100, Geert Uytterhoeven wrote:
> > On Wed, Feb 20, 2019 at 4:42 PM Laurent Pinchart wrote:
> > > On Wed, Feb 20, 2019 at 04:05:31PM +0100, Geert Uytterhoeven wrote:
> > >> During PSCI system suspend, R-Car Gen3 SoCs are powered down, and all
> > >> IPMMU state is lost. Hence after s2ram, devices wired behind an IPMMU,
> > >> and configured to use it, will see their DMA operations hang.
> > >>
> > >> To fix this, restore all IPMMU contexts, and re-enable all active
> > >> micro-TLBs during system resume.
> > >>
> > >> To avoid overhead on platforms not needing it, the resume code has a
> > >> build time dependency on sleep and PSCI support, and a runtime
> > >> dependency on PSCI.
> > >> --- a/drivers/iommu/ipmmu-vmsa.c
> > >> +++ b/drivers/iommu/ipmmu-vmsa.c
> > >> @@ -1158,10 +1166,52 @@ static int ipmmu_remove(struct platform_device *pdev)
> > >> return 0;
> > >> }
> > >>
> > >> +#if defined(CONFIG_PM_SLEEP) && defined(CONFIG_ARM_PSCI_FW)
> > >> +static int ipmmu_resume_noirq(struct device *dev)
> > >> +{
> > >> + struct ipmmu_vmsa_device *mmu = dev_get_drvdata(dev);
> > >> + unsigned int i;
> > >> +
> > >> + /* This is the best we can do to check for the presence of PSCI */
> > >> + if (!psci_ops.cpu_suspend)
> > >> + return 0;
> > >
> > > PSCI suspend disabling power to the SoC completely may be a common
> > > behaviour on our development boards, but isn't mandated by the PSCI
> > > specification if I'm not mistaken. Is there a way to instead detect that
> > > power has been lost, perhaps by checking whether a register has been
> > > reset to its default value ?
> >
> > The approach here is the same as in the clk and pinctrl drivers.
> >
> > I think we could check if the IMCTR registers for allocated domains in root
> > IPMMUs are non-zero. But that's about as expensive as doing the full
> > restore, I think.
>
> Would reading just one register be more expensive that full
> reconfiguration ? Or is there no single register that could serve this
> purpose ?
>
> > And it may have to be done for each and every IPMMU instance, or do
> > you trust caching for this?
>
> If we can find a single register I think that reading it for every IPMMU
> instance wouldn't be an issue.
Upon more thought, probably it can be done by reading the IMCTR
for the first non-zero domain. Will look into it...
> > >> +static const struct dev_pm_ops ipmmu_pm = {
> > >> + SET_NOIRQ_SYSTEM_SLEEP_PM_OPS(NULL, ipmmu_resume_noirq)
> > >> +};
> > >> +#define DEV_PM_OPS &ipmmu_pm
> > >> +#else
> > >> +#define DEV_PM_OPS NULL
> > >> +#endif /* CONFIG_PM_SLEEP && CONFIG_ARM_PSCI_FW */
> > >> +
> > >> static struct platform_driver ipmmu_driver = {
> > >> .driver = {
> > >> .name = "ipmmu-vmsa",
> > >> .of_match_table = of_match_ptr(ipmmu_of_ids),
> > >> + .pm = DEV_PM_OPS,
> > >
> > > I would have used conditional compilation here instead of using a
> > > DEV_PM_OPS macro, as I think the macro decreases readability (and also
> > > given how its generic name could later conflict with something else).
> >
> > You mean
> >
> > #ifdef ...
> > .pm = &ipmmu_pm,
> > #endif
> >
> > and marking ipmmu_pm __maybe_unused__?
>
> Yes. Up to you.
I'm not such a big fan of __maybe_unused. It's easy to add, and too
easy to forget to remove it when it's no longer needed (casts have the
same problem).
Usually people just annotate the actual suspend/resume methods with
__maybe_unused, which still leaves the (large) struct dev_pm_ops in
memory.
So I started preferring the DEV_PM_OPS approach...
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx
In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds