Hi Laurent,
On Wed, Feb 20, 2019 at 4:42 PM Laurent Pinchart
<laurent.pinchart@xxxxxxxxxxxxxxxx> 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.
> >
> > Signed-off-by: Geert Uytterhoeven <geert+renesas@xxxxxxxxx>
> > ---
> > This patch takes a different approach than the BSP, which implements a
> > bulk save/restore of all registers during system suspend/resume.
>
> I like this approach better too.
Thanks ;-)
> > --- a/drivers/iommu/ipmmu-vmsa.c
> > +++ b/drivers/iommu/ipmmu-vmsa.c
> > @@ -58,6 +62,7 @@ struct ipmmu_vmsa_device {
> > spinlock_t lock; /* Protects ctx and domains[] */
> > DECLARE_BITMAP(ctx, IPMMU_CTX_MAX);
> > struct ipmmu_vmsa_domain *domains[IPMMU_CTX_MAX];
> > + s8 utlb_ctx[IPMMU_UTLB_MAX];
>
> How about making this a bitmask instead to save memory ? I would also
> rename it as utlb_ctx doesn't really carry the meaning of the field,
> whose purpose is to store whether the µTLB is enabled or disabled.
This field isn't just a binary flag, but stores the context used for the
uTLB, so we can map from micro-TLB to context.
Given there can be 8 contexts, plus the need to indicate unused contexts,
that means 4 bits/micro-TLB. So the overhead is just 24 bytes per IPMMU
instance.
I considered allocating the array dynamically (by having s8 utlb_ctx[]
at the end of the structure), but didn't go that route, as the domains[]
array already uses more memory.
> > @@ -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. And it may have to be done for each and every IPMMU
instance, or do you trust caching for this?
> > +
> > + /* Reset root MMU and restore contexts */
>
> I think the rest of the code adds a period at the end of sentences in
> comments.
The balance seems to be just under 50% ;-)
> > + if (ipmmu_is_root(mmu)) {
> > + ipmmu_device_reset(mmu);
> > +
> > + for (i = 0; i < mmu->num_ctx; i++) {
> > + if (!mmu->domains[i])
> > + continue;
> > +
> > + ipmmu_context_init(mmu->domains[i]);
> > + }
> > + }
> > +
> > + /* Re-enable active micro-TLBs */
> > + for (i = 0; i < mmu->features->num_utlbs; i++) {
> > + if (mmu->utlb_ctx[i] == IPMMU_CTX_INVALID)
> > + continue;
> > +
> > + ipmmu_utlb_enable(mmu->root->domains[mmu->utlb_ctx[i]], i);
> > + }
> > +
> > + return 0;
> > +}
> > +
> > +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__?
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
