Re: [PATCH v7 2/2] irqchip: Add Qualcomm MPM controller driver

[Shawn Guo]

On Sat, Mar 05, 2022 at 11:05:07AM +0000, Marc Zyngier wrote:
> On Sat, 05 Mar 2022 09:24:20 +0000,
> Shawn Guo <shawn.guo@xxxxxxxxxx> wrote:
> > 
> > On Fri, Mar 04, 2022 at 03:24:43PM +0000, Marc Zyngier wrote:
> > > On Fri, 04 Mar 2022 08:23:42 +0000,
> > > Shawn Guo <shawn.guo@xxxxxxxxxx> wrote:
> > > > 
> > > > On Fri, Mar 04, 2022 at 07:59:15AM +0000, Marc Zyngier wrote:
> > > > > On Thu, 03 Mar 2022 04:02:29 +0000,
> > > > > Shawn Guo <shawn.guo@xxxxxxxxxx> wrote:
> > > > > > 
> > > > > > On Wed, Mar 02, 2022 at 01:57:27PM +0000, Marc Zyngier wrote:
> > > > > > > This code actually makes me ask more questions. Why is it programming
> > > > > > > 2 'pins' for each IRQ?
> > > > > > 
> > > > > > The mapping between MPM pin and GIC IRQ is not strictly 1-1.  There are
> > > > > > some rare case that up to 2 MPM pins map to a single GIC IRQ, for
> > > > > > example the last two in QC2290 'qcom,mpm-pin-map' below.
> > > > > > 
> > > > > > 	qcom,mpm-pin-map = <2 275>,     /* tsens0_tsens_upper_lower_int */
> > > > > > 			   <5 296>,     /* lpass_irq_out_sdc */
> > > > > > 			   <12 422>,    /* b3_lfps_rxterm_irq */
> > > > > > 			   <24 79>,     /* bi_px_lpi_1_aoss_mx */
> > > > > > 			   <86 183>,    /* mpm_wake,spmi_m */
> > > > > > 			   <90 260>,    /* eud_p0_dpse_int_mx */
> > > > > > 			   <91 260>;    /* eud_p0_dmse_int_mx */
> > > > > > 
> > > > > > 
> > > > > > The downstream uses a DT bindings that specifies GIC hwirq number in
> > > > > > client device nodes.  In that case, d->hwirq in the driver is GIC IRQ
> > > > > > number, and the driver will need to query mapping table, find out the
> > > > > > possible 2 MPM pins, and set them up.
> > > > > > 
> > > > > > The patches I'm posting here use a different bindings that specifies MPM
> > > > > > pin instead in client device nodes.  Thus the driver can simply get the
> > > > > > MPM pin from d->hwirq, so that the whole look-up procedure can be saved.
> > > > > 
> > > > > It still remains that there is no 1:1 mapping between input and
> > > > > output, which is the rule #1 to be able to use a hierarchical setup.
> > > > 
> > > > For direction of MPM pin -> GIC interrupt, it's a 1:1 mapping, i.e. for
> > > > given MPM pin, there is only one GIC interrupt.  And that's the
> > > > mapping MPM driver relies on.  For GIC interrupt -> MPM pin, it's not
> > > > a strict 1:1 mapping.
> > > 
> > > Then this isn't a 1:1 mapping *AT ALL*. The hierarchical setup
> > > mandates that the mapping is a bijective function, and that's exactly
> > > what 1:1 means. There is no such thing a 1:1 in a single
> > > direction. When you take an interrupt, all you see is the GIC
> > > interrupt. How do you know which of the *two* pins interrupted you? Oh
> > > wait, you *can't* know. You end-up never servicing one of the two
> > > interrupts
> > 
> > Yes, you are right!  But that might be a problem only in theory.  I
> > checked all the Qualcomm platforms I know built on MPM, and found that
> > the only 2:1 case is USB DP & DM sensing pins.  Since these two pins
> > will be handled by USB driver with a single interrupt handler, it should
> > not cause any problem in practice.  That said, the 2:1 mapping is just
> > a special case specific to USB, and MPM driver can be implemented as if
> > it's just a 1:1 mapping.
> >
> > Shawn
> > 
> > > (and I suspect this results in memory corruption if you
> > > tear a hierarchy down).
> 
> Key point here ^^^^^^^^^^
> 
> You can't have *any* interrupt that fits this 2:1 model if the irqchip
> implements 1:1. Think about the data structures for a second:
> 
> Pins x and y and routed to GIC interrupt z. This results in the
> following irq_data structures:
> 
>    MPM-x ---\
>              GIC-z
>    MPM-y ---/
> 
> Now, the driver using these interrupts is being removed, and the
> hierarchies is being freed. Tearing down the interrupt with pin x will
> result in z being also freed. And then you'll process pin y, which
> will just explode.

I tested with manually unbinding the USB driver and didn't run into any
memory corruption.  If I read irq_domain code right, it seems that
irq_domain_alloc_irq_data() will call into irq_domain_insert_irq_data()
to allocate z irq_data in context of virq x and y respectively.  So x
and y do not share a single parent (z) irq_data but have their own copy
of z irq_data, no?

Shawn