[PATCH 01/18] arm64: initial support for GICv3

Marc Zyngier marc.zyngier at arm.com
Wed Feb 26 07:53:34 EST 2014


On 25/02/14 18:06, Will Deacon wrote:
> Hi Marc,
> 
> On Wed, Feb 05, 2014 at 01:30:33PM +0000, Marc Zyngier wrote:
>> diff --git a/arch/arm64/kernel/head.S b/arch/arm64/kernel/head.S
>> index 0b281ff..522649b 100644
>> --- a/arch/arm64/kernel/head.S
>> +++ b/arch/arm64/kernel/head.S
>> @@ -22,6 +22,7 @@
>>
>>  #include <linux/linkage.h>
>>  #include <linux/init.h>
>> +#include <linux/irqchip/arm-gic-v3.h>
>>
>>  #include <asm/assembler.h>
>>  #include <asm/ptrace.h>
>> @@ -183,6 +184,19 @@ CPU_LE(    bic     x0, x0, #(3 << 24)      )       // Clear the EE and E0E bits for EL1
>>         msr     cnthctl_el2, x0
>>         msr     cntvoff_el2, xzr                // Clear virtual offset
>>
>> +       /* GICv3 system register access */
>> +       mrs     x0, id_aa64pfr0_el1
>> +       ubfx    x0, x0, #24, #4
>> +       cmp     x0, #1
>> +       b.ne    3f
>> +
>> +       mrs     x0, ICC_SRE_EL2                 // Set ICC_SRE_EL2.SRE==1
>> +       orr     x0, x0, #1
> 
> Don't you need to set the Enable bit (bit 3) too? Also, does setting
> ICC_SRE_EL2.SRE force KVM to use the ICH_* registers? If so, that seems
> like a pretty big decision to take this early on (there's not even an
> #ifdef CONFIG_ARM_GICV3).

The test on id_aa64pfr0_el1 is an absolute guarantee that we have the
GICv3 system registers.

As for the Enable bit, this is handle by the magic trapping we have
below, but I decided to get rid of that code (see the previous
discussion with Christoffer).
> 
>> +       msr     ICC_SRE_EL2, x0
>> +       isb                                     // Make sure SRE is now 1
>> +       msr     ICH_HCR_EL2, xzr                // Reset ICC_HCR_EL2 to defaults
>> +
>> +3:
>>         /* Populate ID registers. */
>>         mrs     x0, midr_el1
>>         mrs     x1, mpidr_el1
>> diff --git a/arch/arm64/kernel/hyp-stub.S b/arch/arm64/kernel/hyp-stub.S
>> index 0959611..c850d1e 100644
>> --- a/arch/arm64/kernel/hyp-stub.S
>> +++ b/arch/arm64/kernel/hyp-stub.S
>> @@ -19,6 +19,7 @@
>>
>>  #include <linux/init.h>
>>  #include <linux/linkage.h>
>> +#include <linux/irqchip/arm-gic-v3.h>
>>
>>  #include <asm/assembler.h>
>>  #include <asm/ptrace.h>
>> @@ -51,16 +52,52 @@ ENDPROC(__hyp_stub_vectors)
>>
>>         .align 11
>>
>> +#define sysreg_to_iss(op0, op1, crn, crm, op2) \
>> +       (((op0 & 3) << 20) | ((op2 & 7 ) << 17) | \
>> +        ((op1 & 7 ) << 14) | ((crn & 15) << 10) | \
>> +        ((crm & 15) << 1))
>> +
>>  el1_sync:
>> +       // Hack alert: we don't have a stack, and we can't use the one
>> +       // from EL1 because our MMU is off. Thankfully, we have
>> +       // tpidr_el2, which is under our complete control, and
>> +       // tpidr_el0, which can be used as long as we haven't reached
>> +       // userspace. Oh hum...
> 
> Yuck. Why can't we create a minimal stack when we initialise EL2 when
> booting the host? Even simpler, just use some .quads in here (you're already
> using a literal pool).

I really liked the "funkiness" here, but that code is now gone. RIP! ;-)

>> +       msr     tpidr_el2, x1                   // Preserve x1
>> +
>>         mrs     x1, esr_el2
>>         lsr     x1, x1, #26
>>         cmp     x1, #0x16
>> -       b.ne    2f                              // Not an HVC trap
>> +       b.ne    3f                              // Not an HVC trap
>> +
>> +       // Handle getting/setting vbar_el2
>>         cbz     x0, 1f
>>         msr     vbar_el2, x0                    // Set vbar_el2
>>         b       2f
>>  1:     mrs     x0, vbar_el2                    // Return vbar_el2
>> -2:     eret
>> +2:     mrs     x1, tpidr_el2                   // Restore x1
>> +       eret
>> +
>> +3:     cmp     x1, #0x18                       // MSR/MRS?
>> +       b.ne    2b
>> +
>> +       // Handle trapped sys reg access
>> +       msr     tpidr_el0, x0                   // Preserve x0
>> +       mrs     x1, esr_el2
>> +       ldr     x0, =sysreg_to_iss(3, 7, 15, 15, 7)
>> +       and     x1, x1, x0
>> +       ldr     x0, =sysreg_to_iss(3, 0, 12, 12, 5)
> 
> I think sysreg_to_iss actually makes this *less* readable! Why not simply
> #define the mask to extract OPx, CRx then compare with another #define for
> ICC_SRE_EL1? If we can represent those as immediates, you can avoid the
> literal loads too.

Gone.

>> +       eor     x0, x0, x1                      // Check for ICC_SRE_EL1
>> +       cbnz    x0, 4f                          // Something bad happened
>> +
>> +       // Set ICC_SRE_EL1 access enable
>> +       mrs     x1, ICC_SRE_EL2
>> +       orr     x1, x1, #(1 << 3)
>> +       msr     ICC_SRE_EL2, x1
> 
> Ah, ok, you do this part lazily. Does that make sense for the host? (i.e. is
> it possible for KVM to use SRE at EL2 but not the host at EL1?).

You could, but I'm not sure it is worth supporting. In the interest of
simplicity, I got rid of all that crap.

>> +
>> +4:     mrs     x0, tpidr_el0                   // Restore x0
>> +       mrs     x1, tpidr_el2                   // Restore x1
>> +       eret
>>  ENDPROC(el1_sync)
>>
>>  .macro invalid_vector  label
>> diff --git a/drivers/irqchip/irq-gic-v3.c b/drivers/irqchip/irq-gic-v3.c
>> new file mode 100644
>> index 0000000..487e061
>> --- /dev/null
>> +++ b/drivers/irqchip/irq-gic-v3.c
>> @@ -0,0 +1,665 @@
>> +/*
>> + * Copyright (C) 2013 ARM Limited, All Rights Reserved.
>> + * Author: Marc Zyngier <marc.zyngier at arm.com>
>> + *
>> + * This program is free software; you can redistribute it and/or modify
>> + * it under the terms of the GNU General Public License version 2 as
>> + * published by the Free Software Foundation.
>> + *
>> + * This program is distributed in the hope that it will be useful,
>> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
>> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
>> + * GNU General Public License for more details.
>> + *
>> + * You should have received a copy of the GNU General Public License
>> + * along with this program.  If not, see <http://www.gnu.org/licenses/>.
>> + */
>> +
>> +#include <linux/cpu.h>
>> +#include <linux/interrupt.h>
>> +#include <linux/of.h>
>> +#include <linux/of_address.h>
>> +#include <linux/of_irq.h>
>> +#include <linux/percpu.h>
>> +#include <linux/slab.h>
>> +
>> +#include <linux/irqchip/arm-gic-v3.h>
>> +
>> +#include <asm/cputype.h>
>> +#include <asm/exception.h>
>> +#include <asm/smp_plat.h>
>> +
>> +#include "irqchip.h"
>> +
>> +struct gic_chip_data {
>> +       void __iomem            *dist_base;
>> +       void __iomem            **redist_base;
>> +       void __percpu __iomem   **rdist;
>> +       struct irq_domain       *domain;
>> +       u64                     redist_stride;
>> +       u32                     redist_regions;
>> +       unsigned int            irq_nr;
>> +};
>> +
>> +static struct gic_chip_data gic_data __read_mostly;
>> +
>> +#define gic_data_rdist_rd_base()       (*__this_cpu_ptr(gic_data.rdist))
> 
> You should avoid __this_cpu_ptr (it's being removed afaik). Take a look at
> the patches from Christoph Lameter, e.g. 170d800af83f ("block: Replace
> __get_cpu_var uses").

Ah, interesting. I'll repaint this.

>> +#define gic_data_rdist_sgi_base()      (gic_data_rdist_rd_base() + SZ_64K)
>> +
>> +static DEFINE_RAW_SPINLOCK(dist_lock);
>> +
>> +#define DEFAULT_PMR_VALUE      0xf0
>> +
>> +static inline void __iomem *gic_dist_base(struct irq_data *d)
>> +{
>> +       if (d->hwirq < 32)      /* SGI+PPI -> SGI_base for this CPU */
>> +               return gic_data_rdist_sgi_base();
>> +
>> +       if (d->hwirq <= 1023)   /* SPI -> dist_base */
>> +               return gic_data.dist_base;
>> +
>> +       if (d->hwirq >= 8192)
>> +               BUG();          /* LPI Detected!!! */
>> +
>> +       return NULL;
>> +}
>> +
>> +static inline unsigned int gic_irq(struct irq_data *d)
>> +{
>> +       return d->hwirq;
>> +}
>> +
>> +static void gic_do_wait_for_rwp(void __iomem *base)
>> +{
>> +       u32 val;
>> +
>> +       do {
>> +               val = readl_relaxed(base + GICD_CTLR);
>> +               cpu_relax();
>> +       } while (val & GICD_CTLR_RWP);
> 
> You should add a timeout in here to save us in the face of broken hardware
> (especially as you hold the distributor lock). See, for example,
> arm_smmu_tlb_sync(...) in the arm-smmu driver.

Yup.

> Thinking about it, is this the only point of the dist_lock (to ensure you
> wait only for your request)? Do we actually need locking for that?

You need to ensure that the distributor has propagated the request. I'm
not sure what the effect would be if we had several CPUs writing to the
distributor at once... I'll investigate.

>> +}
>> +
>> +/* Wait for completion of a distributor change */
>> +static void gic_dist_wait_for_rwp(void)
>> +{
>> +       gic_do_wait_for_rwp(gic_data.dist_base);
>> +}
>> +
>> +/* Wait for completion of a redistributor change */
>> +static void gic_redist_wait_for_rwp(void)
>> +{
>> +       gic_do_wait_for_rwp(gic_data_rdist_rd_base());
>> +}
>> +
>> +static void gic_wait_for_rwp(int irq)
>> +{
>> +       if (irq < 32)
> 
> Maybe we should just bite the bullet and have helpers for
> gic_irq_is_sgi(irq) etc. so we can avoid these magic numbers all over the
> place.
> 
>> +               gic_redist_wait_for_rwp();
>> +       else
>> +               gic_dist_wait_for_rwp();
>> +}
> 
> Saying that, I question the value of mixing this into a single function.
> gic_redist_wait_for_rwp absolutely *must* be called with preemption disabled
> and on the appropriate CPU for the IRQ in question, whilst the latter can be
> called anywhere. So even though you can call this single function from a
> bunch of places, you're really relying on the irq parameter being correct
> for the context of the caller.
> 
> I reckon it would actually be clearer to call the relevant wait function
> explicitly in the caller. Same for gic_dist_base.

Fair enough. I'll try to rework that part.

>> +/* Low level accessors */
>> +static u64 gic_read_iar(void)
>> +{
>> +       u64 irqstat;
>> +
>> +       asm volatile("mrs %0, " reg(ICC_IAR1_EL1) : "=r" (irqstat));
> 
> Hmm, reg() is a terrible name for a macro :) Can't you just use stringify
> directly?

Too long? ;-)

>> +       return irqstat;
>> +}
>> +
>> +static void gic_write_pmr(u64 val)
>> +{
>> +       asm volatile("msr " reg(ICC_PMR_EL1) ", %0" : : "r" (val));
>> +       isb();
>> +}
>> +
>> +static void gic_write_ctlr(u64 val)
>> +{
>> +       asm volatile("msr " reg(ICC_CTLR_EL1) ", %0" : : "r" (val));
>> +       isb();
>> +}
>> +
>> +static void gic_write_grpen1(u64 val)
>> +{
>> +       asm volatile("msr " reg(ICC_GRPEN1_EL1) ", %0" : : "r" (val));
>> +       isb();
>> +}
>> +
>> +static void gic_write_sgi1r(u64 val)
>> +{
>> +       asm volatile("msr " reg(ICC_SGI1R_EL1) ", %0" : : "r" (val));
>> +       isb();
>> +}
> 
> I'm not convinced you *always* need these isbs. For example, the
> gic_handle_irq loop probably just needs one at the end, before returning.
> 
> In fact, looking at the spec, ICC_PMR_EL1 is actually self-synchronising so
> you don't need anything there. For SGI1R at least, you need a dsb to ensure
> write observability by the GIC (it's not clear to me whether you also need
> to the isb).

The dsb is actually in gic_send_sgi. We may end-up doing several writes
to ICC_SGI1R_EL1 if the targets are on different Aff{3,2,1} levels, so
it made sense to put it there.

>> +static void gic_enable_sre(void)
>> +{
>> +       u64 val;
>> +
>> +       /* This access is likely to trap into EL2 */
>> +       asm volatile("mrs %0, " reg(ICC_SRE_EL1) : "=r" (val));
>> +       val |= GICC_SRE_EL1_SRE;
>> +       asm volatile("msr " reg(ICC_SRE_EL1) ", %0" : : "r" (val));
>> +       isb();
>> +}
>> +
>> +static void gic_enable_redist(void)
>> +{
>> +       void __iomem *rbase;
>> +       u32 val;
>> +
>> +       rbase = gic_data_rdist_rd_base();
>> +
>> +       /* Wake up this CPU redistributor */
>> +       val = readl_relaxed(rbase + GICR_WAKER);
>> +       val &= ~GICR_WAKER_ProcessorSleep;
>> +       writel_relaxed(val, rbase + GICR_WAKER);
>> +
>> +       do {
>> +               val = readl_relaxed(rbase + GICR_WAKER);
>> +               cpu_relax();
>> +       } while (val & GICR_WAKER_ChildrenAsleep);
> 
> Again, you want a timeout here.

Agreed.

>> +}
>> +
>> +/*
>> + * Routines to acknowledge, disable and enable interrupts
>> + */
>> +static void gic_mask_irq(struct irq_data *d)
>> +{
>> +       u32 mask = 1 << (gic_irq(d) % 32);
>> +
>> +       raw_spin_lock(&dist_lock);
>> +       writel_relaxed(mask, gic_dist_base(d) + GICD_ICENABLER + (gic_irq(d) / 32) * 4);
>> +       gic_wait_for_rwp(gic_irq(d));
>> +       raw_spin_unlock(&dist_lock);
>> +}
>> +
>> +static void gic_unmask_irq(struct irq_data *d)
>> +{
>> +       u32 mask = 1 << (gic_irq(d) % 32);
>> +
>> +       raw_spin_lock(&dist_lock);
>> +       writel_relaxed(mask, gic_dist_base(d) + GICD_ISENABLER + (gic_irq(d) / 32) * 4);
>> +       gic_wait_for_rwp(gic_irq(d));
>> +       raw_spin_unlock(&dist_lock);
>> +}
>> +
>> +static void gic_eoi_irq(struct irq_data *d)
>> +{
>> +       gic_write_eoir(gic_irq(d));
>> +}
> 
> Can you inline the code here (this is the only user) and move the isb out?

Yup.

>> +static int gic_set_type(struct irq_data *d, unsigned int type)
>> +{
>> +       void __iomem *base = gic_dist_base(d);
>> +       unsigned int irq = gic_irq(d);
>> +       u32 enablemask = 1 << (irq % 32);
>> +       u32 enableoff = (irq / 32) * 4;
>> +       u32 confmask = 0x2 << ((irq % 16) * 2);
>> +       u32 confoff = (irq / 16) * 4;
>> +       bool enabled = false;
>> +       u32 val;
>> +
>> +       /* Interrupt configuration for SGIs can't be changed */
>> +       if (irq < 16)
>> +               return -EINVAL;
>> +
>> +       if (type != IRQ_TYPE_LEVEL_HIGH && type != IRQ_TYPE_EDGE_RISING)
>> +               return -EINVAL;
>> +
>> +       raw_spin_lock(&dist_lock);
>> +
>> +       val = readl_relaxed(base + GICD_ICFGR + confoff);
>> +       if (type == IRQ_TYPE_LEVEL_HIGH)
>> +               val &= ~confmask;
>> +       else if (type == IRQ_TYPE_EDGE_RISING)
>> +               val |= confmask;
>> +
>> +       /*
>> +        * As recommended by the spec, disable the interrupt before changing
>> +        * the configuration
>> +        */
>> +       if (readl_relaxed(base + GICD_ISENABLER + enableoff) & enablemask) {
>> +               writel_relaxed(enablemask, base + GICD_ICENABLER + enableoff);
>> +               gic_wait_for_rwp(irq);
>> +               enabled = true;
>> +       }
>> +
>> +       writel_relaxed(val, base + GICD_ICFGR + confoff);
>> +
>> +       if (enabled) {
>> +               writel_relaxed(enablemask, base + GICD_ISENABLER + enableoff);
>> +               gic_wait_for_rwp(irq);
>> +       }
>> +
>> +       raw_spin_unlock(&dist_lock);
>> +
>> +       return 0;
>> +}
> 
> I really think this should be merged with the existing GIC driver -- the
> code is basically identical.

Not sure. I agree that the code is extremely similar, but some things
are slightly different:
- gic_dist_base can actually return a redistributor base address (which
is per-cpu). That doesn't fit with with the GICv2 model very well.
- The synchronization function doesn't exist in GICv2.

The closest thing I can imagine is a primitive like:

void gic_configure_irq(unsigned int irq, unsigned int type,
		       void __iomem *base, void (*sync)(void));

that'd do the low level stuff.

>> +static int gic_retrigger(struct irq_data *d)
>> +{
>> +       return -ENXIO;
>> +}
> 
> Why bother? (just don't assign the pointer in the chip).

Fair enough.

>> +static u64 gic_mpidr_to_affinity(u64 mpidr)
>> +{
>> +       /* Make sure we don't broadcast the interrupt */
>> +       return mpidr & ~GICD_IROUTER_SPI_MODE_ANY;
> 
> Yikes, mpidr meets the GIC! I'd be a lot more comfortable with this if we
> used our mpidr accessors to construct the value which GICD_IROUTER expects
> (e.g. using the MPIDR_AFFINITY_LEVEL macro).

Sure.

>> +}
>> +
>> +static asmlinkage void __exception_irq_entry gic_handle_irq(struct pt_regs *regs)
>> +{
>> +       u64 irqnr;
>> +
>> +       do {
>> +               irqnr = gic_read_iar();
> 
> What stops the CPU speculating this access around the loop?

Probably not much, if ICC_IAR_EL1 is not a bit "special". Speculating
into this would be very dangerous though, but nothing says otherwise.
I'll ask the question.

>> +               if (likely(irqnr > 15 && irqnr < 1021)) {
>> +                       irqnr = irq_find_mapping(gic_data.domain, irqnr);
>> +                       handle_IRQ(irqnr, regs);
>> +                       continue;
>> +               }
>> +               if (irqnr < 16) {
>> +                       gic_write_eoir(irqnr);
>> +#ifdef CONFIG_SMP
>> +                       handle_IPI(irqnr, regs);
>> +#else
>> +                       WARN_ONCE(true, "Unexpected SGI received!\n");
>> +#endif
>> +                       continue;
>> +               }
>> +       } while (irqnr != 0x3ff);
> 
> Any reason to adjust this control-flow from the existing GIC driver?

Mostly to accomodate the ITS code that goes right in the middle of this.

>> +}
>> +
>> +static void __init gic_dist_init(void)
>> +{
>> +       unsigned int i;
>> +       u64 affinity;
>> +       int gic_irqs = gic_data.irq_nr;
>> +       void __iomem *base = gic_data.dist_base;
>> +
>> +       /* Disable the distributor */
>> +       writel_relaxed(0, base + GICD_CTLR);
>> +
>> +       /*
>> +        * Set all global interrupts to be level triggered, active low.
>> +        */
>> +       for (i = 32; i < gic_data.irq_nr; i += 16)
>> +               writel_relaxed(0, base + GICD_ICFGR + i / 4);
>> +
>> +       /*
>> +        * Set priority on all global interrupts.
>> +        */
>> +       for (i = 32; i < gic_irqs; i += 4)
>> +               writel_relaxed(0xa0a0a0a0, base + GICD_IPRIORITYR + i);
>> +
>> +       /*
>> +        * Disable all interrupts.  Leave the PPI and SGIs alone
>> +        * as these enables are banked registers.
>> +        */
>> +       for (i = 32; i < gic_irqs; i += 32)
>> +               writel_relaxed(0xffffffff, base + GICD_ICENABLER + i / 8);
>> +
>> +       gic_dist_wait_for_rwp();
>> +
>> +       writel_relaxed(GICD_CTLR_ARE_NS | GICD_CTLR_ENABLE_G1A | GICD_CTLR_ENABLE_G1,
>> +                      base + GICD_CTLR);
>> +
>> +       /*
>> +        * Set all global interrupts to the boot CPU only. ARE must be
>> +        * enabled.
>> +        */
>> +       affinity = gic_mpidr_to_affinity(read_cpuid_mpidr());
> 
> You can use the logical map here.

OK.

>> +       for (i = 32; i < gic_data.irq_nr; i++)
>> +               writeq_relaxed(affinity, base + GICD_IROUTER + i * 8);
>> +}
> 
> Again, this looks very similar to the existing GIC driver.

I can probably do a similar thing as with gic_set_type, but I'm not sure
it adds much in terms of being readable.

>> +static int __init gic_populate_rdist(void)
>> +{
>> +       u64 mpidr = cpu_logical_map(smp_processor_id());
>> +       u64 typer;
>> +       u64 aff;
>> +       int i;
>> +
>> +       aff  = mpidr & ((1 << 24) - 1);
>> +       aff |= (mpidr >> 8) & (0xffUL << 24);
> 
> We should have helpers for this in cputype.h.

Yes, I fixed that already in the latest version of the code.

> Also, can we absolutely, 100%, "put-your-mother's-house-on-it" guarantee
> that the ITS, GICRn and the CPUs all agree on an identification scheme, as
> described in the mpidr? My reading of the spec doesn't fill me with
> confidence, and retrofitting a scheme in DT will end up with the same mess
> that we have for CPU topology.

My discussions with the architect of this fine piece of [insert
expletive here] has led me to believe that the MPIDR is the only thing
that holds the thing together. Of course, that doesn't preclude an
implementer from wrecking the ship, but I don't plan to support
something that doesn't follow the spec (or at least its intent).

>> +
>> +       for (i = 0; i < gic_data.redist_regions; i++) {
>> +               void __iomem *ptr = gic_data.redist_base[i];
>> +               u32 reg;
>> +
>> +               reg = readl_relaxed(ptr + GICR_PIDR2) & GIC_PIDR2_ARCH_MASK;
>> +               if (reg != 0x30 && reg != 0x40) { /* We're in trouble... */
>> +                       pr_warn("No redistributor present @%p\n", ptr);
>> +                       break;
>> +               }
>> +
>> +               do {
>> +                       typer = readq_relaxed(ptr + GICR_TYPER);
> 
> Not sure if you care, but readq is a show-stopper if you want to build this
> with arch/arm/.

I realized that recently. Any reason why we don't have writeq/readq on
LPAE-capable stuff (other than the obvious runtime detection madness)?

>> +                       if ((typer >> 32) == aff) {
>> +                               gic_data_rdist_rd_base() = ptr;
>> +                               pr_info("CPU%d: found redistributor %llx @%p\n",
>> +                                       smp_processor_id(),
>> +                                       (unsigned long long) mpidr, ptr);
>> +                               return 0;
>> +                       }
>> +
>> +                       if (gic_data.redist_stride) {
>> +                               ptr += gic_data.redist_stride;
>> +                       } else {
>> +                               ptr += SZ_64K * 2; /* Skip RD_base + SGI_base */
>> +                               if (typer & GICR_TYPER_VLPIS)
>> +                                       ptr += SZ_64K * 2; /* Skip VLPI_base + reserved page */
>> +                       }
> 
> VLPIS is RES0 in GICv3.

Indeed. But we still want this driver to run on GICv4 HW, even if we
don't use VLPIs just yet.

>> +               } while (!(typer & GICR_TYPER_LAST));
>> +       }
>> +
>> +       /* We couldn't even deal with ourselves... */
>> +       WARN(true, "CPU%d: mpidr %lx has no re-distributor!\n",
>> +            smp_processor_id(), (unsigned long)mpidr);
> 
> Inconsistency with the type of mpidr (you use unsigned long long above).

Well spotted.

>> +       return -ENODEV;
>> +}
>> +
>> +static void __init gic_cpu_init(void)
>> +{
>> +       void __iomem *rbase;
>> +       int i;
>> +
>> +       /* Register ourselves with the rest of the world */
>> +       if (gic_populate_rdist())
>> +               return;
>> +
>> +       gic_enable_redist();
>> +
>> +       rbase = gic_data_rdist_sgi_base();
>> +
>> +       /*
>> +        * Set priority on PPI and SGI interrupts
>> +        */
>> +       for (i = 0; i < 32; i += 4)
>> +               writel_relaxed(0xa0a0a0a0, rbase + GICR_IPRIORITYR0 + i * 4 / 4);
>> +
>> +       /*
>> +        * Disable all PPI interrupts, ensure all SGI interrupts are
>> +        * enabled.
>> +        */
>> +       writel_relaxed(0xffff0000, rbase + GICR_ICENABLER0);
>> +       writel_relaxed(0x0000ffff, rbase + GICR_ISENABLER0);
> 
> Whilst there are some clear differences between this and the existing
> driver, you could still factor parts of this out into a common file.

OK.

>> +       gic_redist_wait_for_rwp();
>> +
>> +       /* Enable system registers */
>> +       gic_enable_sre();
>> +
>> +       /* Set priority mask register */
>> +       gic_write_pmr(DEFAULT_PMR_VALUE);
>> +
>> +       /* EOI drops priority too (mode 0) */
>> +       gic_write_ctlr(GICC_CTLR_EL1_EOImode_drop_dir);
>> +
>> +       /* ... and let's hit the road... */
>> +       gic_write_grpen1(1);
>> +}
>> +
>> +#ifdef CONFIG_SMP
>> +static int __init gic_secondary_init(struct notifier_block *nfb,
>> +                                    unsigned long action, void *hcpu)
>> +{
>> +       if (action == CPU_STARTING || action == CPU_STARTING_FROZEN)
>> +               gic_cpu_init();
>> +       return NOTIFY_OK;
>> +}
>> +
>> +/*
>> + * Notifier for enabling the GIC CPU interface. Set an arbitrarily high
>> + * priority because the GIC needs to be up before the ARM generic timers.
>> + */
> 
> Why must the GIC be up before the timers? Obviously, the timers won't be
> able to generate interrupts without the GIC initialised, but the comment
> implies that something worse will happen.

They will also fail to request the IRQ (no irq domain present yet),
hence they will give up, and your kernel will grind to a halt...

>> +static struct notifier_block __initdata gic_cpu_notifier = {
>> +       .notifier_call = gic_secondary_init,
>> +       .priority = 100,
>> +};
>> +
>> +static u16 gic_compute_target_list(int *base_cpu, const struct cpumask *mask,
>> +                                  u64 cluster_id)
>> +{
>> +       int cpu = *base_cpu;
>> +       u64 mpidr = cpu_logical_map(cpu);
>> +       u16 tlist = 0;
>> +
>> +       while (cpu < nr_cpu_ids) {
>> +               /*
>> +                * If we ever get a cluster of more than 16 CPUs, just
>> +                * scream and skip that CPU.
>> +                */
>> +               if (WARN_ON((mpidr & 0xff) >= 16))
>> +                       goto out;
> 
> I'm not comfortable with this. The mpidr in ARMv8 is just a unique
> identifier, so relying on aff0 to provide some sort of upper bound on the
> number of CPUs in a cluster feels extremely fragile.

And yet this is the only thing the GICv3 architecture supports. I've
brought this up ages ago, and been told that "it is all fine". See the
note at the bottom of 5.7.29: "this restricts distribution of SGIs to
the first 16 processors of an affinity 1 cluster".

> I think you'll need to deal with SGIs to target lists that you can't
> represent with a single sgi1r value (and instead batch them up and have
> lots of fun dealing with the result).

That's what the code already does. It is just that some processors may
be out of the reach of GICv3 altogether (those with Aff0 >= 16).

>> +
>> +               tlist |= 1 << (mpidr & 0xf);
>> +
>> +               cpu = cpumask_next(cpu, mask);
>> +               mpidr = cpu_logical_map(cpu);
>> +
>> +               if (cluster_id != (mpidr & ~0xffUL)) {
>> +                       cpu--;
>> +                       goto out;
>> +               }
>> +       }
>> +out:
>> +       *base_cpu = cpu;
>> +       return tlist;
>> +}
>> +
>> +static void gic_send_sgi(u64 cluster_id, u16 tlist, unsigned int irq)
>> +{
>> +       u64 val;
>> +
>> +       val  = (cluster_id & 0xff00ff0000UL) << 16; /* Aff3 + Aff2 */
>> +       val |= (cluster_id & 0xff00) << 8;          /* Aff1 */
> 
> Use the helpers.

Yup.

>> +       val |= irq << 24;
>> +       val |= tlist;
>> +
>> +       pr_debug("CPU%d: ICC_SGI1R_EL1 %llx\n", smp_processor_id(), val);
>> +       gic_write_sgi1r(val);
>> +}
>> +
>> +static void gic_raise_softirq(const struct cpumask *mask, unsigned int irq)
>> +{
>> +       int cpu;
>> +
>> +       if (WARN_ON(irq >= 16))
>> +               return;
>> +
>> +       /*
>> +        * Ensure that stores to Normal memory are visible to the
>> +        * other CPUs before issuing the IPI.
>> +        */
>> +       dsb();
> 
> dmb(ishst), obviously (!) ;)

Indeed ;-).

>> +
>> +       for_each_cpu_mask(cpu, *mask) {
>> +               u64 cluster_id = cpu_logical_map(cpu) & ~0xffUL;
>> +               u16 tlist;
>> +
>> +               tlist = gic_compute_target_list(&cpu, mask, cluster_id);
>> +               gic_send_sgi(cluster_id, tlist, irq);
>> +       }
>> +}
>> +
>> +static void gic_smp_init(void)
>> +{
>> +       set_smp_cross_call(gic_raise_softirq);
>> +       register_cpu_notifier(&gic_cpu_notifier);
>> +}
>> +
>> +static int gic_set_affinity(struct irq_data *d, const struct cpumask *mask_val,
>> +                           bool force)
>> +{
>> +       unsigned int cpu = cpumask_any_and(mask_val, cpu_online_mask);
>> +       void __iomem *reg;
>> +       u64 val;
>> +
>> +       if (gic_irq(d) < 32)
>> +               return -EINVAL;
>> +
>> +       reg = gic_dist_base(d) + GICD_IROUTER + (gic_irq(d) * 8);
>> +       val = gic_mpidr_to_affinity(cpu_logical_map(cpu));
>> +
>> +       writeq_relaxed(val, reg);
> 
> How do you ensure that the interrupt has actually moved?

There is no way to ensure it actually did (other than reading back from
the same register, if that's what you're aiming for?).

>> +       return IRQ_SET_MASK_OK;
>> +}
>> +#else
>> +#define gic_set_affinity       NULL
>> +#define gic_smp_init()         do { } while(0)
>> +#endif
>> +
>> +static struct irq_chip gic_chip = {
>> +       .name                   = "GICv3",
>> +       .irq_mask               = gic_mask_irq,
>> +       .irq_unmask             = gic_unmask_irq,
>> +       .irq_eoi                = gic_eoi_irq,
>> +       .irq_set_type           = gic_set_type,
>> +       .irq_retrigger          = gic_retrigger,
>> +       .irq_set_affinity       = gic_set_affinity,
>> +};
>> +
>> +static int gic_irq_domain_map(struct irq_domain *d, unsigned int irq,
>> +                               irq_hw_number_t hw)
>> +{
>> +       if (hw < 32) {
>> +               irq_set_percpu_devid(irq);
>> +               irq_set_chip_and_handler(irq, &gic_chip,
>> +                                        handle_percpu_devid_irq);
>> +               set_irq_flags(irq, IRQF_VALID | IRQF_NOAUTOEN);
>> +       } else {
>> +               irq_set_chip_and_handler(irq, &gic_chip,
>> +                                        handle_fasteoi_irq);
>> +               set_irq_flags(irq, IRQF_VALID | IRQF_PROBE);
>> +       }
>> +       irq_set_chip_data(irq, d->host_data);
>> +       return 0;
>> +}
> 
> More copied GICv2 code...

I'll move it out, then.

>> +static int gic_irq_domain_xlate(struct irq_domain *d,
>> +                               struct device_node *controller,
>> +                               const u32 *intspec, unsigned int intsize,
>> +                               unsigned long *out_hwirq, unsigned int *out_type)
>> +{
>> +       if (d->of_node != controller)
>> +               return -EINVAL;
>> +       if (intsize < 3)
>> +               return -EINVAL;
>> +
>> +       switch(intspec[0]) {
>> +       case 0:                 /* SPI */
>> +               *out_hwirq = intspec[1] + 32;
>> +               break;
>> +       case 1:                 /* PPI */
>> +               *out_hwirq = intspec[1] + 16;
>> +               break;
>> +       default:
>> +               return -EINVAL;
>> +       }
>> +
>> +       *out_type = intspec[2] & IRQ_TYPE_SENSE_MASK;
>> +       return 0;
>> +}
> 
> ... and some slightly mutated but functionally identical GICv2 code ;)

Same.

>> +static const struct irq_domain_ops gic_irq_domain_ops = {
>> +       .map = gic_irq_domain_map,
>> +       .xlate = gic_irq_domain_xlate,
>> +};
>> +
>> +static int __init gic_of_init(struct device_node *node, struct device_node *parent)
>> +{
>> +       void __iomem *dist_base;
>> +       void __iomem **redist_base;
>> +       u64 redist_stride;
>> +       u32 redist_regions;
>> +       u32 reg;
>> +       int gic_irqs;
>> +       int err;
>> +       int i;
>> +
>> +       dist_base = of_iomap(node, 0);
>> +       if (!dist_base) {
>> +               pr_warn("%s: unable to map gic dist registers\n",
>> +                       node->full_name);
>> +               return -ENXIO;
>> +       }
>> +
>> +       reg = readl_relaxed(dist_base + GICD_PIDR2) & GIC_PIDR2_ARCH_MASK;
>> +       if (reg != 0x30 && reg != 0x40) {
>> +               pr_warn("%s: no distributor detected, giving up\n",
>> +                       node->full_name);
>> +               err = -ENODEV;
>> +               goto out_unmap_dist;
>> +       }
>> +
>> +       if (of_property_read_u32(node, "#redistributor-regions", &redist_regions))
>> +               redist_regions = 1;
>> +
>> +       redist_base = kzalloc(sizeof(*redist_base) * redist_regions, GFP_KERNEL);
>> +       if (!redist_base) {
>> +               err = -ENOMEM;
>> +               goto out_unmap_dist;
>> +       }
> 
> Can you use the devm_* managed resource helpers in irqchip init functions? I
> see a couple of drivers doing it, but it's not clear to me whether things
> are actually freed on failure. Anyway, if you can, then it would clean this
> up a bit.

The device model is not available at the stage where we probe the GIC.
If it was, I'd have converted the GIC to be a proper driver a long time
ago...

The only two users are secondary interrupt controllers, which are probed
through DT, when the whole infrastructure is up and running.

>> +       for (i = 0; i < redist_regions; i++) {
>> +               redist_base[i] = of_iomap(node, 1 + i);
>> +               if (!redist_base[i]) {
>> +                       pr_warn("%s: couldn't map region %d\n",
>> +                               node->full_name, i);
>> +                       err = -ENODEV;
>> +                       goto out_unmap_rdist;
>> +               }
>> +       }
>> +
>> +       if (of_property_read_u64(node, "redistributor-stride", &redist_stride))
>> +               redist_stride = 0;
>> +
>> +       gic_data.dist_base = dist_base;
>> +       gic_data.redist_base = redist_base;
>> +       gic_data.redist_regions = redist_regions;
>> +       gic_data.redist_stride = redist_stride;
>> +
>> +       /*
>> +        * Find out how many interrupts are supported.
>> +        * The GIC only supports up to 1020 interrupt sources (SGI+PPI+SPI)
>> +        */
>> +       gic_irqs = readl_relaxed(gic_data.dist_base + GICD_TYPER) & 0x1f;
>> +       gic_irqs = (gic_irqs + 1) * 32;
>> +       if (gic_irqs > 1020)
>> +               gic_irqs = 1020;
>> +       gic_data.irq_nr = gic_irqs;
>> +
>> +       gic_data.domain = irq_domain_add_linear(node, gic_irqs - 16,
>> +                                               &gic_irq_domain_ops, &gic_data);
>> +       gic_data.rdist = alloc_percpu(typeof(*gic_data.rdist));
>> +
>> +       if (WARN_ON(!gic_data.domain) || WARN_ON(!gic_data.rdist)) {
>> +               err = -ENOMEM;
>> +               goto out_free;
>> +       }
>> +
>> +       set_handle_irq(gic_handle_irq);
>> +
>> +       gic_smp_init();
>> +       gic_dist_init();
>> +       gic_cpu_init();
>> +
>> +       return 0;
>> +
>> +out_free:
> 
> remove the irq_domain here?

Yup, good point.

>> +       free_percpu(gic_data.rdist);
>> +out_unmap_rdist:
>> +       for (i = 0; i < redist_regions; i++)
>> +               if (redist_base[i])
>> +                       iounmap(redist_base[i]);
>> +       kfree(redist_base);
>> +out_unmap_dist:
>> +       iounmap(dist_base);
>> +       return err;
>> +}
>> +
>> +IRQCHIP_DECLARE(gic_v3, "arm,gic-v3", gic_of_init);
>> diff --git a/include/linux/irqchip/arm-gic-v3.h b/include/linux/irqchip/arm-gic-v3.h
>> new file mode 100644
>> index 0000000..e4d060a
>> --- /dev/null
>> +++ b/include/linux/irqchip/arm-gic-v3.h
>> @@ -0,0 +1,190 @@
>> +/*
>> + * Copyright (C) 2013 ARM Limited, All Rights Reserved.
>> + * Author: Marc Zyngier <marc.zyngier at arm.com>
>> + *
>> + *
>> + * This program is free software; you can redistribute it and/or modify
>> + * it under the terms of the GNU General Public License version 2 as
>> + * published by the Free Software Foundation.
>> + *
>> + * This program is distributed in the hope that it will be useful,
>> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
>> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
>> + * GNU General Public License for more details.
>> + *
>> + * You should have received a copy of the GNU General Public License
>> + * along with this program.  If not, see <http://www.gnu.org/licenses/>.
>> + */
>> +#ifndef __LINUX_IRQCHIP_ARM_GIC_V3_H
>> +#define __LINUX_IRQCHIP_ARM_GIC_V3_H
>> +
>> +/*
>> + * Distributor registers. We assume we're running non-secure, with ARE
>> + * being set. Secure-only and non-ARE registers are not described.
>> + */
>> +#define GICD_CTLR                      0x0000
>> +#define GICD_TYPER                     0x0004
>> +#define GICD_IIDR                      0x0008
>> +#define GICD_STATUSR                   0x0010
>> +#define GICD_SETSPI_NSR                        0x0040
>> +#define GICD_CLRSPI_NSR                        0x0048
>> +#define GICD_SETSPI_SR                 0x0050
>> +#define GICD_CLRSPI_SR                 0x0058
>> +#define GICD_SEIR                      0x0068
>> +#define GICD_ISENABLER                 0x0100
>> +#define GICD_ICENABLER                 0x0180
>> +#define GICD_ISPENDR                   0x0200
>> +#define GICD_ICPENDR                   0x0280
>> +#define GICD_ISACTIVER                 0x0300
>> +#define GICD_ICACTIVER                 0x0380
>> +#define GICD_IPRIORITYR                        0x0400
>> +#define GICD_ICFGR                     0x0C00
>> +#define GICD_IROUTER                   0x6000
>> +#define GICD_PIDR2                     0xFFE8
>> +
>> +#define GICD_CTLR_RWP                  (1U << 31)
>> +#define GICD_CTLR_ARE_NS               (1U << 4)
>> +#define GICD_CTLR_ENABLE_G1A           (1U << 1)
>> +#define GICD_CTLR_ENABLE_G1            (1U << 0)
>> +
>> +#define GICD_IROUTER_SPI_MODE_ONE      (0U << 31)
>> +#define GICD_IROUTER_SPI_MODE_ANY      (1U << 31)
>> +
>> +#define GIC_PIDR2_ARCH_MASK            0xf0
>> +
>> +/*
>> + * Re-Distributor registers, offsets from RD_base
>> + */
>> +#define GICR_CTLR                      GICD_CTLR
>> +#define GICR_IIDR                      0x0004
>> +#define GICR_TYPER                     0x0008
>> +#define GICR_STATUSR                   GICD_STATUSR
>> +#define GICR_WAKER                     0x0014
>> +#define GICR_SETLPIR                   0x0040
>> +#define GICR_CLRLPIR                   0x0048
>> +#define GICR_SEIR                      GICD_SEIR
>> +#define GICR_PROPBASER                 0x0070
>> +#define GICR_PENDBASER                 0x0078
>> +#define GICR_INVLPIR                   0x00A0
>> +#define GICR_INVALLR                   0x00B0
>> +#define GICR_SYNCR                     0x00C0
>> +#define GICR_MOVLPIR                   0x0100
>> +#define GICR_MOVALLR                   0x0110
>> +#define GICR_PIDR2                     GICD_PIDR2
>> +
>> +#define GICR_WAKER_ProcessorSleep      (1U << 1)
>> +#define GICR_WAKER_ChildrenAsleep      (1U << 2)
>> +
>> +/*
>> + * Re-Distributor registers, offsets from SGI_base
>> + */
>> +#define GICR_ISENABLER0                        GICD_ISENABLER
>> +#define GICR_ICENABLER0                        GICD_ICENABLER
>> +#define GICR_ISPENDR0                  GICD_ISPENDR
>> +#define GICR_ICPENDR0                  GICD_ICPENDR
>> +#define GICR_ISACTIVER0                        GICD_ISACTIVER
>> +#define GICR_ICACTIVER0                        GICD_ICACTIVER
>> +#define GICR_IPRIORITYR0               GICD_IPRIORITYR
>> +#define GICR_ICFGR0                    GICD_ICFGR
>> +
>> +#define GICR_TYPER_VLPIS               (1U << 1)
>> +#define GICR_TYPER_LAST                        (1U << 4)
>> +
>> +/*
>> + * CPU interface registers
>> + */
>> +#define GICC_CTLR_EL1_EOImode_drop_dir (0U << 1)
>> +#define GICC_CTLR_EL1_EOImode_drop     (1U << 1)
>> +#define GICC_SRE_EL1_SRE               (1U << 0)
>> +
>> +/*
>> + * Hypervisor interface registers (SRE only)
>> + */
>> +#define GICH_LR_VIRTUAL_ID_MASK                ((1UL << 32) - 1)
>> +
>> +#define GICH_LR_EOI                    (1UL << 41)
>> +#define GICH_LR_GROUP                  (1UL << 60)
>> +#define GICH_LR_STATE                  (3UL << 62)
>> +#define GICH_LR_PENDING_BIT            (1UL << 62)
>> +#define GICH_LR_ACTIVE_BIT             (1UL << 63)
>> +
>> +#define GICH_MISR_EOI                  (1 << 0)
>> +#define GICH_MISR_U                    (1 << 1)
>> +
>> +#define GICH_HCR_EN                    (1 << 0)
>> +#define GICH_HCR_UIE                   (1 << 1)
>> +
>> +#define GICH_VMCR_CTLR_SHIFT           0
>> +#define GICH_VMCR_CTLR_MASK            (0x21f << GICH_VMCR_CTLR_SHIFT)
>> +#define GICH_VMCR_BPR1_SHIFT           18
>> +#define GICH_VMCR_BPR1_MASK            (7 << GICH_VMCR_BPR1_SHIFT)
>> +#define GICH_VMCR_BPR0_SHIFT           21
>> +#define GICH_VMCR_BPR0_MASK            (7 << GICH_VMCR_BPR0_SHIFT)
>> +#define GICH_VMCR_PMR_SHIFT            24
>> +#define GICH_VMCR_PMR_MASK             (0xffUL << GICH_VMCR_PMR_SHIFT)
>> +
>> +#define ICC_EOIR1_EL1                  S3_0_C12_C12_1
>> +#define ICC_IAR1_EL1                   S3_0_C12_C12_0
>> +#define ICC_SGI1R_EL1                  S3_0_C12_C11_5
>> +#define ICC_PMR_EL1                    S3_0_C4_C6_0
>> +#define ICC_CTLR_EL1                   S3_0_C12_C12_4
>> +#define ICC_SRE_EL1                    S3_0_C12_C12_5
>> +#define ICC_GRPEN1_EL1                 S3_0_C12_C12_7
>> +
>> +#define ICC_SRE_EL2                    S3_4_C12_C9_5
>> +
>> +#define ICH_VSEIR_EL2                  S3_4_C12_C9_4
>> +#define ICH_HCR_EL2                    S3_4_C12_C11_0
>> +#define ICH_VTR_EL2                    S3_4_C12_C11_1
>> +#define ICH_MISR_EL2                   S3_4_C12_C11_2
>> +#define ICH_EISR_EL2                   S3_4_C12_C11_3
>> +#define ICH_ELSR_EL2                   S3_4_C12_C11_5
>> +#define ICH_VMCR_EL2                   S3_4_C12_C11_7
>> +
>> +#define __LR0_EL2(x)                   S3_4_C12_C12_ ## x
>> +#define __LR8_EL2(x)                   S3_4_C12_C13_ ## x
>> +
>> +#define ICH_LR0_EL2                    __LR0_EL2(0)
>> +#define ICH_LR1_EL2                    __LR0_EL2(1)
>> +#define ICH_LR2_EL2                    __LR0_EL2(2)
>> +#define ICH_LR3_EL2                    __LR0_EL2(3)
>> +#define ICH_LR4_EL2                    __LR0_EL2(4)
>> +#define ICH_LR5_EL2                    __LR0_EL2(5)
>> +#define ICH_LR6_EL2                    __LR0_EL2(6)
>> +#define ICH_LR7_EL2                    __LR0_EL2(7)
>> +#define ICH_LR8_EL2                    __LR8_EL2(0)
>> +#define ICH_LR9_EL2                    __LR8_EL2(1)
>> +#define ICH_LR10_EL2                   __LR8_EL2(2)
>> +#define ICH_LR11_EL2                   __LR8_EL2(3)
>> +#define ICH_LR12_EL2                   __LR8_EL2(4)
>> +#define ICH_LR13_EL2                   __LR8_EL2(5)
>> +#define ICH_LR14_EL2                   __LR8_EL2(6)
>> +#define ICH_LR15_EL2                   __LR8_EL2(7)
>> +
>> +#define __AP0Rx_EL2(x)                 S3_4_C12_C8_ ## x
>> +#define ICH_AP0R0_EL2                  __AP0Rx_EL2(0)
>> +#define ICH_AP0R1_EL2                  __AP0Rx_EL2(1)
>> +#define ICH_AP0R2_EL2                  __AP0Rx_EL2(2)
>> +#define ICH_AP0R3_EL2                  __AP0Rx_EL2(3)
>> +
>> +#define __AP1Rx_EL2(x)                 S3_4_C12_C9_ ## x
>> +#define ICH_AP1R0_EL2                  __AP1Rx_EL2(0)
>> +#define ICH_AP1R1_EL2                  __AP1Rx_EL2(1)
>> +#define ICH_AP1R2_EL2                  __AP1Rx_EL2(2)
>> +#define ICH_AP1R3_EL2                  __AP1Rx_EL2(3)
> 
> I'm going to rely on Christoffer having looked at all these magic numbers
> (he at least had ice cream!).

Quite! ;-)

>> +#ifndef __ASSEMBLY__
>> +
>> +#include <linux/stringify.h>
>> +
>> +#define reg(x)         __stringify(x)
>> +
>> +static inline void gic_write_eoir(u64 irq)
>> +{
>> +       asm volatile("msr " reg(ICC_EOIR1_EL1) ", %0" : : "r" (irq));
>> +       isb();
>> +}
> 
> Why is this in the header file?

For the benefit of the ITS part of the driver, which needs to EOI its
own stuff.

Thanks for the excellent review!

	M.
-- 
Jazz is not dead. It just smells funny...



More information about the linux-arm-kernel mailing list