On 11/11/2026 8:31 AM, Jim MacArthur wrote:
> On Fri, Nov 06, 2026 at 11:21:12PM +0100, Gustavo Romero wrote:
>> Hi Jim,
>>
>> On Fri, Oct 16, 2026 at 10:40:52AM +0100, Jim MacArthur wrote:
>>> In our implementation, SMMU Root registers are directly after the
>>> non-optional SMMU registers. Realms registers will be added above the
>>> root registers later.
>>
>> nit: Root registers
> 
> Noted.
> 
>>
>>
>>>
>>> Signed-off-by: Jim MacArthur <[email protected]>
>>> ---
>>>  hw/arm/sbsa-ref.c | 2 +-
>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/hw/arm/sbsa-ref.c b/hw/arm/sbsa-ref.c
>>> index 158b461eb4..0747f5ca9f 100644
>>> --- a/hw/arm/sbsa-ref.c
>>> +++ b/hw/arm/sbsa-ref.c
>>> @@ -129,7 +129,7 @@ static const MemMapEntry sbsa_ref_memmap[] = {
>>>      [SBSA_GPIO] =               { 0x60020000, 0x00001000 },
>>>      [SBSA_SECURE_UART] =        { 0x60030000, 0x00001000 },
>>>      [SBSA_SECURE_UART_MM] =     { 0x60040000, 0x00001000 },
>>> -    [SBSA_SMMU] =               { 0x60050000, 0x00020000 },
>>> +    [SBSA_SMMU] =               { 0x60050000, 0x00040000 },
>>
>> Why +0x20000 bytes here instead of +0x10000 bytes (1 x 64K page)?
> 
> Ultimately the plan is to add the Realm registers in here as well, and
> those have to be at a 128K-aligned offset from the SMMU base, so per the
> original plan they would have started at 0x40000. Now that you mention it
> though, it makes more sense to slot the Realm registers in at offset 0x20000
> and move the Root registers to 0x40000 when we do so - which means we still
> have space in the memory map for a second SMMU. So for now, I'll trim
> this back to +0x10000.
> 
>>
>> Most importantly, incresing the size of the SBSA_SMMU index is not enough, we
>> need to increase also the size of the "arm-smmuv3" container (memory region).
>>
>> With this change applied I still see:
>>
>> 0000000060050000-000000006006ffff (prio 0, i/o): arm-smmuv3
> 
> Thanks, I hadn't spotted that the actual allocation was specified separately;
> I'll fix that in the next version.
> 
>>
>> on (qemu) info mtree.
>>
>> 2 x 64K pages: Page 0 and Page 1. Rather we should see a range to fit 3 x 
>> 64K:
>>
>> 0000000060050000-000000006007ffff (prio 0, i/o): arm-smmuv3
>>
>> Another issue, afaics, is that the Root register page will be added to
>> every machine that uses the SMMUv3 device, right? Because we're not gating 
>> the
>> addition of Root registers based on the presence of RME.
>>
>> So I think we should add one boolean "x-root" property (flag) to the SMMUv3
>> device that will be used to increase the size of the "arm-smmuv3" if it's set
>> and let it by default set to "false".
>>
>> This "x-root" property would be set at the machine level, so via
>> object_property_set_bool() in create_smmu. For example:
>>
>> diff --git a/hw/arm/sbsa-ref.c b/hw/arm/sbsa-ref.c
>> index 0747f5ca9f..36b0164510 100644
>> --- a/hw/arm/sbsa-ref.c
>> +++ b/hw/arm/sbsa-ref.c
>> @@ -630,6 +630,7 @@ static void create_smmu(const SBSAMachineState *sms, 
>> PCIBus *bus,
>>                               &error_abort);
>>      object_property_set_link(OBJECT(dev), "secure-memory", 
>> OBJECT(secure_sysmem),
>>                               &error_abort);
>> +    object_property_set_bool(OBJECT(dev), "x-root", true, &error_abort);
>>      sysbus_realize_and_unref(SYS_BUS_DEVICE(dev), &error_fatal);
>>      sysbus_mmio_map(SYS_BUS_DEVICE(dev), 0, base);
>>      for (i = 0; i < NUM_SMMU_IRQS; i++) {
>> diff --git a/hw/arm/smmuv3.c b/hw/arm/smmuv3.c
>> index 9d9bc2601b..9912fbf920 100644
>> --- a/hw/arm/smmuv3.c
>> +++ b/hw/arm/smmuv3.c
>> @@ -2267,6 +2267,7 @@ static void smmu_realize(DeviceState *d, Error **errp)
>>      SMMUv3Class *c = ARM_SMMUV3_GET_CLASS(s);
>>      SysBusDevice *dev = SYS_BUS_DEVICE(d);
>>      Error *local_err = NULL;
>> +    uint64_t size = SMMUV3_REGS_SIZE;
>>  
>>      if (!smmu_validate_property(s, errp)) {
>>          return;
>> @@ -2291,8 +2292,12 @@ static void smmu_realize(DeviceState *d, Error **errp)
>>  
>>      qemu_mutex_init(&s->mutex);
>>  
>> +    if (s->has_root) {
>> +        size += SMMUV3_ROOT_PAGE_SIZE;
>> +    }
>> +
>>      memory_region_init_io(&sys->iomem, OBJECT(s),
>> -                          &smmu_mem_ops, sys, TYPE_ARM_SMMUV3, 0x20000);
>> +                          &smmu_mem_ops, sys, TYPE_ARM_SMMUV3, size);
>>  
>>      sys->mrtypename = TYPE_SMMUV3_IOMMU_MEMORY_REGION;
>>  
>> @@ -2382,6 +2387,7 @@ static const Property smmuv3_properties[] = {
>>      /* Identifier used for ACPI IORT SMMUv3 (and DSDT for CMDQV) generation 
>> */
>>      DEFINE_PROP_UINT8("identifier", SMMUv3State, identifier, 0),
>>      DEFINE_PROP_BOOL("accel", SMMUv3State, accel, false),
>> +    DEFINE_PROP_BOOL("x-root", SMMUv3State, has_root, false),
>>      /* GPA of MSI doorbell, for SMMUv3 accel use. */
>>      DEFINE_PROP_UINT64("msi-gpa", SMMUv3State, msi_gpa, 0),
>>      /*
>> diff --git a/include/hw/arm/smmuv3-common.h b/include/hw/arm/smmuv3-common.h
>> index a538334066..5d1cb759c3 100644
>> --- a/include/hw/arm/smmuv3-common.h
>> +++ b/include/hw/arm/smmuv3-common.h
>> @@ -411,7 +411,10 @@ REG32(EVENTQ_IRQ_CFG2,     0xbc)
>>  
>>  #define A_IDREGS           0xfd0
>>  
>> -#define SMMU_ROOT_REG_START 0x20000
>> +#define SMMUV3_PAGE_SIZE        0x10000
>> +#define SMMUV3_REGS_SIZE        (2 * SMMUV3_PAGE_SIZE) /* Page 0 and Page 1 
>> */
>> +#define SMMUV3_ROOT_REG_OFFSET  0x20000
>> +#define SMMUV3_ROOT_PAGE_SIZE   SMMUV3_PAGE_SIZE
>>  
>>  REG32(ROOT_IDR0,            0x0)
>>      FIELD(ROOT_IDR0, ROOT_IMPL,   0,  1)
>> diff --git a/include/hw/arm/smmuv3.h b/include/hw/arm/smmuv3.h
>> index e2708e13fe..a8b354438c 100644
>> --- a/include/hw/arm/smmuv3.h
>> +++ b/include/hw/arm/smmuv3.h
>> @@ -92,6 +92,7 @@ struct SMMUv3State {
>>      OnOffAuto cmdqv;
>>  
>>      Notifier machine_done;
>> +    bool has_root;
>>      SMMUv3RootRegBank root;
>>  };
>>
>>
>> Cheers,
>> Gustavo
> 
> Thank you for doing this. The only complication is that we'll probably want
> an x-realms flag as well later, and if we do the rearrangement of the map
> as above then SMMUV3_ROOT_REG_OFFSET will have to be variable. I'll handle
> that when we get to the Realm part.
>

The biggest different with realms is the presence of GPC, for which the
GPT has to be configured correctly to allow accesses.
However, in TF-A, this is gated behind cpu support for realms, which
also require setting up GPT.

```
/* drivers/arm/smmu/smmu_v3.c */
        if (is_feat_rme_present()) {
                if ((mmio_read_32(smmu_base + SMMU_ROOT_IDR0) &
                                  SMMU_ROOT_IDR0_ROOT_IMPL) == 0U) {
                        WARN("Skip SMMU GPC configuration.\n");
                } else {
                        uint64_t gpccr_el3 = read_gpccr_el3();
                        uint64_t gptbr_el3 = read_gptbr_el3();

                        /* SMMU_ROOT_GPT_BASE_CFG[16] is RES0. */
                        gpccr_el3 &= ~(1UL << 16);
                        mmio_write_32(smmu_base +SMMU_ROOT_GPT_BASE_CFG,
                                      gpccr_el3);
                        mmio_write_64(smmu_base + SMMU_ROOT_GPT_BASE,
                                      gptbr_el3 << 12);
                        mmio_setbits_32(smmu_base + SMMU_ROOT_CR0,
                                        SMMU_ROOT_CR0_GPCEN);
                }
        }
```

As a result, once both will enabled (root + realm), I think we can
enable both by default, as they will be automatically depend on cpu
support for realms. For now, root only would have to be false by default
though.

At the time, I could also run all our tests successfully, without having
to do any modification.

> Jim
> 

Regards,
Pierrick

Reply via email to