On 10/8/26 11:45 AM, Khushit Shah wrote:
>
>> On 8 Oct 2026, at 1:22 PM, Eric Auger <[email protected]> wrote:
>>
>> !-------------------------------------------------------------------|
>> CAUTION: External Email
>>
>> |-------------------------------------------------------------------!
>>
>>
>>
>> On 9/30/26 9:54 AM, Khushit Shah wrote:
>>>> On 16 Sep 2026, at 8:15 PM, Eric Auger <[email protected]> wrote:
>>>>
>>>> !-------------------------------------------------------------------|
>>>> CAUTION: External Email
>>>>
>>>> |-------------------------------------------------------------------!
>>>>
>>>> SYSREG prop settings were recorded in a hashtable but they
>>>> are not yet injected on guest. Iterate over all set fields and
>>>> apply the new value through KVM_SET_ONE_REG. The new field
>>>> value overrides the current observable vcpu register value.
>>>>
>>>> Signed-off-by: Eric Auger <[email protected]>
>>>> ---
>>>> target/arm/kvm.c | 65 +++++++++++++++++++++++++++++++++++++++++
>>>> target/arm/trace-events | 1 +
>>>> 2 files changed, 66 insertions(+)
>>>>
>>>> diff --git a/target/arm/kvm.c b/target/arm/kvm.c
>>>> index 3ec2618cc1..ddf320d0e6 100644
>>>> --- a/target/arm/kvm.c
>>>> +++ b/target/arm/kvm.c
>>>> @@ -343,6 +343,9 @@ static ARM64SysRegField *get_field(int i, ARM64SysReg
>>>> *reg)
>>>> #define MAKE_IDREG_KEY(reg_idx, field_shift) \
>>>> (((uint64_t)(reg_idx) << 8) | ((uint64_t)(field_shift) & 0xFF))
>>>>
>>>> +#define KEY_TO_REG_IDX(key) ((uint32_t)(((uint64_t)(key)) >> 8))
>>>> +#define KEY_TO_SHIFT(key) ((uint8_t)(((uint64_t)(key)) & 0xFF))
>>>> +
>>>> static void set_sysreg_prop(Object *obj, Visitor *v,
>>>> const char *name, void *opaque,
>>>> Error **errp)
>>>> @@ -2215,6 +2218,63 @@ int kvm_arch_pre_create_vcpu(CPUState *cpu, Error
>>>> **errp)
>>>> return 0;
>>>> }
>>>>
>>>> +static int kvm_arm_apply_sysreg_props(ARMCPU *cpu, Error **errp)
>>>> +{
>>>> + uint64_t *idregs = cpu->isar.idregs;
>>>> + gpointer key_ptr, value_ptr;
>>>> + CPUState *cs = CPU(cpu);
>>>> + GHashTableIter iter;
>>>> +
>>>> + g_hash_table_iter_init(&iter, cpu->sysreg_props);
>>>> +
>>>> + while (g_hash_table_iter_next(&iter, &key_ptr, &value_ptr)) {
>>>> + uint64_t key = (uint64_t)key_ptr;
>>>> + uint64_t value = (uint64_t)value_ptr;
>>>> + uint32_t reg_idx = KEY_TO_REG_IDX(key);
>>>> + uint8_t lower = KEY_TO_SHIFT(key);
>>>> + ARM64SysReg *reg = &arm64_id_regs[reg_idx];
>>>> + struct kvm_one_reg kvm_reg;
>>>> + uint64_t current, old, mask;
>>>> + uint64_t kvm_idx;
>>>> + ARM64SysRegField *field = get_field(lower, reg);
>>>> + int length = field->length;
>>>> + uint64_t oldfv;
>>>> + int ret;
>>>> +
>>>> + mask = MAKE_64BIT_MASK(lower, length);
>>>> + value = value << lower;
>>>> +
>>>> + old = idregs[reg_idx];
>>>> +
>>>> + kvm_idx = idregs_sysreg_to_kvm_reg(id_register_sysreg[reg_idx]);
>>>> + ret = read_sys_reg64(cs->kvm_fd, ¤t, kvm_idx);
>>>> + if (ret) {
>>>> + error_setg(errp, "failed to read the current value of %s",
>>>> + reg->name);
>>>> + return ret;
>>>> + }
>>>> + oldfv = (current & mask) >> lower;
>>>> +
>>>> + idregs[reg_idx] = current & ~mask;
>>>> + idregs[reg_idx] |= value;
>>>> +
>>>> + kvm_reg.id = kvm_idx;
>>>> + kvm_reg.addr = (uintptr_t)&idregs[reg_idx];
>>>> +
>>>> + ret = ioctl(cs->kvm_fd, KVM_SET_ONE_REG, &kvm_reg);
>>>> + if (ret) {
>>>> + error_setg(errp, "failed to apply new value 0x%"PRIx64" for
>>>> field %s.%s "
>>>> + "(previous is 0x%"PRIx64"): %m", idregs[reg_idx],
>>>> + reg->name, field->name, oldfv);
>>>> + return ret;
>>>> + }
>>> Hi Eric,
>>>
>>> The above is the only KVM-related code here. Can we split this into
>>> 1. Apply SYSREG_ props callback which resides in cpu64.c?
>>> 2. The previous writeback to cpregs function that writes back ID
>>> registers to KVM?
>> the problem is the previous method was copying all the idregs that
>> changed back to cpregs. But those idregs were populated from the "max"
>> vcpu which in some cases led to inconsistency with the final vcpu
>> setting. The setter now registers the new field value in a hash table.
>> Maybe we can just rely on an arch callback to apply them on the guest
>>
> Yes, this is what I am suggesting.
>
> What I was thinking is a KVM-specific code (kvm_arch_init_vcpu) will call
> callback to cpu64.c to set the property values after the vCPUs are init, and
> after the callback finishes, the values are written back to cpregs.
>
> With this the inconsistency problem does not arise as we will only call the
> callback after we properly setup the vCPU.
I have moved most of the code in cpu64, ie. hash table lookup and
cpu->isar.idregs[] update on success. The functional uses an
acceleration specific callback to apply the field value on guest vcpu.
However the arch agnostic function is still called in kvm so that the
kvmlist is synced to cpregs afterwards.
Let's see in v10 if it matches your expectations
Thanks
Eric
>
> FWIU, you are suggesting the same thing?
>
> Thanks,
> Khushit
>> Thanks
>>
>> Eric
>>> The reason being the SYSREG_ props otherwise do not work with any other
>>> accelerator.
>>> We are already trying to make the SYSREG_ getter/setter accel-agnostic.
>>>
>>> The above method still has the nicety that -cpu host without any SYSREG_
>>> prop should
>>> behave the same as previous :)
>>>
>>> Side note:
>>> While we don’t target TCG support currently, I tried your v8 + named models
>>> on TCG
>>> (quick hack; ~20 LOC), and SYSREG_ props/named models works almost out of
>>> the box
>>> there (very minimal testing), which was cool.
>>>
>>> Thanks,
>>> Khushit
>>>
>>>> +
>>>> + trace_apply_sysreg_prop(reg->name, field->name, old, current,
>>>> + mask, value, idregs[reg_idx]);
>>>> + }
>>>> + return 0;
>>>> +}
>>>> +
>>>> int kvm_arch_init_vcpu(CPUState *cs, Error **errp)
>>>> {
>>>> int ret;
>>>> @@ -2315,6 +2375,11 @@ int kvm_arch_init_vcpu(CPUState *cs, Error **errp)
>>>> }
>>>> cpu->mp_affinity = mpidr & ARM64_AFFINITY_MASK;
>>>>
>>>> + ret = kvm_arm_apply_sysreg_props(cpu, errp);
>>>> + if (ret) {
>>>> + return ret;
>>>> + }
>>>> +
>>>> return kvm_arm_init_cpreg_list(cpu);
>>>> }
>>>>
>>>> diff --git a/target/arm/trace-events b/target/arm/trace-events
>>>> index d72ad6b671..b4d2b45a7d 100644
>>>> --- a/target/arm/trace-events
>>>> +++ b/target/arm/trace-events
>>>> @@ -18,6 +18,7 @@ kvm_arm_writable_idregs_to_cpreg_list(const char *name,
>>>> uint64_t previous, uint6
>>>> decode_idreg_writemap(const char* name, int lower, int upper, char
>>>> *prop_name) "%s [%d:%d] is writable (prop %s)"
>>>> get_sysreg_prop(const char *name, uint64_t value) "%s 0x%"PRIx64
>>>> set_sysreg_prop(const char *name, uint64_t old, uint64_t mask, uint64_t
>>>> field_value, uint64_t new) "%s old reg value=0x%"PRIx64" mask=0x%"PRIx64"
>>>> new field value=0x%"PRIx64" new reg value=0x%"PRIx64
>>>> +apply_sysreg_prop(const char *reg_name, const char *field_name, uint64_t
>>>> old, uint64_t current, uint64_t mask, uint64_t field_value, uint64_t new)
>>>> "%s.%s old=0x%"PRIx64" current=0x%"PRIx64" mask=0x%"PRIx64" new field
>>>> value=0x%"PRIx64" new=0x%"PRIx64
>>>> nb_sysreg_props(const char *name, int count) "%s: %d SYSREG properties"
>>>>
>>>> # cpu.c
>>>> --
>>>> 2.53.0
>
