Re: [PATCH 4/8] accel/tcg: introduce tcg_kick_vcpu_thread

["Philippe Mathieu-Daudé" <philmd@linaro.org>]

On 11/8/25 08:10, Paolo Bonzini wrote:
> On 8/10/25 01:26, Richard Henderson wrote:
>> On 8/9/25 04:59, Paolo Bonzini wrote:
>>> +void tcg_kick_vcpu_thread(CPUState *cpu)
>>> +{
>>> +    /*
>>> +     * Ensure cpu_exec will see the reason why the exit request was 
>>> set.
>>> +     * FIXME: this is not always needed.  Other accelerators instead
>>> +     * read interrupt_request and set exit_request on demand from the
>>> +     * CPU thread; see kvm_arch_pre_run() for example.
>>> +     */
>>> +    qatomic_store_release(&cpu->exit_request, 1);
>>> +
>>> +    /* Ensure cpu_exec will see the exit request after TCG has 
>>> exited.  */
>>> +    qatomic_store_release(&cpu->neg.icount_decr.u16.high, -1);
>>> +}
>>
>> So, now both cpu_exit and cpu_kick set exit_request.
>>
>> You ifdef this out again for user-only in patch 7, but this does 
>> suggest that kick and exit are essentially interchangeable.  You 
>> rearrange things a bit in patch 6, but it's still not clear to me what 
>> the difference between the two should be.  There's certainly nothing 
>> at all in include/hw/core/cpu.h to differentiate them.
>>
>> Should we instead eliminate one of kick or exit, unifying the paths?
> In cpu-exec.c terms, qemu_cpu_kick() *should* go out to 
> cpu_handle_interrupt() whereas cpu_exit() *should* go out to 
> cpu_handle_exception().  The difference matters for some accelerators 
> where qemu_cpu_kick() tries not to take the BQL in the vCPU thread.

While I'm still learning this area, this answer makes sense to me.
If this is correct (reviewed by others), can we document it in the
corresponding prototype declarations?

> 
> Until now TCG's implementation of kick_vcpu_thread set both exit_request 
> and interrupt_request, and I'm not changing that yet for system 
> emulation.

For long I tried to figure out why both are set there.

> Patch 7 does that for user-mode emulation, because it's 
> trivial: neither linux-user not bsd-user use qemu_cpu_kick() directly.
> 
> Paolo
> 
>