Re: [RFC PATCH] semihosting: suspend recieving CPU when blocked (HACK, WIP)

[Alex Bennée]

Paolo Bonzini <address@hidden> writes:

> On 17/12/19 15:18, Alex Bennée wrote:
>> 
>> Paolo Bonzini <address@hidden> writes:
>> 
>>> On 17/12/19 14:42, Alex Bennée wrote:
>>>>> Why do you need to set exception_index to something other than -1 (using
>>>>> cpu_loop_exit_noexc for example)?
>>>> If there is no exception to process we won't exit the main loop which we
>>>> need to do if we want to wait until there is data to read.
>>>
>>> Okay.
>>>
>>>>> Using ->stop here is a bit weird, since ->stop is usually related to
>>>>> pause_all_vcpus.
>>>>
>>>> Arguably we could come up with a better API to cpu.c but this allows us
>>>> to use cpu_resume(c->sleeping_cpu) when waking up rather than hand
>>>> rolling our own wake-up mechanism.
>>>
>>> But we already have the right wake-up mechanism, which is
>>> cpu->halted/cpu_has_work.
>> 
>> cpu_has_work is a guest function though and semihosting_console is a
>> common hw module. It can't peek into the guests internal state.
>
> semihosting_console only needs to something like
> cpu_interrupt(cpu->stopped_cpu, CPU_INTERRUPT_SEMIHOST).

As an exception is being delivered we just end up re-executing the
EXCP_SEMIHOST. I still don't see why using cpu_interrupt is an
improvement seeing as it is secondary to exception processing.

> (By the way,
> the stopped_cpu should probably be a list to mimic the condition
> variable---for example a GSList).

ok

>
>> This all
>> comes back to cpu_thread_is_idle anyway in making our decision about if
>> we do or do not sleep on the halt_cond.
>> 
>>> That also makes it possible to just use
>>> EXCP_HALTED instead of adding a new EXCP_BLOCKED.
>> 
>> We can certainly use EXCP_HALTED but maybe come up with a common way of
>> entering the state? There seems to be a combination of messing around
>> with special interrupts and direct poking of cs->halted = 1 while
>> setting the exception. Maybe this could finally clear up the #if
>> defined(TARGET_I386) hacking in cpus.c?
>
> If you're talking accel/tcg/cpu-exec.c, that's different; the issue
> there is that x86 has a kind of warm reset pin that is not equivalent to
> cpu_reset.  Removing that would only entail adding a new member function
> to CPUClass.
>
> Paolo


-- 
Alex Bennée