Re: [Qemu-devel] MTTCG External Halt

[Alistair Francis]

On Thu, Feb 1, 2018 at 4:01 AM, Alex Bennée <address@hidden> wrote:
>
> Alistair Francis <address@hidden> writes:
>
>> On Wed, Jan 31, 2018 at 12:32 PM, Alex Bennée <address@hidden> wrote:
>>>
>>> Alistair Francis <address@hidden> writes:
>>>
>>>> On Tue, Jan 30, 2018 at 8:26 PM, Paolo Bonzini <address@hidden> wrote:
>>>>> On 30/01/2018 18:56, Alistair Francis wrote:
>>>>>>
>>>>>> I don't have a good solution though, as setting CPU_INTERRUPT_RESET
>>>>>> doesn't help (that isn't handled while we are halted) and
>>>>>> async_run_on_cpu()/run_on_cpu() doesn't reliably reset the CPU when we
>>>>>> want.
>>>>>>
>>>>>> I've ever tried pausing all CPUs before reseting the CPU and them
>>>>>> resuming them all but that doesn't seem to to work either.
>>>>>
>>>>> async_safe_run_on_cpu would be like async_run_on_cpu, except that it
>>>>> takes care of stopping all other CPUs while the function runs.
>>>>>
>>>>>> Is there
>>>>>> anything I'm missing? Is there no reliable way to reset a CPU?
>>>>>
>>>>> What do you mean by reliable?  Executing no instruction after the one
>>>>> you were at?
>>>>
>>>> The reset is called by a GPIO line, so I need the reset to be called
>>>> basically as quickly as the GPIO line changes. The async_ and
>>>> async_safe_ functions seem to not run quickly enough, even if I run a
>>>> process_work_queue() function afterwards.
>>>>
>>>> Is there a way to kick the CPU to act on the async_*?
>>>
>>> Define quickly enough? The async_(safe) functions kick the vCPUs so they
>>> will all exit the run loop as they enter the next TB (even if they loop
>>> to themselves).
>>
>> We have a special power controller CPU that wakes all the CPUs up and
>> at boot the async_* functions don't wake the CPUs up. If I just use
>> the cpu_rest() function directly everything starts fine (but then I
>> hit issues later).
>>
>> If I forcefully run process_queued_cpu_work() then I can get the CPUs
>> up, but I don't think that is the right solution.
>>
>>>
>>> From an external vCPUs point of view those extra instructions have
>>> already executed. If the resetting vCPU needs them to have reset by the
>>> time it executes it's next instruction it should either cpu_loop_exit at
>>> that point or ensure it is the last instruction in it's TB (which is
>>> what we do for the MMU flush cases in ARM, they all end the TB at that
>>> point).
>>
>> cpu_loop_exit() sounds like it would help, but as I'm not in the CPU
>> context it just seg faults.
>
> What context are you in? gdb-stub does have to something like this.

gdb-stub just seems to use vm_stop() and vm_start().

That fixes all hangs/asserts, but now Linux only brings up 1 CPU (instead of 4).

Alistair