Re: [Qemu-devel] [RFC PATCH V3 0/3] Multithread TCG async_safe_work part.

[Frederic Konrad <fred.konrad@greensocs.com>]

On 20/07/2015 19:36, Alex Bennée wrote:
> fred.konrad@greensocs.com writes:
>
>> From: KONRAD Frederic <fred.konrad@greensocs.com>
>>
>> This is the async_safe_work introduction bit of the Multithread TCG work.
>> Rebased on current upstream (6169b60285fe1ff730d840a49527e721bfb30899).
>>
>> (Currently untested as I need to rebase MTTCG first.)
>>
>> It can be cloned here:
>> http://git.greensocs.com/fkonrad/mttcg.git branch async_work_v3
>>
>> The first patch introduces a mutex to protect the existing queued_work_*
>> CPUState members against multiple (concurent) access.
>>
>> The second patch introduces a tcg_exec_flag which will be 1 when we are inside
>> cpu_exec(), -1 when we must not enter the cpu execution and 0 when we are
>> allowed to do so. This is required as safe work need to be sure that's all vCPU
>> are outside cpu_exec().
>>
>> The last patch introduces async_safe_work. It allows to add some work which will
>> be done asynchronously but only when all vCPUs are outside cpu_exec(). The tcg
>> thread will wait that no vCPUs have any pending safe work before reentering
>> cpu-exec().
>>
>> Changes V2 -> V3:
>>    * Check atomically we are not in the executiong loop to fix the race condition
>>      which might happen.
>> Changes V1 -> V2:
>>    * Release the lock while running the callback for both async and safe work.
>>
>> KONRAD Frederic (3):
>>    cpus: protect queued_work_* with work_mutex.
>>    cpus: add tcg_exec_flag.
>>    cpus: introduce async_run_safe_work_on_cpu.
> Also currently breaks a lot of targets:
>
> https://travis-ci.org/stsquad/qemu/builds/71797143

oops yes, seems linux-user is not happy at all.
Will fix that.

Thanks,
Fred
>>   cpu-exec.c        |  10 ++++
>>   cpus.c            | 160 ++++++++++++++++++++++++++++++++++++++++--------------
>>   include/qom/cpu.h |  57 +++++++++++++++++++
>>   qom/cpu.c         |  20 +++++++
>>   4 files changed, 207 insertions(+), 40 deletions(-)