Re: [Qemu-devel] [PATCH v3 00/11] Reduce lock contention on TCG hot-path

[Sergey Fedorov]

On 13/07/16 10:39, Paolo Bonzini wrote:
> On 12/07/2016 22:13, Sergey Fedorov wrote:
>> From: Sergey Fedorov <address@hidden>
>>
>> Hi,
>>
>> This is my respin of Alex's v2 series [1].
>>
>> The first 8 patches are preparation for the patch 9, the subject matter
>> of this series, which enables lockless translation block lookup. The
>> main change here is that Paolo's suggestion is implemented: TBs are
>> marked with invalid CPU state early during invalidation. This allows to
>> make lockless lookup safe from races on 'tb_jmp_cache' and direct block
>> chaining.
> Thanks for looking at the suggestion again and especially for perfecting it!
>
>> The patch 10 is a simple solution to avoid unnecessary bouncing on
>> 'tb_lock' between tb_gen_code() and tb_add_jump(). A local variable is
>> used to keep track of whether 'tb_lock' has already been taken.
>>
>> The last patch is my attempt to restructure tb_find_{fast,slow}() into a
>> single function tb_find(). I think it will be easier to follow the
>> locking scheme this way. However, I am afraid this last patch can be
>> controversial, so it can be simply dropped.
> Actually I agree entirely with it.
>
> If anything, for historical reasons one might rename tb_find_physical to
> tb_find_slow and leave the tb_find_fast name, but I think the patch is
> good as is.

Nice to hear that :)

>
> Have you measured performance with the series?  In any case, it's nice
> to see MTTCG finally taking shape!

No, I didn't measured the performance. Maybe Alex can help me with this?

Thanks,
Sergey