On 11/05/16 16:46, Paolo Bonzini wrote:
> On 11/05/2016 15:36, Sergey Fedorov wrote:
>> On 11/05/16 15:58, Paolo Bonzini wrote:
>>> On 06/05/2016 20:22, Sergey Fedorov wrote:
>>>> However, there's no sensible description of what is protected by tb_lock
>>>> and mmap_lock. I think we need to have a clear documented description of
>>>> the TCG locking scheme in order to be sure we do right things in MTTCG.
>>> I think there was such a patch somewhere, but: tb_lock basically
>>> protects tcg_ctx, while mmap_lock protects the user-mode emulation page
>>> table (the equivalent for system emulation is the memory map which is
>>> protected by the BQL). Furthermore, mmap_lock must be taken outside
>>> tb_lock.
>> What's a user-mode emulation page table? 'l1_map'?
> Yes. It's used beyond TCG in user-mode emulation.
>
>> It is used by system
>> emulation to keep track of TBs per page and 'code_bitmap'. Shouldn't it
>> be protected with 'mmap_lock' in system emulation?
> tb_lock is used instead because it's taken everywhere system emulation
> uses l1_map; so tb_lock is protecting l1_map too in system emulation.
>
> As mentioned above, user-mode emulation uses l1_map in linux-user/mmap.c
> via page_{get,set}_flags, which I guess is why the lock is separate.
> None of us was involved in the original multi-threaded linux-user work,
> we're reverse engineering it just like you. :)
While I'm investigating 'tb_lock' and 'mmap_lock' usage I am wondering
why don't put 'l1_map' into 'tcg_ctx'?
Kind regards,
Sergey
