On Mon, Jul 10, 2017 at 17:33:07 -0400, Paolo Bonzini wrote:
>
> > I agree that it would be nice to have the same mechanism for all.
> >
> > The main hurdle I see is how to allow for concurrent code generation while
> > minimizing flushes of the single, fixed-size[*] code_gen_buffer.
> > In user-mode this is tricky because there is no way to bound the number
> > of threads that might be spawned by the guest code (I don't think reading
> > /proc/sys/kernel/threads-max is a viable solution here).
> >
> > Switching to a "__thread *tcg_ctx_ptr" model will help minimize
> > user-mode/softmmu differences though. The only remaining difference would be
> > that user-mode would need tb_lock() around tb_gen_code, whereas softmmu
> > wouldn't, but everything else would be the same.
>
> Hmm, tb_gen_code is already protected by mmap_lock in linux-user, so you
> wouldn't
> get any parallelism. On the other hand, you could just say that the
> fixed-size
> code_gen_buffer is protected by mmap_lock, which doesn't exist for softmmu.
Yes. tb_lock/mmap_lock, or like they're called in some asserts, memory_lock.
A way to get some parallelism in user-mode given the constraints
would be to share regions among TCG threads. Threads would still need to take
a per-region lock, but it wouldn't be a global lock so that would scale better.
I'm not sure we really need that much parallelism for code generation in
user-mode,
though. So I wouldn't focus on this until seeing benchmarks that have a clear
bottleneck due to "memory_lock".
E.
