Re: [Qemu-devel] [RFC 33/38] cpu: introduce cpu_tcg_sched_work to run work while other CPUs sleep

[Emilio G. Cota]

On Sun, Aug 23, 2015 at 18:24:55 -0700, Paolo Bonzini wrote:
> On 23/08/2015 17:24, Emilio G. Cota wrote:
> > This is similar in intent to the async_safe_work mechanism. The main
> > differences are:
> > 
> > - Work is run on a single CPU thread *after* all others are put to sleep
> > 
> > - Sleeping threads are woken up by the worker thread upon completing its job
> > 
> > - A flag as been added to tcg_ctx so that only one thread can schedule
> >   work at a time. The flag is checked every time tb_lock is acquired.
> > 
> > - Handles the possibility of CPU threads being created after the existing
> >   CPUs are put to sleep. This is easily triggered with many threads on
> >   a many-core host in usermode.
> > 
> > - Works for both softmmu and usermode
> > 
> > Signed-off-by: Emilio G. Cota <address@hidden>
> 
> I think this is a duplicate of the existing run_on_cpu code.  If needed
> in user-mode emulation, it should be extracted out of cpus.c.

They're similar, yes.

> Also I think it is dangerous (prone to deadlocks) to wait for other CPUs
> with synchronize_cpu and condvar.

The key to avoid deadlocks is not to hold any locks that might be
acquired within an RCU read critical section when calling
synchronize_rcu(). The condvars are for the sleeping threads so
that they can be woken up; sleepers don't call synchronize_rcu().

>  I would much rather prefer to _halt_
> the CPUs if there is pending work, and keep it halted like this:
> 
>  static inline bool cpu_has_work(CPUState *cpu)
>  {
>      CPUClass *cc = CPU_GET_CLASS(cpu);
> 
> +    if (tcg_ctx.tb_ctx.tcg_has_work) {
> +        return false;
> +    }
>      g_assert(cc->has_work);
>      return cc->has_work(cpu);
>  }
> 
> You can then run flush_queued_work from linux-user/main.c (and
> bsd-user/main.c) when cpu_exec returns EXCP_HALTED.

OK. Will try something like this.

                Emilio