On 12/08/2015 11:58, Paolo Bonzini wrote:
On 11/08/2015 23:34, Frederic Konrad wrote:
Also if qemu_cond_broadcast(&qemu_io_proceeded_cond) is being dropped
there is no point keeping the guff around in qemu_tcg_wait_io_event.
Yes good point.
BTW this leads to high consumption of host CPU eg: 100% per VCPU thread as
the VCPUs thread are no longer waiting for qemu_io_proceeded_cond.
If the guest CPU is busy waiting, that's expected. But if the guest CPU
is halted, it should not have 100% host CPU consumption.
Paolo
Hmm so that's definitely strange... I mean theorically it's the same as
before?
An other thing. It seems that we need to signal the VCPU when the
iothread take
the lock eg:
if (tcg_enabled() && qemu_thread_is_self(&io_thread)) {
CPU_FOREACH(cpu) {
cpu_exit(cpu);
}
}
To make this patch working without MTTCG.
Fred
