• Paolo Bonzini's avatar
    target-i386: yield to another VCPU on PAUSE · 81f3053b
    Paolo Bonzini authored
    After commit b1bbfe72 (aio / timers: On timer modification, qemu_notify
    or aio_notify, 2013-08-21) FreeBSD guests report a huge slowdown.
    The problem shows up as soon as FreeBSD turns out its periodic (~1 ms)
    tick, but the timers are only the trigger for a pre-existing problem.
    Before the offending patch, setting a timer did a timer_settime system call.
    After, setting the timer exits the event loop (which uses poll) and
    reenters it with a new deadline.  This does not cause any slowdown; the
    difference is between one system call (timer_settime and a signal
    delivery (SIGALRM) before the patch, and two system calls afterwards
    (write to a pipe or eventfd + calling poll again when re-entering the
    event loop).
    Unfortunately, the exit/enter causes the main loop to grab the iothread
    lock, which in turns kicks the VCPU thread out of execution.  This
    causes TCG to execute the next VCPU in its round-robin scheduling of
    VCPUS.  When the second VCPU is mostly unused, FreeBSD runs a "pause"
    instruction in its idle loop which only burns cycles without any
    progress.  As soon as the timer tick expires, the first VCPU runs
    the interrupt handler but very soon it sets it again---and QEMU
    then goes back doing nothing in the second VCPU.
    The fix is to make the pause instruction do "cpu_loop_exit".
    Cc: Richard Henderson <rth@twiddle.net>
    Reported-by: default avatarLuigi Rizzo <rizzo@iet.unipi.it>
    Signed-off-by: default avatarPaolo Bonzini <pbonzini@redhat.com>
    Reviewed-by: default avatarRichard Henderson <rth@twiddle.net>
    Message-id: 1384948442-24217-1-git-send-email-pbonzini@redhat.com
    Signed-off-by: default avatarAnthony Liguori <aliguori@amazon.com>
helper.h 7.29 KB