[Qemu-devel] commit b1bbfe72 causes huge slowdown with no kvm

[Luigi Rizzo]

I recently found out that without kvm enabled, and especially

with -smp 2 or greater, qemu becomes incredibly slow

(to the point that you can see kernel messages in the

quest print one character at a time).

This happens with a Linux host (even with -smp 1)

and with FreeBSD host (in this case -smp 2 or greater;

-smp 1 seems to work fine there).

Bisecting points to this commit as the culprit:

commit b1bbfe72ec1ebf302d97f886cc646466c0abd679

Author: Alex Bligh <address@hidden>

Date:   Wed Aug 21 16:02:55 2013 +0100

    aio / timers: On timer modification, qemu_notify or aio_notify

    On qemu_mod_timer_ns, ensure qemu_notify or aio_notify is called to

    end the appropriate poll(), irrespective of use_icount value.

    On qemu_clock_enable, ensure qemu_notify or aio_notify is called for

    all QEMUTimerLists attached to the QEMUClock.

    Signed-off-by: Alex Bligh <address@hidden>

    Signed-off-by: Stefan Hajnoczi <address@hidden>

I could not revert this individual commit into master because

of other changes, but i notice that 

one key

changes of the commit

was

to

 ma

k

e a

call to timerlist_notify() unconditional, whereas

before

it was controlled by the

"

use_icount

"

variable.

So I tried the small patch below and it seems to restore the original

performance, but I have no idea what use_icount does and

whether the change makes sense.

If not, there is an open problem.

Any ideas ?

cheers

luigi

d

iff --git a/qemu-timer.c b/qemu-timer.c

index e15ce47..4180a86 100644

--- a/qemu-timer.c

+++ b/qemu-timer.c

@@ -380,6 +380,7 @@ static void timerlist_rearm(QEMUTimerList *timer_list)

 {

     /* Interrupt execution to force deadline recalculation.  */

     qemu_clock_warp(timer_list->clock->type);

+if (use_icount) // XXX

     timerlist_notify(timer_list);

 }