Messages in this thread | | | Date | Mon, 25 Jul 2011 14:59:02 +0200 | From | Paolo Bonzini <> | Subject | Re: [GIT PULL] Native Linux KVM tool for 3.1 |
| |
On 07/25/2011 09:50 AM, Sasha Levin wrote: > Anthony had a talk on last years KVM forum regarding the QEMU threading > model (slide: > http://www.linux-kvm.org/wiki/images/7/70/2010-forum-threading-qemu.pdf) . > > It was suggested that the KVM part of QEMU is having a hard time > achieving the ideal threading model due to its need to support TCG - > something which has nothing to do with KVM itself.
No, it is not having a hard time. The "foot in the door" that Anthony mentions has been part of QEMU and qemu-kvm for a long time (multi-threading is necessary to support SMP!) and works quite well.
Historically, there were three main loops:
1) QEMU single-threaded;
2) QEMU multi-threaded; clean, but buggy, untested and bitrotting;
3) qemu-kvm multi-threaded, forked from (1), ugly but robust and widely deployed.
In 0.15 the two multi-threaded versions have been unified by Jan Kiszka. I have even ported (2) to Windows with little or no pain; porting to Mac OS X interestingly is harder than Windows, because non-portable Linux assumptions about signal handling have crept in the code (to preempt the objections: they weren't just non-portabilities, they were latent bugs). Windows just does not have signals. :)
So, right now, the only difference is that QEMU is still defaulting to the single-threaded main loop, while qemu-kvm enables multi-threading by default. In some time even QEMU will switch.
Yes, this is of course worse than getting it right in the first place; Nobody is saying the opposite.
Paolo
| |