Re: [Qemu-devel] qemu becomes unresponsive sometimes

[Paul Jakma]

On Sun, 28 Nov 2004, Mulyadi Santosa wrote:

> First, let me ask for confirmation, do you run Qemu on Solaris or 
> Linux host? If this is on solaris, maybe my whole assumption would 
> be useless

On Linux, Linux has bridging. I've not tried to run Linux as a qemu 
instance yet, but i intend to soon (as well as some BSDs).

For me, the problem is easily reproducable by running something very 
intensive in the "qemud" host, eg openssl speed.

> Running strace on binary only change one thing: on every syscall, 
> it will be recorded as mentioned by ptrace behaviour. AFAIK, this 
> recording will make the traced binary (in this case Qemu binary) 
> halt on each syscall invocation and signal receive. (man strace and 
> man ptrace).

And syscall exit.

> After that, the parent process (the shell which fork Qemu)

Well, strace in this case. (and an attached strace - not parent).

> will continue the execution of Qemu binary by sending signal

With ptrace (PTRACE_CONT, ...) actually, AIUI.

> So, my suspicion, somehow Qemu is put into task_interruptible by 
> host kernel (Linux?) no matter how much is the load.

I've no idea unfortunately. I was hoping I could find a clue as to 
what the problem was by stracing qemu, but seeing as how the problem 
disappears when i do that... and I dont know anything about qemu 
internals unfortunately - is there a central scheduler loop somewhere 
I could add debug instrumentation to?

Also, does qemu make heavy use of signals? From an ltrace of qemu, it 
looks like it might be up to some trickery with SIG_IO - but that 
could be SDL. Sadly, I dont know enough of qemu. :(

> But, it is possible that it is not the whole Qemu that is 
> unresponsive, maybe it is just the Qemu monitor/display that is put 
> into "sleep"

The whole qemu host, along with qemu monitor and SDL display is 
unresponsive - my ssh session to the host stop responding, the hosted 
kernel stops responding to pings - and qemu usually takes 99% CPU, so 
its up to something, just not to do with servicing the hosted OS or 
any other external events/IO (AFAICT).

> What do you think ? Fabrice any comment? Maybe we need to force 
> waking up the whole Qemu process (SDL output, monitor, VNC perhaps) 
> everytime there is a "CPU" activity inside the guest kernel?

I'd love to have this fixed. If someone were to, I'd reward them with 
as much Guinness as they could drink in a night ;) (offer redeemable 
at any decent pub in Dublin city or west Dublin county.)

> regards
>
> Mulyadi

regards,
--
Paul Jakma address@hidden address@hidden Key ID: 64A2FF6A
Fortune: 
statistics, n.:
        A system for expressing your political prejudices in convincing
        scientific guise.