Re: [Qemu-devel] [RFC 0/3] Determinitic behaviour with icount.

[Paolo Bonzini <pbonzini@redhat.com>]

Il 18/07/2013 17:06, Peter Maydell ha scritto:
> On 18 July 2013 16:02,  <fred.konrad@greensocs.com> wrote:
>> As I said in the last email, we have issues with determinism with icount.
>> We are wondering if determinism is really ensured with icount?
> 
> My opinion is that it *should* be deterministic but it would
> be unsurprising if the determinism had got broken along the way.

First of all, it can only be deterministic if the guest satisfies (at
least) all the following condition:

1) only uses timer that QEMU bases on vm_clock (which means that you
should use "-rtc clock=vm"---sorry Fred, didn't think about this in the
previous answer);

2) never does any network operation nor any asynchronous disk I/O operation

3) never halts the VCPU waiting for an interrupt


Point 1 is obvious.


To explain points 2, let's consider what happens if a block device uses
synchronous vs. asynchronous I/O.

With synchronous I/O, each block device operation will complete
immediately.  All clocks are stalled during the operation.

With asynchronous I/O, each block device operation will be done while
the CPU is running.  If the CPU is polling a completion flag, the number
of instructions executed (thus icount) depends on how long it takes to
do I/O.


To explain point 3 (which is the only one that _might_ be fixable),
let's see what happens if the VCPU halts waiting for an interrupt.  If
that is the case, and you haven't done any asynchronous I/O, there
should be active vm_clock timers, and you have another possible source
of non-deterministic behavior.

The current QEMU behavior is (and has always been) to start tracking
rt_clock.  This is obviously not deterministic.  Note that with the
switch to separate threads for iothread/VCPU, the algorithm to do this
has become much better.  Let's look at a couple possibilities:

2) jump to the next vm_clock deadline.  This sounds appealing, but it is
still nondeterministic in the general case when the guest *is* doing
asynchronous I/O too.  How many vm_clock timers do you run before I/O
finishes?  Furthermore, the vm_clock might move too fast.  Think of an
RTC clock whose alarm registers are 0/0/0 so it fires at midnight; if it
is the only active vm_clock timer, you end up in 2107 even before the
kernel boots!

3) do not process vm_clock timers at all unless there is no pending I/O
(block/network); if there is none, track rt_clock as in current
behavior.  I just made it up, but it sounds promising and similar to
synchronous I/O.  It should not be extremely hard to implement, and it
can remove this kind of nondeterminism.  But it won't fix the case when
the CPU is polling.

Paolo

ps: I'm not an expert on icount at all, I'm only reasoning of the
possible interactions with the main loop.

>> Both icount and reverse execution need an instruction counter. icount use a
>> count-down mechanism but reverse execution need a continuous counter. For now
>> we have build a separate counter and we think that these two counters can be
>> merged. However we would like feedback about this before modifying this.
> 
> I definitely think that there should only be one counter, not two.
> 
> thanks
> -- PMM
>