Re: [Qemu-devel] [PATCH] runstate: do not discard runstate changes when paused

[Paolo Bonzini <pbonzini@redhat.com>]

On 10/05/2011 08:50 PM, Luiz Capitulino wrote:
>>>> >  >  >  I'm not exactly against the semantics you're proposing, but they don't
>>>> >  >  >  seem to fit today's qemu.
>>> >  >
>>> >  >  Today's qemu is broken here.
>> >
>> >  For me it's broken because it will abort() if you migrate a paused vm, for
>> >  you it seems to be broken at the semantic level.
>> >
>> >  We can fix the semantics without breaking compatibility.
>
> s/We can/ We can't

I think we should divide stop causes into three groups:

1) those that are undone by QEMU itself:
	RSTATE_DEBUG
	RSTATE_SAVEVM
	RSTATE_PRE_MIGRATE
	RSTATE_RESTORE

For these a lock/release scheme is definitely better.  The VM should not 
start until none of these conditions is in effect, even after a "cont" 
command.

2) those that are undone by management:
	RSTATE_IO_ERROR

For this we can add a new "retry" monitor command that guarantees no 
races if the user issues a "stop" or "cont" command while management is 
processing it.  Effectively, it is also a lock/release scheme but 
controlled by management.

3) those that are undone by "cont":
	RSTATE_PRE_LAUNCH
	RSTATE_PAUSED
	RSTATE_WATCHDOG
	RSTATE_POST_MIGRATE
	RSTATE_PANICKED

It put here the three runstates where the VM should really not be 
restarted at all.  We can then add a new "start" command that only flips 
these five to RSTATE_RUNNING.


So the runstate is composed of six elements: five lock/unlock states (of 
which only one can be unlocked by the user), and one running/paused 
state (composed of five pause reasons + "none").  That is, the runstate 
is a tuple like [debug, savevm, pre_migrate, restore, io_error, 
pause_reason] and for the VM to run it must look like [false, false, 
false, false, false, none].

The four monitor commands would be:

1) "stop":
	if runstate[pause_reason] == none then
		runstate[pause_reason] = paused

2) "retry":
	runstate[io_error] = false

3) "start":
	runstate[pause_reason] = none

There could also be a differentiation between "start" and "start -f", 
where "-f" would be needed to get out of RSTATE_POST_MIGRATE, 
RSTATE_PANICKED and probably RSTATE_WATCHDOG too.

4) "cont": backwards compatibility provided by "retry"+"start -f".

How does this look?

Paolo