Re: [PATCH/RFC 3/4]Introduce "account modifiers" mechanism

Rusty Russell wrote:

On Fri, 2007-08-17 at 09:35 +0200, Laurent Vivier wrote:

quoted

Rusty Russell wrote:

quoted

Hi Laurent,

Hi Rusty,
how are your puppies ?

They're getting a little fat, actually.  Too many features ...

quoted

- remove PATCH 3, and add in task_struct a "ktime vtime" where we accumulate
guest time (by calling something like guest_enter() and guest_exit() from the
virtualization engine), and when in account_system_time() we have cputime >
vtime we substrate vtime from cputime and add vtime to user time and guest time.
But doing like this we freeze in kernel/sched.c the link between system time,
user time and guest time (i.e. system time = system time - vtime, user time =
user time + vtime and guest time = guest time + vtime).

Actually, I think we can set a per-cpu "in_guest" flag for the scheduler
code, which then knows to add the tick to the guest time.  That seems
the simplest possible solution.

lguest or kvm would set the flag before running the guest (which is done
with preempt disabled or using preemption hooks), and reset it
afterwards.

Thoughts?

It was my first attempt (except I didn't have a per-cpu flag, but a per-task
flag), it's not visible but I love simplicity... ;-)

A KVM VCPU is stopped by preemption, so when we enter in scheduler we have
exited from VCPU and thus this flags is off (so we account 0 to the guest). What
I did then is "set the flag on when we enter in the VCPU, and
"account_system_time()" sets the flag off when it adds this timeslice to cpustat
(and compute correctly guest, user, system time). But I didn't like this idea
because all code executed after we entered in the VCPU is accounted to the guest
until we have an account_system_time() and I suppose we can have real system
time in this part. And I guess a VCPU can be less than 1 ms (unit of cputime) in
a timeslice.

So ? What's best ?
Laurent
-- 
------------- Laurent.Vivier@bull.net  --------------
          "Software is hard" - Donald Knuth