Re: [PATCH 3/4 changelog-v2] KVM: Switch to srcu-less get_dirty_log()

[Marcelo Tosatti <mtosatti@redhat.com>]

On Thu, Mar 08, 2012 at 10:35:45AM +0900, Takuya Yoshikawa wrote:
> Marcelo Tosatti <mtosatti@redhat.com> wrote:
> 
> > What is worrying are large memory cases: think of the 50GB slot case.
> > 100ms hold time is pretty bad (and reacquiring the lock is relatively
> > simple).
> > 
> 
> OK, I agree basically.
> 
> But let me explain one thing before deciding what I should do next.
> 
> With my method, even when we use a 50GB slot, the hold time will be under
> 10ms -- not 100ms -- if the memory actually updated from the last time is
> 1GB (256K dirty pages).
> 
> > >   8747274.0   10973.3       33.3      -31%    -3%    256K
>   Note that this unit-test was done on my tiny core-i3 32-bit host.
>   On servers which can install more than 50GB memory, this will become
>   much faster: actually my live migration tests done on Xeon saw much
>   better numbers.  Unit-test has been tuned for the worst case.
> 
> I admit that if the dirty memory size is more than 10GB, we may see over
> 100ms hold time you are worrying about.
> 
>   For that I was proposing introducing a new GET_DIRTY_LOG API which can
>   restrict the number of dirty pages we get the log - but this is a long
>   term goal.
> 
> 
> So, I am OK to try to introduce cond_resched_lock_cb() as you suggested.
> But, as I explained above, my current implementation does not introduce
> any real regression concerning to mmu_lock hold time:
> 
>   Before we could see 10ms hold time in real workloads:
>   > funcgraph_entry:      ! 9783.060 us |  kvm_mmu_slot_remove_write_access();
> 
>   I have never seen ms hold time with my method in the same workloads.
> 
> So, it would be helpful if you can apply the patch series and I can work
> on top of that: although I cannot use servers with 100GB memory now,
> migrating a guest with 16GB memory or so may be possible later: I need
> to reserve servers for that.

Makes sense.

It looks good to me, Avi can you review & ack please?

> I also want to know "mmu_lock -- TLB flush"-decoupling plan.  We will not
> need to introduce cond_resched_lock_cb() in sched.h if that is possible.
> 
> Thanks,
> 	Takuya