On Sun, Apr 29, 2012 at 05:50:04PM +0900, Takuya Yoshikawa wrote:
> On Fri, 27 Apr 2012 11:52:13 -0300
> Marcelo Tosatti <mtosatti@xxxxxxxxxx> wrote:
>
> > Yes but the objective you are aiming for is to read and write sptes
> > without mmu_lock. That is, i am not talking about this patch.
> > Please read carefully the two examples i gave (separated by "example)").
>
> The real objective is not still clear.
>
> The ~10% improvement reported before was on macro benchmarks during live
> migration. At least, that optimization was the initial objective.
>
> But at some point, the objective suddenly changed to "lock-less" without
> understanding what introduced the original improvement.
>
> Was the problem really mmu_lock contention?
>
> If the path being introduced by this patch is really fast, isn't it
> possible to achieve the same improvement still using mmu_lock?
Right. Supposedly, mmu_lock cacheline bouncing is the problem. Hum:
$ pahole -C "kvm" /tmp/kvm.ko
struct kvm {
spinlock_t mmu_lock; /* 0 2
*/
/* XXX 6 bytes hole, try to pack */
struct mutex slots_lock; /* 8 32
*/
struct mm_struct * mm; /* 40 8
*/
struct kvm_memslots * memslots; /* 48 8
*/
struct srcu_struct srcu; /* 56 48
*/
/* --- cacheline 1 boundary (64 bytes) was 40 bytes ago --- */
u32 bsp_vcpu_id; /* 104 4
*/
/* XXX 4 bytes hole, try to pack */
Oops. False sharing?
> Note: During live migration, the fact that the guest gets faulted is
> itself a limitation. We could easily see noticeable slowdown of a
> program even if it runs only between two GET_DIRTY_LOGs.
>
>
> > The rules for code under mmu_lock should be:
> >
> > 1) Spte updates under mmu lock must always be atomic and
> > with locked instructions.
> > 2) Spte values must be read once, and appropriate action
> > must be taken when writing them back in case their value
> > has changed (remote TLB flush might be required).
>
> Although I am not certain about what will be really needed in the
> final form, if this kind of maybe-needed-overhead is going to be
> added little by little, I worry about possible regression.
>
> Thanks,
> Takuya
Yes, that is a possibility.
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
