Re: [PATCH] kvm: optimize ISR lookups

From: Michael S. Tsirkin
Date: Mon May 21 2012 - 19:06:25 EST


On Mon, May 21, 2012 at 11:04:25PM +0200, Thomas Gleixner wrote:
> On Mon, 21 May 2012, Michael S. Tsirkin wrote:
>
> > We perform ISR lookups twice: during interrupt
> > injection and on EOI. Typical workloads only have
> > a single bit set there. So we can avoid ISR scans by
> > 1. counting bits as we set/clear them in ISR
> > 2. if count is 1, caching the vector number
> > 3. if count != 1, invalidating the cache
> >
> > The real purpose of this is enabling PV EOI
> > which needs to quickly validate the vector.
> > But non PV guests also benefit.
> >
> > Signed-off-by: Michael S. Tsirkin <mst@xxxxxxxxxx>
> > ---
> > arch/x86/kvm/lapic.c | 51 ++++++++++++++++++++++++++++++++++++++++++++++++-
> > arch/x86/kvm/lapic.h | 2 +
> > 2 files changed, 51 insertions(+), 2 deletions(-)
> >
> > diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
> > index 93c1574..232950a 100644
> > --- a/arch/x86/kvm/lapic.c
> > +++ b/arch/x86/kvm/lapic.c
> > @@ -107,6 +107,16 @@ static inline void apic_clear_vector(int vec, void *bitmap)
> > clear_bit(VEC_POS(vec), (bitmap) + REG_POS(vec));
> > }
> >
> > +static inline int __apic_test_and_set_vector(int vec, void *bitmap)
> > +{
> > + return __test_and_set_bit(VEC_POS(vec), (bitmap) + REG_POS(vec));
> > +}
> > +
> > +static inline int __apic_test_and_clear_vector(int vec, void *bitmap)
> > +{
> > + return __test_and_clear_bit(VEC_POS(vec), (bitmap) + REG_POS(vec));
> > +}
> > +
> > static inline int apic_hw_enabled(struct kvm_lapic *apic)
> > {
> > return (apic)->vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE;
> > @@ -210,6 +220,16 @@ static int find_highest_vector(void *bitmap)
> > return fls(word[word_offset << 2]) - 1 + (word_offset << 5);
> > }
> >
> > +static u8 count_vectors(void *bitmap)
> > +{
> > + u32 *word = bitmap;
> > + int word_offset;
> > + u8 count = 0;
> > + for (word_offset = 0; word_offset < MAX_APIC_VECTOR >> 5; ++word_offset)
> > + count += hweight32(word[word_offset << 2]);
> > + return count;
> > +}
>
> Why do you need to reimplement bitmap_weight()?
>
> Because your bitmap is not a bitmap,

It's an existing "bitmap". It's not my bitmap :)

> but a set of 32bit registers
> which have an offset of 4 words each. I really had to twist my brain
> around this function and the test_and_clr/set ones above just because
> the name bitmap is so horribly misleading. Add an extra bonus for
> using void pointers.
>
> Yes, I know, the existing code is full of this, but that's not an
> excuse to add more of it.

There's no other way to use existing code. If current code
is reworked to use bitmap.h then my patch can use it too.


> This emulation is just silly. Nothing in an emulation where the access
> to the emulated hardware is implemented with vmexits is requiring a
> 1:1 memory layout. It's completely irrelevant whether the hardware has
> an ISR regs offset of 0x10 or not. Nothing prevents the emulation to
> use a consecutive bitmap for the vector bits instead of reimplementing
> a boatload of bitmap operations.
>
> The only justification for having the same layout as the actual
> hardware is when you are going to map the memory into the guest space,
> which is not the case here.

I think the reason is __apic_read which now simply copies the registers
out to guest, this code will become less straight-forward if it's not
1:1.

> And if you look deeper, then you'll notice that _ALL_ APIC registers
> are on a 16 byte boundary (thanks Peter for pointing that out).
>
> So it's even more silly to have a 1:1 representation instead of
> implementing the default emulated apic_read/write functions to access
> (offset >> 2).
>
> And of course, that design decision causes lookups to be slow.

Yes, it might be one of the reasons why my patch helps so
much: it adds a cache in front of this data structure.

> It's
> way faster to scan a consecutive bitmap than iterating through eight
> 32bit words with an offset of 0x10, especially on a 64 bit host.
>
> > static inline int apic_test_and_set_irr(int vec, struct kvm_lapic *apic)
> > {
> > apic->irr_pending = true;
> > @@ -242,6 +262,25 @@ static inline void apic_clear_irr(int vec, struct kvm_lapic *apic)
> > apic->irr_pending = true;
> > }
> >
> > +static inline void apic_set_isr(int vec, struct kvm_lapic *apic)
> > +{
> > + if (!__apic_test_and_set_vector(vec, apic->regs + APIC_ISR))
> > + ++apic->isr_count;
> > + ASSERT(apic->isr_count > MAX_APIC_VECTOR);
>
> I'm really curious what you observed when defining DEBUG in that file.
>
> Clearly you never did.
>
> > + if (likely(apic->isr_count == 1))
> > + apic->isr_cache = vec;
> > + else
> > + apic->isr_cache = -1;
> > +}
> > +
> > +static inline void apic_clear_isr(int vec, struct kvm_lapic *apic)
> > +{
> > + if (__apic_test_and_clear_vector(vec, apic->regs + APIC_ISR))
> > + --apic->isr_count;
> > + ASSERT(apic->isr_count < 0);
>
> Ditto.
>
> > result = find_highest_vector(apic->regs + APIC_ISR);
> > ASSERT(result == -1 || result >= 16);
>
> And obviously none of the folks who added this gem bothered to define
> DEBUG either.
>
> So please instead of working around horrid design decisions and adding
> more mess to the existing one, can we please cleanup the stuff first
> and then decide whether it's worth to add the extra magic?
>
> Thanks,
>
> tglx


So what you propose is in fact to rework apic registers at least for
ISR,IRR,TMR to use a bitmap.
I am fine with this suggestion but would like some feedback from kvm
maintainers on where they want to go before I spend time on that.

--
MST
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/