Re: [PATCH 6/9] KVM: arm64: Split huge pages when dirty logging is enabled

[Marc Zyngier <maz@kernel.org>]

On Thu, 26 Jan 2023 18:45:43 +0000,
Ricardo Koller <ricarkol@google.com> wrote:
> 
> On Tue, Jan 24, 2023 at 2:45 PM Oliver Upton <oliver.upton@linux.dev> wrote:
> >
> > Hi Ricardo,
> >
> > On Fri, Jan 13, 2023 at 03:49:57AM +0000, Ricardo Koller wrote:
> > > Split huge pages eagerly when enabling dirty logging. The goal is to
> > > avoid doing it while faulting on write-protected pages, which
> > > negatively impacts guest performance.
> > >
> > > A memslot marked for dirty logging is split in 1GB pieces at a time.
> > > This is in order to release the mmu_lock and give other kernel threads
> > > the opportunity to run, and also in order to allocate enough pages to
> > > split a 1GB range worth of huge pages (or a single 1GB huge page).
> > > Note that these page allocations can fail, so eager page splitting is
> > > best-effort.  This is not a correctness issue though, as huge pages
> > > can still be split on write-faults.
> > >
> > > The benefits of eager page splitting are the same as in x86, added
> > > with commit a3fe5dbda0a4 ("KVM: x86/mmu: Split huge pages mapped by
> > > the TDP MMU when dirty logging is enabled"). For example, when running
> > > dirty_log_perf_test with 64 virtual CPUs (Ampere Altra), 1GB per vCPU,
> > > 50% reads, and 2MB HugeTLB memory, the time it takes vCPUs to access
> > > all of their memory after dirty logging is enabled decreased by 44%
> > > from 2.58s to 1.42s.
> > >
> > > Signed-off-by: Ricardo Koller <ricarkol@google.com>
> > > ---
> > >  arch/arm64/include/asm/kvm_host.h |  30 ++++++++
> > >  arch/arm64/kvm/mmu.c              | 110 +++++++++++++++++++++++++++++-
> > >  2 files changed, 138 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> > > index 35a159d131b5..6ab37209b1d1 100644
> > > --- a/arch/arm64/include/asm/kvm_host.h
> > > +++ b/arch/arm64/include/asm/kvm_host.h
> > > @@ -153,6 +153,36 @@ struct kvm_s2_mmu {
> > >       /* The last vcpu id that ran on each physical CPU */
> > >       int __percpu *last_vcpu_ran;
> > >
> > > +     /*
> > > +      * Memory cache used to split EAGER_PAGE_SPLIT_CHUNK_SIZE worth of huge
> > > +      * pages. It is used to allocate stage2 page tables while splitting
> > > +      * huge pages. Its capacity should be EAGER_PAGE_SPLIT_CACHE_CAPACITY.
> > > +      * Note that the choice of EAGER_PAGE_SPLIT_CHUNK_SIZE influences both
> > > +      * the capacity of the split page cache (CACHE_CAPACITY), and how often
> > > +      * KVM reschedules. Be wary of raising CHUNK_SIZE too high.
> > > +      *
> > > +      * A good heuristic to pick CHUNK_SIZE is that it should be larger than
> > > +      * all the available huge-page sizes, and be a multiple of all the
> > > +      * other ones; for example, 1GB when all the available huge-page sizes
> > > +      * are (1GB, 2MB, 32MB, 512MB).
> > > +      *
> > > +      * CACHE_CAPACITY should have enough pages to cover CHUNK_SIZE; for
> > > +      * example, 1GB requires the following number of PAGE_SIZE-pages:
> > > +      * - 512 when using 2MB hugepages with 4KB granules (1GB / 2MB).
> > > +      * - 513 when using 1GB hugepages with 4KB granules (1 + (1GB / 2MB)).
> > > +      * - 32 when using 32MB hugepages with 16KB granule (1GB / 32MB).
> > > +      * - 2 when using 512MB hugepages with 64KB granules (1GB / 512MB).
> > > +      * CACHE_CAPACITY below assumes the worst case: 1GB hugepages with 4KB
> > > +      * granules.
> > > +      *
> > > +      * Protected by kvm->slots_lock.
> > > +      */
> > > +#define EAGER_PAGE_SPLIT_CHUNK_SIZE                 SZ_1G
> > > +#define EAGER_PAGE_SPLIT_CACHE_CAPACITY                                      \
> > > +     (DIV_ROUND_UP_ULL(EAGER_PAGE_SPLIT_CHUNK_SIZE, SZ_1G) +         \
> > > +      DIV_ROUND_UP_ULL(EAGER_PAGE_SPLIT_CHUNK_SIZE, SZ_2M))
> >
> > Could you instead make use of the existing KVM_PGTABLE_MIN_BLOCK_LEVEL
> > as the batch size? 513 pages across all page sizes is a non-negligible
> > amount of memory that goes largely unused when PAGE_SIZE != 4K.
> >
> 
> Sounds good, will refine this for v2.
> 
> > With that change it is a lot easier to correctly match the cache
> > capacity to the selected page size. Additionally, we continue to have a
> > single set of batching logic that we can improve later on.
> >
> > > +     struct kvm_mmu_memory_cache split_page_cache;
> > > +
> > >       struct kvm_arch *arch;
> > >  };
> > >
> > > diff --git a/arch/arm64/kvm/mmu.c b/arch/arm64/kvm/mmu.c
> > > index 700c5774b50d..41ee330edae3 100644
> > > --- a/arch/arm64/kvm/mmu.c
> > > +++ b/arch/arm64/kvm/mmu.c
> > > @@ -31,14 +31,24 @@ static phys_addr_t hyp_idmap_vector;
> > >
> > >  static unsigned long io_map_base;
> > >
> > > -static phys_addr_t stage2_range_addr_end(phys_addr_t addr, phys_addr_t end)
> > > +bool __read_mostly eager_page_split = true;
> > > +module_param(eager_page_split, bool, 0644);
> > > +
> >
> > Unless someone is really begging for it I'd prefer we not add a module
> > parameter for this.
> 
> It was mainly to match x86 and because it makes perf testing a bit
> simpler. What do others think?

From my PoV this is a no.

If you have a flag because this is an experimental feature (like NV),
then this is a kernel option, and you taint the kernel when it is set.

If you have a flag because this is a modal option that makes different
use of the HW which cannot be exposed to userspace (like GICv4), then
this also is a kernel option.

This is neither.

The one thing that would convince me to make it an option is the
amount of memory this thing consumes. 512+ pages is a huge amount, and
I'm not overly happy about that. Why can't this be a userspace visible
option, selectable on a per VM (or memslot) basis?

Thanks,

	M.

-- 
Without deviation from the norm, progress is not possible.