On Mon, Feb 28, 2022 at 3:41 PM David Matlack <dmatlack@xxxxxxxxxx> wrote:
>
> On Mon, Feb 28, 2022 at 1:22 PM Ben Gardon <bgardon@xxxxxxxxxx> wrote:
> >
> > On Wed, Feb 2, 2022 at 5:03 PM David Matlack <dmatlack@xxxxxxxxxx> wrote:
> > >
> > > When splitting a huge page we need to add all of the lower level SPTEs
> > > to the memslot rmap. The current implementation of eager page splitting
> > > bails if adding an SPTE would require allocating an extra pte_list_desc
> > > struct. Fix this limitation by allocating enough pte_list_desc structs
> > > before splitting the huge page.
> > >
> > > This eliminates the need for TLB flushing under the MMU lock because the
> > > huge page is always entirely split (no subregion of the huge page is
> > > unmapped).
> > >
> > > Signed-off-by: David Matlack <dmatlack@xxxxxxxxxx>
> > > ---
> > > arch/x86/include/asm/kvm_host.h | 10 ++++
> > > arch/x86/kvm/mmu/mmu.c | 101 ++++++++++++++++++--------------
> > > 2 files changed, 67 insertions(+), 44 deletions(-)
> > >
> > > diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
> > > index d0b12bfe5818..a0f7578f7a26 100644
> > > --- a/arch/x86/include/asm/kvm_host.h
> > > +++ b/arch/x86/include/asm/kvm_host.h
> > > @@ -1232,6 +1232,16 @@ struct kvm_arch {
> > > hpa_t hv_root_tdp;
> > > spinlock_t hv_root_tdp_lock;
> > > #endif
> > > +
> > > + /*
> > > + * Memory cache used to allocate pte_list_desc structs while splitting
> > > + * huge pages. In the worst case, to split one huge page we need 512
> > > + * pte_list_desc structs to add each new lower level leaf sptep to the
> > > + * memslot rmap.
> > > + */
> > > +#define HUGE_PAGE_SPLIT_DESC_CACHE_CAPACITY 512
> > > + __DEFINE_KVM_MMU_MEMORY_CACHE(huge_page_split_desc_cache,
> > > + HUGE_PAGE_SPLIT_DESC_CACHE_CAPACITY);
> > > };
> > >
> > > struct kvm_vm_stat {
> > > diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> > > index 825cfdec589b..c7981a934237 100644
> > > --- a/arch/x86/kvm/mmu/mmu.c
> > > +++ b/arch/x86/kvm/mmu/mmu.c
> > > @@ -5905,6 +5905,11 @@ void kvm_mmu_init_vm(struct kvm *kvm)
> > > node->track_write = kvm_mmu_pte_write;
> > > node->track_flush_slot = kvm_mmu_invalidate_zap_pages_in_memslot;
> > > kvm_page_track_register_notifier(kvm, node);
> > > +
> > > + kvm->arch.huge_page_split_desc_cache.capacity =
> > > + HUGE_PAGE_SPLIT_DESC_CACHE_CAPACITY;
> > > + kvm->arch.huge_page_split_desc_cache.kmem_cache = pte_list_desc_cache;
> > > + kvm->arch.huge_page_split_desc_cache.gfp_zero = __GFP_ZERO;
> > > }
> > >
> > > void kvm_mmu_uninit_vm(struct kvm *kvm)
> > > @@ -6035,9 +6040,42 @@ void kvm_mmu_slot_remove_write_access(struct kvm *kvm,
> > > kvm_arch_flush_remote_tlbs_memslot(kvm, memslot);
> > > }
> > >
> > > +static int min_descs_for_split(const struct kvm_memory_slot *slot, u64 *huge_sptep)
> > > +{
> > > + struct kvm_mmu_page *huge_sp = sptep_to_sp(huge_sptep);
> > > + int split_level = huge_sp->role.level - 1;
> > > + int i, min = 0;
> > > + gfn_t gfn;
> > > +
> > > + gfn = kvm_mmu_page_get_gfn(huge_sp, huge_sptep - huge_sp->spt);
> > >
> > > -static int alloc_memory_for_split(struct kvm *kvm, struct kvm_mmu_page **spp, gfp_t gfp)
> > > + for (i = 0; i < PT64_ENT_PER_PAGE; i++) {
> > > + if (rmap_need_new_pte_list_desc(slot, gfn, split_level))
> > > + min++;
> > > +
> > > + gfn += KVM_PAGES_PER_HPAGE(split_level);
> > > + }
> > > +
> > > + return min;
> > > +}
> >
> > Is this calculation worth doing? It seems like we're doing a lot of
> > work here to calculate exactly how many pages we need to allocate, but
> > if eager splitting we'll be doing this over and over again. It seems
> > like it would be more efficient to just always fill the cache since
> > any extra pages allocated to split one page can be used to split the
> > next one.
>
> topup_huge_page_split_desc_cache() does fill the cache. This
> calculation is just to determine the minimum number of objects needed
> to split the next huge page, so that we can skip refilling the cache
> when its unnecessary.
>
> I think you are suggesting we unconditionally topup the cache and
> hard-code the min to 513 (the capacity of the cache)? That would
> certainly allow us to drop this function (less code complexity) but
> would result in extra unnecessary allocations. If the cost of those
> allocations is negligible then I can see an argument for going with
> your approach.
Right, exactly.
If you're eagerly splitting the entire EPT for a VM, then the number
of extra allocations is bounded at 513 because memory allocated for
one page can be used for the next one if not needed right?
If you check how many you need on each pass, you'll be doing
potentially O(pages split) extra work, so I suspect that
unconditionally filling the cache will scale better.
>
> >
> > > +
> > > +static int topup_huge_page_split_desc_cache(struct kvm *kvm, int min, gfp_t gfp)
> > > +{
> > > + struct kvm_mmu_memory_cache *cache =
> > > + &kvm->arch.huge_page_split_desc_cache;
> > > +
> > > + return __kvm_mmu_topup_memory_cache(cache, min, gfp);
> > > +}
> > > +
> > > +static int alloc_memory_for_split(struct kvm *kvm, struct kvm_mmu_page **spp,
> > > + int min_descs, gfp_t gfp)
> > > {
> > > + int r;
> > > +
> > > + r = topup_huge_page_split_desc_cache(kvm, min_descs, gfp);
> > > + if (r)
> > > + return r;
> > > +
> > > if (*spp)
> > > return 0;
> > >
> > > @@ -6050,9 +6088,9 @@ static int prepare_to_split_huge_page(struct kvm *kvm,
> > > const struct kvm_memory_slot *slot,
> > > u64 *huge_sptep,
> > > struct kvm_mmu_page **spp,
> > > - bool *flush,
> > > bool *dropped_lock)
> > > {
> > > + int min_descs = min_descs_for_split(slot, huge_sptep);
> > > int r = 0;
> > >
> > > *dropped_lock = false;
> > > @@ -6063,22 +6101,18 @@ static int prepare_to_split_huge_page(struct kvm *kvm,
> > > if (need_resched() || rwlock_needbreak(&kvm->mmu_lock))
> > > goto drop_lock;
> > >
> > > - r = alloc_memory_for_split(kvm, spp, GFP_NOWAIT | __GFP_ACCOUNT);
> > > + r = alloc_memory_for_split(kvm, spp, min_descs, GFP_NOWAIT | __GFP_ACCOUNT);
> > > if (r)
> > > goto drop_lock;
> > >
> > > return 0;
> > >
> > > drop_lock:
> > > - if (*flush)
> > > - kvm_arch_flush_remote_tlbs_memslot(kvm, slot);
> > > -
> > > - *flush = false;
> > > *dropped_lock = true;
> > >
> > > write_unlock(&kvm->mmu_lock);
> > > cond_resched();
> > > - r = alloc_memory_for_split(kvm, spp, GFP_KERNEL_ACCOUNT);
> > > + r = alloc_memory_for_split(kvm, spp, min_descs, GFP_KERNEL_ACCOUNT);
> > > write_lock(&kvm->mmu_lock);
> > >
> > > return r;
> > > @@ -6122,10 +6156,10 @@ static struct kvm_mmu_page *kvm_mmu_get_sp_for_split(struct kvm *kvm,
> > >
> > > static int kvm_mmu_split_huge_page(struct kvm *kvm,
> > > const struct kvm_memory_slot *slot,
> > > - u64 *huge_sptep, struct kvm_mmu_page **spp,
> > > - bool *flush)
> > > + u64 *huge_sptep, struct kvm_mmu_page **spp)
> > >
> > > {
> > > + struct kvm_mmu_memory_cache *cache;
> > > struct kvm_mmu_page *split_sp;
> > > u64 huge_spte, split_spte;
> > > int split_level, index;
> > > @@ -6138,9 +6172,9 @@ static int kvm_mmu_split_huge_page(struct kvm *kvm,
> > > return -EOPNOTSUPP;
> > >
> > > /*
> > > - * Since we did not allocate pte_list_desc_structs for the split, we
> > > - * cannot add a new parent SPTE to parent_ptes. This should never happen
> > > - * in practice though since this is a fresh SP.
> > > + * We did not allocate an extra pte_list_desc struct to add huge_sptep
> > > + * to split_sp->parent_ptes. An extra pte_list_desc struct should never
> > > + * be necessary in practice though since split_sp is brand new.
> > > *
> > > * Note, this makes it safe to pass NULL to __link_shadow_page() below.
> > > */
> > > @@ -6151,6 +6185,7 @@ static int kvm_mmu_split_huge_page(struct kvm *kvm,
> > >
> > > split_level = split_sp->role.level;
> > > access = split_sp->role.access;
> > > + cache = &kvm->arch.huge_page_split_desc_cache;
> > >
> > > for (index = 0; index < PT64_ENT_PER_PAGE; index++) {
> > > split_sptep = &split_sp->spt[index];
> > > @@ -6158,25 +6193,11 @@ static int kvm_mmu_split_huge_page(struct kvm *kvm,
> > >
> > > BUG_ON(is_shadow_present_pte(*split_sptep));
> > >
> > > - /*
> > > - * Since we did not allocate pte_list_desc structs for the
> > > - * split, we can't add a new SPTE that maps this GFN.
> > > - * Skipping this SPTE means we're only partially mapping the
> > > - * huge page, which means we'll need to flush TLBs before
> > > - * dropping the MMU lock.
> > > - *
> > > - * Note, this make it safe to pass NULL to __rmap_add() below.
> > > - */
> > > - if (rmap_need_new_pte_list_desc(slot, split_gfn, split_level)) {
> > > - *flush = true;
> > > - continue;
> > > - }
> > > -
> > > split_spte = make_huge_page_split_spte(
> > > huge_spte, split_level + 1, index, access);
> > >
> > > mmu_spte_set(split_sptep, split_spte);
> > > - __rmap_add(kvm, NULL, slot, split_sptep, split_gfn, access);
> > > + __rmap_add(kvm, cache, slot, split_sptep, split_gfn, access);
> > > }
> > >
> > > /*
> > > @@ -6222,7 +6243,6 @@ static bool rmap_try_split_huge_pages(struct kvm *kvm,
> > > struct kvm_mmu_page *sp = NULL;
> > > struct rmap_iterator iter;
> > > u64 *huge_sptep, spte;
> > > - bool flush = false;
> > > bool dropped_lock;
> > > int level;
> > > gfn_t gfn;
> > > @@ -6237,7 +6257,7 @@ static bool rmap_try_split_huge_pages(struct kvm *kvm,
> > > level = sptep_to_sp(huge_sptep)->role.level;
> > > gfn = sptep_to_gfn(huge_sptep);
> > >
> > > - r = prepare_to_split_huge_page(kvm, slot, huge_sptep, &sp, &flush, &dropped_lock);
> > > + r = prepare_to_split_huge_page(kvm, slot, huge_sptep, &sp, &dropped_lock);
> > > if (r) {
> > > trace_kvm_mmu_split_huge_page(gfn, spte, level, r);
> > > break;
> > > @@ -6246,7 +6266,7 @@ static bool rmap_try_split_huge_pages(struct kvm *kvm,
> > > if (dropped_lock)
> > > goto restart;
> > >
> > > - r = kvm_mmu_split_huge_page(kvm, slot, huge_sptep, &sp, &flush);
> > > + r = kvm_mmu_split_huge_page(kvm, slot, huge_sptep, &sp);
> > >
> > > trace_kvm_mmu_split_huge_page(gfn, spte, level, r);
> > >
> > > @@ -6261,7 +6281,7 @@ static bool rmap_try_split_huge_pages(struct kvm *kvm,
> > > if (sp)
> > > kvm_mmu_free_sp(sp);
> > >
> > > - return flush;
> > > + return false;
> > > }
> > >
> > > static void kvm_rmap_try_split_huge_pages(struct kvm *kvm,
> > > @@ -6269,7 +6289,6 @@ static void kvm_rmap_try_split_huge_pages(struct kvm *kvm,
> > > gfn_t start, gfn_t end,
> > > int target_level)
> > > {
> > > - bool flush;
> > > int level;
> > >
> > > /*
> > > @@ -6277,21 +6296,15 @@ static void kvm_rmap_try_split_huge_pages(struct kvm *kvm,
> > > * down to the target level. This ensures pages are recursively split
> > > * all the way to the target level. There's no need to split pages
> > > * already at the target level.
> > > - *
> > > - * Note that TLB flushes must be done before dropping the MMU lock since
> > > - * rmap_try_split_huge_pages() may partially split any given huge page,
> > > - * i.e. it may effectively unmap (make non-present) a portion of the
> > > - * huge page.
> > > */
> > > for (level = KVM_MAX_HUGEPAGE_LEVEL; level > target_level; level--) {
> > > - flush = slot_handle_level_range(kvm, slot,
> > > - rmap_try_split_huge_pages,
> > > - level, level, start, end - 1,
> > > - true, flush);
> > > + slot_handle_level_range(kvm, slot,
> > > + rmap_try_split_huge_pages,
> > > + level, level, start, end - 1,
> > > + true, false);
> > > }
> > >
> > > - if (flush)
> > > - kvm_arch_flush_remote_tlbs_memslot(kvm, slot);
> > > + kvm_mmu_free_memory_cache(&kvm->arch.huge_page_split_desc_cache);
> > > }
> > >
> > > /* Must be called with the mmu_lock held in write-mode. */
> > > --
> > > 2.35.0.rc2.247.g8bbb082509-goog
> > >
