KVM: MMU: large page support
authorMarcelo Tosatti <marcelo@kvack.org>
Sat, 23 Feb 2008 14:44:30 +0000 (11:44 -0300)
committerAvi Kivity <avi@qumranet.com>
Sun, 27 Apr 2008 08:53:25 +0000 (11:53 +0300)
Create large pages mappings if the guest PTE's are marked as such and
the underlying memory is hugetlbfs backed.  If the largepage contains
write-protected pages, a large pte is not used.

Gives a consistent 2% improvement for data copies on ram mounted
filesystem, without NPT/EPT.

Anthony measures a 4% improvement on 4-way kernbench, with NPT.

Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@qumranet.com>
arch/x86/kvm/mmu.c
arch/x86/kvm/paging_tmpl.h
arch/x86/kvm/x86.c
include/asm-x86/kvm_host.h
include/linux/kvm_host.h
virt/kvm/kvm_main.c

index 103d008dab8bab3b29ebddd7e8877a912a03b14b..1932a3aeda1d3651171b401f45e81cdab9d2da19 100644 (file)
@@ -27,6 +27,7 @@
 #include <linux/highmem.h>
 #include <linux/module.h>
 #include <linux/swap.h>
+#include <linux/hugetlb.h>
 
 #include <asm/page.h>
 #include <asm/cmpxchg.h>
@@ -211,6 +212,11 @@ static int is_shadow_present_pte(u64 pte)
                && pte != shadow_notrap_nonpresent_pte;
 }
 
+static int is_large_pte(u64 pte)
+{
+       return pte & PT_PAGE_SIZE_MASK;
+}
+
 static int is_writeble_pte(unsigned long pte)
 {
        return pte & PT_WRITABLE_MASK;
@@ -349,17 +355,101 @@ static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
        kfree(rd);
 }
 
+/*
+ * Return the pointer to the largepage write count for a given
+ * gfn, handling slots that are not large page aligned.
+ */
+static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
+{
+       unsigned long idx;
+
+       idx = (gfn / KVM_PAGES_PER_HPAGE) -
+             (slot->base_gfn / KVM_PAGES_PER_HPAGE);
+       return &slot->lpage_info[idx].write_count;
+}
+
+static void account_shadowed(struct kvm *kvm, gfn_t gfn)
+{
+       int *write_count;
+
+       write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
+       *write_count += 1;
+       WARN_ON(*write_count > KVM_PAGES_PER_HPAGE);
+}
+
+static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
+{
+       int *write_count;
+
+       write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
+       *write_count -= 1;
+       WARN_ON(*write_count < 0);
+}
+
+static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
+{
+       struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
+       int *largepage_idx;
+
+       if (slot) {
+               largepage_idx = slot_largepage_idx(gfn, slot);
+               return *largepage_idx;
+       }
+
+       return 1;
+}
+
+static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
+{
+       struct vm_area_struct *vma;
+       unsigned long addr;
+
+       addr = gfn_to_hva(kvm, gfn);
+       if (kvm_is_error_hva(addr))
+               return 0;
+
+       vma = find_vma(current->mm, addr);
+       if (vma && is_vm_hugetlb_page(vma))
+               return 1;
+
+       return 0;
+}
+
+static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
+{
+       struct kvm_memory_slot *slot;
+
+       if (has_wrprotected_page(vcpu->kvm, large_gfn))
+               return 0;
+
+       if (!host_largepage_backed(vcpu->kvm, large_gfn))
+               return 0;
+
+       slot = gfn_to_memslot(vcpu->kvm, large_gfn);
+       if (slot && slot->dirty_bitmap)
+               return 0;
+
+       return 1;
+}
+
 /*
  * Take gfn and return the reverse mapping to it.
  * Note: gfn must be unaliased before this function get called
  */
 
-static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn)
+static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
 {
        struct kvm_memory_slot *slot;
+       unsigned long idx;
 
        slot = gfn_to_memslot(kvm, gfn);
-       return &slot->rmap[gfn - slot->base_gfn];
+       if (!lpage)
+               return &slot->rmap[gfn - slot->base_gfn];
+
+       idx = (gfn / KVM_PAGES_PER_HPAGE) -
+             (slot->base_gfn / KVM_PAGES_PER_HPAGE);
+
+       return &slot->lpage_info[idx].rmap_pde;
 }
 
 /*
@@ -371,7 +461,7 @@ static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn)
  * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
  * containing more mappings.
  */
-static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
+static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
 {
        struct kvm_mmu_page *sp;
        struct kvm_rmap_desc *desc;
@@ -383,7 +473,7 @@ static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
        gfn = unalias_gfn(vcpu->kvm, gfn);
        sp = page_header(__pa(spte));
        sp->gfns[spte - sp->spt] = gfn;
-       rmapp = gfn_to_rmap(vcpu->kvm, gfn);
+       rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
        if (!*rmapp) {
                rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
                *rmapp = (unsigned long)spte;
@@ -449,7 +539,7 @@ static void rmap_remove(struct kvm *kvm, u64 *spte)
                kvm_release_page_dirty(page);
        else
                kvm_release_page_clean(page);
-       rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt]);
+       rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
        if (!*rmapp) {
                printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
                BUG();
@@ -515,7 +605,7 @@ static void rmap_write_protect(struct kvm *kvm, u64 gfn)
        int write_protected = 0;
 
        gfn = unalias_gfn(kvm, gfn);
-       rmapp = gfn_to_rmap(kvm, gfn);
+       rmapp = gfn_to_rmap(kvm, gfn, 0);
 
        spte = rmap_next(kvm, rmapp, NULL);
        while (spte) {
@@ -528,8 +618,27 @@ static void rmap_write_protect(struct kvm *kvm, u64 gfn)
                }
                spte = rmap_next(kvm, rmapp, spte);
        }
+       /* check for huge page mappings */
+       rmapp = gfn_to_rmap(kvm, gfn, 1);
+       spte = rmap_next(kvm, rmapp, NULL);
+       while (spte) {
+               BUG_ON(!spte);
+               BUG_ON(!(*spte & PT_PRESENT_MASK));
+               BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
+               pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
+               if (is_writeble_pte(*spte)) {
+                       rmap_remove(kvm, spte);
+                       --kvm->stat.lpages;
+                       set_shadow_pte(spte, shadow_trap_nonpresent_pte);
+                       write_protected = 1;
+               }
+               spte = rmap_next(kvm, rmapp, spte);
+       }
+
        if (write_protected)
                kvm_flush_remote_tlbs(kvm);
+
+       account_shadowed(kvm, gfn);
 }
 
 #ifdef MMU_DEBUG
@@ -747,11 +856,17 @@ static void kvm_mmu_page_unlink_children(struct kvm *kvm,
        for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
                ent = pt[i];
 
+               if (is_shadow_present_pte(ent)) {
+                       if (!is_large_pte(ent)) {
+                               ent &= PT64_BASE_ADDR_MASK;
+                               mmu_page_remove_parent_pte(page_header(ent),
+                                                          &pt[i]);
+                       } else {
+                               --kvm->stat.lpages;
+                               rmap_remove(kvm, &pt[i]);
+                       }
+               }
                pt[i] = shadow_trap_nonpresent_pte;
-               if (!is_shadow_present_pte(ent))
-                       continue;
-               ent &= PT64_BASE_ADDR_MASK;
-               mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
        }
        kvm_flush_remote_tlbs(kvm);
 }
@@ -791,6 +906,8 @@ static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
        }
        kvm_mmu_page_unlink_children(kvm, sp);
        if (!sp->root_count) {
+               if (!sp->role.metaphysical)
+                       unaccount_shadowed(kvm, sp->gfn);
                hlist_del(&sp->hash_link);
                kvm_mmu_free_page(kvm, sp);
        } else {
@@ -894,7 +1011,8 @@ struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
                         unsigned pt_access, unsigned pte_access,
                         int user_fault, int write_fault, int dirty,
-                        int *ptwrite, gfn_t gfn, struct page *page)
+                        int *ptwrite, int largepage, gfn_t gfn,
+                        struct page *page)
 {
        u64 spte;
        int was_rmapped = 0;
@@ -907,15 +1025,29 @@ static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
                 write_fault, user_fault, gfn);
 
        if (is_rmap_pte(*shadow_pte)) {
-               if (host_pfn != page_to_pfn(page)) {
+               /*
+                * If we overwrite a PTE page pointer with a 2MB PMD, unlink
+                * the parent of the now unreachable PTE.
+                */
+               if (largepage && !is_large_pte(*shadow_pte)) {
+                       struct kvm_mmu_page *child;
+                       u64 pte = *shadow_pte;
+
+                       child = page_header(pte & PT64_BASE_ADDR_MASK);
+                       mmu_page_remove_parent_pte(child, shadow_pte);
+               } else if (host_pfn != page_to_pfn(page)) {
                        pgprintk("hfn old %lx new %lx\n",
                                 host_pfn, page_to_pfn(page));
                        rmap_remove(vcpu->kvm, shadow_pte);
+               } else {
+                       if (largepage)
+                               was_rmapped = is_large_pte(*shadow_pte);
+                       else
+                               was_rmapped = 1;
                }
-               else
-                       was_rmapped = 1;
        }
 
+
        /*
         * We don't set the accessed bit, since we sometimes want to see
         * whether the guest actually used the pte (in order to detect
@@ -930,6 +1062,8 @@ static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
        spte |= PT_PRESENT_MASK;
        if (pte_access & ACC_USER_MASK)
                spte |= PT_USER_MASK;
+       if (largepage)
+               spte |= PT_PAGE_SIZE_MASK;
 
        spte |= page_to_phys(page);
 
@@ -944,7 +1078,8 @@ static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
                }
 
                shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
-               if (shadow) {
+               if (shadow ||
+                  (largepage && has_wrprotected_page(vcpu->kvm, gfn))) {
                        pgprintk("%s: found shadow page for %lx, marking ro\n",
                                 __FUNCTION__, gfn);
                        pte_access &= ~ACC_WRITE_MASK;
@@ -963,10 +1098,17 @@ unshadowed:
                mark_page_dirty(vcpu->kvm, gfn);
 
        pgprintk("%s: setting spte %llx\n", __FUNCTION__, spte);
+       pgprintk("instantiating %s PTE (%s) at %d (%llx) addr %llx\n",
+                (spte&PT_PAGE_SIZE_MASK)? "2MB" : "4kB",
+                (spte&PT_WRITABLE_MASK)?"RW":"R", gfn, spte, shadow_pte);
        set_shadow_pte(shadow_pte, spte);
+       if (!was_rmapped && (spte & PT_PAGE_SIZE_MASK)
+           && (spte & PT_PRESENT_MASK))
+               ++vcpu->kvm->stat.lpages;
+
        page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
        if (!was_rmapped) {
-               rmap_add(vcpu, shadow_pte, gfn);
+               rmap_add(vcpu, shadow_pte, gfn, largepage);
                if (!is_rmap_pte(*shadow_pte))
                        kvm_release_page_clean(page);
        } else {
@@ -984,7 +1126,8 @@ static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
 }
 
 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
-                          gfn_t gfn, struct page *page, int level)
+                          int largepage, gfn_t gfn, struct page *page,
+                          int level)
 {
        hpa_t table_addr = vcpu->arch.mmu.root_hpa;
        int pt_write = 0;
@@ -998,7 +1141,13 @@ static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
 
                if (level == 1) {
                        mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
-                                    0, write, 1, &pt_write, gfn, page);
+                                    0, write, 1, &pt_write, 0, gfn, page);
+                       return pt_write;
+               }
+
+               if (largepage && level == 2) {
+                       mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
+                                   0, write, 1, &pt_write, 1, gfn, page);
                        return pt_write;
                }
 
@@ -1027,12 +1176,18 @@ static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
 {
        int r;
+       int largepage = 0;
 
        struct page *page;
 
        down_read(&vcpu->kvm->slots_lock);
 
        down_read(&current->mm->mmap_sem);
+       if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
+               gfn &= ~(KVM_PAGES_PER_HPAGE-1);
+               largepage = 1;
+       }
+
        page = gfn_to_page(vcpu->kvm, gfn);
        up_read(&current->mm->mmap_sem);
 
@@ -1045,7 +1200,8 @@ static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
 
        spin_lock(&vcpu->kvm->mmu_lock);
        kvm_mmu_free_some_pages(vcpu);
-       r = __direct_map(vcpu, v, write, gfn, page, PT32E_ROOT_LEVEL);
+       r = __direct_map(vcpu, v, write, largepage, gfn, page,
+                        PT32E_ROOT_LEVEL);
        spin_unlock(&vcpu->kvm->mmu_lock);
 
        up_read(&vcpu->kvm->slots_lock);
@@ -1180,6 +1336,8 @@ static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
 {
        struct page *page;
        int r;
+       int largepage = 0;
+       gfn_t gfn = gpa >> PAGE_SHIFT;
 
        ASSERT(vcpu);
        ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
@@ -1189,7 +1347,11 @@ static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
                return r;
 
        down_read(&current->mm->mmap_sem);
-       page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
+       if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
+               gfn &= ~(KVM_PAGES_PER_HPAGE-1);
+               largepage = 1;
+       }
+       page = gfn_to_page(vcpu->kvm, gfn);
        if (is_error_page(page)) {
                kvm_release_page_clean(page);
                up_read(&current->mm->mmap_sem);
@@ -1198,7 +1360,7 @@ static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
        spin_lock(&vcpu->kvm->mmu_lock);
        kvm_mmu_free_some_pages(vcpu);
        r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
-                        gpa >> PAGE_SHIFT, page, TDP_ROOT_LEVEL);
+                        largepage, gfn, page, TDP_ROOT_LEVEL);
        spin_unlock(&vcpu->kvm->mmu_lock);
        up_read(&current->mm->mmap_sem);
 
@@ -1397,7 +1559,8 @@ static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
 
        pte = *spte;
        if (is_shadow_present_pte(pte)) {
-               if (sp->role.level == PT_PAGE_TABLE_LEVEL)
+               if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
+                   is_large_pte(pte))
                        rmap_remove(vcpu->kvm, spte);
                else {
                        child = page_header(pte & PT64_BASE_ADDR_MASK);
@@ -1405,6 +1568,8 @@ static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
                }
        }
        set_shadow_pte(spte, shadow_trap_nonpresent_pte);
+       if (is_large_pte(pte))
+               --vcpu->kvm->stat.lpages;
 }
 
 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
@@ -1412,7 +1577,8 @@ static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
                                  u64 *spte,
                                  const void *new)
 {
-       if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
+       if ((sp->role.level != PT_PAGE_TABLE_LEVEL)
+           && !vcpu->arch.update_pte.largepage) {
                ++vcpu->kvm->stat.mmu_pde_zapped;
                return;
        }
@@ -1460,6 +1626,8 @@ static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
        u64 gpte = 0;
        struct page *page;
 
+       vcpu->arch.update_pte.largepage = 0;
+
        if (bytes != 4 && bytes != 8)
                return;
 
@@ -1487,9 +1655,13 @@ static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
                return;
        gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
 
-       down_read(&vcpu->kvm->slots_lock);
+       down_read(&current->mm->mmap_sem);
+       if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
+               gfn &= ~(KVM_PAGES_PER_HPAGE-1);
+               vcpu->arch.update_pte.largepage = 1;
+       }
        page = gfn_to_page(vcpu->kvm, gfn);
-       up_read(&vcpu->kvm->slots_lock);
+       up_read(&current->mm->mmap_sem);
 
        if (is_error_page(page)) {
                kvm_release_page_clean(page);
index 4b55f462e2b3bd8912882de0038d8ab10c1c4291..17f9d160ca34edb3c2a1a5f924ebb7f476ae2eed 100644 (file)
@@ -248,6 +248,7 @@ static void FNAME(update_pte)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *page,
        pt_element_t gpte;
        unsigned pte_access;
        struct page *npage;
+       int largepage = vcpu->arch.update_pte.largepage;
 
        gpte = *(const pt_element_t *)pte;
        if (~gpte & (PT_PRESENT_MASK | PT_ACCESSED_MASK)) {
@@ -264,7 +265,8 @@ static void FNAME(update_pte)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *page,
                return;
        get_page(npage);
        mmu_set_spte(vcpu, spte, page->role.access, pte_access, 0, 0,
-                    gpte & PT_DIRTY_MASK, NULL, gpte_to_gfn(gpte), npage);
+                    gpte & PT_DIRTY_MASK, NULL, largepage, gpte_to_gfn(gpte),
+                    npage);
 }
 
 /*
@@ -272,8 +274,8 @@ static void FNAME(update_pte)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *page,
  */
 static u64 *FNAME(fetch)(struct kvm_vcpu *vcpu, gva_t addr,
                         struct guest_walker *walker,
-                        int user_fault, int write_fault, int *ptwrite,
-                        struct page *page)
+                        int user_fault, int write_fault, int largepage,
+                        int *ptwrite, struct page *page)
 {
        hpa_t shadow_addr;
        int level;
@@ -301,11 +303,19 @@ static u64 *FNAME(fetch)(struct kvm_vcpu *vcpu, gva_t addr,
                shadow_ent = ((u64 *)__va(shadow_addr)) + index;
                if (level == PT_PAGE_TABLE_LEVEL)
                        break;
-               if (is_shadow_present_pte(*shadow_ent)) {
+
+               if (largepage && level == PT_DIRECTORY_LEVEL)
+                       break;
+
+               if (is_shadow_present_pte(*shadow_ent)
+                   && !is_large_pte(*shadow_ent)) {
                        shadow_addr = *shadow_ent & PT64_BASE_ADDR_MASK;
                        continue;
                }
 
+               if (is_large_pte(*shadow_ent))
+                       rmap_remove(vcpu->kvm, shadow_ent);
+
                if (level - 1 == PT_PAGE_TABLE_LEVEL
                    && walker->level == PT_DIRECTORY_LEVEL) {
                        metaphysical = 1;
@@ -339,7 +349,7 @@ static u64 *FNAME(fetch)(struct kvm_vcpu *vcpu, gva_t addr,
        mmu_set_spte(vcpu, shadow_ent, access, walker->pte_access & access,
                     user_fault, write_fault,
                     walker->ptes[walker->level-1] & PT_DIRTY_MASK,
-                    ptwrite, walker->gfn, page);
+                    ptwrite, largepage, walker->gfn, page);
 
        return shadow_ent;
 }
@@ -369,6 +379,7 @@ static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gva_t addr,
        int write_pt = 0;
        int r;
        struct page *page;
+       int largepage = 0;
 
        pgprintk("%s: addr %lx err %x\n", __FUNCTION__, addr, error_code);
        kvm_mmu_audit(vcpu, "pre page fault");
@@ -396,6 +407,14 @@ static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gva_t addr,
        }
 
        down_read(&current->mm->mmap_sem);
+       if (walker.level == PT_DIRECTORY_LEVEL) {
+               gfn_t large_gfn;
+               large_gfn = walker.gfn & ~(KVM_PAGES_PER_HPAGE-1);
+               if (is_largepage_backed(vcpu, large_gfn)) {
+                       walker.gfn = large_gfn;
+                       largepage = 1;
+               }
+       }
        page = gfn_to_page(vcpu->kvm, walker.gfn);
        up_read(&current->mm->mmap_sem);
 
@@ -410,7 +429,8 @@ static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gva_t addr,
        spin_lock(&vcpu->kvm->mmu_lock);
        kvm_mmu_free_some_pages(vcpu);
        shadow_pte = FNAME(fetch)(vcpu, addr, &walker, user_fault, write_fault,
-                                 &write_pt, page);
+                                 largepage, &write_pt, page);
+
        pgprintk("%s: shadow pte %p %llx ptwrite %d\n", __FUNCTION__,
                 shadow_pte, *shadow_pte, write_pt);
 
index e8e64927bddcf1bca138ba2c3dc19fce000493a0..0458bd51618502a74117bb5836e410c97c98397a 100644 (file)
@@ -88,6 +88,7 @@ struct kvm_stats_debugfs_item debugfs_entries[] = {
        { "mmu_recycled", VM_STAT(mmu_recycled) },
        { "mmu_cache_miss", VM_STAT(mmu_cache_miss) },
        { "remote_tlb_flush", VM_STAT(remote_tlb_flush) },
+       { "largepages", VM_STAT(lpages) },
        { NULL }
 };
 
index 8c3f74b7352452886767a3b62ef1d86a2a9d1629..95473ef5a906f9d8f039e7a82baf8057c403a76d 100644 (file)
 #define INVALID_PAGE (~(hpa_t)0)
 #define UNMAPPED_GVA (~(gpa_t)0)
 
+/* shadow tables are PAE even on non-PAE hosts */
+#define KVM_HPAGE_SHIFT 21
+#define KVM_HPAGE_SIZE (1UL << KVM_HPAGE_SHIFT)
+#define KVM_HPAGE_MASK (~(KVM_HPAGE_SIZE - 1))
+
+#define KVM_PAGES_PER_HPAGE (KVM_HPAGE_SIZE / PAGE_SIZE)
+
 #define DE_VECTOR 0
 #define UD_VECTOR 6
 #define NM_VECTOR 7
@@ -230,6 +237,7 @@ struct kvm_vcpu_arch {
        struct {
                gfn_t gfn;          /* presumed gfn during guest pte update */
                struct page *page;  /* page corresponding to that gfn */
+               int largepage;
        } update_pte;
 
        struct i387_fxsave_struct host_fx_image;
@@ -307,6 +315,7 @@ struct kvm_vm_stat {
        u32 mmu_recycled;
        u32 mmu_cache_miss;
        u32 remote_tlb_flush;
+       u32 lpages;
 };
 
 struct kvm_vcpu_stat {
index 994278fb5883447e9b9d255b35b66ba0c6d89c19..9750bb3c5a755d082ee9dd4695898a7b7a37e58c 100644 (file)
@@ -103,6 +103,10 @@ struct kvm_memory_slot {
        unsigned long flags;
        unsigned long *rmap;
        unsigned long *dirty_bitmap;
+       struct {
+               unsigned long rmap_pde;
+               int write_count;
+       } *lpage_info;
        unsigned long userspace_addr;
        int user_alloc;
 };
@@ -169,6 +173,7 @@ int kvm_arch_set_memory_region(struct kvm *kvm,
                                int user_alloc);
 gfn_t unalias_gfn(struct kvm *kvm, gfn_t gfn);
 struct page *gfn_to_page(struct kvm *kvm, gfn_t gfn);
+unsigned long gfn_to_hva(struct kvm *kvm, gfn_t gfn);
 void kvm_release_page_clean(struct page *page);
 void kvm_release_page_dirty(struct page *page);
 int kvm_read_guest_page(struct kvm *kvm, gfn_t gfn, void *data, int offset,
index c41eb57ce29b8c7ba6ffd2b78175b90ad8254bae..31db9b4d3016e91843aa94abf631edf3ebe864dc 100644 (file)
@@ -212,9 +212,13 @@ static void kvm_free_physmem_slot(struct kvm_memory_slot *free,
        if (!dont || free->dirty_bitmap != dont->dirty_bitmap)
                vfree(free->dirty_bitmap);
 
+       if (!dont || free->lpage_info != dont->lpage_info)
+               vfree(free->lpage_info);
+
        free->npages = 0;
        free->dirty_bitmap = NULL;
        free->rmap = NULL;
+       free->lpage_info = NULL;
 }
 
 void kvm_free_physmem(struct kvm *kvm)
@@ -324,6 +328,25 @@ int __kvm_set_memory_region(struct kvm *kvm,
                new.user_alloc = user_alloc;
                new.userspace_addr = mem->userspace_addr;
        }
+       if (npages && !new.lpage_info) {
+               int largepages = npages / KVM_PAGES_PER_HPAGE;
+               if (npages % KVM_PAGES_PER_HPAGE)
+                       largepages++;
+               if (base_gfn % KVM_PAGES_PER_HPAGE)
+                       largepages++;
+
+               new.lpage_info = vmalloc(largepages * sizeof(*new.lpage_info));
+
+               if (!new.lpage_info)
+                       goto out_free;
+
+               memset(new.lpage_info, 0, largepages * sizeof(*new.lpage_info));
+
+               if (base_gfn % KVM_PAGES_PER_HPAGE)
+                       new.lpage_info[0].write_count = 1;
+               if ((base_gfn+npages) % KVM_PAGES_PER_HPAGE)
+                       new.lpage_info[largepages-1].write_count = 1;
+       }
 
        /* Allocate page dirty bitmap if needed */
        if ((new.flags & KVM_MEM_LOG_DIRTY_PAGES) && !new.dirty_bitmap) {
@@ -467,7 +490,7 @@ int kvm_is_visible_gfn(struct kvm *kvm, gfn_t gfn)
 }
 EXPORT_SYMBOL_GPL(kvm_is_visible_gfn);
 
-static unsigned long gfn_to_hva(struct kvm *kvm, gfn_t gfn)
+unsigned long gfn_to_hva(struct kvm *kvm, gfn_t gfn)
 {
        struct kvm_memory_slot *slot;