KVM: MMU: Improve iteration through sptes from rmap
Takuya Yoshikawa [Wed, 21 Mar 2012 14:50:34 +0000 (23:50 +0900)]
Iteration using rmap_next(), the actual body is pte_list_next(), is
inefficient: every time we call it we start from checking whether rmap
holds a single spte or points to a descriptor which links more sptes.

In the case of shadow paging, this quadratic total iteration cost is a
problem.  Even for two dimensional paging, with EPT/NPT on, in which we
almost always have a single mapping, the extra checks at the end of the
iteration should be eliminated.

This patch fixes this by introducing rmap_iterator which keeps the
iteration context for the next search.  Furthermore the implementation
of rmap_next() is splitted into two functions, rmap_get_first() and
rmap_get_next(), to avoid repeatedly checking whether the rmap being
iterated on has only one spte.

Although there seemed to be only a slight change for EPT/NPT, the actual
improvement was significant: we observed that GET_DIRTY_LOG for 1GB
dirty memory became 15% faster than before.  This is probably because
the new code is easy to make branch predictions.

Note: we just remove pte_list_next() because we can think of parent_ptes
as a reverse mapping.

Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Signed-off-by: Avi Kivity <avi@redhat.com>

arch/x86/kvm/mmu.c
arch/x86/kvm/mmu_audit.c

index 3213348..29ad6f9 100644 (file)
@@ -842,32 +842,6 @@ static int pte_list_add(struct kvm_vcpu *vcpu, u64 *spte,
        return count;
 }
 
-static u64 *pte_list_next(unsigned long *pte_list, u64 *spte)
-{
-       struct pte_list_desc *desc;
-       u64 *prev_spte;
-       int i;
-
-       if (!*pte_list)
-               return NULL;
-       else if (!(*pte_list & 1)) {
-               if (!spte)
-                       return (u64 *)*pte_list;
-               return NULL;
-       }
-       desc = (struct pte_list_desc *)(*pte_list & ~1ul);
-       prev_spte = NULL;
-       while (desc) {
-               for (i = 0; i < PTE_LIST_EXT && desc->sptes[i]; ++i) {
-                       if (prev_spte == spte)
-                               return desc->sptes[i];
-                       prev_spte = desc->sptes[i];
-               }
-               desc = desc->more;
-       }
-       return NULL;
-}
-
 static void
 pte_list_desc_remove_entry(unsigned long *pte_list, struct pte_list_desc *desc,
                           int i, struct pte_list_desc *prev_desc)
@@ -988,11 +962,6 @@ static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
        return pte_list_add(vcpu, spte, rmapp);
 }
 
-static u64 *rmap_next(unsigned long *rmapp, u64 *spte)
-{
-       return pte_list_next(rmapp, spte);
-}
-
 static void rmap_remove(struct kvm *kvm, u64 *spte)
 {
        struct kvm_mmu_page *sp;
@@ -1005,6 +974,67 @@ static void rmap_remove(struct kvm *kvm, u64 *spte)
        pte_list_remove(spte, rmapp);
 }
 
+/*
+ * Used by the following functions to iterate through the sptes linked by a
+ * rmap.  All fields are private and not assumed to be used outside.
+ */
+struct rmap_iterator {
+       /* private fields */
+       struct pte_list_desc *desc;     /* holds the sptep if not NULL */
+       int pos;                        /* index of the sptep */
+};
+
+/*
+ * Iteration must be started by this function.  This should also be used after
+ * removing/dropping sptes from the rmap link because in such cases the
+ * information in the itererator may not be valid.
+ *
+ * Returns sptep if found, NULL otherwise.
+ */
+static u64 *rmap_get_first(unsigned long rmap, struct rmap_iterator *iter)
+{
+       if (!rmap)
+               return NULL;
+
+       if (!(rmap & 1)) {
+               iter->desc = NULL;
+               return (u64 *)rmap;
+       }
+
+       iter->desc = (struct pte_list_desc *)(rmap & ~1ul);
+       iter->pos = 0;
+       return iter->desc->sptes[iter->pos];
+}
+
+/*
+ * Must be used with a valid iterator: e.g. after rmap_get_first().
+ *
+ * Returns sptep if found, NULL otherwise.
+ */
+static u64 *rmap_get_next(struct rmap_iterator *iter)
+{
+       if (iter->desc) {
+               if (iter->pos < PTE_LIST_EXT - 1) {
+                       u64 *sptep;
+
+                       ++iter->pos;
+                       sptep = iter->desc->sptes[iter->pos];
+                       if (sptep)
+                               return sptep;
+               }
+
+               iter->desc = iter->desc->more;
+
+               if (iter->desc) {
+                       iter->pos = 0;
+                       /* desc->sptes[0] cannot be NULL */
+                       return iter->desc->sptes[iter->pos];
+               }
+       }
+
+       return NULL;
+}
+
 static void drop_spte(struct kvm *kvm, u64 *sptep)
 {
        if (mmu_spte_clear_track_bits(sptep))
@@ -1013,23 +1043,27 @@ static void drop_spte(struct kvm *kvm, u64 *sptep)
 
 static int __rmap_write_protect(struct kvm *kvm, unsigned long *rmapp, int level)
 {
-       u64 *spte = NULL;
+       u64 *sptep;
+       struct rmap_iterator iter;
        int write_protected = 0;
 
-       while ((spte = rmap_next(rmapp, spte))) {
-               BUG_ON(!(*spte & PT_PRESENT_MASK));
-               rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
+       for (sptep = rmap_get_first(*rmapp, &iter); sptep;) {
+               BUG_ON(!(*sptep & PT_PRESENT_MASK));
+               rmap_printk("rmap_write_protect: spte %p %llx\n", sptep, *sptep);
 
-               if (!is_writable_pte(*spte))
+               if (!is_writable_pte(*sptep)) {
+                       sptep = rmap_get_next(&iter);
                        continue;
+               }
 
                if (level == PT_PAGE_TABLE_LEVEL) {
-                       mmu_spte_update(spte, *spte & ~PT_WRITABLE_MASK);
+                       mmu_spte_update(sptep, *sptep & ~PT_WRITABLE_MASK);
+                       sptep = rmap_get_next(&iter);
                } else {
-                       BUG_ON(!is_large_pte(*spte));
-                       drop_spte(kvm, spte);
+                       BUG_ON(!is_large_pte(*sptep));
+                       drop_spte(kvm, sptep);
                        --kvm->stat.lpages;
-                       spte = NULL;
+                       sptep = rmap_get_first(*rmapp, &iter);
                }
 
                write_protected = 1;
@@ -1084,48 +1118,57 @@ static int rmap_write_protect(struct kvm *kvm, u64 gfn)
 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
                           unsigned long data)
 {
-       u64 *spte;
+       u64 *sptep;
+       struct rmap_iterator iter;
        int need_tlb_flush = 0;
 
-       while ((spte = rmap_next(rmapp, NULL))) {
-               BUG_ON(!(*spte & PT_PRESENT_MASK));
-               rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
-               drop_spte(kvm, spte);
+       while ((sptep = rmap_get_first(*rmapp, &iter))) {
+               BUG_ON(!(*sptep & PT_PRESENT_MASK));
+               rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", sptep, *sptep);
+
+               drop_spte(kvm, sptep);
                need_tlb_flush = 1;
        }
+
        return need_tlb_flush;
 }
 
 static int kvm_set_pte_rmapp(struct kvm *kvm, unsigned long *rmapp,
                             unsigned long data)
 {
+       u64 *sptep;
+       struct rmap_iterator iter;
        int need_flush = 0;
-       u64 *spte, new_spte;
+       u64 new_spte;
        pte_t *ptep = (pte_t *)data;
        pfn_t new_pfn;
 
        WARN_ON(pte_huge(*ptep));
        new_pfn = pte_pfn(*ptep);
-       spte = rmap_next(rmapp, NULL);
-       while (spte) {
-               BUG_ON(!is_shadow_present_pte(*spte));
-               rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte, *spte);
+
+       for (sptep = rmap_get_first(*rmapp, &iter); sptep;) {
+               BUG_ON(!is_shadow_present_pte(*sptep));
+               rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", sptep, *sptep);
+
                need_flush = 1;
+
                if (pte_write(*ptep)) {
-                       drop_spte(kvm, spte);
-                       spte = rmap_next(rmapp, NULL);
+                       drop_spte(kvm, sptep);
+                       sptep = rmap_get_first(*rmapp, &iter);
                } else {
-                       new_spte = *spte &~ (PT64_BASE_ADDR_MASK);
+                       new_spte = *sptep & ~PT64_BASE_ADDR_MASK;
                        new_spte |= (u64)new_pfn << PAGE_SHIFT;
 
                        new_spte &= ~PT_WRITABLE_MASK;
                        new_spte &= ~SPTE_HOST_WRITEABLE;
                        new_spte &= ~shadow_accessed_mask;
-                       mmu_spte_clear_track_bits(spte);
-                       mmu_spte_set(spte, new_spte);
-                       spte = rmap_next(rmapp, spte);
+
+                       mmu_spte_clear_track_bits(sptep);
+                       mmu_spte_set(sptep, new_spte);
+                       sptep = rmap_get_next(&iter);
                }
        }
+
        if (need_flush)
                kvm_flush_remote_tlbs(kvm);
 
@@ -1184,7 +1227,8 @@ void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
                         unsigned long data)
 {
-       u64 *spte;
+       u64 *sptep;
+       struct rmap_iterator iter;
        int young = 0;
 
        /*
@@ -1197,25 +1241,24 @@ static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
        if (!shadow_accessed_mask)
                return kvm_unmap_rmapp(kvm, rmapp, data);
 
-       spte = rmap_next(rmapp, NULL);
-       while (spte) {
-               int _young;
-               u64 _spte = *spte;
-               BUG_ON(!(_spte & PT_PRESENT_MASK));
-               _young = _spte & PT_ACCESSED_MASK;
-               if (_young) {
+       for (sptep = rmap_get_first(*rmapp, &iter); sptep;
+            sptep = rmap_get_next(&iter)) {
+               BUG_ON(!(*sptep & PT_PRESENT_MASK));
+
+               if (*sptep & PT_ACCESSED_MASK) {
                        young = 1;
-                       clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
+                       clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)sptep);
                }
-               spte = rmap_next(rmapp, spte);
        }
+
        return young;
 }
 
 static int kvm_test_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
                              unsigned long data)
 {
-       u64 *spte;
+       u64 *sptep;
+       struct rmap_iterator iter;
        int young = 0;
 
        /*
@@ -1226,16 +1269,14 @@ static int kvm_test_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
        if (!shadow_accessed_mask)
                goto out;
 
-       spte = rmap_next(rmapp, NULL);
-       while (spte) {
-               u64 _spte = *spte;
-               BUG_ON(!(_spte & PT_PRESENT_MASK));
-               young = _spte & PT_ACCESSED_MASK;
-               if (young) {
+       for (sptep = rmap_get_first(*rmapp, &iter); sptep;
+            sptep = rmap_get_next(&iter)) {
+               BUG_ON(!(*sptep & PT_PRESENT_MASK));
+
+               if (*sptep & PT_ACCESSED_MASK) {
                        young = 1;
                        break;
                }
-               spte = rmap_next(rmapp, spte);
        }
 out:
        return young;
@@ -1887,10 +1928,11 @@ static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
 
 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
 {
-       u64 *parent_pte;
+       u64 *sptep;
+       struct rmap_iterator iter;
 
-       while ((parent_pte = pte_list_next(&sp->parent_ptes, NULL)))
-               drop_parent_pte(sp, parent_pte);
+       while ((sptep = rmap_get_first(sp->parent_ptes, &iter)))
+               drop_parent_pte(sp, sptep);
 }
 
 static int mmu_zap_unsync_children(struct kvm *kvm,
index 715da5a..7d7d0b9 100644 (file)
@@ -192,7 +192,8 @@ static void audit_write_protection(struct kvm *kvm, struct kvm_mmu_page *sp)
 {
        struct kvm_memory_slot *slot;
        unsigned long *rmapp;
-       u64 *spte;
+       u64 *sptep;
+       struct rmap_iterator iter;
 
        if (sp->role.direct || sp->unsync || sp->role.invalid)
                return;
@@ -200,13 +201,12 @@ static void audit_write_protection(struct kvm *kvm, struct kvm_mmu_page *sp)
        slot = gfn_to_memslot(kvm, sp->gfn);
        rmapp = &slot->rmap[sp->gfn - slot->base_gfn];
 
-       spte = rmap_next(rmapp, NULL);
-       while (spte) {
-               if (is_writable_pte(*spte))
+       for (sptep = rmap_get_first(*rmapp, &iter); sptep;
+            sptep = rmap_get_next(&iter)) {
+               if (is_writable_pte(*sptep))
                        audit_printk(kvm, "shadow page has writable "
                                     "mappings: gfn %llx role %x\n",
                                     sp->gfn, sp->role.word);
-               spte = rmap_next(rmapp, spte);
        }
 }