[PATCH] KVM: MMU: Page table write flood protection
authorAvi Kivity <avi@qumranet.com>
Sat, 6 Jan 2007 00:36:50 +0000 (16:36 -0800)
committerLinus Torvalds <torvalds@woody.osdl.org>
Sat, 6 Jan 2007 07:55:26 +0000 (23:55 -0800)
In fork() (or when we protect a page that is no longer a page table), we can
experience floods of writes to a page, which have to be emulated.  This is
expensive.

So, if we detect such a flood, zap the page so subsequent writes can proceed
natively.

Signed-off-by: Avi Kivity <avi@qumranet.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
drivers/kvm/kvm.h
drivers/kvm/mmu.c

index 6e4daf40414656a62a3b27b741d8e58e9741927d..201b2735ca9194919b135614acb8b25184578fd9 100644 (file)
@@ -238,6 +238,9 @@ struct kvm_vcpu {
        struct kvm_mmu_page page_header_buf[KVM_NUM_MMU_PAGES];
        struct kvm_mmu mmu;
 
+       gfn_t last_pt_write_gfn;
+       int   last_pt_write_count;
+
        struct kvm_guest_debug guest_debug;
 
        char fx_buf[FX_BUF_SIZE];
index 8cf3688f7e708e1e7aee529f3deb7a39adb79e0f..0e44aca9eee7e422f32a34b331c917b83741d198 100644 (file)
@@ -969,8 +969,17 @@ void kvm_mmu_pre_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
        unsigned page_offset;
        unsigned misaligned;
        int level;
+       int flooded = 0;
 
        pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
+       if (gfn == vcpu->last_pt_write_gfn) {
+               ++vcpu->last_pt_write_count;
+               if (vcpu->last_pt_write_count >= 3)
+                       flooded = 1;
+       } else {
+               vcpu->last_pt_write_gfn = gfn;
+               vcpu->last_pt_write_count = 1;
+       }
        index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
        bucket = &vcpu->kvm->mmu_page_hash[index];
        hlist_for_each_entry_safe(page, node, n, bucket, hash_link) {
@@ -978,11 +987,16 @@ void kvm_mmu_pre_write(struct kvm_vcpu *vcpu, gpa_t gpa, int bytes)
                        continue;
                pte_size = page->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
                misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
-               if (misaligned) {
+               if (misaligned || flooded) {
                        /*
                         * Misaligned accesses are too much trouble to fix
                         * up; also, they usually indicate a page is not used
                         * as a page table.
+                        *
+                        * If we're seeing too many writes to a page,
+                        * it may no longer be a page table, or we may be
+                        * forking, in which case it is better to unmap the
+                        * page.
                         */
                        pgprintk("misaligned: gpa %llx bytes %d role %x\n",
                                 gpa, bytes, page->role.word);