2 * Kernel-based Virtual Machine driver for Linux
4 * This module enables machines with Intel VT-x extensions to run virtual
5 * machines without emulation or binary translation.
9 * Copyright (C) 2006 Qumranet, Inc.
12 * Yaniv Kamay <yaniv@qumranet.com>
13 * Avi Kivity <avi@qumranet.com>
15 * This work is licensed under the terms of the GNU GPL, version 2. See
16 * the COPYING file in the top-level directory.
21 * We need the mmu code to access both 32-bit and 64-bit guest ptes,
22 * so the code in this file is compiled twice, once per pte size.
26 #define pt_element_t u64
27 #define guest_walker guest_walker64
28 #define shadow_walker shadow_walker64
29 #define FNAME(name) paging##64_##name
30 #define PT_BASE_ADDR_MASK PT64_BASE_ADDR_MASK
31 #define PT_DIR_BASE_ADDR_MASK PT64_DIR_BASE_ADDR_MASK
32 #define PT_INDEX(addr, level) PT64_INDEX(addr, level)
33 #define PT_LEVEL_MASK(level) PT64_LEVEL_MASK(level)
34 #define PT_LEVEL_BITS PT64_LEVEL_BITS
36 #define PT_MAX_FULL_LEVELS 4
37 #define CMPXCHG cmpxchg
39 #define CMPXCHG cmpxchg64
40 #define PT_MAX_FULL_LEVELS 2
43 #define pt_element_t u32
44 #define guest_walker guest_walker32
45 #define shadow_walker shadow_walker32
46 #define FNAME(name) paging##32_##name
47 #define PT_BASE_ADDR_MASK PT32_BASE_ADDR_MASK
48 #define PT_DIR_BASE_ADDR_MASK PT32_DIR_BASE_ADDR_MASK
49 #define PT_INDEX(addr, level) PT32_INDEX(addr, level)
50 #define PT_LEVEL_MASK(level) PT32_LEVEL_MASK(level)
51 #define PT_LEVEL_BITS PT32_LEVEL_BITS
52 #define PT_MAX_FULL_LEVELS 2
53 #define CMPXCHG cmpxchg
55 #error Invalid PTTYPE value
58 #define gpte_to_gfn FNAME(gpte_to_gfn)
59 #define gpte_to_gfn_pde FNAME(gpte_to_gfn_pde)
62 * The guest_walker structure emulates the behavior of the hardware page
67 gfn_t table_gfn[PT_MAX_FULL_LEVELS];
68 pt_element_t ptes[PT_MAX_FULL_LEVELS];
69 gpa_t pte_gpa[PT_MAX_FULL_LEVELS];
76 struct shadow_walker {
77 struct kvm_shadow_walk walker;
78 struct guest_walker *guest_walker;
88 static gfn_t gpte_to_gfn(pt_element_t gpte)
90 return (gpte & PT_BASE_ADDR_MASK) >> PAGE_SHIFT;
93 static gfn_t gpte_to_gfn_pde(pt_element_t gpte)
95 return (gpte & PT_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
98 static bool FNAME(cmpxchg_gpte)(struct kvm *kvm,
99 gfn_t table_gfn, unsigned index,
100 pt_element_t orig_pte, pt_element_t new_pte)
106 page = gfn_to_page(kvm, table_gfn);
108 table = kmap_atomic(page, KM_USER0);
109 ret = CMPXCHG(&table[index], orig_pte, new_pte);
110 kunmap_atomic(table, KM_USER0);
112 kvm_release_page_dirty(page);
114 return (ret != orig_pte);
117 static unsigned FNAME(gpte_access)(struct kvm_vcpu *vcpu, pt_element_t gpte)
121 access = (gpte & (PT_WRITABLE_MASK | PT_USER_MASK)) | ACC_EXEC_MASK;
124 access &= ~(gpte >> PT64_NX_SHIFT);
130 * Fetch a guest pte for a guest virtual address
132 static int FNAME(walk_addr)(struct guest_walker *walker,
133 struct kvm_vcpu *vcpu, gva_t addr,
134 int write_fault, int user_fault, int fetch_fault)
138 unsigned index, pt_access, pte_access;
141 pgprintk("%s: addr %lx\n", __func__, addr);
143 walker->level = vcpu->arch.mmu.root_level;
144 pte = vcpu->arch.cr3;
146 if (!is_long_mode(vcpu)) {
147 pte = vcpu->arch.pdptrs[(addr >> 30) & 3];
148 if (!is_present_pte(pte))
153 ASSERT((!is_long_mode(vcpu) && is_pae(vcpu)) ||
154 (vcpu->arch.cr3 & CR3_NONPAE_RESERVED_BITS) == 0);
159 index = PT_INDEX(addr, walker->level);
161 table_gfn = gpte_to_gfn(pte);
162 pte_gpa = gfn_to_gpa(table_gfn);
163 pte_gpa += index * sizeof(pt_element_t);
164 walker->table_gfn[walker->level - 1] = table_gfn;
165 walker->pte_gpa[walker->level - 1] = pte_gpa;
166 pgprintk("%s: table_gfn[%d] %lx\n", __func__,
167 walker->level - 1, table_gfn);
169 kvm_read_guest(vcpu->kvm, pte_gpa, &pte, sizeof(pte));
171 if (!is_present_pte(pte))
174 if (write_fault && !is_writeble_pte(pte))
175 if (user_fault || is_write_protection(vcpu))
178 if (user_fault && !(pte & PT_USER_MASK))
182 if (fetch_fault && is_nx(vcpu) && (pte & PT64_NX_MASK))
186 if (!(pte & PT_ACCESSED_MASK)) {
187 mark_page_dirty(vcpu->kvm, table_gfn);
188 if (FNAME(cmpxchg_gpte)(vcpu->kvm, table_gfn,
189 index, pte, pte|PT_ACCESSED_MASK))
191 pte |= PT_ACCESSED_MASK;
194 pte_access = pt_access & FNAME(gpte_access)(vcpu, pte);
196 walker->ptes[walker->level - 1] = pte;
198 if (walker->level == PT_PAGE_TABLE_LEVEL) {
199 walker->gfn = gpte_to_gfn(pte);
203 if (walker->level == PT_DIRECTORY_LEVEL
204 && (pte & PT_PAGE_SIZE_MASK)
205 && (PTTYPE == 64 || is_pse(vcpu))) {
206 walker->gfn = gpte_to_gfn_pde(pte);
207 walker->gfn += PT_INDEX(addr, PT_PAGE_TABLE_LEVEL);
208 if (PTTYPE == 32 && is_cpuid_PSE36())
209 walker->gfn += pse36_gfn_delta(pte);
213 pt_access = pte_access;
217 if (write_fault && !is_dirty_pte(pte)) {
220 mark_page_dirty(vcpu->kvm, table_gfn);
221 ret = FNAME(cmpxchg_gpte)(vcpu->kvm, table_gfn, index, pte,
225 pte |= PT_DIRTY_MASK;
226 kvm_mmu_pte_write(vcpu, pte_gpa, (u8 *)&pte, sizeof(pte), 0);
227 walker->ptes[walker->level - 1] = pte;
230 walker->pt_access = pt_access;
231 walker->pte_access = pte_access;
232 pgprintk("%s: pte %llx pte_access %x pt_access %x\n",
233 __func__, (u64)pte, pt_access, pte_access);
237 walker->error_code = 0;
241 walker->error_code = PFERR_PRESENT_MASK;
245 walker->error_code |= PFERR_WRITE_MASK;
247 walker->error_code |= PFERR_USER_MASK;
249 walker->error_code |= PFERR_FETCH_MASK;
253 static void FNAME(update_pte)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *page,
254 u64 *spte, const void *pte)
259 int largepage = vcpu->arch.update_pte.largepage;
261 gpte = *(const pt_element_t *)pte;
262 if (~gpte & (PT_PRESENT_MASK | PT_ACCESSED_MASK)) {
263 if (!is_present_pte(gpte))
264 set_shadow_pte(spte, shadow_notrap_nonpresent_pte);
267 pgprintk("%s: gpte %llx spte %p\n", __func__, (u64)gpte, spte);
268 pte_access = page->role.access & FNAME(gpte_access)(vcpu, gpte);
269 if (gpte_to_gfn(gpte) != vcpu->arch.update_pte.gfn)
271 pfn = vcpu->arch.update_pte.pfn;
272 if (is_error_pfn(pfn))
274 if (mmu_notifier_retry(vcpu, vcpu->arch.update_pte.mmu_seq))
277 mmu_set_spte(vcpu, spte, page->role.access, pte_access, 0, 0,
278 gpte & PT_DIRTY_MASK, NULL, largepage,
279 gpte & PT_GLOBAL_MASK, gpte_to_gfn(gpte),
284 * Fetch a shadow pte for a specific level in the paging hierarchy.
286 static int FNAME(shadow_walk_entry)(struct kvm_shadow_walk *_sw,
287 struct kvm_vcpu *vcpu, u64 addr,
288 u64 *sptep, int level)
290 struct shadow_walker *sw =
291 container_of(_sw, struct shadow_walker, walker);
292 struct guest_walker *gw = sw->guest_walker;
293 unsigned access = gw->pt_access;
294 struct kvm_mmu_page *shadow_page;
299 pt_element_t curr_pte;
301 if (level == PT_PAGE_TABLE_LEVEL
302 || (sw->largepage && level == PT_DIRECTORY_LEVEL)) {
303 mmu_set_spte(vcpu, sptep, access, gw->pte_access & access,
304 sw->user_fault, sw->write_fault,
305 gw->ptes[gw->level-1] & PT_DIRTY_MASK,
306 sw->ptwrite, sw->largepage,
307 gw->ptes[gw->level-1] & PT_GLOBAL_MASK,
308 gw->gfn, sw->pfn, false);
313 if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep))
316 if (is_large_pte(*sptep)) {
317 set_shadow_pte(sptep, shadow_trap_nonpresent_pte);
318 kvm_flush_remote_tlbs(vcpu->kvm);
319 rmap_remove(vcpu->kvm, sptep);
322 if (level == PT_DIRECTORY_LEVEL && gw->level == PT_DIRECTORY_LEVEL) {
324 if (!is_dirty_pte(gw->ptes[level - 1]))
325 access &= ~ACC_WRITE_MASK;
326 table_gfn = gpte_to_gfn(gw->ptes[level - 1]);
329 table_gfn = gw->table_gfn[level - 2];
331 shadow_page = kvm_mmu_get_page(vcpu, table_gfn, (gva_t)addr, level-1,
332 metaphysical, access, sptep);
334 r = kvm_read_guest_atomic(vcpu->kvm, gw->pte_gpa[level - 2],
335 &curr_pte, sizeof(curr_pte));
336 if (r || curr_pte != gw->ptes[level - 2]) {
337 kvm_mmu_put_page(shadow_page, sptep);
338 kvm_release_pfn_clean(sw->pfn);
344 spte = __pa(shadow_page->spt) | PT_PRESENT_MASK | PT_ACCESSED_MASK
345 | PT_WRITABLE_MASK | PT_USER_MASK;
350 static u64 *FNAME(fetch)(struct kvm_vcpu *vcpu, gva_t addr,
351 struct guest_walker *guest_walker,
352 int user_fault, int write_fault, int largepage,
353 int *ptwrite, pfn_t pfn)
355 struct shadow_walker walker = {
356 .walker = { .entry = FNAME(shadow_walk_entry), },
357 .guest_walker = guest_walker,
358 .user_fault = user_fault,
359 .write_fault = write_fault,
360 .largepage = largepage,
365 if (!is_present_pte(guest_walker->ptes[guest_walker->level - 1]))
368 walk_shadow(&walker.walker, vcpu, addr);
374 * Page fault handler. There are several causes for a page fault:
375 * - there is no shadow pte for the guest pte
376 * - write access through a shadow pte marked read only so that we can set
378 * - write access to a shadow pte marked read only so we can update the page
379 * dirty bitmap, when userspace requests it
380 * - mmio access; in this case we will never install a present shadow pte
381 * - normal guest page fault due to the guest pte marked not present, not
382 * writable, or not executable
384 * Returns: 1 if we need to emulate the instruction, 0 otherwise, or
385 * a negative value on error.
387 static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gva_t addr,
390 int write_fault = error_code & PFERR_WRITE_MASK;
391 int user_fault = error_code & PFERR_USER_MASK;
392 int fetch_fault = error_code & PFERR_FETCH_MASK;
393 struct guest_walker walker;
399 unsigned long mmu_seq;
401 pgprintk("%s: addr %lx err %x\n", __func__, addr, error_code);
402 kvm_mmu_audit(vcpu, "pre page fault");
404 r = mmu_topup_memory_caches(vcpu);
409 * Look up the shadow pte for the faulting address.
411 r = FNAME(walk_addr)(&walker, vcpu, addr, write_fault, user_fault,
415 * The page is not mapped by the guest. Let the guest handle it.
418 pgprintk("%s: guest page fault\n", __func__);
419 inject_page_fault(vcpu, addr, walker.error_code);
420 vcpu->arch.last_pt_write_count = 0; /* reset fork detector */
424 if (walker.level == PT_DIRECTORY_LEVEL) {
426 large_gfn = walker.gfn & ~(KVM_PAGES_PER_HPAGE-1);
427 if (is_largepage_backed(vcpu, large_gfn)) {
428 walker.gfn = large_gfn;
432 mmu_seq = vcpu->kvm->mmu_notifier_seq;
434 pfn = gfn_to_pfn(vcpu->kvm, walker.gfn);
437 if (is_error_pfn(pfn)) {
438 pgprintk("gfn %lx is mmio\n", walker.gfn);
439 kvm_release_pfn_clean(pfn);
443 spin_lock(&vcpu->kvm->mmu_lock);
444 if (mmu_notifier_retry(vcpu, mmu_seq))
446 kvm_mmu_free_some_pages(vcpu);
447 shadow_pte = FNAME(fetch)(vcpu, addr, &walker, user_fault, write_fault,
448 largepage, &write_pt, pfn);
450 pgprintk("%s: shadow pte %p %llx ptwrite %d\n", __func__,
451 shadow_pte, *shadow_pte, write_pt);
454 vcpu->arch.last_pt_write_count = 0; /* reset fork detector */
456 ++vcpu->stat.pf_fixed;
457 kvm_mmu_audit(vcpu, "post page fault (fixed)");
458 spin_unlock(&vcpu->kvm->mmu_lock);
463 spin_unlock(&vcpu->kvm->mmu_lock);
464 kvm_release_pfn_clean(pfn);
468 static int FNAME(shadow_invlpg_entry)(struct kvm_shadow_walk *_sw,
469 struct kvm_vcpu *vcpu, u64 addr,
470 u64 *sptep, int level)
472 struct shadow_walker *sw =
473 container_of(_sw, struct shadow_walker, walker);
475 /* FIXME: properly handle invlpg on large guest pages */
476 if (level == PT_PAGE_TABLE_LEVEL ||
477 ((level == PT_DIRECTORY_LEVEL) && is_large_pte(*sptep))) {
478 struct kvm_mmu_page *sp = page_header(__pa(sptep));
480 sw->pte_gpa = (sp->gfn << PAGE_SHIFT);
481 sw->pte_gpa += (sptep - sp->spt) * sizeof(pt_element_t);
483 if (is_shadow_present_pte(*sptep)) {
484 rmap_remove(vcpu->kvm, sptep);
485 if (is_large_pte(*sptep))
486 --vcpu->kvm->stat.lpages;
488 set_shadow_pte(sptep, shadow_trap_nonpresent_pte);
491 if (!is_shadow_present_pte(*sptep))
496 static void FNAME(invlpg)(struct kvm_vcpu *vcpu, gva_t gva)
499 struct shadow_walker walker = {
500 .walker = { .entry = FNAME(shadow_invlpg_entry), },
504 spin_lock(&vcpu->kvm->mmu_lock);
505 walk_shadow(&walker.walker, vcpu, gva);
506 spin_unlock(&vcpu->kvm->mmu_lock);
507 if (walker.pte_gpa == -1)
509 if (kvm_read_guest_atomic(vcpu->kvm, walker.pte_gpa, &gpte,
510 sizeof(pt_element_t)))
512 if (is_present_pte(gpte) && (gpte & PT_ACCESSED_MASK)) {
513 if (mmu_topup_memory_caches(vcpu))
515 kvm_mmu_pte_write(vcpu, walker.pte_gpa, (const u8 *)&gpte,
516 sizeof(pt_element_t), 0);
520 static gpa_t FNAME(gva_to_gpa)(struct kvm_vcpu *vcpu, gva_t vaddr)
522 struct guest_walker walker;
523 gpa_t gpa = UNMAPPED_GVA;
526 r = FNAME(walk_addr)(&walker, vcpu, vaddr, 0, 0, 0);
529 gpa = gfn_to_gpa(walker.gfn);
530 gpa |= vaddr & ~PAGE_MASK;
536 static void FNAME(prefetch_page)(struct kvm_vcpu *vcpu,
537 struct kvm_mmu_page *sp)
540 pt_element_t pt[256 / sizeof(pt_element_t)];
543 if (sp->role.metaphysical
544 || (PTTYPE == 32 && sp->role.level > PT_PAGE_TABLE_LEVEL)) {
545 nonpaging_prefetch_page(vcpu, sp);
549 pte_gpa = gfn_to_gpa(sp->gfn);
551 offset = sp->role.quadrant << PT64_LEVEL_BITS;
552 pte_gpa += offset * sizeof(pt_element_t);
555 for (i = 0; i < PT64_ENT_PER_PAGE; i += ARRAY_SIZE(pt)) {
556 r = kvm_read_guest_atomic(vcpu->kvm, pte_gpa, pt, sizeof pt);
557 pte_gpa += ARRAY_SIZE(pt) * sizeof(pt_element_t);
558 for (j = 0; j < ARRAY_SIZE(pt); ++j)
559 if (r || is_present_pte(pt[j]))
560 sp->spt[i+j] = shadow_trap_nonpresent_pte;
562 sp->spt[i+j] = shadow_notrap_nonpresent_pte;
567 * Using the cached information from sp->gfns is safe because:
568 * - The spte has a reference to the struct page, so the pfn for a given gfn
569 * can't change unless all sptes pointing to it are nuked first.
570 * - Alias changes zap the entire shadow cache.
572 static int FNAME(sync_page)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
574 int i, offset, nr_present;
576 offset = nr_present = 0;
579 offset = sp->role.quadrant << PT64_LEVEL_BITS;
581 for (i = 0; i < PT64_ENT_PER_PAGE; i++) {
585 gfn_t gfn = sp->gfns[i];
587 if (!is_shadow_present_pte(sp->spt[i]))
590 pte_gpa = gfn_to_gpa(sp->gfn);
591 pte_gpa += (i+offset) * sizeof(pt_element_t);
593 if (kvm_read_guest_atomic(vcpu->kvm, pte_gpa, &gpte,
594 sizeof(pt_element_t)))
597 if (gpte_to_gfn(gpte) != gfn || !is_present_pte(gpte) ||
598 !(gpte & PT_ACCESSED_MASK)) {
601 rmap_remove(vcpu->kvm, &sp->spt[i]);
602 if (is_present_pte(gpte))
603 nonpresent = shadow_trap_nonpresent_pte;
605 nonpresent = shadow_notrap_nonpresent_pte;
606 set_shadow_pte(&sp->spt[i], nonpresent);
611 pte_access = sp->role.access & FNAME(gpte_access)(vcpu, gpte);
612 set_spte(vcpu, &sp->spt[i], pte_access, 0, 0,
613 is_dirty_pte(gpte), 0, gpte & PT_GLOBAL_MASK, gfn,
614 spte_to_pfn(sp->spt[i]), true, false);
624 #undef PT_BASE_ADDR_MASK
627 #undef PT_DIR_BASE_ADDR_MASK
629 #undef PT_MAX_FULL_LEVELS
631 #undef gpte_to_gfn_pde