?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
KVM: MMU: Fix SMP shadow instantiation race
[linux-2.6/kvm.git] / drivers / kvm / mmu.c
blob92ac0d1106b417aeb2884833f530bf75cde649b2
1 /*
2 * Kernel-based Virtual Machine driver for Linux
3 *
4 * This module enables machines with Intel VT-x extensions to run virtual
5 * machines without emulation or binary translation.
6 *
7 * MMU support
8 *
9 * Copyright (C) 2006 Qumranet, Inc.
10 *
11 * Authors:
12 * Yaniv Kamay <yaniv@qumranet.com>
13 * Avi Kivity <avi@qumranet.com>
14 *
15 * This work is licensed under the terms of the GNU GPL, version 2. See
16 * the COPYING file in the top-level directory.
17 *
18 */
19
20 #include "vmx.h"
21 #include "kvm.h"
22 #include "x86.h"
23
24 #include <linux/types.h>
25 #include <linux/string.h>
26 #include <linux/mm.h>
27 #include <linux/highmem.h>
28 #include <linux/module.h>
29 #include <linux/swap.h>
30
31 #include <asm/page.h>
32 #include <asm/cmpxchg.h>
33 #include <asm/io.h>
34
35 #undef MMU_DEBUG
36
37 #undef AUDIT
38
39 #ifdef AUDIT
40 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
41 #else
42 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
43 #endif
44
45 #ifdef MMU_DEBUG
46
47 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
48 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
49
50 #else
51
52 #define pgprintk(x...) do { } while (0)
53 #define rmap_printk(x...) do { } while (0)
54
55 #endif
56
57 #if defined(MMU_DEBUG) || defined(AUDIT)
58 static int dbg = 1;
59 #endif
60
61 #ifndef MMU_DEBUG
62 #define ASSERT(x) do { } while (0)
63 #else
64 #define ASSERT(x) \
65 if (!(x)) { \
66 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
67 __FILE__, __LINE__, #x); \
68 }
69 #endif
70
71 #define PT64_PT_BITS 9
72 #define PT64_ENT_PER_PAGE (1 << PT64_PT_BITS)
73 #define PT32_PT_BITS 10
74 #define PT32_ENT_PER_PAGE (1 << PT32_PT_BITS)
75
76 #define PT_WRITABLE_SHIFT 1
77
78 #define PT_PRESENT_MASK (1ULL << 0)
79 #define PT_WRITABLE_MASK (1ULL << PT_WRITABLE_SHIFT)
80 #define PT_USER_MASK (1ULL << 2)
81 #define PT_PWT_MASK (1ULL << 3)
82 #define PT_PCD_MASK (1ULL << 4)
83 #define PT_ACCESSED_MASK (1ULL << 5)
84 #define PT_DIRTY_MASK (1ULL << 6)
85 #define PT_PAGE_SIZE_MASK (1ULL << 7)
86 #define PT_PAT_MASK (1ULL << 7)
87 #define PT_GLOBAL_MASK (1ULL << 8)
88 #define PT64_NX_SHIFT 63
89 #define PT64_NX_MASK (1ULL << PT64_NX_SHIFT)
90
91 #define PT_PAT_SHIFT 7
92 #define PT_DIR_PAT_SHIFT 12
93 #define PT_DIR_PAT_MASK (1ULL << PT_DIR_PAT_SHIFT)
94
95 #define PT32_DIR_PSE36_SIZE 4
96 #define PT32_DIR_PSE36_SHIFT 13
97 #define PT32_DIR_PSE36_MASK \
98 (((1ULL << PT32_DIR_PSE36_SIZE) - 1) << PT32_DIR_PSE36_SHIFT)
99
100
101 #define PT_FIRST_AVAIL_BITS_SHIFT 9
102 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
103
104 #define PT_SHADOW_IO_MARK (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
105
106 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
107
108 #define PT64_LEVEL_BITS 9
109
110 #define PT64_LEVEL_SHIFT(level) \
111 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
112
113 #define PT64_LEVEL_MASK(level) \
114 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
115
116 #define PT64_INDEX(address, level)\
117 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
118
119
120 #define PT32_LEVEL_BITS 10
121
122 #define PT32_LEVEL_SHIFT(level) \
123 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
124
125 #define PT32_LEVEL_MASK(level) \
126 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
127
128 #define PT32_INDEX(address, level)\
129 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
130
131
132 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
133 #define PT64_DIR_BASE_ADDR_MASK \
134 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
135
136 #define PT32_BASE_ADDR_MASK PAGE_MASK
137 #define PT32_DIR_BASE_ADDR_MASK \
138 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
139
140 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
141 | PT64_NX_MASK)
142
143 #define PFERR_PRESENT_MASK (1U << 0)
144 #define PFERR_WRITE_MASK (1U << 1)
145 #define PFERR_USER_MASK (1U << 2)
146 #define PFERR_FETCH_MASK (1U << 4)
147
148 #define PT64_ROOT_LEVEL 4
149 #define PT32_ROOT_LEVEL 2
150 #define PT32E_ROOT_LEVEL 3
151
152 #define PT_DIRECTORY_LEVEL 2
153 #define PT_PAGE_TABLE_LEVEL 1
154
155 #define RMAP_EXT 4
156
157 #define ACC_EXEC_MASK 1
158 #define ACC_WRITE_MASK PT_WRITABLE_MASK
159 #define ACC_USER_MASK PT_USER_MASK
160 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
161
162 struct kvm_rmap_desc {
163 u64 *shadow_ptes[RMAP_EXT];
164 struct kvm_rmap_desc *more;
165 };
166
167 static struct kmem_cache *pte_chain_cache;
168 static struct kmem_cache *rmap_desc_cache;
169 static struct kmem_cache *mmu_page_header_cache;
170
171 static u64 __read_mostly shadow_trap_nonpresent_pte;
172 static u64 __read_mostly shadow_notrap_nonpresent_pte;
173
174 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
175 {
176 shadow_trap_nonpresent_pte = trap_pte;
177 shadow_notrap_nonpresent_pte = notrap_pte;
178 }
179 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
180
181 static int is_write_protection(struct kvm_vcpu *vcpu)
182 {
183 return vcpu->cr0 & X86_CR0_WP;
184 }
185
186 static int is_cpuid_PSE36(void)
187 {
188 return 1;
189 }
190
191 static int is_nx(struct kvm_vcpu *vcpu)
192 {
193 return vcpu->shadow_efer & EFER_NX;
194 }
195
196 static int is_present_pte(unsigned long pte)
197 {
198 return pte & PT_PRESENT_MASK;
199 }
200
201 static int is_shadow_present_pte(u64 pte)
202 {
203 pte &= ~PT_SHADOW_IO_MARK;
204 return pte != shadow_trap_nonpresent_pte
205 && pte != shadow_notrap_nonpresent_pte;
206 }
207
208 static int is_writeble_pte(unsigned long pte)
209 {
210 return pte & PT_WRITABLE_MASK;
211 }
212
213 static int is_dirty_pte(unsigned long pte)
214 {
215 return pte & PT_DIRTY_MASK;
216 }
217
218 static int is_io_pte(unsigned long pte)
219 {
220 return pte & PT_SHADOW_IO_MARK;
221 }
222
223 static int is_rmap_pte(u64 pte)
224 {
225 return pte != shadow_trap_nonpresent_pte
226 && pte != shadow_notrap_nonpresent_pte;
227 }
228
229 static gfn_t pse36_gfn_delta(u32 gpte)
230 {
231 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
232
233 return (gpte & PT32_DIR_PSE36_MASK) << shift;
234 }
235
236 static void set_shadow_pte(u64 *sptep, u64 spte)
237 {
238 #ifdef CONFIG_X86_64
239 set_64bit((unsigned long *)sptep, spte);
240 #else
241 set_64bit((unsigned long long *)sptep, spte);
242 #endif
243 }
244
245 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
246 struct kmem_cache *base_cache, int min)
247 {
248 void *obj;
249
250 if (cache->nobjs >= min)
251 return 0;
252 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
253 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
254 if (!obj)
255 return -ENOMEM;
256 cache->objects[cache->nobjs++] = obj;
257 }
258 return 0;
259 }
260
261 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
262 {
263 while (mc->nobjs)
264 kfree(mc->objects[--mc->nobjs]);
265 }
266
267 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
268 int min)
269 {
270 struct page *page;
271
272 if (cache->nobjs >= min)
273 return 0;
274 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
275 page = alloc_page(GFP_KERNEL);
276 if (!page)
277 return -ENOMEM;
278 set_page_private(page, 0);
279 cache->objects[cache->nobjs++] = page_address(page);
280 }
281 return 0;
282 }
283
284 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
285 {
286 while (mc->nobjs)
287 free_page((unsigned long)mc->objects[--mc->nobjs]);
288 }
289
290 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
291 {
292 int r;
293
294 kvm_mmu_free_some_pages(vcpu);
295 r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
296 pte_chain_cache, 4);
297 if (r)
298 goto out;
299 r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
300 rmap_desc_cache, 1);
301 if (r)
302 goto out;
303 r = mmu_topup_memory_cache_page(&vcpu->mmu_page_cache, 8);
304 if (r)
305 goto out;
306 r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
307 mmu_page_header_cache, 4);
308 out:
309 return r;
310 }
311
312 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
313 {
314 mmu_free_memory_cache(&vcpu->mmu_pte_chain_cache);
315 mmu_free_memory_cache(&vcpu->mmu_rmap_desc_cache);
316 mmu_free_memory_cache_page(&vcpu->mmu_page_cache);
317 mmu_free_memory_cache(&vcpu->mmu_page_header_cache);
318 }
319
320 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
321 size_t size)
322 {
323 void *p;
324
325 BUG_ON(!mc->nobjs);
326 p = mc->objects[--mc->nobjs];
327 memset(p, 0, size);
328 return p;
329 }
330
331 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
332 {
333 return mmu_memory_cache_alloc(&vcpu->mmu_pte_chain_cache,
334 sizeof(struct kvm_pte_chain));
335 }
336
337 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
338 {
339 kfree(pc);
340 }
341
342 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
343 {
344 return mmu_memory_cache_alloc(&vcpu->mmu_rmap_desc_cache,
345 sizeof(struct kvm_rmap_desc));
346 }
347
348 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
349 {
350 kfree(rd);
351 }
352
353 /*
354 * Take gfn and return the reverse mapping to it.
355 * Note: gfn must be unaliased before this function get called
356 */
357
358 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn)
359 {
360 struct kvm_memory_slot *slot;
361
362 slot = gfn_to_memslot(kvm, gfn);
363 return &slot->rmap[gfn - slot->base_gfn];
364 }
365
366 /*
367 * Reverse mapping data structures:
368 *
369 * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
370 * that points to page_address(page).
371 *
372 * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
373 * containing more mappings.
374 */
375 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
376 {
377 struct kvm_mmu_page *sp;
378 struct kvm_rmap_desc *desc;
379 unsigned long *rmapp;
380 int i;
381
382 if (!is_rmap_pte(*spte))
383 return;
384 gfn = unalias_gfn(vcpu->kvm, gfn);
385 sp = page_header(__pa(spte));
386 sp->gfns[spte - sp->spt] = gfn;
387 rmapp = gfn_to_rmap(vcpu->kvm, gfn);
388 if (!*rmapp) {
389 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
390 *rmapp = (unsigned long)spte;
391 } else if (!(*rmapp & 1)) {
392 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
393 desc = mmu_alloc_rmap_desc(vcpu);
394 desc->shadow_ptes[0] = (u64 *)*rmapp;
395 desc->shadow_ptes[1] = spte;
396 *rmapp = (unsigned long)desc | 1;
397 } else {
398 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
399 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
400 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
401 desc = desc->more;
402 if (desc->shadow_ptes[RMAP_EXT-1]) {
403 desc->more = mmu_alloc_rmap_desc(vcpu);
404 desc = desc->more;
405 }
406 for (i = 0; desc->shadow_ptes[i]; ++i)
407 ;
408 desc->shadow_ptes[i] = spte;
409 }
410 }
411
412 static void rmap_desc_remove_entry(unsigned long *rmapp,
413 struct kvm_rmap_desc *desc,
414 int i,
415 struct kvm_rmap_desc *prev_desc)
416 {
417 int j;
418
419 for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
420 ;
421 desc->shadow_ptes[i] = desc->shadow_ptes[j];
422 desc->shadow_ptes[j] = NULL;
423 if (j != 0)
424 return;
425 if (!prev_desc && !desc->more)
426 *rmapp = (unsigned long)desc->shadow_ptes[0];
427 else
428 if (prev_desc)
429 prev_desc->more = desc->more;
430 else
431 *rmapp = (unsigned long)desc->more | 1;
432 mmu_free_rmap_desc(desc);
433 }
434
435 static void rmap_remove(struct kvm *kvm, u64 *spte)
436 {
437 struct kvm_rmap_desc *desc;
438 struct kvm_rmap_desc *prev_desc;
439 struct kvm_mmu_page *sp;
440 struct page *page;
441 unsigned long *rmapp;
442 int i;
443
444 if (!is_rmap_pte(*spte))
445 return;
446 sp = page_header(__pa(spte));
447 page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
448 mark_page_accessed(page);
449 if (is_writeble_pte(*spte))
450 kvm_release_page_dirty(page);
451 else
452 kvm_release_page_clean(page);
453 rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt]);
454 if (!*rmapp) {
455 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
456 BUG();
457 } else if (!(*rmapp & 1)) {
458 rmap_printk("rmap_remove: %p %llx 1->0\n", spte, *spte);
459 if ((u64 *)*rmapp != spte) {
460 printk(KERN_ERR "rmap_remove: %p %llx 1->BUG\n",
461 spte, *spte);
462 BUG();
463 }
464 *rmapp = 0;
465 } else {
466 rmap_printk("rmap_remove: %p %llx many->many\n", spte, *spte);
467 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
468 prev_desc = NULL;
469 while (desc) {
470 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
471 if (desc->shadow_ptes[i] == spte) {
472 rmap_desc_remove_entry(rmapp,
473 desc, i,
474 prev_desc);
475 return;
476 }
477 prev_desc = desc;
478 desc = desc->more;
479 }
480 BUG();
481 }
482 }
483
484 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
485 {
486 struct kvm_rmap_desc *desc;
487 struct kvm_rmap_desc *prev_desc;
488 u64 *prev_spte;
489 int i;
490
491 if (!*rmapp)
492 return NULL;
493 else if (!(*rmapp & 1)) {
494 if (!spte)
495 return (u64 *)*rmapp;
496 return NULL;
497 }
498 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
499 prev_desc = NULL;
500 prev_spte = NULL;
501 while (desc) {
502 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
503 if (prev_spte == spte)
504 return desc->shadow_ptes[i];
505 prev_spte = desc->shadow_ptes[i];
506 }
507 desc = desc->more;
508 }
509 return NULL;
510 }
511
512 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
513 {
514 unsigned long *rmapp;
515 u64 *spte;
516
517 gfn = unalias_gfn(kvm, gfn);
518 rmapp = gfn_to_rmap(kvm, gfn);
519
520 spte = rmap_next(kvm, rmapp, NULL);
521 while (spte) {
522 BUG_ON(!spte);
523 BUG_ON(!(*spte & PT_PRESENT_MASK));
524 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
525 if (is_writeble_pte(*spte))
526 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
527 kvm_flush_remote_tlbs(kvm);
528 spte = rmap_next(kvm, rmapp, spte);
529 }
530 }
531
532 #ifdef MMU_DEBUG
533 static int is_empty_shadow_page(u64 *spt)
534 {
535 u64 *pos;
536 u64 *end;
537
538 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
539 if ((*pos & ~PT_SHADOW_IO_MARK) != shadow_trap_nonpresent_pte) {
540 printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
541 pos, *pos);
542 return 0;
543 }
544 return 1;
545 }
546 #endif
547
548 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
549 {
550 ASSERT(is_empty_shadow_page(sp->spt));
551 list_del(&sp->link);
552 __free_page(virt_to_page(sp->spt));
553 __free_page(virt_to_page(sp->gfns));
554 kfree(sp);
555 ++kvm->n_free_mmu_pages;
556 }
557
558 static unsigned kvm_page_table_hashfn(gfn_t gfn)
559 {
560 return gfn;
561 }
562
563 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
564 u64 *parent_pte)
565 {
566 struct kvm_mmu_page *sp;
567
568 if (!vcpu->kvm->n_free_mmu_pages)
569 return NULL;
570
571 sp = mmu_memory_cache_alloc(&vcpu->mmu_page_header_cache, sizeof *sp);
572 sp->spt = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
573 sp->gfns = mmu_memory_cache_alloc(&vcpu->mmu_page_cache, PAGE_SIZE);
574 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
575 list_add(&sp->link, &vcpu->kvm->active_mmu_pages);
576 ASSERT(is_empty_shadow_page(sp->spt));
577 sp->slot_bitmap = 0;
578 sp->multimapped = 0;
579 sp->parent_pte = parent_pte;
580 --vcpu->kvm->n_free_mmu_pages;
581 return sp;
582 }
583
584 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
585 struct kvm_mmu_page *sp, u64 *parent_pte)
586 {
587 struct kvm_pte_chain *pte_chain;
588 struct hlist_node *node;
589 int i;
590
591 if (!parent_pte)
592 return;
593 if (!sp->multimapped) {
594 u64 *old = sp->parent_pte;
595
596 if (!old) {
597 sp->parent_pte = parent_pte;
598 return;
599 }
600 sp->multimapped = 1;
601 pte_chain = mmu_alloc_pte_chain(vcpu);
602 INIT_HLIST_HEAD(&sp->parent_ptes);
603 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
604 pte_chain->parent_ptes[0] = old;
605 }
606 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
607 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
608 continue;
609 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
610 if (!pte_chain->parent_ptes[i]) {
611 pte_chain->parent_ptes[i] = parent_pte;
612 return;
613 }
614 }
615 pte_chain = mmu_alloc_pte_chain(vcpu);
616 BUG_ON(!pte_chain);
617 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
618 pte_chain->parent_ptes[0] = parent_pte;
619 }
620
621 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
622 u64 *parent_pte)
623 {
624 struct kvm_pte_chain *pte_chain;
625 struct hlist_node *node;
626 int i;
627
628 if (!sp->multimapped) {
629 BUG_ON(sp->parent_pte != parent_pte);
630 sp->parent_pte = NULL;
631 return;
632 }
633 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
634 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
635 if (!pte_chain->parent_ptes[i])
636 break;
637 if (pte_chain->parent_ptes[i] != parent_pte)
638 continue;
639 while (i + 1 < NR_PTE_CHAIN_ENTRIES
640 && pte_chain->parent_ptes[i + 1]) {
641 pte_chain->parent_ptes[i]
642 = pte_chain->parent_ptes[i + 1];
643 ++i;
644 }
645 pte_chain->parent_ptes[i] = NULL;
646 if (i == 0) {
647 hlist_del(&pte_chain->link);
648 mmu_free_pte_chain(pte_chain);
649 if (hlist_empty(&sp->parent_ptes)) {
650 sp->multimapped = 0;
651 sp->parent_pte = NULL;
652 }
653 }
654 return;
655 }
656 BUG();
657 }
658
659 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
660 {
661 unsigned index;
662 struct hlist_head *bucket;
663 struct kvm_mmu_page *sp;
664 struct hlist_node *node;
665
666 pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
667 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
668 bucket = &kvm->mmu_page_hash[index];
669 hlist_for_each_entry(sp, node, bucket, hash_link)
670 if (sp->gfn == gfn && !sp->role.metaphysical) {
671 pgprintk("%s: found role %x\n",
672 __FUNCTION__, sp->role.word);
673 return sp;
674 }
675 return NULL;
676 }
677
678 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
679 gfn_t gfn,
680 gva_t gaddr,
681 unsigned level,
682 int metaphysical,
683 unsigned access,
684 u64 *parent_pte,
685 bool *new_page)
686 {
687 union kvm_mmu_page_role role;
688 unsigned index;
689 unsigned quadrant;
690 struct hlist_head *bucket;
691 struct kvm_mmu_page *sp;
692 struct hlist_node *node;
693
694 role.word = 0;
695 role.glevels = vcpu->mmu.root_level;
696 role.level = level;
697 role.metaphysical = metaphysical;
698 role.access = access;
699 if (vcpu->mmu.root_level <= PT32_ROOT_LEVEL) {
700 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
701 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
702 role.quadrant = quadrant;
703 }
704 pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
705 gfn, role.word);
706 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
707 bucket = &vcpu->kvm->mmu_page_hash[index];
708 hlist_for_each_entry(sp, node, bucket, hash_link)
709 if (sp->gfn == gfn && sp->role.word == role.word) {
710 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
711 pgprintk("%s: found\n", __FUNCTION__);
712 return sp;
713 }
714 sp = kvm_mmu_alloc_page(vcpu, parent_pte);
715 if (!sp)
716 return sp;
717 pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
718 sp->gfn = gfn;
719 sp->role = role;
720 hlist_add_head(&sp->hash_link, bucket);
721 vcpu->mmu.prefetch_page(vcpu, sp);
722 if (!metaphysical)
723 rmap_write_protect(vcpu->kvm, gfn);
724 if (new_page)
725 *new_page = 1;
726 return sp;
727 }
728
729 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
730 struct kvm_mmu_page *sp)
731 {
732 unsigned i;
733 u64 *pt;
734 u64 ent;
735
736 pt = sp->spt;
737
738 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
739 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
740 if (is_shadow_present_pte(pt[i]))
741 rmap_remove(kvm, &pt[i]);
742 pt[i] = shadow_trap_nonpresent_pte;
743 }
744 kvm_flush_remote_tlbs(kvm);
745 return;
746 }
747
748 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
749 ent = pt[i];
750
751 pt[i] = shadow_trap_nonpresent_pte;
752 if (!is_shadow_present_pte(ent))
753 continue;
754 ent &= PT64_BASE_ADDR_MASK;
755 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
756 }
757 kvm_flush_remote_tlbs(kvm);
758 }
759
760 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
761 {
762 mmu_page_remove_parent_pte(sp, parent_pte);
763 }
764
765 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
766 {
767 int i;
768
769 for (i = 0; i < KVM_MAX_VCPUS; ++i)
770 if (kvm->vcpus[i])
771 kvm->vcpus[i]->last_pte_updated = NULL;
772 }
773
774 static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
775 {
776 u64 *parent_pte;
777
778 ++kvm->stat.mmu_shadow_zapped;
779 while (sp->multimapped || sp->parent_pte) {
780 if (!sp->multimapped)
781 parent_pte = sp->parent_pte;
782 else {
783 struct kvm_pte_chain *chain;
784
785 chain = container_of(sp->parent_ptes.first,
786 struct kvm_pte_chain, link);
787 parent_pte = chain->parent_ptes[0];
788 }
789 BUG_ON(!parent_pte);
790 kvm_mmu_put_page(sp, parent_pte);
791 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
792 }
793 kvm_mmu_page_unlink_children(kvm, sp);
794 if (!sp->root_count) {
795 hlist_del(&sp->hash_link);
796 kvm_mmu_free_page(kvm, sp);
797 } else
798 list_move(&sp->link, &kvm->active_mmu_pages);
799 kvm_mmu_reset_last_pte_updated(kvm);
800 }
801
802 /*
803 * Changing the number of mmu pages allocated to the vm
804 * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
805 */
806 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
807 {
808 /*
809 * If we set the number of mmu pages to be smaller be than the
810 * number of actived pages , we must to free some mmu pages before we
811 * change the value
812 */
813
814 if ((kvm->n_alloc_mmu_pages - kvm->n_free_mmu_pages) >
815 kvm_nr_mmu_pages) {
816 int n_used_mmu_pages = kvm->n_alloc_mmu_pages
817 - kvm->n_free_mmu_pages;
818
819 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
820 struct kvm_mmu_page *page;
821
822 page = container_of(kvm->active_mmu_pages.prev,
823 struct kvm_mmu_page, link);
824 kvm_mmu_zap_page(kvm, page);
825 n_used_mmu_pages--;
826 }
827 kvm->n_free_mmu_pages = 0;
828 }
829 else
830 kvm->n_free_mmu_pages += kvm_nr_mmu_pages
831 - kvm->n_alloc_mmu_pages;
832
833 kvm->n_alloc_mmu_pages = kvm_nr_mmu_pages;
834 }
835
836 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
837 {
838 unsigned index;
839 struct hlist_head *bucket;
840 struct kvm_mmu_page *sp;
841 struct hlist_node *node, *n;
842 int r;
843
844 pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
845 r = 0;
846 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
847 bucket = &kvm->mmu_page_hash[index];
848 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
849 if (sp->gfn == gfn && !sp->role.metaphysical) {
850 pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
851 sp->role.word);
852 kvm_mmu_zap_page(kvm, sp);
853 r = 1;
854 }
855 return r;
856 }
857
858 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
859 {
860 struct kvm_mmu_page *sp;
861
862 while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
863 pgprintk("%s: zap %lx %x\n", __FUNCTION__, gfn, sp->role.word);
864 kvm_mmu_zap_page(kvm, sp);
865 }
866 }
867
868 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
869 {
870 int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
871 struct kvm_mmu_page *sp = page_header(__pa(pte));
872
873 __set_bit(slot, &sp->slot_bitmap);
874 }
875
876 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
877 {
878 gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
879
880 if (gpa == UNMAPPED_GVA)
881 return NULL;
882 return gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
883 }
884
885 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
886 unsigned pt_access, unsigned pte_access,
887 int user_fault, int write_fault, int dirty,
888 int *ptwrite, gfn_t gfn)
889 {
890 u64 spte;
891 int was_rmapped = is_rmap_pte(*shadow_pte);
892 struct page *page;
893
894 pgprintk("%s: spte %llx access %x write_fault %d"
895 " user_fault %d gfn %lx\n",
896 __FUNCTION__, *shadow_pte, pt_access,
897 write_fault, user_fault, gfn);
898
899 /*
900 * We don't set the accessed bit, since we sometimes want to see
901 * whether the guest actually used the pte (in order to detect
902 * demand paging).
903 */
904 spte = PT_PRESENT_MASK | PT_DIRTY_MASK;
905 if (!dirty)
906 pte_access &= ~ACC_WRITE_MASK;
907 if (!(pte_access & ACC_EXEC_MASK))
908 spte |= PT64_NX_MASK;
909
910 page = gfn_to_page(vcpu->kvm, gfn);
911
912 spte |= PT_PRESENT_MASK;
913 if (pte_access & ACC_USER_MASK)
914 spte |= PT_USER_MASK;
915
916 if (is_error_page(page)) {
917 set_shadow_pte(shadow_pte,
918 shadow_trap_nonpresent_pte | PT_SHADOW_IO_MARK);
919 kvm_release_page_clean(page);
920 return;
921 }
922
923 spte |= page_to_phys(page);
924
925 if ((pte_access & ACC_WRITE_MASK)
926 || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
927 struct kvm_mmu_page *shadow;
928
929 spte |= PT_WRITABLE_MASK;
930 if (user_fault) {
931 mmu_unshadow(vcpu->kvm, gfn);
932 goto unshadowed;
933 }
934
935 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
936 if (shadow) {
937 pgprintk("%s: found shadow page for %lx, marking ro\n",
938 __FUNCTION__, gfn);
939 pte_access &= ~ACC_WRITE_MASK;
940 if (is_writeble_pte(spte)) {
941 spte &= ~PT_WRITABLE_MASK;
942 kvm_x86_ops->tlb_flush(vcpu);
943 }
944 if (write_fault)
945 *ptwrite = 1;
946 }
947 }
948
949 unshadowed:
950
951 if (pte_access & ACC_WRITE_MASK)
952 mark_page_dirty(vcpu->kvm, gfn);
953
954 pgprintk("%s: setting spte %llx\n", __FUNCTION__, spte);
955 set_shadow_pte(shadow_pte, spte);
956 page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
957 if (!was_rmapped) {
958 rmap_add(vcpu, shadow_pte, gfn);
959 if (!is_rmap_pte(*shadow_pte))
960 kvm_release_page_clean(page);
961 }
962 else
963 kvm_release_page_clean(page);
964 if (!ptwrite || !*ptwrite)
965 vcpu->last_pte_updated = shadow_pte;
966 }
967
968 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
969 {
970 }
971
972 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
973 {
974 int level = PT32E_ROOT_LEVEL;
975 hpa_t table_addr = vcpu->mmu.root_hpa;
976 int pt_write = 0;
977
978 for (; ; level--) {
979 u32 index = PT64_INDEX(v, level);
980 u64 *table;
981
982 ASSERT(VALID_PAGE(table_addr));
983 table = __va(table_addr);
984
985 if (level == 1) {
986 mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
987 0, write, 1, &pt_write, gfn);
988 return pt_write || is_io_pte(table[index]);
989 }
990
991 if (table[index] == shadow_trap_nonpresent_pte) {
992 struct kvm_mmu_page *new_table;
993 gfn_t pseudo_gfn;
994
995 pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
996 >> PAGE_SHIFT;
997 new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
998 v, level - 1,
999 1, ACC_ALL, &table[index],
1000 NULL);
1001 if (!new_table) {
1002 pgprintk("nonpaging_map: ENOMEM\n");
1003 return -ENOMEM;
1004 }
1005
1006 table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
1007 | PT_WRITABLE_MASK | PT_USER_MASK;
1008 }
1009 table_addr = table[index] & PT64_BASE_ADDR_MASK;
1010 }
1011 }
1012
1013 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1014 struct kvm_mmu_page *sp)
1015 {
1016 int i;
1017
1018 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1019 sp->spt[i] = shadow_trap_nonpresent_pte;
1020 }
1021
1022 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1023 {
1024 int i;
1025 struct kvm_mmu_page *sp;
1026
1027 if (!VALID_PAGE(vcpu->mmu.root_hpa))
1028 return;
1029 #ifdef CONFIG_X86_64
1030 if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1031 hpa_t root = vcpu->mmu.root_hpa;
1032
1033 sp = page_header(root);
1034 --sp->root_count;
1035 vcpu->mmu.root_hpa = INVALID_PAGE;
1036 return;
1037 }
1038 #endif
1039 for (i = 0; i < 4; ++i) {
1040 hpa_t root = vcpu->mmu.pae_root[i];
1041
1042 if (root) {
1043 root &= PT64_BASE_ADDR_MASK;
1044 sp = page_header(root);
1045 --sp->root_count;
1046 }
1047 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1048 }
1049 vcpu->mmu.root_hpa = INVALID_PAGE;
1050 }
1051
1052 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1053 {
1054 int i;
1055 gfn_t root_gfn;
1056 struct kvm_mmu_page *sp;
1057
1058 root_gfn = vcpu->cr3 >> PAGE_SHIFT;
1059
1060 #ifdef CONFIG_X86_64
1061 if (vcpu->mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1062 hpa_t root = vcpu->mmu.root_hpa;
1063
1064 ASSERT(!VALID_PAGE(root));
1065 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1066 PT64_ROOT_LEVEL, 0, ACC_ALL, NULL, NULL);
1067 root = __pa(sp->spt);
1068 ++sp->root_count;
1069 vcpu->mmu.root_hpa = root;
1070 return;
1071 }
1072 #endif
1073 for (i = 0; i < 4; ++i) {
1074 hpa_t root = vcpu->mmu.pae_root[i];
1075
1076 ASSERT(!VALID_PAGE(root));
1077 if (vcpu->mmu.root_level == PT32E_ROOT_LEVEL) {
1078 if (!is_present_pte(vcpu->pdptrs[i])) {
1079 vcpu->mmu.pae_root[i] = 0;
1080 continue;
1081 }
1082 root_gfn = vcpu->pdptrs[i] >> PAGE_SHIFT;
1083 } else if (vcpu->mmu.root_level == 0)
1084 root_gfn = 0;
1085 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1086 PT32_ROOT_LEVEL, !is_paging(vcpu),
1087 ACC_ALL, NULL, NULL);
1088 root = __pa(sp->spt);
1089 ++sp->root_count;
1090 vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
1091 }
1092 vcpu->mmu.root_hpa = __pa(vcpu->mmu.pae_root);
1093 }
1094
1095 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1096 {
1097 return vaddr;
1098 }
1099
1100 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1101 u32 error_code)
1102 {
1103 gfn_t gfn;
1104 int r;
1105
1106 pgprintk("%s: gva %lx error %x\n", __FUNCTION__, gva, error_code);
1107 r = mmu_topup_memory_caches(vcpu);
1108 if (r)
1109 return r;
1110
1111 ASSERT(vcpu);
1112 ASSERT(VALID_PAGE(vcpu->mmu.root_hpa));
1113
1114 gfn = gva >> PAGE_SHIFT;
1115
1116 return nonpaging_map(vcpu, gva & PAGE_MASK,
1117 error_code & PFERR_WRITE_MASK, gfn);
1118 }
1119
1120 static void nonpaging_free(struct kvm_vcpu *vcpu)
1121 {
1122 mmu_free_roots(vcpu);
1123 }
1124
1125 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1126 {
1127 struct kvm_mmu *context = &vcpu->mmu;
1128
1129 context->new_cr3 = nonpaging_new_cr3;
1130 context->page_fault = nonpaging_page_fault;
1131 context->gva_to_gpa = nonpaging_gva_to_gpa;
1132 context->free = nonpaging_free;
1133 context->prefetch_page = nonpaging_prefetch_page;
1134 context->root_level = 0;
1135 context->shadow_root_level = PT32E_ROOT_LEVEL;
1136 context->root_hpa = INVALID_PAGE;
1137 return 0;
1138 }
1139
1140 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1141 {
1142 ++vcpu->stat.tlb_flush;
1143 kvm_x86_ops->tlb_flush(vcpu);
1144 }
1145
1146 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1147 {
1148 pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
1149 mmu_free_roots(vcpu);
1150 }
1151
1152 static void inject_page_fault(struct kvm_vcpu *vcpu,
1153 u64 addr,
1154 u32 err_code)
1155 {
1156 kvm_inject_page_fault(vcpu, addr, err_code);
1157 }
1158
1159 static void paging_free(struct kvm_vcpu *vcpu)
1160 {
1161 nonpaging_free(vcpu);
1162 }
1163
1164 #define PTTYPE 64
1165 #include "paging_tmpl.h"
1166 #undef PTTYPE
1167
1168 #define PTTYPE 32
1169 #include "paging_tmpl.h"
1170 #undef PTTYPE
1171
1172 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1173 {
1174 struct kvm_mmu *context = &vcpu->mmu;
1175
1176 ASSERT(is_pae(vcpu));
1177 context->new_cr3 = paging_new_cr3;
1178 context->page_fault = paging64_page_fault;
1179 context->gva_to_gpa = paging64_gva_to_gpa;
1180 context->prefetch_page = paging64_prefetch_page;
1181 context->free = paging_free;
1182 context->root_level = level;
1183 context->shadow_root_level = level;
1184 context->root_hpa = INVALID_PAGE;
1185 return 0;
1186 }
1187
1188 static int paging64_init_context(struct kvm_vcpu *vcpu)
1189 {
1190 return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1191 }
1192
1193 static int paging32_init_context(struct kvm_vcpu *vcpu)
1194 {
1195 struct kvm_mmu *context = &vcpu->mmu;
1196
1197 context->new_cr3 = paging_new_cr3;
1198 context->page_fault = paging32_page_fault;
1199 context->gva_to_gpa = paging32_gva_to_gpa;
1200 context->free = paging_free;
1201 context->prefetch_page = paging32_prefetch_page;
1202 context->root_level = PT32_ROOT_LEVEL;
1203 context->shadow_root_level = PT32E_ROOT_LEVEL;
1204 context->root_hpa = INVALID_PAGE;
1205 return 0;
1206 }
1207
1208 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1209 {
1210 return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1211 }
1212
1213 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1214 {
1215 ASSERT(vcpu);
1216 ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1217
1218 if (!is_paging(vcpu))
1219 return nonpaging_init_context(vcpu);
1220 else if (is_long_mode(vcpu))
1221 return paging64_init_context(vcpu);
1222 else if (is_pae(vcpu))
1223 return paging32E_init_context(vcpu);
1224 else
1225 return paging32_init_context(vcpu);
1226 }
1227
1228 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1229 {
1230 ASSERT(vcpu);
1231 if (VALID_PAGE(vcpu->mmu.root_hpa)) {
1232 vcpu->mmu.free(vcpu);
1233 vcpu->mmu.root_hpa = INVALID_PAGE;
1234 }
1235 }
1236
1237 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1238 {
1239 destroy_kvm_mmu(vcpu);
1240 return init_kvm_mmu(vcpu);
1241 }
1242 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1243
1244 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1245 {
1246 int r;
1247
1248 mutex_lock(&vcpu->kvm->lock);
1249 r = mmu_topup_memory_caches(vcpu);
1250 if (r)
1251 goto out;
1252 mmu_alloc_roots(vcpu);
1253 kvm_x86_ops->set_cr3(vcpu, vcpu->mmu.root_hpa);
1254 kvm_mmu_flush_tlb(vcpu);
1255 out:
1256 mutex_unlock(&vcpu->kvm->lock);
1257 return r;
1258 }
1259 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1260
1261 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1262 {
1263 mmu_free_roots(vcpu);
1264 }
1265
1266 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1267 struct kvm_mmu_page *sp,
1268 u64 *spte)
1269 {
1270 u64 pte;
1271 struct kvm_mmu_page *child;
1272
1273 pte = *spte;
1274 if (is_shadow_present_pte(pte)) {
1275 if (sp->role.level == PT_PAGE_TABLE_LEVEL)
1276 rmap_remove(vcpu->kvm, spte);
1277 else {
1278 child = page_header(pte & PT64_BASE_ADDR_MASK);
1279 mmu_page_remove_parent_pte(child, spte);
1280 }
1281 }
1282 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1283 }
1284
1285 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1286 struct kvm_mmu_page *sp,
1287 u64 *spte,
1288 const void *new, int bytes,
1289 int offset_in_pte)
1290 {
1291 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
1292 ++vcpu->kvm->stat.mmu_pde_zapped;
1293 return;
1294 }
1295
1296 ++vcpu->kvm->stat.mmu_pte_updated;
1297 if (sp->role.glevels == PT32_ROOT_LEVEL)
1298 paging32_update_pte(vcpu, sp, spte, new, bytes, offset_in_pte);
1299 else
1300 paging64_update_pte(vcpu, sp, spte, new, bytes, offset_in_pte);
1301 }
1302
1303 static bool need_remote_flush(u64 old, u64 new)
1304 {
1305 if (!is_shadow_present_pte(old))
1306 return false;
1307 if (!is_shadow_present_pte(new))
1308 return true;
1309 if ((old ^ new) & PT64_BASE_ADDR_MASK)
1310 return true;
1311 old ^= PT64_NX_MASK;
1312 new ^= PT64_NX_MASK;
1313 return (old & ~new & PT64_PERM_MASK) != 0;
1314 }
1315
1316 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1317 {
1318 if (need_remote_flush(old, new))
1319 kvm_flush_remote_tlbs(vcpu->kvm);
1320 else
1321 kvm_mmu_flush_tlb(vcpu);
1322 }
1323
1324 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1325 {
1326 u64 *spte = vcpu->last_pte_updated;
1327
1328 return !!(spte && (*spte & PT_ACCESSED_MASK));
1329 }
1330
1331 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1332 const u8 *new, int bytes)
1333 {
1334 gfn_t gfn = gpa >> PAGE_SHIFT;
1335 struct kvm_mmu_page *sp;
1336 struct hlist_node *node, *n;
1337 struct hlist_head *bucket;
1338 unsigned index;
1339 u64 entry;
1340 u64 *spte;
1341 unsigned offset = offset_in_page(gpa);
1342 unsigned pte_size;
1343 unsigned page_offset;
1344 unsigned misaligned;
1345 unsigned quadrant;
1346 int level;
1347 int flooded = 0;
1348 int npte;
1349
1350 pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1351 ++vcpu->kvm->stat.mmu_pte_write;
1352 kvm_mmu_audit(vcpu, "pre pte write");
1353 if (gfn == vcpu->last_pt_write_gfn
1354 && !last_updated_pte_accessed(vcpu)) {
1355 ++vcpu->last_pt_write_count;
1356 if (vcpu->last_pt_write_count >= 3)
1357 flooded = 1;
1358 } else {
1359 vcpu->last_pt_write_gfn = gfn;
1360 vcpu->last_pt_write_count = 1;
1361 vcpu->last_pte_updated = NULL;
1362 }
1363 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1364 bucket = &vcpu->kvm->mmu_page_hash[index];
1365 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1366 if (sp->gfn != gfn || sp->role.metaphysical)
1367 continue;
1368 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1369 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1370 misaligned |= bytes < 4;
1371 if (misaligned || flooded) {
1372 /*
1373 * Misaligned accesses are too much trouble to fix
1374 * up; also, they usually indicate a page is not used
1375 * as a page table.
1376 *
1377 * If we're seeing too many writes to a page,
1378 * it may no longer be a page table, or we may be
1379 * forking, in which case it is better to unmap the
1380 * page.
1381 */
1382 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1383 gpa, bytes, sp->role.word);
1384 kvm_mmu_zap_page(vcpu->kvm, sp);
1385 ++vcpu->kvm->stat.mmu_flooded;
1386 continue;
1387 }
1388 page_offset = offset;
1389 level = sp->role.level;
1390 npte = 1;
1391 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1392 page_offset <<= 1; /* 32->64 */
1393 /*
1394 * A 32-bit pde maps 4MB while the shadow pdes map
1395 * only 2MB. So we need to double the offset again
1396 * and zap two pdes instead of one.
1397 */
1398 if (level == PT32_ROOT_LEVEL) {
1399 page_offset &= ~7; /* kill rounding error */
1400 page_offset <<= 1;
1401 npte = 2;
1402 }
1403 quadrant = page_offset >> PAGE_SHIFT;
1404 page_offset &= ~PAGE_MASK;
1405 if (quadrant != sp->role.quadrant)
1406 continue;
1407 }
1408 spte = &sp->spt[page_offset / sizeof(*spte)];
1409 while (npte--) {
1410 entry = *spte;
1411 mmu_pte_write_zap_pte(vcpu, sp, spte);
1412 mmu_pte_write_new_pte(vcpu, sp, spte, new, bytes,
1413 page_offset & (pte_size - 1));
1414 mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1415 ++spte;
1416 }
1417 }
1418 kvm_mmu_audit(vcpu, "post pte write");
1419 }
1420
1421 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1422 {
1423 gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, gva);
1424
1425 return kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1426 }
1427
1428 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1429 {
1430 while (vcpu->kvm->n_free_mmu_pages < KVM_REFILL_PAGES) {
1431 struct kvm_mmu_page *sp;
1432
1433 sp = container_of(vcpu->kvm->active_mmu_pages.prev,
1434 struct kvm_mmu_page, link);
1435 kvm_mmu_zap_page(vcpu->kvm, sp);
1436 ++vcpu->kvm->stat.mmu_recycled;
1437 }
1438 }
1439
1440 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1441 {
1442 int r;
1443 enum emulation_result er;
1444
1445 mutex_lock(&vcpu->kvm->lock);
1446 r = vcpu->mmu.page_fault(vcpu, cr2, error_code);
1447 if (r < 0)
1448 goto out;
1449
1450 if (!r) {
1451 r = 1;
1452 goto out;
1453 }
1454
1455 r = mmu_topup_memory_caches(vcpu);
1456 if (r)
1457 goto out;
1458
1459 er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1460 mutex_unlock(&vcpu->kvm->lock);
1461
1462 switch (er) {
1463 case EMULATE_DONE:
1464 return 1;
1465 case EMULATE_DO_MMIO:
1466 ++vcpu->stat.mmio_exits;
1467 return 0;
1468 case EMULATE_FAIL:
1469 kvm_report_emulation_failure(vcpu, "pagetable");
1470 return 1;
1471 default:
1472 BUG();
1473 }
1474 out:
1475 mutex_unlock(&vcpu->kvm->lock);
1476 return r;
1477 }
1478 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1479
1480 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1481 {
1482 struct kvm_mmu_page *sp;
1483
1484 while (!list_empty(&vcpu->kvm->active_mmu_pages)) {
1485 sp = container_of(vcpu->kvm->active_mmu_pages.next,
1486 struct kvm_mmu_page, link);
1487 kvm_mmu_zap_page(vcpu->kvm, sp);
1488 }
1489 free_page((unsigned long)vcpu->mmu.pae_root);
1490 }
1491
1492 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1493 {
1494 struct page *page;
1495 int i;
1496
1497 ASSERT(vcpu);
1498
1499 if (vcpu->kvm->n_requested_mmu_pages)
1500 vcpu->kvm->n_free_mmu_pages = vcpu->kvm->n_requested_mmu_pages;
1501 else
1502 vcpu->kvm->n_free_mmu_pages = vcpu->kvm->n_alloc_mmu_pages;
1503 /*
1504 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1505 * Therefore we need to allocate shadow page tables in the first
1506 * 4GB of memory, which happens to fit the DMA32 zone.
1507 */
1508 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1509 if (!page)
1510 goto error_1;
1511 vcpu->mmu.pae_root = page_address(page);
1512 for (i = 0; i < 4; ++i)
1513 vcpu->mmu.pae_root[i] = INVALID_PAGE;
1514
1515 return 0;
1516
1517 error_1:
1518 free_mmu_pages(vcpu);
1519 return -ENOMEM;
1520 }
1521
1522 int kvm_mmu_create(struct kvm_vcpu *vcpu)
1523 {
1524 ASSERT(vcpu);
1525 ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1526
1527 return alloc_mmu_pages(vcpu);
1528 }
1529
1530 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1531 {
1532 ASSERT(vcpu);
1533 ASSERT(!VALID_PAGE(vcpu->mmu.root_hpa));
1534
1535 return init_kvm_mmu(vcpu);
1536 }
1537
1538 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1539 {
1540 ASSERT(vcpu);
1541
1542 destroy_kvm_mmu(vcpu);
1543 free_mmu_pages(vcpu);
1544 mmu_free_memory_caches(vcpu);
1545 }
1546
1547 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1548 {
1549 struct kvm_mmu_page *sp;
1550
1551 list_for_each_entry(sp, &kvm->active_mmu_pages, link) {
1552 int i;
1553 u64 *pt;
1554
1555 if (!test_bit(slot, &sp->slot_bitmap))
1556 continue;
1557
1558 pt = sp->spt;
1559 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1560 /* avoid RMW */
1561 if (pt[i] & PT_WRITABLE_MASK)
1562 pt[i] &= ~PT_WRITABLE_MASK;
1563 }
1564 }
1565
1566 void kvm_mmu_zap_all(struct kvm *kvm)
1567 {
1568 struct kvm_mmu_page *sp, *node;
1569
1570 list_for_each_entry_safe(sp, node, &kvm->active_mmu_pages, link)
1571 kvm_mmu_zap_page(kvm, sp);
1572
1573 kvm_flush_remote_tlbs(kvm);
1574 }
1575
1576 void kvm_mmu_module_exit(void)
1577 {
1578 if (pte_chain_cache)
1579 kmem_cache_destroy(pte_chain_cache);
1580 if (rmap_desc_cache)
1581 kmem_cache_destroy(rmap_desc_cache);
1582 if (mmu_page_header_cache)
1583 kmem_cache_destroy(mmu_page_header_cache);
1584 }
1585
1586 int kvm_mmu_module_init(void)
1587 {
1588 pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1589 sizeof(struct kvm_pte_chain),
1590 0, 0, NULL);
1591 if (!pte_chain_cache)
1592 goto nomem;
1593 rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1594 sizeof(struct kvm_rmap_desc),
1595 0, 0, NULL);
1596 if (!rmap_desc_cache)
1597 goto nomem;
1598
1599 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1600 sizeof(struct kvm_mmu_page),
1601 0, 0, NULL);
1602 if (!mmu_page_header_cache)
1603 goto nomem;
1604
1605 return 0;
1606
1607 nomem:
1608 kvm_mmu_module_exit();
1609 return -ENOMEM;
1610 }
1611
1612 /*
1613 * Caculate mmu pages needed for kvm.
1614 */
1615 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
1616 {
1617 int i;
1618 unsigned int nr_mmu_pages;
1619 unsigned int nr_pages = 0;
1620
1621 for (i = 0; i < kvm->nmemslots; i++)
1622 nr_pages += kvm->memslots[i].npages;
1623
1624 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
1625 nr_mmu_pages = max(nr_mmu_pages,
1626 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
1627
1628 return nr_mmu_pages;
1629 }
1630
1631 #ifdef AUDIT
1632
1633 static const char *audit_msg;
1634
1635 static gva_t canonicalize(gva_t gva)
1636 {
1637 #ifdef CONFIG_X86_64
1638 gva = (long long)(gva << 16) >> 16;
1639 #endif
1640 return gva;
1641 }
1642
1643 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1644 gva_t va, int level)
1645 {
1646 u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1647 int i;
1648 gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1649
1650 for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1651 u64 ent = pt[i];
1652
1653 if (ent == shadow_trap_nonpresent_pte)
1654 continue;
1655
1656 va = canonicalize(va);
1657 if (level > 1) {
1658 if (ent == shadow_notrap_nonpresent_pte)
1659 printk(KERN_ERR "audit: (%s) nontrapping pte"
1660 " in nonleaf level: levels %d gva %lx"
1661 " level %d pte %llx\n", audit_msg,
1662 vcpu->mmu.root_level, va, level, ent);
1663
1664 audit_mappings_page(vcpu, ent, va, level - 1);
1665 } else {
1666 gpa_t gpa = vcpu->mmu.gva_to_gpa(vcpu, va);
1667 struct page *page = gpa_to_page(vcpu, gpa);
1668 hpa_t hpa = page_to_phys(page);
1669
1670 if (is_shadow_present_pte(ent)
1671 && (ent & PT64_BASE_ADDR_MASK) != hpa)
1672 printk(KERN_ERR "xx audit error: (%s) levels %d"
1673 " gva %lx gpa %llx hpa %llx ent %llx %d\n",
1674 audit_msg, vcpu->mmu.root_level,
1675 va, gpa, hpa, ent,
1676 is_shadow_present_pte(ent));
1677 else if (ent == shadow_notrap_nonpresent_pte
1678 && !is_error_hpa(hpa))
1679 printk(KERN_ERR "audit: (%s) notrap shadow,"
1680 " valid guest gva %lx\n", audit_msg, va);
1681 kvm_release_page_clean(page);
1682
1683 }
1684 }
1685 }
1686
1687 static void audit_mappings(struct kvm_vcpu *vcpu)
1688 {
1689 unsigned i;
1690
1691 if (vcpu->mmu.root_level == 4)
1692 audit_mappings_page(vcpu, vcpu->mmu.root_hpa, 0, 4);
1693 else
1694 for (i = 0; i < 4; ++i)
1695 if (vcpu->mmu.pae_root[i] & PT_PRESENT_MASK)
1696 audit_mappings_page(vcpu,
1697 vcpu->mmu.pae_root[i],
1698 i << 30,
1699 2);
1700 }
1701
1702 static int count_rmaps(struct kvm_vcpu *vcpu)
1703 {
1704 int nmaps = 0;
1705 int i, j, k;
1706
1707 for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1708 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1709 struct kvm_rmap_desc *d;
1710
1711 for (j = 0; j < m->npages; ++j) {
1712 unsigned long *rmapp = &m->rmap[j];
1713
1714 if (!*rmapp)
1715 continue;
1716 if (!(*rmapp & 1)) {
1717 ++nmaps;
1718 continue;
1719 }
1720 d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
1721 while (d) {
1722 for (k = 0; k < RMAP_EXT; ++k)
1723 if (d->shadow_ptes[k])
1724 ++nmaps;
1725 else
1726 break;
1727 d = d->more;
1728 }
1729 }
1730 }
1731 return nmaps;
1732 }
1733
1734 static int count_writable_mappings(struct kvm_vcpu *vcpu)
1735 {
1736 int nmaps = 0;
1737 struct kvm_mmu_page *sp;
1738 int i;
1739
1740 list_for_each_entry(sp, &vcpu->kvm->active_mmu_pages, link) {
1741 u64 *pt = sp->spt;
1742
1743 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
1744 continue;
1745
1746 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1747 u64 ent = pt[i];
1748
1749 if (!(ent & PT_PRESENT_MASK))
1750 continue;
1751 if (!(ent & PT_WRITABLE_MASK))
1752 continue;
1753 ++nmaps;
1754 }
1755 }
1756 return nmaps;
1757 }
1758
1759 static void audit_rmap(struct kvm_vcpu *vcpu)
1760 {
1761 int n_rmap = count_rmaps(vcpu);
1762 int n_actual = count_writable_mappings(vcpu);
1763
1764 if (n_rmap != n_actual)
1765 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1766 __FUNCTION__, audit_msg, n_rmap, n_actual);
1767 }
1768
1769 static void audit_write_protection(struct kvm_vcpu *vcpu)
1770 {
1771 struct kvm_mmu_page *sp;
1772 struct kvm_memory_slot *slot;
1773 unsigned long *rmapp;
1774 gfn_t gfn;
1775
1776 list_for_each_entry(sp, &vcpu->kvm->active_mmu_pages, link) {
1777 if (sp->role.metaphysical)
1778 continue;
1779
1780 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
1781 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
1782 rmapp = &slot->rmap[gfn - slot->base_gfn];
1783 if (*rmapp)
1784 printk(KERN_ERR "%s: (%s) shadow page has writable"
1785 " mappings: gfn %lx role %x\n",
1786 __FUNCTION__, audit_msg, sp->gfn,
1787 sp->role.word);
1788 }
1789 }
1790
1791 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1792 {
1793 int olddbg = dbg;
1794
1795 dbg = 0;
1796 audit_msg = msg;
1797 audit_rmap(vcpu);
1798 audit_write_protection(vcpu);
1799 audit_mappings(vcpu);
1800 dbg = olddbg;
1801 }
1802
1803 #endif
kernel-based virtual machine