aboutsummaryrefslogtreecommitdiffstats
path: root/arch/x86/kvm/mmu.c
diff options
context:
space:
mode:
Diffstat (limited to 'arch/x86/kvm/mmu.c')
-rw-r--r--arch/x86/kvm/mmu.c1885
1 files changed, 1885 insertions, 0 deletions
diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
new file mode 100644
index 000000000000..8efdcdbebb03
--- /dev/null
+++ b/arch/x86/kvm/mmu.c
@@ -0,0 +1,1885 @@
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 "mmu.h"
22
23#include <linux/kvm_host.h>
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
40static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
41#else
42static 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)
58static 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
162struct kvm_rmap_desc {
163 u64 *shadow_ptes[RMAP_EXT];
164 struct kvm_rmap_desc *more;
165};
166
167static struct kmem_cache *pte_chain_cache;
168static struct kmem_cache *rmap_desc_cache;
169static struct kmem_cache *mmu_page_header_cache;
170
171static u64 __read_mostly shadow_trap_nonpresent_pte;
172static u64 __read_mostly shadow_notrap_nonpresent_pte;
173
174void 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}
179EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
180
181static int is_write_protection(struct kvm_vcpu *vcpu)
182{
183 return vcpu->arch.cr0 & X86_CR0_WP;
184}
185
186static int is_cpuid_PSE36(void)
187{
188 return 1;
189}
190
191static int is_nx(struct kvm_vcpu *vcpu)
192{
193 return vcpu->arch.shadow_efer & EFER_NX;
194}
195
196static int is_present_pte(unsigned long pte)
197{
198 return pte & PT_PRESENT_MASK;
199}
200
201static 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
208static int is_writeble_pte(unsigned long pte)
209{
210 return pte & PT_WRITABLE_MASK;
211}
212
213static int is_dirty_pte(unsigned long pte)
214{
215 return pte & PT_DIRTY_MASK;
216}
217
218static int is_io_pte(unsigned long pte)
219{
220 return pte & PT_SHADOW_IO_MARK;
221}
222
223static int is_rmap_pte(u64 pte)
224{
225 return pte != shadow_trap_nonpresent_pte
226 && pte != shadow_notrap_nonpresent_pte;
227}
228
229static 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
236static 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
245static 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
261static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
262{
263 while (mc->nobjs)
264 kfree(mc->objects[--mc->nobjs]);
265}
266
267static 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
284static 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
290static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
291{
292 int r;
293
294 r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
295 pte_chain_cache, 4);
296 if (r)
297 goto out;
298 r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
299 rmap_desc_cache, 1);
300 if (r)
301 goto out;
302 r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
303 if (r)
304 goto out;
305 r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
306 mmu_page_header_cache, 4);
307out:
308 return r;
309}
310
311static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
312{
313 mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
314 mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
315 mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
316 mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
317}
318
319static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
320 size_t size)
321{
322 void *p;
323
324 BUG_ON(!mc->nobjs);
325 p = mc->objects[--mc->nobjs];
326 memset(p, 0, size);
327 return p;
328}
329
330static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
331{
332 return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
333 sizeof(struct kvm_pte_chain));
334}
335
336static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
337{
338 kfree(pc);
339}
340
341static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
342{
343 return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
344 sizeof(struct kvm_rmap_desc));
345}
346
347static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
348{
349 kfree(rd);
350}
351
352/*
353 * Take gfn and return the reverse mapping to it.
354 * Note: gfn must be unaliased before this function get called
355 */
356
357static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn)
358{
359 struct kvm_memory_slot *slot;
360
361 slot = gfn_to_memslot(kvm, gfn);
362 return &slot->rmap[gfn - slot->base_gfn];
363}
364
365/*
366 * Reverse mapping data structures:
367 *
368 * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
369 * that points to page_address(page).
370 *
371 * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
372 * containing more mappings.
373 */
374static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
375{
376 struct kvm_mmu_page *sp;
377 struct kvm_rmap_desc *desc;
378 unsigned long *rmapp;
379 int i;
380
381 if (!is_rmap_pte(*spte))
382 return;
383 gfn = unalias_gfn(vcpu->kvm, gfn);
384 sp = page_header(__pa(spte));
385 sp->gfns[spte - sp->spt] = gfn;
386 rmapp = gfn_to_rmap(vcpu->kvm, gfn);
387 if (!*rmapp) {
388 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
389 *rmapp = (unsigned long)spte;
390 } else if (!(*rmapp & 1)) {
391 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
392 desc = mmu_alloc_rmap_desc(vcpu);
393 desc->shadow_ptes[0] = (u64 *)*rmapp;
394 desc->shadow_ptes[1] = spte;
395 *rmapp = (unsigned long)desc | 1;
396 } else {
397 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
398 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
399 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
400 desc = desc->more;
401 if (desc->shadow_ptes[RMAP_EXT-1]) {
402 desc->more = mmu_alloc_rmap_desc(vcpu);
403 desc = desc->more;
404 }
405 for (i = 0; desc->shadow_ptes[i]; ++i)
406 ;
407 desc->shadow_ptes[i] = spte;
408 }
409}
410
411static void rmap_desc_remove_entry(unsigned long *rmapp,
412 struct kvm_rmap_desc *desc,
413 int i,
414 struct kvm_rmap_desc *prev_desc)
415{
416 int j;
417
418 for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
419 ;
420 desc->shadow_ptes[i] = desc->shadow_ptes[j];
421 desc->shadow_ptes[j] = NULL;
422 if (j != 0)
423 return;
424 if (!prev_desc && !desc->more)
425 *rmapp = (unsigned long)desc->shadow_ptes[0];
426 else
427 if (prev_desc)
428 prev_desc->more = desc->more;
429 else
430 *rmapp = (unsigned long)desc->more | 1;
431 mmu_free_rmap_desc(desc);
432}
433
434static void rmap_remove(struct kvm *kvm, u64 *spte)
435{
436 struct kvm_rmap_desc *desc;
437 struct kvm_rmap_desc *prev_desc;
438 struct kvm_mmu_page *sp;
439 struct page *page;
440 unsigned long *rmapp;
441 int i;
442
443 if (!is_rmap_pte(*spte))
444 return;
445 sp = page_header(__pa(spte));
446 page = pfn_to_page((*spte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT);
447 mark_page_accessed(page);
448 if (is_writeble_pte(*spte))
449 kvm_release_page_dirty(page);
450 else
451 kvm_release_page_clean(page);
452 rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt]);
453 if (!*rmapp) {
454 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
455 BUG();
456 } else if (!(*rmapp & 1)) {
457 rmap_printk("rmap_remove: %p %llx 1->0\n", spte, *spte);
458 if ((u64 *)*rmapp != spte) {
459 printk(KERN_ERR "rmap_remove: %p %llx 1->BUG\n",
460 spte, *spte);
461 BUG();
462 }
463 *rmapp = 0;
464 } else {
465 rmap_printk("rmap_remove: %p %llx many->many\n", spte, *spte);
466 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
467 prev_desc = NULL;
468 while (desc) {
469 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
470 if (desc->shadow_ptes[i] == spte) {
471 rmap_desc_remove_entry(rmapp,
472 desc, i,
473 prev_desc);
474 return;
475 }
476 prev_desc = desc;
477 desc = desc->more;
478 }
479 BUG();
480 }
481}
482
483static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
484{
485 struct kvm_rmap_desc *desc;
486 struct kvm_rmap_desc *prev_desc;
487 u64 *prev_spte;
488 int i;
489
490 if (!*rmapp)
491 return NULL;
492 else if (!(*rmapp & 1)) {
493 if (!spte)
494 return (u64 *)*rmapp;
495 return NULL;
496 }
497 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
498 prev_desc = NULL;
499 prev_spte = NULL;
500 while (desc) {
501 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
502 if (prev_spte == spte)
503 return desc->shadow_ptes[i];
504 prev_spte = desc->shadow_ptes[i];
505 }
506 desc = desc->more;
507 }
508 return NULL;
509}
510
511static void rmap_write_protect(struct kvm *kvm, u64 gfn)
512{
513 unsigned long *rmapp;
514 u64 *spte;
515 int write_protected = 0;
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 write_protected = 1;
528 }
529 spte = rmap_next(kvm, rmapp, spte);
530 }
531 if (write_protected)
532 kvm_flush_remote_tlbs(kvm);
533}
534
535#ifdef MMU_DEBUG
536static int is_empty_shadow_page(u64 *spt)
537{
538 u64 *pos;
539 u64 *end;
540
541 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
542 if ((*pos & ~PT_SHADOW_IO_MARK) != shadow_trap_nonpresent_pte) {
543 printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
544 pos, *pos);
545 return 0;
546 }
547 return 1;
548}
549#endif
550
551static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
552{
553 ASSERT(is_empty_shadow_page(sp->spt));
554 list_del(&sp->link);
555 __free_page(virt_to_page(sp->spt));
556 __free_page(virt_to_page(sp->gfns));
557 kfree(sp);
558 ++kvm->arch.n_free_mmu_pages;
559}
560
561static unsigned kvm_page_table_hashfn(gfn_t gfn)
562{
563 return gfn;
564}
565
566static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
567 u64 *parent_pte)
568{
569 struct kvm_mmu_page *sp;
570
571 sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
572 sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
573 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
574 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
575 list_add(&sp->link, &vcpu->kvm->arch.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->arch.n_free_mmu_pages;
581 return sp;
582}
583
584static 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
621static 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
659static 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->arch.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
678static 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->arch.mmu.root_level;
696 role.level = level;
697 role.metaphysical = metaphysical;
698 role.access = access;
699 if (vcpu->arch.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->arch.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 ++vcpu->kvm->stat.mmu_cache_miss;
715 sp = kvm_mmu_alloc_page(vcpu, parent_pte);
716 if (!sp)
717 return sp;
718 pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
719 sp->gfn = gfn;
720 sp->role = role;
721 hlist_add_head(&sp->hash_link, bucket);
722 vcpu->arch.mmu.prefetch_page(vcpu, sp);
723 if (!metaphysical)
724 rmap_write_protect(vcpu->kvm, gfn);
725 if (new_page)
726 *new_page = 1;
727 return sp;
728}
729
730static void kvm_mmu_page_unlink_children(struct kvm *kvm,
731 struct kvm_mmu_page *sp)
732{
733 unsigned i;
734 u64 *pt;
735 u64 ent;
736
737 pt = sp->spt;
738
739 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
740 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
741 if (is_shadow_present_pte(pt[i]))
742 rmap_remove(kvm, &pt[i]);
743 pt[i] = shadow_trap_nonpresent_pte;
744 }
745 kvm_flush_remote_tlbs(kvm);
746 return;
747 }
748
749 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
750 ent = pt[i];
751
752 pt[i] = shadow_trap_nonpresent_pte;
753 if (!is_shadow_present_pte(ent))
754 continue;
755 ent &= PT64_BASE_ADDR_MASK;
756 mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
757 }
758 kvm_flush_remote_tlbs(kvm);
759}
760
761static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
762{
763 mmu_page_remove_parent_pte(sp, parent_pte);
764}
765
766static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
767{
768 int i;
769
770 for (i = 0; i < KVM_MAX_VCPUS; ++i)
771 if (kvm->vcpus[i])
772 kvm->vcpus[i]->arch.last_pte_updated = NULL;
773}
774
775static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
776{
777 u64 *parent_pte;
778
779 ++kvm->stat.mmu_shadow_zapped;
780 while (sp->multimapped || sp->parent_pte) {
781 if (!sp->multimapped)
782 parent_pte = sp->parent_pte;
783 else {
784 struct kvm_pte_chain *chain;
785
786 chain = container_of(sp->parent_ptes.first,
787 struct kvm_pte_chain, link);
788 parent_pte = chain->parent_ptes[0];
789 }
790 BUG_ON(!parent_pte);
791 kvm_mmu_put_page(sp, parent_pte);
792 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
793 }
794 kvm_mmu_page_unlink_children(kvm, sp);
795 if (!sp->root_count) {
796 hlist_del(&sp->hash_link);
797 kvm_mmu_free_page(kvm, sp);
798 } else
799 list_move(&sp->link, &kvm->arch.active_mmu_pages);
800 kvm_mmu_reset_last_pte_updated(kvm);
801}
802
803/*
804 * Changing the number of mmu pages allocated to the vm
805 * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
806 */
807void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
808{
809 /*
810 * If we set the number of mmu pages to be smaller be than the
811 * number of actived pages , we must to free some mmu pages before we
812 * change the value
813 */
814
815 if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
816 kvm_nr_mmu_pages) {
817 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
818 - kvm->arch.n_free_mmu_pages;
819
820 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
821 struct kvm_mmu_page *page;
822
823 page = container_of(kvm->arch.active_mmu_pages.prev,
824 struct kvm_mmu_page, link);
825 kvm_mmu_zap_page(kvm, page);
826 n_used_mmu_pages--;
827 }
828 kvm->arch.n_free_mmu_pages = 0;
829 }
830 else
831 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
832 - kvm->arch.n_alloc_mmu_pages;
833
834 kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
835}
836
837static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
838{
839 unsigned index;
840 struct hlist_head *bucket;
841 struct kvm_mmu_page *sp;
842 struct hlist_node *node, *n;
843 int r;
844
845 pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
846 r = 0;
847 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
848 bucket = &kvm->arch.mmu_page_hash[index];
849 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
850 if (sp->gfn == gfn && !sp->role.metaphysical) {
851 pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
852 sp->role.word);
853 kvm_mmu_zap_page(kvm, sp);
854 r = 1;
855 }
856 return r;
857}
858
859static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
860{
861 struct kvm_mmu_page *sp;
862
863 while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
864 pgprintk("%s: zap %lx %x\n", __FUNCTION__, gfn, sp->role.word);
865 kvm_mmu_zap_page(kvm, sp);
866 }
867}
868
869static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
870{
871 int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
872 struct kvm_mmu_page *sp = page_header(__pa(pte));
873
874 __set_bit(slot, &sp->slot_bitmap);
875}
876
877struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
878{
879 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
880
881 if (gpa == UNMAPPED_GVA)
882 return NULL;
883 return gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
884}
885
886static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
887 unsigned pt_access, unsigned pte_access,
888 int user_fault, int write_fault, int dirty,
889 int *ptwrite, gfn_t gfn, struct page *page)
890{
891 u64 spte;
892 int was_rmapped = is_rmap_pte(*shadow_pte);
893 int was_writeble = is_writeble_pte(*shadow_pte);
894
895 pgprintk("%s: spte %llx access %x write_fault %d"
896 " user_fault %d gfn %lx\n",
897 __FUNCTION__, *shadow_pte, pt_access,
898 write_fault, user_fault, gfn);
899
900 /*
901 * We don't set the accessed bit, since we sometimes want to see
902 * whether the guest actually used the pte (in order to detect
903 * demand paging).
904 */
905 spte = PT_PRESENT_MASK | PT_DIRTY_MASK;
906 if (!dirty)
907 pte_access &= ~ACC_WRITE_MASK;
908 if (!(pte_access & ACC_EXEC_MASK))
909 spte |= PT64_NX_MASK;
910
911 spte |= PT_PRESENT_MASK;
912 if (pte_access & ACC_USER_MASK)
913 spte |= PT_USER_MASK;
914
915 if (is_error_page(page)) {
916 set_shadow_pte(shadow_pte,
917 shadow_trap_nonpresent_pte | PT_SHADOW_IO_MARK);
918 kvm_release_page_clean(page);
919 return;
920 }
921
922 spte |= page_to_phys(page);
923
924 if ((pte_access & ACC_WRITE_MASK)
925 || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
926 struct kvm_mmu_page *shadow;
927
928 spte |= PT_WRITABLE_MASK;
929 if (user_fault) {
930 mmu_unshadow(vcpu->kvm, gfn);
931 goto unshadowed;
932 }
933
934 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
935 if (shadow) {
936 pgprintk("%s: found shadow page for %lx, marking ro\n",
937 __FUNCTION__, gfn);
938 pte_access &= ~ACC_WRITE_MASK;
939 if (is_writeble_pte(spte)) {
940 spte &= ~PT_WRITABLE_MASK;
941 kvm_x86_ops->tlb_flush(vcpu);
942 }
943 if (write_fault)
944 *ptwrite = 1;
945 }
946 }
947
948unshadowed:
949
950 if (pte_access & ACC_WRITE_MASK)
951 mark_page_dirty(vcpu->kvm, gfn);
952
953 pgprintk("%s: setting spte %llx\n", __FUNCTION__, spte);
954 set_shadow_pte(shadow_pte, spte);
955 page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
956 if (!was_rmapped) {
957 rmap_add(vcpu, shadow_pte, gfn);
958 if (!is_rmap_pte(*shadow_pte))
959 kvm_release_page_clean(page);
960 } else {
961 if (was_writeble)
962 kvm_release_page_dirty(page);
963 else
964 kvm_release_page_clean(page);
965 }
966 if (!ptwrite || !*ptwrite)
967 vcpu->arch.last_pte_updated = shadow_pte;
968}
969
970static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
971{
972}
973
974static int __nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write,
975 gfn_t gfn, struct page *page)
976{
977 int level = PT32E_ROOT_LEVEL;
978 hpa_t table_addr = vcpu->arch.mmu.root_hpa;
979 int pt_write = 0;
980
981 for (; ; level--) {
982 u32 index = PT64_INDEX(v, level);
983 u64 *table;
984
985 ASSERT(VALID_PAGE(table_addr));
986 table = __va(table_addr);
987
988 if (level == 1) {
989 mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
990 0, write, 1, &pt_write, gfn, page);
991 return pt_write || is_io_pte(table[index]);
992 }
993
994 if (table[index] == shadow_trap_nonpresent_pte) {
995 struct kvm_mmu_page *new_table;
996 gfn_t pseudo_gfn;
997
998 pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
999 >> PAGE_SHIFT;
1000 new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
1001 v, level - 1,
1002 1, ACC_ALL, &table[index],
1003 NULL);
1004 if (!new_table) {
1005 pgprintk("nonpaging_map: ENOMEM\n");
1006 kvm_release_page_clean(page);
1007 return -ENOMEM;
1008 }
1009
1010 table[index] = __pa(new_table->spt) | PT_PRESENT_MASK
1011 | PT_WRITABLE_MASK | PT_USER_MASK;
1012 }
1013 table_addr = table[index] & PT64_BASE_ADDR_MASK;
1014 }
1015}
1016
1017static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1018{
1019 int r;
1020
1021 struct page *page;
1022
1023 down_read(&current->mm->mmap_sem);
1024 page = gfn_to_page(vcpu->kvm, gfn);
1025
1026 spin_lock(&vcpu->kvm->mmu_lock);
1027 kvm_mmu_free_some_pages(vcpu);
1028 r = __nonpaging_map(vcpu, v, write, gfn, page);
1029 spin_unlock(&vcpu->kvm->mmu_lock);
1030
1031 up_read(&current->mm->mmap_sem);
1032
1033 return r;
1034}
1035
1036
1037static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1038 struct kvm_mmu_page *sp)
1039{
1040 int i;
1041
1042 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1043 sp->spt[i] = shadow_trap_nonpresent_pte;
1044}
1045
1046static void mmu_free_roots(struct kvm_vcpu *vcpu)
1047{
1048 int i;
1049 struct kvm_mmu_page *sp;
1050
1051 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1052 return;
1053 spin_lock(&vcpu->kvm->mmu_lock);
1054#ifdef CONFIG_X86_64
1055 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1056 hpa_t root = vcpu->arch.mmu.root_hpa;
1057
1058 sp = page_header(root);
1059 --sp->root_count;
1060 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1061 spin_unlock(&vcpu->kvm->mmu_lock);
1062 return;
1063 }
1064#endif
1065 for (i = 0; i < 4; ++i) {
1066 hpa_t root = vcpu->arch.mmu.pae_root[i];
1067
1068 if (root) {
1069 root &= PT64_BASE_ADDR_MASK;
1070 sp = page_header(root);
1071 --sp->root_count;
1072 }
1073 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1074 }
1075 spin_unlock(&vcpu->kvm->mmu_lock);
1076 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1077}
1078
1079static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1080{
1081 int i;
1082 gfn_t root_gfn;
1083 struct kvm_mmu_page *sp;
1084
1085 root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1086
1087#ifdef CONFIG_X86_64
1088 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1089 hpa_t root = vcpu->arch.mmu.root_hpa;
1090
1091 ASSERT(!VALID_PAGE(root));
1092 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1093 PT64_ROOT_LEVEL, 0, ACC_ALL, NULL, NULL);
1094 root = __pa(sp->spt);
1095 ++sp->root_count;
1096 vcpu->arch.mmu.root_hpa = root;
1097 return;
1098 }
1099#endif
1100 for (i = 0; i < 4; ++i) {
1101 hpa_t root = vcpu->arch.mmu.pae_root[i];
1102
1103 ASSERT(!VALID_PAGE(root));
1104 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1105 if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1106 vcpu->arch.mmu.pae_root[i] = 0;
1107 continue;
1108 }
1109 root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1110 } else if (vcpu->arch.mmu.root_level == 0)
1111 root_gfn = 0;
1112 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1113 PT32_ROOT_LEVEL, !is_paging(vcpu),
1114 ACC_ALL, NULL, NULL);
1115 root = __pa(sp->spt);
1116 ++sp->root_count;
1117 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1118 }
1119 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1120}
1121
1122static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1123{
1124 return vaddr;
1125}
1126
1127static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1128 u32 error_code)
1129{
1130 gfn_t gfn;
1131 int r;
1132
1133 pgprintk("%s: gva %lx error %x\n", __FUNCTION__, gva, error_code);
1134 r = mmu_topup_memory_caches(vcpu);
1135 if (r)
1136 return r;
1137
1138 ASSERT(vcpu);
1139 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1140
1141 gfn = gva >> PAGE_SHIFT;
1142
1143 return nonpaging_map(vcpu, gva & PAGE_MASK,
1144 error_code & PFERR_WRITE_MASK, gfn);
1145}
1146
1147static void nonpaging_free(struct kvm_vcpu *vcpu)
1148{
1149 mmu_free_roots(vcpu);
1150}
1151
1152static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1153{
1154 struct kvm_mmu *context = &vcpu->arch.mmu;
1155
1156 context->new_cr3 = nonpaging_new_cr3;
1157 context->page_fault = nonpaging_page_fault;
1158 context->gva_to_gpa = nonpaging_gva_to_gpa;
1159 context->free = nonpaging_free;
1160 context->prefetch_page = nonpaging_prefetch_page;
1161 context->root_level = 0;
1162 context->shadow_root_level = PT32E_ROOT_LEVEL;
1163 context->root_hpa = INVALID_PAGE;
1164 return 0;
1165}
1166
1167void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1168{
1169 ++vcpu->stat.tlb_flush;
1170 kvm_x86_ops->tlb_flush(vcpu);
1171}
1172
1173static void paging_new_cr3(struct kvm_vcpu *vcpu)
1174{
1175 pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
1176 mmu_free_roots(vcpu);
1177}
1178
1179static void inject_page_fault(struct kvm_vcpu *vcpu,
1180 u64 addr,
1181 u32 err_code)
1182{
1183 kvm_inject_page_fault(vcpu, addr, err_code);
1184}
1185
1186static void paging_free(struct kvm_vcpu *vcpu)
1187{
1188 nonpaging_free(vcpu);
1189}
1190
1191#define PTTYPE 64
1192#include "paging_tmpl.h"
1193#undef PTTYPE
1194
1195#define PTTYPE 32
1196#include "paging_tmpl.h"
1197#undef PTTYPE
1198
1199static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1200{
1201 struct kvm_mmu *context = &vcpu->arch.mmu;
1202
1203 ASSERT(is_pae(vcpu));
1204 context->new_cr3 = paging_new_cr3;
1205 context->page_fault = paging64_page_fault;
1206 context->gva_to_gpa = paging64_gva_to_gpa;
1207 context->prefetch_page = paging64_prefetch_page;
1208 context->free = paging_free;
1209 context->root_level = level;
1210 context->shadow_root_level = level;
1211 context->root_hpa = INVALID_PAGE;
1212 return 0;
1213}
1214
1215static int paging64_init_context(struct kvm_vcpu *vcpu)
1216{
1217 return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1218}
1219
1220static int paging32_init_context(struct kvm_vcpu *vcpu)
1221{
1222 struct kvm_mmu *context = &vcpu->arch.mmu;
1223
1224 context->new_cr3 = paging_new_cr3;
1225 context->page_fault = paging32_page_fault;
1226 context->gva_to_gpa = paging32_gva_to_gpa;
1227 context->free = paging_free;
1228 context->prefetch_page = paging32_prefetch_page;
1229 context->root_level = PT32_ROOT_LEVEL;
1230 context->shadow_root_level = PT32E_ROOT_LEVEL;
1231 context->root_hpa = INVALID_PAGE;
1232 return 0;
1233}
1234
1235static int paging32E_init_context(struct kvm_vcpu *vcpu)
1236{
1237 return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1238}
1239
1240static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1241{
1242 ASSERT(vcpu);
1243 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1244
1245 if (!is_paging(vcpu))
1246 return nonpaging_init_context(vcpu);
1247 else if (is_long_mode(vcpu))
1248 return paging64_init_context(vcpu);
1249 else if (is_pae(vcpu))
1250 return paging32E_init_context(vcpu);
1251 else
1252 return paging32_init_context(vcpu);
1253}
1254
1255static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1256{
1257 ASSERT(vcpu);
1258 if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1259 vcpu->arch.mmu.free(vcpu);
1260 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1261 }
1262}
1263
1264int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1265{
1266 destroy_kvm_mmu(vcpu);
1267 return init_kvm_mmu(vcpu);
1268}
1269EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1270
1271int kvm_mmu_load(struct kvm_vcpu *vcpu)
1272{
1273 int r;
1274
1275 r = mmu_topup_memory_caches(vcpu);
1276 if (r)
1277 goto out;
1278 spin_lock(&vcpu->kvm->mmu_lock);
1279 kvm_mmu_free_some_pages(vcpu);
1280 mmu_alloc_roots(vcpu);
1281 spin_unlock(&vcpu->kvm->mmu_lock);
1282 kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1283 kvm_mmu_flush_tlb(vcpu);
1284out:
1285 return r;
1286}
1287EXPORT_SYMBOL_GPL(kvm_mmu_load);
1288
1289void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1290{
1291 mmu_free_roots(vcpu);
1292}
1293
1294static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1295 struct kvm_mmu_page *sp,
1296 u64 *spte)
1297{
1298 u64 pte;
1299 struct kvm_mmu_page *child;
1300
1301 pte = *spte;
1302 if (is_shadow_present_pte(pte)) {
1303 if (sp->role.level == PT_PAGE_TABLE_LEVEL)
1304 rmap_remove(vcpu->kvm, spte);
1305 else {
1306 child = page_header(pte & PT64_BASE_ADDR_MASK);
1307 mmu_page_remove_parent_pte(child, spte);
1308 }
1309 }
1310 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1311}
1312
1313static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1314 struct kvm_mmu_page *sp,
1315 u64 *spte,
1316 const void *new, int bytes,
1317 int offset_in_pte)
1318{
1319 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
1320 ++vcpu->kvm->stat.mmu_pde_zapped;
1321 return;
1322 }
1323
1324 ++vcpu->kvm->stat.mmu_pte_updated;
1325 if (sp->role.glevels == PT32_ROOT_LEVEL)
1326 paging32_update_pte(vcpu, sp, spte, new, bytes, offset_in_pte);
1327 else
1328 paging64_update_pte(vcpu, sp, spte, new, bytes, offset_in_pte);
1329}
1330
1331static bool need_remote_flush(u64 old, u64 new)
1332{
1333 if (!is_shadow_present_pte(old))
1334 return false;
1335 if (!is_shadow_present_pte(new))
1336 return true;
1337 if ((old ^ new) & PT64_BASE_ADDR_MASK)
1338 return true;
1339 old ^= PT64_NX_MASK;
1340 new ^= PT64_NX_MASK;
1341 return (old & ~new & PT64_PERM_MASK) != 0;
1342}
1343
1344static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1345{
1346 if (need_remote_flush(old, new))
1347 kvm_flush_remote_tlbs(vcpu->kvm);
1348 else
1349 kvm_mmu_flush_tlb(vcpu);
1350}
1351
1352static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1353{
1354 u64 *spte = vcpu->arch.last_pte_updated;
1355
1356 return !!(spte && (*spte & PT_ACCESSED_MASK));
1357}
1358
1359static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1360 const u8 *new, int bytes)
1361{
1362 gfn_t gfn;
1363 int r;
1364 u64 gpte = 0;
1365
1366 if (bytes != 4 && bytes != 8)
1367 return;
1368
1369 /*
1370 * Assume that the pte write on a page table of the same type
1371 * as the current vcpu paging mode. This is nearly always true
1372 * (might be false while changing modes). Note it is verified later
1373 * by update_pte().
1374 */
1375 if (is_pae(vcpu)) {
1376 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1377 if ((bytes == 4) && (gpa % 4 == 0)) {
1378 r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
1379 if (r)
1380 return;
1381 memcpy((void *)&gpte + (gpa % 8), new, 4);
1382 } else if ((bytes == 8) && (gpa % 8 == 0)) {
1383 memcpy((void *)&gpte, new, 8);
1384 }
1385 } else {
1386 if ((bytes == 4) && (gpa % 4 == 0))
1387 memcpy((void *)&gpte, new, 4);
1388 }
1389 if (!is_present_pte(gpte))
1390 return;
1391 gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
1392 vcpu->arch.update_pte.gfn = gfn;
1393 vcpu->arch.update_pte.page = gfn_to_page(vcpu->kvm, gfn);
1394}
1395
1396void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1397 const u8 *new, int bytes)
1398{
1399 gfn_t gfn = gpa >> PAGE_SHIFT;
1400 struct kvm_mmu_page *sp;
1401 struct hlist_node *node, *n;
1402 struct hlist_head *bucket;
1403 unsigned index;
1404 u64 entry;
1405 u64 *spte;
1406 unsigned offset = offset_in_page(gpa);
1407 unsigned pte_size;
1408 unsigned page_offset;
1409 unsigned misaligned;
1410 unsigned quadrant;
1411 int level;
1412 int flooded = 0;
1413 int npte;
1414
1415 pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1416 mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
1417 spin_lock(&vcpu->kvm->mmu_lock);
1418 kvm_mmu_free_some_pages(vcpu);
1419 ++vcpu->kvm->stat.mmu_pte_write;
1420 kvm_mmu_audit(vcpu, "pre pte write");
1421 if (gfn == vcpu->arch.last_pt_write_gfn
1422 && !last_updated_pte_accessed(vcpu)) {
1423 ++vcpu->arch.last_pt_write_count;
1424 if (vcpu->arch.last_pt_write_count >= 3)
1425 flooded = 1;
1426 } else {
1427 vcpu->arch.last_pt_write_gfn = gfn;
1428 vcpu->arch.last_pt_write_count = 1;
1429 vcpu->arch.last_pte_updated = NULL;
1430 }
1431 index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1432 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1433 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1434 if (sp->gfn != gfn || sp->role.metaphysical)
1435 continue;
1436 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1437 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1438 misaligned |= bytes < 4;
1439 if (misaligned || flooded) {
1440 /*
1441 * Misaligned accesses are too much trouble to fix
1442 * up; also, they usually indicate a page is not used
1443 * as a page table.
1444 *
1445 * If we're seeing too many writes to a page,
1446 * it may no longer be a page table, or we may be
1447 * forking, in which case it is better to unmap the
1448 * page.
1449 */
1450 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1451 gpa, bytes, sp->role.word);
1452 kvm_mmu_zap_page(vcpu->kvm, sp);
1453 ++vcpu->kvm->stat.mmu_flooded;
1454 continue;
1455 }
1456 page_offset = offset;
1457 level = sp->role.level;
1458 npte = 1;
1459 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1460 page_offset <<= 1; /* 32->64 */
1461 /*
1462 * A 32-bit pde maps 4MB while the shadow pdes map
1463 * only 2MB. So we need to double the offset again
1464 * and zap two pdes instead of one.
1465 */
1466 if (level == PT32_ROOT_LEVEL) {
1467 page_offset &= ~7; /* kill rounding error */
1468 page_offset <<= 1;
1469 npte = 2;
1470 }
1471 quadrant = page_offset >> PAGE_SHIFT;
1472 page_offset &= ~PAGE_MASK;
1473 if (quadrant != sp->role.quadrant)
1474 continue;
1475 }
1476 spte = &sp->spt[page_offset / sizeof(*spte)];
1477 while (npte--) {
1478 entry = *spte;
1479 mmu_pte_write_zap_pte(vcpu, sp, spte);
1480 mmu_pte_write_new_pte(vcpu, sp, spte, new, bytes,
1481 page_offset & (pte_size - 1));
1482 mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1483 ++spte;
1484 }
1485 }
1486 kvm_mmu_audit(vcpu, "post pte write");
1487 spin_unlock(&vcpu->kvm->mmu_lock);
1488 if (vcpu->arch.update_pte.page) {
1489 kvm_release_page_clean(vcpu->arch.update_pte.page);
1490 vcpu->arch.update_pte.page = NULL;
1491 }
1492}
1493
1494int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1495{
1496 gpa_t gpa;
1497 int r;
1498
1499 down_read(&current->mm->mmap_sem);
1500 gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1501 up_read(&current->mm->mmap_sem);
1502
1503 spin_lock(&vcpu->kvm->mmu_lock);
1504 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1505 spin_unlock(&vcpu->kvm->mmu_lock);
1506 return r;
1507}
1508
1509void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1510{
1511 while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
1512 struct kvm_mmu_page *sp;
1513
1514 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
1515 struct kvm_mmu_page, link);
1516 kvm_mmu_zap_page(vcpu->kvm, sp);
1517 ++vcpu->kvm->stat.mmu_recycled;
1518 }
1519}
1520
1521int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1522{
1523 int r;
1524 enum emulation_result er;
1525
1526 r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
1527 if (r < 0)
1528 goto out;
1529
1530 if (!r) {
1531 r = 1;
1532 goto out;
1533 }
1534
1535 r = mmu_topup_memory_caches(vcpu);
1536 if (r)
1537 goto out;
1538
1539 er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1540
1541 switch (er) {
1542 case EMULATE_DONE:
1543 return 1;
1544 case EMULATE_DO_MMIO:
1545 ++vcpu->stat.mmio_exits;
1546 return 0;
1547 case EMULATE_FAIL:
1548 kvm_report_emulation_failure(vcpu, "pagetable");
1549 return 1;
1550 default:
1551 BUG();
1552 }
1553out:
1554 return r;
1555}
1556EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1557
1558static void free_mmu_pages(struct kvm_vcpu *vcpu)
1559{
1560 struct kvm_mmu_page *sp;
1561
1562 while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
1563 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
1564 struct kvm_mmu_page, link);
1565 kvm_mmu_zap_page(vcpu->kvm, sp);
1566 }
1567 free_page((unsigned long)vcpu->arch.mmu.pae_root);
1568}
1569
1570static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1571{
1572 struct page *page;
1573 int i;
1574
1575 ASSERT(vcpu);
1576
1577 if (vcpu->kvm->arch.n_requested_mmu_pages)
1578 vcpu->kvm->arch.n_free_mmu_pages =
1579 vcpu->kvm->arch.n_requested_mmu_pages;
1580 else
1581 vcpu->kvm->arch.n_free_mmu_pages =
1582 vcpu->kvm->arch.n_alloc_mmu_pages;
1583 /*
1584 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1585 * Therefore we need to allocate shadow page tables in the first
1586 * 4GB of memory, which happens to fit the DMA32 zone.
1587 */
1588 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1589 if (!page)
1590 goto error_1;
1591 vcpu->arch.mmu.pae_root = page_address(page);
1592 for (i = 0; i < 4; ++i)
1593 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1594
1595 return 0;
1596
1597error_1:
1598 free_mmu_pages(vcpu);
1599 return -ENOMEM;
1600}
1601
1602int kvm_mmu_create(struct kvm_vcpu *vcpu)
1603{
1604 ASSERT(vcpu);
1605 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1606
1607 return alloc_mmu_pages(vcpu);
1608}
1609
1610int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1611{
1612 ASSERT(vcpu);
1613 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1614
1615 return init_kvm_mmu(vcpu);
1616}
1617
1618void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1619{
1620 ASSERT(vcpu);
1621
1622 destroy_kvm_mmu(vcpu);
1623 free_mmu_pages(vcpu);
1624 mmu_free_memory_caches(vcpu);
1625}
1626
1627void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1628{
1629 struct kvm_mmu_page *sp;
1630
1631 list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
1632 int i;
1633 u64 *pt;
1634
1635 if (!test_bit(slot, &sp->slot_bitmap))
1636 continue;
1637
1638 pt = sp->spt;
1639 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1640 /* avoid RMW */
1641 if (pt[i] & PT_WRITABLE_MASK)
1642 pt[i] &= ~PT_WRITABLE_MASK;
1643 }
1644}
1645
1646void kvm_mmu_zap_all(struct kvm *kvm)
1647{
1648 struct kvm_mmu_page *sp, *node;
1649
1650 spin_lock(&kvm->mmu_lock);
1651 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
1652 kvm_mmu_zap_page(kvm, sp);
1653 spin_unlock(&kvm->mmu_lock);
1654
1655 kvm_flush_remote_tlbs(kvm);
1656}
1657
1658void kvm_mmu_module_exit(void)
1659{
1660 if (pte_chain_cache)
1661 kmem_cache_destroy(pte_chain_cache);
1662 if (rmap_desc_cache)
1663 kmem_cache_destroy(rmap_desc_cache);
1664 if (mmu_page_header_cache)
1665 kmem_cache_destroy(mmu_page_header_cache);
1666}
1667
1668int kvm_mmu_module_init(void)
1669{
1670 pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1671 sizeof(struct kvm_pte_chain),
1672 0, 0, NULL);
1673 if (!pte_chain_cache)
1674 goto nomem;
1675 rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1676 sizeof(struct kvm_rmap_desc),
1677 0, 0, NULL);
1678 if (!rmap_desc_cache)
1679 goto nomem;
1680
1681 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1682 sizeof(struct kvm_mmu_page),
1683 0, 0, NULL);
1684 if (!mmu_page_header_cache)
1685 goto nomem;
1686
1687 return 0;
1688
1689nomem:
1690 kvm_mmu_module_exit();
1691 return -ENOMEM;
1692}
1693
1694/*
1695 * Caculate mmu pages needed for kvm.
1696 */
1697unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
1698{
1699 int i;
1700 unsigned int nr_mmu_pages;
1701 unsigned int nr_pages = 0;
1702
1703 for (i = 0; i < kvm->nmemslots; i++)
1704 nr_pages += kvm->memslots[i].npages;
1705
1706 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
1707 nr_mmu_pages = max(nr_mmu_pages,
1708 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
1709
1710 return nr_mmu_pages;
1711}
1712
1713#ifdef AUDIT
1714
1715static const char *audit_msg;
1716
1717static gva_t canonicalize(gva_t gva)
1718{
1719#ifdef CONFIG_X86_64
1720 gva = (long long)(gva << 16) >> 16;
1721#endif
1722 return gva;
1723}
1724
1725static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1726 gva_t va, int level)
1727{
1728 u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1729 int i;
1730 gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1731
1732 for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1733 u64 ent = pt[i];
1734
1735 if (ent == shadow_trap_nonpresent_pte)
1736 continue;
1737
1738 va = canonicalize(va);
1739 if (level > 1) {
1740 if (ent == shadow_notrap_nonpresent_pte)
1741 printk(KERN_ERR "audit: (%s) nontrapping pte"
1742 " in nonleaf level: levels %d gva %lx"
1743 " level %d pte %llx\n", audit_msg,
1744 vcpu->arch.mmu.root_level, va, level, ent);
1745
1746 audit_mappings_page(vcpu, ent, va, level - 1);
1747 } else {
1748 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
1749 struct page *page = gpa_to_page(vcpu, gpa);
1750 hpa_t hpa = page_to_phys(page);
1751
1752 if (is_shadow_present_pte(ent)
1753 && (ent & PT64_BASE_ADDR_MASK) != hpa)
1754 printk(KERN_ERR "xx audit error: (%s) levels %d"
1755 " gva %lx gpa %llx hpa %llx ent %llx %d\n",
1756 audit_msg, vcpu->arch.mmu.root_level,
1757 va, gpa, hpa, ent,
1758 is_shadow_present_pte(ent));
1759 else if (ent == shadow_notrap_nonpresent_pte
1760 && !is_error_hpa(hpa))
1761 printk(KERN_ERR "audit: (%s) notrap shadow,"
1762 " valid guest gva %lx\n", audit_msg, va);
1763 kvm_release_page_clean(page);
1764
1765 }
1766 }
1767}
1768
1769static void audit_mappings(struct kvm_vcpu *vcpu)
1770{
1771 unsigned i;
1772
1773 if (vcpu->arch.mmu.root_level == 4)
1774 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
1775 else
1776 for (i = 0; i < 4; ++i)
1777 if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
1778 audit_mappings_page(vcpu,
1779 vcpu->arch.mmu.pae_root[i],
1780 i << 30,
1781 2);
1782}
1783
1784static int count_rmaps(struct kvm_vcpu *vcpu)
1785{
1786 int nmaps = 0;
1787 int i, j, k;
1788
1789 for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1790 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1791 struct kvm_rmap_desc *d;
1792
1793 for (j = 0; j < m->npages; ++j) {
1794 unsigned long *rmapp = &m->rmap[j];
1795
1796 if (!*rmapp)
1797 continue;
1798 if (!(*rmapp & 1)) {
1799 ++nmaps;
1800 continue;
1801 }
1802 d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
1803 while (d) {
1804 for (k = 0; k < RMAP_EXT; ++k)
1805 if (d->shadow_ptes[k])
1806 ++nmaps;
1807 else
1808 break;
1809 d = d->more;
1810 }
1811 }
1812 }
1813 return nmaps;
1814}
1815
1816static int count_writable_mappings(struct kvm_vcpu *vcpu)
1817{
1818 int nmaps = 0;
1819 struct kvm_mmu_page *sp;
1820 int i;
1821
1822 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
1823 u64 *pt = sp->spt;
1824
1825 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
1826 continue;
1827
1828 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1829 u64 ent = pt[i];
1830
1831 if (!(ent & PT_PRESENT_MASK))
1832 continue;
1833 if (!(ent & PT_WRITABLE_MASK))
1834 continue;
1835 ++nmaps;
1836 }
1837 }
1838 return nmaps;
1839}
1840
1841static void audit_rmap(struct kvm_vcpu *vcpu)
1842{
1843 int n_rmap = count_rmaps(vcpu);
1844 int n_actual = count_writable_mappings(vcpu);
1845
1846 if (n_rmap != n_actual)
1847 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1848 __FUNCTION__, audit_msg, n_rmap, n_actual);
1849}
1850
1851static void audit_write_protection(struct kvm_vcpu *vcpu)
1852{
1853 struct kvm_mmu_page *sp;
1854 struct kvm_memory_slot *slot;
1855 unsigned long *rmapp;
1856 gfn_t gfn;
1857
1858 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
1859 if (sp->role.metaphysical)
1860 continue;
1861
1862 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
1863 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
1864 rmapp = &slot->rmap[gfn - slot->base_gfn];
1865 if (*rmapp)
1866 printk(KERN_ERR "%s: (%s) shadow page has writable"
1867 " mappings: gfn %lx role %x\n",
1868 __FUNCTION__, audit_msg, sp->gfn,
1869 sp->role.word);
1870 }
1871}
1872
1873static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1874{
1875 int olddbg = dbg;
1876
1877 dbg = 0;
1878 audit_msg = msg;
1879 audit_rmap(vcpu);
1880 audit_write_protection(vcpu);
1881 audit_mappings(vcpu);
1882 dbg = olddbg;
1883}
1884
1885#endif
generated by cgit v1.2.2 (git 2.25.0) at 2025-09-12 00:25:27 -0400