« back to all changes in this revision
Viewing changes to kernel/tmp/mmu.c
- browse files at revision 22
- compare with another revision
- download diff
- download tarball
- view history from revision 22
- Committer: Bazaar Package Importer
- Author(s): Soren Hansen
- Date: 2008-01-03 10:39:25 UTC
- mfrom: (1.1.16 upstream)
- Revision ID: james.westby@ubuntu.com-20080103103925-8480u7sq2646hvbh
Tags: 1:59+dfsg-0ubuntu1
* New upstream release
* Build with alsa support (cherry pick from 57+dfsg-2)
* Build with alsa support (cherry pick from 57+dfsg-2)
- files added:
- bios/vapic.S
- user/test/powerpc
- user/test/powerpc/44x
- vgabios
- vgabios/tests
- vgabios/tests/lfbprof
- files removed:
- kernel/tmp
- kernel/tmp/include
- kernel/tmp/include/asm-x86
- kernel/tmp/include/linux
- files modified:
- Makefile
added
removed
kernel/tmp/mmu.c
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
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->arch.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->arch.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->arch.mmu_pte_chain_cache,
296
pte_chain_cache, 4);
297
if (r)
298
goto out;
299
r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
300
rmap_desc_cache, 1);
301
if (r)
302
goto out;
303
r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
304
if (r)
305
goto out;
306
r = mmu_topup_memory_cache(&vcpu->arch.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->arch.mmu_pte_chain_cache);
315
mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
316
mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
317
mmu_free_memory_cache(&vcpu->arch.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->arch.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->arch.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
int write_protected = 0;
517
518
gfn = unalias_gfn(kvm, gfn);
519
rmapp = gfn_to_rmap(kvm, gfn);
520
521
spte = rmap_next(kvm, rmapp, NULL);
522
while (spte) {
523
BUG_ON(!spte);
524
BUG_ON(!(*spte & PT_PRESENT_MASK));
525
rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
526
if (is_writeble_pte(*spte)) {
527
set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
528
write_protected = 1;
529
}
530
spte = rmap_next(kvm, rmapp, spte);
531
}
532
if (write_protected)
533
kvm_flush_remote_tlbs(kvm);
534
}
535
536
#ifdef MMU_DEBUG
537
static int is_empty_shadow_page(u64 *spt)
538
{
539
u64 *pos;
540
u64 *end;
541
542
for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
543
if ((*pos & ~PT_SHADOW_IO_MARK) != shadow_trap_nonpresent_pte) {
544
printk(KERN_ERR "%s: %p %llx\n", __FUNCTION__,
545
pos, *pos);
546
return 0;
547
}
548
return 1;
549
}
550
#endif
551
552
static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
553
{
554
ASSERT(is_empty_shadow_page(sp->spt));
555
list_del(&sp->link);
556
__free_page(virt_to_page(sp->spt));
557
__free_page(virt_to_page(sp->gfns));
558
kfree(sp);
559
++kvm->arch.n_free_mmu_pages;
560
}
561
562
static unsigned kvm_page_table_hashfn(gfn_t gfn)
563
{
564
return gfn;
565
}
566
567
static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
568
u64 *parent_pte)
569
{
570
struct kvm_mmu_page *sp;
571
572
if (!vcpu->kvm->arch.n_free_mmu_pages)
573
return NULL;
574
575
sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
576
sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
577
sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
578
set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
579
list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
580
ASSERT(is_empty_shadow_page(sp->spt));
581
sp->slot_bitmap = 0;
582
sp->multimapped = 0;
583
sp->parent_pte = parent_pte;
584
--vcpu->kvm->arch.n_free_mmu_pages;
585
return sp;
586
}
587
588
static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
589
struct kvm_mmu_page *sp, u64 *parent_pte)
590
{
591
struct kvm_pte_chain *pte_chain;
592
struct hlist_node *node;
593
int i;
594
595
if (!parent_pte)
596
return;
597
if (!sp->multimapped) {
598
u64 *old = sp->parent_pte;
599
600
if (!old) {
601
sp->parent_pte = parent_pte;
602
return;
603
}
604
sp->multimapped = 1;
605
pte_chain = mmu_alloc_pte_chain(vcpu);
606
INIT_HLIST_HEAD(&sp->parent_ptes);
607
hlist_add_head(&pte_chain->link, &sp->parent_ptes);
608
pte_chain->parent_ptes[0] = old;
609
}
610
hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
611
if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
612
continue;
613
for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
614
if (!pte_chain->parent_ptes[i]) {
615
pte_chain->parent_ptes[i] = parent_pte;
616
return;
617
}
618
}
619
pte_chain = mmu_alloc_pte_chain(vcpu);
620
BUG_ON(!pte_chain);
621
hlist_add_head(&pte_chain->link, &sp->parent_ptes);
622
pte_chain->parent_ptes[0] = parent_pte;
623
}
624
625
static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
626
u64 *parent_pte)
627
{
628
struct kvm_pte_chain *pte_chain;
629
struct hlist_node *node;
630
int i;
631
632
if (!sp->multimapped) {
633
BUG_ON(sp->parent_pte != parent_pte);
634
sp->parent_pte = NULL;
635
return;
636
}
637
hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
638
for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
639
if (!pte_chain->parent_ptes[i])
640
break;
641
if (pte_chain->parent_ptes[i] != parent_pte)
642
continue;
643
while (i + 1 < NR_PTE_CHAIN_ENTRIES
644
&& pte_chain->parent_ptes[i + 1]) {
645
pte_chain->parent_ptes[i]
646
= pte_chain->parent_ptes[i + 1];
647
++i;
648
}
649
pte_chain->parent_ptes[i] = NULL;
650
if (i == 0) {
651
hlist_del(&pte_chain->link);
652
mmu_free_pte_chain(pte_chain);
653
if (hlist_empty(&sp->parent_ptes)) {
654
sp->multimapped = 0;
655
sp->parent_pte = NULL;
656
}
657
}
658
return;
659
}
660
BUG();
661
}
662
663
static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
664
{
665
unsigned index;
666
struct hlist_head *bucket;
667
struct kvm_mmu_page *sp;
668
struct hlist_node *node;
669
670
pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
671
index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
672
bucket = &kvm->arch.mmu_page_hash[index];
673
hlist_for_each_entry(sp, node, bucket, hash_link)
674
if (sp->gfn == gfn && !sp->role.metaphysical) {
675
pgprintk("%s: found role %x\n",
676
__FUNCTION__, sp->role.word);
677
return sp;
678
}
679
return NULL;
680
}
681
682
static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
683
gfn_t gfn,
684
gva_t gaddr,
685
unsigned level,
686
int metaphysical,
687
unsigned access,
688
u64 *parent_pte,
689
bool *new_page)
690
{
691
union kvm_mmu_page_role role;
692
unsigned index;
693
unsigned quadrant;
694
struct hlist_head *bucket;
695
struct kvm_mmu_page *sp;
696
struct hlist_node *node;
697
698
role.word = 0;
699
role.glevels = vcpu->arch.mmu.root_level;
700
role.level = level;
701
role.metaphysical = metaphysical;
702
role.access = access;
703
if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
704
quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
705
quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
706
role.quadrant = quadrant;
707
}
708
pgprintk("%s: looking gfn %lx role %x\n", __FUNCTION__,
709
gfn, role.word);
710
index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
711
bucket = &vcpu->kvm->arch.mmu_page_hash[index];
712
hlist_for_each_entry(sp, node, bucket, hash_link)
713
if (sp->gfn == gfn && sp->role.word == role.word) {
714
mmu_page_add_parent_pte(vcpu, sp, parent_pte);
715
pgprintk("%s: found\n", __FUNCTION__);
716
return sp;
717
}
718
sp = kvm_mmu_alloc_page(vcpu, parent_pte);
719
if (!sp)
720
return sp;
721
pgprintk("%s: adding gfn %lx role %x\n", __FUNCTION__, gfn, role.word);
722
sp->gfn = gfn;
723
sp->role = role;
724
hlist_add_head(&sp->hash_link, bucket);
725
vcpu->arch.mmu.prefetch_page(vcpu, sp);
726
if (!metaphysical)
727
rmap_write_protect(vcpu->kvm, gfn);
728
if (new_page)
729
*new_page = 1;
730
return sp;
731
}
732
733
static void kvm_mmu_page_unlink_children(struct kvm *kvm,
734
struct kvm_mmu_page *sp)
735
{
736
unsigned i;
737
u64 *pt;
738
u64 ent;
739
740
pt = sp->spt;
741
742
if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
743
for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
744
if (is_shadow_present_pte(pt[i]))
745
rmap_remove(kvm, &pt[i]);
746
pt[i] = shadow_trap_nonpresent_pte;
747
}
748
kvm_flush_remote_tlbs(kvm);
749
return;
750
}
751
752
for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
753
ent = pt[i];
754
755
pt[i] = shadow_trap_nonpresent_pte;
756
if (!is_shadow_present_pte(ent))
757
continue;
758
ent &= PT64_BASE_ADDR_MASK;
759
mmu_page_remove_parent_pte(page_header(ent), &pt[i]);
760
}
761
kvm_flush_remote_tlbs(kvm);
762
}
763
764
static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
765
{
766
mmu_page_remove_parent_pte(sp, parent_pte);
767
}
768
769
static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
770
{
771
int i;
772
773
for (i = 0; i < KVM_MAX_VCPUS; ++i)
774
if (kvm->vcpus[i])
775
kvm->vcpus[i]->arch.last_pte_updated = NULL;
776
}
777
778
static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
779
{
780
u64 *parent_pte;
781
782
++kvm->stat.mmu_shadow_zapped;
783
while (sp->multimapped || sp->parent_pte) {
784
if (!sp->multimapped)
785
parent_pte = sp->parent_pte;
786
else {
787
struct kvm_pte_chain *chain;
788
789
chain = container_of(sp->parent_ptes.first,
790
struct kvm_pte_chain, link);
791
parent_pte = chain->parent_ptes[0];
792
}
793
BUG_ON(!parent_pte);
794
kvm_mmu_put_page(sp, parent_pte);
795
set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
796
}
797
kvm_mmu_page_unlink_children(kvm, sp);
798
if (!sp->root_count) {
799
hlist_del(&sp->hash_link);
800
kvm_mmu_free_page(kvm, sp);
801
} else
802
list_move(&sp->link, &kvm->arch.active_mmu_pages);
803
kvm_mmu_reset_last_pte_updated(kvm);
804
}
805
806
/*
807
* Changing the number of mmu pages allocated to the vm
808
* Note: if kvm_nr_mmu_pages is too small, you will get dead lock
809
*/
810
void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
811
{
812
/*
813
* If we set the number of mmu pages to be smaller be than the
814
* number of actived pages , we must to free some mmu pages before we
815
* change the value
816
*/
817
818
if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
819
kvm_nr_mmu_pages) {
820
int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
821
- kvm->arch.n_free_mmu_pages;
822
823
while (n_used_mmu_pages > kvm_nr_mmu_pages) {
824
struct kvm_mmu_page *page;
825
826
page = container_of(kvm->arch.active_mmu_pages.prev,
827
struct kvm_mmu_page, link);
828
kvm_mmu_zap_page(kvm, page);
829
n_used_mmu_pages--;
830
}
831
kvm->arch.n_free_mmu_pages = 0;
832
}
833
else
834
kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
835
- kvm->arch.n_alloc_mmu_pages;
836
837
kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
838
}
839
840
static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
841
{
842
unsigned index;
843
struct hlist_head *bucket;
844
struct kvm_mmu_page *sp;
845
struct hlist_node *node, *n;
846
int r;
847
848
pgprintk("%s: looking for gfn %lx\n", __FUNCTION__, gfn);
849
r = 0;
850
index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
851
bucket = &kvm->arch.mmu_page_hash[index];
852
hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
853
if (sp->gfn == gfn && !sp->role.metaphysical) {
854
pgprintk("%s: gfn %lx role %x\n", __FUNCTION__, gfn,
855
sp->role.word);
856
kvm_mmu_zap_page(kvm, sp);
857
r = 1;
858
}
859
return r;
860
}
861
862
static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
863
{
864
struct kvm_mmu_page *sp;
865
866
while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
867
pgprintk("%s: zap %lx %x\n", __FUNCTION__, gfn, sp->role.word);
868
kvm_mmu_zap_page(kvm, sp);
869
}
870
}
871
872
static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
873
{
874
int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
875
struct kvm_mmu_page *sp = page_header(__pa(pte));
876
877
__set_bit(slot, &sp->slot_bitmap);
878
}
879
880
struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
881
{
882
gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
883
884
if (gpa == UNMAPPED_GVA)
885
return NULL;
886
return gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
887
}
888
889
static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
890
unsigned pt_access, unsigned pte_access,
891
int user_fault, int write_fault, int dirty,
892
int *ptwrite, gfn_t gfn)
893
{
894
u64 spte;
895
int was_rmapped = is_rmap_pte(*shadow_pte);
896
struct page *page;
897
898
pgprintk("%s: spte %llx access %x write_fault %d"
899
" user_fault %d gfn %lx\n",
900
__FUNCTION__, *shadow_pte, pt_access,
901
write_fault, user_fault, gfn);
902
903
/*
904
* We don't set the accessed bit, since we sometimes want to see
905
* whether the guest actually used the pte (in order to detect
906
* demand paging).
907
*/
908
spte = PT_PRESENT_MASK | PT_DIRTY_MASK;
909
if (!dirty)
910
pte_access &= ~ACC_WRITE_MASK;
911
if (!(pte_access & ACC_EXEC_MASK))
912
spte |= PT64_NX_MASK;
913
914
page = gfn_to_page(vcpu->kvm, gfn);
915
916
spte |= PT_PRESENT_MASK;
917
if (pte_access & ACC_USER_MASK)
918
spte |= PT_USER_MASK;
919
920
if (is_error_page(page)) {
921
set_shadow_pte(shadow_pte,
922
shadow_trap_nonpresent_pte | PT_SHADOW_IO_MARK);
923
kvm_release_page_clean(page);
924
return;
925
}
926
927
spte |= page_to_phys(page);
928
929
if ((pte_access & ACC_WRITE_MASK)
930
|| (write_fault && !is_write_protection(vcpu) && !user_fault)) {
931
struct kvm_mmu_page *shadow;
932
933
spte |= PT_WRITABLE_MASK;
934
if (user_fault) {
935
mmu_unshadow(vcpu->kvm, gfn);
936
goto unshadowed;
937
}
938
939
shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
940
if (shadow) {
941
pgprintk("%s: found shadow page for %lx, marking ro\n",
942
__FUNCTION__, gfn);
943
pte_access &= ~ACC_WRITE_MASK;
944
if (is_writeble_pte(spte)) {
945
spte &= ~PT_WRITABLE_MASK;
946
kvm_x86_ops->tlb_flush(vcpu);
947
}
948
if (write_fault)
949
*ptwrite = 1;
950
}
951
}
952
953
unshadowed:
954
955
if (pte_access & ACC_WRITE_MASK)
956
mark_page_dirty(vcpu->kvm, gfn);
957
958
pgprintk("%s: setting spte %llx\n", __FUNCTION__, spte);
959
set_shadow_pte(shadow_pte, spte);
960
page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
961
if (!was_rmapped) {
962
rmap_add(vcpu, shadow_pte, gfn);
963
if (!is_rmap_pte(*shadow_pte))
964
kvm_release_page_clean(page);
965
}
966
else
967
kvm_release_page_clean(page);
968
if (!ptwrite || !*ptwrite)
969
vcpu->arch.last_pte_updated = shadow_pte;
970
}
971
972
static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
973
{
974
}
975
976
static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
977
{
978
int level = PT32E_ROOT_LEVEL;
979
hpa_t table_addr = vcpu->arch.mmu.root_hpa;
980
int pt_write = 0;
981
982
for (; ; level--) {
983
u32 index = PT64_INDEX(v, level);
984
u64 *table;
985
986
ASSERT(VALID_PAGE(table_addr));
987
table = __va(table_addr);
988
989
if (level == 1) {
990
mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
991
0, write, 1, &pt_write, gfn);
992
return pt_write || is_io_pte(table[index]);
993
}
994
995
if (table[index] == shadow_trap_nonpresent_pte) {
996
struct kvm_mmu_page *new_table;
997
gfn_t pseudo_gfn;
998
999
pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
1000
>> PAGE_SHIFT;
1001
new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
1002
v, level - 1,
1003
1, ACC_ALL, &table[index],
1004
NULL);
1005
if (!new_table) {
1006
pgprintk("nonpaging_map: ENOMEM\n");
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
1017
static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
1018
struct kvm_mmu_page *sp)
1019
{
1020
int i;
1021
1022
for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1023
sp->spt[i] = shadow_trap_nonpresent_pte;
1024
}
1025
1026
static void mmu_free_roots(struct kvm_vcpu *vcpu)
1027
{
1028
int i;
1029
struct kvm_mmu_page *sp;
1030
1031
if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1032
return;
1033
#ifdef CONFIG_X86_64
1034
if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1035
hpa_t root = vcpu->arch.mmu.root_hpa;
1036
1037
sp = page_header(root);
1038
--sp->root_count;
1039
vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1040
return;
1041
}
1042
#endif
1043
for (i = 0; i < 4; ++i) {
1044
hpa_t root = vcpu->arch.mmu.pae_root[i];
1045
1046
if (root) {
1047
root &= PT64_BASE_ADDR_MASK;
1048
sp = page_header(root);
1049
--sp->root_count;
1050
}
1051
vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1052
}
1053
vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1054
}
1055
1056
static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1057
{
1058
int i;
1059
gfn_t root_gfn;
1060
struct kvm_mmu_page *sp;
1061
1062
root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1063
1064
#ifdef CONFIG_X86_64
1065
if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1066
hpa_t root = vcpu->arch.mmu.root_hpa;
1067
1068
ASSERT(!VALID_PAGE(root));
1069
sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1070
PT64_ROOT_LEVEL, 0, ACC_ALL, NULL, NULL);
1071
root = __pa(sp->spt);
1072
++sp->root_count;
1073
vcpu->arch.mmu.root_hpa = root;
1074
return;
1075
}
1076
#endif
1077
for (i = 0; i < 4; ++i) {
1078
hpa_t root = vcpu->arch.mmu.pae_root[i];
1079
1080
ASSERT(!VALID_PAGE(root));
1081
if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1082
if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1083
vcpu->arch.mmu.pae_root[i] = 0;
1084
continue;
1085
}
1086
root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1087
} else if (vcpu->arch.mmu.root_level == 0)
1088
root_gfn = 0;
1089
sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1090
PT32_ROOT_LEVEL, !is_paging(vcpu),
1091
ACC_ALL, NULL, NULL);
1092
root = __pa(sp->spt);
1093
++sp->root_count;
1094
vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1095
}
1096
vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1097
}
1098
1099
static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1100
{
1101
return vaddr;
1102
}
1103
1104
static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1105
u32 error_code)
1106
{
1107
gfn_t gfn;
1108
int r;
1109
1110
pgprintk("%s: gva %lx error %x\n", __FUNCTION__, gva, error_code);
1111
r = mmu_topup_memory_caches(vcpu);
1112
if (r)
1113
return r;
1114
1115
ASSERT(vcpu);
1116
ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1117
1118
gfn = gva >> PAGE_SHIFT;
1119
1120
return nonpaging_map(vcpu, gva & PAGE_MASK,
1121
error_code & PFERR_WRITE_MASK, gfn);
1122
}
1123
1124
static void nonpaging_free(struct kvm_vcpu *vcpu)
1125
{
1126
mmu_free_roots(vcpu);
1127
}
1128
1129
static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1130
{
1131
struct kvm_mmu *context = &vcpu->arch.mmu;
1132
1133
context->new_cr3 = nonpaging_new_cr3;
1134
context->page_fault = nonpaging_page_fault;
1135
context->gva_to_gpa = nonpaging_gva_to_gpa;
1136
context->free = nonpaging_free;
1137
context->prefetch_page = nonpaging_prefetch_page;
1138
context->root_level = 0;
1139
context->shadow_root_level = PT32E_ROOT_LEVEL;
1140
context->root_hpa = INVALID_PAGE;
1141
return 0;
1142
}
1143
1144
void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1145
{
1146
++vcpu->stat.tlb_flush;
1147
kvm_x86_ops->tlb_flush(vcpu);
1148
}
1149
1150
static void paging_new_cr3(struct kvm_vcpu *vcpu)
1151
{
1152
pgprintk("%s: cr3 %lx\n", __FUNCTION__, vcpu->cr3);
1153
mmu_free_roots(vcpu);
1154
}
1155
1156
static void inject_page_fault(struct kvm_vcpu *vcpu,
1157
u64 addr,
1158
u32 err_code)
1159
{
1160
kvm_inject_page_fault(vcpu, addr, err_code);
1161
}
1162
1163
static void paging_free(struct kvm_vcpu *vcpu)
1164
{
1165
nonpaging_free(vcpu);
1166
}
1167
1168
#define PTTYPE 64
1169
#include "paging_tmpl.h"
1170
#undef PTTYPE
1171
1172
#define PTTYPE 32
1173
#include "paging_tmpl.h"
1174
#undef PTTYPE
1175
1176
static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1177
{
1178
struct kvm_mmu *context = &vcpu->arch.mmu;
1179
1180
ASSERT(is_pae(vcpu));
1181
context->new_cr3 = paging_new_cr3;
1182
context->page_fault = paging64_page_fault;
1183
context->gva_to_gpa = paging64_gva_to_gpa;
1184
context->prefetch_page = paging64_prefetch_page;
1185
context->free = paging_free;
1186
context->root_level = level;
1187
context->shadow_root_level = level;
1188
context->root_hpa = INVALID_PAGE;
1189
return 0;
1190
}
1191
1192
static int paging64_init_context(struct kvm_vcpu *vcpu)
1193
{
1194
return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1195
}
1196
1197
static int paging32_init_context(struct kvm_vcpu *vcpu)
1198
{
1199
struct kvm_mmu *context = &vcpu->arch.mmu;
1200
1201
context->new_cr3 = paging_new_cr3;
1202
context->page_fault = paging32_page_fault;
1203
context->gva_to_gpa = paging32_gva_to_gpa;
1204
context->free = paging_free;
1205
context->prefetch_page = paging32_prefetch_page;
1206
context->root_level = PT32_ROOT_LEVEL;
1207
context->shadow_root_level = PT32E_ROOT_LEVEL;
1208
context->root_hpa = INVALID_PAGE;
1209
return 0;
1210
}
1211
1212
static int paging32E_init_context(struct kvm_vcpu *vcpu)
1213
{
1214
return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1215
}
1216
1217
static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1218
{
1219
ASSERT(vcpu);
1220
ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1221
1222
if (!is_paging(vcpu))
1223
return nonpaging_init_context(vcpu);
1224
else if (is_long_mode(vcpu))
1225
return paging64_init_context(vcpu);
1226
else if (is_pae(vcpu))
1227
return paging32E_init_context(vcpu);
1228
else
1229
return paging32_init_context(vcpu);
1230
}
1231
1232
static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1233
{
1234
ASSERT(vcpu);
1235
if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1236
vcpu->arch.mmu.free(vcpu);
1237
vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1238
}
1239
}
1240
1241
int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1242
{
1243
destroy_kvm_mmu(vcpu);
1244
return init_kvm_mmu(vcpu);
1245
}
1246
EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1247
1248
int kvm_mmu_load(struct kvm_vcpu *vcpu)
1249
{
1250
int r;
1251
1252
mutex_lock(&vcpu->kvm->lock);
1253
r = mmu_topup_memory_caches(vcpu);
1254
if (r)
1255
goto out;
1256
mmu_alloc_roots(vcpu);
1257
kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1258
kvm_mmu_flush_tlb(vcpu);
1259
out:
1260
mutex_unlock(&vcpu->kvm->lock);
1261
return r;
1262
}
1263
EXPORT_SYMBOL_GPL(kvm_mmu_load);
1264
1265
void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1266
{
1267
mmu_free_roots(vcpu);
1268
}
1269
1270
static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1271
struct kvm_mmu_page *sp,
1272
u64 *spte)
1273
{
1274
u64 pte;
1275
struct kvm_mmu_page *child;
1276
1277
pte = *spte;
1278
if (is_shadow_present_pte(pte)) {
1279
if (sp->role.level == PT_PAGE_TABLE_LEVEL)
1280
rmap_remove(vcpu->kvm, spte);
1281
else {
1282
child = page_header(pte & PT64_BASE_ADDR_MASK);
1283
mmu_page_remove_parent_pte(child, spte);
1284
}
1285
}
1286
set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1287
}
1288
1289
static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1290
struct kvm_mmu_page *sp,
1291
u64 *spte,
1292
const void *new, int bytes,
1293
int offset_in_pte)
1294
{
1295
if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
1296
++vcpu->kvm->stat.mmu_pde_zapped;
1297
return;
1298
}
1299
1300
++vcpu->kvm->stat.mmu_pte_updated;
1301
if (sp->role.glevels == PT32_ROOT_LEVEL)
1302
paging32_update_pte(vcpu, sp, spte, new, bytes, offset_in_pte);
1303
else
1304
paging64_update_pte(vcpu, sp, spte, new, bytes, offset_in_pte);
1305
}
1306
1307
static bool need_remote_flush(u64 old, u64 new)
1308
{
1309
if (!is_shadow_present_pte(old))
1310
return false;
1311
if (!is_shadow_present_pte(new))
1312
return true;
1313
if ((old ^ new) & PT64_BASE_ADDR_MASK)
1314
return true;
1315
old ^= PT64_NX_MASK;
1316
new ^= PT64_NX_MASK;
1317
return (old & ~new & PT64_PERM_MASK) != 0;
1318
}
1319
1320
static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1321
{
1322
if (need_remote_flush(old, new))
1323
kvm_flush_remote_tlbs(vcpu->kvm);
1324
else
1325
kvm_mmu_flush_tlb(vcpu);
1326
}
1327
1328
static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1329
{
1330
u64 *spte = vcpu->arch.last_pte_updated;
1331
1332
return !!(spte && (*spte & PT_ACCESSED_MASK));
1333
}
1334
1335
void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1336
const u8 *new, int bytes)
1337
{
1338
gfn_t gfn = gpa >> PAGE_SHIFT;
1339
struct kvm_mmu_page *sp;
1340
struct hlist_node *node, *n;
1341
struct hlist_head *bucket;
1342
unsigned index;
1343
u64 entry;
1344
u64 *spte;
1345
unsigned offset = offset_in_page(gpa);
1346
unsigned pte_size;
1347
unsigned page_offset;
1348
unsigned misaligned;
1349
unsigned quadrant;
1350
int level;
1351
int flooded = 0;
1352
int npte;
1353
1354
pgprintk("%s: gpa %llx bytes %d\n", __FUNCTION__, gpa, bytes);
1355
++vcpu->kvm->stat.mmu_pte_write;
1356
kvm_mmu_audit(vcpu, "pre pte write");
1357
if (gfn == vcpu->arch.last_pt_write_gfn
1358
&& !last_updated_pte_accessed(vcpu)) {
1359
++vcpu->arch.last_pt_write_count;
1360
if (vcpu->arch.last_pt_write_count >= 3)
1361
flooded = 1;
1362
} else {
1363
vcpu->arch.last_pt_write_gfn = gfn;
1364
vcpu->arch.last_pt_write_count = 1;
1365
vcpu->arch.last_pte_updated = NULL;
1366
}
1367
index = kvm_page_table_hashfn(gfn) % KVM_NUM_MMU_PAGES;
1368
bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1369
hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1370
if (sp->gfn != gfn || sp->role.metaphysical)
1371
continue;
1372
pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1373
misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1374
misaligned |= bytes < 4;
1375
if (misaligned || flooded) {
1376
/*
1377
* Misaligned accesses are too much trouble to fix
1378
* up; also, they usually indicate a page is not used
1379
* as a page table.
1380
*
1381
* If we're seeing too many writes to a page,
1382
* it may no longer be a page table, or we may be
1383
* forking, in which case it is better to unmap the
1384
* page.
1385
*/
1386
pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1387
gpa, bytes, sp->role.word);
1388
kvm_mmu_zap_page(vcpu->kvm, sp);
1389
++vcpu->kvm->stat.mmu_flooded;
1390
continue;
1391
}
1392
page_offset = offset;
1393
level = sp->role.level;
1394
npte = 1;
1395
if (sp->role.glevels == PT32_ROOT_LEVEL) {
1396
page_offset <<= 1; /* 32->64 */
1397
/*
1398
* A 32-bit pde maps 4MB while the shadow pdes map
1399
* only 2MB. So we need to double the offset again
1400
* and zap two pdes instead of one.
1401
*/
1402
if (level == PT32_ROOT_LEVEL) {
1403
page_offset &= ~7; /* kill rounding error */
1404
page_offset <<= 1;
1405
npte = 2;
1406
}
1407
quadrant = page_offset >> PAGE_SHIFT;
1408
page_offset &= ~PAGE_MASK;
1409
if (quadrant != sp->role.quadrant)
1410
continue;
1411
}
1412
spte = &sp->spt[page_offset / sizeof(*spte)];
1413
while (npte--) {
1414
entry = *spte;
1415
mmu_pte_write_zap_pte(vcpu, sp, spte);
1416
mmu_pte_write_new_pte(vcpu, sp, spte, new, bytes,
1417
page_offset & (pte_size - 1));
1418
mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1419
++spte;
1420
}
1421
}
1422
kvm_mmu_audit(vcpu, "post pte write");
1423
}
1424
1425
int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1426
{
1427
gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1428
1429
return kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1430
}
1431
1432
void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1433
{
1434
while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
1435
struct kvm_mmu_page *sp;
1436
1437
sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
1438
struct kvm_mmu_page, link);
1439
kvm_mmu_zap_page(vcpu->kvm, sp);
1440
++vcpu->kvm->stat.mmu_recycled;
1441
}
1442
}
1443
1444
int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1445
{
1446
int r;
1447
enum emulation_result er;
1448
1449
mutex_lock(&vcpu->kvm->lock);
1450
r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
1451
if (r < 0)
1452
goto out;
1453
1454
if (!r) {
1455
r = 1;
1456
goto out;
1457
}
1458
1459
r = mmu_topup_memory_caches(vcpu);
1460
if (r)
1461
goto out;
1462
1463
er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1464
mutex_unlock(&vcpu->kvm->lock);
1465
1466
switch (er) {
1467
case EMULATE_DONE:
1468
return 1;
1469
case EMULATE_DO_MMIO:
1470
++vcpu->stat.mmio_exits;
1471
return 0;
1472
case EMULATE_FAIL:
1473
kvm_report_emulation_failure(vcpu, "pagetable");
1474
return 1;
1475
default:
1476
BUG();
1477
}
1478
out:
1479
mutex_unlock(&vcpu->kvm->lock);
1480
return r;
1481
}
1482
EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1483
1484
static void free_mmu_pages(struct kvm_vcpu *vcpu)
1485
{
1486
struct kvm_mmu_page *sp;
1487
1488
while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
1489
sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
1490
struct kvm_mmu_page, link);
1491
kvm_mmu_zap_page(vcpu->kvm, sp);
1492
}
1493
free_page((unsigned long)vcpu->arch.mmu.pae_root);
1494
}
1495
1496
static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1497
{
1498
struct page *page;
1499
int i;
1500
1501
ASSERT(vcpu);
1502
1503
if (vcpu->kvm->arch.n_requested_mmu_pages)
1504
vcpu->kvm->arch.n_free_mmu_pages =
1505
vcpu->kvm->arch.n_requested_mmu_pages;
1506
else
1507
vcpu->kvm->arch.n_free_mmu_pages =
1508
vcpu->kvm->arch.n_alloc_mmu_pages;
1509
/*
1510
* When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1511
* Therefore we need to allocate shadow page tables in the first
1512
* 4GB of memory, which happens to fit the DMA32 zone.
1513
*/
1514
page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1515
if (!page)
1516
goto error_1;
1517
vcpu->arch.mmu.pae_root = page_address(page);
1518
for (i = 0; i < 4; ++i)
1519
vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1520
1521
return 0;
1522
1523
error_1:
1524
free_mmu_pages(vcpu);
1525
return -ENOMEM;
1526
}
1527
1528
int kvm_mmu_create(struct kvm_vcpu *vcpu)
1529
{
1530
ASSERT(vcpu);
1531
ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1532
1533
return alloc_mmu_pages(vcpu);
1534
}
1535
1536
int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1537
{
1538
ASSERT(vcpu);
1539
ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1540
1541
return init_kvm_mmu(vcpu);
1542
}
1543
1544
void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1545
{
1546
ASSERT(vcpu);
1547
1548
destroy_kvm_mmu(vcpu);
1549
free_mmu_pages(vcpu);
1550
mmu_free_memory_caches(vcpu);
1551
}
1552
1553
void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1554
{
1555
struct kvm_mmu_page *sp;
1556
1557
list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
1558
int i;
1559
u64 *pt;
1560
1561
if (!test_bit(slot, &sp->slot_bitmap))
1562
continue;
1563
1564
pt = sp->spt;
1565
for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1566
/* avoid RMW */
1567
if (pt[i] & PT_WRITABLE_MASK)
1568
pt[i] &= ~PT_WRITABLE_MASK;
1569
}
1570
}
1571
1572
void kvm_mmu_zap_all(struct kvm *kvm)
1573
{
1574
struct kvm_mmu_page *sp, *node;
1575
1576
list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
1577
kvm_mmu_zap_page(kvm, sp);
1578
1579
kvm_flush_remote_tlbs(kvm);
1580
}
1581
1582
void kvm_mmu_module_exit(void)
1583
{
1584
if (pte_chain_cache)
1585
kmem_cache_destroy(pte_chain_cache);
1586
if (rmap_desc_cache)
1587
kmem_cache_destroy(rmap_desc_cache);
1588
if (mmu_page_header_cache)
1589
kmem_cache_destroy(mmu_page_header_cache);
1590
}
1591
1592
int kvm_mmu_module_init(void)
1593
{
1594
pte_chain_cache = kmem_cache_create("kvm_pte_chain",
1595
sizeof(struct kvm_pte_chain),
1596
0, 0, NULL);
1597
if (!pte_chain_cache)
1598
goto nomem;
1599
rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
1600
sizeof(struct kvm_rmap_desc),
1601
0, 0, NULL);
1602
if (!rmap_desc_cache)
1603
goto nomem;
1604
1605
mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
1606
sizeof(struct kvm_mmu_page),
1607
0, 0, NULL);
1608
if (!mmu_page_header_cache)
1609
goto nomem;
1610
1611
return 0;
1612
1613
nomem:
1614
kvm_mmu_module_exit();
1615
return -ENOMEM;
1616
}
1617
1618
/*
1619
* Caculate mmu pages needed for kvm.
1620
*/
1621
unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
1622
{
1623
int i;
1624
unsigned int nr_mmu_pages;
1625
unsigned int nr_pages = 0;
1626
1627
for (i = 0; i < kvm->nmemslots; i++)
1628
nr_pages += kvm->memslots[i].npages;
1629
1630
nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
1631
nr_mmu_pages = max(nr_mmu_pages,
1632
(unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
1633
1634
return nr_mmu_pages;
1635
}
1636
1637
#ifdef AUDIT
1638
1639
static const char *audit_msg;
1640
1641
static gva_t canonicalize(gva_t gva)
1642
{
1643
#ifdef CONFIG_X86_64
1644
gva = (long long)(gva << 16) >> 16;
1645
#endif
1646
return gva;
1647
}
1648
1649
static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
1650
gva_t va, int level)
1651
{
1652
u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
1653
int i;
1654
gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
1655
1656
for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
1657
u64 ent = pt[i];
1658
1659
if (ent == shadow_trap_nonpresent_pte)
1660
continue;
1661
1662
va = canonicalize(va);
1663
if (level > 1) {
1664
if (ent == shadow_notrap_nonpresent_pte)
1665
printk(KERN_ERR "audit: (%s) nontrapping pte"
1666
" in nonleaf level: levels %d gva %lx"
1667
" level %d pte %llx\n", audit_msg,
1668
vcpu->arch.mmu.root_level, va, level, ent);
1669
1670
audit_mappings_page(vcpu, ent, va, level - 1);
1671
} else {
1672
gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
1673
struct page *page = gpa_to_page(vcpu, gpa);
1674
hpa_t hpa = page_to_phys(page);
1675
1676
if (is_shadow_present_pte(ent)
1677
&& (ent & PT64_BASE_ADDR_MASK) != hpa)
1678
printk(KERN_ERR "xx audit error: (%s) levels %d"
1679
" gva %lx gpa %llx hpa %llx ent %llx %d\n",
1680
audit_msg, vcpu->arch.mmu.root_level,
1681
va, gpa, hpa, ent,
1682
is_shadow_present_pte(ent));
1683
else if (ent == shadow_notrap_nonpresent_pte
1684
&& !is_error_hpa(hpa))
1685
printk(KERN_ERR "audit: (%s) notrap shadow,"
1686
" valid guest gva %lx\n", audit_msg, va);
1687
kvm_release_page_clean(page);
1688
1689
}
1690
}
1691
}
1692
1693
static void audit_mappings(struct kvm_vcpu *vcpu)
1694
{
1695
unsigned i;
1696
1697
if (vcpu->arch.mmu.root_level == 4)
1698
audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
1699
else
1700
for (i = 0; i < 4; ++i)
1701
if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
1702
audit_mappings_page(vcpu,
1703
vcpu->arch.mmu.pae_root[i],
1704
i << 30,
1705
2);
1706
}
1707
1708
static int count_rmaps(struct kvm_vcpu *vcpu)
1709
{
1710
int nmaps = 0;
1711
int i, j, k;
1712
1713
for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
1714
struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
1715
struct kvm_rmap_desc *d;
1716
1717
for (j = 0; j < m->npages; ++j) {
1718
unsigned long *rmapp = &m->rmap[j];
1719
1720
if (!*rmapp)
1721
continue;
1722
if (!(*rmapp & 1)) {
1723
++nmaps;
1724
continue;
1725
}
1726
d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
1727
while (d) {
1728
for (k = 0; k < RMAP_EXT; ++k)
1729
if (d->shadow_ptes[k])
1730
++nmaps;
1731
else
1732
break;
1733
d = d->more;
1734
}
1735
}
1736
}
1737
return nmaps;
1738
}
1739
1740
static int count_writable_mappings(struct kvm_vcpu *vcpu)
1741
{
1742
int nmaps = 0;
1743
struct kvm_mmu_page *sp;
1744
int i;
1745
1746
list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
1747
u64 *pt = sp->spt;
1748
1749
if (sp->role.level != PT_PAGE_TABLE_LEVEL)
1750
continue;
1751
1752
for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1753
u64 ent = pt[i];
1754
1755
if (!(ent & PT_PRESENT_MASK))
1756
continue;
1757
if (!(ent & PT_WRITABLE_MASK))
1758
continue;
1759
++nmaps;
1760
}
1761
}
1762
return nmaps;
1763
}
1764
1765
static void audit_rmap(struct kvm_vcpu *vcpu)
1766
{
1767
int n_rmap = count_rmaps(vcpu);
1768
int n_actual = count_writable_mappings(vcpu);
1769
1770
if (n_rmap != n_actual)
1771
printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
1772
__FUNCTION__, audit_msg, n_rmap, n_actual);
1773
}
1774
1775
static void audit_write_protection(struct kvm_vcpu *vcpu)
1776
{
1777
struct kvm_mmu_page *sp;
1778
struct kvm_memory_slot *slot;
1779
unsigned long *rmapp;
1780
gfn_t gfn;
1781
1782
list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
1783
if (sp->role.metaphysical)
1784
continue;
1785
1786
slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
1787
gfn = unalias_gfn(vcpu->kvm, sp->gfn);
1788
rmapp = &slot->rmap[gfn - slot->base_gfn];
1789
if (*rmapp)
1790
printk(KERN_ERR "%s: (%s) shadow page has writable"
1791
" mappings: gfn %lx role %x\n",
1792
__FUNCTION__, audit_msg, sp->gfn,
1793
sp->role.word);
1794
}
1795
}
1796
1797
static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
1798
{
1799
int olddbg = dbg;
1800
1801
dbg = 0;
1802
audit_msg = msg;
1803
audit_rmap(vcpu);
1804
audit_write_protection(vcpu);
1805
audit_mappings(vcpu);
1806
dbg = olddbg;
1807
}
1808
1809
#endif
Loggerhead is a web-based interface for Breezy
Version: 2.0.1