?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
KVM: Add local get_mtrr_type() to support MTRR
[linux-2.6/mini2440.git] / arch / x86 / kvm / mmu.c
blobac2304fd173ed81366dd325bdbc83b180378c9be
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 #include <linux/hugetlb.h>
31 #include <linux/compiler.h>
32
33 #include <asm/page.h>
34 #include <asm/cmpxchg.h>
35 #include <asm/io.h>
36
37 /*
38 * When setting this variable to true it enables Two-Dimensional-Paging
39 * where the hardware walks 2 page tables:
40 * 1. the guest-virtual to guest-physical
41 * 2. while doing 1. it walks guest-physical to host-physical
42 * If the hardware supports that we don't need to do shadow paging.
43 */
44 bool tdp_enabled = false;
45
46 #undef MMU_DEBUG
47
48 #undef AUDIT
49
50 #ifdef AUDIT
51 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
52 #else
53 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg) {}
54 #endif
55
56 #ifdef MMU_DEBUG
57
58 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
59 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
60
61 #else
62
63 #define pgprintk(x...) do { } while (0)
64 #define rmap_printk(x...) do { } while (0)
65
66 #endif
67
68 #if defined(MMU_DEBUG) || defined(AUDIT)
69 static int dbg = 0;
70 module_param(dbg, bool, 0644);
71 #endif
72
73 static int oos_shadow = 1;
74 module_param(oos_shadow, bool, 0644);
75
76 #ifndef MMU_DEBUG
77 #define ASSERT(x) do { } while (0)
78 #else
79 #define ASSERT(x) \
80 if (!(x)) { \
81 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
82 __FILE__, __LINE__, #x); \
83 }
84 #endif
85
86 #define PT_FIRST_AVAIL_BITS_SHIFT 9
87 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
88
89 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
90
91 #define PT64_LEVEL_BITS 9
92
93 #define PT64_LEVEL_SHIFT(level) \
94 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
95
96 #define PT64_LEVEL_MASK(level) \
97 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
98
99 #define PT64_INDEX(address, level)\
100 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
101
102
103 #define PT32_LEVEL_BITS 10
104
105 #define PT32_LEVEL_SHIFT(level) \
106 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
107
108 #define PT32_LEVEL_MASK(level) \
109 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
110
111 #define PT32_INDEX(address, level)\
112 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
113
114
115 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
116 #define PT64_DIR_BASE_ADDR_MASK \
117 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
118
119 #define PT32_BASE_ADDR_MASK PAGE_MASK
120 #define PT32_DIR_BASE_ADDR_MASK \
121 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
122
123 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
124 | PT64_NX_MASK)
125
126 #define PFERR_PRESENT_MASK (1U << 0)
127 #define PFERR_WRITE_MASK (1U << 1)
128 #define PFERR_USER_MASK (1U << 2)
129 #define PFERR_FETCH_MASK (1U << 4)
130
131 #define PT_DIRECTORY_LEVEL 2
132 #define PT_PAGE_TABLE_LEVEL 1
133
134 #define RMAP_EXT 4
135
136 #define ACC_EXEC_MASK 1
137 #define ACC_WRITE_MASK PT_WRITABLE_MASK
138 #define ACC_USER_MASK PT_USER_MASK
139 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
140
141 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
142
143 struct kvm_rmap_desc {
144 u64 *shadow_ptes[RMAP_EXT];
145 struct kvm_rmap_desc *more;
146 };
147
148 struct kvm_shadow_walk {
149 int (*entry)(struct kvm_shadow_walk *walk, struct kvm_vcpu *vcpu,
150 u64 addr, u64 *spte, int level);
151 };
152
153 struct kvm_unsync_walk {
154 int (*entry) (struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk);
155 };
156
157 typedef int (*mmu_parent_walk_fn) (struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp);
158
159 static struct kmem_cache *pte_chain_cache;
160 static struct kmem_cache *rmap_desc_cache;
161 static struct kmem_cache *mmu_page_header_cache;
162
163 static u64 __read_mostly shadow_trap_nonpresent_pte;
164 static u64 __read_mostly shadow_notrap_nonpresent_pte;
165 static u64 __read_mostly shadow_base_present_pte;
166 static u64 __read_mostly shadow_nx_mask;
167 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
168 static u64 __read_mostly shadow_user_mask;
169 static u64 __read_mostly shadow_accessed_mask;
170 static u64 __read_mostly shadow_dirty_mask;
171
172 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
173 {
174 shadow_trap_nonpresent_pte = trap_pte;
175 shadow_notrap_nonpresent_pte = notrap_pte;
176 }
177 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
178
179 void kvm_mmu_set_base_ptes(u64 base_pte)
180 {
181 shadow_base_present_pte = base_pte;
182 }
183 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
184
185 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
186 u64 dirty_mask, u64 nx_mask, u64 x_mask)
187 {
188 shadow_user_mask = user_mask;
189 shadow_accessed_mask = accessed_mask;
190 shadow_dirty_mask = dirty_mask;
191 shadow_nx_mask = nx_mask;
192 shadow_x_mask = x_mask;
193 }
194 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
195
196 static int is_write_protection(struct kvm_vcpu *vcpu)
197 {
198 return vcpu->arch.cr0 & X86_CR0_WP;
199 }
200
201 static int is_cpuid_PSE36(void)
202 {
203 return 1;
204 }
205
206 static int is_nx(struct kvm_vcpu *vcpu)
207 {
208 return vcpu->arch.shadow_efer & EFER_NX;
209 }
210
211 static int is_present_pte(unsigned long pte)
212 {
213 return pte & PT_PRESENT_MASK;
214 }
215
216 static int is_shadow_present_pte(u64 pte)
217 {
218 return pte != shadow_trap_nonpresent_pte
219 && pte != shadow_notrap_nonpresent_pte;
220 }
221
222 static int is_large_pte(u64 pte)
223 {
224 return pte & PT_PAGE_SIZE_MASK;
225 }
226
227 static int is_writeble_pte(unsigned long pte)
228 {
229 return pte & PT_WRITABLE_MASK;
230 }
231
232 static int is_dirty_pte(unsigned long pte)
233 {
234 return pte & shadow_dirty_mask;
235 }
236
237 static int is_rmap_pte(u64 pte)
238 {
239 return is_shadow_present_pte(pte);
240 }
241
242 static pfn_t spte_to_pfn(u64 pte)
243 {
244 return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
245 }
246
247 static gfn_t pse36_gfn_delta(u32 gpte)
248 {
249 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
250
251 return (gpte & PT32_DIR_PSE36_MASK) << shift;
252 }
253
254 static void set_shadow_pte(u64 *sptep, u64 spte)
255 {
256 #ifdef CONFIG_X86_64
257 set_64bit((unsigned long *)sptep, spte);
258 #else
259 set_64bit((unsigned long long *)sptep, spte);
260 #endif
261 }
262
263 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
264 struct kmem_cache *base_cache, int min)
265 {
266 void *obj;
267
268 if (cache->nobjs >= min)
269 return 0;
270 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
271 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
272 if (!obj)
273 return -ENOMEM;
274 cache->objects[cache->nobjs++] = obj;
275 }
276 return 0;
277 }
278
279 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
280 {
281 while (mc->nobjs)
282 kfree(mc->objects[--mc->nobjs]);
283 }
284
285 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
286 int min)
287 {
288 struct page *page;
289
290 if (cache->nobjs >= min)
291 return 0;
292 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
293 page = alloc_page(GFP_KERNEL);
294 if (!page)
295 return -ENOMEM;
296 set_page_private(page, 0);
297 cache->objects[cache->nobjs++] = page_address(page);
298 }
299 return 0;
300 }
301
302 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
303 {
304 while (mc->nobjs)
305 free_page((unsigned long)mc->objects[--mc->nobjs]);
306 }
307
308 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
309 {
310 int r;
311
312 r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
313 pte_chain_cache, 4);
314 if (r)
315 goto out;
316 r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
317 rmap_desc_cache, 4);
318 if (r)
319 goto out;
320 r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
321 if (r)
322 goto out;
323 r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
324 mmu_page_header_cache, 4);
325 out:
326 return r;
327 }
328
329 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
330 {
331 mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
332 mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
333 mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
334 mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
335 }
336
337 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
338 size_t size)
339 {
340 void *p;
341
342 BUG_ON(!mc->nobjs);
343 p = mc->objects[--mc->nobjs];
344 memset(p, 0, size);
345 return p;
346 }
347
348 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
349 {
350 return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
351 sizeof(struct kvm_pte_chain));
352 }
353
354 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
355 {
356 kfree(pc);
357 }
358
359 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
360 {
361 return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
362 sizeof(struct kvm_rmap_desc));
363 }
364
365 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
366 {
367 kfree(rd);
368 }
369
370 /*
371 * Return the pointer to the largepage write count for a given
372 * gfn, handling slots that are not large page aligned.
373 */
374 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
375 {
376 unsigned long idx;
377
378 idx = (gfn / KVM_PAGES_PER_HPAGE) -
379 (slot->base_gfn / KVM_PAGES_PER_HPAGE);
380 return &slot->lpage_info[idx].write_count;
381 }
382
383 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
384 {
385 int *write_count;
386
387 write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
388 *write_count += 1;
389 }
390
391 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
392 {
393 int *write_count;
394
395 write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
396 *write_count -= 1;
397 WARN_ON(*write_count < 0);
398 }
399
400 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
401 {
402 struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
403 int *largepage_idx;
404
405 if (slot) {
406 largepage_idx = slot_largepage_idx(gfn, slot);
407 return *largepage_idx;
408 }
409
410 return 1;
411 }
412
413 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
414 {
415 struct vm_area_struct *vma;
416 unsigned long addr;
417 int ret = 0;
418
419 addr = gfn_to_hva(kvm, gfn);
420 if (kvm_is_error_hva(addr))
421 return ret;
422
423 down_read(&current->mm->mmap_sem);
424 vma = find_vma(current->mm, addr);
425 if (vma && is_vm_hugetlb_page(vma))
426 ret = 1;
427 up_read(&current->mm->mmap_sem);
428
429 return ret;
430 }
431
432 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
433 {
434 struct kvm_memory_slot *slot;
435
436 if (has_wrprotected_page(vcpu->kvm, large_gfn))
437 return 0;
438
439 if (!host_largepage_backed(vcpu->kvm, large_gfn))
440 return 0;
441
442 slot = gfn_to_memslot(vcpu->kvm, large_gfn);
443 if (slot && slot->dirty_bitmap)
444 return 0;
445
446 return 1;
447 }
448
449 /*
450 * Take gfn and return the reverse mapping to it.
451 * Note: gfn must be unaliased before this function get called
452 */
453
454 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
455 {
456 struct kvm_memory_slot *slot;
457 unsigned long idx;
458
459 slot = gfn_to_memslot(kvm, gfn);
460 if (!lpage)
461 return &slot->rmap[gfn - slot->base_gfn];
462
463 idx = (gfn / KVM_PAGES_PER_HPAGE) -
464 (slot->base_gfn / KVM_PAGES_PER_HPAGE);
465
466 return &slot->lpage_info[idx].rmap_pde;
467 }
468
469 /*
470 * Reverse mapping data structures:
471 *
472 * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
473 * that points to page_address(page).
474 *
475 * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
476 * containing more mappings.
477 */
478 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
479 {
480 struct kvm_mmu_page *sp;
481 struct kvm_rmap_desc *desc;
482 unsigned long *rmapp;
483 int i;
484
485 if (!is_rmap_pte(*spte))
486 return;
487 gfn = unalias_gfn(vcpu->kvm, gfn);
488 sp = page_header(__pa(spte));
489 sp->gfns[spte - sp->spt] = gfn;
490 rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
491 if (!*rmapp) {
492 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
493 *rmapp = (unsigned long)spte;
494 } else if (!(*rmapp & 1)) {
495 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
496 desc = mmu_alloc_rmap_desc(vcpu);
497 desc->shadow_ptes[0] = (u64 *)*rmapp;
498 desc->shadow_ptes[1] = spte;
499 *rmapp = (unsigned long)desc | 1;
500 } else {
501 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
502 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
503 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
504 desc = desc->more;
505 if (desc->shadow_ptes[RMAP_EXT-1]) {
506 desc->more = mmu_alloc_rmap_desc(vcpu);
507 desc = desc->more;
508 }
509 for (i = 0; desc->shadow_ptes[i]; ++i)
510 ;
511 desc->shadow_ptes[i] = spte;
512 }
513 }
514
515 static void rmap_desc_remove_entry(unsigned long *rmapp,
516 struct kvm_rmap_desc *desc,
517 int i,
518 struct kvm_rmap_desc *prev_desc)
519 {
520 int j;
521
522 for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
523 ;
524 desc->shadow_ptes[i] = desc->shadow_ptes[j];
525 desc->shadow_ptes[j] = NULL;
526 if (j != 0)
527 return;
528 if (!prev_desc && !desc->more)
529 *rmapp = (unsigned long)desc->shadow_ptes[0];
530 else
531 if (prev_desc)
532 prev_desc->more = desc->more;
533 else
534 *rmapp = (unsigned long)desc->more | 1;
535 mmu_free_rmap_desc(desc);
536 }
537
538 static void rmap_remove(struct kvm *kvm, u64 *spte)
539 {
540 struct kvm_rmap_desc *desc;
541 struct kvm_rmap_desc *prev_desc;
542 struct kvm_mmu_page *sp;
543 pfn_t pfn;
544 unsigned long *rmapp;
545 int i;
546
547 if (!is_rmap_pte(*spte))
548 return;
549 sp = page_header(__pa(spte));
550 pfn = spte_to_pfn(*spte);
551 if (*spte & shadow_accessed_mask)
552 kvm_set_pfn_accessed(pfn);
553 if (is_writeble_pte(*spte))
554 kvm_release_pfn_dirty(pfn);
555 else
556 kvm_release_pfn_clean(pfn);
557 rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
558 if (!*rmapp) {
559 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
560 BUG();
561 } else if (!(*rmapp & 1)) {
562 rmap_printk("rmap_remove: %p %llx 1->0\n", spte, *spte);
563 if ((u64 *)*rmapp != spte) {
564 printk(KERN_ERR "rmap_remove: %p %llx 1->BUG\n",
565 spte, *spte);
566 BUG();
567 }
568 *rmapp = 0;
569 } else {
570 rmap_printk("rmap_remove: %p %llx many->many\n", spte, *spte);
571 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
572 prev_desc = NULL;
573 while (desc) {
574 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
575 if (desc->shadow_ptes[i] == spte) {
576 rmap_desc_remove_entry(rmapp,
577 desc, i,
578 prev_desc);
579 return;
580 }
581 prev_desc = desc;
582 desc = desc->more;
583 }
584 BUG();
585 }
586 }
587
588 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
589 {
590 struct kvm_rmap_desc *desc;
591 struct kvm_rmap_desc *prev_desc;
592 u64 *prev_spte;
593 int i;
594
595 if (!*rmapp)
596 return NULL;
597 else if (!(*rmapp & 1)) {
598 if (!spte)
599 return (u64 *)*rmapp;
600 return NULL;
601 }
602 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
603 prev_desc = NULL;
604 prev_spte = NULL;
605 while (desc) {
606 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
607 if (prev_spte == spte)
608 return desc->shadow_ptes[i];
609 prev_spte = desc->shadow_ptes[i];
610 }
611 desc = desc->more;
612 }
613 return NULL;
614 }
615
616 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
617 {
618 unsigned long *rmapp;
619 u64 *spte;
620 int write_protected = 0;
621
622 gfn = unalias_gfn(kvm, gfn);
623 rmapp = gfn_to_rmap(kvm, gfn, 0);
624
625 spte = rmap_next(kvm, rmapp, NULL);
626 while (spte) {
627 BUG_ON(!spte);
628 BUG_ON(!(*spte & PT_PRESENT_MASK));
629 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
630 if (is_writeble_pte(*spte)) {
631 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
632 write_protected = 1;
633 }
634 spte = rmap_next(kvm, rmapp, spte);
635 }
636 if (write_protected) {
637 pfn_t pfn;
638
639 spte = rmap_next(kvm, rmapp, NULL);
640 pfn = spte_to_pfn(*spte);
641 kvm_set_pfn_dirty(pfn);
642 }
643
644 /* check for huge page mappings */
645 rmapp = gfn_to_rmap(kvm, gfn, 1);
646 spte = rmap_next(kvm, rmapp, NULL);
647 while (spte) {
648 BUG_ON(!spte);
649 BUG_ON(!(*spte & PT_PRESENT_MASK));
650 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
651 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
652 if (is_writeble_pte(*spte)) {
653 rmap_remove(kvm, spte);
654 --kvm->stat.lpages;
655 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
656 spte = NULL;
657 write_protected = 1;
658 }
659 spte = rmap_next(kvm, rmapp, spte);
660 }
661
662 if (write_protected)
663 kvm_flush_remote_tlbs(kvm);
664 }
665
666 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
667 {
668 u64 *spte;
669 int need_tlb_flush = 0;
670
671 while ((spte = rmap_next(kvm, rmapp, NULL))) {
672 BUG_ON(!(*spte & PT_PRESENT_MASK));
673 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
674 rmap_remove(kvm, spte);
675 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
676 need_tlb_flush = 1;
677 }
678 return need_tlb_flush;
679 }
680
681 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
682 int (*handler)(struct kvm *kvm, unsigned long *rmapp))
683 {
684 int i;
685 int retval = 0;
686
687 /*
688 * If mmap_sem isn't taken, we can look the memslots with only
689 * the mmu_lock by skipping over the slots with userspace_addr == 0.
690 */
691 for (i = 0; i < kvm->nmemslots; i++) {
692 struct kvm_memory_slot *memslot = &kvm->memslots[i];
693 unsigned long start = memslot->userspace_addr;
694 unsigned long end;
695
696 /* mmu_lock protects userspace_addr */
697 if (!start)
698 continue;
699
700 end = start + (memslot->npages << PAGE_SHIFT);
701 if (hva >= start && hva < end) {
702 gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
703 retval |= handler(kvm, &memslot->rmap[gfn_offset]);
704 retval |= handler(kvm,
705 &memslot->lpage_info[
706 gfn_offset /
707 KVM_PAGES_PER_HPAGE].rmap_pde);
708 }
709 }
710
711 return retval;
712 }
713
714 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
715 {
716 return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
717 }
718
719 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
720 {
721 u64 *spte;
722 int young = 0;
723
724 /* always return old for EPT */
725 if (!shadow_accessed_mask)
726 return 0;
727
728 spte = rmap_next(kvm, rmapp, NULL);
729 while (spte) {
730 int _young;
731 u64 _spte = *spte;
732 BUG_ON(!(_spte & PT_PRESENT_MASK));
733 _young = _spte & PT_ACCESSED_MASK;
734 if (_young) {
735 young = 1;
736 clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
737 }
738 spte = rmap_next(kvm, rmapp, spte);
739 }
740 return young;
741 }
742
743 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
744 {
745 return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
746 }
747
748 #ifdef MMU_DEBUG
749 static int is_empty_shadow_page(u64 *spt)
750 {
751 u64 *pos;
752 u64 *end;
753
754 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
755 if (is_shadow_present_pte(*pos)) {
756 printk(KERN_ERR "%s: %p %llx\n", __func__,
757 pos, *pos);
758 return 0;
759 }
760 return 1;
761 }
762 #endif
763
764 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
765 {
766 ASSERT(is_empty_shadow_page(sp->spt));
767 list_del(&sp->link);
768 __free_page(virt_to_page(sp->spt));
769 __free_page(virt_to_page(sp->gfns));
770 kfree(sp);
771 ++kvm->arch.n_free_mmu_pages;
772 }
773
774 static unsigned kvm_page_table_hashfn(gfn_t gfn)
775 {
776 return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
777 }
778
779 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
780 u64 *parent_pte)
781 {
782 struct kvm_mmu_page *sp;
783
784 sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
785 sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
786 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
787 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
788 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
789 ASSERT(is_empty_shadow_page(sp->spt));
790 sp->slot_bitmap = 0;
791 sp->multimapped = 0;
792 sp->parent_pte = parent_pte;
793 --vcpu->kvm->arch.n_free_mmu_pages;
794 return sp;
795 }
796
797 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
798 struct kvm_mmu_page *sp, u64 *parent_pte)
799 {
800 struct kvm_pte_chain *pte_chain;
801 struct hlist_node *node;
802 int i;
803
804 if (!parent_pte)
805 return;
806 if (!sp->multimapped) {
807 u64 *old = sp->parent_pte;
808
809 if (!old) {
810 sp->parent_pte = parent_pte;
811 return;
812 }
813 sp->multimapped = 1;
814 pte_chain = mmu_alloc_pte_chain(vcpu);
815 INIT_HLIST_HEAD(&sp->parent_ptes);
816 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
817 pte_chain->parent_ptes[0] = old;
818 }
819 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
820 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
821 continue;
822 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
823 if (!pte_chain->parent_ptes[i]) {
824 pte_chain->parent_ptes[i] = parent_pte;
825 return;
826 }
827 }
828 pte_chain = mmu_alloc_pte_chain(vcpu);
829 BUG_ON(!pte_chain);
830 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
831 pte_chain->parent_ptes[0] = parent_pte;
832 }
833
834 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
835 u64 *parent_pte)
836 {
837 struct kvm_pte_chain *pte_chain;
838 struct hlist_node *node;
839 int i;
840
841 if (!sp->multimapped) {
842 BUG_ON(sp->parent_pte != parent_pte);
843 sp->parent_pte = NULL;
844 return;
845 }
846 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
847 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
848 if (!pte_chain->parent_ptes[i])
849 break;
850 if (pte_chain->parent_ptes[i] != parent_pte)
851 continue;
852 while (i + 1 < NR_PTE_CHAIN_ENTRIES
853 && pte_chain->parent_ptes[i + 1]) {
854 pte_chain->parent_ptes[i]
855 = pte_chain->parent_ptes[i + 1];
856 ++i;
857 }
858 pte_chain->parent_ptes[i] = NULL;
859 if (i == 0) {
860 hlist_del(&pte_chain->link);
861 mmu_free_pte_chain(pte_chain);
862 if (hlist_empty(&sp->parent_ptes)) {
863 sp->multimapped = 0;
864 sp->parent_pte = NULL;
865 }
866 }
867 return;
868 }
869 BUG();
870 }
871
872
873 static void mmu_parent_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
874 mmu_parent_walk_fn fn)
875 {
876 struct kvm_pte_chain *pte_chain;
877 struct hlist_node *node;
878 struct kvm_mmu_page *parent_sp;
879 int i;
880
881 if (!sp->multimapped && sp->parent_pte) {
882 parent_sp = page_header(__pa(sp->parent_pte));
883 fn(vcpu, parent_sp);
884 mmu_parent_walk(vcpu, parent_sp, fn);
885 return;
886 }
887 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
888 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
889 if (!pte_chain->parent_ptes[i])
890 break;
891 parent_sp = page_header(__pa(pte_chain->parent_ptes[i]));
892 fn(vcpu, parent_sp);
893 mmu_parent_walk(vcpu, parent_sp, fn);
894 }
895 }
896
897 static void kvm_mmu_update_unsync_bitmap(u64 *spte)
898 {
899 unsigned int index;
900 struct kvm_mmu_page *sp = page_header(__pa(spte));
901
902 index = spte - sp->spt;
903 __set_bit(index, sp->unsync_child_bitmap);
904 sp->unsync_children = 1;
905 }
906
907 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page *sp)
908 {
909 struct kvm_pte_chain *pte_chain;
910 struct hlist_node *node;
911 int i;
912
913 if (!sp->parent_pte)
914 return;
915
916 if (!sp->multimapped) {
917 kvm_mmu_update_unsync_bitmap(sp->parent_pte);
918 return;
919 }
920
921 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
922 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
923 if (!pte_chain->parent_ptes[i])
924 break;
925 kvm_mmu_update_unsync_bitmap(pte_chain->parent_ptes[i]);
926 }
927 }
928
929 static int unsync_walk_fn(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
930 {
931 sp->unsync_children = 1;
932 kvm_mmu_update_parents_unsync(sp);
933 return 1;
934 }
935
936 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu *vcpu,
937 struct kvm_mmu_page *sp)
938 {
939 mmu_parent_walk(vcpu, sp, unsync_walk_fn);
940 kvm_mmu_update_parents_unsync(sp);
941 }
942
943 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
944 struct kvm_mmu_page *sp)
945 {
946 int i;
947
948 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
949 sp->spt[i] = shadow_trap_nonpresent_pte;
950 }
951
952 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
953 struct kvm_mmu_page *sp)
954 {
955 return 1;
956 }
957
958 static void nonpaging_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
959 {
960 }
961
962 #define for_each_unsync_children(bitmap, idx) \
963 for (idx = find_first_bit(bitmap, 512); \
964 idx < 512; \
965 idx = find_next_bit(bitmap, 512, idx+1))
966
967 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
968 struct kvm_unsync_walk *walker)
969 {
970 int i, ret;
971
972 if (!sp->unsync_children)
973 return 0;
974
975 for_each_unsync_children(sp->unsync_child_bitmap, i) {
976 u64 ent = sp->spt[i];
977
978 if (is_shadow_present_pte(ent)) {
979 struct kvm_mmu_page *child;
980 child = page_header(ent & PT64_BASE_ADDR_MASK);
981
982 if (child->unsync_children) {
983 ret = mmu_unsync_walk(child, walker);
984 if (ret)
985 return ret;
986 __clear_bit(i, sp->unsync_child_bitmap);
987 }
988
989 if (child->unsync) {
990 ret = walker->entry(child, walker);
991 __clear_bit(i, sp->unsync_child_bitmap);
992 if (ret)
993 return ret;
994 }
995 }
996 }
997
998 if (find_first_bit(sp->unsync_child_bitmap, 512) == 512)
999 sp->unsync_children = 0;
1000
1001 return 0;
1002 }
1003
1004 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
1005 {
1006 unsigned index;
1007 struct hlist_head *bucket;
1008 struct kvm_mmu_page *sp;
1009 struct hlist_node *node;
1010
1011 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1012 index = kvm_page_table_hashfn(gfn);
1013 bucket = &kvm->arch.mmu_page_hash[index];
1014 hlist_for_each_entry(sp, node, bucket, hash_link)
1015 if (sp->gfn == gfn && !sp->role.metaphysical
1016 && !sp->role.invalid) {
1017 pgprintk("%s: found role %x\n",
1018 __func__, sp->role.word);
1019 return sp;
1020 }
1021 return NULL;
1022 }
1023
1024 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1025 {
1026 WARN_ON(!sp->unsync);
1027 sp->unsync = 0;
1028 --kvm->stat.mmu_unsync;
1029 }
1030
1031 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp);
1032
1033 static int kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1034 {
1035 if (sp->role.glevels != vcpu->arch.mmu.root_level) {
1036 kvm_mmu_zap_page(vcpu->kvm, sp);
1037 return 1;
1038 }
1039
1040 rmap_write_protect(vcpu->kvm, sp->gfn);
1041 kvm_unlink_unsync_page(vcpu->kvm, sp);
1042 if (vcpu->arch.mmu.sync_page(vcpu, sp)) {
1043 kvm_mmu_zap_page(vcpu->kvm, sp);
1044 return 1;
1045 }
1046
1047 kvm_mmu_flush_tlb(vcpu);
1048 return 0;
1049 }
1050
1051 struct sync_walker {
1052 struct kvm_vcpu *vcpu;
1053 struct kvm_unsync_walk walker;
1054 };
1055
1056 static int mmu_sync_fn(struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk)
1057 {
1058 struct sync_walker *sync_walk = container_of(walk, struct sync_walker,
1059 walker);
1060 struct kvm_vcpu *vcpu = sync_walk->vcpu;
1061
1062 kvm_sync_page(vcpu, sp);
1063 return (need_resched() || spin_needbreak(&vcpu->kvm->mmu_lock));
1064 }
1065
1066 static void mmu_sync_children(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1067 {
1068 struct sync_walker walker = {
1069 .walker = { .entry = mmu_sync_fn, },
1070 .vcpu = vcpu,
1071 };
1072
1073 while (mmu_unsync_walk(sp, &walker.walker))
1074 cond_resched_lock(&vcpu->kvm->mmu_lock);
1075 }
1076
1077 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
1078 gfn_t gfn,
1079 gva_t gaddr,
1080 unsigned level,
1081 int metaphysical,
1082 unsigned access,
1083 u64 *parent_pte)
1084 {
1085 union kvm_mmu_page_role role;
1086 unsigned index;
1087 unsigned quadrant;
1088 struct hlist_head *bucket;
1089 struct kvm_mmu_page *sp;
1090 struct hlist_node *node, *tmp;
1091
1092 role.word = 0;
1093 role.glevels = vcpu->arch.mmu.root_level;
1094 role.level = level;
1095 role.metaphysical = metaphysical;
1096 role.access = access;
1097 if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
1098 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
1099 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
1100 role.quadrant = quadrant;
1101 }
1102 pgprintk("%s: looking gfn %lx role %x\n", __func__,
1103 gfn, role.word);
1104 index = kvm_page_table_hashfn(gfn);
1105 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1106 hlist_for_each_entry_safe(sp, node, tmp, bucket, hash_link)
1107 if (sp->gfn == gfn) {
1108 if (sp->unsync)
1109 if (kvm_sync_page(vcpu, sp))
1110 continue;
1111
1112 if (sp->role.word != role.word)
1113 continue;
1114
1115 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
1116 if (sp->unsync_children) {
1117 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
1118 kvm_mmu_mark_parents_unsync(vcpu, sp);
1119 }
1120 pgprintk("%s: found\n", __func__);
1121 return sp;
1122 }
1123 ++vcpu->kvm->stat.mmu_cache_miss;
1124 sp = kvm_mmu_alloc_page(vcpu, parent_pte);
1125 if (!sp)
1126 return sp;
1127 pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
1128 sp->gfn = gfn;
1129 sp->role = role;
1130 hlist_add_head(&sp->hash_link, bucket);
1131 if (!metaphysical) {
1132 rmap_write_protect(vcpu->kvm, gfn);
1133 account_shadowed(vcpu->kvm, gfn);
1134 }
1135 if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
1136 vcpu->arch.mmu.prefetch_page(vcpu, sp);
1137 else
1138 nonpaging_prefetch_page(vcpu, sp);
1139 return sp;
1140 }
1141
1142 static int walk_shadow(struct kvm_shadow_walk *walker,
1143 struct kvm_vcpu *vcpu, u64 addr)
1144 {
1145 hpa_t shadow_addr;
1146 int level;
1147 int r;
1148 u64 *sptep;
1149 unsigned index;
1150
1151 shadow_addr = vcpu->arch.mmu.root_hpa;
1152 level = vcpu->arch.mmu.shadow_root_level;
1153 if (level == PT32E_ROOT_LEVEL) {
1154 shadow_addr = vcpu->arch.mmu.pae_root[(addr >> 30) & 3];
1155 shadow_addr &= PT64_BASE_ADDR_MASK;
1156 --level;
1157 }
1158
1159 while (level >= PT_PAGE_TABLE_LEVEL) {
1160 index = SHADOW_PT_INDEX(addr, level);
1161 sptep = ((u64 *)__va(shadow_addr)) + index;
1162 r = walker->entry(walker, vcpu, addr, sptep, level);
1163 if (r)
1164 return r;
1165 shadow_addr = *sptep & PT64_BASE_ADDR_MASK;
1166 --level;
1167 }
1168 return 0;
1169 }
1170
1171 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
1172 struct kvm_mmu_page *sp)
1173 {
1174 unsigned i;
1175 u64 *pt;
1176 u64 ent;
1177
1178 pt = sp->spt;
1179
1180 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
1181 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1182 if (is_shadow_present_pte(pt[i]))
1183 rmap_remove(kvm, &pt[i]);
1184 pt[i] = shadow_trap_nonpresent_pte;
1185 }
1186 return;
1187 }
1188
1189 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1190 ent = pt[i];
1191
1192 if (is_shadow_present_pte(ent)) {
1193 if (!is_large_pte(ent)) {
1194 ent &= PT64_BASE_ADDR_MASK;
1195 mmu_page_remove_parent_pte(page_header(ent),
1196 &pt[i]);
1197 } else {
1198 --kvm->stat.lpages;
1199 rmap_remove(kvm, &pt[i]);
1200 }
1201 }
1202 pt[i] = shadow_trap_nonpresent_pte;
1203 }
1204 }
1205
1206 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
1207 {
1208 mmu_page_remove_parent_pte(sp, parent_pte);
1209 }
1210
1211 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
1212 {
1213 int i;
1214
1215 for (i = 0; i < KVM_MAX_VCPUS; ++i)
1216 if (kvm->vcpus[i])
1217 kvm->vcpus[i]->arch.last_pte_updated = NULL;
1218 }
1219
1220 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
1221 {
1222 u64 *parent_pte;
1223
1224 while (sp->multimapped || sp->parent_pte) {
1225 if (!sp->multimapped)
1226 parent_pte = sp->parent_pte;
1227 else {
1228 struct kvm_pte_chain *chain;
1229
1230 chain = container_of(sp->parent_ptes.first,
1231 struct kvm_pte_chain, link);
1232 parent_pte = chain->parent_ptes[0];
1233 }
1234 BUG_ON(!parent_pte);
1235 kvm_mmu_put_page(sp, parent_pte);
1236 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
1237 }
1238 }
1239
1240 struct zap_walker {
1241 struct kvm_unsync_walk walker;
1242 struct kvm *kvm;
1243 int zapped;
1244 };
1245
1246 static int mmu_zap_fn(struct kvm_mmu_page *sp, struct kvm_unsync_walk *walk)
1247 {
1248 struct zap_walker *zap_walk = container_of(walk, struct zap_walker,
1249 walker);
1250 kvm_mmu_zap_page(zap_walk->kvm, sp);
1251 zap_walk->zapped = 1;
1252 return 0;
1253 }
1254
1255 static int mmu_zap_unsync_children(struct kvm *kvm, struct kvm_mmu_page *sp)
1256 {
1257 struct zap_walker walker = {
1258 .walker = { .entry = mmu_zap_fn, },
1259 .kvm = kvm,
1260 .zapped = 0,
1261 };
1262
1263 if (sp->role.level == PT_PAGE_TABLE_LEVEL)
1264 return 0;
1265 mmu_unsync_walk(sp, &walker.walker);
1266 return walker.zapped;
1267 }
1268
1269 static int kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1270 {
1271 int ret;
1272 ++kvm->stat.mmu_shadow_zapped;
1273 ret = mmu_zap_unsync_children(kvm, sp);
1274 kvm_mmu_page_unlink_children(kvm, sp);
1275 kvm_mmu_unlink_parents(kvm, sp);
1276 kvm_flush_remote_tlbs(kvm);
1277 if (!sp->role.invalid && !sp->role.metaphysical)
1278 unaccount_shadowed(kvm, sp->gfn);
1279 if (sp->unsync)
1280 kvm_unlink_unsync_page(kvm, sp);
1281 if (!sp->root_count) {
1282 hlist_del(&sp->hash_link);
1283 kvm_mmu_free_page(kvm, sp);
1284 } else {
1285 sp->role.invalid = 1;
1286 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1287 kvm_reload_remote_mmus(kvm);
1288 }
1289 kvm_mmu_reset_last_pte_updated(kvm);
1290 return ret;
1291 }
1292
1293 /*
1294 * Changing the number of mmu pages allocated to the vm
1295 * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1296 */
1297 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1298 {
1299 /*
1300 * If we set the number of mmu pages to be smaller be than the
1301 * number of actived pages , we must to free some mmu pages before we
1302 * change the value
1303 */
1304
1305 if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
1306 kvm_nr_mmu_pages) {
1307 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
1308 - kvm->arch.n_free_mmu_pages;
1309
1310 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
1311 struct kvm_mmu_page *page;
1312
1313 page = container_of(kvm->arch.active_mmu_pages.prev,
1314 struct kvm_mmu_page, link);
1315 kvm_mmu_zap_page(kvm, page);
1316 n_used_mmu_pages--;
1317 }
1318 kvm->arch.n_free_mmu_pages = 0;
1319 }
1320 else
1321 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1322 - kvm->arch.n_alloc_mmu_pages;
1323
1324 kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1325 }
1326
1327 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1328 {
1329 unsigned index;
1330 struct hlist_head *bucket;
1331 struct kvm_mmu_page *sp;
1332 struct hlist_node *node, *n;
1333 int r;
1334
1335 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1336 r = 0;
1337 index = kvm_page_table_hashfn(gfn);
1338 bucket = &kvm->arch.mmu_page_hash[index];
1339 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1340 if (sp->gfn == gfn && !sp->role.metaphysical) {
1341 pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1342 sp->role.word);
1343 r = 1;
1344 if (kvm_mmu_zap_page(kvm, sp))
1345 n = bucket->first;
1346 }
1347 return r;
1348 }
1349
1350 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1351 {
1352 struct kvm_mmu_page *sp;
1353
1354 while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1355 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1356 kvm_mmu_zap_page(kvm, sp);
1357 }
1358 }
1359
1360 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1361 {
1362 int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1363 struct kvm_mmu_page *sp = page_header(__pa(pte));
1364
1365 __set_bit(slot, &sp->slot_bitmap);
1366 }
1367
1368 static void mmu_convert_notrap(struct kvm_mmu_page *sp)
1369 {
1370 int i;
1371 u64 *pt = sp->spt;
1372
1373 if (shadow_trap_nonpresent_pte == shadow_notrap_nonpresent_pte)
1374 return;
1375
1376 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
1377 if (pt[i] == shadow_notrap_nonpresent_pte)
1378 set_shadow_pte(&pt[i], shadow_trap_nonpresent_pte);
1379 }
1380 }
1381
1382 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1383 {
1384 struct page *page;
1385
1386 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1387
1388 if (gpa == UNMAPPED_GVA)
1389 return NULL;
1390
1391 page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1392
1393 return page;
1394 }
1395
1396 /*
1397 * The function is based on mtrr_type_lookup() in
1398 * arch/x86/kernel/cpu/mtrr/generic.c
1399 */
1400 static int get_mtrr_type(struct mtrr_state_type *mtrr_state,
1401 u64 start, u64 end)
1402 {
1403 int i;
1404 u64 base, mask;
1405 u8 prev_match, curr_match;
1406 int num_var_ranges = KVM_NR_VAR_MTRR;
1407
1408 if (!mtrr_state->enabled)
1409 return 0xFF;
1410
1411 /* Make end inclusive end, instead of exclusive */
1412 end--;
1413
1414 /* Look in fixed ranges. Just return the type as per start */
1415 if (mtrr_state->have_fixed && (start < 0x100000)) {
1416 int idx;
1417
1418 if (start < 0x80000) {
1419 idx = 0;
1420 idx += (start >> 16);
1421 return mtrr_state->fixed_ranges[idx];
1422 } else if (start < 0xC0000) {
1423 idx = 1 * 8;
1424 idx += ((start - 0x80000) >> 14);
1425 return mtrr_state->fixed_ranges[idx];
1426 } else if (start < 0x1000000) {
1427 idx = 3 * 8;
1428 idx += ((start - 0xC0000) >> 12);
1429 return mtrr_state->fixed_ranges[idx];
1430 }
1431 }
1432
1433 /*
1434 * Look in variable ranges
1435 * Look of multiple ranges matching this address and pick type
1436 * as per MTRR precedence
1437 */
1438 if (!(mtrr_state->enabled & 2))
1439 return mtrr_state->def_type;
1440
1441 prev_match = 0xFF;
1442 for (i = 0; i < num_var_ranges; ++i) {
1443 unsigned short start_state, end_state;
1444
1445 if (!(mtrr_state->var_ranges[i].mask_lo & (1 << 11)))
1446 continue;
1447
1448 base = (((u64)mtrr_state->var_ranges[i].base_hi) << 32) +
1449 (mtrr_state->var_ranges[i].base_lo & PAGE_MASK);
1450 mask = (((u64)mtrr_state->var_ranges[i].mask_hi) << 32) +
1451 (mtrr_state->var_ranges[i].mask_lo & PAGE_MASK);
1452
1453 start_state = ((start & mask) == (base & mask));
1454 end_state = ((end & mask) == (base & mask));
1455 if (start_state != end_state)
1456 return 0xFE;
1457
1458 if ((start & mask) != (base & mask))
1459 continue;
1460
1461 curr_match = mtrr_state->var_ranges[i].base_lo & 0xff;
1462 if (prev_match == 0xFF) {
1463 prev_match = curr_match;
1464 continue;
1465 }
1466
1467 if (prev_match == MTRR_TYPE_UNCACHABLE ||
1468 curr_match == MTRR_TYPE_UNCACHABLE)
1469 return MTRR_TYPE_UNCACHABLE;
1470
1471 if ((prev_match == MTRR_TYPE_WRBACK &&
1472 curr_match == MTRR_TYPE_WRTHROUGH) ||
1473 (prev_match == MTRR_TYPE_WRTHROUGH &&
1474 curr_match == MTRR_TYPE_WRBACK)) {
1475 prev_match = MTRR_TYPE_WRTHROUGH;
1476 curr_match = MTRR_TYPE_WRTHROUGH;
1477 }
1478
1479 if (prev_match != curr_match)
1480 return MTRR_TYPE_UNCACHABLE;
1481 }
1482
1483 if (prev_match != 0xFF)
1484 return prev_match;
1485
1486 return mtrr_state->def_type;
1487 }
1488
1489 static u8 get_memory_type(struct kvm_vcpu *vcpu, gfn_t gfn)
1490 {
1491 u8 mtrr;
1492
1493 mtrr = get_mtrr_type(&vcpu->arch.mtrr_state, gfn << PAGE_SHIFT,
1494 (gfn << PAGE_SHIFT) + PAGE_SIZE);
1495 if (mtrr == 0xfe || mtrr == 0xff)
1496 mtrr = MTRR_TYPE_WRBACK;
1497 return mtrr;
1498 }
1499
1500 static int kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1501 {
1502 unsigned index;
1503 struct hlist_head *bucket;
1504 struct kvm_mmu_page *s;
1505 struct hlist_node *node, *n;
1506
1507 index = kvm_page_table_hashfn(sp->gfn);
1508 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1509 /* don't unsync if pagetable is shadowed with multiple roles */
1510 hlist_for_each_entry_safe(s, node, n, bucket, hash_link) {
1511 if (s->gfn != sp->gfn || s->role.metaphysical)
1512 continue;
1513 if (s->role.word != sp->role.word)
1514 return 1;
1515 }
1516 kvm_mmu_mark_parents_unsync(vcpu, sp);
1517 ++vcpu->kvm->stat.mmu_unsync;
1518 sp->unsync = 1;
1519 mmu_convert_notrap(sp);
1520 return 0;
1521 }
1522
1523 static int mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
1524 bool can_unsync)
1525 {
1526 struct kvm_mmu_page *shadow;
1527
1528 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1529 if (shadow) {
1530 if (shadow->role.level != PT_PAGE_TABLE_LEVEL)
1531 return 1;
1532 if (shadow->unsync)
1533 return 0;
1534 if (can_unsync && oos_shadow)
1535 return kvm_unsync_page(vcpu, shadow);
1536 return 1;
1537 }
1538 return 0;
1539 }
1540
1541 static int set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1542 unsigned pte_access, int user_fault,
1543 int write_fault, int dirty, int largepage,
1544 gfn_t gfn, pfn_t pfn, bool speculative,
1545 bool can_unsync)
1546 {
1547 u64 spte;
1548 int ret = 0;
1549 /*
1550 * We don't set the accessed bit, since we sometimes want to see
1551 * whether the guest actually used the pte (in order to detect
1552 * demand paging).
1553 */
1554 spte = shadow_base_present_pte | shadow_dirty_mask;
1555 if (!speculative)
1556 spte |= shadow_accessed_mask;
1557 if (!dirty)
1558 pte_access &= ~ACC_WRITE_MASK;
1559 if (pte_access & ACC_EXEC_MASK)
1560 spte |= shadow_x_mask;
1561 else
1562 spte |= shadow_nx_mask;
1563 if (pte_access & ACC_USER_MASK)
1564 spte |= shadow_user_mask;
1565 if (largepage)
1566 spte |= PT_PAGE_SIZE_MASK;
1567
1568 spte |= (u64)pfn << PAGE_SHIFT;
1569
1570 if ((pte_access & ACC_WRITE_MASK)
1571 || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1572
1573 if (largepage && has_wrprotected_page(vcpu->kvm, gfn)) {
1574 ret = 1;
1575 spte = shadow_trap_nonpresent_pte;
1576 goto set_pte;
1577 }
1578
1579 spte |= PT_WRITABLE_MASK;
1580
1581 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
1582 pgprintk("%s: found shadow page for %lx, marking ro\n",
1583 __func__, gfn);
1584 ret = 1;
1585 pte_access &= ~ACC_WRITE_MASK;
1586 if (is_writeble_pte(spte))
1587 spte &= ~PT_WRITABLE_MASK;
1588 }
1589 }
1590
1591 if (pte_access & ACC_WRITE_MASK)
1592 mark_page_dirty(vcpu->kvm, gfn);
1593
1594 set_pte:
1595 set_shadow_pte(shadow_pte, spte);
1596 return ret;
1597 }
1598
1599 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1600 unsigned pt_access, unsigned pte_access,
1601 int user_fault, int write_fault, int dirty,
1602 int *ptwrite, int largepage, gfn_t gfn,
1603 pfn_t pfn, bool speculative)
1604 {
1605 int was_rmapped = 0;
1606 int was_writeble = is_writeble_pte(*shadow_pte);
1607
1608 pgprintk("%s: spte %llx access %x write_fault %d"
1609 " user_fault %d gfn %lx\n",
1610 __func__, *shadow_pte, pt_access,
1611 write_fault, user_fault, gfn);
1612
1613 if (is_rmap_pte(*shadow_pte)) {
1614 /*
1615 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1616 * the parent of the now unreachable PTE.
1617 */
1618 if (largepage && !is_large_pte(*shadow_pte)) {
1619 struct kvm_mmu_page *child;
1620 u64 pte = *shadow_pte;
1621
1622 child = page_header(pte & PT64_BASE_ADDR_MASK);
1623 mmu_page_remove_parent_pte(child, shadow_pte);
1624 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1625 pgprintk("hfn old %lx new %lx\n",
1626 spte_to_pfn(*shadow_pte), pfn);
1627 rmap_remove(vcpu->kvm, shadow_pte);
1628 } else {
1629 if (largepage)
1630 was_rmapped = is_large_pte(*shadow_pte);
1631 else
1632 was_rmapped = 1;
1633 }
1634 }
1635 if (set_spte(vcpu, shadow_pte, pte_access, user_fault, write_fault,
1636 dirty, largepage, gfn, pfn, speculative, true)) {
1637 if (write_fault)
1638 *ptwrite = 1;
1639 kvm_x86_ops->tlb_flush(vcpu);
1640 }
1641
1642 pgprintk("%s: setting spte %llx\n", __func__, *shadow_pte);
1643 pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1644 is_large_pte(*shadow_pte)? "2MB" : "4kB",
1645 is_present_pte(*shadow_pte)?"RW":"R", gfn,
1646 *shadow_pte, shadow_pte);
1647 if (!was_rmapped && is_large_pte(*shadow_pte))
1648 ++vcpu->kvm->stat.lpages;
1649
1650 page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1651 if (!was_rmapped) {
1652 rmap_add(vcpu, shadow_pte, gfn, largepage);
1653 if (!is_rmap_pte(*shadow_pte))
1654 kvm_release_pfn_clean(pfn);
1655 } else {
1656 if (was_writeble)
1657 kvm_release_pfn_dirty(pfn);
1658 else
1659 kvm_release_pfn_clean(pfn);
1660 }
1661 if (speculative) {
1662 vcpu->arch.last_pte_updated = shadow_pte;
1663 vcpu->arch.last_pte_gfn = gfn;
1664 }
1665 }
1666
1667 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1668 {
1669 }
1670
1671 struct direct_shadow_walk {
1672 struct kvm_shadow_walk walker;
1673 pfn_t pfn;
1674 int write;
1675 int largepage;
1676 int pt_write;
1677 };
1678
1679 static int direct_map_entry(struct kvm_shadow_walk *_walk,
1680 struct kvm_vcpu *vcpu,
1681 u64 addr, u64 *sptep, int level)
1682 {
1683 struct direct_shadow_walk *walk =
1684 container_of(_walk, struct direct_shadow_walk, walker);
1685 struct kvm_mmu_page *sp;
1686 gfn_t pseudo_gfn;
1687 gfn_t gfn = addr >> PAGE_SHIFT;
1688
1689 if (level == PT_PAGE_TABLE_LEVEL
1690 || (walk->largepage && level == PT_DIRECTORY_LEVEL)) {
1691 mmu_set_spte(vcpu, sptep, ACC_ALL, ACC_ALL,
1692 0, walk->write, 1, &walk->pt_write,
1693 walk->largepage, gfn, walk->pfn, false);
1694 ++vcpu->stat.pf_fixed;
1695 return 1;
1696 }
1697
1698 if (*sptep == shadow_trap_nonpresent_pte) {
1699 pseudo_gfn = (addr & PT64_DIR_BASE_ADDR_MASK) >> PAGE_SHIFT;
1700 sp = kvm_mmu_get_page(vcpu, pseudo_gfn, (gva_t)addr, level - 1,
1701 1, ACC_ALL, sptep);
1702 if (!sp) {
1703 pgprintk("nonpaging_map: ENOMEM\n");
1704 kvm_release_pfn_clean(walk->pfn);
1705 return -ENOMEM;
1706 }
1707
1708 set_shadow_pte(sptep,
1709 __pa(sp->spt)
1710 | PT_PRESENT_MASK | PT_WRITABLE_MASK
1711 | shadow_user_mask | shadow_x_mask);
1712 }
1713 return 0;
1714 }
1715
1716 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1717 int largepage, gfn_t gfn, pfn_t pfn)
1718 {
1719 int r;
1720 struct direct_shadow_walk walker = {
1721 .walker = { .entry = direct_map_entry, },
1722 .pfn = pfn,
1723 .largepage = largepage,
1724 .write = write,
1725 .pt_write = 0,
1726 };
1727
1728 r = walk_shadow(&walker.walker, vcpu, gfn << PAGE_SHIFT);
1729 if (r < 0)
1730 return r;
1731 return walker.pt_write;
1732 }
1733
1734 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1735 {
1736 int r;
1737 int largepage = 0;
1738 pfn_t pfn;
1739 unsigned long mmu_seq;
1740
1741 if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1742 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1743 largepage = 1;
1744 }
1745
1746 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1747 smp_rmb();
1748 pfn = gfn_to_pfn(vcpu->kvm, gfn);
1749
1750 /* mmio */
1751 if (is_error_pfn(pfn)) {
1752 kvm_release_pfn_clean(pfn);
1753 return 1;
1754 }
1755
1756 spin_lock(&vcpu->kvm->mmu_lock);
1757 if (mmu_notifier_retry(vcpu, mmu_seq))
1758 goto out_unlock;
1759 kvm_mmu_free_some_pages(vcpu);
1760 r = __direct_map(vcpu, v, write, largepage, gfn, pfn);
1761 spin_unlock(&vcpu->kvm->mmu_lock);
1762
1763
1764 return r;
1765
1766 out_unlock:
1767 spin_unlock(&vcpu->kvm->mmu_lock);
1768 kvm_release_pfn_clean(pfn);
1769 return 0;
1770 }
1771
1772
1773 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1774 {
1775 int i;
1776 struct kvm_mmu_page *sp;
1777
1778 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1779 return;
1780 spin_lock(&vcpu->kvm->mmu_lock);
1781 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1782 hpa_t root = vcpu->arch.mmu.root_hpa;
1783
1784 sp = page_header(root);
1785 --sp->root_count;
1786 if (!sp->root_count && sp->role.invalid)
1787 kvm_mmu_zap_page(vcpu->kvm, sp);
1788 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1789 spin_unlock(&vcpu->kvm->mmu_lock);
1790 return;
1791 }
1792 for (i = 0; i < 4; ++i) {
1793 hpa_t root = vcpu->arch.mmu.pae_root[i];
1794
1795 if (root) {
1796 root &= PT64_BASE_ADDR_MASK;
1797 sp = page_header(root);
1798 --sp->root_count;
1799 if (!sp->root_count && sp->role.invalid)
1800 kvm_mmu_zap_page(vcpu->kvm, sp);
1801 }
1802 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1803 }
1804 spin_unlock(&vcpu->kvm->mmu_lock);
1805 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1806 }
1807
1808 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1809 {
1810 int i;
1811 gfn_t root_gfn;
1812 struct kvm_mmu_page *sp;
1813 int metaphysical = 0;
1814
1815 root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1816
1817 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1818 hpa_t root = vcpu->arch.mmu.root_hpa;
1819
1820 ASSERT(!VALID_PAGE(root));
1821 if (tdp_enabled)
1822 metaphysical = 1;
1823 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1824 PT64_ROOT_LEVEL, metaphysical,
1825 ACC_ALL, NULL);
1826 root = __pa(sp->spt);
1827 ++sp->root_count;
1828 vcpu->arch.mmu.root_hpa = root;
1829 return;
1830 }
1831 metaphysical = !is_paging(vcpu);
1832 if (tdp_enabled)
1833 metaphysical = 1;
1834 for (i = 0; i < 4; ++i) {
1835 hpa_t root = vcpu->arch.mmu.pae_root[i];
1836
1837 ASSERT(!VALID_PAGE(root));
1838 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1839 if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1840 vcpu->arch.mmu.pae_root[i] = 0;
1841 continue;
1842 }
1843 root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1844 } else if (vcpu->arch.mmu.root_level == 0)
1845 root_gfn = 0;
1846 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1847 PT32_ROOT_LEVEL, metaphysical,
1848 ACC_ALL, NULL);
1849 root = __pa(sp->spt);
1850 ++sp->root_count;
1851 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1852 }
1853 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1854 }
1855
1856 static void mmu_sync_roots(struct kvm_vcpu *vcpu)
1857 {
1858 int i;
1859 struct kvm_mmu_page *sp;
1860
1861 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1862 return;
1863 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1864 hpa_t root = vcpu->arch.mmu.root_hpa;
1865 sp = page_header(root);
1866 mmu_sync_children(vcpu, sp);
1867 return;
1868 }
1869 for (i = 0; i < 4; ++i) {
1870 hpa_t root = vcpu->arch.mmu.pae_root[i];
1871
1872 if (root) {
1873 root &= PT64_BASE_ADDR_MASK;
1874 sp = page_header(root);
1875 mmu_sync_children(vcpu, sp);
1876 }
1877 }
1878 }
1879
1880 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
1881 {
1882 spin_lock(&vcpu->kvm->mmu_lock);
1883 mmu_sync_roots(vcpu);
1884 spin_unlock(&vcpu->kvm->mmu_lock);
1885 }
1886
1887 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1888 {
1889 return vaddr;
1890 }
1891
1892 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1893 u32 error_code)
1894 {
1895 gfn_t gfn;
1896 int r;
1897
1898 pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
1899 r = mmu_topup_memory_caches(vcpu);
1900 if (r)
1901 return r;
1902
1903 ASSERT(vcpu);
1904 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1905
1906 gfn = gva >> PAGE_SHIFT;
1907
1908 return nonpaging_map(vcpu, gva & PAGE_MASK,
1909 error_code & PFERR_WRITE_MASK, gfn);
1910 }
1911
1912 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1913 u32 error_code)
1914 {
1915 pfn_t pfn;
1916 int r;
1917 int largepage = 0;
1918 gfn_t gfn = gpa >> PAGE_SHIFT;
1919 unsigned long mmu_seq;
1920
1921 ASSERT(vcpu);
1922 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1923
1924 r = mmu_topup_memory_caches(vcpu);
1925 if (r)
1926 return r;
1927
1928 if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1929 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1930 largepage = 1;
1931 }
1932 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1933 smp_rmb();
1934 pfn = gfn_to_pfn(vcpu->kvm, gfn);
1935 if (is_error_pfn(pfn)) {
1936 kvm_release_pfn_clean(pfn);
1937 return 1;
1938 }
1939 spin_lock(&vcpu->kvm->mmu_lock);
1940 if (mmu_notifier_retry(vcpu, mmu_seq))
1941 goto out_unlock;
1942 kvm_mmu_free_some_pages(vcpu);
1943 r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1944 largepage, gfn, pfn);
1945 spin_unlock(&vcpu->kvm->mmu_lock);
1946
1947 return r;
1948
1949 out_unlock:
1950 spin_unlock(&vcpu->kvm->mmu_lock);
1951 kvm_release_pfn_clean(pfn);
1952 return 0;
1953 }
1954
1955 static void nonpaging_free(struct kvm_vcpu *vcpu)
1956 {
1957 mmu_free_roots(vcpu);
1958 }
1959
1960 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1961 {
1962 struct kvm_mmu *context = &vcpu->arch.mmu;
1963
1964 context->new_cr3 = nonpaging_new_cr3;
1965 context->page_fault = nonpaging_page_fault;
1966 context->gva_to_gpa = nonpaging_gva_to_gpa;
1967 context->free = nonpaging_free;
1968 context->prefetch_page = nonpaging_prefetch_page;
1969 context->sync_page = nonpaging_sync_page;
1970 context->invlpg = nonpaging_invlpg;
1971 context->root_level = 0;
1972 context->shadow_root_level = PT32E_ROOT_LEVEL;
1973 context->root_hpa = INVALID_PAGE;
1974 return 0;
1975 }
1976
1977 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1978 {
1979 ++vcpu->stat.tlb_flush;
1980 kvm_x86_ops->tlb_flush(vcpu);
1981 }
1982
1983 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1984 {
1985 pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
1986 mmu_free_roots(vcpu);
1987 }
1988
1989 static void inject_page_fault(struct kvm_vcpu *vcpu,
1990 u64 addr,
1991 u32 err_code)
1992 {
1993 kvm_inject_page_fault(vcpu, addr, err_code);
1994 }
1995
1996 static void paging_free(struct kvm_vcpu *vcpu)
1997 {
1998 nonpaging_free(vcpu);
1999 }
2000
2001 #define PTTYPE 64
2002 #include "paging_tmpl.h"
2003 #undef PTTYPE
2004
2005 #define PTTYPE 32
2006 #include "paging_tmpl.h"
2007 #undef PTTYPE
2008
2009 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
2010 {
2011 struct kvm_mmu *context = &vcpu->arch.mmu;
2012
2013 ASSERT(is_pae(vcpu));
2014 context->new_cr3 = paging_new_cr3;
2015 context->page_fault = paging64_page_fault;
2016 context->gva_to_gpa = paging64_gva_to_gpa;
2017 context->prefetch_page = paging64_prefetch_page;
2018 context->sync_page = paging64_sync_page;
2019 context->invlpg = paging64_invlpg;
2020 context->free = paging_free;
2021 context->root_level = level;
2022 context->shadow_root_level = level;
2023 context->root_hpa = INVALID_PAGE;
2024 return 0;
2025 }
2026
2027 static int paging64_init_context(struct kvm_vcpu *vcpu)
2028 {
2029 return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
2030 }
2031
2032 static int paging32_init_context(struct kvm_vcpu *vcpu)
2033 {
2034 struct kvm_mmu *context = &vcpu->arch.mmu;
2035
2036 context->new_cr3 = paging_new_cr3;
2037 context->page_fault = paging32_page_fault;
2038 context->gva_to_gpa = paging32_gva_to_gpa;
2039 context->free = paging_free;
2040 context->prefetch_page = paging32_prefetch_page;
2041 context->sync_page = paging32_sync_page;
2042 context->invlpg = paging32_invlpg;
2043 context->root_level = PT32_ROOT_LEVEL;
2044 context->shadow_root_level = PT32E_ROOT_LEVEL;
2045 context->root_hpa = INVALID_PAGE;
2046 return 0;
2047 }
2048
2049 static int paging32E_init_context(struct kvm_vcpu *vcpu)
2050 {
2051 return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
2052 }
2053
2054 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
2055 {
2056 struct kvm_mmu *context = &vcpu->arch.mmu;
2057
2058 context->new_cr3 = nonpaging_new_cr3;
2059 context->page_fault = tdp_page_fault;
2060 context->free = nonpaging_free;
2061 context->prefetch_page = nonpaging_prefetch_page;
2062 context->sync_page = nonpaging_sync_page;
2063 context->invlpg = nonpaging_invlpg;
2064 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
2065 context->root_hpa = INVALID_PAGE;
2066
2067 if (!is_paging(vcpu)) {
2068 context->gva_to_gpa = nonpaging_gva_to_gpa;
2069 context->root_level = 0;
2070 } else if (is_long_mode(vcpu)) {
2071 context->gva_to_gpa = paging64_gva_to_gpa;
2072 context->root_level = PT64_ROOT_LEVEL;
2073 } else if (is_pae(vcpu)) {
2074 context->gva_to_gpa = paging64_gva_to_gpa;
2075 context->root_level = PT32E_ROOT_LEVEL;
2076 } else {
2077 context->gva_to_gpa = paging32_gva_to_gpa;
2078 context->root_level = PT32_ROOT_LEVEL;
2079 }
2080
2081 return 0;
2082 }
2083
2084 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
2085 {
2086 ASSERT(vcpu);
2087 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2088
2089 if (!is_paging(vcpu))
2090 return nonpaging_init_context(vcpu);
2091 else if (is_long_mode(vcpu))
2092 return paging64_init_context(vcpu);
2093 else if (is_pae(vcpu))
2094 return paging32E_init_context(vcpu);
2095 else
2096 return paging32_init_context(vcpu);
2097 }
2098
2099 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
2100 {
2101 vcpu->arch.update_pte.pfn = bad_pfn;
2102
2103 if (tdp_enabled)
2104 return init_kvm_tdp_mmu(vcpu);
2105 else
2106 return init_kvm_softmmu(vcpu);
2107 }
2108
2109 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
2110 {
2111 ASSERT(vcpu);
2112 if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
2113 vcpu->arch.mmu.free(vcpu);
2114 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
2115 }
2116 }
2117
2118 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
2119 {
2120 destroy_kvm_mmu(vcpu);
2121 return init_kvm_mmu(vcpu);
2122 }
2123 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
2124
2125 int kvm_mmu_load(struct kvm_vcpu *vcpu)
2126 {
2127 int r;
2128
2129 r = mmu_topup_memory_caches(vcpu);
2130 if (r)
2131 goto out;
2132 spin_lock(&vcpu->kvm->mmu_lock);
2133 kvm_mmu_free_some_pages(vcpu);
2134 mmu_alloc_roots(vcpu);
2135 mmu_sync_roots(vcpu);
2136 spin_unlock(&vcpu->kvm->mmu_lock);
2137 kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
2138 kvm_mmu_flush_tlb(vcpu);
2139 out:
2140 return r;
2141 }
2142 EXPORT_SYMBOL_GPL(kvm_mmu_load);
2143
2144 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
2145 {
2146 mmu_free_roots(vcpu);
2147 }
2148
2149 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
2150 struct kvm_mmu_page *sp,
2151 u64 *spte)
2152 {
2153 u64 pte;
2154 struct kvm_mmu_page *child;
2155
2156 pte = *spte;
2157 if (is_shadow_present_pte(pte)) {
2158 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
2159 is_large_pte(pte))
2160 rmap_remove(vcpu->kvm, spte);
2161 else {
2162 child = page_header(pte & PT64_BASE_ADDR_MASK);
2163 mmu_page_remove_parent_pte(child, spte);
2164 }
2165 }
2166 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
2167 if (is_large_pte(pte))
2168 --vcpu->kvm->stat.lpages;
2169 }
2170
2171 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
2172 struct kvm_mmu_page *sp,
2173 u64 *spte,
2174 const void *new)
2175 {
2176 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
2177 if (!vcpu->arch.update_pte.largepage ||
2178 sp->role.glevels == PT32_ROOT_LEVEL) {
2179 ++vcpu->kvm->stat.mmu_pde_zapped;
2180 return;
2181 }
2182 }
2183
2184 ++vcpu->kvm->stat.mmu_pte_updated;
2185 if (sp->role.glevels == PT32_ROOT_LEVEL)
2186 paging32_update_pte(vcpu, sp, spte, new);
2187 else
2188 paging64_update_pte(vcpu, sp, spte, new);
2189 }
2190
2191 static bool need_remote_flush(u64 old, u64 new)
2192 {
2193 if (!is_shadow_present_pte(old))
2194 return false;
2195 if (!is_shadow_present_pte(new))
2196 return true;
2197 if ((old ^ new) & PT64_BASE_ADDR_MASK)
2198 return true;
2199 old ^= PT64_NX_MASK;
2200 new ^= PT64_NX_MASK;
2201 return (old & ~new & PT64_PERM_MASK) != 0;
2202 }
2203
2204 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
2205 {
2206 if (need_remote_flush(old, new))
2207 kvm_flush_remote_tlbs(vcpu->kvm);
2208 else
2209 kvm_mmu_flush_tlb(vcpu);
2210 }
2211
2212 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
2213 {
2214 u64 *spte = vcpu->arch.last_pte_updated;
2215
2216 return !!(spte && (*spte & shadow_accessed_mask));
2217 }
2218
2219 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2220 const u8 *new, int bytes)
2221 {
2222 gfn_t gfn;
2223 int r;
2224 u64 gpte = 0;
2225 pfn_t pfn;
2226
2227 vcpu->arch.update_pte.largepage = 0;
2228
2229 if (bytes != 4 && bytes != 8)
2230 return;
2231
2232 /*
2233 * Assume that the pte write on a page table of the same type
2234 * as the current vcpu paging mode. This is nearly always true
2235 * (might be false while changing modes). Note it is verified later
2236 * by update_pte().
2237 */
2238 if (is_pae(vcpu)) {
2239 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2240 if ((bytes == 4) && (gpa % 4 == 0)) {
2241 r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
2242 if (r)
2243 return;
2244 memcpy((void *)&gpte + (gpa % 8), new, 4);
2245 } else if ((bytes == 8) && (gpa % 8 == 0)) {
2246 memcpy((void *)&gpte, new, 8);
2247 }
2248 } else {
2249 if ((bytes == 4) && (gpa % 4 == 0))
2250 memcpy((void *)&gpte, new, 4);
2251 }
2252 if (!is_present_pte(gpte))
2253 return;
2254 gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
2255
2256 if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
2257 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
2258 vcpu->arch.update_pte.largepage = 1;
2259 }
2260 vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
2261 smp_rmb();
2262 pfn = gfn_to_pfn(vcpu->kvm, gfn);
2263
2264 if (is_error_pfn(pfn)) {
2265 kvm_release_pfn_clean(pfn);
2266 return;
2267 }
2268 vcpu->arch.update_pte.gfn = gfn;
2269 vcpu->arch.update_pte.pfn = pfn;
2270 }
2271
2272 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
2273 {
2274 u64 *spte = vcpu->arch.last_pte_updated;
2275
2276 if (spte
2277 && vcpu->arch.last_pte_gfn == gfn
2278 && shadow_accessed_mask
2279 && !(*spte & shadow_accessed_mask)
2280 && is_shadow_present_pte(*spte))
2281 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
2282 }
2283
2284 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
2285 const u8 *new, int bytes)
2286 {
2287 gfn_t gfn = gpa >> PAGE_SHIFT;
2288 struct kvm_mmu_page *sp;
2289 struct hlist_node *node, *n;
2290 struct hlist_head *bucket;
2291 unsigned index;
2292 u64 entry, gentry;
2293 u64 *spte;
2294 unsigned offset = offset_in_page(gpa);
2295 unsigned pte_size;
2296 unsigned page_offset;
2297 unsigned misaligned;
2298 unsigned quadrant;
2299 int level;
2300 int flooded = 0;
2301 int npte;
2302 int r;
2303
2304 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
2305 mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
2306 spin_lock(&vcpu->kvm->mmu_lock);
2307 kvm_mmu_access_page(vcpu, gfn);
2308 kvm_mmu_free_some_pages(vcpu);
2309 ++vcpu->kvm->stat.mmu_pte_write;
2310 kvm_mmu_audit(vcpu, "pre pte write");
2311 if (gfn == vcpu->arch.last_pt_write_gfn
2312 && !last_updated_pte_accessed(vcpu)) {
2313 ++vcpu->arch.last_pt_write_count;
2314 if (vcpu->arch.last_pt_write_count >= 3)
2315 flooded = 1;
2316 } else {
2317 vcpu->arch.last_pt_write_gfn = gfn;
2318 vcpu->arch.last_pt_write_count = 1;
2319 vcpu->arch.last_pte_updated = NULL;
2320 }
2321 index = kvm_page_table_hashfn(gfn);
2322 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
2323 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
2324 if (sp->gfn != gfn || sp->role.metaphysical || sp->role.invalid)
2325 continue;
2326 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
2327 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
2328 misaligned |= bytes < 4;
2329 if (misaligned || flooded) {
2330 /*
2331 * Misaligned accesses are too much trouble to fix
2332 * up; also, they usually indicate a page is not used
2333 * as a page table.
2334 *
2335 * If we're seeing too many writes to a page,
2336 * it may no longer be a page table, or we may be
2337 * forking, in which case it is better to unmap the
2338 * page.
2339 */
2340 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2341 gpa, bytes, sp->role.word);
2342 if (kvm_mmu_zap_page(vcpu->kvm, sp))
2343 n = bucket->first;
2344 ++vcpu->kvm->stat.mmu_flooded;
2345 continue;
2346 }
2347 page_offset = offset;
2348 level = sp->role.level;
2349 npte = 1;
2350 if (sp->role.glevels == PT32_ROOT_LEVEL) {
2351 page_offset <<= 1; /* 32->64 */
2352 /*
2353 * A 32-bit pde maps 4MB while the shadow pdes map
2354 * only 2MB. So we need to double the offset again
2355 * and zap two pdes instead of one.
2356 */
2357 if (level == PT32_ROOT_LEVEL) {
2358 page_offset &= ~7; /* kill rounding error */
2359 page_offset <<= 1;
2360 npte = 2;
2361 }
2362 quadrant = page_offset >> PAGE_SHIFT;
2363 page_offset &= ~PAGE_MASK;
2364 if (quadrant != sp->role.quadrant)
2365 continue;
2366 }
2367 spte = &sp->spt[page_offset / sizeof(*spte)];
2368 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
2369 gentry = 0;
2370 r = kvm_read_guest_atomic(vcpu->kvm,
2371 gpa & ~(u64)(pte_size - 1),
2372 &gentry, pte_size);
2373 new = (const void *)&gentry;
2374 if (r < 0)
2375 new = NULL;
2376 }
2377 while (npte--) {
2378 entry = *spte;
2379 mmu_pte_write_zap_pte(vcpu, sp, spte);
2380 if (new)
2381 mmu_pte_write_new_pte(vcpu, sp, spte, new);
2382 mmu_pte_write_flush_tlb(vcpu, entry, *spte);
2383 ++spte;
2384 }
2385 }
2386 kvm_mmu_audit(vcpu, "post pte write");
2387 spin_unlock(&vcpu->kvm->mmu_lock);
2388 if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
2389 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
2390 vcpu->arch.update_pte.pfn = bad_pfn;
2391 }
2392 }
2393
2394 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2395 {
2396 gpa_t gpa;
2397 int r;
2398
2399 gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
2400
2401 spin_lock(&vcpu->kvm->mmu_lock);
2402 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2403 spin_unlock(&vcpu->kvm->mmu_lock);
2404 return r;
2405 }
2406 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
2407
2408 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
2409 {
2410 while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
2411 struct kvm_mmu_page *sp;
2412
2413 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
2414 struct kvm_mmu_page, link);
2415 kvm_mmu_zap_page(vcpu->kvm, sp);
2416 ++vcpu->kvm->stat.mmu_recycled;
2417 }
2418 }
2419
2420 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
2421 {
2422 int r;
2423 enum emulation_result er;
2424
2425 r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
2426 if (r < 0)
2427 goto out;
2428
2429 if (!r) {
2430 r = 1;
2431 goto out;
2432 }
2433
2434 r = mmu_topup_memory_caches(vcpu);
2435 if (r)
2436 goto out;
2437
2438 er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
2439
2440 switch (er) {
2441 case EMULATE_DONE:
2442 return 1;
2443 case EMULATE_DO_MMIO:
2444 ++vcpu->stat.mmio_exits;
2445 return 0;
2446 case EMULATE_FAIL:
2447 kvm_report_emulation_failure(vcpu, "pagetable");
2448 return 1;
2449 default:
2450 BUG();
2451 }
2452 out:
2453 return r;
2454 }
2455 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
2456
2457 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
2458 {
2459 spin_lock(&vcpu->kvm->mmu_lock);
2460 vcpu->arch.mmu.invlpg(vcpu, gva);
2461 spin_unlock(&vcpu->kvm->mmu_lock);
2462 kvm_mmu_flush_tlb(vcpu);
2463 ++vcpu->stat.invlpg;
2464 }
2465 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
2466
2467 void kvm_enable_tdp(void)
2468 {
2469 tdp_enabled = true;
2470 }
2471 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
2472
2473 void kvm_disable_tdp(void)
2474 {
2475 tdp_enabled = false;
2476 }
2477 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
2478
2479 static void free_mmu_pages(struct kvm_vcpu *vcpu)
2480 {
2481 struct kvm_mmu_page *sp;
2482
2483 while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
2484 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
2485 struct kvm_mmu_page, link);
2486 kvm_mmu_zap_page(vcpu->kvm, sp);
2487 cond_resched();
2488 }
2489 free_page((unsigned long)vcpu->arch.mmu.pae_root);
2490 }
2491
2492 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
2493 {
2494 struct page *page;
2495 int i;
2496
2497 ASSERT(vcpu);
2498
2499 if (vcpu->kvm->arch.n_requested_mmu_pages)
2500 vcpu->kvm->arch.n_free_mmu_pages =
2501 vcpu->kvm->arch.n_requested_mmu_pages;
2502 else
2503 vcpu->kvm->arch.n_free_mmu_pages =
2504 vcpu->kvm->arch.n_alloc_mmu_pages;
2505 /*
2506 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2507 * Therefore we need to allocate shadow page tables in the first
2508 * 4GB of memory, which happens to fit the DMA32 zone.
2509 */
2510 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2511 if (!page)
2512 goto error_1;
2513 vcpu->arch.mmu.pae_root = page_address(page);
2514 for (i = 0; i < 4; ++i)
2515 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2516
2517 return 0;
2518
2519 error_1:
2520 free_mmu_pages(vcpu);
2521 return -ENOMEM;
2522 }
2523
2524 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2525 {
2526 ASSERT(vcpu);
2527 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2528
2529 return alloc_mmu_pages(vcpu);
2530 }
2531
2532 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2533 {
2534 ASSERT(vcpu);
2535 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2536
2537 return init_kvm_mmu(vcpu);
2538 }
2539
2540 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2541 {
2542 ASSERT(vcpu);
2543
2544 destroy_kvm_mmu(vcpu);
2545 free_mmu_pages(vcpu);
2546 mmu_free_memory_caches(vcpu);
2547 }
2548
2549 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2550 {
2551 struct kvm_mmu_page *sp;
2552
2553 spin_lock(&kvm->mmu_lock);
2554 list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2555 int i;
2556 u64 *pt;
2557
2558 if (!test_bit(slot, &sp->slot_bitmap))
2559 continue;
2560
2561 pt = sp->spt;
2562 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2563 /* avoid RMW */
2564 if (pt[i] & PT_WRITABLE_MASK)
2565 pt[i] &= ~PT_WRITABLE_MASK;
2566 }
2567 kvm_flush_remote_tlbs(kvm);
2568 spin_unlock(&kvm->mmu_lock);
2569 }
2570
2571 void kvm_mmu_zap_all(struct kvm *kvm)
2572 {
2573 struct kvm_mmu_page *sp, *node;
2574
2575 spin_lock(&kvm->mmu_lock);
2576 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2577 if (kvm_mmu_zap_page(kvm, sp))
2578 node = container_of(kvm->arch.active_mmu_pages.next,
2579 struct kvm_mmu_page, link);
2580 spin_unlock(&kvm->mmu_lock);
2581
2582 kvm_flush_remote_tlbs(kvm);
2583 }
2584
2585 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2586 {
2587 struct kvm_mmu_page *page;
2588
2589 page = container_of(kvm->arch.active_mmu_pages.prev,
2590 struct kvm_mmu_page, link);
2591 kvm_mmu_zap_page(kvm, page);
2592 }
2593
2594 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2595 {
2596 struct kvm *kvm;
2597 struct kvm *kvm_freed = NULL;
2598 int cache_count = 0;
2599
2600 spin_lock(&kvm_lock);
2601
2602 list_for_each_entry(kvm, &vm_list, vm_list) {
2603 int npages;
2604
2605 if (!down_read_trylock(&kvm->slots_lock))
2606 continue;
2607 spin_lock(&kvm->mmu_lock);
2608 npages = kvm->arch.n_alloc_mmu_pages -
2609 kvm->arch.n_free_mmu_pages;
2610 cache_count += npages;
2611 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2612 kvm_mmu_remove_one_alloc_mmu_page(kvm);
2613 cache_count--;
2614 kvm_freed = kvm;
2615 }
2616 nr_to_scan--;
2617
2618 spin_unlock(&kvm->mmu_lock);
2619 up_read(&kvm->slots_lock);
2620 }
2621 if (kvm_freed)
2622 list_move_tail(&kvm_freed->vm_list, &vm_list);
2623
2624 spin_unlock(&kvm_lock);
2625
2626 return cache_count;
2627 }
2628
2629 static struct shrinker mmu_shrinker = {
2630 .shrink = mmu_shrink,
2631 .seeks = DEFAULT_SEEKS * 10,
2632 };
2633
2634 static void mmu_destroy_caches(void)
2635 {
2636 if (pte_chain_cache)
2637 kmem_cache_destroy(pte_chain_cache);
2638 if (rmap_desc_cache)
2639 kmem_cache_destroy(rmap_desc_cache);
2640 if (mmu_page_header_cache)
2641 kmem_cache_destroy(mmu_page_header_cache);
2642 }
2643
2644 void kvm_mmu_module_exit(void)
2645 {
2646 mmu_destroy_caches();
2647 unregister_shrinker(&mmu_shrinker);
2648 }
2649
2650 int kvm_mmu_module_init(void)
2651 {
2652 pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2653 sizeof(struct kvm_pte_chain),
2654 0, 0, NULL);
2655 if (!pte_chain_cache)
2656 goto nomem;
2657 rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2658 sizeof(struct kvm_rmap_desc),
2659 0, 0, NULL);
2660 if (!rmap_desc_cache)
2661 goto nomem;
2662
2663 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2664 sizeof(struct kvm_mmu_page),
2665 0, 0, NULL);
2666 if (!mmu_page_header_cache)
2667 goto nomem;
2668
2669 register_shrinker(&mmu_shrinker);
2670
2671 return 0;
2672
2673 nomem:
2674 mmu_destroy_caches();
2675 return -ENOMEM;
2676 }
2677
2678 /*
2679 * Caculate mmu pages needed for kvm.
2680 */
2681 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2682 {
2683 int i;
2684 unsigned int nr_mmu_pages;
2685 unsigned int nr_pages = 0;
2686
2687 for (i = 0; i < kvm->nmemslots; i++)
2688 nr_pages += kvm->memslots[i].npages;
2689
2690 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2691 nr_mmu_pages = max(nr_mmu_pages,
2692 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2693
2694 return nr_mmu_pages;
2695 }
2696
2697 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2698 unsigned len)
2699 {
2700 if (len > buffer->len)
2701 return NULL;
2702 return buffer->ptr;
2703 }
2704
2705 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2706 unsigned len)
2707 {
2708 void *ret;
2709
2710 ret = pv_mmu_peek_buffer(buffer, len);
2711 if (!ret)
2712 return ret;
2713 buffer->ptr += len;
2714 buffer->len -= len;
2715 buffer->processed += len;
2716 return ret;
2717 }
2718
2719 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2720 gpa_t addr, gpa_t value)
2721 {
2722 int bytes = 8;
2723 int r;
2724
2725 if (!is_long_mode(vcpu) && !is_pae(vcpu))
2726 bytes = 4;
2727
2728 r = mmu_topup_memory_caches(vcpu);
2729 if (r)
2730 return r;
2731
2732 if (!emulator_write_phys(vcpu, addr, &value, bytes))
2733 return -EFAULT;
2734
2735 return 1;
2736 }
2737
2738 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2739 {
2740 kvm_x86_ops->tlb_flush(vcpu);
2741 set_bit(KVM_REQ_MMU_SYNC, &vcpu->requests);
2742 return 1;
2743 }
2744
2745 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2746 {
2747 spin_lock(&vcpu->kvm->mmu_lock);
2748 mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2749 spin_unlock(&vcpu->kvm->mmu_lock);
2750 return 1;
2751 }
2752
2753 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2754 struct kvm_pv_mmu_op_buffer *buffer)
2755 {
2756 struct kvm_mmu_op_header *header;
2757
2758 header = pv_mmu_peek_buffer(buffer, sizeof *header);
2759 if (!header)
2760 return 0;
2761 switch (header->op) {
2762 case KVM_MMU_OP_WRITE_PTE: {
2763 struct kvm_mmu_op_write_pte *wpte;
2764
2765 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2766 if (!wpte)
2767 return 0;
2768 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2769 wpte->pte_val);
2770 }
2771 case KVM_MMU_OP_FLUSH_TLB: {
2772 struct kvm_mmu_op_flush_tlb *ftlb;
2773
2774 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2775 if (!ftlb)
2776 return 0;
2777 return kvm_pv_mmu_flush_tlb(vcpu);
2778 }
2779 case KVM_MMU_OP_RELEASE_PT: {
2780 struct kvm_mmu_op_release_pt *rpt;
2781
2782 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2783 if (!rpt)
2784 return 0;
2785 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2786 }
2787 default: return 0;
2788 }
2789 }
2790
2791 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2792 gpa_t addr, unsigned long *ret)
2793 {
2794 int r;
2795 struct kvm_pv_mmu_op_buffer *buffer = &vcpu->arch.mmu_op_buffer;
2796
2797 buffer->ptr = buffer->buf;
2798 buffer->len = min_t(unsigned long, bytes, sizeof buffer->buf);
2799 buffer->processed = 0;
2800
2801 r = kvm_read_guest(vcpu->kvm, addr, buffer->buf, buffer->len);
2802 if (r)
2803 goto out;
2804
2805 while (buffer->len) {
2806 r = kvm_pv_mmu_op_one(vcpu, buffer);
2807 if (r < 0)
2808 goto out;
2809 if (r == 0)
2810 break;
2811 }
2812
2813 r = 1;
2814 out:
2815 *ret = buffer->processed;
2816 return r;
2817 }
2818
2819 #ifdef AUDIT
2820
2821 static const char *audit_msg;
2822
2823 static gva_t canonicalize(gva_t gva)
2824 {
2825 #ifdef CONFIG_X86_64
2826 gva = (long long)(gva << 16) >> 16;
2827 #endif
2828 return gva;
2829 }
2830
2831 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2832 gva_t va, int level)
2833 {
2834 u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2835 int i;
2836 gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2837
2838 for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2839 u64 ent = pt[i];
2840
2841 if (ent == shadow_trap_nonpresent_pte)
2842 continue;
2843
2844 va = canonicalize(va);
2845 if (level > 1) {
2846 if (ent == shadow_notrap_nonpresent_pte)
2847 printk(KERN_ERR "audit: (%s) nontrapping pte"
2848 " in nonleaf level: levels %d gva %lx"
2849 " level %d pte %llx\n", audit_msg,
2850 vcpu->arch.mmu.root_level, va, level, ent);
2851
2852 audit_mappings_page(vcpu, ent, va, level - 1);
2853 } else {
2854 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2855 hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2856
2857 if (is_shadow_present_pte(ent)
2858 && (ent & PT64_BASE_ADDR_MASK) != hpa)
2859 printk(KERN_ERR "xx audit error: (%s) levels %d"
2860 " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2861 audit_msg, vcpu->arch.mmu.root_level,
2862 va, gpa, hpa, ent,
2863 is_shadow_present_pte(ent));
2864 else if (ent == shadow_notrap_nonpresent_pte
2865 && !is_error_hpa(hpa))
2866 printk(KERN_ERR "audit: (%s) notrap shadow,"
2867 " valid guest gva %lx\n", audit_msg, va);
2868 kvm_release_pfn_clean(pfn);
2869
2870 }
2871 }
2872 }
2873
2874 static void audit_mappings(struct kvm_vcpu *vcpu)
2875 {
2876 unsigned i;
2877
2878 if (vcpu->arch.mmu.root_level == 4)
2879 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2880 else
2881 for (i = 0; i < 4; ++i)
2882 if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2883 audit_mappings_page(vcpu,
2884 vcpu->arch.mmu.pae_root[i],
2885 i << 30,
2886 2);
2887 }
2888
2889 static int count_rmaps(struct kvm_vcpu *vcpu)
2890 {
2891 int nmaps = 0;
2892 int i, j, k;
2893
2894 for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
2895 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
2896 struct kvm_rmap_desc *d;
2897
2898 for (j = 0; j < m->npages; ++j) {
2899 unsigned long *rmapp = &m->rmap[j];
2900
2901 if (!*rmapp)
2902 continue;
2903 if (!(*rmapp & 1)) {
2904 ++nmaps;
2905 continue;
2906 }
2907 d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
2908 while (d) {
2909 for (k = 0; k < RMAP_EXT; ++k)
2910 if (d->shadow_ptes[k])
2911 ++nmaps;
2912 else
2913 break;
2914 d = d->more;
2915 }
2916 }
2917 }
2918 return nmaps;
2919 }
2920
2921 static int count_writable_mappings(struct kvm_vcpu *vcpu)
2922 {
2923 int nmaps = 0;
2924 struct kvm_mmu_page *sp;
2925 int i;
2926
2927 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2928 u64 *pt = sp->spt;
2929
2930 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
2931 continue;
2932
2933 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
2934 u64 ent = pt[i];
2935
2936 if (!(ent & PT_PRESENT_MASK))
2937 continue;
2938 if (!(ent & PT_WRITABLE_MASK))
2939 continue;
2940 ++nmaps;
2941 }
2942 }
2943 return nmaps;
2944 }
2945
2946 static void audit_rmap(struct kvm_vcpu *vcpu)
2947 {
2948 int n_rmap = count_rmaps(vcpu);
2949 int n_actual = count_writable_mappings(vcpu);
2950
2951 if (n_rmap != n_actual)
2952 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
2953 __func__, audit_msg, n_rmap, n_actual);
2954 }
2955
2956 static void audit_write_protection(struct kvm_vcpu *vcpu)
2957 {
2958 struct kvm_mmu_page *sp;
2959 struct kvm_memory_slot *slot;
2960 unsigned long *rmapp;
2961 gfn_t gfn;
2962
2963 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2964 if (sp->role.metaphysical)
2965 continue;
2966
2967 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
2968 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
2969 rmapp = &slot->rmap[gfn - slot->base_gfn];
2970 if (*rmapp)
2971 printk(KERN_ERR "%s: (%s) shadow page has writable"
2972 " mappings: gfn %lx role %x\n",
2973 __func__, audit_msg, sp->gfn,
2974 sp->role.word);
2975 }
2976 }
2977
2978 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
2979 {
2980 int olddbg = dbg;
2981
2982 dbg = 0;
2983 audit_msg = msg;
2984 audit_rmap(vcpu);
2985 audit_write_protection(vcpu);
2986 audit_mappings(vcpu);
2987 dbg = olddbg;
2988 }
2989
2990 #endif
Support for the Chinese Samsung S3C2440 based development boards