projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
KVM: Reduce stack usage in kvm_arch_vcpu_ioctl()
[deliverable/linux.git] / arch / x86 / kvm / 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 #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 #ifndef MMU_DEBUG
74 #define ASSERT(x) do { } while (0)
75 #else
76 #define ASSERT(x) \
77 if (!(x)) { \
78 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
79 __FILE__, __LINE__, #x); \
80 }
81 #endif
82
83 #define PT_FIRST_AVAIL_BITS_SHIFT 9
84 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
85
86 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
87
88 #define PT64_LEVEL_BITS 9
89
90 #define PT64_LEVEL_SHIFT(level) \
91 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
92
93 #define PT64_LEVEL_MASK(level) \
94 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
95
96 #define PT64_INDEX(address, level)\
97 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
98
99
100 #define PT32_LEVEL_BITS 10
101
102 #define PT32_LEVEL_SHIFT(level) \
103 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
104
105 #define PT32_LEVEL_MASK(level) \
106 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
107
108 #define PT32_INDEX(address, level)\
109 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
110
111
112 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
113 #define PT64_DIR_BASE_ADDR_MASK \
114 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
115
116 #define PT32_BASE_ADDR_MASK PAGE_MASK
117 #define PT32_DIR_BASE_ADDR_MASK \
118 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
119
120 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
121 | PT64_NX_MASK)
122
123 #define PFERR_PRESENT_MASK (1U << 0)
124 #define PFERR_WRITE_MASK (1U << 1)
125 #define PFERR_USER_MASK (1U << 2)
126 #define PFERR_FETCH_MASK (1U << 4)
127
128 #define PT_DIRECTORY_LEVEL 2
129 #define PT_PAGE_TABLE_LEVEL 1
130
131 #define RMAP_EXT 4
132
133 #define ACC_EXEC_MASK 1
134 #define ACC_WRITE_MASK PT_WRITABLE_MASK
135 #define ACC_USER_MASK PT_USER_MASK
136 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
137
138 struct kvm_pv_mmu_op_buffer {
139 void *ptr;
140 unsigned len;
141 unsigned processed;
142 char buf[512] __aligned(sizeof(long));
143 };
144
145 struct kvm_rmap_desc {
146 u64 *shadow_ptes[RMAP_EXT];
147 struct kvm_rmap_desc *more;
148 };
149
150 static struct kmem_cache *pte_chain_cache;
151 static struct kmem_cache *rmap_desc_cache;
152 static struct kmem_cache *mmu_page_header_cache;
153
154 static u64 __read_mostly shadow_trap_nonpresent_pte;
155 static u64 __read_mostly shadow_notrap_nonpresent_pte;
156 static u64 __read_mostly shadow_base_present_pte;
157 static u64 __read_mostly shadow_nx_mask;
158 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
159 static u64 __read_mostly shadow_user_mask;
160 static u64 __read_mostly shadow_accessed_mask;
161 static u64 __read_mostly shadow_dirty_mask;
162
163 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
164 {
165 shadow_trap_nonpresent_pte = trap_pte;
166 shadow_notrap_nonpresent_pte = notrap_pte;
167 }
168 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
169
170 void kvm_mmu_set_base_ptes(u64 base_pte)
171 {
172 shadow_base_present_pte = base_pte;
173 }
174 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
175
176 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
177 u64 dirty_mask, u64 nx_mask, u64 x_mask)
178 {
179 shadow_user_mask = user_mask;
180 shadow_accessed_mask = accessed_mask;
181 shadow_dirty_mask = dirty_mask;
182 shadow_nx_mask = nx_mask;
183 shadow_x_mask = x_mask;
184 }
185 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
186
187 static int is_write_protection(struct kvm_vcpu *vcpu)
188 {
189 return vcpu->arch.cr0 & X86_CR0_WP;
190 }
191
192 static int is_cpuid_PSE36(void)
193 {
194 return 1;
195 }
196
197 static int is_nx(struct kvm_vcpu *vcpu)
198 {
199 return vcpu->arch.shadow_efer & EFER_NX;
200 }
201
202 static int is_present_pte(unsigned long pte)
203 {
204 return pte & PT_PRESENT_MASK;
205 }
206
207 static int is_shadow_present_pte(u64 pte)
208 {
209 return pte != shadow_trap_nonpresent_pte
210 && pte != shadow_notrap_nonpresent_pte;
211 }
212
213 static int is_large_pte(u64 pte)
214 {
215 return pte & PT_PAGE_SIZE_MASK;
216 }
217
218 static int is_writeble_pte(unsigned long pte)
219 {
220 return pte & PT_WRITABLE_MASK;
221 }
222
223 static int is_dirty_pte(unsigned long pte)
224 {
225 return pte & shadow_dirty_mask;
226 }
227
228 static int is_rmap_pte(u64 pte)
229 {
230 return is_shadow_present_pte(pte);
231 }
232
233 static pfn_t spte_to_pfn(u64 pte)
234 {
235 return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
236 }
237
238 static gfn_t pse36_gfn_delta(u32 gpte)
239 {
240 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
241
242 return (gpte & PT32_DIR_PSE36_MASK) << shift;
243 }
244
245 static void set_shadow_pte(u64 *sptep, u64 spte)
246 {
247 #ifdef CONFIG_X86_64
248 set_64bit((unsigned long *)sptep, spte);
249 #else
250 set_64bit((unsigned long long *)sptep, spte);
251 #endif
252 }
253
254 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
255 struct kmem_cache *base_cache, int min)
256 {
257 void *obj;
258
259 if (cache->nobjs >= min)
260 return 0;
261 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
262 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
263 if (!obj)
264 return -ENOMEM;
265 cache->objects[cache->nobjs++] = obj;
266 }
267 return 0;
268 }
269
270 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
271 {
272 while (mc->nobjs)
273 kfree(mc->objects[--mc->nobjs]);
274 }
275
276 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
277 int min)
278 {
279 struct page *page;
280
281 if (cache->nobjs >= min)
282 return 0;
283 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
284 page = alloc_page(GFP_KERNEL);
285 if (!page)
286 return -ENOMEM;
287 set_page_private(page, 0);
288 cache->objects[cache->nobjs++] = page_address(page);
289 }
290 return 0;
291 }
292
293 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
294 {
295 while (mc->nobjs)
296 free_page((unsigned long)mc->objects[--mc->nobjs]);
297 }
298
299 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
300 {
301 int r;
302
303 r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
304 pte_chain_cache, 4);
305 if (r)
306 goto out;
307 r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
308 rmap_desc_cache, 1);
309 if (r)
310 goto out;
311 r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
312 if (r)
313 goto out;
314 r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
315 mmu_page_header_cache, 4);
316 out:
317 return r;
318 }
319
320 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
321 {
322 mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
323 mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
324 mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
325 mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
326 }
327
328 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
329 size_t size)
330 {
331 void *p;
332
333 BUG_ON(!mc->nobjs);
334 p = mc->objects[--mc->nobjs];
335 memset(p, 0, size);
336 return p;
337 }
338
339 static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
340 {
341 return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
342 sizeof(struct kvm_pte_chain));
343 }
344
345 static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
346 {
347 kfree(pc);
348 }
349
350 static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
351 {
352 return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
353 sizeof(struct kvm_rmap_desc));
354 }
355
356 static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
357 {
358 kfree(rd);
359 }
360
361 /*
362 * Return the pointer to the largepage write count for a given
363 * gfn, handling slots that are not large page aligned.
364 */
365 static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
366 {
367 unsigned long idx;
368
369 idx = (gfn / KVM_PAGES_PER_HPAGE) -
370 (slot->base_gfn / KVM_PAGES_PER_HPAGE);
371 return &slot->lpage_info[idx].write_count;
372 }
373
374 static void account_shadowed(struct kvm *kvm, gfn_t gfn)
375 {
376 int *write_count;
377
378 write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
379 *write_count += 1;
380 }
381
382 static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
383 {
384 int *write_count;
385
386 write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
387 *write_count -= 1;
388 WARN_ON(*write_count < 0);
389 }
390
391 static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
392 {
393 struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
394 int *largepage_idx;
395
396 if (slot) {
397 largepage_idx = slot_largepage_idx(gfn, slot);
398 return *largepage_idx;
399 }
400
401 return 1;
402 }
403
404 static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
405 {
406 struct vm_area_struct *vma;
407 unsigned long addr;
408
409 addr = gfn_to_hva(kvm, gfn);
410 if (kvm_is_error_hva(addr))
411 return 0;
412
413 vma = find_vma(current->mm, addr);
414 if (vma && is_vm_hugetlb_page(vma))
415 return 1;
416
417 return 0;
418 }
419
420 static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
421 {
422 struct kvm_memory_slot *slot;
423
424 if (has_wrprotected_page(vcpu->kvm, large_gfn))
425 return 0;
426
427 if (!host_largepage_backed(vcpu->kvm, large_gfn))
428 return 0;
429
430 slot = gfn_to_memslot(vcpu->kvm, large_gfn);
431 if (slot && slot->dirty_bitmap)
432 return 0;
433
434 return 1;
435 }
436
437 /*
438 * Take gfn and return the reverse mapping to it.
439 * Note: gfn must be unaliased before this function get called
440 */
441
442 static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
443 {
444 struct kvm_memory_slot *slot;
445 unsigned long idx;
446
447 slot = gfn_to_memslot(kvm, gfn);
448 if (!lpage)
449 return &slot->rmap[gfn - slot->base_gfn];
450
451 idx = (gfn / KVM_PAGES_PER_HPAGE) -
452 (slot->base_gfn / KVM_PAGES_PER_HPAGE);
453
454 return &slot->lpage_info[idx].rmap_pde;
455 }
456
457 /*
458 * Reverse mapping data structures:
459 *
460 * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
461 * that points to page_address(page).
462 *
463 * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
464 * containing more mappings.
465 */
466 static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
467 {
468 struct kvm_mmu_page *sp;
469 struct kvm_rmap_desc *desc;
470 unsigned long *rmapp;
471 int i;
472
473 if (!is_rmap_pte(*spte))
474 return;
475 gfn = unalias_gfn(vcpu->kvm, gfn);
476 sp = page_header(__pa(spte));
477 sp->gfns[spte - sp->spt] = gfn;
478 rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
479 if (!*rmapp) {
480 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
481 *rmapp = (unsigned long)spte;
482 } else if (!(*rmapp & 1)) {
483 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
484 desc = mmu_alloc_rmap_desc(vcpu);
485 desc->shadow_ptes[0] = (u64 *)*rmapp;
486 desc->shadow_ptes[1] = spte;
487 *rmapp = (unsigned long)desc | 1;
488 } else {
489 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
490 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
491 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
492 desc = desc->more;
493 if (desc->shadow_ptes[RMAP_EXT-1]) {
494 desc->more = mmu_alloc_rmap_desc(vcpu);
495 desc = desc->more;
496 }
497 for (i = 0; desc->shadow_ptes[i]; ++i)
498 ;
499 desc->shadow_ptes[i] = spte;
500 }
501 }
502
503 static void rmap_desc_remove_entry(unsigned long *rmapp,
504 struct kvm_rmap_desc *desc,
505 int i,
506 struct kvm_rmap_desc *prev_desc)
507 {
508 int j;
509
510 for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
511 ;
512 desc->shadow_ptes[i] = desc->shadow_ptes[j];
513 desc->shadow_ptes[j] = NULL;
514 if (j != 0)
515 return;
516 if (!prev_desc && !desc->more)
517 *rmapp = (unsigned long)desc->shadow_ptes[0];
518 else
519 if (prev_desc)
520 prev_desc->more = desc->more;
521 else
522 *rmapp = (unsigned long)desc->more | 1;
523 mmu_free_rmap_desc(desc);
524 }
525
526 static void rmap_remove(struct kvm *kvm, u64 *spte)
527 {
528 struct kvm_rmap_desc *desc;
529 struct kvm_rmap_desc *prev_desc;
530 struct kvm_mmu_page *sp;
531 pfn_t pfn;
532 unsigned long *rmapp;
533 int i;
534
535 if (!is_rmap_pte(*spte))
536 return;
537 sp = page_header(__pa(spte));
538 pfn = spte_to_pfn(*spte);
539 if (*spte & shadow_accessed_mask)
540 kvm_set_pfn_accessed(pfn);
541 if (is_writeble_pte(*spte))
542 kvm_release_pfn_dirty(pfn);
543 else
544 kvm_release_pfn_clean(pfn);
545 rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
546 if (!*rmapp) {
547 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
548 BUG();
549 } else if (!(*rmapp & 1)) {
550 rmap_printk("rmap_remove: %p %llx 1->0\n", spte, *spte);
551 if ((u64 *)*rmapp != spte) {
552 printk(KERN_ERR "rmap_remove: %p %llx 1->BUG\n",
553 spte, *spte);
554 BUG();
555 }
556 *rmapp = 0;
557 } else {
558 rmap_printk("rmap_remove: %p %llx many->many\n", spte, *spte);
559 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
560 prev_desc = NULL;
561 while (desc) {
562 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
563 if (desc->shadow_ptes[i] == spte) {
564 rmap_desc_remove_entry(rmapp,
565 desc, i,
566 prev_desc);
567 return;
568 }
569 prev_desc = desc;
570 desc = desc->more;
571 }
572 BUG();
573 }
574 }
575
576 static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
577 {
578 struct kvm_rmap_desc *desc;
579 struct kvm_rmap_desc *prev_desc;
580 u64 *prev_spte;
581 int i;
582
583 if (!*rmapp)
584 return NULL;
585 else if (!(*rmapp & 1)) {
586 if (!spte)
587 return (u64 *)*rmapp;
588 return NULL;
589 }
590 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
591 prev_desc = NULL;
592 prev_spte = NULL;
593 while (desc) {
594 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
595 if (prev_spte == spte)
596 return desc->shadow_ptes[i];
597 prev_spte = desc->shadow_ptes[i];
598 }
599 desc = desc->more;
600 }
601 return NULL;
602 }
603
604 static void rmap_write_protect(struct kvm *kvm, u64 gfn)
605 {
606 unsigned long *rmapp;
607 u64 *spte;
608 int write_protected = 0;
609
610 gfn = unalias_gfn(kvm, gfn);
611 rmapp = gfn_to_rmap(kvm, gfn, 0);
612
613 spte = rmap_next(kvm, rmapp, NULL);
614 while (spte) {
615 BUG_ON(!spte);
616 BUG_ON(!(*spte & PT_PRESENT_MASK));
617 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
618 if (is_writeble_pte(*spte)) {
619 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
620 write_protected = 1;
621 }
622 spte = rmap_next(kvm, rmapp, spte);
623 }
624 if (write_protected) {
625 pfn_t pfn;
626
627 spte = rmap_next(kvm, rmapp, NULL);
628 pfn = spte_to_pfn(*spte);
629 kvm_set_pfn_dirty(pfn);
630 }
631
632 /* check for huge page mappings */
633 rmapp = gfn_to_rmap(kvm, gfn, 1);
634 spte = rmap_next(kvm, rmapp, NULL);
635 while (spte) {
636 BUG_ON(!spte);
637 BUG_ON(!(*spte & PT_PRESENT_MASK));
638 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
639 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
640 if (is_writeble_pte(*spte)) {
641 rmap_remove(kvm, spte);
642 --kvm->stat.lpages;
643 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
644 spte = NULL;
645 write_protected = 1;
646 }
647 spte = rmap_next(kvm, rmapp, spte);
648 }
649
650 if (write_protected)
651 kvm_flush_remote_tlbs(kvm);
652
653 account_shadowed(kvm, gfn);
654 }
655
656 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp)
657 {
658 u64 *spte;
659 int need_tlb_flush = 0;
660
661 while ((spte = rmap_next(kvm, rmapp, NULL))) {
662 BUG_ON(!(*spte & PT_PRESENT_MASK));
663 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte, *spte);
664 rmap_remove(kvm, spte);
665 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
666 need_tlb_flush = 1;
667 }
668 return need_tlb_flush;
669 }
670
671 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
672 int (*handler)(struct kvm *kvm, unsigned long *rmapp))
673 {
674 int i;
675 int retval = 0;
676
677 /*
678 * If mmap_sem isn't taken, we can look the memslots with only
679 * the mmu_lock by skipping over the slots with userspace_addr == 0.
680 */
681 for (i = 0; i < kvm->nmemslots; i++) {
682 struct kvm_memory_slot *memslot = &kvm->memslots[i];
683 unsigned long start = memslot->userspace_addr;
684 unsigned long end;
685
686 /* mmu_lock protects userspace_addr */
687 if (!start)
688 continue;
689
690 end = start + (memslot->npages << PAGE_SHIFT);
691 if (hva >= start && hva < end) {
692 gfn_t gfn_offset = (hva - start) >> PAGE_SHIFT;
693 retval |= handler(kvm, &memslot->rmap[gfn_offset]);
694 retval |= handler(kvm,
695 &memslot->lpage_info[
696 gfn_offset /
697 KVM_PAGES_PER_HPAGE].rmap_pde);
698 }
699 }
700
701 return retval;
702 }
703
704 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
705 {
706 return kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
707 }
708
709 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp)
710 {
711 u64 *spte;
712 int young = 0;
713
714 /* always return old for EPT */
715 if (!shadow_accessed_mask)
716 return 0;
717
718 spte = rmap_next(kvm, rmapp, NULL);
719 while (spte) {
720 int _young;
721 u64 _spte = *spte;
722 BUG_ON(!(_spte & PT_PRESENT_MASK));
723 _young = _spte & PT_ACCESSED_MASK;
724 if (_young) {
725 young = 1;
726 clear_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
727 }
728 spte = rmap_next(kvm, rmapp, spte);
729 }
730 return young;
731 }
732
733 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
734 {
735 return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
736 }
737
738 #ifdef MMU_DEBUG
739 static int is_empty_shadow_page(u64 *spt)
740 {
741 u64 *pos;
742 u64 *end;
743
744 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
745 if (is_shadow_present_pte(*pos)) {
746 printk(KERN_ERR "%s: %p %llx\n", __func__,
747 pos, *pos);
748 return 0;
749 }
750 return 1;
751 }
752 #endif
753
754 static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
755 {
756 ASSERT(is_empty_shadow_page(sp->spt));
757 list_del(&sp->link);
758 __free_page(virt_to_page(sp->spt));
759 __free_page(virt_to_page(sp->gfns));
760 kfree(sp);
761 ++kvm->arch.n_free_mmu_pages;
762 }
763
764 static unsigned kvm_page_table_hashfn(gfn_t gfn)
765 {
766 return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
767 }
768
769 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
770 u64 *parent_pte)
771 {
772 struct kvm_mmu_page *sp;
773
774 sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
775 sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
776 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
777 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
778 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
779 ASSERT(is_empty_shadow_page(sp->spt));
780 sp->slot_bitmap = 0;
781 sp->multimapped = 0;
782 sp->parent_pte = parent_pte;
783 --vcpu->kvm->arch.n_free_mmu_pages;
784 return sp;
785 }
786
787 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
788 struct kvm_mmu_page *sp, u64 *parent_pte)
789 {
790 struct kvm_pte_chain *pte_chain;
791 struct hlist_node *node;
792 int i;
793
794 if (!parent_pte)
795 return;
796 if (!sp->multimapped) {
797 u64 *old = sp->parent_pte;
798
799 if (!old) {
800 sp->parent_pte = parent_pte;
801 return;
802 }
803 sp->multimapped = 1;
804 pte_chain = mmu_alloc_pte_chain(vcpu);
805 INIT_HLIST_HEAD(&sp->parent_ptes);
806 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
807 pte_chain->parent_ptes[0] = old;
808 }
809 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
810 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
811 continue;
812 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
813 if (!pte_chain->parent_ptes[i]) {
814 pte_chain->parent_ptes[i] = parent_pte;
815 return;
816 }
817 }
818 pte_chain = mmu_alloc_pte_chain(vcpu);
819 BUG_ON(!pte_chain);
820 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
821 pte_chain->parent_ptes[0] = parent_pte;
822 }
823
824 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
825 u64 *parent_pte)
826 {
827 struct kvm_pte_chain *pte_chain;
828 struct hlist_node *node;
829 int i;
830
831 if (!sp->multimapped) {
832 BUG_ON(sp->parent_pte != parent_pte);
833 sp->parent_pte = NULL;
834 return;
835 }
836 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
837 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
838 if (!pte_chain->parent_ptes[i])
839 break;
840 if (pte_chain->parent_ptes[i] != parent_pte)
841 continue;
842 while (i + 1 < NR_PTE_CHAIN_ENTRIES
843 && pte_chain->parent_ptes[i + 1]) {
844 pte_chain->parent_ptes[i]
845 = pte_chain->parent_ptes[i + 1];
846 ++i;
847 }
848 pte_chain->parent_ptes[i] = NULL;
849 if (i == 0) {
850 hlist_del(&pte_chain->link);
851 mmu_free_pte_chain(pte_chain);
852 if (hlist_empty(&sp->parent_ptes)) {
853 sp->multimapped = 0;
854 sp->parent_pte = NULL;
855 }
856 }
857 return;
858 }
859 BUG();
860 }
861
862 static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
863 struct kvm_mmu_page *sp)
864 {
865 int i;
866
867 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
868 sp->spt[i] = shadow_trap_nonpresent_pte;
869 }
870
871 static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
872 {
873 unsigned index;
874 struct hlist_head *bucket;
875 struct kvm_mmu_page *sp;
876 struct hlist_node *node;
877
878 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
879 index = kvm_page_table_hashfn(gfn);
880 bucket = &kvm->arch.mmu_page_hash[index];
881 hlist_for_each_entry(sp, node, bucket, hash_link)
882 if (sp->gfn == gfn && !sp->role.metaphysical
883 && !sp->role.invalid) {
884 pgprintk("%s: found role %x\n",
885 __func__, sp->role.word);
886 return sp;
887 }
888 return NULL;
889 }
890
891 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
892 gfn_t gfn,
893 gva_t gaddr,
894 unsigned level,
895 int metaphysical,
896 unsigned access,
897 u64 *parent_pte)
898 {
899 union kvm_mmu_page_role role;
900 unsigned index;
901 unsigned quadrant;
902 struct hlist_head *bucket;
903 struct kvm_mmu_page *sp;
904 struct hlist_node *node;
905
906 role.word = 0;
907 role.glevels = vcpu->arch.mmu.root_level;
908 role.level = level;
909 role.metaphysical = metaphysical;
910 role.access = access;
911 if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
912 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
913 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
914 role.quadrant = quadrant;
915 }
916 pgprintk("%s: looking gfn %lx role %x\n", __func__,
917 gfn, role.word);
918 index = kvm_page_table_hashfn(gfn);
919 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
920 hlist_for_each_entry(sp, node, bucket, hash_link)
921 if (sp->gfn == gfn && sp->role.word == role.word) {
922 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
923 pgprintk("%s: found\n", __func__);
924 return sp;
925 }
926 ++vcpu->kvm->stat.mmu_cache_miss;
927 sp = kvm_mmu_alloc_page(vcpu, parent_pte);
928 if (!sp)
929 return sp;
930 pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
931 sp->gfn = gfn;
932 sp->role = role;
933 hlist_add_head(&sp->hash_link, bucket);
934 if (!metaphysical)
935 rmap_write_protect(vcpu->kvm, gfn);
936 if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
937 vcpu->arch.mmu.prefetch_page(vcpu, sp);
938 else
939 nonpaging_prefetch_page(vcpu, sp);
940 return sp;
941 }
942
943 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
944 struct kvm_mmu_page *sp)
945 {
946 unsigned i;
947 u64 *pt;
948 u64 ent;
949
950 pt = sp->spt;
951
952 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
953 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
954 if (is_shadow_present_pte(pt[i]))
955 rmap_remove(kvm, &pt[i]);
956 pt[i] = shadow_trap_nonpresent_pte;
957 }
958 return;
959 }
960
961 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
962 ent = pt[i];
963
964 if (is_shadow_present_pte(ent)) {
965 if (!is_large_pte(ent)) {
966 ent &= PT64_BASE_ADDR_MASK;
967 mmu_page_remove_parent_pte(page_header(ent),
968 &pt[i]);
969 } else {
970 --kvm->stat.lpages;
971 rmap_remove(kvm, &pt[i]);
972 }
973 }
974 pt[i] = shadow_trap_nonpresent_pte;
975 }
976 }
977
978 static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
979 {
980 mmu_page_remove_parent_pte(sp, parent_pte);
981 }
982
983 static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
984 {
985 int i;
986
987 for (i = 0; i < KVM_MAX_VCPUS; ++i)
988 if (kvm->vcpus[i])
989 kvm->vcpus[i]->arch.last_pte_updated = NULL;
990 }
991
992 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
993 {
994 u64 *parent_pte;
995
996 while (sp->multimapped || sp->parent_pte) {
997 if (!sp->multimapped)
998 parent_pte = sp->parent_pte;
999 else {
1000 struct kvm_pte_chain *chain;
1001
1002 chain = container_of(sp->parent_ptes.first,
1003 struct kvm_pte_chain, link);
1004 parent_pte = chain->parent_ptes[0];
1005 }
1006 BUG_ON(!parent_pte);
1007 kvm_mmu_put_page(sp, parent_pte);
1008 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
1009 }
1010 }
1011
1012 static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1013 {
1014 ++kvm->stat.mmu_shadow_zapped;
1015 kvm_mmu_page_unlink_children(kvm, sp);
1016 kvm_mmu_unlink_parents(kvm, sp);
1017 kvm_flush_remote_tlbs(kvm);
1018 if (!sp->role.invalid && !sp->role.metaphysical)
1019 unaccount_shadowed(kvm, sp->gfn);
1020 if (!sp->root_count) {
1021 hlist_del(&sp->hash_link);
1022 kvm_mmu_free_page(kvm, sp);
1023 } else {
1024 sp->role.invalid = 1;
1025 list_move(&sp->link, &kvm->arch.active_mmu_pages);
1026 kvm_reload_remote_mmus(kvm);
1027 }
1028 kvm_mmu_reset_last_pte_updated(kvm);
1029 }
1030
1031 /*
1032 * Changing the number of mmu pages allocated to the vm
1033 * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1034 */
1035 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
1036 {
1037 /*
1038 * If we set the number of mmu pages to be smaller be than the
1039 * number of actived pages , we must to free some mmu pages before we
1040 * change the value
1041 */
1042
1043 if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
1044 kvm_nr_mmu_pages) {
1045 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
1046 - kvm->arch.n_free_mmu_pages;
1047
1048 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
1049 struct kvm_mmu_page *page;
1050
1051 page = container_of(kvm->arch.active_mmu_pages.prev,
1052 struct kvm_mmu_page, link);
1053 kvm_mmu_zap_page(kvm, page);
1054 n_used_mmu_pages--;
1055 }
1056 kvm->arch.n_free_mmu_pages = 0;
1057 }
1058 else
1059 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
1060 - kvm->arch.n_alloc_mmu_pages;
1061
1062 kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
1063 }
1064
1065 static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
1066 {
1067 unsigned index;
1068 struct hlist_head *bucket;
1069 struct kvm_mmu_page *sp;
1070 struct hlist_node *node, *n;
1071 int r;
1072
1073 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
1074 r = 0;
1075 index = kvm_page_table_hashfn(gfn);
1076 bucket = &kvm->arch.mmu_page_hash[index];
1077 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
1078 if (sp->gfn == gfn && !sp->role.metaphysical) {
1079 pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
1080 sp->role.word);
1081 kvm_mmu_zap_page(kvm, sp);
1082 r = 1;
1083 }
1084 return r;
1085 }
1086
1087 static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1088 {
1089 struct kvm_mmu_page *sp;
1090
1091 while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1092 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1093 kvm_mmu_zap_page(kvm, sp);
1094 }
1095 }
1096
1097 static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1098 {
1099 int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1100 struct kvm_mmu_page *sp = page_header(__pa(pte));
1101
1102 __set_bit(slot, &sp->slot_bitmap);
1103 }
1104
1105 struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1106 {
1107 struct page *page;
1108
1109 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1110
1111 if (gpa == UNMAPPED_GVA)
1112 return NULL;
1113
1114 down_read(&current->mm->mmap_sem);
1115 page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1116 up_read(&current->mm->mmap_sem);
1117
1118 return page;
1119 }
1120
1121 static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1122 unsigned pt_access, unsigned pte_access,
1123 int user_fault, int write_fault, int dirty,
1124 int *ptwrite, int largepage, gfn_t gfn,
1125 pfn_t pfn, bool speculative)
1126 {
1127 u64 spte;
1128 int was_rmapped = 0;
1129 int was_writeble = is_writeble_pte(*shadow_pte);
1130
1131 pgprintk("%s: spte %llx access %x write_fault %d"
1132 " user_fault %d gfn %lx\n",
1133 __func__, *shadow_pte, pt_access,
1134 write_fault, user_fault, gfn);
1135
1136 if (is_rmap_pte(*shadow_pte)) {
1137 /*
1138 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1139 * the parent of the now unreachable PTE.
1140 */
1141 if (largepage && !is_large_pte(*shadow_pte)) {
1142 struct kvm_mmu_page *child;
1143 u64 pte = *shadow_pte;
1144
1145 child = page_header(pte & PT64_BASE_ADDR_MASK);
1146 mmu_page_remove_parent_pte(child, shadow_pte);
1147 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1148 pgprintk("hfn old %lx new %lx\n",
1149 spte_to_pfn(*shadow_pte), pfn);
1150 rmap_remove(vcpu->kvm, shadow_pte);
1151 } else {
1152 if (largepage)
1153 was_rmapped = is_large_pte(*shadow_pte);
1154 else
1155 was_rmapped = 1;
1156 }
1157 }
1158
1159 /*
1160 * We don't set the accessed bit, since we sometimes want to see
1161 * whether the guest actually used the pte (in order to detect
1162 * demand paging).
1163 */
1164 spte = shadow_base_present_pte | shadow_dirty_mask;
1165 if (!speculative)
1166 pte_access |= PT_ACCESSED_MASK;
1167 if (!dirty)
1168 pte_access &= ~ACC_WRITE_MASK;
1169 if (pte_access & ACC_EXEC_MASK)
1170 spte |= shadow_x_mask;
1171 else
1172 spte |= shadow_nx_mask;
1173 if (pte_access & ACC_USER_MASK)
1174 spte |= shadow_user_mask;
1175 if (largepage)
1176 spte |= PT_PAGE_SIZE_MASK;
1177
1178 spte |= (u64)pfn << PAGE_SHIFT;
1179
1180 if ((pte_access & ACC_WRITE_MASK)
1181 || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1182 struct kvm_mmu_page *shadow;
1183
1184 spte |= PT_WRITABLE_MASK;
1185
1186 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1187 if (shadow ||
1188 (largepage && has_wrprotected_page(vcpu->kvm, gfn))) {
1189 pgprintk("%s: found shadow page for %lx, marking ro\n",
1190 __func__, gfn);
1191 pte_access &= ~ACC_WRITE_MASK;
1192 if (is_writeble_pte(spte)) {
1193 spte &= ~PT_WRITABLE_MASK;
1194 kvm_x86_ops->tlb_flush(vcpu);
1195 }
1196 if (write_fault)
1197 *ptwrite = 1;
1198 }
1199 }
1200
1201 if (pte_access & ACC_WRITE_MASK)
1202 mark_page_dirty(vcpu->kvm, gfn);
1203
1204 pgprintk("%s: setting spte %llx\n", __func__, spte);
1205 pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1206 (spte&PT_PAGE_SIZE_MASK)? "2MB" : "4kB",
1207 (spte&PT_WRITABLE_MASK)?"RW":"R", gfn, spte, shadow_pte);
1208 set_shadow_pte(shadow_pte, spte);
1209 if (!was_rmapped && (spte & PT_PAGE_SIZE_MASK)
1210 && (spte & PT_PRESENT_MASK))
1211 ++vcpu->kvm->stat.lpages;
1212
1213 page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1214 if (!was_rmapped) {
1215 rmap_add(vcpu, shadow_pte, gfn, largepage);
1216 if (!is_rmap_pte(*shadow_pte))
1217 kvm_release_pfn_clean(pfn);
1218 } else {
1219 if (was_writeble)
1220 kvm_release_pfn_dirty(pfn);
1221 else
1222 kvm_release_pfn_clean(pfn);
1223 }
1224 if (speculative) {
1225 vcpu->arch.last_pte_updated = shadow_pte;
1226 vcpu->arch.last_pte_gfn = gfn;
1227 }
1228 }
1229
1230 static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1231 {
1232 }
1233
1234 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1235 int largepage, gfn_t gfn, pfn_t pfn,
1236 int level)
1237 {
1238 hpa_t table_addr = vcpu->arch.mmu.root_hpa;
1239 int pt_write = 0;
1240
1241 for (; ; level--) {
1242 u32 index = PT64_INDEX(v, level);
1243 u64 *table;
1244
1245 ASSERT(VALID_PAGE(table_addr));
1246 table = __va(table_addr);
1247
1248 if (level == 1) {
1249 mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1250 0, write, 1, &pt_write, 0, gfn, pfn, false);
1251 return pt_write;
1252 }
1253
1254 if (largepage && level == 2) {
1255 mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1256 0, write, 1, &pt_write, 1, gfn, pfn, false);
1257 return pt_write;
1258 }
1259
1260 if (table[index] == shadow_trap_nonpresent_pte) {
1261 struct kvm_mmu_page *new_table;
1262 gfn_t pseudo_gfn;
1263
1264 pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
1265 >> PAGE_SHIFT;
1266 new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
1267 v, level - 1,
1268 1, ACC_ALL, &table[index]);
1269 if (!new_table) {
1270 pgprintk("nonpaging_map: ENOMEM\n");
1271 kvm_release_pfn_clean(pfn);
1272 return -ENOMEM;
1273 }
1274
1275 set_shadow_pte(&table[index],
1276 __pa(new_table->spt)
1277 | PT_PRESENT_MASK | PT_WRITABLE_MASK
1278 | shadow_user_mask | shadow_x_mask);
1279 }
1280 table_addr = table[index] & PT64_BASE_ADDR_MASK;
1281 }
1282 }
1283
1284 static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1285 {
1286 int r;
1287 int largepage = 0;
1288 pfn_t pfn;
1289 unsigned long mmu_seq;
1290
1291 down_read(&current->mm->mmap_sem);
1292 if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1293 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1294 largepage = 1;
1295 }
1296
1297 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1298 /* implicit mb(), we'll read before PT lock is unlocked */
1299 pfn = gfn_to_pfn(vcpu->kvm, gfn);
1300 up_read(&current->mm->mmap_sem);
1301
1302 /* mmio */
1303 if (is_error_pfn(pfn)) {
1304 kvm_release_pfn_clean(pfn);
1305 return 1;
1306 }
1307
1308 spin_lock(&vcpu->kvm->mmu_lock);
1309 if (mmu_notifier_retry(vcpu, mmu_seq))
1310 goto out_unlock;
1311 kvm_mmu_free_some_pages(vcpu);
1312 r = __direct_map(vcpu, v, write, largepage, gfn, pfn,
1313 PT32E_ROOT_LEVEL);
1314 spin_unlock(&vcpu->kvm->mmu_lock);
1315
1316
1317 return r;
1318
1319 out_unlock:
1320 spin_unlock(&vcpu->kvm->mmu_lock);
1321 kvm_release_pfn_clean(pfn);
1322 return 0;
1323 }
1324
1325
1326 static void mmu_free_roots(struct kvm_vcpu *vcpu)
1327 {
1328 int i;
1329 struct kvm_mmu_page *sp;
1330
1331 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1332 return;
1333 spin_lock(&vcpu->kvm->mmu_lock);
1334 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1335 hpa_t root = vcpu->arch.mmu.root_hpa;
1336
1337 sp = page_header(root);
1338 --sp->root_count;
1339 if (!sp->root_count && sp->role.invalid)
1340 kvm_mmu_zap_page(vcpu->kvm, sp);
1341 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1342 spin_unlock(&vcpu->kvm->mmu_lock);
1343 return;
1344 }
1345 for (i = 0; i < 4; ++i) {
1346 hpa_t root = vcpu->arch.mmu.pae_root[i];
1347
1348 if (root) {
1349 root &= PT64_BASE_ADDR_MASK;
1350 sp = page_header(root);
1351 --sp->root_count;
1352 if (!sp->root_count && sp->role.invalid)
1353 kvm_mmu_zap_page(vcpu->kvm, sp);
1354 }
1355 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1356 }
1357 spin_unlock(&vcpu->kvm->mmu_lock);
1358 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1359 }
1360
1361 static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1362 {
1363 int i;
1364 gfn_t root_gfn;
1365 struct kvm_mmu_page *sp;
1366 int metaphysical = 0;
1367
1368 root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1369
1370 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1371 hpa_t root = vcpu->arch.mmu.root_hpa;
1372
1373 ASSERT(!VALID_PAGE(root));
1374 if (tdp_enabled)
1375 metaphysical = 1;
1376 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1377 PT64_ROOT_LEVEL, metaphysical,
1378 ACC_ALL, NULL);
1379 root = __pa(sp->spt);
1380 ++sp->root_count;
1381 vcpu->arch.mmu.root_hpa = root;
1382 return;
1383 }
1384 metaphysical = !is_paging(vcpu);
1385 if (tdp_enabled)
1386 metaphysical = 1;
1387 for (i = 0; i < 4; ++i) {
1388 hpa_t root = vcpu->arch.mmu.pae_root[i];
1389
1390 ASSERT(!VALID_PAGE(root));
1391 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1392 if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1393 vcpu->arch.mmu.pae_root[i] = 0;
1394 continue;
1395 }
1396 root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1397 } else if (vcpu->arch.mmu.root_level == 0)
1398 root_gfn = 0;
1399 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1400 PT32_ROOT_LEVEL, metaphysical,
1401 ACC_ALL, NULL);
1402 root = __pa(sp->spt);
1403 ++sp->root_count;
1404 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1405 }
1406 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1407 }
1408
1409 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1410 {
1411 return vaddr;
1412 }
1413
1414 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1415 u32 error_code)
1416 {
1417 gfn_t gfn;
1418 int r;
1419
1420 pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
1421 r = mmu_topup_memory_caches(vcpu);
1422 if (r)
1423 return r;
1424
1425 ASSERT(vcpu);
1426 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1427
1428 gfn = gva >> PAGE_SHIFT;
1429
1430 return nonpaging_map(vcpu, gva & PAGE_MASK,
1431 error_code & PFERR_WRITE_MASK, gfn);
1432 }
1433
1434 static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1435 u32 error_code)
1436 {
1437 pfn_t pfn;
1438 int r;
1439 int largepage = 0;
1440 gfn_t gfn = gpa >> PAGE_SHIFT;
1441 unsigned long mmu_seq;
1442
1443 ASSERT(vcpu);
1444 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1445
1446 r = mmu_topup_memory_caches(vcpu);
1447 if (r)
1448 return r;
1449
1450 down_read(&current->mm->mmap_sem);
1451 if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1452 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1453 largepage = 1;
1454 }
1455 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1456 /* implicit mb(), we'll read before PT lock is unlocked */
1457 pfn = gfn_to_pfn(vcpu->kvm, gfn);
1458 up_read(&current->mm->mmap_sem);
1459 if (is_error_pfn(pfn)) {
1460 kvm_release_pfn_clean(pfn);
1461 return 1;
1462 }
1463 spin_lock(&vcpu->kvm->mmu_lock);
1464 if (mmu_notifier_retry(vcpu, mmu_seq))
1465 goto out_unlock;
1466 kvm_mmu_free_some_pages(vcpu);
1467 r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1468 largepage, gfn, pfn, kvm_x86_ops->get_tdp_level());
1469 spin_unlock(&vcpu->kvm->mmu_lock);
1470
1471 return r;
1472
1473 out_unlock:
1474 spin_unlock(&vcpu->kvm->mmu_lock);
1475 kvm_release_pfn_clean(pfn);
1476 return 0;
1477 }
1478
1479 static void nonpaging_free(struct kvm_vcpu *vcpu)
1480 {
1481 mmu_free_roots(vcpu);
1482 }
1483
1484 static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1485 {
1486 struct kvm_mmu *context = &vcpu->arch.mmu;
1487
1488 context->new_cr3 = nonpaging_new_cr3;
1489 context->page_fault = nonpaging_page_fault;
1490 context->gva_to_gpa = nonpaging_gva_to_gpa;
1491 context->free = nonpaging_free;
1492 context->prefetch_page = nonpaging_prefetch_page;
1493 context->root_level = 0;
1494 context->shadow_root_level = PT32E_ROOT_LEVEL;
1495 context->root_hpa = INVALID_PAGE;
1496 return 0;
1497 }
1498
1499 void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1500 {
1501 ++vcpu->stat.tlb_flush;
1502 kvm_x86_ops->tlb_flush(vcpu);
1503 }
1504
1505 static void paging_new_cr3(struct kvm_vcpu *vcpu)
1506 {
1507 pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
1508 mmu_free_roots(vcpu);
1509 }
1510
1511 static void inject_page_fault(struct kvm_vcpu *vcpu,
1512 u64 addr,
1513 u32 err_code)
1514 {
1515 kvm_inject_page_fault(vcpu, addr, err_code);
1516 }
1517
1518 static void paging_free(struct kvm_vcpu *vcpu)
1519 {
1520 nonpaging_free(vcpu);
1521 }
1522
1523 #define PTTYPE 64
1524 #include "paging_tmpl.h"
1525 #undef PTTYPE
1526
1527 #define PTTYPE 32
1528 #include "paging_tmpl.h"
1529 #undef PTTYPE
1530
1531 static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1532 {
1533 struct kvm_mmu *context = &vcpu->arch.mmu;
1534
1535 ASSERT(is_pae(vcpu));
1536 context->new_cr3 = paging_new_cr3;
1537 context->page_fault = paging64_page_fault;
1538 context->gva_to_gpa = paging64_gva_to_gpa;
1539 context->prefetch_page = paging64_prefetch_page;
1540 context->free = paging_free;
1541 context->root_level = level;
1542 context->shadow_root_level = level;
1543 context->root_hpa = INVALID_PAGE;
1544 return 0;
1545 }
1546
1547 static int paging64_init_context(struct kvm_vcpu *vcpu)
1548 {
1549 return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1550 }
1551
1552 static int paging32_init_context(struct kvm_vcpu *vcpu)
1553 {
1554 struct kvm_mmu *context = &vcpu->arch.mmu;
1555
1556 context->new_cr3 = paging_new_cr3;
1557 context->page_fault = paging32_page_fault;
1558 context->gva_to_gpa = paging32_gva_to_gpa;
1559 context->free = paging_free;
1560 context->prefetch_page = paging32_prefetch_page;
1561 context->root_level = PT32_ROOT_LEVEL;
1562 context->shadow_root_level = PT32E_ROOT_LEVEL;
1563 context->root_hpa = INVALID_PAGE;
1564 return 0;
1565 }
1566
1567 static int paging32E_init_context(struct kvm_vcpu *vcpu)
1568 {
1569 return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1570 }
1571
1572 static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
1573 {
1574 struct kvm_mmu *context = &vcpu->arch.mmu;
1575
1576 context->new_cr3 = nonpaging_new_cr3;
1577 context->page_fault = tdp_page_fault;
1578 context->free = nonpaging_free;
1579 context->prefetch_page = nonpaging_prefetch_page;
1580 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
1581 context->root_hpa = INVALID_PAGE;
1582
1583 if (!is_paging(vcpu)) {
1584 context->gva_to_gpa = nonpaging_gva_to_gpa;
1585 context->root_level = 0;
1586 } else if (is_long_mode(vcpu)) {
1587 context->gva_to_gpa = paging64_gva_to_gpa;
1588 context->root_level = PT64_ROOT_LEVEL;
1589 } else if (is_pae(vcpu)) {
1590 context->gva_to_gpa = paging64_gva_to_gpa;
1591 context->root_level = PT32E_ROOT_LEVEL;
1592 } else {
1593 context->gva_to_gpa = paging32_gva_to_gpa;
1594 context->root_level = PT32_ROOT_LEVEL;
1595 }
1596
1597 return 0;
1598 }
1599
1600 static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
1601 {
1602 ASSERT(vcpu);
1603 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1604
1605 if (!is_paging(vcpu))
1606 return nonpaging_init_context(vcpu);
1607 else if (is_long_mode(vcpu))
1608 return paging64_init_context(vcpu);
1609 else if (is_pae(vcpu))
1610 return paging32E_init_context(vcpu);
1611 else
1612 return paging32_init_context(vcpu);
1613 }
1614
1615 static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1616 {
1617 vcpu->arch.update_pte.pfn = bad_pfn;
1618
1619 if (tdp_enabled)
1620 return init_kvm_tdp_mmu(vcpu);
1621 else
1622 return init_kvm_softmmu(vcpu);
1623 }
1624
1625 static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1626 {
1627 ASSERT(vcpu);
1628 if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1629 vcpu->arch.mmu.free(vcpu);
1630 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1631 }
1632 }
1633
1634 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1635 {
1636 destroy_kvm_mmu(vcpu);
1637 return init_kvm_mmu(vcpu);
1638 }
1639 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1640
1641 int kvm_mmu_load(struct kvm_vcpu *vcpu)
1642 {
1643 int r;
1644
1645 r = mmu_topup_memory_caches(vcpu);
1646 if (r)
1647 goto out;
1648 spin_lock(&vcpu->kvm->mmu_lock);
1649 kvm_mmu_free_some_pages(vcpu);
1650 mmu_alloc_roots(vcpu);
1651 spin_unlock(&vcpu->kvm->mmu_lock);
1652 kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1653 kvm_mmu_flush_tlb(vcpu);
1654 out:
1655 return r;
1656 }
1657 EXPORT_SYMBOL_GPL(kvm_mmu_load);
1658
1659 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1660 {
1661 mmu_free_roots(vcpu);
1662 }
1663
1664 static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1665 struct kvm_mmu_page *sp,
1666 u64 *spte)
1667 {
1668 u64 pte;
1669 struct kvm_mmu_page *child;
1670
1671 pte = *spte;
1672 if (is_shadow_present_pte(pte)) {
1673 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
1674 is_large_pte(pte))
1675 rmap_remove(vcpu->kvm, spte);
1676 else {
1677 child = page_header(pte & PT64_BASE_ADDR_MASK);
1678 mmu_page_remove_parent_pte(child, spte);
1679 }
1680 }
1681 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1682 if (is_large_pte(pte))
1683 --vcpu->kvm->stat.lpages;
1684 }
1685
1686 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1687 struct kvm_mmu_page *sp,
1688 u64 *spte,
1689 const void *new)
1690 {
1691 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
1692 if (!vcpu->arch.update_pte.largepage ||
1693 sp->role.glevels == PT32_ROOT_LEVEL) {
1694 ++vcpu->kvm->stat.mmu_pde_zapped;
1695 return;
1696 }
1697 }
1698
1699 ++vcpu->kvm->stat.mmu_pte_updated;
1700 if (sp->role.glevels == PT32_ROOT_LEVEL)
1701 paging32_update_pte(vcpu, sp, spte, new);
1702 else
1703 paging64_update_pte(vcpu, sp, spte, new);
1704 }
1705
1706 static bool need_remote_flush(u64 old, u64 new)
1707 {
1708 if (!is_shadow_present_pte(old))
1709 return false;
1710 if (!is_shadow_present_pte(new))
1711 return true;
1712 if ((old ^ new) & PT64_BASE_ADDR_MASK)
1713 return true;
1714 old ^= PT64_NX_MASK;
1715 new ^= PT64_NX_MASK;
1716 return (old & ~new & PT64_PERM_MASK) != 0;
1717 }
1718
1719 static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1720 {
1721 if (need_remote_flush(old, new))
1722 kvm_flush_remote_tlbs(vcpu->kvm);
1723 else
1724 kvm_mmu_flush_tlb(vcpu);
1725 }
1726
1727 static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1728 {
1729 u64 *spte = vcpu->arch.last_pte_updated;
1730
1731 return !!(spte && (*spte & shadow_accessed_mask));
1732 }
1733
1734 static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1735 const u8 *new, int bytes)
1736 {
1737 gfn_t gfn;
1738 int r;
1739 u64 gpte = 0;
1740 pfn_t pfn;
1741
1742 vcpu->arch.update_pte.largepage = 0;
1743
1744 if (bytes != 4 && bytes != 8)
1745 return;
1746
1747 /*
1748 * Assume that the pte write on a page table of the same type
1749 * as the current vcpu paging mode. This is nearly always true
1750 * (might be false while changing modes). Note it is verified later
1751 * by update_pte().
1752 */
1753 if (is_pae(vcpu)) {
1754 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1755 if ((bytes == 4) && (gpa % 4 == 0)) {
1756 r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
1757 if (r)
1758 return;
1759 memcpy((void *)&gpte + (gpa % 8), new, 4);
1760 } else if ((bytes == 8) && (gpa % 8 == 0)) {
1761 memcpy((void *)&gpte, new, 8);
1762 }
1763 } else {
1764 if ((bytes == 4) && (gpa % 4 == 0))
1765 memcpy((void *)&gpte, new, 4);
1766 }
1767 if (!is_present_pte(gpte))
1768 return;
1769 gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
1770
1771 down_read(&current->mm->mmap_sem);
1772 if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
1773 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1774 vcpu->arch.update_pte.largepage = 1;
1775 }
1776 vcpu->arch.update_pte.mmu_seq = vcpu->kvm->mmu_notifier_seq;
1777 /* implicit mb(), we'll read before PT lock is unlocked */
1778 pfn = gfn_to_pfn(vcpu->kvm, gfn);
1779 up_read(&current->mm->mmap_sem);
1780
1781 if (is_error_pfn(pfn)) {
1782 kvm_release_pfn_clean(pfn);
1783 return;
1784 }
1785 vcpu->arch.update_pte.gfn = gfn;
1786 vcpu->arch.update_pte.pfn = pfn;
1787 }
1788
1789 static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
1790 {
1791 u64 *spte = vcpu->arch.last_pte_updated;
1792
1793 if (spte
1794 && vcpu->arch.last_pte_gfn == gfn
1795 && shadow_accessed_mask
1796 && !(*spte & shadow_accessed_mask)
1797 && is_shadow_present_pte(*spte))
1798 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
1799 }
1800
1801 void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1802 const u8 *new, int bytes)
1803 {
1804 gfn_t gfn = gpa >> PAGE_SHIFT;
1805 struct kvm_mmu_page *sp;
1806 struct hlist_node *node, *n;
1807 struct hlist_head *bucket;
1808 unsigned index;
1809 u64 entry, gentry;
1810 u64 *spte;
1811 unsigned offset = offset_in_page(gpa);
1812 unsigned pte_size;
1813 unsigned page_offset;
1814 unsigned misaligned;
1815 unsigned quadrant;
1816 int level;
1817 int flooded = 0;
1818 int npte;
1819 int r;
1820
1821 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
1822 mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
1823 spin_lock(&vcpu->kvm->mmu_lock);
1824 kvm_mmu_access_page(vcpu, gfn);
1825 kvm_mmu_free_some_pages(vcpu);
1826 ++vcpu->kvm->stat.mmu_pte_write;
1827 kvm_mmu_audit(vcpu, "pre pte write");
1828 if (gfn == vcpu->arch.last_pt_write_gfn
1829 && !last_updated_pte_accessed(vcpu)) {
1830 ++vcpu->arch.last_pt_write_count;
1831 if (vcpu->arch.last_pt_write_count >= 3)
1832 flooded = 1;
1833 } else {
1834 vcpu->arch.last_pt_write_gfn = gfn;
1835 vcpu->arch.last_pt_write_count = 1;
1836 vcpu->arch.last_pte_updated = NULL;
1837 }
1838 index = kvm_page_table_hashfn(gfn);
1839 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1840 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1841 if (sp->gfn != gfn || sp->role.metaphysical || sp->role.invalid)
1842 continue;
1843 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1844 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1845 misaligned |= bytes < 4;
1846 if (misaligned || flooded) {
1847 /*
1848 * Misaligned accesses are too much trouble to fix
1849 * up; also, they usually indicate a page is not used
1850 * as a page table.
1851 *
1852 * If we're seeing too many writes to a page,
1853 * it may no longer be a page table, or we may be
1854 * forking, in which case it is better to unmap the
1855 * page.
1856 */
1857 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1858 gpa, bytes, sp->role.word);
1859 kvm_mmu_zap_page(vcpu->kvm, sp);
1860 ++vcpu->kvm->stat.mmu_flooded;
1861 continue;
1862 }
1863 page_offset = offset;
1864 level = sp->role.level;
1865 npte = 1;
1866 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1867 page_offset <<= 1; /* 32->64 */
1868 /*
1869 * A 32-bit pde maps 4MB while the shadow pdes map
1870 * only 2MB. So we need to double the offset again
1871 * and zap two pdes instead of one.
1872 */
1873 if (level == PT32_ROOT_LEVEL) {
1874 page_offset &= ~7; /* kill rounding error */
1875 page_offset <<= 1;
1876 npte = 2;
1877 }
1878 quadrant = page_offset >> PAGE_SHIFT;
1879 page_offset &= ~PAGE_MASK;
1880 if (quadrant != sp->role.quadrant)
1881 continue;
1882 }
1883 spte = &sp->spt[page_offset / sizeof(*spte)];
1884 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
1885 gentry = 0;
1886 r = kvm_read_guest_atomic(vcpu->kvm,
1887 gpa & ~(u64)(pte_size - 1),
1888 &gentry, pte_size);
1889 new = (const void *)&gentry;
1890 if (r < 0)
1891 new = NULL;
1892 }
1893 while (npte--) {
1894 entry = *spte;
1895 mmu_pte_write_zap_pte(vcpu, sp, spte);
1896 if (new)
1897 mmu_pte_write_new_pte(vcpu, sp, spte, new);
1898 mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1899 ++spte;
1900 }
1901 }
1902 kvm_mmu_audit(vcpu, "post pte write");
1903 spin_unlock(&vcpu->kvm->mmu_lock);
1904 if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
1905 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
1906 vcpu->arch.update_pte.pfn = bad_pfn;
1907 }
1908 }
1909
1910 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1911 {
1912 gpa_t gpa;
1913 int r;
1914
1915 gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1916
1917 spin_lock(&vcpu->kvm->mmu_lock);
1918 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1919 spin_unlock(&vcpu->kvm->mmu_lock);
1920 return r;
1921 }
1922 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
1923
1924 void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1925 {
1926 while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
1927 struct kvm_mmu_page *sp;
1928
1929 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
1930 struct kvm_mmu_page, link);
1931 kvm_mmu_zap_page(vcpu->kvm, sp);
1932 ++vcpu->kvm->stat.mmu_recycled;
1933 }
1934 }
1935
1936 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1937 {
1938 int r;
1939 enum emulation_result er;
1940
1941 r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
1942 if (r < 0)
1943 goto out;
1944
1945 if (!r) {
1946 r = 1;
1947 goto out;
1948 }
1949
1950 r = mmu_topup_memory_caches(vcpu);
1951 if (r)
1952 goto out;
1953
1954 er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1955
1956 switch (er) {
1957 case EMULATE_DONE:
1958 return 1;
1959 case EMULATE_DO_MMIO:
1960 ++vcpu->stat.mmio_exits;
1961 return 0;
1962 case EMULATE_FAIL:
1963 kvm_report_emulation_failure(vcpu, "pagetable");
1964 return 1;
1965 default:
1966 BUG();
1967 }
1968 out:
1969 return r;
1970 }
1971 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1972
1973 void kvm_enable_tdp(void)
1974 {
1975 tdp_enabled = true;
1976 }
1977 EXPORT_SYMBOL_GPL(kvm_enable_tdp);
1978
1979 void kvm_disable_tdp(void)
1980 {
1981 tdp_enabled = false;
1982 }
1983 EXPORT_SYMBOL_GPL(kvm_disable_tdp);
1984
1985 static void free_mmu_pages(struct kvm_vcpu *vcpu)
1986 {
1987 struct kvm_mmu_page *sp;
1988
1989 while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
1990 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
1991 struct kvm_mmu_page, link);
1992 kvm_mmu_zap_page(vcpu->kvm, sp);
1993 cond_resched();
1994 }
1995 free_page((unsigned long)vcpu->arch.mmu.pae_root);
1996 }
1997
1998 static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1999 {
2000 struct page *page;
2001 int i;
2002
2003 ASSERT(vcpu);
2004
2005 if (vcpu->kvm->arch.n_requested_mmu_pages)
2006 vcpu->kvm->arch.n_free_mmu_pages =
2007 vcpu->kvm->arch.n_requested_mmu_pages;
2008 else
2009 vcpu->kvm->arch.n_free_mmu_pages =
2010 vcpu->kvm->arch.n_alloc_mmu_pages;
2011 /*
2012 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2013 * Therefore we need to allocate shadow page tables in the first
2014 * 4GB of memory, which happens to fit the DMA32 zone.
2015 */
2016 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
2017 if (!page)
2018 goto error_1;
2019 vcpu->arch.mmu.pae_root = page_address(page);
2020 for (i = 0; i < 4; ++i)
2021 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
2022
2023 return 0;
2024
2025 error_1:
2026 free_mmu_pages(vcpu);
2027 return -ENOMEM;
2028 }
2029
2030 int kvm_mmu_create(struct kvm_vcpu *vcpu)
2031 {
2032 ASSERT(vcpu);
2033 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2034
2035 return alloc_mmu_pages(vcpu);
2036 }
2037
2038 int kvm_mmu_setup(struct kvm_vcpu *vcpu)
2039 {
2040 ASSERT(vcpu);
2041 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
2042
2043 return init_kvm_mmu(vcpu);
2044 }
2045
2046 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
2047 {
2048 ASSERT(vcpu);
2049
2050 destroy_kvm_mmu(vcpu);
2051 free_mmu_pages(vcpu);
2052 mmu_free_memory_caches(vcpu);
2053 }
2054
2055 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
2056 {
2057 struct kvm_mmu_page *sp;
2058
2059 list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
2060 int i;
2061 u64 *pt;
2062
2063 if (!test_bit(slot, &sp->slot_bitmap))
2064 continue;
2065
2066 pt = sp->spt;
2067 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2068 /* avoid RMW */
2069 if (pt[i] & PT_WRITABLE_MASK)
2070 pt[i] &= ~PT_WRITABLE_MASK;
2071 }
2072 }
2073
2074 void kvm_mmu_zap_all(struct kvm *kvm)
2075 {
2076 struct kvm_mmu_page *sp, *node;
2077
2078 spin_lock(&kvm->mmu_lock);
2079 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
2080 kvm_mmu_zap_page(kvm, sp);
2081 spin_unlock(&kvm->mmu_lock);
2082
2083 kvm_flush_remote_tlbs(kvm);
2084 }
2085
2086 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
2087 {
2088 struct kvm_mmu_page *page;
2089
2090 page = container_of(kvm->arch.active_mmu_pages.prev,
2091 struct kvm_mmu_page, link);
2092 kvm_mmu_zap_page(kvm, page);
2093 }
2094
2095 static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
2096 {
2097 struct kvm *kvm;
2098 struct kvm *kvm_freed = NULL;
2099 int cache_count = 0;
2100
2101 spin_lock(&kvm_lock);
2102
2103 list_for_each_entry(kvm, &vm_list, vm_list) {
2104 int npages;
2105
2106 if (!down_read_trylock(&kvm->slots_lock))
2107 continue;
2108 spin_lock(&kvm->mmu_lock);
2109 npages = kvm->arch.n_alloc_mmu_pages -
2110 kvm->arch.n_free_mmu_pages;
2111 cache_count += npages;
2112 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
2113 kvm_mmu_remove_one_alloc_mmu_page(kvm);
2114 cache_count--;
2115 kvm_freed = kvm;
2116 }
2117 nr_to_scan--;
2118
2119 spin_unlock(&kvm->mmu_lock);
2120 up_read(&kvm->slots_lock);
2121 }
2122 if (kvm_freed)
2123 list_move_tail(&kvm_freed->vm_list, &vm_list);
2124
2125 spin_unlock(&kvm_lock);
2126
2127 return cache_count;
2128 }
2129
2130 static struct shrinker mmu_shrinker = {
2131 .shrink = mmu_shrink,
2132 .seeks = DEFAULT_SEEKS * 10,
2133 };
2134
2135 static void mmu_destroy_caches(void)
2136 {
2137 if (pte_chain_cache)
2138 kmem_cache_destroy(pte_chain_cache);
2139 if (rmap_desc_cache)
2140 kmem_cache_destroy(rmap_desc_cache);
2141 if (mmu_page_header_cache)
2142 kmem_cache_destroy(mmu_page_header_cache);
2143 }
2144
2145 void kvm_mmu_module_exit(void)
2146 {
2147 mmu_destroy_caches();
2148 unregister_shrinker(&mmu_shrinker);
2149 }
2150
2151 int kvm_mmu_module_init(void)
2152 {
2153 pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2154 sizeof(struct kvm_pte_chain),
2155 0, 0, NULL);
2156 if (!pte_chain_cache)
2157 goto nomem;
2158 rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2159 sizeof(struct kvm_rmap_desc),
2160 0, 0, NULL);
2161 if (!rmap_desc_cache)
2162 goto nomem;
2163
2164 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2165 sizeof(struct kvm_mmu_page),
2166 0, 0, NULL);
2167 if (!mmu_page_header_cache)
2168 goto nomem;
2169
2170 register_shrinker(&mmu_shrinker);
2171
2172 return 0;
2173
2174 nomem:
2175 mmu_destroy_caches();
2176 return -ENOMEM;
2177 }
2178
2179 /*
2180 * Caculate mmu pages needed for kvm.
2181 */
2182 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2183 {
2184 int i;
2185 unsigned int nr_mmu_pages;
2186 unsigned int nr_pages = 0;
2187
2188 for (i = 0; i < kvm->nmemslots; i++)
2189 nr_pages += kvm->memslots[i].npages;
2190
2191 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2192 nr_mmu_pages = max(nr_mmu_pages,
2193 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2194
2195 return nr_mmu_pages;
2196 }
2197
2198 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2199 unsigned len)
2200 {
2201 if (len > buffer->len)
2202 return NULL;
2203 return buffer->ptr;
2204 }
2205
2206 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2207 unsigned len)
2208 {
2209 void *ret;
2210
2211 ret = pv_mmu_peek_buffer(buffer, len);
2212 if (!ret)
2213 return ret;
2214 buffer->ptr += len;
2215 buffer->len -= len;
2216 buffer->processed += len;
2217 return ret;
2218 }
2219
2220 static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2221 gpa_t addr, gpa_t value)
2222 {
2223 int bytes = 8;
2224 int r;
2225
2226 if (!is_long_mode(vcpu) && !is_pae(vcpu))
2227 bytes = 4;
2228
2229 r = mmu_topup_memory_caches(vcpu);
2230 if (r)
2231 return r;
2232
2233 if (!emulator_write_phys(vcpu, addr, &value, bytes))
2234 return -EFAULT;
2235
2236 return 1;
2237 }
2238
2239 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2240 {
2241 kvm_x86_ops->tlb_flush(vcpu);
2242 return 1;
2243 }
2244
2245 static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2246 {
2247 spin_lock(&vcpu->kvm->mmu_lock);
2248 mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2249 spin_unlock(&vcpu->kvm->mmu_lock);
2250 return 1;
2251 }
2252
2253 static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2254 struct kvm_pv_mmu_op_buffer *buffer)
2255 {
2256 struct kvm_mmu_op_header *header;
2257
2258 header = pv_mmu_peek_buffer(buffer, sizeof *header);
2259 if (!header)
2260 return 0;
2261 switch (header->op) {
2262 case KVM_MMU_OP_WRITE_PTE: {
2263 struct kvm_mmu_op_write_pte *wpte;
2264
2265 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2266 if (!wpte)
2267 return 0;
2268 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2269 wpte->pte_val);
2270 }
2271 case KVM_MMU_OP_FLUSH_TLB: {
2272 struct kvm_mmu_op_flush_tlb *ftlb;
2273
2274 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2275 if (!ftlb)
2276 return 0;
2277 return kvm_pv_mmu_flush_tlb(vcpu);
2278 }
2279 case KVM_MMU_OP_RELEASE_PT: {
2280 struct kvm_mmu_op_release_pt *rpt;
2281
2282 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2283 if (!rpt)
2284 return 0;
2285 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2286 }
2287 default: return 0;
2288 }
2289 }
2290
2291 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2292 gpa_t addr, unsigned long *ret)
2293 {
2294 int r;
2295 struct kvm_pv_mmu_op_buffer buffer;
2296
2297 buffer.ptr = buffer.buf;
2298 buffer.len = min_t(unsigned long, bytes, sizeof buffer.buf);
2299 buffer.processed = 0;
2300
2301 r = kvm_read_guest(vcpu->kvm, addr, buffer.buf, buffer.len);
2302 if (r)
2303 goto out;
2304
2305 while (buffer.len) {
2306 r = kvm_pv_mmu_op_one(vcpu, &buffer);
2307 if (r < 0)
2308 goto out;
2309 if (r == 0)
2310 break;
2311 }
2312
2313 r = 1;
2314 out:
2315 *ret = buffer.processed;
2316 return r;
2317 }
2318
2319 #ifdef AUDIT
2320
2321 static const char *audit_msg;
2322
2323 static gva_t canonicalize(gva_t gva)
2324 {
2325 #ifdef CONFIG_X86_64
2326 gva = (long long)(gva << 16) >> 16;
2327 #endif
2328 return gva;
2329 }
2330
2331 static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2332 gva_t va, int level)
2333 {
2334 u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2335 int i;
2336 gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2337
2338 for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2339 u64 ent = pt[i];
2340
2341 if (ent == shadow_trap_nonpresent_pte)
2342 continue;
2343
2344 va = canonicalize(va);
2345 if (level > 1) {
2346 if (ent == shadow_notrap_nonpresent_pte)
2347 printk(KERN_ERR "audit: (%s) nontrapping pte"
2348 " in nonleaf level: levels %d gva %lx"
2349 " level %d pte %llx\n", audit_msg,
2350 vcpu->arch.mmu.root_level, va, level, ent);
2351
2352 audit_mappings_page(vcpu, ent, va, level - 1);
2353 } else {
2354 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2355 hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2356
2357 if (is_shadow_present_pte(ent)
2358 && (ent & PT64_BASE_ADDR_MASK) != hpa)
2359 printk(KERN_ERR "xx audit error: (%s) levels %d"
2360 " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2361 audit_msg, vcpu->arch.mmu.root_level,
2362 va, gpa, hpa, ent,
2363 is_shadow_present_pte(ent));
2364 else if (ent == shadow_notrap_nonpresent_pte
2365 && !is_error_hpa(hpa))
2366 printk(KERN_ERR "audit: (%s) notrap shadow,"
2367 " valid guest gva %lx\n", audit_msg, va);
2368 kvm_release_pfn_clean(pfn);
2369
2370 }
2371 }
2372 }
2373
2374 static void audit_mappings(struct kvm_vcpu *vcpu)
2375 {
2376 unsigned i;
2377
2378 if (vcpu->arch.mmu.root_level == 4)
2379 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2380 else
2381 for (i = 0; i < 4; ++i)
2382 if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2383 audit_mappings_page(vcpu,
2384 vcpu->arch.mmu.pae_root[i],
2385 i << 30,
2386 2);
2387 }
2388
2389 static int count_rmaps(struct kvm_vcpu *vcpu)
2390 {
2391 int nmaps = 0;
2392 int i, j, k;
2393
2394 for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
2395 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
2396 struct kvm_rmap_desc *d;
2397
2398 for (j = 0; j < m->npages; ++j) {
2399 unsigned long *rmapp = &m->rmap[j];
2400
2401 if (!*rmapp)
2402 continue;
2403 if (!(*rmapp & 1)) {
2404 ++nmaps;
2405 continue;
2406 }
2407 d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
2408 while (d) {
2409 for (k = 0; k < RMAP_EXT; ++k)
2410 if (d->shadow_ptes[k])
2411 ++nmaps;
2412 else
2413 break;
2414 d = d->more;
2415 }
2416 }
2417 }
2418 return nmaps;
2419 }
2420
2421 static int count_writable_mappings(struct kvm_vcpu *vcpu)
2422 {
2423 int nmaps = 0;
2424 struct kvm_mmu_page *sp;
2425 int i;
2426
2427 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2428 u64 *pt = sp->spt;
2429
2430 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
2431 continue;
2432
2433 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
2434 u64 ent = pt[i];
2435
2436 if (!(ent & PT_PRESENT_MASK))
2437 continue;
2438 if (!(ent & PT_WRITABLE_MASK))
2439 continue;
2440 ++nmaps;
2441 }
2442 }
2443 return nmaps;
2444 }
2445
2446 static void audit_rmap(struct kvm_vcpu *vcpu)
2447 {
2448 int n_rmap = count_rmaps(vcpu);
2449 int n_actual = count_writable_mappings(vcpu);
2450
2451 if (n_rmap != n_actual)
2452 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
2453 __func__, audit_msg, n_rmap, n_actual);
2454 }
2455
2456 static void audit_write_protection(struct kvm_vcpu *vcpu)
2457 {
2458 struct kvm_mmu_page *sp;
2459 struct kvm_memory_slot *slot;
2460 unsigned long *rmapp;
2461 gfn_t gfn;
2462
2463 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2464 if (sp->role.metaphysical)
2465 continue;
2466
2467 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
2468 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
2469 rmapp = &slot->rmap[gfn - slot->base_gfn];
2470 if (*rmapp)
2471 printk(KERN_ERR "%s: (%s) shadow page has writable"
2472 " mappings: gfn %lx role %x\n",
2473 __func__, audit_msg, sp->gfn,
2474 sp->role.word);
2475 }
2476 }
2477
2478 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
2479 {
2480 int olddbg = dbg;
2481
2482 dbg = 0;
2483 audit_msg = msg;
2484 audit_rmap(vcpu);
2485 audit_write_protection(vcpu);
2486 audit_mappings(vcpu);
2487 dbg = olddbg;
2488 }
2489
2490 #endif
Linux kernel - Deliverables
This page took 0.081226 seconds and 6 git commands to generate.