projects / linux-block.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
KVM: x86 emulator: lazily evaluate segment registers
[linux-block.git] / arch / x86 / kvm / mmu.c
... / ...
CommitLineData
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 */
44bool tdp_enabled = false;
45
46#undef MMU_DEBUG
47
48#undef AUDIT
49
50#ifdef AUDIT
51static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg);
52#else
53static 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)
69static int dbg = 1;
70#endif
71
72#ifndef MMU_DEBUG
73#define ASSERT(x) do { } while (0)
74#else
75#define ASSERT(x) \
76 if (!(x)) { \
77 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
78 __FILE__, __LINE__, #x); \
79 }
80#endif
81
82#define PT_FIRST_AVAIL_BITS_SHIFT 9
83#define PT64_SECOND_AVAIL_BITS_SHIFT 52
84
85#define VALID_PAGE(x) ((x) != INVALID_PAGE)
86
87#define PT64_LEVEL_BITS 9
88
89#define PT64_LEVEL_SHIFT(level) \
90 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
91
92#define PT64_LEVEL_MASK(level) \
93 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
94
95#define PT64_INDEX(address, level)\
96 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
97
98
99#define PT32_LEVEL_BITS 10
100
101#define PT32_LEVEL_SHIFT(level) \
102 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
103
104#define PT32_LEVEL_MASK(level) \
105 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
106
107#define PT32_INDEX(address, level)\
108 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
109
110
111#define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
112#define PT64_DIR_BASE_ADDR_MASK \
113 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
114
115#define PT32_BASE_ADDR_MASK PAGE_MASK
116#define PT32_DIR_BASE_ADDR_MASK \
117 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
118
119#define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
120 | PT64_NX_MASK)
121
122#define PFERR_PRESENT_MASK (1U << 0)
123#define PFERR_WRITE_MASK (1U << 1)
124#define PFERR_USER_MASK (1U << 2)
125#define PFERR_FETCH_MASK (1U << 4)
126
127#define PT_DIRECTORY_LEVEL 2
128#define PT_PAGE_TABLE_LEVEL 1
129
130#define RMAP_EXT 4
131
132#define ACC_EXEC_MASK 1
133#define ACC_WRITE_MASK PT_WRITABLE_MASK
134#define ACC_USER_MASK PT_USER_MASK
135#define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
136
137struct kvm_pv_mmu_op_buffer {
138 void *ptr;
139 unsigned len;
140 unsigned processed;
141 char buf[512] __aligned(sizeof(long));
142};
143
144struct kvm_rmap_desc {
145 u64 *shadow_ptes[RMAP_EXT];
146 struct kvm_rmap_desc *more;
147};
148
149static struct kmem_cache *pte_chain_cache;
150static struct kmem_cache *rmap_desc_cache;
151static struct kmem_cache *mmu_page_header_cache;
152
153static u64 __read_mostly shadow_trap_nonpresent_pte;
154static u64 __read_mostly shadow_notrap_nonpresent_pte;
155static u64 __read_mostly shadow_base_present_pte;
156static u64 __read_mostly shadow_nx_mask;
157static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
158static u64 __read_mostly shadow_user_mask;
159static u64 __read_mostly shadow_accessed_mask;
160static u64 __read_mostly shadow_dirty_mask;
161
162void kvm_mmu_set_nonpresent_ptes(u64 trap_pte, u64 notrap_pte)
163{
164 shadow_trap_nonpresent_pte = trap_pte;
165 shadow_notrap_nonpresent_pte = notrap_pte;
166}
167EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes);
168
169void kvm_mmu_set_base_ptes(u64 base_pte)
170{
171 shadow_base_present_pte = base_pte;
172}
173EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes);
174
175void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
176 u64 dirty_mask, u64 nx_mask, u64 x_mask)
177{
178 shadow_user_mask = user_mask;
179 shadow_accessed_mask = accessed_mask;
180 shadow_dirty_mask = dirty_mask;
181 shadow_nx_mask = nx_mask;
182 shadow_x_mask = x_mask;
183}
184EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
185
186static int is_write_protection(struct kvm_vcpu *vcpu)
187{
188 return vcpu->arch.cr0 & X86_CR0_WP;
189}
190
191static int is_cpuid_PSE36(void)
192{
193 return 1;
194}
195
196static int is_nx(struct kvm_vcpu *vcpu)
197{
198 return vcpu->arch.shadow_efer & EFER_NX;
199}
200
201static int is_present_pte(unsigned long pte)
202{
203 return pte & PT_PRESENT_MASK;
204}
205
206static int is_shadow_present_pte(u64 pte)
207{
208 return pte != shadow_trap_nonpresent_pte
209 && pte != shadow_notrap_nonpresent_pte;
210}
211
212static int is_large_pte(u64 pte)
213{
214 return pte & PT_PAGE_SIZE_MASK;
215}
216
217static int is_writeble_pte(unsigned long pte)
218{
219 return pte & PT_WRITABLE_MASK;
220}
221
222static int is_dirty_pte(unsigned long pte)
223{
224 return pte & shadow_dirty_mask;
225}
226
227static int is_rmap_pte(u64 pte)
228{
229 return is_shadow_present_pte(pte);
230}
231
232static pfn_t spte_to_pfn(u64 pte)
233{
234 return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
235}
236
237static gfn_t pse36_gfn_delta(u32 gpte)
238{
239 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
240
241 return (gpte & PT32_DIR_PSE36_MASK) << shift;
242}
243
244static void set_shadow_pte(u64 *sptep, u64 spte)
245{
246#ifdef CONFIG_X86_64
247 set_64bit((unsigned long *)sptep, spte);
248#else
249 set_64bit((unsigned long long *)sptep, spte);
250#endif
251}
252
253static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
254 struct kmem_cache *base_cache, int min)
255{
256 void *obj;
257
258 if (cache->nobjs >= min)
259 return 0;
260 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
261 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL);
262 if (!obj)
263 return -ENOMEM;
264 cache->objects[cache->nobjs++] = obj;
265 }
266 return 0;
267}
268
269static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
270{
271 while (mc->nobjs)
272 kfree(mc->objects[--mc->nobjs]);
273}
274
275static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
276 int min)
277{
278 struct page *page;
279
280 if (cache->nobjs >= min)
281 return 0;
282 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
283 page = alloc_page(GFP_KERNEL);
284 if (!page)
285 return -ENOMEM;
286 set_page_private(page, 0);
287 cache->objects[cache->nobjs++] = page_address(page);
288 }
289 return 0;
290}
291
292static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
293{
294 while (mc->nobjs)
295 free_page((unsigned long)mc->objects[--mc->nobjs]);
296}
297
298static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
299{
300 int r;
301
302 r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_chain_cache,
303 pte_chain_cache, 4);
304 if (r)
305 goto out;
306 r = mmu_topup_memory_cache(&vcpu->arch.mmu_rmap_desc_cache,
307 rmap_desc_cache, 1);
308 if (r)
309 goto out;
310 r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
311 if (r)
312 goto out;
313 r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
314 mmu_page_header_cache, 4);
315out:
316 return r;
317}
318
319static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
320{
321 mmu_free_memory_cache(&vcpu->arch.mmu_pte_chain_cache);
322 mmu_free_memory_cache(&vcpu->arch.mmu_rmap_desc_cache);
323 mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
324 mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
325}
326
327static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc,
328 size_t size)
329{
330 void *p;
331
332 BUG_ON(!mc->nobjs);
333 p = mc->objects[--mc->nobjs];
334 memset(p, 0, size);
335 return p;
336}
337
338static struct kvm_pte_chain *mmu_alloc_pte_chain(struct kvm_vcpu *vcpu)
339{
340 return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_chain_cache,
341 sizeof(struct kvm_pte_chain));
342}
343
344static void mmu_free_pte_chain(struct kvm_pte_chain *pc)
345{
346 kfree(pc);
347}
348
349static struct kvm_rmap_desc *mmu_alloc_rmap_desc(struct kvm_vcpu *vcpu)
350{
351 return mmu_memory_cache_alloc(&vcpu->arch.mmu_rmap_desc_cache,
352 sizeof(struct kvm_rmap_desc));
353}
354
355static void mmu_free_rmap_desc(struct kvm_rmap_desc *rd)
356{
357 kfree(rd);
358}
359
360/*
361 * Return the pointer to the largepage write count for a given
362 * gfn, handling slots that are not large page aligned.
363 */
364static int *slot_largepage_idx(gfn_t gfn, struct kvm_memory_slot *slot)
365{
366 unsigned long idx;
367
368 idx = (gfn / KVM_PAGES_PER_HPAGE) -
369 (slot->base_gfn / KVM_PAGES_PER_HPAGE);
370 return &slot->lpage_info[idx].write_count;
371}
372
373static void account_shadowed(struct kvm *kvm, gfn_t gfn)
374{
375 int *write_count;
376
377 write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
378 *write_count += 1;
379}
380
381static void unaccount_shadowed(struct kvm *kvm, gfn_t gfn)
382{
383 int *write_count;
384
385 write_count = slot_largepage_idx(gfn, gfn_to_memslot(kvm, gfn));
386 *write_count -= 1;
387 WARN_ON(*write_count < 0);
388}
389
390static int has_wrprotected_page(struct kvm *kvm, gfn_t gfn)
391{
392 struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
393 int *largepage_idx;
394
395 if (slot) {
396 largepage_idx = slot_largepage_idx(gfn, slot);
397 return *largepage_idx;
398 }
399
400 return 1;
401}
402
403static int host_largepage_backed(struct kvm *kvm, gfn_t gfn)
404{
405 struct vm_area_struct *vma;
406 unsigned long addr;
407
408 addr = gfn_to_hva(kvm, gfn);
409 if (kvm_is_error_hva(addr))
410 return 0;
411
412 vma = find_vma(current->mm, addr);
413 if (vma && is_vm_hugetlb_page(vma))
414 return 1;
415
416 return 0;
417}
418
419static int is_largepage_backed(struct kvm_vcpu *vcpu, gfn_t large_gfn)
420{
421 struct kvm_memory_slot *slot;
422
423 if (has_wrprotected_page(vcpu->kvm, large_gfn))
424 return 0;
425
426 if (!host_largepage_backed(vcpu->kvm, large_gfn))
427 return 0;
428
429 slot = gfn_to_memslot(vcpu->kvm, large_gfn);
430 if (slot && slot->dirty_bitmap)
431 return 0;
432
433 return 1;
434}
435
436/*
437 * Take gfn and return the reverse mapping to it.
438 * Note: gfn must be unaliased before this function get called
439 */
440
441static unsigned long *gfn_to_rmap(struct kvm *kvm, gfn_t gfn, int lpage)
442{
443 struct kvm_memory_slot *slot;
444 unsigned long idx;
445
446 slot = gfn_to_memslot(kvm, gfn);
447 if (!lpage)
448 return &slot->rmap[gfn - slot->base_gfn];
449
450 idx = (gfn / KVM_PAGES_PER_HPAGE) -
451 (slot->base_gfn / KVM_PAGES_PER_HPAGE);
452
453 return &slot->lpage_info[idx].rmap_pde;
454}
455
456/*
457 * Reverse mapping data structures:
458 *
459 * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
460 * that points to page_address(page).
461 *
462 * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
463 * containing more mappings.
464 */
465static void rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn, int lpage)
466{
467 struct kvm_mmu_page *sp;
468 struct kvm_rmap_desc *desc;
469 unsigned long *rmapp;
470 int i;
471
472 if (!is_rmap_pte(*spte))
473 return;
474 gfn = unalias_gfn(vcpu->kvm, gfn);
475 sp = page_header(__pa(spte));
476 sp->gfns[spte - sp->spt] = gfn;
477 rmapp = gfn_to_rmap(vcpu->kvm, gfn, lpage);
478 if (!*rmapp) {
479 rmap_printk("rmap_add: %p %llx 0->1\n", spte, *spte);
480 *rmapp = (unsigned long)spte;
481 } else if (!(*rmapp & 1)) {
482 rmap_printk("rmap_add: %p %llx 1->many\n", spte, *spte);
483 desc = mmu_alloc_rmap_desc(vcpu);
484 desc->shadow_ptes[0] = (u64 *)*rmapp;
485 desc->shadow_ptes[1] = spte;
486 *rmapp = (unsigned long)desc | 1;
487 } else {
488 rmap_printk("rmap_add: %p %llx many->many\n", spte, *spte);
489 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
490 while (desc->shadow_ptes[RMAP_EXT-1] && desc->more)
491 desc = desc->more;
492 if (desc->shadow_ptes[RMAP_EXT-1]) {
493 desc->more = mmu_alloc_rmap_desc(vcpu);
494 desc = desc->more;
495 }
496 for (i = 0; desc->shadow_ptes[i]; ++i)
497 ;
498 desc->shadow_ptes[i] = spte;
499 }
500}
501
502static void rmap_desc_remove_entry(unsigned long *rmapp,
503 struct kvm_rmap_desc *desc,
504 int i,
505 struct kvm_rmap_desc *prev_desc)
506{
507 int j;
508
509 for (j = RMAP_EXT - 1; !desc->shadow_ptes[j] && j > i; --j)
510 ;
511 desc->shadow_ptes[i] = desc->shadow_ptes[j];
512 desc->shadow_ptes[j] = NULL;
513 if (j != 0)
514 return;
515 if (!prev_desc && !desc->more)
516 *rmapp = (unsigned long)desc->shadow_ptes[0];
517 else
518 if (prev_desc)
519 prev_desc->more = desc->more;
520 else
521 *rmapp = (unsigned long)desc->more | 1;
522 mmu_free_rmap_desc(desc);
523}
524
525static void rmap_remove(struct kvm *kvm, u64 *spte)
526{
527 struct kvm_rmap_desc *desc;
528 struct kvm_rmap_desc *prev_desc;
529 struct kvm_mmu_page *sp;
530 pfn_t pfn;
531 unsigned long *rmapp;
532 int i;
533
534 if (!is_rmap_pte(*spte))
535 return;
536 sp = page_header(__pa(spte));
537 pfn = spte_to_pfn(*spte);
538 if (*spte & shadow_accessed_mask)
539 kvm_set_pfn_accessed(pfn);
540 if (is_writeble_pte(*spte))
541 kvm_release_pfn_dirty(pfn);
542 else
543 kvm_release_pfn_clean(pfn);
544 rmapp = gfn_to_rmap(kvm, sp->gfns[spte - sp->spt], is_large_pte(*spte));
545 if (!*rmapp) {
546 printk(KERN_ERR "rmap_remove: %p %llx 0->BUG\n", spte, *spte);
547 BUG();
548 } else if (!(*rmapp & 1)) {
549 rmap_printk("rmap_remove: %p %llx 1->0\n", spte, *spte);
550 if ((u64 *)*rmapp != spte) {
551 printk(KERN_ERR "rmap_remove: %p %llx 1->BUG\n",
552 spte, *spte);
553 BUG();
554 }
555 *rmapp = 0;
556 } else {
557 rmap_printk("rmap_remove: %p %llx many->many\n", spte, *spte);
558 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
559 prev_desc = NULL;
560 while (desc) {
561 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i)
562 if (desc->shadow_ptes[i] == spte) {
563 rmap_desc_remove_entry(rmapp,
564 desc, i,
565 prev_desc);
566 return;
567 }
568 prev_desc = desc;
569 desc = desc->more;
570 }
571 BUG();
572 }
573}
574
575static u64 *rmap_next(struct kvm *kvm, unsigned long *rmapp, u64 *spte)
576{
577 struct kvm_rmap_desc *desc;
578 struct kvm_rmap_desc *prev_desc;
579 u64 *prev_spte;
580 int i;
581
582 if (!*rmapp)
583 return NULL;
584 else if (!(*rmapp & 1)) {
585 if (!spte)
586 return (u64 *)*rmapp;
587 return NULL;
588 }
589 desc = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
590 prev_desc = NULL;
591 prev_spte = NULL;
592 while (desc) {
593 for (i = 0; i < RMAP_EXT && desc->shadow_ptes[i]; ++i) {
594 if (prev_spte == spte)
595 return desc->shadow_ptes[i];
596 prev_spte = desc->shadow_ptes[i];
597 }
598 desc = desc->more;
599 }
600 return NULL;
601}
602
603static void rmap_write_protect(struct kvm *kvm, u64 gfn)
604{
605 unsigned long *rmapp;
606 u64 *spte;
607 int write_protected = 0;
608
609 gfn = unalias_gfn(kvm, gfn);
610 rmapp = gfn_to_rmap(kvm, gfn, 0);
611
612 spte = rmap_next(kvm, rmapp, NULL);
613 while (spte) {
614 BUG_ON(!spte);
615 BUG_ON(!(*spte & PT_PRESENT_MASK));
616 rmap_printk("rmap_write_protect: spte %p %llx\n", spte, *spte);
617 if (is_writeble_pte(*spte)) {
618 set_shadow_pte(spte, *spte & ~PT_WRITABLE_MASK);
619 write_protected = 1;
620 }
621 spte = rmap_next(kvm, rmapp, spte);
622 }
623 if (write_protected) {
624 pfn_t pfn;
625
626 spte = rmap_next(kvm, rmapp, NULL);
627 pfn = spte_to_pfn(*spte);
628 kvm_set_pfn_dirty(pfn);
629 }
630
631 /* check for huge page mappings */
632 rmapp = gfn_to_rmap(kvm, gfn, 1);
633 spte = rmap_next(kvm, rmapp, NULL);
634 while (spte) {
635 BUG_ON(!spte);
636 BUG_ON(!(*spte & PT_PRESENT_MASK));
637 BUG_ON((*spte & (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK)) != (PT_PAGE_SIZE_MASK|PT_PRESENT_MASK));
638 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte, *spte, gfn);
639 if (is_writeble_pte(*spte)) {
640 rmap_remove(kvm, spte);
641 --kvm->stat.lpages;
642 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
643 spte = NULL;
644 write_protected = 1;
645 }
646 spte = rmap_next(kvm, rmapp, spte);
647 }
648
649 if (write_protected)
650 kvm_flush_remote_tlbs(kvm);
651
652 account_shadowed(kvm, gfn);
653}
654
655#ifdef MMU_DEBUG
656static int is_empty_shadow_page(u64 *spt)
657{
658 u64 *pos;
659 u64 *end;
660
661 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
662 if (is_shadow_present_pte(*pos)) {
663 printk(KERN_ERR "%s: %p %llx\n", __func__,
664 pos, *pos);
665 return 0;
666 }
667 return 1;
668}
669#endif
670
671static void kvm_mmu_free_page(struct kvm *kvm, struct kvm_mmu_page *sp)
672{
673 ASSERT(is_empty_shadow_page(sp->spt));
674 list_del(&sp->link);
675 __free_page(virt_to_page(sp->spt));
676 __free_page(virt_to_page(sp->gfns));
677 kfree(sp);
678 ++kvm->arch.n_free_mmu_pages;
679}
680
681static unsigned kvm_page_table_hashfn(gfn_t gfn)
682{
683 return gfn & ((1 << KVM_MMU_HASH_SHIFT) - 1);
684}
685
686static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu,
687 u64 *parent_pte)
688{
689 struct kvm_mmu_page *sp;
690
691 sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache, sizeof *sp);
692 sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
693 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache, PAGE_SIZE);
694 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
695 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
696 ASSERT(is_empty_shadow_page(sp->spt));
697 sp->slot_bitmap = 0;
698 sp->multimapped = 0;
699 sp->parent_pte = parent_pte;
700 --vcpu->kvm->arch.n_free_mmu_pages;
701 return sp;
702}
703
704static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
705 struct kvm_mmu_page *sp, u64 *parent_pte)
706{
707 struct kvm_pte_chain *pte_chain;
708 struct hlist_node *node;
709 int i;
710
711 if (!parent_pte)
712 return;
713 if (!sp->multimapped) {
714 u64 *old = sp->parent_pte;
715
716 if (!old) {
717 sp->parent_pte = parent_pte;
718 return;
719 }
720 sp->multimapped = 1;
721 pte_chain = mmu_alloc_pte_chain(vcpu);
722 INIT_HLIST_HEAD(&sp->parent_ptes);
723 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
724 pte_chain->parent_ptes[0] = old;
725 }
726 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link) {
727 if (pte_chain->parent_ptes[NR_PTE_CHAIN_ENTRIES-1])
728 continue;
729 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i)
730 if (!pte_chain->parent_ptes[i]) {
731 pte_chain->parent_ptes[i] = parent_pte;
732 return;
733 }
734 }
735 pte_chain = mmu_alloc_pte_chain(vcpu);
736 BUG_ON(!pte_chain);
737 hlist_add_head(&pte_chain->link, &sp->parent_ptes);
738 pte_chain->parent_ptes[0] = parent_pte;
739}
740
741static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
742 u64 *parent_pte)
743{
744 struct kvm_pte_chain *pte_chain;
745 struct hlist_node *node;
746 int i;
747
748 if (!sp->multimapped) {
749 BUG_ON(sp->parent_pte != parent_pte);
750 sp->parent_pte = NULL;
751 return;
752 }
753 hlist_for_each_entry(pte_chain, node, &sp->parent_ptes, link)
754 for (i = 0; i < NR_PTE_CHAIN_ENTRIES; ++i) {
755 if (!pte_chain->parent_ptes[i])
756 break;
757 if (pte_chain->parent_ptes[i] != parent_pte)
758 continue;
759 while (i + 1 < NR_PTE_CHAIN_ENTRIES
760 && pte_chain->parent_ptes[i + 1]) {
761 pte_chain->parent_ptes[i]
762 = pte_chain->parent_ptes[i + 1];
763 ++i;
764 }
765 pte_chain->parent_ptes[i] = NULL;
766 if (i == 0) {
767 hlist_del(&pte_chain->link);
768 mmu_free_pte_chain(pte_chain);
769 if (hlist_empty(&sp->parent_ptes)) {
770 sp->multimapped = 0;
771 sp->parent_pte = NULL;
772 }
773 }
774 return;
775 }
776 BUG();
777}
778
779static void nonpaging_prefetch_page(struct kvm_vcpu *vcpu,
780 struct kvm_mmu_page *sp)
781{
782 int i;
783
784 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
785 sp->spt[i] = shadow_trap_nonpresent_pte;
786}
787
788static struct kvm_mmu_page *kvm_mmu_lookup_page(struct kvm *kvm, gfn_t gfn)
789{
790 unsigned index;
791 struct hlist_head *bucket;
792 struct kvm_mmu_page *sp;
793 struct hlist_node *node;
794
795 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
796 index = kvm_page_table_hashfn(gfn);
797 bucket = &kvm->arch.mmu_page_hash[index];
798 hlist_for_each_entry(sp, node, bucket, hash_link)
799 if (sp->gfn == gfn && !sp->role.metaphysical
800 && !sp->role.invalid) {
801 pgprintk("%s: found role %x\n",
802 __func__, sp->role.word);
803 return sp;
804 }
805 return NULL;
806}
807
808static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
809 gfn_t gfn,
810 gva_t gaddr,
811 unsigned level,
812 int metaphysical,
813 unsigned access,
814 u64 *parent_pte)
815{
816 union kvm_mmu_page_role role;
817 unsigned index;
818 unsigned quadrant;
819 struct hlist_head *bucket;
820 struct kvm_mmu_page *sp;
821 struct hlist_node *node;
822
823 role.word = 0;
824 role.glevels = vcpu->arch.mmu.root_level;
825 role.level = level;
826 role.metaphysical = metaphysical;
827 role.access = access;
828 if (vcpu->arch.mmu.root_level <= PT32_ROOT_LEVEL) {
829 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
830 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
831 role.quadrant = quadrant;
832 }
833 pgprintk("%s: looking gfn %lx role %x\n", __func__,
834 gfn, role.word);
835 index = kvm_page_table_hashfn(gfn);
836 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
837 hlist_for_each_entry(sp, node, bucket, hash_link)
838 if (sp->gfn == gfn && sp->role.word == role.word) {
839 mmu_page_add_parent_pte(vcpu, sp, parent_pte);
840 pgprintk("%s: found\n", __func__);
841 return sp;
842 }
843 ++vcpu->kvm->stat.mmu_cache_miss;
844 sp = kvm_mmu_alloc_page(vcpu, parent_pte);
845 if (!sp)
846 return sp;
847 pgprintk("%s: adding gfn %lx role %x\n", __func__, gfn, role.word);
848 sp->gfn = gfn;
849 sp->role = role;
850 hlist_add_head(&sp->hash_link, bucket);
851 if (!metaphysical)
852 rmap_write_protect(vcpu->kvm, gfn);
853 if (shadow_trap_nonpresent_pte != shadow_notrap_nonpresent_pte)
854 vcpu->arch.mmu.prefetch_page(vcpu, sp);
855 else
856 nonpaging_prefetch_page(vcpu, sp);
857 return sp;
858}
859
860static void kvm_mmu_page_unlink_children(struct kvm *kvm,
861 struct kvm_mmu_page *sp)
862{
863 unsigned i;
864 u64 *pt;
865 u64 ent;
866
867 pt = sp->spt;
868
869 if (sp->role.level == PT_PAGE_TABLE_LEVEL) {
870 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
871 if (is_shadow_present_pte(pt[i]))
872 rmap_remove(kvm, &pt[i]);
873 pt[i] = shadow_trap_nonpresent_pte;
874 }
875 kvm_flush_remote_tlbs(kvm);
876 return;
877 }
878
879 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
880 ent = pt[i];
881
882 if (is_shadow_present_pte(ent)) {
883 if (!is_large_pte(ent)) {
884 ent &= PT64_BASE_ADDR_MASK;
885 mmu_page_remove_parent_pte(page_header(ent),
886 &pt[i]);
887 } else {
888 --kvm->stat.lpages;
889 rmap_remove(kvm, &pt[i]);
890 }
891 }
892 pt[i] = shadow_trap_nonpresent_pte;
893 }
894 kvm_flush_remote_tlbs(kvm);
895}
896
897static void kvm_mmu_put_page(struct kvm_mmu_page *sp, u64 *parent_pte)
898{
899 mmu_page_remove_parent_pte(sp, parent_pte);
900}
901
902static void kvm_mmu_reset_last_pte_updated(struct kvm *kvm)
903{
904 int i;
905
906 for (i = 0; i < KVM_MAX_VCPUS; ++i)
907 if (kvm->vcpus[i])
908 kvm->vcpus[i]->arch.last_pte_updated = NULL;
909}
910
911static void kvm_mmu_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp)
912{
913 u64 *parent_pte;
914
915 ++kvm->stat.mmu_shadow_zapped;
916 while (sp->multimapped || sp->parent_pte) {
917 if (!sp->multimapped)
918 parent_pte = sp->parent_pte;
919 else {
920 struct kvm_pte_chain *chain;
921
922 chain = container_of(sp->parent_ptes.first,
923 struct kvm_pte_chain, link);
924 parent_pte = chain->parent_ptes[0];
925 }
926 BUG_ON(!parent_pte);
927 kvm_mmu_put_page(sp, parent_pte);
928 set_shadow_pte(parent_pte, shadow_trap_nonpresent_pte);
929 }
930 kvm_mmu_page_unlink_children(kvm, sp);
931 if (!sp->root_count) {
932 if (!sp->role.metaphysical)
933 unaccount_shadowed(kvm, sp->gfn);
934 hlist_del(&sp->hash_link);
935 kvm_mmu_free_page(kvm, sp);
936 } else {
937 list_move(&sp->link, &kvm->arch.active_mmu_pages);
938 sp->role.invalid = 1;
939 kvm_reload_remote_mmus(kvm);
940 }
941 kvm_mmu_reset_last_pte_updated(kvm);
942}
943
944/*
945 * Changing the number of mmu pages allocated to the vm
946 * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
947 */
948void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned int kvm_nr_mmu_pages)
949{
950 /*
951 * If we set the number of mmu pages to be smaller be than the
952 * number of actived pages , we must to free some mmu pages before we
953 * change the value
954 */
955
956 if ((kvm->arch.n_alloc_mmu_pages - kvm->arch.n_free_mmu_pages) >
957 kvm_nr_mmu_pages) {
958 int n_used_mmu_pages = kvm->arch.n_alloc_mmu_pages
959 - kvm->arch.n_free_mmu_pages;
960
961 while (n_used_mmu_pages > kvm_nr_mmu_pages) {
962 struct kvm_mmu_page *page;
963
964 page = container_of(kvm->arch.active_mmu_pages.prev,
965 struct kvm_mmu_page, link);
966 kvm_mmu_zap_page(kvm, page);
967 n_used_mmu_pages--;
968 }
969 kvm->arch.n_free_mmu_pages = 0;
970 }
971 else
972 kvm->arch.n_free_mmu_pages += kvm_nr_mmu_pages
973 - kvm->arch.n_alloc_mmu_pages;
974
975 kvm->arch.n_alloc_mmu_pages = kvm_nr_mmu_pages;
976}
977
978static int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
979{
980 unsigned index;
981 struct hlist_head *bucket;
982 struct kvm_mmu_page *sp;
983 struct hlist_node *node, *n;
984 int r;
985
986 pgprintk("%s: looking for gfn %lx\n", __func__, gfn);
987 r = 0;
988 index = kvm_page_table_hashfn(gfn);
989 bucket = &kvm->arch.mmu_page_hash[index];
990 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link)
991 if (sp->gfn == gfn && !sp->role.metaphysical) {
992 pgprintk("%s: gfn %lx role %x\n", __func__, gfn,
993 sp->role.word);
994 kvm_mmu_zap_page(kvm, sp);
995 r = 1;
996 }
997 return r;
998}
999
1000static void mmu_unshadow(struct kvm *kvm, gfn_t gfn)
1001{
1002 struct kvm_mmu_page *sp;
1003
1004 while ((sp = kvm_mmu_lookup_page(kvm, gfn)) != NULL) {
1005 pgprintk("%s: zap %lx %x\n", __func__, gfn, sp->role.word);
1006 kvm_mmu_zap_page(kvm, sp);
1007 }
1008}
1009
1010static void page_header_update_slot(struct kvm *kvm, void *pte, gfn_t gfn)
1011{
1012 int slot = memslot_id(kvm, gfn_to_memslot(kvm, gfn));
1013 struct kvm_mmu_page *sp = page_header(__pa(pte));
1014
1015 __set_bit(slot, &sp->slot_bitmap);
1016}
1017
1018struct page *gva_to_page(struct kvm_vcpu *vcpu, gva_t gva)
1019{
1020 struct page *page;
1021
1022 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1023
1024 if (gpa == UNMAPPED_GVA)
1025 return NULL;
1026
1027 down_read(&current->mm->mmap_sem);
1028 page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1029 up_read(&current->mm->mmap_sem);
1030
1031 return page;
1032}
1033
1034static void mmu_set_spte(struct kvm_vcpu *vcpu, u64 *shadow_pte,
1035 unsigned pt_access, unsigned pte_access,
1036 int user_fault, int write_fault, int dirty,
1037 int *ptwrite, int largepage, gfn_t gfn,
1038 pfn_t pfn, bool speculative)
1039{
1040 u64 spte;
1041 int was_rmapped = 0;
1042 int was_writeble = is_writeble_pte(*shadow_pte);
1043
1044 pgprintk("%s: spte %llx access %x write_fault %d"
1045 " user_fault %d gfn %lx\n",
1046 __func__, *shadow_pte, pt_access,
1047 write_fault, user_fault, gfn);
1048
1049 if (is_rmap_pte(*shadow_pte)) {
1050 /*
1051 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1052 * the parent of the now unreachable PTE.
1053 */
1054 if (largepage && !is_large_pte(*shadow_pte)) {
1055 struct kvm_mmu_page *child;
1056 u64 pte = *shadow_pte;
1057
1058 child = page_header(pte & PT64_BASE_ADDR_MASK);
1059 mmu_page_remove_parent_pte(child, shadow_pte);
1060 } else if (pfn != spte_to_pfn(*shadow_pte)) {
1061 pgprintk("hfn old %lx new %lx\n",
1062 spte_to_pfn(*shadow_pte), pfn);
1063 rmap_remove(vcpu->kvm, shadow_pte);
1064 } else {
1065 if (largepage)
1066 was_rmapped = is_large_pte(*shadow_pte);
1067 else
1068 was_rmapped = 1;
1069 }
1070 }
1071
1072 /*
1073 * We don't set the accessed bit, since we sometimes want to see
1074 * whether the guest actually used the pte (in order to detect
1075 * demand paging).
1076 */
1077 spte = shadow_base_present_pte | shadow_dirty_mask;
1078 if (!speculative)
1079 pte_access |= PT_ACCESSED_MASK;
1080 if (!dirty)
1081 pte_access &= ~ACC_WRITE_MASK;
1082 if (pte_access & ACC_EXEC_MASK)
1083 spte |= shadow_x_mask;
1084 else
1085 spte |= shadow_nx_mask;
1086 if (pte_access & ACC_USER_MASK)
1087 spte |= shadow_user_mask;
1088 if (largepage)
1089 spte |= PT_PAGE_SIZE_MASK;
1090
1091 spte |= (u64)pfn << PAGE_SHIFT;
1092
1093 if ((pte_access & ACC_WRITE_MASK)
1094 || (write_fault && !is_write_protection(vcpu) && !user_fault)) {
1095 struct kvm_mmu_page *shadow;
1096
1097 spte |= PT_WRITABLE_MASK;
1098
1099 shadow = kvm_mmu_lookup_page(vcpu->kvm, gfn);
1100 if (shadow ||
1101 (largepage && has_wrprotected_page(vcpu->kvm, gfn))) {
1102 pgprintk("%s: found shadow page for %lx, marking ro\n",
1103 __func__, gfn);
1104 pte_access &= ~ACC_WRITE_MASK;
1105 if (is_writeble_pte(spte)) {
1106 spte &= ~PT_WRITABLE_MASK;
1107 kvm_x86_ops->tlb_flush(vcpu);
1108 }
1109 if (write_fault)
1110 *ptwrite = 1;
1111 }
1112 }
1113
1114 if (pte_access & ACC_WRITE_MASK)
1115 mark_page_dirty(vcpu->kvm, gfn);
1116
1117 pgprintk("%s: setting spte %llx\n", __func__, spte);
1118 pgprintk("instantiating %s PTE (%s) at %d (%llx) addr %llx\n",
1119 (spte&PT_PAGE_SIZE_MASK)? "2MB" : "4kB",
1120 (spte&PT_WRITABLE_MASK)?"RW":"R", gfn, spte, shadow_pte);
1121 set_shadow_pte(shadow_pte, spte);
1122 if (!was_rmapped && (spte & PT_PAGE_SIZE_MASK)
1123 && (spte & PT_PRESENT_MASK))
1124 ++vcpu->kvm->stat.lpages;
1125
1126 page_header_update_slot(vcpu->kvm, shadow_pte, gfn);
1127 if (!was_rmapped) {
1128 rmap_add(vcpu, shadow_pte, gfn, largepage);
1129 if (!is_rmap_pte(*shadow_pte))
1130 kvm_release_pfn_clean(pfn);
1131 } else {
1132 if (was_writeble)
1133 kvm_release_pfn_dirty(pfn);
1134 else
1135 kvm_release_pfn_clean(pfn);
1136 }
1137 if (speculative) {
1138 vcpu->arch.last_pte_updated = shadow_pte;
1139 vcpu->arch.last_pte_gfn = gfn;
1140 }
1141}
1142
1143static void nonpaging_new_cr3(struct kvm_vcpu *vcpu)
1144{
1145}
1146
1147static int __direct_map(struct kvm_vcpu *vcpu, gpa_t v, int write,
1148 int largepage, gfn_t gfn, pfn_t pfn,
1149 int level)
1150{
1151 hpa_t table_addr = vcpu->arch.mmu.root_hpa;
1152 int pt_write = 0;
1153
1154 for (; ; level--) {
1155 u32 index = PT64_INDEX(v, level);
1156 u64 *table;
1157
1158 ASSERT(VALID_PAGE(table_addr));
1159 table = __va(table_addr);
1160
1161 if (level == 1) {
1162 mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1163 0, write, 1, &pt_write, 0, gfn, pfn, false);
1164 return pt_write;
1165 }
1166
1167 if (largepage && level == 2) {
1168 mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
1169 0, write, 1, &pt_write, 1, gfn, pfn, false);
1170 return pt_write;
1171 }
1172
1173 if (table[index] == shadow_trap_nonpresent_pte) {
1174 struct kvm_mmu_page *new_table;
1175 gfn_t pseudo_gfn;
1176
1177 pseudo_gfn = (v & PT64_DIR_BASE_ADDR_MASK)
1178 >> PAGE_SHIFT;
1179 new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
1180 v, level - 1,
1181 1, ACC_ALL, &table[index]);
1182 if (!new_table) {
1183 pgprintk("nonpaging_map: ENOMEM\n");
1184 kvm_release_pfn_clean(pfn);
1185 return -ENOMEM;
1186 }
1187
1188 table[index] = __pa(new_table->spt)
1189 | PT_PRESENT_MASK | PT_WRITABLE_MASK
1190 | shadow_user_mask | shadow_x_mask;
1191 }
1192 table_addr = table[index] & PT64_BASE_ADDR_MASK;
1193 }
1194}
1195
1196static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
1197{
1198 int r;
1199 int largepage = 0;
1200 pfn_t pfn;
1201
1202 down_read(&current->mm->mmap_sem);
1203 if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1204 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1205 largepage = 1;
1206 }
1207
1208 pfn = gfn_to_pfn(vcpu->kvm, gfn);
1209 up_read(&current->mm->mmap_sem);
1210
1211 /* mmio */
1212 if (is_error_pfn(pfn)) {
1213 kvm_release_pfn_clean(pfn);
1214 return 1;
1215 }
1216
1217 spin_lock(&vcpu->kvm->mmu_lock);
1218 kvm_mmu_free_some_pages(vcpu);
1219 r = __direct_map(vcpu, v, write, largepage, gfn, pfn,
1220 PT32E_ROOT_LEVEL);
1221 spin_unlock(&vcpu->kvm->mmu_lock);
1222
1223
1224 return r;
1225}
1226
1227
1228static void mmu_free_roots(struct kvm_vcpu *vcpu)
1229{
1230 int i;
1231 struct kvm_mmu_page *sp;
1232
1233 if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
1234 return;
1235 spin_lock(&vcpu->kvm->mmu_lock);
1236 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1237 hpa_t root = vcpu->arch.mmu.root_hpa;
1238
1239 sp = page_header(root);
1240 --sp->root_count;
1241 if (!sp->root_count && sp->role.invalid)
1242 kvm_mmu_zap_page(vcpu->kvm, sp);
1243 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1244 spin_unlock(&vcpu->kvm->mmu_lock);
1245 return;
1246 }
1247 for (i = 0; i < 4; ++i) {
1248 hpa_t root = vcpu->arch.mmu.pae_root[i];
1249
1250 if (root) {
1251 root &= PT64_BASE_ADDR_MASK;
1252 sp = page_header(root);
1253 --sp->root_count;
1254 if (!sp->root_count && sp->role.invalid)
1255 kvm_mmu_zap_page(vcpu->kvm, sp);
1256 }
1257 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1258 }
1259 spin_unlock(&vcpu->kvm->mmu_lock);
1260 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1261}
1262
1263static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
1264{
1265 int i;
1266 gfn_t root_gfn;
1267 struct kvm_mmu_page *sp;
1268 int metaphysical = 0;
1269
1270 root_gfn = vcpu->arch.cr3 >> PAGE_SHIFT;
1271
1272 if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
1273 hpa_t root = vcpu->arch.mmu.root_hpa;
1274
1275 ASSERT(!VALID_PAGE(root));
1276 if (tdp_enabled)
1277 metaphysical = 1;
1278 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
1279 PT64_ROOT_LEVEL, metaphysical,
1280 ACC_ALL, NULL);
1281 root = __pa(sp->spt);
1282 ++sp->root_count;
1283 vcpu->arch.mmu.root_hpa = root;
1284 return;
1285 }
1286 metaphysical = !is_paging(vcpu);
1287 if (tdp_enabled)
1288 metaphysical = 1;
1289 for (i = 0; i < 4; ++i) {
1290 hpa_t root = vcpu->arch.mmu.pae_root[i];
1291
1292 ASSERT(!VALID_PAGE(root));
1293 if (vcpu->arch.mmu.root_level == PT32E_ROOT_LEVEL) {
1294 if (!is_present_pte(vcpu->arch.pdptrs[i])) {
1295 vcpu->arch.mmu.pae_root[i] = 0;
1296 continue;
1297 }
1298 root_gfn = vcpu->arch.pdptrs[i] >> PAGE_SHIFT;
1299 } else if (vcpu->arch.mmu.root_level == 0)
1300 root_gfn = 0;
1301 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
1302 PT32_ROOT_LEVEL, metaphysical,
1303 ACC_ALL, NULL);
1304 root = __pa(sp->spt);
1305 ++sp->root_count;
1306 vcpu->arch.mmu.pae_root[i] = root | PT_PRESENT_MASK;
1307 }
1308 vcpu->arch.mmu.root_hpa = __pa(vcpu->arch.mmu.pae_root);
1309}
1310
1311static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gva_t vaddr)
1312{
1313 return vaddr;
1314}
1315
1316static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
1317 u32 error_code)
1318{
1319 gfn_t gfn;
1320 int r;
1321
1322 pgprintk("%s: gva %lx error %x\n", __func__, gva, error_code);
1323 r = mmu_topup_memory_caches(vcpu);
1324 if (r)
1325 return r;
1326
1327 ASSERT(vcpu);
1328 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1329
1330 gfn = gva >> PAGE_SHIFT;
1331
1332 return nonpaging_map(vcpu, gva & PAGE_MASK,
1333 error_code & PFERR_WRITE_MASK, gfn);
1334}
1335
1336static int tdp_page_fault(struct kvm_vcpu *vcpu, gva_t gpa,
1337 u32 error_code)
1338{
1339 pfn_t pfn;
1340 int r;
1341 int largepage = 0;
1342 gfn_t gfn = gpa >> PAGE_SHIFT;
1343
1344 ASSERT(vcpu);
1345 ASSERT(VALID_PAGE(vcpu->arch.mmu.root_hpa));
1346
1347 r = mmu_topup_memory_caches(vcpu);
1348 if (r)
1349 return r;
1350
1351 down_read(&current->mm->mmap_sem);
1352 if (is_largepage_backed(vcpu, gfn & ~(KVM_PAGES_PER_HPAGE-1))) {
1353 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1354 largepage = 1;
1355 }
1356 pfn = gfn_to_pfn(vcpu->kvm, gfn);
1357 up_read(&current->mm->mmap_sem);
1358 if (is_error_pfn(pfn)) {
1359 kvm_release_pfn_clean(pfn);
1360 return 1;
1361 }
1362 spin_lock(&vcpu->kvm->mmu_lock);
1363 kvm_mmu_free_some_pages(vcpu);
1364 r = __direct_map(vcpu, gpa, error_code & PFERR_WRITE_MASK,
1365 largepage, gfn, pfn, kvm_x86_ops->get_tdp_level());
1366 spin_unlock(&vcpu->kvm->mmu_lock);
1367
1368 return r;
1369}
1370
1371static void nonpaging_free(struct kvm_vcpu *vcpu)
1372{
1373 mmu_free_roots(vcpu);
1374}
1375
1376static int nonpaging_init_context(struct kvm_vcpu *vcpu)
1377{
1378 struct kvm_mmu *context = &vcpu->arch.mmu;
1379
1380 context->new_cr3 = nonpaging_new_cr3;
1381 context->page_fault = nonpaging_page_fault;
1382 context->gva_to_gpa = nonpaging_gva_to_gpa;
1383 context->free = nonpaging_free;
1384 context->prefetch_page = nonpaging_prefetch_page;
1385 context->root_level = 0;
1386 context->shadow_root_level = PT32E_ROOT_LEVEL;
1387 context->root_hpa = INVALID_PAGE;
1388 return 0;
1389}
1390
1391void kvm_mmu_flush_tlb(struct kvm_vcpu *vcpu)
1392{
1393 ++vcpu->stat.tlb_flush;
1394 kvm_x86_ops->tlb_flush(vcpu);
1395}
1396
1397static void paging_new_cr3(struct kvm_vcpu *vcpu)
1398{
1399 pgprintk("%s: cr3 %lx\n", __func__, vcpu->arch.cr3);
1400 mmu_free_roots(vcpu);
1401}
1402
1403static void inject_page_fault(struct kvm_vcpu *vcpu,
1404 u64 addr,
1405 u32 err_code)
1406{
1407 kvm_inject_page_fault(vcpu, addr, err_code);
1408}
1409
1410static void paging_free(struct kvm_vcpu *vcpu)
1411{
1412 nonpaging_free(vcpu);
1413}
1414
1415#define PTTYPE 64
1416#include "paging_tmpl.h"
1417#undef PTTYPE
1418
1419#define PTTYPE 32
1420#include "paging_tmpl.h"
1421#undef PTTYPE
1422
1423static int paging64_init_context_common(struct kvm_vcpu *vcpu, int level)
1424{
1425 struct kvm_mmu *context = &vcpu->arch.mmu;
1426
1427 ASSERT(is_pae(vcpu));
1428 context->new_cr3 = paging_new_cr3;
1429 context->page_fault = paging64_page_fault;
1430 context->gva_to_gpa = paging64_gva_to_gpa;
1431 context->prefetch_page = paging64_prefetch_page;
1432 context->free = paging_free;
1433 context->root_level = level;
1434 context->shadow_root_level = level;
1435 context->root_hpa = INVALID_PAGE;
1436 return 0;
1437}
1438
1439static int paging64_init_context(struct kvm_vcpu *vcpu)
1440{
1441 return paging64_init_context_common(vcpu, PT64_ROOT_LEVEL);
1442}
1443
1444static int paging32_init_context(struct kvm_vcpu *vcpu)
1445{
1446 struct kvm_mmu *context = &vcpu->arch.mmu;
1447
1448 context->new_cr3 = paging_new_cr3;
1449 context->page_fault = paging32_page_fault;
1450 context->gva_to_gpa = paging32_gva_to_gpa;
1451 context->free = paging_free;
1452 context->prefetch_page = paging32_prefetch_page;
1453 context->root_level = PT32_ROOT_LEVEL;
1454 context->shadow_root_level = PT32E_ROOT_LEVEL;
1455 context->root_hpa = INVALID_PAGE;
1456 return 0;
1457}
1458
1459static int paging32E_init_context(struct kvm_vcpu *vcpu)
1460{
1461 return paging64_init_context_common(vcpu, PT32E_ROOT_LEVEL);
1462}
1463
1464static int init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
1465{
1466 struct kvm_mmu *context = &vcpu->arch.mmu;
1467
1468 context->new_cr3 = nonpaging_new_cr3;
1469 context->page_fault = tdp_page_fault;
1470 context->free = nonpaging_free;
1471 context->prefetch_page = nonpaging_prefetch_page;
1472 context->shadow_root_level = kvm_x86_ops->get_tdp_level();
1473 context->root_hpa = INVALID_PAGE;
1474
1475 if (!is_paging(vcpu)) {
1476 context->gva_to_gpa = nonpaging_gva_to_gpa;
1477 context->root_level = 0;
1478 } else if (is_long_mode(vcpu)) {
1479 context->gva_to_gpa = paging64_gva_to_gpa;
1480 context->root_level = PT64_ROOT_LEVEL;
1481 } else if (is_pae(vcpu)) {
1482 context->gva_to_gpa = paging64_gva_to_gpa;
1483 context->root_level = PT32E_ROOT_LEVEL;
1484 } else {
1485 context->gva_to_gpa = paging32_gva_to_gpa;
1486 context->root_level = PT32_ROOT_LEVEL;
1487 }
1488
1489 return 0;
1490}
1491
1492static int init_kvm_softmmu(struct kvm_vcpu *vcpu)
1493{
1494 ASSERT(vcpu);
1495 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1496
1497 if (!is_paging(vcpu))
1498 return nonpaging_init_context(vcpu);
1499 else if (is_long_mode(vcpu))
1500 return paging64_init_context(vcpu);
1501 else if (is_pae(vcpu))
1502 return paging32E_init_context(vcpu);
1503 else
1504 return paging32_init_context(vcpu);
1505}
1506
1507static int init_kvm_mmu(struct kvm_vcpu *vcpu)
1508{
1509 vcpu->arch.update_pte.pfn = bad_pfn;
1510
1511 if (tdp_enabled)
1512 return init_kvm_tdp_mmu(vcpu);
1513 else
1514 return init_kvm_softmmu(vcpu);
1515}
1516
1517static void destroy_kvm_mmu(struct kvm_vcpu *vcpu)
1518{
1519 ASSERT(vcpu);
1520 if (VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
1521 vcpu->arch.mmu.free(vcpu);
1522 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
1523 }
1524}
1525
1526int kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
1527{
1528 destroy_kvm_mmu(vcpu);
1529 return init_kvm_mmu(vcpu);
1530}
1531EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
1532
1533int kvm_mmu_load(struct kvm_vcpu *vcpu)
1534{
1535 int r;
1536
1537 r = mmu_topup_memory_caches(vcpu);
1538 if (r)
1539 goto out;
1540 spin_lock(&vcpu->kvm->mmu_lock);
1541 kvm_mmu_free_some_pages(vcpu);
1542 mmu_alloc_roots(vcpu);
1543 spin_unlock(&vcpu->kvm->mmu_lock);
1544 kvm_x86_ops->set_cr3(vcpu, vcpu->arch.mmu.root_hpa);
1545 kvm_mmu_flush_tlb(vcpu);
1546out:
1547 return r;
1548}
1549EXPORT_SYMBOL_GPL(kvm_mmu_load);
1550
1551void kvm_mmu_unload(struct kvm_vcpu *vcpu)
1552{
1553 mmu_free_roots(vcpu);
1554}
1555
1556static void mmu_pte_write_zap_pte(struct kvm_vcpu *vcpu,
1557 struct kvm_mmu_page *sp,
1558 u64 *spte)
1559{
1560 u64 pte;
1561 struct kvm_mmu_page *child;
1562
1563 pte = *spte;
1564 if (is_shadow_present_pte(pte)) {
1565 if (sp->role.level == PT_PAGE_TABLE_LEVEL ||
1566 is_large_pte(pte))
1567 rmap_remove(vcpu->kvm, spte);
1568 else {
1569 child = page_header(pte & PT64_BASE_ADDR_MASK);
1570 mmu_page_remove_parent_pte(child, spte);
1571 }
1572 }
1573 set_shadow_pte(spte, shadow_trap_nonpresent_pte);
1574 if (is_large_pte(pte))
1575 --vcpu->kvm->stat.lpages;
1576}
1577
1578static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
1579 struct kvm_mmu_page *sp,
1580 u64 *spte,
1581 const void *new)
1582{
1583 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
1584 if (!vcpu->arch.update_pte.largepage ||
1585 sp->role.glevels == PT32_ROOT_LEVEL) {
1586 ++vcpu->kvm->stat.mmu_pde_zapped;
1587 return;
1588 }
1589 }
1590
1591 ++vcpu->kvm->stat.mmu_pte_updated;
1592 if (sp->role.glevels == PT32_ROOT_LEVEL)
1593 paging32_update_pte(vcpu, sp, spte, new);
1594 else
1595 paging64_update_pte(vcpu, sp, spte, new);
1596}
1597
1598static bool need_remote_flush(u64 old, u64 new)
1599{
1600 if (!is_shadow_present_pte(old))
1601 return false;
1602 if (!is_shadow_present_pte(new))
1603 return true;
1604 if ((old ^ new) & PT64_BASE_ADDR_MASK)
1605 return true;
1606 old ^= PT64_NX_MASK;
1607 new ^= PT64_NX_MASK;
1608 return (old & ~new & PT64_PERM_MASK) != 0;
1609}
1610
1611static void mmu_pte_write_flush_tlb(struct kvm_vcpu *vcpu, u64 old, u64 new)
1612{
1613 if (need_remote_flush(old, new))
1614 kvm_flush_remote_tlbs(vcpu->kvm);
1615 else
1616 kvm_mmu_flush_tlb(vcpu);
1617}
1618
1619static bool last_updated_pte_accessed(struct kvm_vcpu *vcpu)
1620{
1621 u64 *spte = vcpu->arch.last_pte_updated;
1622
1623 return !!(spte && (*spte & shadow_accessed_mask));
1624}
1625
1626static void mmu_guess_page_from_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1627 const u8 *new, int bytes)
1628{
1629 gfn_t gfn;
1630 int r;
1631 u64 gpte = 0;
1632 pfn_t pfn;
1633
1634 vcpu->arch.update_pte.largepage = 0;
1635
1636 if (bytes != 4 && bytes != 8)
1637 return;
1638
1639 /*
1640 * Assume that the pte write on a page table of the same type
1641 * as the current vcpu paging mode. This is nearly always true
1642 * (might be false while changing modes). Note it is verified later
1643 * by update_pte().
1644 */
1645 if (is_pae(vcpu)) {
1646 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
1647 if ((bytes == 4) && (gpa % 4 == 0)) {
1648 r = kvm_read_guest(vcpu->kvm, gpa & ~(u64)7, &gpte, 8);
1649 if (r)
1650 return;
1651 memcpy((void *)&gpte + (gpa % 8), new, 4);
1652 } else if ((bytes == 8) && (gpa % 8 == 0)) {
1653 memcpy((void *)&gpte, new, 8);
1654 }
1655 } else {
1656 if ((bytes == 4) && (gpa % 4 == 0))
1657 memcpy((void *)&gpte, new, 4);
1658 }
1659 if (!is_present_pte(gpte))
1660 return;
1661 gfn = (gpte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
1662
1663 down_read(&current->mm->mmap_sem);
1664 if (is_large_pte(gpte) && is_largepage_backed(vcpu, gfn)) {
1665 gfn &= ~(KVM_PAGES_PER_HPAGE-1);
1666 vcpu->arch.update_pte.largepage = 1;
1667 }
1668 pfn = gfn_to_pfn(vcpu->kvm, gfn);
1669 up_read(&current->mm->mmap_sem);
1670
1671 if (is_error_pfn(pfn)) {
1672 kvm_release_pfn_clean(pfn);
1673 return;
1674 }
1675 vcpu->arch.update_pte.gfn = gfn;
1676 vcpu->arch.update_pte.pfn = pfn;
1677}
1678
1679static void kvm_mmu_access_page(struct kvm_vcpu *vcpu, gfn_t gfn)
1680{
1681 u64 *spte = vcpu->arch.last_pte_updated;
1682
1683 if (spte
1684 && vcpu->arch.last_pte_gfn == gfn
1685 && shadow_accessed_mask
1686 && !(*spte & shadow_accessed_mask)
1687 && is_shadow_present_pte(*spte))
1688 set_bit(PT_ACCESSED_SHIFT, (unsigned long *)spte);
1689}
1690
1691void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1692 const u8 *new, int bytes)
1693{
1694 gfn_t gfn = gpa >> PAGE_SHIFT;
1695 struct kvm_mmu_page *sp;
1696 struct hlist_node *node, *n;
1697 struct hlist_head *bucket;
1698 unsigned index;
1699 u64 entry, gentry;
1700 u64 *spte;
1701 unsigned offset = offset_in_page(gpa);
1702 unsigned pte_size;
1703 unsigned page_offset;
1704 unsigned misaligned;
1705 unsigned quadrant;
1706 int level;
1707 int flooded = 0;
1708 int npte;
1709 int r;
1710
1711 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
1712 mmu_guess_page_from_pte_write(vcpu, gpa, new, bytes);
1713 spin_lock(&vcpu->kvm->mmu_lock);
1714 kvm_mmu_access_page(vcpu, gfn);
1715 kvm_mmu_free_some_pages(vcpu);
1716 ++vcpu->kvm->stat.mmu_pte_write;
1717 kvm_mmu_audit(vcpu, "pre pte write");
1718 if (gfn == vcpu->arch.last_pt_write_gfn
1719 && !last_updated_pte_accessed(vcpu)) {
1720 ++vcpu->arch.last_pt_write_count;
1721 if (vcpu->arch.last_pt_write_count >= 3)
1722 flooded = 1;
1723 } else {
1724 vcpu->arch.last_pt_write_gfn = gfn;
1725 vcpu->arch.last_pt_write_count = 1;
1726 vcpu->arch.last_pte_updated = NULL;
1727 }
1728 index = kvm_page_table_hashfn(gfn);
1729 bucket = &vcpu->kvm->arch.mmu_page_hash[index];
1730 hlist_for_each_entry_safe(sp, node, n, bucket, hash_link) {
1731 if (sp->gfn != gfn || sp->role.metaphysical)
1732 continue;
1733 pte_size = sp->role.glevels == PT32_ROOT_LEVEL ? 4 : 8;
1734 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
1735 misaligned |= bytes < 4;
1736 if (misaligned || flooded) {
1737 /*
1738 * Misaligned accesses are too much trouble to fix
1739 * up; also, they usually indicate a page is not used
1740 * as a page table.
1741 *
1742 * If we're seeing too many writes to a page,
1743 * it may no longer be a page table, or we may be
1744 * forking, in which case it is better to unmap the
1745 * page.
1746 */
1747 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
1748 gpa, bytes, sp->role.word);
1749 kvm_mmu_zap_page(vcpu->kvm, sp);
1750 ++vcpu->kvm->stat.mmu_flooded;
1751 continue;
1752 }
1753 page_offset = offset;
1754 level = sp->role.level;
1755 npte = 1;
1756 if (sp->role.glevels == PT32_ROOT_LEVEL) {
1757 page_offset <<= 1; /* 32->64 */
1758 /*
1759 * A 32-bit pde maps 4MB while the shadow pdes map
1760 * only 2MB. So we need to double the offset again
1761 * and zap two pdes instead of one.
1762 */
1763 if (level == PT32_ROOT_LEVEL) {
1764 page_offset &= ~7; /* kill rounding error */
1765 page_offset <<= 1;
1766 npte = 2;
1767 }
1768 quadrant = page_offset >> PAGE_SHIFT;
1769 page_offset &= ~PAGE_MASK;
1770 if (quadrant != sp->role.quadrant)
1771 continue;
1772 }
1773 spte = &sp->spt[page_offset / sizeof(*spte)];
1774 if ((gpa & (pte_size - 1)) || (bytes < pte_size)) {
1775 gentry = 0;
1776 r = kvm_read_guest_atomic(vcpu->kvm,
1777 gpa & ~(u64)(pte_size - 1),
1778 &gentry, pte_size);
1779 new = (const void *)&gentry;
1780 if (r < 0)
1781 new = NULL;
1782 }
1783 while (npte--) {
1784 entry = *spte;
1785 mmu_pte_write_zap_pte(vcpu, sp, spte);
1786 if (new)
1787 mmu_pte_write_new_pte(vcpu, sp, spte, new);
1788 mmu_pte_write_flush_tlb(vcpu, entry, *spte);
1789 ++spte;
1790 }
1791 }
1792 kvm_mmu_audit(vcpu, "post pte write");
1793 spin_unlock(&vcpu->kvm->mmu_lock);
1794 if (!is_error_pfn(vcpu->arch.update_pte.pfn)) {
1795 kvm_release_pfn_clean(vcpu->arch.update_pte.pfn);
1796 vcpu->arch.update_pte.pfn = bad_pfn;
1797 }
1798}
1799
1800int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
1801{
1802 gpa_t gpa;
1803 int r;
1804
1805 gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, gva);
1806
1807 spin_lock(&vcpu->kvm->mmu_lock);
1808 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
1809 spin_unlock(&vcpu->kvm->mmu_lock);
1810 return r;
1811}
1812
1813void __kvm_mmu_free_some_pages(struct kvm_vcpu *vcpu)
1814{
1815 while (vcpu->kvm->arch.n_free_mmu_pages < KVM_REFILL_PAGES) {
1816 struct kvm_mmu_page *sp;
1817
1818 sp = container_of(vcpu->kvm->arch.active_mmu_pages.prev,
1819 struct kvm_mmu_page, link);
1820 kvm_mmu_zap_page(vcpu->kvm, sp);
1821 ++vcpu->kvm->stat.mmu_recycled;
1822 }
1823}
1824
1825int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u32 error_code)
1826{
1827 int r;
1828 enum emulation_result er;
1829
1830 r = vcpu->arch.mmu.page_fault(vcpu, cr2, error_code);
1831 if (r < 0)
1832 goto out;
1833
1834 if (!r) {
1835 r = 1;
1836 goto out;
1837 }
1838
1839 r = mmu_topup_memory_caches(vcpu);
1840 if (r)
1841 goto out;
1842
1843 er = emulate_instruction(vcpu, vcpu->run, cr2, error_code, 0);
1844
1845 switch (er) {
1846 case EMULATE_DONE:
1847 return 1;
1848 case EMULATE_DO_MMIO:
1849 ++vcpu->stat.mmio_exits;
1850 return 0;
1851 case EMULATE_FAIL:
1852 kvm_report_emulation_failure(vcpu, "pagetable");
1853 return 1;
1854 default:
1855 BUG();
1856 }
1857out:
1858 return r;
1859}
1860EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
1861
1862void kvm_enable_tdp(void)
1863{
1864 tdp_enabled = true;
1865}
1866EXPORT_SYMBOL_GPL(kvm_enable_tdp);
1867
1868static void free_mmu_pages(struct kvm_vcpu *vcpu)
1869{
1870 struct kvm_mmu_page *sp;
1871
1872 while (!list_empty(&vcpu->kvm->arch.active_mmu_pages)) {
1873 sp = container_of(vcpu->kvm->arch.active_mmu_pages.next,
1874 struct kvm_mmu_page, link);
1875 kvm_mmu_zap_page(vcpu->kvm, sp);
1876 cond_resched();
1877 }
1878 free_page((unsigned long)vcpu->arch.mmu.pae_root);
1879}
1880
1881static int alloc_mmu_pages(struct kvm_vcpu *vcpu)
1882{
1883 struct page *page;
1884 int i;
1885
1886 ASSERT(vcpu);
1887
1888 if (vcpu->kvm->arch.n_requested_mmu_pages)
1889 vcpu->kvm->arch.n_free_mmu_pages =
1890 vcpu->kvm->arch.n_requested_mmu_pages;
1891 else
1892 vcpu->kvm->arch.n_free_mmu_pages =
1893 vcpu->kvm->arch.n_alloc_mmu_pages;
1894 /*
1895 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
1896 * Therefore we need to allocate shadow page tables in the first
1897 * 4GB of memory, which happens to fit the DMA32 zone.
1898 */
1899 page = alloc_page(GFP_KERNEL | __GFP_DMA32);
1900 if (!page)
1901 goto error_1;
1902 vcpu->arch.mmu.pae_root = page_address(page);
1903 for (i = 0; i < 4; ++i)
1904 vcpu->arch.mmu.pae_root[i] = INVALID_PAGE;
1905
1906 return 0;
1907
1908error_1:
1909 free_mmu_pages(vcpu);
1910 return -ENOMEM;
1911}
1912
1913int kvm_mmu_create(struct kvm_vcpu *vcpu)
1914{
1915 ASSERT(vcpu);
1916 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1917
1918 return alloc_mmu_pages(vcpu);
1919}
1920
1921int kvm_mmu_setup(struct kvm_vcpu *vcpu)
1922{
1923 ASSERT(vcpu);
1924 ASSERT(!VALID_PAGE(vcpu->arch.mmu.root_hpa));
1925
1926 return init_kvm_mmu(vcpu);
1927}
1928
1929void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
1930{
1931 ASSERT(vcpu);
1932
1933 destroy_kvm_mmu(vcpu);
1934 free_mmu_pages(vcpu);
1935 mmu_free_memory_caches(vcpu);
1936}
1937
1938void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
1939{
1940 struct kvm_mmu_page *sp;
1941
1942 list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link) {
1943 int i;
1944 u64 *pt;
1945
1946 if (!test_bit(slot, &sp->slot_bitmap))
1947 continue;
1948
1949 pt = sp->spt;
1950 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
1951 /* avoid RMW */
1952 if (pt[i] & PT_WRITABLE_MASK)
1953 pt[i] &= ~PT_WRITABLE_MASK;
1954 }
1955}
1956
1957void kvm_mmu_zap_all(struct kvm *kvm)
1958{
1959 struct kvm_mmu_page *sp, *node;
1960
1961 spin_lock(&kvm->mmu_lock);
1962 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link)
1963 kvm_mmu_zap_page(kvm, sp);
1964 spin_unlock(&kvm->mmu_lock);
1965
1966 kvm_flush_remote_tlbs(kvm);
1967}
1968
1969static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm *kvm)
1970{
1971 struct kvm_mmu_page *page;
1972
1973 page = container_of(kvm->arch.active_mmu_pages.prev,
1974 struct kvm_mmu_page, link);
1975 kvm_mmu_zap_page(kvm, page);
1976}
1977
1978static int mmu_shrink(int nr_to_scan, gfp_t gfp_mask)
1979{
1980 struct kvm *kvm;
1981 struct kvm *kvm_freed = NULL;
1982 int cache_count = 0;
1983
1984 spin_lock(&kvm_lock);
1985
1986 list_for_each_entry(kvm, &vm_list, vm_list) {
1987 int npages;
1988
1989 spin_lock(&kvm->mmu_lock);
1990 npages = kvm->arch.n_alloc_mmu_pages -
1991 kvm->arch.n_free_mmu_pages;
1992 cache_count += npages;
1993 if (!kvm_freed && nr_to_scan > 0 && npages > 0) {
1994 kvm_mmu_remove_one_alloc_mmu_page(kvm);
1995 cache_count--;
1996 kvm_freed = kvm;
1997 }
1998 nr_to_scan--;
1999
2000 spin_unlock(&kvm->mmu_lock);
2001 }
2002 if (kvm_freed)
2003 list_move_tail(&kvm_freed->vm_list, &vm_list);
2004
2005 spin_unlock(&kvm_lock);
2006
2007 return cache_count;
2008}
2009
2010static struct shrinker mmu_shrinker = {
2011 .shrink = mmu_shrink,
2012 .seeks = DEFAULT_SEEKS * 10,
2013};
2014
2015static void mmu_destroy_caches(void)
2016{
2017 if (pte_chain_cache)
2018 kmem_cache_destroy(pte_chain_cache);
2019 if (rmap_desc_cache)
2020 kmem_cache_destroy(rmap_desc_cache);
2021 if (mmu_page_header_cache)
2022 kmem_cache_destroy(mmu_page_header_cache);
2023}
2024
2025void kvm_mmu_module_exit(void)
2026{
2027 mmu_destroy_caches();
2028 unregister_shrinker(&mmu_shrinker);
2029}
2030
2031int kvm_mmu_module_init(void)
2032{
2033 pte_chain_cache = kmem_cache_create("kvm_pte_chain",
2034 sizeof(struct kvm_pte_chain),
2035 0, 0, NULL);
2036 if (!pte_chain_cache)
2037 goto nomem;
2038 rmap_desc_cache = kmem_cache_create("kvm_rmap_desc",
2039 sizeof(struct kvm_rmap_desc),
2040 0, 0, NULL);
2041 if (!rmap_desc_cache)
2042 goto nomem;
2043
2044 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
2045 sizeof(struct kvm_mmu_page),
2046 0, 0, NULL);
2047 if (!mmu_page_header_cache)
2048 goto nomem;
2049
2050 register_shrinker(&mmu_shrinker);
2051
2052 return 0;
2053
2054nomem:
2055 mmu_destroy_caches();
2056 return -ENOMEM;
2057}
2058
2059/*
2060 * Caculate mmu pages needed for kvm.
2061 */
2062unsigned int kvm_mmu_calculate_mmu_pages(struct kvm *kvm)
2063{
2064 int i;
2065 unsigned int nr_mmu_pages;
2066 unsigned int nr_pages = 0;
2067
2068 for (i = 0; i < kvm->nmemslots; i++)
2069 nr_pages += kvm->memslots[i].npages;
2070
2071 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
2072 nr_mmu_pages = max(nr_mmu_pages,
2073 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES);
2074
2075 return nr_mmu_pages;
2076}
2077
2078static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2079 unsigned len)
2080{
2081 if (len > buffer->len)
2082 return NULL;
2083 return buffer->ptr;
2084}
2085
2086static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer *buffer,
2087 unsigned len)
2088{
2089 void *ret;
2090
2091 ret = pv_mmu_peek_buffer(buffer, len);
2092 if (!ret)
2093 return ret;
2094 buffer->ptr += len;
2095 buffer->len -= len;
2096 buffer->processed += len;
2097 return ret;
2098}
2099
2100static int kvm_pv_mmu_write(struct kvm_vcpu *vcpu,
2101 gpa_t addr, gpa_t value)
2102{
2103 int bytes = 8;
2104 int r;
2105
2106 if (!is_long_mode(vcpu) && !is_pae(vcpu))
2107 bytes = 4;
2108
2109 r = mmu_topup_memory_caches(vcpu);
2110 if (r)
2111 return r;
2112
2113 if (!emulator_write_phys(vcpu, addr, &value, bytes))
2114 return -EFAULT;
2115
2116 return 1;
2117}
2118
2119static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu *vcpu)
2120{
2121 kvm_x86_ops->tlb_flush(vcpu);
2122 return 1;
2123}
2124
2125static int kvm_pv_mmu_release_pt(struct kvm_vcpu *vcpu, gpa_t addr)
2126{
2127 spin_lock(&vcpu->kvm->mmu_lock);
2128 mmu_unshadow(vcpu->kvm, addr >> PAGE_SHIFT);
2129 spin_unlock(&vcpu->kvm->mmu_lock);
2130 return 1;
2131}
2132
2133static int kvm_pv_mmu_op_one(struct kvm_vcpu *vcpu,
2134 struct kvm_pv_mmu_op_buffer *buffer)
2135{
2136 struct kvm_mmu_op_header *header;
2137
2138 header = pv_mmu_peek_buffer(buffer, sizeof *header);
2139 if (!header)
2140 return 0;
2141 switch (header->op) {
2142 case KVM_MMU_OP_WRITE_PTE: {
2143 struct kvm_mmu_op_write_pte *wpte;
2144
2145 wpte = pv_mmu_read_buffer(buffer, sizeof *wpte);
2146 if (!wpte)
2147 return 0;
2148 return kvm_pv_mmu_write(vcpu, wpte->pte_phys,
2149 wpte->pte_val);
2150 }
2151 case KVM_MMU_OP_FLUSH_TLB: {
2152 struct kvm_mmu_op_flush_tlb *ftlb;
2153
2154 ftlb = pv_mmu_read_buffer(buffer, sizeof *ftlb);
2155 if (!ftlb)
2156 return 0;
2157 return kvm_pv_mmu_flush_tlb(vcpu);
2158 }
2159 case KVM_MMU_OP_RELEASE_PT: {
2160 struct kvm_mmu_op_release_pt *rpt;
2161
2162 rpt = pv_mmu_read_buffer(buffer, sizeof *rpt);
2163 if (!rpt)
2164 return 0;
2165 return kvm_pv_mmu_release_pt(vcpu, rpt->pt_phys);
2166 }
2167 default: return 0;
2168 }
2169}
2170
2171int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
2172 gpa_t addr, unsigned long *ret)
2173{
2174 int r;
2175 struct kvm_pv_mmu_op_buffer buffer;
2176
2177 buffer.ptr = buffer.buf;
2178 buffer.len = min_t(unsigned long, bytes, sizeof buffer.buf);
2179 buffer.processed = 0;
2180
2181 r = kvm_read_guest(vcpu->kvm, addr, buffer.buf, buffer.len);
2182 if (r)
2183 goto out;
2184
2185 while (buffer.len) {
2186 r = kvm_pv_mmu_op_one(vcpu, &buffer);
2187 if (r < 0)
2188 goto out;
2189 if (r == 0)
2190 break;
2191 }
2192
2193 r = 1;
2194out:
2195 *ret = buffer.processed;
2196 return r;
2197}
2198
2199#ifdef AUDIT
2200
2201static const char *audit_msg;
2202
2203static gva_t canonicalize(gva_t gva)
2204{
2205#ifdef CONFIG_X86_64
2206 gva = (long long)(gva << 16) >> 16;
2207#endif
2208 return gva;
2209}
2210
2211static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
2212 gva_t va, int level)
2213{
2214 u64 *pt = __va(page_pte & PT64_BASE_ADDR_MASK);
2215 int i;
2216 gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
2217
2218 for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
2219 u64 ent = pt[i];
2220
2221 if (ent == shadow_trap_nonpresent_pte)
2222 continue;
2223
2224 va = canonicalize(va);
2225 if (level > 1) {
2226 if (ent == shadow_notrap_nonpresent_pte)
2227 printk(KERN_ERR "audit: (%s) nontrapping pte"
2228 " in nonleaf level: levels %d gva %lx"
2229 " level %d pte %llx\n", audit_msg,
2230 vcpu->arch.mmu.root_level, va, level, ent);
2231
2232 audit_mappings_page(vcpu, ent, va, level - 1);
2233 } else {
2234 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, va);
2235 hpa_t hpa = (hpa_t)gpa_to_pfn(vcpu, gpa) << PAGE_SHIFT;
2236
2237 if (is_shadow_present_pte(ent)
2238 && (ent & PT64_BASE_ADDR_MASK) != hpa)
2239 printk(KERN_ERR "xx audit error: (%s) levels %d"
2240 " gva %lx gpa %llx hpa %llx ent %llx %d\n",
2241 audit_msg, vcpu->arch.mmu.root_level,
2242 va, gpa, hpa, ent,
2243 is_shadow_present_pte(ent));
2244 else if (ent == shadow_notrap_nonpresent_pte
2245 && !is_error_hpa(hpa))
2246 printk(KERN_ERR "audit: (%s) notrap shadow,"
2247 " valid guest gva %lx\n", audit_msg, va);
2248 kvm_release_pfn_clean(pfn);
2249
2250 }
2251 }
2252}
2253
2254static void audit_mappings(struct kvm_vcpu *vcpu)
2255{
2256 unsigned i;
2257
2258 if (vcpu->arch.mmu.root_level == 4)
2259 audit_mappings_page(vcpu, vcpu->arch.mmu.root_hpa, 0, 4);
2260 else
2261 for (i = 0; i < 4; ++i)
2262 if (vcpu->arch.mmu.pae_root[i] & PT_PRESENT_MASK)
2263 audit_mappings_page(vcpu,
2264 vcpu->arch.mmu.pae_root[i],
2265 i << 30,
2266 2);
2267}
2268
2269static int count_rmaps(struct kvm_vcpu *vcpu)
2270{
2271 int nmaps = 0;
2272 int i, j, k;
2273
2274 for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
2275 struct kvm_memory_slot *m = &vcpu->kvm->memslots[i];
2276 struct kvm_rmap_desc *d;
2277
2278 for (j = 0; j < m->npages; ++j) {
2279 unsigned long *rmapp = &m->rmap[j];
2280
2281 if (!*rmapp)
2282 continue;
2283 if (!(*rmapp & 1)) {
2284 ++nmaps;
2285 continue;
2286 }
2287 d = (struct kvm_rmap_desc *)(*rmapp & ~1ul);
2288 while (d) {
2289 for (k = 0; k < RMAP_EXT; ++k)
2290 if (d->shadow_ptes[k])
2291 ++nmaps;
2292 else
2293 break;
2294 d = d->more;
2295 }
2296 }
2297 }
2298 return nmaps;
2299}
2300
2301static int count_writable_mappings(struct kvm_vcpu *vcpu)
2302{
2303 int nmaps = 0;
2304 struct kvm_mmu_page *sp;
2305 int i;
2306
2307 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2308 u64 *pt = sp->spt;
2309
2310 if (sp->role.level != PT_PAGE_TABLE_LEVEL)
2311 continue;
2312
2313 for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
2314 u64 ent = pt[i];
2315
2316 if (!(ent & PT_PRESENT_MASK))
2317 continue;
2318 if (!(ent & PT_WRITABLE_MASK))
2319 continue;
2320 ++nmaps;
2321 }
2322 }
2323 return nmaps;
2324}
2325
2326static void audit_rmap(struct kvm_vcpu *vcpu)
2327{
2328 int n_rmap = count_rmaps(vcpu);
2329 int n_actual = count_writable_mappings(vcpu);
2330
2331 if (n_rmap != n_actual)
2332 printk(KERN_ERR "%s: (%s) rmap %d actual %d\n",
2333 __func__, audit_msg, n_rmap, n_actual);
2334}
2335
2336static void audit_write_protection(struct kvm_vcpu *vcpu)
2337{
2338 struct kvm_mmu_page *sp;
2339 struct kvm_memory_slot *slot;
2340 unsigned long *rmapp;
2341 gfn_t gfn;
2342
2343 list_for_each_entry(sp, &vcpu->kvm->arch.active_mmu_pages, link) {
2344 if (sp->role.metaphysical)
2345 continue;
2346
2347 slot = gfn_to_memslot(vcpu->kvm, sp->gfn);
2348 gfn = unalias_gfn(vcpu->kvm, sp->gfn);
2349 rmapp = &slot->rmap[gfn - slot->base_gfn];
2350 if (*rmapp)
2351 printk(KERN_ERR "%s: (%s) shadow page has writable"
2352 " mappings: gfn %lx role %x\n",
2353 __func__, audit_msg, sp->gfn,
2354 sp->role.word);
2355 }
2356}
2357
2358static void kvm_mmu_audit(struct kvm_vcpu *vcpu, const char *msg)
2359{
2360 int olddbg = dbg;
2361
2362 dbg = 0;
2363 audit_msg = msg;
2364 audit_rmap(vcpu);
2365 audit_write_protection(vcpu);
2366 audit_mappings(vcpu);
2367 dbg = olddbg;
2368}
2369
2370#endif
Linux 6.x block layer and io_uring tree(s)