1 /* SPDX-License-Identifier: GPL-2.0-only */
2 /*
3 * Kernel-based Virtual Machine driver for Linux
4 *
5 * This module enables machines with Intel VT-x extensions to run virtual
6 * machines without emulation or binary translation.
7 *
8 * MMU support
9 *
10 * Copyright (C) 2006 Qumranet, Inc.
11 * Copyright 2010 Red Hat, Inc. and/or its affiliates.
12 *
13 * Authors:
14 * Yaniv Kamay <yaniv@qumranet.com>
15 * Avi Kivity <avi@qumranet.com>
16 */
17
18 /*
19 * We need the mmu code to access both 32-bit and 64-bit guest ptes,
20 * so the code in this file is compiled twice, once per pte size.
21 */
22
23 #if PTTYPE == 64
24 #define pt_element_t u64
25 #define guest_walker guest_walker64
26 #define FNAME(name) paging##64_##name
27 #define PT_BASE_ADDR_MASK GUEST_PT64_BASE_ADDR_MASK
28 #define PT_LVL_ADDR_MASK(lvl) PT64_LVL_ADDR_MASK(lvl)
29 #define PT_LVL_OFFSET_MASK(lvl) PT64_LVL_OFFSET_MASK(lvl)
30 #define PT_INDEX(addr, level) PT64_INDEX(addr, level)
31 #define PT_LEVEL_BITS PT64_LEVEL_BITS
32 #define PT_GUEST_DIRTY_SHIFT PT_DIRTY_SHIFT
33 #define PT_GUEST_ACCESSED_SHIFT PT_ACCESSED_SHIFT
34 #define PT_HAVE_ACCESSED_DIRTY(mmu) true
35 #ifdef CONFIG_X86_64
36 #define PT_MAX_FULL_LEVELS PT64_ROOT_MAX_LEVEL
37 #define CMPXCHG "cmpxchgq"
38 #else
39 #define PT_MAX_FULL_LEVELS 2
40 #endif
41 #elif PTTYPE == 32
42 #define pt_element_t u32
43 #define guest_walker guest_walker32
44 #define FNAME(name) paging##32_##name
45 #define PT_BASE_ADDR_MASK PT32_BASE_ADDR_MASK
46 #define PT_LVL_ADDR_MASK(lvl) PT32_LVL_ADDR_MASK(lvl)
47 #define PT_LVL_OFFSET_MASK(lvl) PT32_LVL_OFFSET_MASK(lvl)
48 #define PT_INDEX(addr, level) PT32_INDEX(addr, level)
49 #define PT_LEVEL_BITS PT32_LEVEL_BITS
50 #define PT_MAX_FULL_LEVELS 2
51 #define PT_GUEST_DIRTY_SHIFT PT_DIRTY_SHIFT
52 #define PT_GUEST_ACCESSED_SHIFT PT_ACCESSED_SHIFT
53 #define PT_HAVE_ACCESSED_DIRTY(mmu) true
54 #define CMPXCHG "cmpxchgl"
55 #elif PTTYPE == PTTYPE_EPT
56 #define pt_element_t u64
57 #define guest_walker guest_walkerEPT
58 #define FNAME(name) ept_##name
59 #define PT_BASE_ADDR_MASK GUEST_PT64_BASE_ADDR_MASK
60 #define PT_LVL_ADDR_MASK(lvl) PT64_LVL_ADDR_MASK(lvl)
61 #define PT_LVL_OFFSET_MASK(lvl) PT64_LVL_OFFSET_MASK(lvl)
62 #define PT_INDEX(addr, level) PT64_INDEX(addr, level)
63 #define PT_LEVEL_BITS PT64_LEVEL_BITS
64 #define PT_GUEST_DIRTY_SHIFT 9
65 #define PT_GUEST_ACCESSED_SHIFT 8
66 #define PT_HAVE_ACCESSED_DIRTY(mmu) ((mmu)->ept_ad)
67 #ifdef CONFIG_X86_64
68 #define CMPXCHG "cmpxchgq"
69 #endif
70 #define PT_MAX_FULL_LEVELS PT64_ROOT_MAX_LEVEL
71 #else
72 #error Invalid PTTYPE value
73 #endif
74
75 #define PT_GUEST_DIRTY_MASK (1 << PT_GUEST_DIRTY_SHIFT)
76 #define PT_GUEST_ACCESSED_MASK (1 << PT_GUEST_ACCESSED_SHIFT)
77
78 #define gpte_to_gfn_lvl FNAME(gpte_to_gfn_lvl)
79 #define gpte_to_gfn(pte) gpte_to_gfn_lvl((pte), PG_LEVEL_4K)
80
81 /*
82 * The guest_walker structure emulates the behavior of the hardware page
83 * table walker.
84 */
85 struct guest_walker {
86 int level;
87 unsigned max_level;
88 gfn_t table_gfn[PT_MAX_FULL_LEVELS];
89 pt_element_t ptes[PT_MAX_FULL_LEVELS];
90 pt_element_t prefetch_ptes[PTE_PREFETCH_NUM];
91 gpa_t pte_gpa[PT_MAX_FULL_LEVELS];
92 pt_element_t __user *ptep_user[PT_MAX_FULL_LEVELS];
93 bool pte_writable[PT_MAX_FULL_LEVELS];
94 unsigned int pt_access[PT_MAX_FULL_LEVELS];
95 unsigned int pte_access;
96 gfn_t gfn;
97 struct x86_exception fault;
98 };
99
gpte_to_gfn_lvl(pt_element_t gpte,int lvl)100 static gfn_t gpte_to_gfn_lvl(pt_element_t gpte, int lvl)
101 {
102 return (gpte & PT_LVL_ADDR_MASK(lvl)) >> PAGE_SHIFT;
103 }
104
FNAME(protect_clean_gpte)105 static inline void FNAME(protect_clean_gpte)(struct kvm_mmu *mmu, unsigned *access,
106 unsigned gpte)
107 {
108 unsigned mask;
109
110 /* dirty bit is not supported, so no need to track it */
111 if (!PT_HAVE_ACCESSED_DIRTY(mmu))
112 return;
113
114 BUILD_BUG_ON(PT_WRITABLE_MASK != ACC_WRITE_MASK);
115
116 mask = (unsigned)~ACC_WRITE_MASK;
117 /* Allow write access to dirty gptes */
118 mask |= (gpte >> (PT_GUEST_DIRTY_SHIFT - PT_WRITABLE_SHIFT)) &
119 PT_WRITABLE_MASK;
120 *access &= mask;
121 }
122
FNAME(is_present_gpte)123 static inline int FNAME(is_present_gpte)(unsigned long pte)
124 {
125 #if PTTYPE != PTTYPE_EPT
126 return pte & PT_PRESENT_MASK;
127 #else
128 return pte & 7;
129 #endif
130 }
131
FNAME(is_bad_mt_xwr)132 static bool FNAME(is_bad_mt_xwr)(struct rsvd_bits_validate *rsvd_check, u64 gpte)
133 {
134 #if PTTYPE != PTTYPE_EPT
135 return false;
136 #else
137 return __is_bad_mt_xwr(rsvd_check, gpte);
138 #endif
139 }
140
FNAME(is_rsvd_bits_set)141 static bool FNAME(is_rsvd_bits_set)(struct kvm_mmu *mmu, u64 gpte, int level)
142 {
143 return __is_rsvd_bits_set(&mmu->guest_rsvd_check, gpte, level) ||
144 FNAME(is_bad_mt_xwr)(&mmu->guest_rsvd_check, gpte);
145 }
146
FNAME(cmpxchg_gpte)147 static int FNAME(cmpxchg_gpte)(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
148 pt_element_t __user *ptep_user, unsigned index,
149 pt_element_t orig_pte, pt_element_t new_pte)
150 {
151 int r = -EFAULT;
152
153 if (!user_access_begin(ptep_user, sizeof(pt_element_t)))
154 return -EFAULT;
155
156 #ifdef CMPXCHG
157 asm volatile("1:" LOCK_PREFIX CMPXCHG " %[new], %[ptr]\n"
158 "mov $0, %[r]\n"
159 "setnz %b[r]\n"
160 "2:"
161 _ASM_EXTABLE_UA(1b, 2b)
162 : [ptr] "+m" (*ptep_user),
163 [old] "+a" (orig_pte),
164 [r] "+q" (r)
165 : [new] "r" (new_pte)
166 : "memory");
167 #else
168 asm volatile("1:" LOCK_PREFIX "cmpxchg8b %[ptr]\n"
169 "movl $0, %[r]\n"
170 "jz 2f\n"
171 "incl %[r]\n"
172 "2:"
173 _ASM_EXTABLE_UA(1b, 2b)
174 : [ptr] "+m" (*ptep_user),
175 [old] "+A" (orig_pte),
176 [r] "+rm" (r)
177 : [new_lo] "b" ((u32)new_pte),
178 [new_hi] "c" ((u32)(new_pte >> 32))
179 : "memory");
180 #endif
181
182 user_access_end();
183 return r;
184 }
185
FNAME(prefetch_invalid_gpte)186 static bool FNAME(prefetch_invalid_gpte)(struct kvm_vcpu *vcpu,
187 struct kvm_mmu_page *sp, u64 *spte,
188 u64 gpte)
189 {
190 if (!FNAME(is_present_gpte)(gpte))
191 goto no_present;
192
193 /* if accessed bit is not supported prefetch non accessed gpte */
194 if (PT_HAVE_ACCESSED_DIRTY(vcpu->arch.mmu) &&
195 !(gpte & PT_GUEST_ACCESSED_MASK))
196 goto no_present;
197
198 if (FNAME(is_rsvd_bits_set)(vcpu->arch.mmu, gpte, PG_LEVEL_4K))
199 goto no_present;
200
201 return false;
202
203 no_present:
204 drop_spte(vcpu->kvm, spte);
205 return true;
206 }
207
208 /*
209 * For PTTYPE_EPT, a page table can be executable but not readable
210 * on supported processors. Therefore, set_spte does not automatically
211 * set bit 0 if execute only is supported. Here, we repurpose ACC_USER_MASK
212 * to signify readability since it isn't used in the EPT case
213 */
FNAME(gpte_access)214 static inline unsigned FNAME(gpte_access)(u64 gpte)
215 {
216 unsigned access;
217 #if PTTYPE == PTTYPE_EPT
218 access = ((gpte & VMX_EPT_WRITABLE_MASK) ? ACC_WRITE_MASK : 0) |
219 ((gpte & VMX_EPT_EXECUTABLE_MASK) ? ACC_EXEC_MASK : 0) |
220 ((gpte & VMX_EPT_READABLE_MASK) ? ACC_USER_MASK : 0);
221 #else
222 BUILD_BUG_ON(ACC_EXEC_MASK != PT_PRESENT_MASK);
223 BUILD_BUG_ON(ACC_EXEC_MASK != 1);
224 access = gpte & (PT_WRITABLE_MASK | PT_USER_MASK | PT_PRESENT_MASK);
225 /* Combine NX with P (which is set here) to get ACC_EXEC_MASK. */
226 access ^= (gpte >> PT64_NX_SHIFT);
227 #endif
228
229 return access;
230 }
231
FNAME(update_accessed_dirty_bits)232 static int FNAME(update_accessed_dirty_bits)(struct kvm_vcpu *vcpu,
233 struct kvm_mmu *mmu,
234 struct guest_walker *walker,
235 gpa_t addr, int write_fault)
236 {
237 unsigned level, index;
238 pt_element_t pte, orig_pte;
239 pt_element_t __user *ptep_user;
240 gfn_t table_gfn;
241 int ret;
242
243 /* dirty/accessed bits are not supported, so no need to update them */
244 if (!PT_HAVE_ACCESSED_DIRTY(mmu))
245 return 0;
246
247 for (level = walker->max_level; level >= walker->level; --level) {
248 pte = orig_pte = walker->ptes[level - 1];
249 table_gfn = walker->table_gfn[level - 1];
250 ptep_user = walker->ptep_user[level - 1];
251 index = offset_in_page(ptep_user) / sizeof(pt_element_t);
252 if (!(pte & PT_GUEST_ACCESSED_MASK)) {
253 trace_kvm_mmu_set_accessed_bit(table_gfn, index, sizeof(pte));
254 pte |= PT_GUEST_ACCESSED_MASK;
255 }
256 if (level == walker->level && write_fault &&
257 !(pte & PT_GUEST_DIRTY_MASK)) {
258 trace_kvm_mmu_set_dirty_bit(table_gfn, index, sizeof(pte));
259 #if PTTYPE == PTTYPE_EPT
260 if (kvm_x86_ops.nested_ops->write_log_dirty(vcpu, addr))
261 return -EINVAL;
262 #endif
263 pte |= PT_GUEST_DIRTY_MASK;
264 }
265 if (pte == orig_pte)
266 continue;
267
268 /*
269 * If the slot is read-only, simply do not process the accessed
270 * and dirty bits. This is the correct thing to do if the slot
271 * is ROM, and page tables in read-as-ROM/write-as-MMIO slots
272 * are only supported if the accessed and dirty bits are already
273 * set in the ROM (so that MMIO writes are never needed).
274 *
275 * Note that NPT does not allow this at all and faults, since
276 * it always wants nested page table entries for the guest
277 * page tables to be writable. And EPT works but will simply
278 * overwrite the read-only memory to set the accessed and dirty
279 * bits.
280 */
281 if (unlikely(!walker->pte_writable[level - 1]))
282 continue;
283
284 ret = FNAME(cmpxchg_gpte)(vcpu, mmu, ptep_user, index, orig_pte, pte);
285 if (ret)
286 return ret;
287
288 kvm_vcpu_mark_page_dirty(vcpu, table_gfn);
289 walker->ptes[level - 1] = pte;
290 }
291 return 0;
292 }
293
FNAME(gpte_pkeys)294 static inline unsigned FNAME(gpte_pkeys)(struct kvm_vcpu *vcpu, u64 gpte)
295 {
296 unsigned pkeys = 0;
297 #if PTTYPE == 64
298 pte_t pte = {.pte = gpte};
299
300 pkeys = pte_flags_pkey(pte_flags(pte));
301 #endif
302 return pkeys;
303 }
304
305 /*
306 * Fetch a guest pte for a guest virtual address, or for an L2's GPA.
307 */
FNAME(walk_addr_generic)308 static int FNAME(walk_addr_generic)(struct guest_walker *walker,
309 struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
310 gpa_t addr, u32 access)
311 {
312 int ret;
313 pt_element_t pte;
314 pt_element_t __user *ptep_user;
315 gfn_t table_gfn;
316 u64 pt_access, pte_access;
317 unsigned index, accessed_dirty, pte_pkey;
318 unsigned nested_access;
319 gpa_t pte_gpa;
320 bool have_ad;
321 int offset;
322 u64 walk_nx_mask = 0;
323 const int write_fault = access & PFERR_WRITE_MASK;
324 const int user_fault = access & PFERR_USER_MASK;
325 const int fetch_fault = access & PFERR_FETCH_MASK;
326 u16 errcode = 0;
327 gpa_t real_gpa;
328 gfn_t gfn;
329
330 trace_kvm_mmu_pagetable_walk(addr, access);
331 retry_walk:
332 walker->level = mmu->root_level;
333 pte = mmu->get_guest_pgd(vcpu);
334 have_ad = PT_HAVE_ACCESSED_DIRTY(mmu);
335
336 #if PTTYPE == 64
337 walk_nx_mask = 1ULL << PT64_NX_SHIFT;
338 if (walker->level == PT32E_ROOT_LEVEL) {
339 pte = mmu->get_pdptr(vcpu, (addr >> 30) & 3);
340 trace_kvm_mmu_paging_element(pte, walker->level);
341 if (!FNAME(is_present_gpte)(pte))
342 goto error;
343 --walker->level;
344 }
345 #endif
346 walker->max_level = walker->level;
347 ASSERT(!(is_long_mode(vcpu) && !is_pae(vcpu)));
348
349 /*
350 * FIXME: on Intel processors, loads of the PDPTE registers for PAE paging
351 * by the MOV to CR instruction are treated as reads and do not cause the
352 * processor to set the dirty flag in any EPT paging-structure entry.
353 */
354 nested_access = (have_ad ? PFERR_WRITE_MASK : 0) | PFERR_USER_MASK;
355
356 pte_access = ~0;
357 ++walker->level;
358
359 do {
360 unsigned long host_addr;
361
362 pt_access = pte_access;
363 --walker->level;
364
365 index = PT_INDEX(addr, walker->level);
366 table_gfn = gpte_to_gfn(pte);
367 offset = index * sizeof(pt_element_t);
368 pte_gpa = gfn_to_gpa(table_gfn) + offset;
369
370 BUG_ON(walker->level < 1);
371 walker->table_gfn[walker->level - 1] = table_gfn;
372 walker->pte_gpa[walker->level - 1] = pte_gpa;
373
374 real_gpa = mmu->translate_gpa(vcpu, gfn_to_gpa(table_gfn),
375 nested_access,
376 &walker->fault);
377
378 /*
379 * FIXME: This can happen if emulation (for of an INS/OUTS
380 * instruction) triggers a nested page fault. The exit
381 * qualification / exit info field will incorrectly have
382 * "guest page access" as the nested page fault's cause,
383 * instead of "guest page structure access". To fix this,
384 * the x86_exception struct should be augmented with enough
385 * information to fix the exit_qualification or exit_info_1
386 * fields.
387 */
388 if (unlikely(real_gpa == UNMAPPED_GVA))
389 return 0;
390
391 host_addr = kvm_vcpu_gfn_to_hva_prot(vcpu, gpa_to_gfn(real_gpa),
392 &walker->pte_writable[walker->level - 1]);
393 if (unlikely(kvm_is_error_hva(host_addr)))
394 goto error;
395
396 ptep_user = (pt_element_t __user *)((void *)host_addr + offset);
397 if (unlikely(__get_user(pte, ptep_user)))
398 goto error;
399 walker->ptep_user[walker->level - 1] = ptep_user;
400
401 trace_kvm_mmu_paging_element(pte, walker->level);
402
403 /*
404 * Inverting the NX it lets us AND it like other
405 * permission bits.
406 */
407 pte_access = pt_access & (pte ^ walk_nx_mask);
408
409 if (unlikely(!FNAME(is_present_gpte)(pte)))
410 goto error;
411
412 if (unlikely(FNAME(is_rsvd_bits_set)(mmu, pte, walker->level))) {
413 errcode = PFERR_RSVD_MASK | PFERR_PRESENT_MASK;
414 goto error;
415 }
416
417 walker->ptes[walker->level - 1] = pte;
418
419 /* Convert to ACC_*_MASK flags for struct guest_walker. */
420 walker->pt_access[walker->level - 1] = FNAME(gpte_access)(pt_access ^ walk_nx_mask);
421 } while (!is_last_gpte(mmu, walker->level, pte));
422
423 pte_pkey = FNAME(gpte_pkeys)(vcpu, pte);
424 accessed_dirty = have_ad ? pte_access & PT_GUEST_ACCESSED_MASK : 0;
425
426 /* Convert to ACC_*_MASK flags for struct guest_walker. */
427 walker->pte_access = FNAME(gpte_access)(pte_access ^ walk_nx_mask);
428 errcode = permission_fault(vcpu, mmu, walker->pte_access, pte_pkey, access);
429 if (unlikely(errcode))
430 goto error;
431
432 gfn = gpte_to_gfn_lvl(pte, walker->level);
433 gfn += (addr & PT_LVL_OFFSET_MASK(walker->level)) >> PAGE_SHIFT;
434
435 if (PTTYPE == 32 && walker->level > PG_LEVEL_4K && is_cpuid_PSE36())
436 gfn += pse36_gfn_delta(pte);
437
438 real_gpa = mmu->translate_gpa(vcpu, gfn_to_gpa(gfn), access, &walker->fault);
439 if (real_gpa == UNMAPPED_GVA)
440 return 0;
441
442 walker->gfn = real_gpa >> PAGE_SHIFT;
443
444 if (!write_fault)
445 FNAME(protect_clean_gpte)(mmu, &walker->pte_access, pte);
446 else
447 /*
448 * On a write fault, fold the dirty bit into accessed_dirty.
449 * For modes without A/D bits support accessed_dirty will be
450 * always clear.
451 */
452 accessed_dirty &= pte >>
453 (PT_GUEST_DIRTY_SHIFT - PT_GUEST_ACCESSED_SHIFT);
454
455 if (unlikely(!accessed_dirty)) {
456 ret = FNAME(update_accessed_dirty_bits)(vcpu, mmu, walker,
457 addr, write_fault);
458 if (unlikely(ret < 0))
459 goto error;
460 else if (ret)
461 goto retry_walk;
462 }
463
464 pgprintk("%s: pte %llx pte_access %x pt_access %x\n",
465 __func__, (u64)pte, walker->pte_access,
466 walker->pt_access[walker->level - 1]);
467 return 1;
468
469 error:
470 errcode |= write_fault | user_fault;
471 if (fetch_fault && (mmu->nx || mmu->mmu_role.ext.cr4_smep))
472 errcode |= PFERR_FETCH_MASK;
473
474 walker->fault.vector = PF_VECTOR;
475 walker->fault.error_code_valid = true;
476 walker->fault.error_code = errcode;
477
478 #if PTTYPE == PTTYPE_EPT
479 /*
480 * Use PFERR_RSVD_MASK in error_code to to tell if EPT
481 * misconfiguration requires to be injected. The detection is
482 * done by is_rsvd_bits_set() above.
483 *
484 * We set up the value of exit_qualification to inject:
485 * [2:0] - Derive from the access bits. The exit_qualification might be
486 * out of date if it is serving an EPT misconfiguration.
487 * [5:3] - Calculated by the page walk of the guest EPT page tables
488 * [7:8] - Derived from [7:8] of real exit_qualification
489 *
490 * The other bits are set to 0.
491 */
492 if (!(errcode & PFERR_RSVD_MASK)) {
493 vcpu->arch.exit_qualification &= 0x180;
494 if (write_fault)
495 vcpu->arch.exit_qualification |= EPT_VIOLATION_ACC_WRITE;
496 if (user_fault)
497 vcpu->arch.exit_qualification |= EPT_VIOLATION_ACC_READ;
498 if (fetch_fault)
499 vcpu->arch.exit_qualification |= EPT_VIOLATION_ACC_INSTR;
500 vcpu->arch.exit_qualification |= (pte_access & 0x7) << 3;
501 }
502 #endif
503 walker->fault.address = addr;
504 walker->fault.nested_page_fault = mmu != vcpu->arch.walk_mmu;
505
506 trace_kvm_mmu_walker_error(walker->fault.error_code);
507 return 0;
508 }
509
FNAME(walk_addr)510 static int FNAME(walk_addr)(struct guest_walker *walker,
511 struct kvm_vcpu *vcpu, gpa_t addr, u32 access)
512 {
513 return FNAME(walk_addr_generic)(walker, vcpu, vcpu->arch.mmu, addr,
514 access);
515 }
516
517 #if PTTYPE != PTTYPE_EPT
FNAME(walk_addr_nested)518 static int FNAME(walk_addr_nested)(struct guest_walker *walker,
519 struct kvm_vcpu *vcpu, gva_t addr,
520 u32 access)
521 {
522 return FNAME(walk_addr_generic)(walker, vcpu, &vcpu->arch.nested_mmu,
523 addr, access);
524 }
525 #endif
526
527 static bool
FNAME(prefetch_gpte)528 FNAME(prefetch_gpte)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
529 u64 *spte, pt_element_t gpte, bool no_dirty_log)
530 {
531 unsigned pte_access;
532 gfn_t gfn;
533 kvm_pfn_t pfn;
534
535 if (FNAME(prefetch_invalid_gpte)(vcpu, sp, spte, gpte))
536 return false;
537
538 pgprintk("%s: gpte %llx spte %p\n", __func__, (u64)gpte, spte);
539
540 gfn = gpte_to_gfn(gpte);
541 pte_access = sp->role.access & FNAME(gpte_access)(gpte);
542 FNAME(protect_clean_gpte)(vcpu->arch.mmu, &pte_access, gpte);
543 pfn = pte_prefetch_gfn_to_pfn(vcpu, gfn,
544 no_dirty_log && (pte_access & ACC_WRITE_MASK));
545 if (is_error_pfn(pfn))
546 return false;
547
548 /*
549 * we call mmu_set_spte() with host_writable = true because
550 * pte_prefetch_gfn_to_pfn always gets a writable pfn.
551 */
552 mmu_set_spte(vcpu, spte, pte_access, false, PG_LEVEL_4K, gfn, pfn,
553 true, true);
554
555 kvm_release_pfn_clean(pfn);
556 return true;
557 }
558
FNAME(update_pte)559 static void FNAME(update_pte)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
560 u64 *spte, const void *pte)
561 {
562 pt_element_t gpte = *(const pt_element_t *)pte;
563
564 FNAME(prefetch_gpte)(vcpu, sp, spte, gpte, false);
565 }
566
FNAME(gpte_changed)567 static bool FNAME(gpte_changed)(struct kvm_vcpu *vcpu,
568 struct guest_walker *gw, int level)
569 {
570 pt_element_t curr_pte;
571 gpa_t base_gpa, pte_gpa = gw->pte_gpa[level - 1];
572 u64 mask;
573 int r, index;
574
575 if (level == PG_LEVEL_4K) {
576 mask = PTE_PREFETCH_NUM * sizeof(pt_element_t) - 1;
577 base_gpa = pte_gpa & ~mask;
578 index = (pte_gpa - base_gpa) / sizeof(pt_element_t);
579
580 r = kvm_vcpu_read_guest_atomic(vcpu, base_gpa,
581 gw->prefetch_ptes, sizeof(gw->prefetch_ptes));
582 curr_pte = gw->prefetch_ptes[index];
583 } else
584 r = kvm_vcpu_read_guest_atomic(vcpu, pte_gpa,
585 &curr_pte, sizeof(curr_pte));
586
587 return r || curr_pte != gw->ptes[level - 1];
588 }
589
FNAME(pte_prefetch)590 static void FNAME(pte_prefetch)(struct kvm_vcpu *vcpu, struct guest_walker *gw,
591 u64 *sptep)
592 {
593 struct kvm_mmu_page *sp;
594 pt_element_t *gptep = gw->prefetch_ptes;
595 u64 *spte;
596 int i;
597
598 sp = sptep_to_sp(sptep);
599
600 if (sp->role.level > PG_LEVEL_4K)
601 return;
602
603 if (sp->role.direct)
604 return __direct_pte_prefetch(vcpu, sp, sptep);
605
606 i = (sptep - sp->spt) & ~(PTE_PREFETCH_NUM - 1);
607 spte = sp->spt + i;
608
609 for (i = 0; i < PTE_PREFETCH_NUM; i++, spte++) {
610 if (spte == sptep)
611 continue;
612
613 if (is_shadow_present_pte(*spte))
614 continue;
615
616 if (!FNAME(prefetch_gpte)(vcpu, sp, spte, gptep[i], true))
617 break;
618 }
619 }
620
621 /*
622 * Fetch a shadow pte for a specific level in the paging hierarchy.
623 * If the guest tries to write a write-protected page, we need to
624 * emulate this operation, return 1 to indicate this case.
625 */
FNAME(fetch)626 static int FNAME(fetch)(struct kvm_vcpu *vcpu, gpa_t addr,
627 struct guest_walker *gw, u32 error_code,
628 int max_level, kvm_pfn_t pfn, bool map_writable,
629 bool prefault)
630 {
631 bool nx_huge_page_workaround_enabled = is_nx_huge_page_enabled();
632 bool write_fault = error_code & PFERR_WRITE_MASK;
633 bool exec = error_code & PFERR_FETCH_MASK;
634 bool huge_page_disallowed = exec && nx_huge_page_workaround_enabled;
635 struct kvm_mmu_page *sp = NULL;
636 struct kvm_shadow_walk_iterator it;
637 unsigned int direct_access, access;
638 int top_level, level, req_level, ret;
639 gfn_t base_gfn = gw->gfn;
640
641 direct_access = gw->pte_access;
642
643 top_level = vcpu->arch.mmu->root_level;
644 if (top_level == PT32E_ROOT_LEVEL)
645 top_level = PT32_ROOT_LEVEL;
646 /*
647 * Verify that the top-level gpte is still there. Since the page
648 * is a root page, it is either write protected (and cannot be
649 * changed from now on) or it is invalid (in which case, we don't
650 * really care if it changes underneath us after this point).
651 */
652 if (FNAME(gpte_changed)(vcpu, gw, top_level))
653 goto out_gpte_changed;
654
655 if (WARN_ON(!VALID_PAGE(vcpu->arch.mmu->root_hpa)))
656 goto out_gpte_changed;
657
658 for (shadow_walk_init(&it, vcpu, addr);
659 shadow_walk_okay(&it) && it.level > gw->level;
660 shadow_walk_next(&it)) {
661 gfn_t table_gfn;
662
663 clear_sp_write_flooding_count(it.sptep);
664 drop_large_spte(vcpu, it.sptep);
665
666 sp = NULL;
667 if (!is_shadow_present_pte(*it.sptep)) {
668 table_gfn = gw->table_gfn[it.level - 2];
669 access = gw->pt_access[it.level - 2];
670 sp = kvm_mmu_get_page(vcpu, table_gfn, addr, it.level-1,
671 false, access);
672 }
673
674 /*
675 * Verify that the gpte in the page we've just write
676 * protected is still there.
677 */
678 if (FNAME(gpte_changed)(vcpu, gw, it.level - 1))
679 goto out_gpte_changed;
680
681 if (sp)
682 link_shadow_page(vcpu, it.sptep, sp);
683 }
684
685 level = kvm_mmu_hugepage_adjust(vcpu, gw->gfn, max_level, &pfn,
686 huge_page_disallowed, &req_level);
687
688 trace_kvm_mmu_spte_requested(addr, gw->level, pfn);
689
690 for (; shadow_walk_okay(&it); shadow_walk_next(&it)) {
691 clear_sp_write_flooding_count(it.sptep);
692
693 /*
694 * We cannot overwrite existing page tables with an NX
695 * large page, as the leaf could be executable.
696 */
697 if (nx_huge_page_workaround_enabled)
698 disallowed_hugepage_adjust(*it.sptep, gw->gfn, it.level,
699 &pfn, &level);
700
701 base_gfn = gw->gfn & ~(KVM_PAGES_PER_HPAGE(it.level) - 1);
702 if (it.level == level)
703 break;
704
705 validate_direct_spte(vcpu, it.sptep, direct_access);
706
707 drop_large_spte(vcpu, it.sptep);
708
709 if (!is_shadow_present_pte(*it.sptep)) {
710 sp = kvm_mmu_get_page(vcpu, base_gfn, addr,
711 it.level - 1, true, direct_access);
712 link_shadow_page(vcpu, it.sptep, sp);
713 if (huge_page_disallowed && req_level >= it.level)
714 account_huge_nx_page(vcpu->kvm, sp);
715 }
716 }
717
718 ret = mmu_set_spte(vcpu, it.sptep, gw->pte_access, write_fault,
719 it.level, base_gfn, pfn, prefault, map_writable);
720 if (ret == RET_PF_SPURIOUS)
721 return ret;
722
723 FNAME(pte_prefetch)(vcpu, gw, it.sptep);
724 ++vcpu->stat.pf_fixed;
725 return ret;
726
727 out_gpte_changed:
728 return RET_PF_RETRY;
729 }
730
731 /*
732 * To see whether the mapped gfn can write its page table in the current
733 * mapping.
734 *
735 * It is the helper function of FNAME(page_fault). When guest uses large page
736 * size to map the writable gfn which is used as current page table, we should
737 * force kvm to use small page size to map it because new shadow page will be
738 * created when kvm establishes shadow page table that stop kvm using large
739 * page size. Do it early can avoid unnecessary #PF and emulation.
740 *
741 * @write_fault_to_shadow_pgtable will return true if the fault gfn is
742 * currently used as its page table.
743 *
744 * Note: the PDPT page table is not checked for PAE-32 bit guest. It is ok
745 * since the PDPT is always shadowed, that means, we can not use large page
746 * size to map the gfn which is used as PDPT.
747 */
748 static bool
FNAME(is_self_change_mapping)749 FNAME(is_self_change_mapping)(struct kvm_vcpu *vcpu,
750 struct guest_walker *walker, bool user_fault,
751 bool *write_fault_to_shadow_pgtable)
752 {
753 int level;
754 gfn_t mask = ~(KVM_PAGES_PER_HPAGE(walker->level) - 1);
755 bool self_changed = false;
756
757 if (!(walker->pte_access & ACC_WRITE_MASK ||
758 (!is_write_protection(vcpu) && !user_fault)))
759 return false;
760
761 for (level = walker->level; level <= walker->max_level; level++) {
762 gfn_t gfn = walker->gfn ^ walker->table_gfn[level - 1];
763
764 self_changed |= !(gfn & mask);
765 *write_fault_to_shadow_pgtable |= !gfn;
766 }
767
768 return self_changed;
769 }
770
771 /*
772 * Page fault handler. There are several causes for a page fault:
773 * - there is no shadow pte for the guest pte
774 * - write access through a shadow pte marked read only so that we can set
775 * the dirty bit
776 * - write access to a shadow pte marked read only so we can update the page
777 * dirty bitmap, when userspace requests it
778 * - mmio access; in this case we will never install a present shadow pte
779 * - normal guest page fault due to the guest pte marked not present, not
780 * writable, or not executable
781 *
782 * Returns: 1 if we need to emulate the instruction, 0 otherwise, or
783 * a negative value on error.
784 */
FNAME(page_fault)785 static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gpa_t addr, u32 error_code,
786 bool prefault)
787 {
788 bool write_fault = error_code & PFERR_WRITE_MASK;
789 bool user_fault = error_code & PFERR_USER_MASK;
790 struct guest_walker walker;
791 int r;
792 kvm_pfn_t pfn;
793 unsigned long mmu_seq;
794 bool map_writable, is_self_change_mapping;
795 int max_level;
796
797 pgprintk("%s: addr %lx err %x\n", __func__, addr, error_code);
798
799 /*
800 * If PFEC.RSVD is set, this is a shadow page fault.
801 * The bit needs to be cleared before walking guest page tables.
802 */
803 error_code &= ~PFERR_RSVD_MASK;
804
805 /*
806 * Look up the guest pte for the faulting address.
807 */
808 r = FNAME(walk_addr)(&walker, vcpu, addr, error_code);
809
810 /*
811 * The page is not mapped by the guest. Let the guest handle it.
812 */
813 if (!r) {
814 pgprintk("%s: guest page fault\n", __func__);
815 if (!prefault)
816 kvm_inject_emulated_page_fault(vcpu, &walker.fault);
817
818 return RET_PF_RETRY;
819 }
820
821 if (page_fault_handle_page_track(vcpu, error_code, walker.gfn)) {
822 shadow_page_table_clear_flood(vcpu, addr);
823 return RET_PF_EMULATE;
824 }
825
826 r = mmu_topup_memory_caches(vcpu, true);
827 if (r)
828 return r;
829
830 vcpu->arch.write_fault_to_shadow_pgtable = false;
831
832 is_self_change_mapping = FNAME(is_self_change_mapping)(vcpu,
833 &walker, user_fault, &vcpu->arch.write_fault_to_shadow_pgtable);
834
835 if (is_self_change_mapping)
836 max_level = PG_LEVEL_4K;
837 else
838 max_level = walker.level;
839
840 mmu_seq = vcpu->kvm->mmu_notifier_seq;
841 smp_rmb();
842
843 if (try_async_pf(vcpu, prefault, walker.gfn, addr, &pfn, write_fault,
844 &map_writable))
845 return RET_PF_RETRY;
846
847 if (handle_abnormal_pfn(vcpu, addr, walker.gfn, pfn, walker.pte_access, &r))
848 return r;
849
850 /*
851 * Do not change pte_access if the pfn is a mmio page, otherwise
852 * we will cache the incorrect access into mmio spte.
853 */
854 if (write_fault && !(walker.pte_access & ACC_WRITE_MASK) &&
855 !is_write_protection(vcpu) && !user_fault &&
856 !is_noslot_pfn(pfn)) {
857 walker.pte_access |= ACC_WRITE_MASK;
858 walker.pte_access &= ~ACC_USER_MASK;
859
860 /*
861 * If we converted a user page to a kernel page,
862 * so that the kernel can write to it when cr0.wp=0,
863 * then we should prevent the kernel from executing it
864 * if SMEP is enabled.
865 */
866 if (kvm_read_cr4_bits(vcpu, X86_CR4_SMEP))
867 walker.pte_access &= ~ACC_EXEC_MASK;
868 }
869
870 r = RET_PF_RETRY;
871 spin_lock(&vcpu->kvm->mmu_lock);
872 if (mmu_notifier_retry(vcpu->kvm, mmu_seq))
873 goto out_unlock;
874
875 kvm_mmu_audit(vcpu, AUDIT_PRE_PAGE_FAULT);
876 r = make_mmu_pages_available(vcpu);
877 if (r)
878 goto out_unlock;
879 r = FNAME(fetch)(vcpu, addr, &walker, error_code, max_level, pfn,
880 map_writable, prefault);
881 kvm_mmu_audit(vcpu, AUDIT_POST_PAGE_FAULT);
882
883 out_unlock:
884 spin_unlock(&vcpu->kvm->mmu_lock);
885 kvm_release_pfn_clean(pfn);
886 return r;
887 }
888
FNAME(get_level1_sp_gpa)889 static gpa_t FNAME(get_level1_sp_gpa)(struct kvm_mmu_page *sp)
890 {
891 int offset = 0;
892
893 WARN_ON(sp->role.level != PG_LEVEL_4K);
894
895 if (PTTYPE == 32)
896 offset = sp->role.quadrant << PT64_LEVEL_BITS;
897
898 return gfn_to_gpa(sp->gfn) + offset * sizeof(pt_element_t);
899 }
900
FNAME(invlpg)901 static void FNAME(invlpg)(struct kvm_vcpu *vcpu, gva_t gva, hpa_t root_hpa)
902 {
903 struct kvm_shadow_walk_iterator iterator;
904 struct kvm_mmu_page *sp;
905 u64 old_spte;
906 int level;
907 u64 *sptep;
908
909 vcpu_clear_mmio_info(vcpu, gva);
910
911 /*
912 * No need to check return value here, rmap_can_add() can
913 * help us to skip pte prefetch later.
914 */
915 mmu_topup_memory_caches(vcpu, true);
916
917 if (!VALID_PAGE(root_hpa)) {
918 WARN_ON(1);
919 return;
920 }
921
922 spin_lock(&vcpu->kvm->mmu_lock);
923 for_each_shadow_entry_using_root(vcpu, root_hpa, gva, iterator) {
924 level = iterator.level;
925 sptep = iterator.sptep;
926
927 sp = sptep_to_sp(sptep);
928 old_spte = *sptep;
929 if (is_last_spte(old_spte, level)) {
930 pt_element_t gpte;
931 gpa_t pte_gpa;
932
933 if (!sp->unsync)
934 break;
935
936 pte_gpa = FNAME(get_level1_sp_gpa)(sp);
937 pte_gpa += (sptep - sp->spt) * sizeof(pt_element_t);
938
939 mmu_page_zap_pte(vcpu->kvm, sp, sptep, NULL);
940 if (is_shadow_present_pte(old_spte))
941 kvm_flush_remote_tlbs_with_address(vcpu->kvm,
942 sp->gfn, KVM_PAGES_PER_HPAGE(sp->role.level));
943
944 if (!rmap_can_add(vcpu))
945 break;
946
947 if (kvm_vcpu_read_guest_atomic(vcpu, pte_gpa, &gpte,
948 sizeof(pt_element_t)))
949 break;
950
951 FNAME(update_pte)(vcpu, sp, sptep, &gpte);
952 }
953
954 if (!is_shadow_present_pte(*sptep) || !sp->unsync_children)
955 break;
956 }
957 spin_unlock(&vcpu->kvm->mmu_lock);
958 }
959
960 /* Note, @addr is a GPA when gva_to_gpa() translates an L2 GPA to an L1 GPA. */
FNAME(gva_to_gpa)961 static gpa_t FNAME(gva_to_gpa)(struct kvm_vcpu *vcpu, gpa_t addr, u32 access,
962 struct x86_exception *exception)
963 {
964 struct guest_walker walker;
965 gpa_t gpa = UNMAPPED_GVA;
966 int r;
967
968 r = FNAME(walk_addr)(&walker, vcpu, addr, access);
969
970 if (r) {
971 gpa = gfn_to_gpa(walker.gfn);
972 gpa |= addr & ~PAGE_MASK;
973 } else if (exception)
974 *exception = walker.fault;
975
976 return gpa;
977 }
978
979 #if PTTYPE != PTTYPE_EPT
980 /* Note, gva_to_gpa_nested() is only used to translate L2 GVAs. */
FNAME(gva_to_gpa_nested)981 static gpa_t FNAME(gva_to_gpa_nested)(struct kvm_vcpu *vcpu, gpa_t vaddr,
982 u32 access,
983 struct x86_exception *exception)
984 {
985 struct guest_walker walker;
986 gpa_t gpa = UNMAPPED_GVA;
987 int r;
988
989 #ifndef CONFIG_X86_64
990 /* A 64-bit GVA should be impossible on 32-bit KVM. */
991 WARN_ON_ONCE(vaddr >> 32);
992 #endif
993
994 r = FNAME(walk_addr_nested)(&walker, vcpu, vaddr, access);
995
996 if (r) {
997 gpa = gfn_to_gpa(walker.gfn);
998 gpa |= vaddr & ~PAGE_MASK;
999 } else if (exception)
1000 *exception = walker.fault;
1001
1002 return gpa;
1003 }
1004 #endif
1005
1006 /*
1007 * Using the cached information from sp->gfns is safe because:
1008 * - The spte has a reference to the struct page, so the pfn for a given gfn
1009 * can't change unless all sptes pointing to it are nuked first.
1010 *
1011 * Note:
1012 * We should flush all tlbs if spte is dropped even though guest is
1013 * responsible for it. Since if we don't, kvm_mmu_notifier_invalidate_page
1014 * and kvm_mmu_notifier_invalidate_range_start detect the mapping page isn't
1015 * used by guest then tlbs are not flushed, so guest is allowed to access the
1016 * freed pages.
1017 * And we increase kvm->tlbs_dirty to delay tlbs flush in this case.
1018 */
FNAME(sync_page)1019 static int FNAME(sync_page)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
1020 {
1021 int i, nr_present = 0;
1022 bool host_writable;
1023 gpa_t first_pte_gpa;
1024 int set_spte_ret = 0;
1025
1026 /* direct kvm_mmu_page can not be unsync. */
1027 BUG_ON(sp->role.direct);
1028
1029 first_pte_gpa = FNAME(get_level1_sp_gpa)(sp);
1030
1031 for (i = 0; i < PT64_ENT_PER_PAGE; i++) {
1032 unsigned pte_access;
1033 pt_element_t gpte;
1034 gpa_t pte_gpa;
1035 gfn_t gfn;
1036
1037 if (!sp->spt[i])
1038 continue;
1039
1040 pte_gpa = first_pte_gpa + i * sizeof(pt_element_t);
1041
1042 if (kvm_vcpu_read_guest_atomic(vcpu, pte_gpa, &gpte,
1043 sizeof(pt_element_t)))
1044 return 0;
1045
1046 if (FNAME(prefetch_invalid_gpte)(vcpu, sp, &sp->spt[i], gpte)) {
1047 /*
1048 * Update spte before increasing tlbs_dirty to make
1049 * sure no tlb flush is lost after spte is zapped; see
1050 * the comments in kvm_flush_remote_tlbs().
1051 */
1052 smp_wmb();
1053 vcpu->kvm->tlbs_dirty++;
1054 continue;
1055 }
1056
1057 gfn = gpte_to_gfn(gpte);
1058 pte_access = sp->role.access;
1059 pte_access &= FNAME(gpte_access)(gpte);
1060 FNAME(protect_clean_gpte)(vcpu->arch.mmu, &pte_access, gpte);
1061
1062 if (sync_mmio_spte(vcpu, &sp->spt[i], gfn, pte_access,
1063 &nr_present))
1064 continue;
1065
1066 if (gfn != sp->gfns[i]) {
1067 drop_spte(vcpu->kvm, &sp->spt[i]);
1068 /*
1069 * The same as above where we are doing
1070 * prefetch_invalid_gpte().
1071 */
1072 smp_wmb();
1073 vcpu->kvm->tlbs_dirty++;
1074 continue;
1075 }
1076
1077 nr_present++;
1078
1079 host_writable = sp->spt[i] & SPTE_HOST_WRITEABLE;
1080
1081 set_spte_ret |= set_spte(vcpu, &sp->spt[i],
1082 pte_access, PG_LEVEL_4K,
1083 gfn, spte_to_pfn(sp->spt[i]),
1084 true, false, host_writable);
1085 }
1086
1087 if (set_spte_ret & SET_SPTE_NEED_REMOTE_TLB_FLUSH)
1088 kvm_flush_remote_tlbs(vcpu->kvm);
1089
1090 return nr_present;
1091 }
1092
1093 #undef pt_element_t
1094 #undef guest_walker
1095 #undef FNAME
1096 #undef PT_BASE_ADDR_MASK
1097 #undef PT_INDEX
1098 #undef PT_LVL_ADDR_MASK
1099 #undef PT_LVL_OFFSET_MASK
1100 #undef PT_LEVEL_BITS
1101 #undef PT_MAX_FULL_LEVELS
1102 #undef gpte_to_gfn
1103 #undef gpte_to_gfn_lvl
1104 #undef CMPXCHG
1105 #undef PT_GUEST_ACCESSED_MASK
1106 #undef PT_GUEST_DIRTY_MASK
1107 #undef PT_GUEST_DIRTY_SHIFT
1108 #undef PT_GUEST_ACCESSED_SHIFT
1109 #undef PT_HAVE_ACCESSED_DIRTY
1110