• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
11  *
12  * Authors:
13  *   Yaniv Kamay  <yaniv@qumranet.com>
14  *   Avi Kivity   <avi@qumranet.com>
15  *
16  * This work is licensed under the terms of the GNU GPL, version 2.  See
17  * the COPYING file in the top-level directory.
18  *
19  */
20 
21 /*
22  * We need the mmu code to access both 32-bit and 64-bit guest ptes,
23  * so the code in this file is compiled twice, once per pte size.
24  */
25 
26 /*
27  * This is used to catch non optimized PT_GUEST_(DIRTY|ACCESS)_SHIFT macro
28  * uses for EPT without A/D paging type.
29  */
30 extern u64 __pure __using_nonexistent_pte_bit(void)
31 	       __compiletime_error("wrong use of PT_GUEST_(DIRTY|ACCESS)_SHIFT");
32 
33 #if PTTYPE == 64
34 	#define pt_element_t u64
35 	#define guest_walker guest_walker64
36 	#define FNAME(name) paging##64_##name
37 	#define PT_BASE_ADDR_MASK PT64_BASE_ADDR_MASK
38 	#define PT_LVL_ADDR_MASK(lvl) PT64_LVL_ADDR_MASK(lvl)
39 	#define PT_LVL_OFFSET_MASK(lvl) PT64_LVL_OFFSET_MASK(lvl)
40 	#define PT_INDEX(addr, level) PT64_INDEX(addr, level)
41 	#define PT_LEVEL_BITS PT64_LEVEL_BITS
42 	#define PT_GUEST_ACCESSED_MASK PT_ACCESSED_MASK
43 	#define PT_GUEST_DIRTY_MASK PT_DIRTY_MASK
44 	#define PT_GUEST_DIRTY_SHIFT PT_DIRTY_SHIFT
45 	#define PT_GUEST_ACCESSED_SHIFT PT_ACCESSED_SHIFT
46 	#ifdef CONFIG_X86_64
47 	#define PT_MAX_FULL_LEVELS 4
48 	#define CMPXCHG cmpxchg
49 	#else
50 	#define CMPXCHG cmpxchg64
51 	#define PT_MAX_FULL_LEVELS 2
52 	#endif
53 #elif PTTYPE == 32
54 	#define pt_element_t u32
55 	#define guest_walker guest_walker32
56 	#define FNAME(name) paging##32_##name
57 	#define PT_BASE_ADDR_MASK PT32_BASE_ADDR_MASK
58 	#define PT_LVL_ADDR_MASK(lvl) PT32_LVL_ADDR_MASK(lvl)
59 	#define PT_LVL_OFFSET_MASK(lvl) PT32_LVL_OFFSET_MASK(lvl)
60 	#define PT_INDEX(addr, level) PT32_INDEX(addr, level)
61 	#define PT_LEVEL_BITS PT32_LEVEL_BITS
62 	#define PT_MAX_FULL_LEVELS 2
63 	#define PT_GUEST_ACCESSED_MASK PT_ACCESSED_MASK
64 	#define PT_GUEST_DIRTY_MASK PT_DIRTY_MASK
65 	#define PT_GUEST_DIRTY_SHIFT PT_DIRTY_SHIFT
66 	#define PT_GUEST_ACCESSED_SHIFT PT_ACCESSED_SHIFT
67 	#define CMPXCHG cmpxchg
68 #elif PTTYPE == PTTYPE_EPT
69 	#define pt_element_t u64
70 	#define guest_walker guest_walkerEPT
71 	#define FNAME(name) ept_##name
72 	#define PT_BASE_ADDR_MASK PT64_BASE_ADDR_MASK
73 	#define PT_LVL_ADDR_MASK(lvl) PT64_LVL_ADDR_MASK(lvl)
74 	#define PT_LVL_OFFSET_MASK(lvl) PT64_LVL_OFFSET_MASK(lvl)
75 	#define PT_INDEX(addr, level) PT64_INDEX(addr, level)
76 	#define PT_LEVEL_BITS PT64_LEVEL_BITS
77 	#define PT_GUEST_ACCESSED_MASK 0
78 	#define PT_GUEST_DIRTY_MASK 0
79 	#define PT_GUEST_DIRTY_SHIFT __using_nonexistent_pte_bit()
80 	#define PT_GUEST_ACCESSED_SHIFT __using_nonexistent_pte_bit()
81 	#define CMPXCHG cmpxchg64
82 	#define PT_MAX_FULL_LEVELS 4
83 #else
84 	#error Invalid PTTYPE value
85 #endif
86 
87 #define gpte_to_gfn_lvl FNAME(gpte_to_gfn_lvl)
88 #define gpte_to_gfn(pte) gpte_to_gfn_lvl((pte), PT_PAGE_TABLE_LEVEL)
89 
90 /*
91  * The guest_walker structure emulates the behavior of the hardware page
92  * table walker.
93  */
94 struct guest_walker {
95 	int level;
96 	unsigned max_level;
97 	gfn_t table_gfn[PT_MAX_FULL_LEVELS];
98 	pt_element_t ptes[PT_MAX_FULL_LEVELS];
99 	pt_element_t prefetch_ptes[PTE_PREFETCH_NUM];
100 	gpa_t pte_gpa[PT_MAX_FULL_LEVELS];
101 	pt_element_t __user *ptep_user[PT_MAX_FULL_LEVELS];
102 	bool pte_writable[PT_MAX_FULL_LEVELS];
103 	unsigned pt_access;
104 	unsigned pte_access;
105 	gfn_t gfn;
106 	struct x86_exception fault;
107 };
108 
gpte_to_gfn_lvl(pt_element_t gpte,int lvl)109 static gfn_t gpte_to_gfn_lvl(pt_element_t gpte, int lvl)
110 {
111 	return (gpte & PT_LVL_ADDR_MASK(lvl)) >> PAGE_SHIFT;
112 }
113 
FNAME(protect_clean_gpte)114 static inline void FNAME(protect_clean_gpte)(unsigned *access, unsigned gpte)
115 {
116 	unsigned mask;
117 
118 	/* dirty bit is not supported, so no need to track it */
119 	if (!PT_GUEST_DIRTY_MASK)
120 		return;
121 
122 	BUILD_BUG_ON(PT_WRITABLE_MASK != ACC_WRITE_MASK);
123 
124 	mask = (unsigned)~ACC_WRITE_MASK;
125 	/* Allow write access to dirty gptes */
126 	mask |= (gpte >> (PT_GUEST_DIRTY_SHIFT - PT_WRITABLE_SHIFT)) &
127 		PT_WRITABLE_MASK;
128 	*access &= mask;
129 }
130 
FNAME(is_present_gpte)131 static inline int FNAME(is_present_gpte)(unsigned long pte)
132 {
133 #if PTTYPE != PTTYPE_EPT
134 	return is_present_gpte(pte);
135 #else
136 	return pte & 7;
137 #endif
138 }
139 
FNAME(cmpxchg_gpte)140 static int FNAME(cmpxchg_gpte)(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
141 			       pt_element_t __user *ptep_user, unsigned index,
142 			       pt_element_t orig_pte, pt_element_t new_pte)
143 {
144 	int npages;
145 	pt_element_t ret;
146 	pt_element_t *table;
147 	struct page *page;
148 
149 	npages = get_user_pages_fast((unsigned long)ptep_user, 1, 1, &page);
150 	/* Check if the user is doing something meaningless. */
151 	if (unlikely(npages != 1))
152 		return -EFAULT;
153 
154 	table = kmap_atomic(page);
155 	ret = CMPXCHG(&table[index], orig_pte, new_pte);
156 	kunmap_atomic(table);
157 
158 	kvm_release_page_dirty(page);
159 
160 	return (ret != orig_pte);
161 }
162 
FNAME(prefetch_invalid_gpte)163 static bool FNAME(prefetch_invalid_gpte)(struct kvm_vcpu *vcpu,
164 				  struct kvm_mmu_page *sp, u64 *spte,
165 				  u64 gpte)
166 {
167 	if (is_rsvd_bits_set(&vcpu->arch.mmu, gpte, PT_PAGE_TABLE_LEVEL))
168 		goto no_present;
169 
170 	if (!FNAME(is_present_gpte)(gpte))
171 		goto no_present;
172 
173 	/* if accessed bit is not supported prefetch non accessed gpte */
174 	if (PT_GUEST_ACCESSED_MASK && !(gpte & PT_GUEST_ACCESSED_MASK))
175 		goto no_present;
176 
177 	return false;
178 
179 no_present:
180 	drop_spte(vcpu->kvm, spte);
181 	return true;
182 }
183 
FNAME(gpte_access)184 static inline unsigned FNAME(gpte_access)(struct kvm_vcpu *vcpu, u64 gpte)
185 {
186 	unsigned access;
187 #if PTTYPE == PTTYPE_EPT
188 	access = ((gpte & VMX_EPT_WRITABLE_MASK) ? ACC_WRITE_MASK : 0) |
189 		((gpte & VMX_EPT_EXECUTABLE_MASK) ? ACC_EXEC_MASK : 0) |
190 		ACC_USER_MASK;
191 #else
192 	access = (gpte & (PT_WRITABLE_MASK | PT_USER_MASK)) | ACC_EXEC_MASK;
193 	access &= ~(gpte >> PT64_NX_SHIFT);
194 #endif
195 
196 	return access;
197 }
198 
FNAME(update_accessed_dirty_bits)199 static int FNAME(update_accessed_dirty_bits)(struct kvm_vcpu *vcpu,
200 					     struct kvm_mmu *mmu,
201 					     struct guest_walker *walker,
202 					     int write_fault)
203 {
204 	unsigned level, index;
205 	pt_element_t pte, orig_pte;
206 	pt_element_t __user *ptep_user;
207 	gfn_t table_gfn;
208 	int ret;
209 
210 	/* dirty/accessed bits are not supported, so no need to update them */
211 	if (!PT_GUEST_DIRTY_MASK)
212 		return 0;
213 
214 	for (level = walker->max_level; level >= walker->level; --level) {
215 		pte = orig_pte = walker->ptes[level - 1];
216 		table_gfn = walker->table_gfn[level - 1];
217 		ptep_user = walker->ptep_user[level - 1];
218 		index = offset_in_page(ptep_user) / sizeof(pt_element_t);
219 		if (!(pte & PT_GUEST_ACCESSED_MASK)) {
220 			trace_kvm_mmu_set_accessed_bit(table_gfn, index, sizeof(pte));
221 			pte |= PT_GUEST_ACCESSED_MASK;
222 		}
223 		if (level == walker->level && write_fault &&
224 				!(pte & PT_GUEST_DIRTY_MASK)) {
225 			trace_kvm_mmu_set_dirty_bit(table_gfn, index, sizeof(pte));
226 			pte |= PT_GUEST_DIRTY_MASK;
227 		}
228 		if (pte == orig_pte)
229 			continue;
230 
231 		/*
232 		 * If the slot is read-only, simply do not process the accessed
233 		 * and dirty bits.  This is the correct thing to do if the slot
234 		 * is ROM, and page tables in read-as-ROM/write-as-MMIO slots
235 		 * are only supported if the accessed and dirty bits are already
236 		 * set in the ROM (so that MMIO writes are never needed).
237 		 *
238 		 * Note that NPT does not allow this at all and faults, since
239 		 * it always wants nested page table entries for the guest
240 		 * page tables to be writable.  And EPT works but will simply
241 		 * overwrite the read-only memory to set the accessed and dirty
242 		 * bits.
243 		 */
244 		if (unlikely(!walker->pte_writable[level - 1]))
245 			continue;
246 
247 		ret = FNAME(cmpxchg_gpte)(vcpu, mmu, ptep_user, index, orig_pte, pte);
248 		if (ret)
249 			return ret;
250 
251 		kvm_vcpu_mark_page_dirty(vcpu, table_gfn);
252 		walker->ptes[level - 1] = pte;
253 	}
254 	return 0;
255 }
256 
257 /*
258  * Fetch a guest pte for a guest virtual address
259  */
FNAME(walk_addr_generic)260 static int FNAME(walk_addr_generic)(struct guest_walker *walker,
261 				    struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
262 				    gva_t addr, u32 access)
263 {
264 	int ret;
265 	pt_element_t pte;
266 	pt_element_t __user *uninitialized_var(ptep_user);
267 	gfn_t table_gfn;
268 	unsigned index, pt_access, pte_access, accessed_dirty;
269 	gpa_t pte_gpa;
270 	int offset;
271 	const int write_fault = access & PFERR_WRITE_MASK;
272 	const int user_fault  = access & PFERR_USER_MASK;
273 	const int fetch_fault = access & PFERR_FETCH_MASK;
274 	u16 errcode = 0;
275 	gpa_t real_gpa;
276 	gfn_t gfn;
277 
278 	trace_kvm_mmu_pagetable_walk(addr, access);
279 retry_walk:
280 	walker->level = mmu->root_level;
281 	pte           = mmu->get_cr3(vcpu);
282 
283 #if PTTYPE == 64
284 	if (walker->level == PT32E_ROOT_LEVEL) {
285 		pte = mmu->get_pdptr(vcpu, (addr >> 30) & 3);
286 		trace_kvm_mmu_paging_element(pte, walker->level);
287 		if (!FNAME(is_present_gpte)(pte))
288 			goto error;
289 		--walker->level;
290 	}
291 #endif
292 	walker->max_level = walker->level;
293 	ASSERT(!(is_long_mode(vcpu) && !is_pae(vcpu)));
294 
295 	accessed_dirty = PT_GUEST_ACCESSED_MASK;
296 	pt_access = pte_access = ACC_ALL;
297 	++walker->level;
298 
299 	do {
300 		gfn_t real_gfn;
301 		unsigned long host_addr;
302 
303 		pt_access &= pte_access;
304 		--walker->level;
305 
306 		index = PT_INDEX(addr, walker->level);
307 
308 		table_gfn = gpte_to_gfn(pte);
309 		offset    = index * sizeof(pt_element_t);
310 		pte_gpa   = gfn_to_gpa(table_gfn) + offset;
311 		walker->table_gfn[walker->level - 1] = table_gfn;
312 		walker->pte_gpa[walker->level - 1] = pte_gpa;
313 
314 		real_gfn = mmu->translate_gpa(vcpu, gfn_to_gpa(table_gfn),
315 					      PFERR_USER_MASK|PFERR_WRITE_MASK,
316 					      &walker->fault);
317 
318 		/*
319 		 * FIXME: This can happen if emulation (for of an INS/OUTS
320 		 * instruction) triggers a nested page fault.  The exit
321 		 * qualification / exit info field will incorrectly have
322 		 * "guest page access" as the nested page fault's cause,
323 		 * instead of "guest page structure access".  To fix this,
324 		 * the x86_exception struct should be augmented with enough
325 		 * information to fix the exit_qualification or exit_info_1
326 		 * fields.
327 		 */
328 		if (unlikely(real_gfn == UNMAPPED_GVA))
329 			return 0;
330 
331 		real_gfn = gpa_to_gfn(real_gfn);
332 
333 		host_addr = kvm_vcpu_gfn_to_hva_prot(vcpu, real_gfn,
334 					    &walker->pte_writable[walker->level - 1]);
335 		if (unlikely(kvm_is_error_hva(host_addr)))
336 			goto error;
337 
338 		ptep_user = (pt_element_t __user *)((void *)host_addr + offset);
339 		if (unlikely(__copy_from_user(&pte, ptep_user, sizeof(pte))))
340 			goto error;
341 		walker->ptep_user[walker->level - 1] = ptep_user;
342 
343 		trace_kvm_mmu_paging_element(pte, walker->level);
344 
345 		if (unlikely(!FNAME(is_present_gpte)(pte)))
346 			goto error;
347 
348 		if (unlikely(is_rsvd_bits_set(mmu, pte, walker->level))) {
349 			errcode |= PFERR_RSVD_MASK | PFERR_PRESENT_MASK;
350 			goto error;
351 		}
352 
353 		accessed_dirty &= pte;
354 		pte_access = pt_access & FNAME(gpte_access)(vcpu, pte);
355 
356 		walker->ptes[walker->level - 1] = pte;
357 	} while (!is_last_gpte(mmu, walker->level, pte));
358 
359 	if (unlikely(permission_fault(vcpu, mmu, pte_access, access))) {
360 		errcode |= PFERR_PRESENT_MASK;
361 		goto error;
362 	}
363 
364 	gfn = gpte_to_gfn_lvl(pte, walker->level);
365 	gfn += (addr & PT_LVL_OFFSET_MASK(walker->level)) >> PAGE_SHIFT;
366 
367 	if (PTTYPE == 32 && walker->level == PT_DIRECTORY_LEVEL && is_cpuid_PSE36())
368 		gfn += pse36_gfn_delta(pte);
369 
370 	real_gpa = mmu->translate_gpa(vcpu, gfn_to_gpa(gfn), access, &walker->fault);
371 	if (real_gpa == UNMAPPED_GVA)
372 		return 0;
373 
374 	walker->gfn = real_gpa >> PAGE_SHIFT;
375 
376 	if (!write_fault)
377 		FNAME(protect_clean_gpte)(&pte_access, pte);
378 	else
379 		/*
380 		 * On a write fault, fold the dirty bit into accessed_dirty.
381 		 * For modes without A/D bits support accessed_dirty will be
382 		 * always clear.
383 		 */
384 		accessed_dirty &= pte >>
385 			(PT_GUEST_DIRTY_SHIFT - PT_GUEST_ACCESSED_SHIFT);
386 
387 	if (unlikely(!accessed_dirty)) {
388 		ret = FNAME(update_accessed_dirty_bits)(vcpu, mmu, walker, write_fault);
389 		if (unlikely(ret < 0))
390 			goto error;
391 		else if (ret)
392 			goto retry_walk;
393 	}
394 
395 	walker->pt_access = pt_access;
396 	walker->pte_access = pte_access;
397 	pgprintk("%s: pte %llx pte_access %x pt_access %x\n",
398 		 __func__, (u64)pte, pte_access, pt_access);
399 	return 1;
400 
401 error:
402 	errcode |= write_fault | user_fault;
403 	if (fetch_fault && (mmu->nx ||
404 			    kvm_read_cr4_bits(vcpu, X86_CR4_SMEP)))
405 		errcode |= PFERR_FETCH_MASK;
406 
407 	walker->fault.vector = PF_VECTOR;
408 	walker->fault.error_code_valid = true;
409 	walker->fault.error_code = errcode;
410 
411 #if PTTYPE == PTTYPE_EPT
412 	/*
413 	 * Use PFERR_RSVD_MASK in error_code to to tell if EPT
414 	 * misconfiguration requires to be injected. The detection is
415 	 * done by is_rsvd_bits_set() above.
416 	 *
417 	 * We set up the value of exit_qualification to inject:
418 	 * [2:0] - Derive from [2:0] of real exit_qualification at EPT violation
419 	 * [5:3] - Calculated by the page walk of the guest EPT page tables
420 	 * [7:8] - Derived from [7:8] of real exit_qualification
421 	 *
422 	 * The other bits are set to 0.
423 	 */
424 	if (!(errcode & PFERR_RSVD_MASK)) {
425 		vcpu->arch.exit_qualification &= 0x187;
426 		vcpu->arch.exit_qualification |= ((pt_access & pte) & 0x7) << 3;
427 	}
428 #endif
429 	walker->fault.address = addr;
430 	walker->fault.nested_page_fault = mmu != vcpu->arch.walk_mmu;
431 
432 	trace_kvm_mmu_walker_error(walker->fault.error_code);
433 	return 0;
434 }
435 
FNAME(walk_addr)436 static int FNAME(walk_addr)(struct guest_walker *walker,
437 			    struct kvm_vcpu *vcpu, gva_t addr, u32 access)
438 {
439 	return FNAME(walk_addr_generic)(walker, vcpu, &vcpu->arch.mmu, addr,
440 					access);
441 }
442 
443 #if PTTYPE != PTTYPE_EPT
FNAME(walk_addr_nested)444 static int FNAME(walk_addr_nested)(struct guest_walker *walker,
445 				   struct kvm_vcpu *vcpu, gva_t addr,
446 				   u32 access)
447 {
448 	return FNAME(walk_addr_generic)(walker, vcpu, &vcpu->arch.nested_mmu,
449 					addr, access);
450 }
451 #endif
452 
453 static bool
FNAME(prefetch_gpte)454 FNAME(prefetch_gpte)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
455 		     u64 *spte, pt_element_t gpte, bool no_dirty_log)
456 {
457 	unsigned pte_access;
458 	gfn_t gfn;
459 	pfn_t pfn;
460 
461 	if (FNAME(prefetch_invalid_gpte)(vcpu, sp, spte, gpte))
462 		return false;
463 
464 	pgprintk("%s: gpte %llx spte %p\n", __func__, (u64)gpte, spte);
465 
466 	gfn = gpte_to_gfn(gpte);
467 	pte_access = sp->role.access & FNAME(gpte_access)(vcpu, gpte);
468 	FNAME(protect_clean_gpte)(&pte_access, gpte);
469 	pfn = pte_prefetch_gfn_to_pfn(vcpu, gfn,
470 			no_dirty_log && (pte_access & ACC_WRITE_MASK));
471 	if (is_error_pfn(pfn))
472 		return false;
473 
474 	/*
475 	 * we call mmu_set_spte() with host_writable = true because
476 	 * pte_prefetch_gfn_to_pfn always gets a writable pfn.
477 	 */
478 	mmu_set_spte(vcpu, spte, pte_access, 0, NULL, PT_PAGE_TABLE_LEVEL,
479 		     gfn, pfn, true, true);
480 
481 	return true;
482 }
483 
FNAME(update_pte)484 static void FNAME(update_pte)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
485 			      u64 *spte, const void *pte)
486 {
487 	pt_element_t gpte = *(const pt_element_t *)pte;
488 
489 	FNAME(prefetch_gpte)(vcpu, sp, spte, gpte, false);
490 }
491 
FNAME(gpte_changed)492 static bool FNAME(gpte_changed)(struct kvm_vcpu *vcpu,
493 				struct guest_walker *gw, int level)
494 {
495 	pt_element_t curr_pte;
496 	gpa_t base_gpa, pte_gpa = gw->pte_gpa[level - 1];
497 	u64 mask;
498 	int r, index;
499 
500 	if (level == PT_PAGE_TABLE_LEVEL) {
501 		mask = PTE_PREFETCH_NUM * sizeof(pt_element_t) - 1;
502 		base_gpa = pte_gpa & ~mask;
503 		index = (pte_gpa - base_gpa) / sizeof(pt_element_t);
504 
505 		r = kvm_vcpu_read_guest_atomic(vcpu, base_gpa,
506 				gw->prefetch_ptes, sizeof(gw->prefetch_ptes));
507 		curr_pte = gw->prefetch_ptes[index];
508 	} else
509 		r = kvm_vcpu_read_guest_atomic(vcpu, pte_gpa,
510 				  &curr_pte, sizeof(curr_pte));
511 
512 	return r || curr_pte != gw->ptes[level - 1];
513 }
514 
FNAME(pte_prefetch)515 static void FNAME(pte_prefetch)(struct kvm_vcpu *vcpu, struct guest_walker *gw,
516 				u64 *sptep)
517 {
518 	struct kvm_mmu_page *sp;
519 	pt_element_t *gptep = gw->prefetch_ptes;
520 	u64 *spte;
521 	int i;
522 
523 	sp = page_header(__pa(sptep));
524 
525 	if (sp->role.level > PT_PAGE_TABLE_LEVEL)
526 		return;
527 
528 	if (sp->role.direct)
529 		return __direct_pte_prefetch(vcpu, sp, sptep);
530 
531 	i = (sptep - sp->spt) & ~(PTE_PREFETCH_NUM - 1);
532 	spte = sp->spt + i;
533 
534 	for (i = 0; i < PTE_PREFETCH_NUM; i++, spte++) {
535 		if (spte == sptep)
536 			continue;
537 
538 		if (is_shadow_present_pte(*spte))
539 			continue;
540 
541 		if (!FNAME(prefetch_gpte)(vcpu, sp, spte, gptep[i], true))
542 			break;
543 	}
544 }
545 
546 /*
547  * Fetch a shadow pte for a specific level in the paging hierarchy.
548  * If the guest tries to write a write-protected page, we need to
549  * emulate this operation, return 1 to indicate this case.
550  */
FNAME(fetch)551 static int FNAME(fetch)(struct kvm_vcpu *vcpu, gva_t addr,
552 			 struct guest_walker *gw,
553 			 int write_fault, int hlevel,
554 			 pfn_t pfn, bool map_writable, bool prefault)
555 {
556 	struct kvm_mmu_page *sp = NULL;
557 	struct kvm_shadow_walk_iterator it;
558 	unsigned direct_access, access = gw->pt_access;
559 	int top_level, emulate = 0;
560 
561 	direct_access = gw->pte_access;
562 
563 	top_level = vcpu->arch.mmu.root_level;
564 	if (top_level == PT32E_ROOT_LEVEL)
565 		top_level = PT32_ROOT_LEVEL;
566 	/*
567 	 * Verify that the top-level gpte is still there.  Since the page
568 	 * is a root page, it is either write protected (and cannot be
569 	 * changed from now on) or it is invalid (in which case, we don't
570 	 * really care if it changes underneath us after this point).
571 	 */
572 	if (FNAME(gpte_changed)(vcpu, gw, top_level))
573 		goto out_gpte_changed;
574 
575 	if (!VALID_PAGE(vcpu->arch.mmu.root_hpa))
576 		goto out_gpte_changed;
577 
578 	for (shadow_walk_init(&it, vcpu, addr);
579 	     shadow_walk_okay(&it) && it.level > gw->level;
580 	     shadow_walk_next(&it)) {
581 		gfn_t table_gfn;
582 
583 		clear_sp_write_flooding_count(it.sptep);
584 		drop_large_spte(vcpu, it.sptep);
585 
586 		sp = NULL;
587 		if (!is_shadow_present_pte(*it.sptep)) {
588 			table_gfn = gw->table_gfn[it.level - 2];
589 			sp = kvm_mmu_get_page(vcpu, table_gfn, addr, it.level-1,
590 					      false, access, it.sptep);
591 		}
592 
593 		/*
594 		 * Verify that the gpte in the page we've just write
595 		 * protected is still there.
596 		 */
597 		if (FNAME(gpte_changed)(vcpu, gw, it.level - 1))
598 			goto out_gpte_changed;
599 
600 		if (sp)
601 			link_shadow_page(it.sptep, sp, PT_GUEST_ACCESSED_MASK);
602 	}
603 
604 	for (;
605 	     shadow_walk_okay(&it) && it.level > hlevel;
606 	     shadow_walk_next(&it)) {
607 		gfn_t direct_gfn;
608 
609 		clear_sp_write_flooding_count(it.sptep);
610 		validate_direct_spte(vcpu, it.sptep, direct_access);
611 
612 		drop_large_spte(vcpu, it.sptep);
613 
614 		if (is_shadow_present_pte(*it.sptep))
615 			continue;
616 
617 		direct_gfn = gw->gfn & ~(KVM_PAGES_PER_HPAGE(it.level) - 1);
618 
619 		sp = kvm_mmu_get_page(vcpu, direct_gfn, addr, it.level-1,
620 				      true, direct_access, it.sptep);
621 		link_shadow_page(it.sptep, sp, PT_GUEST_ACCESSED_MASK);
622 	}
623 
624 	clear_sp_write_flooding_count(it.sptep);
625 	mmu_set_spte(vcpu, it.sptep, gw->pte_access, write_fault, &emulate,
626 		     it.level, gw->gfn, pfn, prefault, map_writable);
627 	FNAME(pte_prefetch)(vcpu, gw, it.sptep);
628 
629 	return emulate;
630 
631 out_gpte_changed:
632 	if (sp)
633 		kvm_mmu_put_page(sp, it.sptep);
634 	kvm_release_pfn_clean(pfn);
635 	return 0;
636 }
637 
638  /*
639  * To see whether the mapped gfn can write its page table in the current
640  * mapping.
641  *
642  * It is the helper function of FNAME(page_fault). When guest uses large page
643  * size to map the writable gfn which is used as current page table, we should
644  * force kvm to use small page size to map it because new shadow page will be
645  * created when kvm establishes shadow page table that stop kvm using large
646  * page size. Do it early can avoid unnecessary #PF and emulation.
647  *
648  * @write_fault_to_shadow_pgtable will return true if the fault gfn is
649  * currently used as its page table.
650  *
651  * Note: the PDPT page table is not checked for PAE-32 bit guest. It is ok
652  * since the PDPT is always shadowed, that means, we can not use large page
653  * size to map the gfn which is used as PDPT.
654  */
655 static bool
FNAME(is_self_change_mapping)656 FNAME(is_self_change_mapping)(struct kvm_vcpu *vcpu,
657 			      struct guest_walker *walker, int user_fault,
658 			      bool *write_fault_to_shadow_pgtable)
659 {
660 	int level;
661 	gfn_t mask = ~(KVM_PAGES_PER_HPAGE(walker->level) - 1);
662 	bool self_changed = false;
663 
664 	if (!(walker->pte_access & ACC_WRITE_MASK ||
665 	      (!is_write_protection(vcpu) && !user_fault)))
666 		return false;
667 
668 	for (level = walker->level; level <= walker->max_level; level++) {
669 		gfn_t gfn = walker->gfn ^ walker->table_gfn[level - 1];
670 
671 		self_changed |= !(gfn & mask);
672 		*write_fault_to_shadow_pgtable |= !gfn;
673 	}
674 
675 	return self_changed;
676 }
677 
678 /*
679  * Page fault handler.  There are several causes for a page fault:
680  *   - there is no shadow pte for the guest pte
681  *   - write access through a shadow pte marked read only so that we can set
682  *     the dirty bit
683  *   - write access to a shadow pte marked read only so we can update the page
684  *     dirty bitmap, when userspace requests it
685  *   - mmio access; in this case we will never install a present shadow pte
686  *   - normal guest page fault due to the guest pte marked not present, not
687  *     writable, or not executable
688  *
689  *  Returns: 1 if we need to emulate the instruction, 0 otherwise, or
690  *           a negative value on error.
691  */
FNAME(page_fault)692 static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gva_t addr, u32 error_code,
693 			     bool prefault)
694 {
695 	int write_fault = error_code & PFERR_WRITE_MASK;
696 	int user_fault = error_code & PFERR_USER_MASK;
697 	struct guest_walker walker;
698 	int r;
699 	pfn_t pfn;
700 	int level = PT_PAGE_TABLE_LEVEL;
701 	bool force_pt_level = false;
702 	unsigned long mmu_seq;
703 	bool map_writable, is_self_change_mapping;
704 
705 	pgprintk("%s: addr %lx err %x\n", __func__, addr, error_code);
706 
707 	if (unlikely(error_code & PFERR_RSVD_MASK)) {
708 		r = handle_mmio_page_fault(vcpu, addr, mmu_is_nested(vcpu));
709 		if (likely(r != RET_MMIO_PF_INVALID))
710 			return r;
711 
712 		/*
713 		 * page fault with PFEC.RSVD  = 1 is caused by shadow
714 		 * page fault, should not be used to walk guest page
715 		 * table.
716 		 */
717 		error_code &= ~PFERR_RSVD_MASK;
718 	};
719 
720 	r = mmu_topup_memory_caches(vcpu);
721 	if (r)
722 		return r;
723 
724 	/*
725 	 * Look up the guest pte for the faulting address.
726 	 */
727 	r = FNAME(walk_addr)(&walker, vcpu, addr, error_code);
728 
729 	/*
730 	 * The page is not mapped by the guest.  Let the guest handle it.
731 	 */
732 	if (!r) {
733 		pgprintk("%s: guest page fault\n", __func__);
734 		if (!prefault)
735 			inject_page_fault(vcpu, &walker.fault);
736 
737 		return 0;
738 	}
739 
740 	vcpu->arch.write_fault_to_shadow_pgtable = false;
741 
742 	is_self_change_mapping = FNAME(is_self_change_mapping)(vcpu,
743 	      &walker, user_fault, &vcpu->arch.write_fault_to_shadow_pgtable);
744 
745 	if (walker.level >= PT_DIRECTORY_LEVEL && !is_self_change_mapping) {
746 		level = mapping_level(vcpu, walker.gfn, &force_pt_level);
747 		if (likely(!force_pt_level)) {
748 			level = min(walker.level, level);
749 			walker.gfn = walker.gfn & ~(KVM_PAGES_PER_HPAGE(level) - 1);
750 		}
751 	} else
752 		force_pt_level = true;
753 
754 	mmu_seq = vcpu->kvm->mmu_notifier_seq;
755 	smp_rmb();
756 
757 	if (try_async_pf(vcpu, prefault, walker.gfn, addr, &pfn, write_fault,
758 			 &map_writable))
759 		return 0;
760 
761 	if (handle_abnormal_pfn(vcpu, mmu_is_nested(vcpu) ? 0 : addr,
762 				walker.gfn, pfn, walker.pte_access, &r))
763 		return r;
764 
765 	/*
766 	 * Do not change pte_access if the pfn is a mmio page, otherwise
767 	 * we will cache the incorrect access into mmio spte.
768 	 */
769 	if (write_fault && !(walker.pte_access & ACC_WRITE_MASK) &&
770 	     !is_write_protection(vcpu) && !user_fault &&
771 	      !is_noslot_pfn(pfn)) {
772 		walker.pte_access |= ACC_WRITE_MASK;
773 		walker.pte_access &= ~ACC_USER_MASK;
774 
775 		/*
776 		 * If we converted a user page to a kernel page,
777 		 * so that the kernel can write to it when cr0.wp=0,
778 		 * then we should prevent the kernel from executing it
779 		 * if SMEP is enabled.
780 		 */
781 		if (kvm_read_cr4_bits(vcpu, X86_CR4_SMEP))
782 			walker.pte_access &= ~ACC_EXEC_MASK;
783 	}
784 
785 	spin_lock(&vcpu->kvm->mmu_lock);
786 	if (mmu_notifier_retry(vcpu->kvm, mmu_seq))
787 		goto out_unlock;
788 
789 	kvm_mmu_audit(vcpu, AUDIT_PRE_PAGE_FAULT);
790 	make_mmu_pages_available(vcpu);
791 	if (!force_pt_level)
792 		transparent_hugepage_adjust(vcpu, &walker.gfn, &pfn, &level);
793 	r = FNAME(fetch)(vcpu, addr, &walker, write_fault,
794 			 level, pfn, map_writable, prefault);
795 	++vcpu->stat.pf_fixed;
796 	kvm_mmu_audit(vcpu, AUDIT_POST_PAGE_FAULT);
797 	spin_unlock(&vcpu->kvm->mmu_lock);
798 
799 	return r;
800 
801 out_unlock:
802 	spin_unlock(&vcpu->kvm->mmu_lock);
803 	kvm_release_pfn_clean(pfn);
804 	return 0;
805 }
806 
FNAME(get_level1_sp_gpa)807 static gpa_t FNAME(get_level1_sp_gpa)(struct kvm_mmu_page *sp)
808 {
809 	int offset = 0;
810 
811 	WARN_ON(sp->role.level != PT_PAGE_TABLE_LEVEL);
812 
813 	if (PTTYPE == 32)
814 		offset = sp->role.quadrant << PT64_LEVEL_BITS;
815 
816 	return gfn_to_gpa(sp->gfn) + offset * sizeof(pt_element_t);
817 }
818 
FNAME(invlpg)819 static void FNAME(invlpg)(struct kvm_vcpu *vcpu, gva_t gva)
820 {
821 	struct kvm_shadow_walk_iterator iterator;
822 	struct kvm_mmu_page *sp;
823 	int level;
824 	u64 *sptep;
825 
826 	vcpu_clear_mmio_info(vcpu, gva);
827 
828 	/*
829 	 * No need to check return value here, rmap_can_add() can
830 	 * help us to skip pte prefetch later.
831 	 */
832 	mmu_topup_memory_caches(vcpu);
833 
834 	if (!VALID_PAGE(vcpu->arch.mmu.root_hpa)) {
835 		WARN_ON(1);
836 		return;
837 	}
838 
839 	spin_lock(&vcpu->kvm->mmu_lock);
840 	for_each_shadow_entry(vcpu, gva, iterator) {
841 		level = iterator.level;
842 		sptep = iterator.sptep;
843 
844 		sp = page_header(__pa(sptep));
845 		if (is_last_spte(*sptep, level)) {
846 			pt_element_t gpte;
847 			gpa_t pte_gpa;
848 
849 			if (!sp->unsync)
850 				break;
851 
852 			pte_gpa = FNAME(get_level1_sp_gpa)(sp);
853 			pte_gpa += (sptep - sp->spt) * sizeof(pt_element_t);
854 
855 			if (mmu_page_zap_pte(vcpu->kvm, sp, sptep))
856 				kvm_flush_remote_tlbs(vcpu->kvm);
857 
858 			if (!rmap_can_add(vcpu))
859 				break;
860 
861 			if (kvm_vcpu_read_guest_atomic(vcpu, pte_gpa, &gpte,
862 						       sizeof(pt_element_t)))
863 				break;
864 
865 			FNAME(update_pte)(vcpu, sp, sptep, &gpte);
866 		}
867 
868 		if (!is_shadow_present_pte(*sptep) || !sp->unsync_children)
869 			break;
870 	}
871 	spin_unlock(&vcpu->kvm->mmu_lock);
872 }
873 
FNAME(gva_to_gpa)874 static gpa_t FNAME(gva_to_gpa)(struct kvm_vcpu *vcpu, gva_t vaddr, u32 access,
875 			       struct x86_exception *exception)
876 {
877 	struct guest_walker walker;
878 	gpa_t gpa = UNMAPPED_GVA;
879 	int r;
880 
881 	r = FNAME(walk_addr)(&walker, vcpu, vaddr, access);
882 
883 	if (r) {
884 		gpa = gfn_to_gpa(walker.gfn);
885 		gpa |= vaddr & ~PAGE_MASK;
886 	} else if (exception)
887 		*exception = walker.fault;
888 
889 	return gpa;
890 }
891 
892 #if PTTYPE != PTTYPE_EPT
FNAME(gva_to_gpa_nested)893 static gpa_t FNAME(gva_to_gpa_nested)(struct kvm_vcpu *vcpu, gva_t vaddr,
894 				      u32 access,
895 				      struct x86_exception *exception)
896 {
897 	struct guest_walker walker;
898 	gpa_t gpa = UNMAPPED_GVA;
899 	int r;
900 
901 	r = FNAME(walk_addr_nested)(&walker, vcpu, vaddr, access);
902 
903 	if (r) {
904 		gpa = gfn_to_gpa(walker.gfn);
905 		gpa |= vaddr & ~PAGE_MASK;
906 	} else if (exception)
907 		*exception = walker.fault;
908 
909 	return gpa;
910 }
911 #endif
912 
913 /*
914  * Using the cached information from sp->gfns is safe because:
915  * - The spte has a reference to the struct page, so the pfn for a given gfn
916  *   can't change unless all sptes pointing to it are nuked first.
917  *
918  * Note:
919  *   We should flush all tlbs if spte is dropped even though guest is
920  *   responsible for it. Since if we don't, kvm_mmu_notifier_invalidate_page
921  *   and kvm_mmu_notifier_invalidate_range_start detect the mapping page isn't
922  *   used by guest then tlbs are not flushed, so guest is allowed to access the
923  *   freed pages.
924  *   And we increase kvm->tlbs_dirty to delay tlbs flush in this case.
925  */
FNAME(sync_page)926 static int FNAME(sync_page)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
927 {
928 	int i, nr_present = 0;
929 	bool host_writable;
930 	gpa_t first_pte_gpa;
931 
932 	/* direct kvm_mmu_page can not be unsync. */
933 	BUG_ON(sp->role.direct);
934 
935 	first_pte_gpa = FNAME(get_level1_sp_gpa)(sp);
936 
937 	for (i = 0; i < PT64_ENT_PER_PAGE; i++) {
938 		unsigned pte_access;
939 		pt_element_t gpte;
940 		gpa_t pte_gpa;
941 		gfn_t gfn;
942 
943 		if (!sp->spt[i])
944 			continue;
945 
946 		pte_gpa = first_pte_gpa + i * sizeof(pt_element_t);
947 
948 		if (kvm_vcpu_read_guest_atomic(vcpu, pte_gpa, &gpte,
949 					       sizeof(pt_element_t)))
950 			return -EINVAL;
951 
952 		if (FNAME(prefetch_invalid_gpte)(vcpu, sp, &sp->spt[i], gpte)) {
953 			vcpu->kvm->tlbs_dirty++;
954 			continue;
955 		}
956 
957 		gfn = gpte_to_gfn(gpte);
958 		pte_access = sp->role.access;
959 		pte_access &= FNAME(gpte_access)(vcpu, gpte);
960 		FNAME(protect_clean_gpte)(&pte_access, gpte);
961 
962 		if (sync_mmio_spte(vcpu, &sp->spt[i], gfn, pte_access,
963 		      &nr_present))
964 			continue;
965 
966 		if (gfn != sp->gfns[i]) {
967 			drop_spte(vcpu->kvm, &sp->spt[i]);
968 			vcpu->kvm->tlbs_dirty++;
969 			continue;
970 		}
971 
972 		nr_present++;
973 
974 		host_writable = sp->spt[i] & SPTE_HOST_WRITEABLE;
975 
976 		set_spte(vcpu, &sp->spt[i], pte_access,
977 			 PT_PAGE_TABLE_LEVEL, gfn,
978 			 spte_to_pfn(sp->spt[i]), true, false,
979 			 host_writable);
980 	}
981 
982 	return !nr_present;
983 }
984 
985 #undef pt_element_t
986 #undef guest_walker
987 #undef FNAME
988 #undef PT_BASE_ADDR_MASK
989 #undef PT_INDEX
990 #undef PT_LVL_ADDR_MASK
991 #undef PT_LVL_OFFSET_MASK
992 #undef PT_LEVEL_BITS
993 #undef PT_MAX_FULL_LEVELS
994 #undef gpte_to_gfn
995 #undef gpte_to_gfn_lvl
996 #undef CMPXCHG
997 #undef PT_GUEST_ACCESSED_MASK
998 #undef PT_GUEST_DIRTY_MASK
999 #undef PT_GUEST_DIRTY_SHIFT
1000 #undef PT_GUEST_ACCESSED_SHIFT
1001