1 // SPDX-License-Identifier: GPL-2.0-only
3 * Kernel-based Virtual Machine driver for Linux
5 * This module enables machines with Intel VT-x extensions to run virtual
6 * machines without emulation or binary translation.
10 * Copyright (C) 2006 Qumranet, Inc.
11 * Copyright 2010 Red Hat, Inc. and/or its affiliates.
14 * Yaniv Kamay <yaniv@qumranet.com>
15 * Avi Kivity <avi@qumranet.com>
21 #include "kvm_cache_regs.h"
22 #include "kvm_emulate.h"
25 #include <linux/kvm_host.h>
26 #include <linux/types.h>
27 #include <linux/string.h>
29 #include <linux/highmem.h>
30 #include <linux/moduleparam.h>
31 #include <linux/export.h>
32 #include <linux/swap.h>
33 #include <linux/hugetlb.h>
34 #include <linux/compiler.h>
35 #include <linux/srcu.h>
36 #include <linux/slab.h>
37 #include <linux/sched/signal.h>
38 #include <linux/uaccess.h>
39 #include <linux/hash.h>
40 #include <linux/kern_levels.h>
41 #include <linux/kthread.h>
44 #include <asm/memtype.h>
45 #include <asm/cmpxchg.h>
46 #include <asm/e820/api.h>
49 #include <asm/kvm_page_track.h>
52 extern bool itlb_multihit_kvm_mitigation;
54 static int __read_mostly nx_huge_pages = -1;
55 #ifdef CONFIG_PREEMPT_RT
56 /* Recovery can cause latency spikes, disable it for PREEMPT_RT. */
57 static uint __read_mostly nx_huge_pages_recovery_ratio = 0;
59 static uint __read_mostly nx_huge_pages_recovery_ratio = 60;
62 static int set_nx_huge_pages(const char *val, const struct kernel_param *kp);
63 static int set_nx_huge_pages_recovery_ratio(const char *val, const struct kernel_param *kp);
65 static struct kernel_param_ops nx_huge_pages_ops = {
66 .set = set_nx_huge_pages,
67 .get = param_get_bool,
70 static struct kernel_param_ops nx_huge_pages_recovery_ratio_ops = {
71 .set = set_nx_huge_pages_recovery_ratio,
72 .get = param_get_uint,
75 module_param_cb(nx_huge_pages, &nx_huge_pages_ops, &nx_huge_pages, 0644);
76 __MODULE_PARM_TYPE(nx_huge_pages, "bool");
77 module_param_cb(nx_huge_pages_recovery_ratio, &nx_huge_pages_recovery_ratio_ops,
78 &nx_huge_pages_recovery_ratio, 0644);
79 __MODULE_PARM_TYPE(nx_huge_pages_recovery_ratio, "uint");
82 * When setting this variable to true it enables Two-Dimensional-Paging
83 * where the hardware walks 2 page tables:
84 * 1. the guest-virtual to guest-physical
85 * 2. while doing 1. it walks guest-physical to host-physical
86 * If the hardware supports that we don't need to do shadow paging.
88 bool tdp_enabled = false;
90 static int max_page_level __read_mostly;
94 AUDIT_POST_PAGE_FAULT,
105 module_param(dbg, bool, 0644);
107 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
108 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
109 #define MMU_WARN_ON(x) WARN_ON(x)
111 #define pgprintk(x...) do { } while (0)
112 #define rmap_printk(x...) do { } while (0)
113 #define MMU_WARN_ON(x) do { } while (0)
116 #define PTE_PREFETCH_NUM 8
118 #define PT_FIRST_AVAIL_BITS_SHIFT 10
119 #define PT64_SECOND_AVAIL_BITS_SHIFT 54
122 * The mask used to denote special SPTEs, which can be either MMIO SPTEs or
123 * Access Tracking SPTEs.
125 #define SPTE_SPECIAL_MASK (3ULL << 52)
126 #define SPTE_AD_ENABLED_MASK (0ULL << 52)
127 #define SPTE_AD_DISABLED_MASK (1ULL << 52)
128 #define SPTE_AD_WRPROT_ONLY_MASK (2ULL << 52)
129 #define SPTE_MMIO_MASK (3ULL << 52)
131 #define PT64_LEVEL_BITS 9
133 #define PT64_LEVEL_SHIFT(level) \
134 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
136 #define PT64_INDEX(address, level)\
137 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
140 #define PT32_LEVEL_BITS 10
142 #define PT32_LEVEL_SHIFT(level) \
143 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
145 #define PT32_LVL_OFFSET_MASK(level) \
146 (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
147 * PT32_LEVEL_BITS))) - 1))
149 #define PT32_INDEX(address, level)\
150 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
153 #ifdef CONFIG_DYNAMIC_PHYSICAL_MASK
154 #define PT64_BASE_ADDR_MASK (physical_mask & ~(u64)(PAGE_SIZE-1))
156 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
158 #define PT64_LVL_ADDR_MASK(level) \
159 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
160 * PT64_LEVEL_BITS))) - 1))
161 #define PT64_LVL_OFFSET_MASK(level) \
162 (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
163 * PT64_LEVEL_BITS))) - 1))
165 #define PT32_BASE_ADDR_MASK PAGE_MASK
166 #define PT32_DIR_BASE_ADDR_MASK \
167 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
168 #define PT32_LVL_ADDR_MASK(level) \
169 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
170 * PT32_LEVEL_BITS))) - 1))
172 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | shadow_user_mask \
173 | shadow_x_mask | shadow_nx_mask | shadow_me_mask)
175 #define ACC_EXEC_MASK 1
176 #define ACC_WRITE_MASK PT_WRITABLE_MASK
177 #define ACC_USER_MASK PT_USER_MASK
178 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
180 /* The mask for the R/X bits in EPT PTEs */
181 #define PT64_EPT_READABLE_MASK 0x1ull
182 #define PT64_EPT_EXECUTABLE_MASK 0x4ull
184 #include <trace/events/kvm.h>
186 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
187 #define SPTE_MMU_WRITEABLE (1ULL << (PT_FIRST_AVAIL_BITS_SHIFT + 1))
189 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
191 /* make pte_list_desc fit well in cache line */
192 #define PTE_LIST_EXT 3
195 * Return values of handle_mmio_page_fault and mmu.page_fault:
196 * RET_PF_RETRY: let CPU fault again on the address.
197 * RET_PF_EMULATE: mmio page fault, emulate the instruction directly.
199 * For handle_mmio_page_fault only:
200 * RET_PF_INVALID: the spte is invalid, let the real page fault path update it.
208 struct pte_list_desc {
209 u64 *sptes[PTE_LIST_EXT];
210 struct pte_list_desc *more;
213 struct kvm_shadow_walk_iterator {
221 #define for_each_shadow_entry_using_root(_vcpu, _root, _addr, _walker) \
222 for (shadow_walk_init_using_root(&(_walker), (_vcpu), \
224 shadow_walk_okay(&(_walker)); \
225 shadow_walk_next(&(_walker)))
227 #define for_each_shadow_entry(_vcpu, _addr, _walker) \
228 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
229 shadow_walk_okay(&(_walker)); \
230 shadow_walk_next(&(_walker)))
232 #define for_each_shadow_entry_lockless(_vcpu, _addr, _walker, spte) \
233 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
234 shadow_walk_okay(&(_walker)) && \
235 ({ spte = mmu_spte_get_lockless(_walker.sptep); 1; }); \
236 __shadow_walk_next(&(_walker), spte))
238 static struct kmem_cache *pte_list_desc_cache;
239 static struct kmem_cache *mmu_page_header_cache;
240 static struct percpu_counter kvm_total_used_mmu_pages;
242 static u64 __read_mostly shadow_nx_mask;
243 static u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
244 static u64 __read_mostly shadow_user_mask;
245 static u64 __read_mostly shadow_accessed_mask;
246 static u64 __read_mostly shadow_dirty_mask;
247 static u64 __read_mostly shadow_mmio_mask;
248 static u64 __read_mostly shadow_mmio_value;
249 static u64 __read_mostly shadow_mmio_access_mask;
250 static u64 __read_mostly shadow_present_mask;
251 static u64 __read_mostly shadow_me_mask;
254 * SPTEs used by MMUs without A/D bits are marked with SPTE_AD_DISABLED_MASK;
255 * shadow_acc_track_mask is the set of bits to be cleared in non-accessed
258 static u64 __read_mostly shadow_acc_track_mask;
261 * The mask/shift to use for saving the original R/X bits when marking the PTE
262 * as not-present for access tracking purposes. We do not save the W bit as the
263 * PTEs being access tracked also need to be dirty tracked, so the W bit will be
264 * restored only when a write is attempted to the page.
266 static const u64 shadow_acc_track_saved_bits_mask = PT64_EPT_READABLE_MASK |
267 PT64_EPT_EXECUTABLE_MASK;
268 static const u64 shadow_acc_track_saved_bits_shift = PT64_SECOND_AVAIL_BITS_SHIFT;
271 * This mask must be set on all non-zero Non-Present or Reserved SPTEs in order
272 * to guard against L1TF attacks.
274 static u64 __read_mostly shadow_nonpresent_or_rsvd_mask;
277 * The number of high-order 1 bits to use in the mask above.
279 static const u64 shadow_nonpresent_or_rsvd_mask_len = 5;
282 * In some cases, we need to preserve the GFN of a non-present or reserved
283 * SPTE when we usurp the upper five bits of the physical address space to
284 * defend against L1TF, e.g. for MMIO SPTEs. To preserve the GFN, we'll
285 * shift bits of the GFN that overlap with shadow_nonpresent_or_rsvd_mask
286 * left into the reserved bits, i.e. the GFN in the SPTE will be split into
287 * high and low parts. This mask covers the lower bits of the GFN.
289 static u64 __read_mostly shadow_nonpresent_or_rsvd_lower_gfn_mask;
292 * The number of non-reserved physical address bits irrespective of features
293 * that repurpose legal bits, e.g. MKTME.
295 static u8 __read_mostly shadow_phys_bits;
297 static void mmu_spte_set(u64 *sptep, u64 spte);
298 static bool is_executable_pte(u64 spte);
299 static union kvm_mmu_page_role
300 kvm_mmu_calc_root_page_role(struct kvm_vcpu *vcpu);
302 #define CREATE_TRACE_POINTS
303 #include "mmutrace.h"
306 static inline bool kvm_available_flush_tlb_with_range(void)
308 return kvm_x86_ops.tlb_remote_flush_with_range;
311 static void kvm_flush_remote_tlbs_with_range(struct kvm *kvm,
312 struct kvm_tlb_range *range)
316 if (range && kvm_x86_ops.tlb_remote_flush_with_range)
317 ret = kvm_x86_ops.tlb_remote_flush_with_range(kvm, range);
320 kvm_flush_remote_tlbs(kvm);
323 static void kvm_flush_remote_tlbs_with_address(struct kvm *kvm,
324 u64 start_gfn, u64 pages)
326 struct kvm_tlb_range range;
328 range.start_gfn = start_gfn;
331 kvm_flush_remote_tlbs_with_range(kvm, &range);
334 void kvm_mmu_set_mmio_spte_mask(u64 mmio_mask, u64 mmio_value, u64 access_mask)
336 BUG_ON((u64)(unsigned)access_mask != access_mask);
337 BUG_ON((mmio_mask & mmio_value) != mmio_value);
338 shadow_mmio_value = mmio_value | SPTE_MMIO_MASK;
339 shadow_mmio_mask = mmio_mask | SPTE_SPECIAL_MASK;
340 shadow_mmio_access_mask = access_mask;
342 EXPORT_SYMBOL_GPL(kvm_mmu_set_mmio_spte_mask);
344 static bool is_mmio_spte(u64 spte)
346 return (spte & shadow_mmio_mask) == shadow_mmio_value;
349 static inline bool sp_ad_disabled(struct kvm_mmu_page *sp)
351 return sp->role.ad_disabled;
354 static inline bool kvm_vcpu_ad_need_write_protect(struct kvm_vcpu *vcpu)
357 * When using the EPT page-modification log, the GPAs in the log
358 * would come from L2 rather than L1. Therefore, we need to rely
359 * on write protection to record dirty pages. This also bypasses
360 * PML, since writes now result in a vmexit.
362 return vcpu->arch.mmu == &vcpu->arch.guest_mmu;
365 static inline bool spte_ad_enabled(u64 spte)
367 MMU_WARN_ON(is_mmio_spte(spte));
368 return (spte & SPTE_SPECIAL_MASK) != SPTE_AD_DISABLED_MASK;
371 static inline bool spte_ad_need_write_protect(u64 spte)
373 MMU_WARN_ON(is_mmio_spte(spte));
374 return (spte & SPTE_SPECIAL_MASK) != SPTE_AD_ENABLED_MASK;
377 static bool is_nx_huge_page_enabled(void)
379 return READ_ONCE(nx_huge_pages);
382 static inline u64 spte_shadow_accessed_mask(u64 spte)
384 MMU_WARN_ON(is_mmio_spte(spte));
385 return spte_ad_enabled(spte) ? shadow_accessed_mask : 0;
388 static inline u64 spte_shadow_dirty_mask(u64 spte)
390 MMU_WARN_ON(is_mmio_spte(spte));
391 return spte_ad_enabled(spte) ? shadow_dirty_mask : 0;
394 static inline bool is_access_track_spte(u64 spte)
396 return !spte_ad_enabled(spte) && (spte & shadow_acc_track_mask) == 0;
400 * Due to limited space in PTEs, the MMIO generation is a 19 bit subset of
401 * the memslots generation and is derived as follows:
403 * Bits 0-8 of the MMIO generation are propagated to spte bits 3-11
404 * Bits 9-18 of the MMIO generation are propagated to spte bits 52-61
406 * The KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS flag is intentionally not included in
407 * the MMIO generation number, as doing so would require stealing a bit from
408 * the "real" generation number and thus effectively halve the maximum number
409 * of MMIO generations that can be handled before encountering a wrap (which
410 * requires a full MMU zap). The flag is instead explicitly queried when
411 * checking for MMIO spte cache hits.
413 #define MMIO_SPTE_GEN_MASK GENMASK_ULL(17, 0)
415 #define MMIO_SPTE_GEN_LOW_START 3
416 #define MMIO_SPTE_GEN_LOW_END 11
417 #define MMIO_SPTE_GEN_LOW_MASK GENMASK_ULL(MMIO_SPTE_GEN_LOW_END, \
418 MMIO_SPTE_GEN_LOW_START)
420 #define MMIO_SPTE_GEN_HIGH_START PT64_SECOND_AVAIL_BITS_SHIFT
421 #define MMIO_SPTE_GEN_HIGH_END 62
422 #define MMIO_SPTE_GEN_HIGH_MASK GENMASK_ULL(MMIO_SPTE_GEN_HIGH_END, \
423 MMIO_SPTE_GEN_HIGH_START)
425 static u64 generation_mmio_spte_mask(u64 gen)
429 WARN_ON(gen & ~MMIO_SPTE_GEN_MASK);
430 BUILD_BUG_ON((MMIO_SPTE_GEN_HIGH_MASK | MMIO_SPTE_GEN_LOW_MASK) & SPTE_SPECIAL_MASK);
432 mask = (gen << MMIO_SPTE_GEN_LOW_START) & MMIO_SPTE_GEN_LOW_MASK;
433 mask |= (gen << MMIO_SPTE_GEN_HIGH_START) & MMIO_SPTE_GEN_HIGH_MASK;
437 static u64 get_mmio_spte_generation(u64 spte)
441 gen = (spte & MMIO_SPTE_GEN_LOW_MASK) >> MMIO_SPTE_GEN_LOW_START;
442 gen |= (spte & MMIO_SPTE_GEN_HIGH_MASK) >> MMIO_SPTE_GEN_HIGH_START;
446 static u64 make_mmio_spte(struct kvm_vcpu *vcpu, u64 gfn, unsigned int access)
449 u64 gen = kvm_vcpu_memslots(vcpu)->generation & MMIO_SPTE_GEN_MASK;
450 u64 mask = generation_mmio_spte_mask(gen);
451 u64 gpa = gfn << PAGE_SHIFT;
453 access &= shadow_mmio_access_mask;
454 mask |= shadow_mmio_value | access;
455 mask |= gpa | shadow_nonpresent_or_rsvd_mask;
456 mask |= (gpa & shadow_nonpresent_or_rsvd_mask)
457 << shadow_nonpresent_or_rsvd_mask_len;
462 static void mark_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, u64 gfn,
465 u64 mask = make_mmio_spte(vcpu, gfn, access);
466 unsigned int gen = get_mmio_spte_generation(mask);
468 access = mask & ACC_ALL;
470 trace_mark_mmio_spte(sptep, gfn, access, gen);
471 mmu_spte_set(sptep, mask);
474 static gfn_t get_mmio_spte_gfn(u64 spte)
476 u64 gpa = spte & shadow_nonpresent_or_rsvd_lower_gfn_mask;
478 gpa |= (spte >> shadow_nonpresent_or_rsvd_mask_len)
479 & shadow_nonpresent_or_rsvd_mask;
481 return gpa >> PAGE_SHIFT;
484 static unsigned get_mmio_spte_access(u64 spte)
486 return spte & shadow_mmio_access_mask;
489 static bool set_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, gfn_t gfn,
490 kvm_pfn_t pfn, unsigned int access)
492 if (unlikely(is_noslot_pfn(pfn))) {
493 mark_mmio_spte(vcpu, sptep, gfn, access);
500 static bool check_mmio_spte(struct kvm_vcpu *vcpu, u64 spte)
502 u64 kvm_gen, spte_gen, gen;
504 gen = kvm_vcpu_memslots(vcpu)->generation;
505 if (unlikely(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS))
508 kvm_gen = gen & MMIO_SPTE_GEN_MASK;
509 spte_gen = get_mmio_spte_generation(spte);
511 trace_check_mmio_spte(spte, kvm_gen, spte_gen);
512 return likely(kvm_gen == spte_gen);
516 * Sets the shadow PTE masks used by the MMU.
519 * - Setting either @accessed_mask or @dirty_mask requires setting both
520 * - At least one of @accessed_mask or @acc_track_mask must be set
522 void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
523 u64 dirty_mask, u64 nx_mask, u64 x_mask, u64 p_mask,
524 u64 acc_track_mask, u64 me_mask)
526 BUG_ON(!dirty_mask != !accessed_mask);
527 BUG_ON(!accessed_mask && !acc_track_mask);
528 BUG_ON(acc_track_mask & SPTE_SPECIAL_MASK);
530 shadow_user_mask = user_mask;
531 shadow_accessed_mask = accessed_mask;
532 shadow_dirty_mask = dirty_mask;
533 shadow_nx_mask = nx_mask;
534 shadow_x_mask = x_mask;
535 shadow_present_mask = p_mask;
536 shadow_acc_track_mask = acc_track_mask;
537 shadow_me_mask = me_mask;
539 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes);
541 static u8 kvm_get_shadow_phys_bits(void)
544 * boot_cpu_data.x86_phys_bits is reduced when MKTME or SME are detected
545 * in CPU detection code, but the processor treats those reduced bits as
546 * 'keyID' thus they are not reserved bits. Therefore KVM needs to look at
547 * the physical address bits reported by CPUID.
549 if (likely(boot_cpu_data.extended_cpuid_level >= 0x80000008))
550 return cpuid_eax(0x80000008) & 0xff;
553 * Quite weird to have VMX or SVM but not MAXPHYADDR; probably a VM with
554 * custom CPUID. Proceed with whatever the kernel found since these features
555 * aren't virtualizable (SME/SEV also require CPUIDs higher than 0x80000008).
557 return boot_cpu_data.x86_phys_bits;
560 static void kvm_mmu_reset_all_pte_masks(void)
564 shadow_user_mask = 0;
565 shadow_accessed_mask = 0;
566 shadow_dirty_mask = 0;
569 shadow_mmio_mask = 0;
570 shadow_present_mask = 0;
571 shadow_acc_track_mask = 0;
573 shadow_phys_bits = kvm_get_shadow_phys_bits();
576 * If the CPU has 46 or less physical address bits, then set an
577 * appropriate mask to guard against L1TF attacks. Otherwise, it is
578 * assumed that the CPU is not vulnerable to L1TF.
580 * Some Intel CPUs address the L1 cache using more PA bits than are
581 * reported by CPUID. Use the PA width of the L1 cache when possible
582 * to achieve more effective mitigation, e.g. if system RAM overlaps
583 * the most significant bits of legal physical address space.
585 shadow_nonpresent_or_rsvd_mask = 0;
586 low_phys_bits = boot_cpu_data.x86_cache_bits;
587 if (boot_cpu_data.x86_cache_bits <
588 52 - shadow_nonpresent_or_rsvd_mask_len) {
589 shadow_nonpresent_or_rsvd_mask =
590 rsvd_bits(boot_cpu_data.x86_cache_bits -
591 shadow_nonpresent_or_rsvd_mask_len,
592 boot_cpu_data.x86_cache_bits - 1);
593 low_phys_bits -= shadow_nonpresent_or_rsvd_mask_len;
595 WARN_ON_ONCE(boot_cpu_has_bug(X86_BUG_L1TF));
597 shadow_nonpresent_or_rsvd_lower_gfn_mask =
598 GENMASK_ULL(low_phys_bits - 1, PAGE_SHIFT);
601 static int is_cpuid_PSE36(void)
606 static int is_nx(struct kvm_vcpu *vcpu)
608 return vcpu->arch.efer & EFER_NX;
611 static int is_shadow_present_pte(u64 pte)
613 return (pte != 0) && !is_mmio_spte(pte);
616 static int is_large_pte(u64 pte)
618 return pte & PT_PAGE_SIZE_MASK;
621 static int is_last_spte(u64 pte, int level)
623 if (level == PT_PAGE_TABLE_LEVEL)
625 if (is_large_pte(pte))
630 static bool is_executable_pte(u64 spte)
632 return (spte & (shadow_x_mask | shadow_nx_mask)) == shadow_x_mask;
635 static kvm_pfn_t spte_to_pfn(u64 pte)
637 return (pte & PT64_BASE_ADDR_MASK) >> PAGE_SHIFT;
640 static gfn_t pse36_gfn_delta(u32 gpte)
642 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
644 return (gpte & PT32_DIR_PSE36_MASK) << shift;
648 static void __set_spte(u64 *sptep, u64 spte)
650 WRITE_ONCE(*sptep, spte);
653 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
655 WRITE_ONCE(*sptep, spte);
658 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
660 return xchg(sptep, spte);
663 static u64 __get_spte_lockless(u64 *sptep)
665 return READ_ONCE(*sptep);
676 static void count_spte_clear(u64 *sptep, u64 spte)
678 struct kvm_mmu_page *sp = page_header(__pa(sptep));
680 if (is_shadow_present_pte(spte))
683 /* Ensure the spte is completely set before we increase the count */
685 sp->clear_spte_count++;
688 static void __set_spte(u64 *sptep, u64 spte)
690 union split_spte *ssptep, sspte;
692 ssptep = (union split_spte *)sptep;
693 sspte = (union split_spte)spte;
695 ssptep->spte_high = sspte.spte_high;
698 * If we map the spte from nonpresent to present, We should store
699 * the high bits firstly, then set present bit, so cpu can not
700 * fetch this spte while we are setting the spte.
704 WRITE_ONCE(ssptep->spte_low, sspte.spte_low);
707 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
709 union split_spte *ssptep, sspte;
711 ssptep = (union split_spte *)sptep;
712 sspte = (union split_spte)spte;
714 WRITE_ONCE(ssptep->spte_low, sspte.spte_low);
717 * If we map the spte from present to nonpresent, we should clear
718 * present bit firstly to avoid vcpu fetch the old high bits.
722 ssptep->spte_high = sspte.spte_high;
723 count_spte_clear(sptep, spte);
726 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
728 union split_spte *ssptep, sspte, orig;
730 ssptep = (union split_spte *)sptep;
731 sspte = (union split_spte)spte;
733 /* xchg acts as a barrier before the setting of the high bits */
734 orig.spte_low = xchg(&ssptep->spte_low, sspte.spte_low);
735 orig.spte_high = ssptep->spte_high;
736 ssptep->spte_high = sspte.spte_high;
737 count_spte_clear(sptep, spte);
743 * The idea using the light way get the spte on x86_32 guest is from
744 * gup_get_pte (mm/gup.c).
746 * An spte tlb flush may be pending, because kvm_set_pte_rmapp
747 * coalesces them and we are running out of the MMU lock. Therefore
748 * we need to protect against in-progress updates of the spte.
750 * Reading the spte while an update is in progress may get the old value
751 * for the high part of the spte. The race is fine for a present->non-present
752 * change (because the high part of the spte is ignored for non-present spte),
753 * but for a present->present change we must reread the spte.
755 * All such changes are done in two steps (present->non-present and
756 * non-present->present), hence it is enough to count the number of
757 * present->non-present updates: if it changed while reading the spte,
758 * we might have hit the race. This is done using clear_spte_count.
760 static u64 __get_spte_lockless(u64 *sptep)
762 struct kvm_mmu_page *sp = page_header(__pa(sptep));
763 union split_spte spte, *orig = (union split_spte *)sptep;
767 count = sp->clear_spte_count;
770 spte.spte_low = orig->spte_low;
773 spte.spte_high = orig->spte_high;
776 if (unlikely(spte.spte_low != orig->spte_low ||
777 count != sp->clear_spte_count))
784 static bool spte_can_locklessly_be_made_writable(u64 spte)
786 return (spte & (SPTE_HOST_WRITEABLE | SPTE_MMU_WRITEABLE)) ==
787 (SPTE_HOST_WRITEABLE | SPTE_MMU_WRITEABLE);
790 static bool spte_has_volatile_bits(u64 spte)
792 if (!is_shadow_present_pte(spte))
796 * Always atomically update spte if it can be updated
797 * out of mmu-lock, it can ensure dirty bit is not lost,
798 * also, it can help us to get a stable is_writable_pte()
799 * to ensure tlb flush is not missed.
801 if (spte_can_locklessly_be_made_writable(spte) ||
802 is_access_track_spte(spte))
805 if (spte_ad_enabled(spte)) {
806 if ((spte & shadow_accessed_mask) == 0 ||
807 (is_writable_pte(spte) && (spte & shadow_dirty_mask) == 0))
814 static bool is_accessed_spte(u64 spte)
816 u64 accessed_mask = spte_shadow_accessed_mask(spte);
818 return accessed_mask ? spte & accessed_mask
819 : !is_access_track_spte(spte);
822 static bool is_dirty_spte(u64 spte)
824 u64 dirty_mask = spte_shadow_dirty_mask(spte);
826 return dirty_mask ? spte & dirty_mask : spte & PT_WRITABLE_MASK;
829 /* Rules for using mmu_spte_set:
830 * Set the sptep from nonpresent to present.
831 * Note: the sptep being assigned *must* be either not present
832 * or in a state where the hardware will not attempt to update
835 static void mmu_spte_set(u64 *sptep, u64 new_spte)
837 WARN_ON(is_shadow_present_pte(*sptep));
838 __set_spte(sptep, new_spte);
842 * Update the SPTE (excluding the PFN), but do not track changes in its
843 * accessed/dirty status.
845 static u64 mmu_spte_update_no_track(u64 *sptep, u64 new_spte)
847 u64 old_spte = *sptep;
849 WARN_ON(!is_shadow_present_pte(new_spte));
851 if (!is_shadow_present_pte(old_spte)) {
852 mmu_spte_set(sptep, new_spte);
856 if (!spte_has_volatile_bits(old_spte))
857 __update_clear_spte_fast(sptep, new_spte);
859 old_spte = __update_clear_spte_slow(sptep, new_spte);
861 WARN_ON(spte_to_pfn(old_spte) != spte_to_pfn(new_spte));
866 /* Rules for using mmu_spte_update:
867 * Update the state bits, it means the mapped pfn is not changed.
869 * Whenever we overwrite a writable spte with a read-only one we
870 * should flush remote TLBs. Otherwise rmap_write_protect
871 * will find a read-only spte, even though the writable spte
872 * might be cached on a CPU's TLB, the return value indicates this
875 * Returns true if the TLB needs to be flushed
877 static bool mmu_spte_update(u64 *sptep, u64 new_spte)
880 u64 old_spte = mmu_spte_update_no_track(sptep, new_spte);
882 if (!is_shadow_present_pte(old_spte))
886 * For the spte updated out of mmu-lock is safe, since
887 * we always atomically update it, see the comments in
888 * spte_has_volatile_bits().
890 if (spte_can_locklessly_be_made_writable(old_spte) &&
891 !is_writable_pte(new_spte))
895 * Flush TLB when accessed/dirty states are changed in the page tables,
896 * to guarantee consistency between TLB and page tables.
899 if (is_accessed_spte(old_spte) && !is_accessed_spte(new_spte)) {
901 kvm_set_pfn_accessed(spte_to_pfn(old_spte));
904 if (is_dirty_spte(old_spte) && !is_dirty_spte(new_spte)) {
906 kvm_set_pfn_dirty(spte_to_pfn(old_spte));
913 * Rules for using mmu_spte_clear_track_bits:
914 * It sets the sptep from present to nonpresent, and track the
915 * state bits, it is used to clear the last level sptep.
916 * Returns non-zero if the PTE was previously valid.
918 static int mmu_spte_clear_track_bits(u64 *sptep)
921 u64 old_spte = *sptep;
923 if (!spte_has_volatile_bits(old_spte))
924 __update_clear_spte_fast(sptep, 0ull);
926 old_spte = __update_clear_spte_slow(sptep, 0ull);
928 if (!is_shadow_present_pte(old_spte))
931 pfn = spte_to_pfn(old_spte);
934 * KVM does not hold the refcount of the page used by
935 * kvm mmu, before reclaiming the page, we should
936 * unmap it from mmu first.
938 WARN_ON(!kvm_is_reserved_pfn(pfn) && !page_count(pfn_to_page(pfn)));
940 if (is_accessed_spte(old_spte))
941 kvm_set_pfn_accessed(pfn);
943 if (is_dirty_spte(old_spte))
944 kvm_set_pfn_dirty(pfn);
950 * Rules for using mmu_spte_clear_no_track:
951 * Directly clear spte without caring the state bits of sptep,
952 * it is used to set the upper level spte.
954 static void mmu_spte_clear_no_track(u64 *sptep)
956 __update_clear_spte_fast(sptep, 0ull);
959 static u64 mmu_spte_get_lockless(u64 *sptep)
961 return __get_spte_lockless(sptep);
964 static u64 mark_spte_for_access_track(u64 spte)
966 if (spte_ad_enabled(spte))
967 return spte & ~shadow_accessed_mask;
969 if (is_access_track_spte(spte))
973 * Making an Access Tracking PTE will result in removal of write access
974 * from the PTE. So, verify that we will be able to restore the write
975 * access in the fast page fault path later on.
977 WARN_ONCE((spte & PT_WRITABLE_MASK) &&
978 !spte_can_locklessly_be_made_writable(spte),
979 "kvm: Writable SPTE is not locklessly dirty-trackable\n");
981 WARN_ONCE(spte & (shadow_acc_track_saved_bits_mask <<
982 shadow_acc_track_saved_bits_shift),
983 "kvm: Access Tracking saved bit locations are not zero\n");
985 spte |= (spte & shadow_acc_track_saved_bits_mask) <<
986 shadow_acc_track_saved_bits_shift;
987 spte &= ~shadow_acc_track_mask;
992 /* Restore an acc-track PTE back to a regular PTE */
993 static u64 restore_acc_track_spte(u64 spte)
996 u64 saved_bits = (spte >> shadow_acc_track_saved_bits_shift)
997 & shadow_acc_track_saved_bits_mask;
999 WARN_ON_ONCE(spte_ad_enabled(spte));
1000 WARN_ON_ONCE(!is_access_track_spte(spte));
1002 new_spte &= ~shadow_acc_track_mask;
1003 new_spte &= ~(shadow_acc_track_saved_bits_mask <<
1004 shadow_acc_track_saved_bits_shift);
1005 new_spte |= saved_bits;
1010 /* Returns the Accessed status of the PTE and resets it at the same time. */
1011 static bool mmu_spte_age(u64 *sptep)
1013 u64 spte = mmu_spte_get_lockless(sptep);
1015 if (!is_accessed_spte(spte))
1018 if (spte_ad_enabled(spte)) {
1019 clear_bit((ffs(shadow_accessed_mask) - 1),
1020 (unsigned long *)sptep);
1023 * Capture the dirty status of the page, so that it doesn't get
1024 * lost when the SPTE is marked for access tracking.
1026 if (is_writable_pte(spte))
1027 kvm_set_pfn_dirty(spte_to_pfn(spte));
1029 spte = mark_spte_for_access_track(spte);
1030 mmu_spte_update_no_track(sptep, spte);
1036 static void walk_shadow_page_lockless_begin(struct kvm_vcpu *vcpu)
1039 * Prevent page table teardown by making any free-er wait during
1040 * kvm_flush_remote_tlbs() IPI to all active vcpus.
1042 local_irq_disable();
1045 * Make sure a following spte read is not reordered ahead of the write
1048 smp_store_mb(vcpu->mode, READING_SHADOW_PAGE_TABLES);
1051 static void walk_shadow_page_lockless_end(struct kvm_vcpu *vcpu)
1054 * Make sure the write to vcpu->mode is not reordered in front of
1055 * reads to sptes. If it does, kvm_mmu_commit_zap_page() can see us
1056 * OUTSIDE_GUEST_MODE and proceed to free the shadow page table.
1058 smp_store_release(&vcpu->mode, OUTSIDE_GUEST_MODE);
1062 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
1063 struct kmem_cache *base_cache, int min)
1067 if (cache->nobjs >= min)
1069 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
1070 obj = kmem_cache_zalloc(base_cache, GFP_KERNEL_ACCOUNT);
1072 return cache->nobjs >= min ? 0 : -ENOMEM;
1073 cache->objects[cache->nobjs++] = obj;
1078 static int mmu_memory_cache_free_objects(struct kvm_mmu_memory_cache *cache)
1080 return cache->nobjs;
1083 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc,
1084 struct kmem_cache *cache)
1087 kmem_cache_free(cache, mc->objects[--mc->nobjs]);
1090 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache *cache,
1095 if (cache->nobjs >= min)
1097 while (cache->nobjs < ARRAY_SIZE(cache->objects)) {
1098 page = (void *)__get_free_page(GFP_KERNEL_ACCOUNT);
1100 return cache->nobjs >= min ? 0 : -ENOMEM;
1101 cache->objects[cache->nobjs++] = page;
1106 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache *mc)
1109 free_page((unsigned long)mc->objects[--mc->nobjs]);
1112 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
1116 r = mmu_topup_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache,
1117 pte_list_desc_cache, 8 + PTE_PREFETCH_NUM);
1120 r = mmu_topup_memory_cache_page(&vcpu->arch.mmu_page_cache, 8);
1123 r = mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
1124 mmu_page_header_cache, 4);
1129 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1131 mmu_free_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache,
1132 pte_list_desc_cache);
1133 mmu_free_memory_cache_page(&vcpu->arch.mmu_page_cache);
1134 mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache,
1135 mmu_page_header_cache);
1138 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
1143 p = mc->objects[--mc->nobjs];
1147 static struct pte_list_desc *mmu_alloc_pte_list_desc(struct kvm_vcpu *vcpu)
1149 return mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_list_desc_cache);
1152 static void mmu_free_pte_list_desc(struct pte_list_desc *pte_list_desc)
1154 kmem_cache_free(pte_list_desc_cache, pte_list_desc);
1157 static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index)
1159 if (!sp->role.direct)
1160 return sp->gfns[index];
1162 return sp->gfn + (index << ((sp->role.level - 1) * PT64_LEVEL_BITS));
1165 static void kvm_mmu_page_set_gfn(struct kvm_mmu_page *sp, int index, gfn_t gfn)
1167 if (!sp->role.direct) {
1168 sp->gfns[index] = gfn;
1172 if (WARN_ON(gfn != kvm_mmu_page_get_gfn(sp, index)))
1173 pr_err_ratelimited("gfn mismatch under direct page %llx "
1174 "(expected %llx, got %llx)\n",
1176 kvm_mmu_page_get_gfn(sp, index), gfn);
1180 * Return the pointer to the large page information for a given gfn,
1181 * handling slots that are not large page aligned.
1183 static struct kvm_lpage_info *lpage_info_slot(gfn_t gfn,
1184 struct kvm_memory_slot *slot,
1189 idx = gfn_to_index(gfn, slot->base_gfn, level);
1190 return &slot->arch.lpage_info[level - 2][idx];
1193 static void update_gfn_disallow_lpage_count(struct kvm_memory_slot *slot,
1194 gfn_t gfn, int count)
1196 struct kvm_lpage_info *linfo;
1199 for (i = PT_DIRECTORY_LEVEL; i <= PT_MAX_HUGEPAGE_LEVEL; ++i) {
1200 linfo = lpage_info_slot(gfn, slot, i);
1201 linfo->disallow_lpage += count;
1202 WARN_ON(linfo->disallow_lpage < 0);
1206 void kvm_mmu_gfn_disallow_lpage(struct kvm_memory_slot *slot, gfn_t gfn)
1208 update_gfn_disallow_lpage_count(slot, gfn, 1);
1211 void kvm_mmu_gfn_allow_lpage(struct kvm_memory_slot *slot, gfn_t gfn)
1213 update_gfn_disallow_lpage_count(slot, gfn, -1);
1216 static void account_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp)
1218 struct kvm_memslots *slots;
1219 struct kvm_memory_slot *slot;
1222 kvm->arch.indirect_shadow_pages++;
1224 slots = kvm_memslots_for_spte_role(kvm, sp->role);
1225 slot = __gfn_to_memslot(slots, gfn);
1227 /* the non-leaf shadow pages are keeping readonly. */
1228 if (sp->role.level > PT_PAGE_TABLE_LEVEL)
1229 return kvm_slot_page_track_add_page(kvm, slot, gfn,
1230 KVM_PAGE_TRACK_WRITE);
1232 kvm_mmu_gfn_disallow_lpage(slot, gfn);
1235 static void account_huge_nx_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1237 if (sp->lpage_disallowed)
1240 ++kvm->stat.nx_lpage_splits;
1241 list_add_tail(&sp->lpage_disallowed_link,
1242 &kvm->arch.lpage_disallowed_mmu_pages);
1243 sp->lpage_disallowed = true;
1246 static void unaccount_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp)
1248 struct kvm_memslots *slots;
1249 struct kvm_memory_slot *slot;
1252 kvm->arch.indirect_shadow_pages--;
1254 slots = kvm_memslots_for_spte_role(kvm, sp->role);
1255 slot = __gfn_to_memslot(slots, gfn);
1256 if (sp->role.level > PT_PAGE_TABLE_LEVEL)
1257 return kvm_slot_page_track_remove_page(kvm, slot, gfn,
1258 KVM_PAGE_TRACK_WRITE);
1260 kvm_mmu_gfn_allow_lpage(slot, gfn);
1263 static void unaccount_huge_nx_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1265 --kvm->stat.nx_lpage_splits;
1266 sp->lpage_disallowed = false;
1267 list_del(&sp->lpage_disallowed_link);
1270 static struct kvm_memory_slot *
1271 gfn_to_memslot_dirty_bitmap(struct kvm_vcpu *vcpu, gfn_t gfn,
1274 struct kvm_memory_slot *slot;
1276 slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1277 if (!slot || slot->flags & KVM_MEMSLOT_INVALID)
1279 if (no_dirty_log && slot->dirty_bitmap)
1286 * About rmap_head encoding:
1288 * If the bit zero of rmap_head->val is clear, then it points to the only spte
1289 * in this rmap chain. Otherwise, (rmap_head->val & ~1) points to a struct
1290 * pte_list_desc containing more mappings.
1294 * Returns the number of pointers in the rmap chain, not counting the new one.
1296 static int pte_list_add(struct kvm_vcpu *vcpu, u64 *spte,
1297 struct kvm_rmap_head *rmap_head)
1299 struct pte_list_desc *desc;
1302 if (!rmap_head->val) {
1303 rmap_printk("pte_list_add: %p %llx 0->1\n", spte, *spte);
1304 rmap_head->val = (unsigned long)spte;
1305 } else if (!(rmap_head->val & 1)) {
1306 rmap_printk("pte_list_add: %p %llx 1->many\n", spte, *spte);
1307 desc = mmu_alloc_pte_list_desc(vcpu);
1308 desc->sptes[0] = (u64 *)rmap_head->val;
1309 desc->sptes[1] = spte;
1310 rmap_head->val = (unsigned long)desc | 1;
1313 rmap_printk("pte_list_add: %p %llx many->many\n", spte, *spte);
1314 desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
1315 while (desc->sptes[PTE_LIST_EXT-1] && desc->more) {
1317 count += PTE_LIST_EXT;
1319 if (desc->sptes[PTE_LIST_EXT-1]) {
1320 desc->more = mmu_alloc_pte_list_desc(vcpu);
1323 for (i = 0; desc->sptes[i]; ++i)
1325 desc->sptes[i] = spte;
1331 pte_list_desc_remove_entry(struct kvm_rmap_head *rmap_head,
1332 struct pte_list_desc *desc, int i,
1333 struct pte_list_desc *prev_desc)
1337 for (j = PTE_LIST_EXT - 1; !desc->sptes[j] && j > i; --j)
1339 desc->sptes[i] = desc->sptes[j];
1340 desc->sptes[j] = NULL;
1343 if (!prev_desc && !desc->more)
1347 prev_desc->more = desc->more;
1349 rmap_head->val = (unsigned long)desc->more | 1;
1350 mmu_free_pte_list_desc(desc);
1353 static void __pte_list_remove(u64 *spte, struct kvm_rmap_head *rmap_head)
1355 struct pte_list_desc *desc;
1356 struct pte_list_desc *prev_desc;
1359 if (!rmap_head->val) {
1360 pr_err("%s: %p 0->BUG\n", __func__, spte);
1362 } else if (!(rmap_head->val & 1)) {
1363 rmap_printk("%s: %p 1->0\n", __func__, spte);
1364 if ((u64 *)rmap_head->val != spte) {
1365 pr_err("%s: %p 1->BUG\n", __func__, spte);
1370 rmap_printk("%s: %p many->many\n", __func__, spte);
1371 desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
1374 for (i = 0; i < PTE_LIST_EXT && desc->sptes[i]; ++i) {
1375 if (desc->sptes[i] == spte) {
1376 pte_list_desc_remove_entry(rmap_head,
1377 desc, i, prev_desc);
1384 pr_err("%s: %p many->many\n", __func__, spte);
1389 static void pte_list_remove(struct kvm_rmap_head *rmap_head, u64 *sptep)
1391 mmu_spte_clear_track_bits(sptep);
1392 __pte_list_remove(sptep, rmap_head);
1395 static struct kvm_rmap_head *__gfn_to_rmap(gfn_t gfn, int level,
1396 struct kvm_memory_slot *slot)
1400 idx = gfn_to_index(gfn, slot->base_gfn, level);
1401 return &slot->arch.rmap[level - PT_PAGE_TABLE_LEVEL][idx];
1404 static struct kvm_rmap_head *gfn_to_rmap(struct kvm *kvm, gfn_t gfn,
1405 struct kvm_mmu_page *sp)
1407 struct kvm_memslots *slots;
1408 struct kvm_memory_slot *slot;
1410 slots = kvm_memslots_for_spte_role(kvm, sp->role);
1411 slot = __gfn_to_memslot(slots, gfn);
1412 return __gfn_to_rmap(gfn, sp->role.level, slot);
1415 static bool rmap_can_add(struct kvm_vcpu *vcpu)
1417 struct kvm_mmu_memory_cache *cache;
1419 cache = &vcpu->arch.mmu_pte_list_desc_cache;
1420 return mmu_memory_cache_free_objects(cache);
1423 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
1425 struct kvm_mmu_page *sp;
1426 struct kvm_rmap_head *rmap_head;
1428 sp = page_header(__pa(spte));
1429 kvm_mmu_page_set_gfn(sp, spte - sp->spt, gfn);
1430 rmap_head = gfn_to_rmap(vcpu->kvm, gfn, sp);
1431 return pte_list_add(vcpu, spte, rmap_head);
1434 static void rmap_remove(struct kvm *kvm, u64 *spte)
1436 struct kvm_mmu_page *sp;
1438 struct kvm_rmap_head *rmap_head;
1440 sp = page_header(__pa(spte));
1441 gfn = kvm_mmu_page_get_gfn(sp, spte - sp->spt);
1442 rmap_head = gfn_to_rmap(kvm, gfn, sp);
1443 __pte_list_remove(spte, rmap_head);
1447 * Used by the following functions to iterate through the sptes linked by a
1448 * rmap. All fields are private and not assumed to be used outside.
1450 struct rmap_iterator {
1451 /* private fields */
1452 struct pte_list_desc *desc; /* holds the sptep if not NULL */
1453 int pos; /* index of the sptep */
1457 * Iteration must be started by this function. This should also be used after
1458 * removing/dropping sptes from the rmap link because in such cases the
1459 * information in the iterator may not be valid.
1461 * Returns sptep if found, NULL otherwise.
1463 static u64 *rmap_get_first(struct kvm_rmap_head *rmap_head,
1464 struct rmap_iterator *iter)
1468 if (!rmap_head->val)
1471 if (!(rmap_head->val & 1)) {
1473 sptep = (u64 *)rmap_head->val;
1477 iter->desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
1479 sptep = iter->desc->sptes[iter->pos];
1481 BUG_ON(!is_shadow_present_pte(*sptep));
1486 * Must be used with a valid iterator: e.g. after rmap_get_first().
1488 * Returns sptep if found, NULL otherwise.
1490 static u64 *rmap_get_next(struct rmap_iterator *iter)
1495 if (iter->pos < PTE_LIST_EXT - 1) {
1497 sptep = iter->desc->sptes[iter->pos];
1502 iter->desc = iter->desc->more;
1506 /* desc->sptes[0] cannot be NULL */
1507 sptep = iter->desc->sptes[iter->pos];
1514 BUG_ON(!is_shadow_present_pte(*sptep));
1518 #define for_each_rmap_spte(_rmap_head_, _iter_, _spte_) \
1519 for (_spte_ = rmap_get_first(_rmap_head_, _iter_); \
1520 _spte_; _spte_ = rmap_get_next(_iter_))
1522 static void drop_spte(struct kvm *kvm, u64 *sptep)
1524 if (mmu_spte_clear_track_bits(sptep))
1525 rmap_remove(kvm, sptep);
1529 static bool __drop_large_spte(struct kvm *kvm, u64 *sptep)
1531 if (is_large_pte(*sptep)) {
1532 WARN_ON(page_header(__pa(sptep))->role.level ==
1533 PT_PAGE_TABLE_LEVEL);
1534 drop_spte(kvm, sptep);
1542 static void drop_large_spte(struct kvm_vcpu *vcpu, u64 *sptep)
1544 if (__drop_large_spte(vcpu->kvm, sptep)) {
1545 struct kvm_mmu_page *sp = page_header(__pa(sptep));
1547 kvm_flush_remote_tlbs_with_address(vcpu->kvm, sp->gfn,
1548 KVM_PAGES_PER_HPAGE(sp->role.level));
1553 * Write-protect on the specified @sptep, @pt_protect indicates whether
1554 * spte write-protection is caused by protecting shadow page table.
1556 * Note: write protection is difference between dirty logging and spte
1558 * - for dirty logging, the spte can be set to writable at anytime if
1559 * its dirty bitmap is properly set.
1560 * - for spte protection, the spte can be writable only after unsync-ing
1563 * Return true if tlb need be flushed.
1565 static bool spte_write_protect(u64 *sptep, bool pt_protect)
1569 if (!is_writable_pte(spte) &&
1570 !(pt_protect && spte_can_locklessly_be_made_writable(spte)))
1573 rmap_printk("rmap_write_protect: spte %p %llx\n", sptep, *sptep);
1576 spte &= ~SPTE_MMU_WRITEABLE;
1577 spte = spte & ~PT_WRITABLE_MASK;
1579 return mmu_spte_update(sptep, spte);
1582 static bool __rmap_write_protect(struct kvm *kvm,
1583 struct kvm_rmap_head *rmap_head,
1587 struct rmap_iterator iter;
1590 for_each_rmap_spte(rmap_head, &iter, sptep)
1591 flush |= spte_write_protect(sptep, pt_protect);
1596 static bool spte_clear_dirty(u64 *sptep)
1600 rmap_printk("rmap_clear_dirty: spte %p %llx\n", sptep, *sptep);
1602 MMU_WARN_ON(!spte_ad_enabled(spte));
1603 spte &= ~shadow_dirty_mask;
1604 return mmu_spte_update(sptep, spte);
1607 static bool spte_wrprot_for_clear_dirty(u64 *sptep)
1609 bool was_writable = test_and_clear_bit(PT_WRITABLE_SHIFT,
1610 (unsigned long *)sptep);
1611 if (was_writable && !spte_ad_enabled(*sptep))
1612 kvm_set_pfn_dirty(spte_to_pfn(*sptep));
1614 return was_writable;
1618 * Gets the GFN ready for another round of dirty logging by clearing the
1619 * - D bit on ad-enabled SPTEs, and
1620 * - W bit on ad-disabled SPTEs.
1621 * Returns true iff any D or W bits were cleared.
1623 static bool __rmap_clear_dirty(struct kvm *kvm, struct kvm_rmap_head *rmap_head)
1626 struct rmap_iterator iter;
1629 for_each_rmap_spte(rmap_head, &iter, sptep)
1630 if (spte_ad_need_write_protect(*sptep))
1631 flush |= spte_wrprot_for_clear_dirty(sptep);
1633 flush |= spte_clear_dirty(sptep);
1638 static bool spte_set_dirty(u64 *sptep)
1642 rmap_printk("rmap_set_dirty: spte %p %llx\n", sptep, *sptep);
1645 * Similar to the !kvm_x86_ops.slot_disable_log_dirty case,
1646 * do not bother adding back write access to pages marked
1647 * SPTE_AD_WRPROT_ONLY_MASK.
1649 spte |= shadow_dirty_mask;
1651 return mmu_spte_update(sptep, spte);
1654 static bool __rmap_set_dirty(struct kvm *kvm, struct kvm_rmap_head *rmap_head)
1657 struct rmap_iterator iter;
1660 for_each_rmap_spte(rmap_head, &iter, sptep)
1661 if (spte_ad_enabled(*sptep))
1662 flush |= spte_set_dirty(sptep);
1668 * kvm_mmu_write_protect_pt_masked - write protect selected PT level pages
1669 * @kvm: kvm instance
1670 * @slot: slot to protect
1671 * @gfn_offset: start of the BITS_PER_LONG pages we care about
1672 * @mask: indicates which pages we should protect
1674 * Used when we do not need to care about huge page mappings: e.g. during dirty
1675 * logging we do not have any such mappings.
1677 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1678 struct kvm_memory_slot *slot,
1679 gfn_t gfn_offset, unsigned long mask)
1681 struct kvm_rmap_head *rmap_head;
1684 rmap_head = __gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask),
1685 PT_PAGE_TABLE_LEVEL, slot);
1686 __rmap_write_protect(kvm, rmap_head, false);
1688 /* clear the first set bit */
1694 * kvm_mmu_clear_dirty_pt_masked - clear MMU D-bit for PT level pages, or write
1695 * protect the page if the D-bit isn't supported.
1696 * @kvm: kvm instance
1697 * @slot: slot to clear D-bit
1698 * @gfn_offset: start of the BITS_PER_LONG pages we care about
1699 * @mask: indicates which pages we should clear D-bit
1701 * Used for PML to re-log the dirty GPAs after userspace querying dirty_bitmap.
1703 void kvm_mmu_clear_dirty_pt_masked(struct kvm *kvm,
1704 struct kvm_memory_slot *slot,
1705 gfn_t gfn_offset, unsigned long mask)
1707 struct kvm_rmap_head *rmap_head;
1710 rmap_head = __gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask),
1711 PT_PAGE_TABLE_LEVEL, slot);
1712 __rmap_clear_dirty(kvm, rmap_head);
1714 /* clear the first set bit */
1718 EXPORT_SYMBOL_GPL(kvm_mmu_clear_dirty_pt_masked);
1721 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1724 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1725 * enable dirty logging for them.
1727 * Used when we do not need to care about huge page mappings: e.g. during dirty
1728 * logging we do not have any such mappings.
1730 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1731 struct kvm_memory_slot *slot,
1732 gfn_t gfn_offset, unsigned long mask)
1734 if (kvm_x86_ops.enable_log_dirty_pt_masked)
1735 kvm_x86_ops.enable_log_dirty_pt_masked(kvm, slot, gfn_offset,
1738 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1742 * kvm_arch_write_log_dirty - emulate dirty page logging
1743 * @vcpu: Guest mode vcpu
1745 * Emulate arch specific page modification logging for the
1748 int kvm_arch_write_log_dirty(struct kvm_vcpu *vcpu)
1750 if (kvm_x86_ops.write_log_dirty)
1751 return kvm_x86_ops.write_log_dirty(vcpu);
1756 bool kvm_mmu_slot_gfn_write_protect(struct kvm *kvm,
1757 struct kvm_memory_slot *slot, u64 gfn)
1759 struct kvm_rmap_head *rmap_head;
1761 bool write_protected = false;
1763 for (i = PT_PAGE_TABLE_LEVEL; i <= PT_MAX_HUGEPAGE_LEVEL; ++i) {
1764 rmap_head = __gfn_to_rmap(gfn, i, slot);
1765 write_protected |= __rmap_write_protect(kvm, rmap_head, true);
1768 return write_protected;
1771 static bool rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
1773 struct kvm_memory_slot *slot;
1775 slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1776 return kvm_mmu_slot_gfn_write_protect(vcpu->kvm, slot, gfn);
1779 static bool kvm_zap_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head)
1782 struct rmap_iterator iter;
1785 while ((sptep = rmap_get_first(rmap_head, &iter))) {
1786 rmap_printk("%s: spte %p %llx.\n", __func__, sptep, *sptep);
1788 pte_list_remove(rmap_head, sptep);
1795 static int kvm_unmap_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1796 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1799 return kvm_zap_rmapp(kvm, rmap_head);
1802 static int kvm_set_pte_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1803 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1807 struct rmap_iterator iter;
1810 pte_t *ptep = (pte_t *)data;
1813 WARN_ON(pte_huge(*ptep));
1814 new_pfn = pte_pfn(*ptep);
1817 for_each_rmap_spte(rmap_head, &iter, sptep) {
1818 rmap_printk("kvm_set_pte_rmapp: spte %p %llx gfn %llx (%d)\n",
1819 sptep, *sptep, gfn, level);
1823 if (pte_write(*ptep)) {
1824 pte_list_remove(rmap_head, sptep);
1827 new_spte = *sptep & ~PT64_BASE_ADDR_MASK;
1828 new_spte |= (u64)new_pfn << PAGE_SHIFT;
1830 new_spte &= ~PT_WRITABLE_MASK;
1831 new_spte &= ~SPTE_HOST_WRITEABLE;
1833 new_spte = mark_spte_for_access_track(new_spte);
1835 mmu_spte_clear_track_bits(sptep);
1836 mmu_spte_set(sptep, new_spte);
1840 if (need_flush && kvm_available_flush_tlb_with_range()) {
1841 kvm_flush_remote_tlbs_with_address(kvm, gfn, 1);
1848 struct slot_rmap_walk_iterator {
1850 struct kvm_memory_slot *slot;
1856 /* output fields. */
1858 struct kvm_rmap_head *rmap;
1861 /* private field. */
1862 struct kvm_rmap_head *end_rmap;
1866 rmap_walk_init_level(struct slot_rmap_walk_iterator *iterator, int level)
1868 iterator->level = level;
1869 iterator->gfn = iterator->start_gfn;
1870 iterator->rmap = __gfn_to_rmap(iterator->gfn, level, iterator->slot);
1871 iterator->end_rmap = __gfn_to_rmap(iterator->end_gfn, level,
1876 slot_rmap_walk_init(struct slot_rmap_walk_iterator *iterator,
1877 struct kvm_memory_slot *slot, int start_level,
1878 int end_level, gfn_t start_gfn, gfn_t end_gfn)
1880 iterator->slot = slot;
1881 iterator->start_level = start_level;
1882 iterator->end_level = end_level;
1883 iterator->start_gfn = start_gfn;
1884 iterator->end_gfn = end_gfn;
1886 rmap_walk_init_level(iterator, iterator->start_level);
1889 static bool slot_rmap_walk_okay(struct slot_rmap_walk_iterator *iterator)
1891 return !!iterator->rmap;
1894 static void slot_rmap_walk_next(struct slot_rmap_walk_iterator *iterator)
1896 if (++iterator->rmap <= iterator->end_rmap) {
1897 iterator->gfn += (1UL << KVM_HPAGE_GFN_SHIFT(iterator->level));
1901 if (++iterator->level > iterator->end_level) {
1902 iterator->rmap = NULL;
1906 rmap_walk_init_level(iterator, iterator->level);
1909 #define for_each_slot_rmap_range(_slot_, _start_level_, _end_level_, \
1910 _start_gfn, _end_gfn, _iter_) \
1911 for (slot_rmap_walk_init(_iter_, _slot_, _start_level_, \
1912 _end_level_, _start_gfn, _end_gfn); \
1913 slot_rmap_walk_okay(_iter_); \
1914 slot_rmap_walk_next(_iter_))
1916 static int kvm_handle_hva_range(struct kvm *kvm,
1917 unsigned long start,
1920 int (*handler)(struct kvm *kvm,
1921 struct kvm_rmap_head *rmap_head,
1922 struct kvm_memory_slot *slot,
1925 unsigned long data))
1927 struct kvm_memslots *slots;
1928 struct kvm_memory_slot *memslot;
1929 struct slot_rmap_walk_iterator iterator;
1933 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
1934 slots = __kvm_memslots(kvm, i);
1935 kvm_for_each_memslot(memslot, slots) {
1936 unsigned long hva_start, hva_end;
1937 gfn_t gfn_start, gfn_end;
1939 hva_start = max(start, memslot->userspace_addr);
1940 hva_end = min(end, memslot->userspace_addr +
1941 (memslot->npages << PAGE_SHIFT));
1942 if (hva_start >= hva_end)
1945 * {gfn(page) | page intersects with [hva_start, hva_end)} =
1946 * {gfn_start, gfn_start+1, ..., gfn_end-1}.
1948 gfn_start = hva_to_gfn_memslot(hva_start, memslot);
1949 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
1951 for_each_slot_rmap_range(memslot, PT_PAGE_TABLE_LEVEL,
1952 PT_MAX_HUGEPAGE_LEVEL,
1953 gfn_start, gfn_end - 1,
1955 ret |= handler(kvm, iterator.rmap, memslot,
1956 iterator.gfn, iterator.level, data);
1963 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
1965 int (*handler)(struct kvm *kvm,
1966 struct kvm_rmap_head *rmap_head,
1967 struct kvm_memory_slot *slot,
1968 gfn_t gfn, int level,
1969 unsigned long data))
1971 return kvm_handle_hva_range(kvm, hva, hva + 1, data, handler);
1974 int kvm_unmap_hva_range(struct kvm *kvm, unsigned long start, unsigned long end)
1976 return kvm_handle_hva_range(kvm, start, end, 0, kvm_unmap_rmapp);
1979 int kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1981 return kvm_handle_hva(kvm, hva, (unsigned long)&pte, kvm_set_pte_rmapp);
1984 static int kvm_age_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1985 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1989 struct rmap_iterator uninitialized_var(iter);
1992 for_each_rmap_spte(rmap_head, &iter, sptep)
1993 young |= mmu_spte_age(sptep);
1995 trace_kvm_age_page(gfn, level, slot, young);
1999 static int kvm_test_age_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
2000 struct kvm_memory_slot *slot, gfn_t gfn,
2001 int level, unsigned long data)
2004 struct rmap_iterator iter;
2006 for_each_rmap_spte(rmap_head, &iter, sptep)
2007 if (is_accessed_spte(*sptep))
2012 #define RMAP_RECYCLE_THRESHOLD 1000
2014 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
2016 struct kvm_rmap_head *rmap_head;
2017 struct kvm_mmu_page *sp;
2019 sp = page_header(__pa(spte));
2021 rmap_head = gfn_to_rmap(vcpu->kvm, gfn, sp);
2023 kvm_unmap_rmapp(vcpu->kvm, rmap_head, NULL, gfn, sp->role.level, 0);
2024 kvm_flush_remote_tlbs_with_address(vcpu->kvm, sp->gfn,
2025 KVM_PAGES_PER_HPAGE(sp->role.level));
2028 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
2030 return kvm_handle_hva_range(kvm, start, end, 0, kvm_age_rmapp);
2033 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
2035 return kvm_handle_hva(kvm, hva, 0, kvm_test_age_rmapp);
2039 static int is_empty_shadow_page(u64 *spt)
2044 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
2045 if (is_shadow_present_pte(*pos)) {
2046 printk(KERN_ERR "%s: %p %llx\n", __func__,
2055 * This value is the sum of all of the kvm instances's
2056 * kvm->arch.n_used_mmu_pages values. We need a global,
2057 * aggregate version in order to make the slab shrinker
2060 static inline void kvm_mod_used_mmu_pages(struct kvm *kvm, unsigned long nr)
2062 kvm->arch.n_used_mmu_pages += nr;
2063 percpu_counter_add(&kvm_total_used_mmu_pages, nr);
2066 static void kvm_mmu_free_page(struct kvm_mmu_page *sp)
2068 MMU_WARN_ON(!is_empty_shadow_page(sp->spt));
2069 hlist_del(&sp->hash_link);
2070 list_del(&sp->link);
2071 free_page((unsigned long)sp->spt);
2072 if (!sp->role.direct)
2073 free_page((unsigned long)sp->gfns);
2074 kmem_cache_free(mmu_page_header_cache, sp);
2077 static unsigned kvm_page_table_hashfn(gfn_t gfn)
2079 return hash_64(gfn, KVM_MMU_HASH_SHIFT);
2082 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
2083 struct kvm_mmu_page *sp, u64 *parent_pte)
2088 pte_list_add(vcpu, parent_pte, &sp->parent_ptes);
2091 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
2094 __pte_list_remove(parent_pte, &sp->parent_ptes);
2097 static void drop_parent_pte(struct kvm_mmu_page *sp,
2100 mmu_page_remove_parent_pte(sp, parent_pte);
2101 mmu_spte_clear_no_track(parent_pte);
2104 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu, int direct)
2106 struct kvm_mmu_page *sp;
2108 sp = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache);
2109 sp->spt = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache);
2111 sp->gfns = mmu_memory_cache_alloc(&vcpu->arch.mmu_page_cache);
2112 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
2115 * active_mmu_pages must be a FIFO list, as kvm_zap_obsolete_pages()
2116 * depends on valid pages being added to the head of the list. See
2117 * comments in kvm_zap_obsolete_pages().
2119 sp->mmu_valid_gen = vcpu->kvm->arch.mmu_valid_gen;
2120 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
2121 kvm_mod_used_mmu_pages(vcpu->kvm, +1);
2125 static void mark_unsync(u64 *spte);
2126 static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp)
2129 struct rmap_iterator iter;
2131 for_each_rmap_spte(&sp->parent_ptes, &iter, sptep) {
2136 static void mark_unsync(u64 *spte)
2138 struct kvm_mmu_page *sp;
2141 sp = page_header(__pa(spte));
2142 index = spte - sp->spt;
2143 if (__test_and_set_bit(index, sp->unsync_child_bitmap))
2145 if (sp->unsync_children++)
2147 kvm_mmu_mark_parents_unsync(sp);
2150 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
2151 struct kvm_mmu_page *sp)
2156 static void nonpaging_update_pte(struct kvm_vcpu *vcpu,
2157 struct kvm_mmu_page *sp, u64 *spte,
2163 #define KVM_PAGE_ARRAY_NR 16
2165 struct kvm_mmu_pages {
2166 struct mmu_page_and_offset {
2167 struct kvm_mmu_page *sp;
2169 } page[KVM_PAGE_ARRAY_NR];
2173 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
2179 for (i=0; i < pvec->nr; i++)
2180 if (pvec->page[i].sp == sp)
2183 pvec->page[pvec->nr].sp = sp;
2184 pvec->page[pvec->nr].idx = idx;
2186 return (pvec->nr == KVM_PAGE_ARRAY_NR);
2189 static inline void clear_unsync_child_bit(struct kvm_mmu_page *sp, int idx)
2191 --sp->unsync_children;
2192 WARN_ON((int)sp->unsync_children < 0);
2193 __clear_bit(idx, sp->unsync_child_bitmap);
2196 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
2197 struct kvm_mmu_pages *pvec)
2199 int i, ret, nr_unsync_leaf = 0;
2201 for_each_set_bit(i, sp->unsync_child_bitmap, 512) {
2202 struct kvm_mmu_page *child;
2203 u64 ent = sp->spt[i];
2205 if (!is_shadow_present_pte(ent) || is_large_pte(ent)) {
2206 clear_unsync_child_bit(sp, i);
2210 child = page_header(ent & PT64_BASE_ADDR_MASK);
2212 if (child->unsync_children) {
2213 if (mmu_pages_add(pvec, child, i))
2216 ret = __mmu_unsync_walk(child, pvec);
2218 clear_unsync_child_bit(sp, i);
2220 } else if (ret > 0) {
2221 nr_unsync_leaf += ret;
2224 } else if (child->unsync) {
2226 if (mmu_pages_add(pvec, child, i))
2229 clear_unsync_child_bit(sp, i);
2232 return nr_unsync_leaf;
2235 #define INVALID_INDEX (-1)
2237 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
2238 struct kvm_mmu_pages *pvec)
2241 if (!sp->unsync_children)
2244 mmu_pages_add(pvec, sp, INVALID_INDEX);
2245 return __mmu_unsync_walk(sp, pvec);
2248 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
2250 WARN_ON(!sp->unsync);
2251 trace_kvm_mmu_sync_page(sp);
2253 --kvm->stat.mmu_unsync;
2256 static bool kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
2257 struct list_head *invalid_list);
2258 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
2259 struct list_head *invalid_list);
2262 #define for_each_valid_sp(_kvm, _sp, _gfn) \
2263 hlist_for_each_entry(_sp, \
2264 &(_kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(_gfn)], hash_link) \
2265 if (is_obsolete_sp((_kvm), (_sp))) { \
2268 #define for_each_gfn_indirect_valid_sp(_kvm, _sp, _gfn) \
2269 for_each_valid_sp(_kvm, _sp, _gfn) \
2270 if ((_sp)->gfn != (_gfn) || (_sp)->role.direct) {} else
2272 static inline bool is_ept_sp(struct kvm_mmu_page *sp)
2274 return sp->role.cr0_wp && sp->role.smap_andnot_wp;
2277 /* @sp->gfn should be write-protected at the call site */
2278 static bool __kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
2279 struct list_head *invalid_list)
2281 if ((!is_ept_sp(sp) && sp->role.gpte_is_8_bytes != !!is_pae(vcpu)) ||
2282 vcpu->arch.mmu->sync_page(vcpu, sp) == 0) {
2283 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
2290 static bool kvm_mmu_remote_flush_or_zap(struct kvm *kvm,
2291 struct list_head *invalid_list,
2294 if (!remote_flush && list_empty(invalid_list))
2297 if (!list_empty(invalid_list))
2298 kvm_mmu_commit_zap_page(kvm, invalid_list);
2300 kvm_flush_remote_tlbs(kvm);
2304 static void kvm_mmu_flush_or_zap(struct kvm_vcpu *vcpu,
2305 struct list_head *invalid_list,
2306 bool remote_flush, bool local_flush)
2308 if (kvm_mmu_remote_flush_or_zap(vcpu->kvm, invalid_list, remote_flush))
2312 kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
2315 #ifdef CONFIG_KVM_MMU_AUDIT
2316 #include "mmu_audit.c"
2318 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, int point) { }
2319 static void mmu_audit_disable(void) { }
2322 static bool is_obsolete_sp(struct kvm *kvm, struct kvm_mmu_page *sp)
2324 return sp->role.invalid ||
2325 unlikely(sp->mmu_valid_gen != kvm->arch.mmu_valid_gen);
2328 static bool kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
2329 struct list_head *invalid_list)
2331 kvm_unlink_unsync_page(vcpu->kvm, sp);
2332 return __kvm_sync_page(vcpu, sp, invalid_list);
2335 /* @gfn should be write-protected at the call site */
2336 static bool kvm_sync_pages(struct kvm_vcpu *vcpu, gfn_t gfn,
2337 struct list_head *invalid_list)
2339 struct kvm_mmu_page *s;
2342 for_each_gfn_indirect_valid_sp(vcpu->kvm, s, gfn) {
2346 WARN_ON(s->role.level != PT_PAGE_TABLE_LEVEL);
2347 ret |= kvm_sync_page(vcpu, s, invalid_list);
2353 struct mmu_page_path {
2354 struct kvm_mmu_page *parent[PT64_ROOT_MAX_LEVEL];
2355 unsigned int idx[PT64_ROOT_MAX_LEVEL];
2358 #define for_each_sp(pvec, sp, parents, i) \
2359 for (i = mmu_pages_first(&pvec, &parents); \
2360 i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \
2361 i = mmu_pages_next(&pvec, &parents, i))
2363 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
2364 struct mmu_page_path *parents,
2369 for (n = i+1; n < pvec->nr; n++) {
2370 struct kvm_mmu_page *sp = pvec->page[n].sp;
2371 unsigned idx = pvec->page[n].idx;
2372 int level = sp->role.level;
2374 parents->idx[level-1] = idx;
2375 if (level == PT_PAGE_TABLE_LEVEL)
2378 parents->parent[level-2] = sp;
2384 static int mmu_pages_first(struct kvm_mmu_pages *pvec,
2385 struct mmu_page_path *parents)
2387 struct kvm_mmu_page *sp;
2393 WARN_ON(pvec->page[0].idx != INVALID_INDEX);
2395 sp = pvec->page[0].sp;
2396 level = sp->role.level;
2397 WARN_ON(level == PT_PAGE_TABLE_LEVEL);
2399 parents->parent[level-2] = sp;
2401 /* Also set up a sentinel. Further entries in pvec are all
2402 * children of sp, so this element is never overwritten.
2404 parents->parent[level-1] = NULL;
2405 return mmu_pages_next(pvec, parents, 0);
2408 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
2410 struct kvm_mmu_page *sp;
2411 unsigned int level = 0;
2414 unsigned int idx = parents->idx[level];
2415 sp = parents->parent[level];
2419 WARN_ON(idx == INVALID_INDEX);
2420 clear_unsync_child_bit(sp, idx);
2422 } while (!sp->unsync_children);
2425 static void mmu_sync_children(struct kvm_vcpu *vcpu,
2426 struct kvm_mmu_page *parent)
2429 struct kvm_mmu_page *sp;
2430 struct mmu_page_path parents;
2431 struct kvm_mmu_pages pages;
2432 LIST_HEAD(invalid_list);
2435 while (mmu_unsync_walk(parent, &pages)) {
2436 bool protected = false;
2438 for_each_sp(pages, sp, parents, i)
2439 protected |= rmap_write_protect(vcpu, sp->gfn);
2442 kvm_flush_remote_tlbs(vcpu->kvm);
2446 for_each_sp(pages, sp, parents, i) {
2447 flush |= kvm_sync_page(vcpu, sp, &invalid_list);
2448 mmu_pages_clear_parents(&parents);
2450 if (need_resched() || spin_needbreak(&vcpu->kvm->mmu_lock)) {
2451 kvm_mmu_flush_or_zap(vcpu, &invalid_list, false, flush);
2452 cond_resched_lock(&vcpu->kvm->mmu_lock);
2457 kvm_mmu_flush_or_zap(vcpu, &invalid_list, false, flush);
2460 static void __clear_sp_write_flooding_count(struct kvm_mmu_page *sp)
2462 atomic_set(&sp->write_flooding_count, 0);
2465 static void clear_sp_write_flooding_count(u64 *spte)
2467 struct kvm_mmu_page *sp = page_header(__pa(spte));
2469 __clear_sp_write_flooding_count(sp);
2472 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
2477 unsigned int access)
2479 union kvm_mmu_page_role role;
2481 struct kvm_mmu_page *sp;
2482 bool need_sync = false;
2485 LIST_HEAD(invalid_list);
2487 role = vcpu->arch.mmu->mmu_role.base;
2489 role.direct = direct;
2491 role.gpte_is_8_bytes = true;
2492 role.access = access;
2493 if (!vcpu->arch.mmu->direct_map
2494 && vcpu->arch.mmu->root_level <= PT32_ROOT_LEVEL) {
2495 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
2496 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
2497 role.quadrant = quadrant;
2499 for_each_valid_sp(vcpu->kvm, sp, gfn) {
2500 if (sp->gfn != gfn) {
2505 if (!need_sync && sp->unsync)
2508 if (sp->role.word != role.word)
2512 /* The page is good, but __kvm_sync_page might still end
2513 * up zapping it. If so, break in order to rebuild it.
2515 if (!__kvm_sync_page(vcpu, sp, &invalid_list))
2518 WARN_ON(!list_empty(&invalid_list));
2519 kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
2522 if (sp->unsync_children)
2523 kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
2525 __clear_sp_write_flooding_count(sp);
2526 trace_kvm_mmu_get_page(sp, false);
2530 ++vcpu->kvm->stat.mmu_cache_miss;
2532 sp = kvm_mmu_alloc_page(vcpu, direct);
2536 hlist_add_head(&sp->hash_link,
2537 &vcpu->kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)]);
2540 * we should do write protection before syncing pages
2541 * otherwise the content of the synced shadow page may
2542 * be inconsistent with guest page table.
2544 account_shadowed(vcpu->kvm, sp);
2545 if (level == PT_PAGE_TABLE_LEVEL &&
2546 rmap_write_protect(vcpu, gfn))
2547 kvm_flush_remote_tlbs_with_address(vcpu->kvm, gfn, 1);
2549 if (level > PT_PAGE_TABLE_LEVEL && need_sync)
2550 flush |= kvm_sync_pages(vcpu, gfn, &invalid_list);
2552 clear_page(sp->spt);
2553 trace_kvm_mmu_get_page(sp, true);
2555 kvm_mmu_flush_or_zap(vcpu, &invalid_list, false, flush);
2557 if (collisions > vcpu->kvm->stat.max_mmu_page_hash_collisions)
2558 vcpu->kvm->stat.max_mmu_page_hash_collisions = collisions;
2562 static void shadow_walk_init_using_root(struct kvm_shadow_walk_iterator *iterator,
2563 struct kvm_vcpu *vcpu, hpa_t root,
2566 iterator->addr = addr;
2567 iterator->shadow_addr = root;
2568 iterator->level = vcpu->arch.mmu->shadow_root_level;
2570 if (iterator->level == PT64_ROOT_4LEVEL &&
2571 vcpu->arch.mmu->root_level < PT64_ROOT_4LEVEL &&
2572 !vcpu->arch.mmu->direct_map)
2575 if (iterator->level == PT32E_ROOT_LEVEL) {
2577 * prev_root is currently only used for 64-bit hosts. So only
2578 * the active root_hpa is valid here.
2580 BUG_ON(root != vcpu->arch.mmu->root_hpa);
2582 iterator->shadow_addr
2583 = vcpu->arch.mmu->pae_root[(addr >> 30) & 3];
2584 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
2586 if (!iterator->shadow_addr)
2587 iterator->level = 0;
2591 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
2592 struct kvm_vcpu *vcpu, u64 addr)
2594 shadow_walk_init_using_root(iterator, vcpu, vcpu->arch.mmu->root_hpa,
2598 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
2600 if (iterator->level < PT_PAGE_TABLE_LEVEL)
2603 iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
2604 iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
2608 static void __shadow_walk_next(struct kvm_shadow_walk_iterator *iterator,
2611 if (is_last_spte(spte, iterator->level)) {
2612 iterator->level = 0;
2616 iterator->shadow_addr = spte & PT64_BASE_ADDR_MASK;
2620 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
2622 __shadow_walk_next(iterator, *iterator->sptep);
2625 static void link_shadow_page(struct kvm_vcpu *vcpu, u64 *sptep,
2626 struct kvm_mmu_page *sp)
2630 BUILD_BUG_ON(VMX_EPT_WRITABLE_MASK != PT_WRITABLE_MASK);
2632 spte = __pa(sp->spt) | shadow_present_mask | PT_WRITABLE_MASK |
2633 shadow_user_mask | shadow_x_mask | shadow_me_mask;
2635 if (sp_ad_disabled(sp))
2636 spte |= SPTE_AD_DISABLED_MASK;
2638 spte |= shadow_accessed_mask;
2640 mmu_spte_set(sptep, spte);
2642 mmu_page_add_parent_pte(vcpu, sp, sptep);
2644 if (sp->unsync_children || sp->unsync)
2648 static void validate_direct_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2649 unsigned direct_access)
2651 if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep)) {
2652 struct kvm_mmu_page *child;
2655 * For the direct sp, if the guest pte's dirty bit
2656 * changed form clean to dirty, it will corrupt the
2657 * sp's access: allow writable in the read-only sp,
2658 * so we should update the spte at this point to get
2659 * a new sp with the correct access.
2661 child = page_header(*sptep & PT64_BASE_ADDR_MASK);
2662 if (child->role.access == direct_access)
2665 drop_parent_pte(child, sptep);
2666 kvm_flush_remote_tlbs_with_address(vcpu->kvm, child->gfn, 1);
2670 static bool mmu_page_zap_pte(struct kvm *kvm, struct kvm_mmu_page *sp,
2674 struct kvm_mmu_page *child;
2677 if (is_shadow_present_pte(pte)) {
2678 if (is_last_spte(pte, sp->role.level)) {
2679 drop_spte(kvm, spte);
2680 if (is_large_pte(pte))
2683 child = page_header(pte & PT64_BASE_ADDR_MASK);
2684 drop_parent_pte(child, spte);
2689 if (is_mmio_spte(pte))
2690 mmu_spte_clear_no_track(spte);
2695 static void kvm_mmu_page_unlink_children(struct kvm *kvm,
2696 struct kvm_mmu_page *sp)
2700 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2701 mmu_page_zap_pte(kvm, sp, sp->spt + i);
2704 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
2707 struct rmap_iterator iter;
2709 while ((sptep = rmap_get_first(&sp->parent_ptes, &iter)))
2710 drop_parent_pte(sp, sptep);
2713 static int mmu_zap_unsync_children(struct kvm *kvm,
2714 struct kvm_mmu_page *parent,
2715 struct list_head *invalid_list)
2718 struct mmu_page_path parents;
2719 struct kvm_mmu_pages pages;
2721 if (parent->role.level == PT_PAGE_TABLE_LEVEL)
2724 while (mmu_unsync_walk(parent, &pages)) {
2725 struct kvm_mmu_page *sp;
2727 for_each_sp(pages, sp, parents, i) {
2728 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
2729 mmu_pages_clear_parents(&parents);
2737 static bool __kvm_mmu_prepare_zap_page(struct kvm *kvm,
2738 struct kvm_mmu_page *sp,
2739 struct list_head *invalid_list,
2744 trace_kvm_mmu_prepare_zap_page(sp);
2745 ++kvm->stat.mmu_shadow_zapped;
2746 *nr_zapped = mmu_zap_unsync_children(kvm, sp, invalid_list);
2747 kvm_mmu_page_unlink_children(kvm, sp);
2748 kvm_mmu_unlink_parents(kvm, sp);
2750 /* Zapping children means active_mmu_pages has become unstable. */
2751 list_unstable = *nr_zapped;
2753 if (!sp->role.invalid && !sp->role.direct)
2754 unaccount_shadowed(kvm, sp);
2757 kvm_unlink_unsync_page(kvm, sp);
2758 if (!sp->root_count) {
2761 list_move(&sp->link, invalid_list);
2762 kvm_mod_used_mmu_pages(kvm, -1);
2764 list_move(&sp->link, &kvm->arch.active_mmu_pages);
2767 * Obsolete pages cannot be used on any vCPUs, see the comment
2768 * in kvm_mmu_zap_all_fast(). Note, is_obsolete_sp() also
2769 * treats invalid shadow pages as being obsolete.
2771 if (!is_obsolete_sp(kvm, sp))
2772 kvm_reload_remote_mmus(kvm);
2775 if (sp->lpage_disallowed)
2776 unaccount_huge_nx_page(kvm, sp);
2778 sp->role.invalid = 1;
2779 return list_unstable;
2782 static bool kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
2783 struct list_head *invalid_list)
2787 __kvm_mmu_prepare_zap_page(kvm, sp, invalid_list, &nr_zapped);
2791 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
2792 struct list_head *invalid_list)
2794 struct kvm_mmu_page *sp, *nsp;
2796 if (list_empty(invalid_list))
2800 * We need to make sure everyone sees our modifications to
2801 * the page tables and see changes to vcpu->mode here. The barrier
2802 * in the kvm_flush_remote_tlbs() achieves this. This pairs
2803 * with vcpu_enter_guest and walk_shadow_page_lockless_begin/end.
2805 * In addition, kvm_flush_remote_tlbs waits for all vcpus to exit
2806 * guest mode and/or lockless shadow page table walks.
2808 kvm_flush_remote_tlbs(kvm);
2810 list_for_each_entry_safe(sp, nsp, invalid_list, link) {
2811 WARN_ON(!sp->role.invalid || sp->root_count);
2812 kvm_mmu_free_page(sp);
2816 static bool prepare_zap_oldest_mmu_page(struct kvm *kvm,
2817 struct list_head *invalid_list)
2819 struct kvm_mmu_page *sp;
2821 if (list_empty(&kvm->arch.active_mmu_pages))
2824 sp = list_last_entry(&kvm->arch.active_mmu_pages,
2825 struct kvm_mmu_page, link);
2826 return kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
2829 static int make_mmu_pages_available(struct kvm_vcpu *vcpu)
2831 LIST_HEAD(invalid_list);
2833 if (likely(kvm_mmu_available_pages(vcpu->kvm) >= KVM_MIN_FREE_MMU_PAGES))
2836 while (kvm_mmu_available_pages(vcpu->kvm) < KVM_REFILL_PAGES) {
2837 if (!prepare_zap_oldest_mmu_page(vcpu->kvm, &invalid_list))
2840 ++vcpu->kvm->stat.mmu_recycled;
2842 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2844 if (!kvm_mmu_available_pages(vcpu->kvm))
2850 * Changing the number of mmu pages allocated to the vm
2851 * Note: if goal_nr_mmu_pages is too small, you will get dead lock
2853 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned long goal_nr_mmu_pages)
2855 LIST_HEAD(invalid_list);
2857 spin_lock(&kvm->mmu_lock);
2859 if (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages) {
2860 /* Need to free some mmu pages to achieve the goal. */
2861 while (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages)
2862 if (!prepare_zap_oldest_mmu_page(kvm, &invalid_list))
2865 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2866 goal_nr_mmu_pages = kvm->arch.n_used_mmu_pages;
2869 kvm->arch.n_max_mmu_pages = goal_nr_mmu_pages;
2871 spin_unlock(&kvm->mmu_lock);
2874 int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
2876 struct kvm_mmu_page *sp;
2877 LIST_HEAD(invalid_list);
2880 pgprintk("%s: looking for gfn %llx\n", __func__, gfn);
2882 spin_lock(&kvm->mmu_lock);
2883 for_each_gfn_indirect_valid_sp(kvm, sp, gfn) {
2884 pgprintk("%s: gfn %llx role %x\n", __func__, gfn,
2887 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
2889 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2890 spin_unlock(&kvm->mmu_lock);
2894 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page);
2896 static void kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
2898 trace_kvm_mmu_unsync_page(sp);
2899 ++vcpu->kvm->stat.mmu_unsync;
2902 kvm_mmu_mark_parents_unsync(sp);
2905 static bool mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn,
2908 struct kvm_mmu_page *sp;
2910 if (kvm_page_track_is_active(vcpu, gfn, KVM_PAGE_TRACK_WRITE))
2913 for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn) {
2920 WARN_ON(sp->role.level != PT_PAGE_TABLE_LEVEL);
2921 kvm_unsync_page(vcpu, sp);
2925 * We need to ensure that the marking of unsync pages is visible
2926 * before the SPTE is updated to allow writes because
2927 * kvm_mmu_sync_roots() checks the unsync flags without holding
2928 * the MMU lock and so can race with this. If the SPTE was updated
2929 * before the page had been marked as unsync-ed, something like the
2930 * following could happen:
2933 * ---------------------------------------------------------------------
2934 * 1.2 Host updates SPTE
2936 * 2.1 Guest writes a GPTE for GVA X.
2937 * (GPTE being in the guest page table shadowed
2938 * by the SP from CPU 1.)
2939 * This reads SPTE during the page table walk.
2940 * Since SPTE.W is read as 1, there is no
2943 * 2.2 Guest issues TLB flush.
2944 * That causes a VM Exit.
2946 * 2.3 kvm_mmu_sync_pages() reads sp->unsync.
2947 * Since it is false, so it just returns.
2949 * 2.4 Guest accesses GVA X.
2950 * Since the mapping in the SP was not updated,
2951 * so the old mapping for GVA X incorrectly
2955 * (sp->unsync = true)
2957 * The write barrier below ensures that 1.1 happens before 1.2 and thus
2958 * the situation in 2.4 does not arise. The implicit barrier in 2.2
2959 * pairs with this write barrier.
2966 static bool kvm_is_mmio_pfn(kvm_pfn_t pfn)
2969 return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) &&
2971 * Some reserved pages, such as those from NVDIMM
2972 * DAX devices, are not for MMIO, and can be mapped
2973 * with cached memory type for better performance.
2974 * However, the above check misconceives those pages
2975 * as MMIO, and results in KVM mapping them with UC
2976 * memory type, which would hurt the performance.
2977 * Therefore, we check the host memory type in addition
2978 * and only treat UC/UC-/WC pages as MMIO.
2980 (!pat_enabled() || pat_pfn_immune_to_uc_mtrr(pfn));
2982 return !e820__mapped_raw_any(pfn_to_hpa(pfn),
2983 pfn_to_hpa(pfn + 1) - 1,
2987 /* Bits which may be returned by set_spte() */
2988 #define SET_SPTE_WRITE_PROTECTED_PT BIT(0)
2989 #define SET_SPTE_NEED_REMOTE_TLB_FLUSH BIT(1)
2991 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2992 unsigned int pte_access, int level,
2993 gfn_t gfn, kvm_pfn_t pfn, bool speculative,
2994 bool can_unsync, bool host_writable)
2998 struct kvm_mmu_page *sp;
3000 if (set_mmio_spte(vcpu, sptep, gfn, pfn, pte_access))
3003 sp = page_header(__pa(sptep));
3004 if (sp_ad_disabled(sp))
3005 spte |= SPTE_AD_DISABLED_MASK;
3006 else if (kvm_vcpu_ad_need_write_protect(vcpu))
3007 spte |= SPTE_AD_WRPROT_ONLY_MASK;
3010 * For the EPT case, shadow_present_mask is 0 if hardware
3011 * supports exec-only page table entries. In that case,
3012 * ACC_USER_MASK and shadow_user_mask are used to represent
3013 * read access. See FNAME(gpte_access) in paging_tmpl.h.
3015 spte |= shadow_present_mask;
3017 spte |= spte_shadow_accessed_mask(spte);
3019 if (level > PT_PAGE_TABLE_LEVEL && (pte_access & ACC_EXEC_MASK) &&
3020 is_nx_huge_page_enabled()) {
3021 pte_access &= ~ACC_EXEC_MASK;
3024 if (pte_access & ACC_EXEC_MASK)
3025 spte |= shadow_x_mask;
3027 spte |= shadow_nx_mask;
3029 if (pte_access & ACC_USER_MASK)
3030 spte |= shadow_user_mask;
3032 if (level > PT_PAGE_TABLE_LEVEL)
3033 spte |= PT_PAGE_SIZE_MASK;
3035 spte |= kvm_x86_ops.get_mt_mask(vcpu, gfn,
3036 kvm_is_mmio_pfn(pfn));
3039 spte |= SPTE_HOST_WRITEABLE;
3041 pte_access &= ~ACC_WRITE_MASK;
3043 if (!kvm_is_mmio_pfn(pfn))
3044 spte |= shadow_me_mask;
3046 spte |= (u64)pfn << PAGE_SHIFT;
3048 if (pte_access & ACC_WRITE_MASK) {
3049 spte |= PT_WRITABLE_MASK | SPTE_MMU_WRITEABLE;
3052 * Optimization: for pte sync, if spte was writable the hash
3053 * lookup is unnecessary (and expensive). Write protection
3054 * is responsibility of mmu_get_page / kvm_sync_page.
3055 * Same reasoning can be applied to dirty page accounting.
3057 if (!can_unsync && is_writable_pte(*sptep))
3060 if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
3061 pgprintk("%s: found shadow page for %llx, marking ro\n",
3063 ret |= SET_SPTE_WRITE_PROTECTED_PT;
3064 pte_access &= ~ACC_WRITE_MASK;
3065 spte &= ~(PT_WRITABLE_MASK | SPTE_MMU_WRITEABLE);
3069 if (pte_access & ACC_WRITE_MASK) {
3070 kvm_vcpu_mark_page_dirty(vcpu, gfn);
3071 spte |= spte_shadow_dirty_mask(spte);
3075 spte = mark_spte_for_access_track(spte);
3078 if (mmu_spte_update(sptep, spte))
3079 ret |= SET_SPTE_NEED_REMOTE_TLB_FLUSH;
3083 static int mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
3084 unsigned int pte_access, int write_fault, int level,
3085 gfn_t gfn, kvm_pfn_t pfn, bool speculative,
3088 int was_rmapped = 0;
3091 int ret = RET_PF_RETRY;
3094 pgprintk("%s: spte %llx write_fault %d gfn %llx\n", __func__,
3095 *sptep, write_fault, gfn);
3097 if (is_shadow_present_pte(*sptep)) {
3099 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
3100 * the parent of the now unreachable PTE.
3102 if (level > PT_PAGE_TABLE_LEVEL &&
3103 !is_large_pte(*sptep)) {
3104 struct kvm_mmu_page *child;
3107 child = page_header(pte & PT64_BASE_ADDR_MASK);
3108 drop_parent_pte(child, sptep);
3110 } else if (pfn != spte_to_pfn(*sptep)) {
3111 pgprintk("hfn old %llx new %llx\n",
3112 spte_to_pfn(*sptep), pfn);
3113 drop_spte(vcpu->kvm, sptep);
3119 set_spte_ret = set_spte(vcpu, sptep, pte_access, level, gfn, pfn,
3120 speculative, true, host_writable);
3121 if (set_spte_ret & SET_SPTE_WRITE_PROTECTED_PT) {
3123 ret = RET_PF_EMULATE;
3124 kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
3127 if (set_spte_ret & SET_SPTE_NEED_REMOTE_TLB_FLUSH || flush)
3128 kvm_flush_remote_tlbs_with_address(vcpu->kvm, gfn,
3129 KVM_PAGES_PER_HPAGE(level));
3131 if (unlikely(is_mmio_spte(*sptep)))
3132 ret = RET_PF_EMULATE;
3134 pgprintk("%s: setting spte %llx\n", __func__, *sptep);
3135 trace_kvm_mmu_set_spte(level, gfn, sptep);
3136 if (!was_rmapped && is_large_pte(*sptep))
3137 ++vcpu->kvm->stat.lpages;
3139 if (is_shadow_present_pte(*sptep)) {
3141 rmap_count = rmap_add(vcpu, sptep, gfn);
3142 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
3143 rmap_recycle(vcpu, sptep, gfn);
3150 static kvm_pfn_t pte_prefetch_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn,
3153 struct kvm_memory_slot *slot;
3155 slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, no_dirty_log);
3157 return KVM_PFN_ERR_FAULT;
3159 return gfn_to_pfn_memslot_atomic(slot, gfn);
3162 static int direct_pte_prefetch_many(struct kvm_vcpu *vcpu,
3163 struct kvm_mmu_page *sp,
3164 u64 *start, u64 *end)
3166 struct page *pages[PTE_PREFETCH_NUM];
3167 struct kvm_memory_slot *slot;
3168 unsigned int access = sp->role.access;
3172 gfn = kvm_mmu_page_get_gfn(sp, start - sp->spt);
3173 slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, access & ACC_WRITE_MASK);
3177 ret = gfn_to_page_many_atomic(slot, gfn, pages, end - start);
3181 for (i = 0; i < ret; i++, gfn++, start++) {
3182 mmu_set_spte(vcpu, start, access, 0, sp->role.level, gfn,
3183 page_to_pfn(pages[i]), true, true);
3190 static void __direct_pte_prefetch(struct kvm_vcpu *vcpu,
3191 struct kvm_mmu_page *sp, u64 *sptep)
3193 u64 *spte, *start = NULL;
3196 WARN_ON(!sp->role.direct);
3198 i = (sptep - sp->spt) & ~(PTE_PREFETCH_NUM - 1);
3201 for (i = 0; i < PTE_PREFETCH_NUM; i++, spte++) {
3202 if (is_shadow_present_pte(*spte) || spte == sptep) {
3205 if (direct_pte_prefetch_many(vcpu, sp, start, spte) < 0)
3213 static void direct_pte_prefetch(struct kvm_vcpu *vcpu, u64 *sptep)
3215 struct kvm_mmu_page *sp;
3217 sp = page_header(__pa(sptep));
3220 * Without accessed bits, there's no way to distinguish between
3221 * actually accessed translations and prefetched, so disable pte
3222 * prefetch if accessed bits aren't available.
3224 if (sp_ad_disabled(sp))
3227 if (sp->role.level > PT_PAGE_TABLE_LEVEL)
3230 __direct_pte_prefetch(vcpu, sp, sptep);
3233 static int host_pfn_mapping_level(struct kvm_vcpu *vcpu, gfn_t gfn,
3234 kvm_pfn_t pfn, struct kvm_memory_slot *slot)
3240 BUILD_BUG_ON(PT_PAGE_TABLE_LEVEL != (int)PG_LEVEL_4K ||
3241 PT_DIRECTORY_LEVEL != (int)PG_LEVEL_2M ||
3242 PT_PDPE_LEVEL != (int)PG_LEVEL_1G);
3244 if (!PageCompound(pfn_to_page(pfn)) && !kvm_is_zone_device_pfn(pfn))
3245 return PT_PAGE_TABLE_LEVEL;
3248 * Note, using the already-retrieved memslot and __gfn_to_hva_memslot()
3249 * is not solely for performance, it's also necessary to avoid the
3250 * "writable" check in __gfn_to_hva_many(), which will always fail on
3251 * read-only memslots due to gfn_to_hva() assuming writes. Earlier
3252 * page fault steps have already verified the guest isn't writing a
3253 * read-only memslot.
3255 hva = __gfn_to_hva_memslot(slot, gfn);
3257 pte = lookup_address_in_mm(vcpu->kvm->mm, hva, &level);
3259 return PT_PAGE_TABLE_LEVEL;
3264 static int kvm_mmu_hugepage_adjust(struct kvm_vcpu *vcpu, gfn_t gfn,
3265 int max_level, kvm_pfn_t *pfnp)
3267 struct kvm_memory_slot *slot;
3268 struct kvm_lpage_info *linfo;
3269 kvm_pfn_t pfn = *pfnp;
3273 if (unlikely(max_level == PT_PAGE_TABLE_LEVEL))
3274 return PT_PAGE_TABLE_LEVEL;
3276 if (is_error_noslot_pfn(pfn) || kvm_is_reserved_pfn(pfn))
3277 return PT_PAGE_TABLE_LEVEL;
3279 slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, true);
3281 return PT_PAGE_TABLE_LEVEL;
3283 max_level = min(max_level, max_page_level);
3284 for ( ; max_level > PT_PAGE_TABLE_LEVEL; max_level--) {
3285 linfo = lpage_info_slot(gfn, slot, max_level);
3286 if (!linfo->disallow_lpage)
3290 if (max_level == PT_PAGE_TABLE_LEVEL)
3291 return PT_PAGE_TABLE_LEVEL;
3293 level = host_pfn_mapping_level(vcpu, gfn, pfn, slot);
3294 if (level == PT_PAGE_TABLE_LEVEL)
3297 level = min(level, max_level);
3300 * mmu_notifier_retry() was successful and mmu_lock is held, so
3301 * the pmd can't be split from under us.
3303 mask = KVM_PAGES_PER_HPAGE(level) - 1;
3304 VM_BUG_ON((gfn & mask) != (pfn & mask));
3305 *pfnp = pfn & ~mask;
3310 static void disallowed_hugepage_adjust(struct kvm_shadow_walk_iterator it,
3311 gfn_t gfn, kvm_pfn_t *pfnp, int *levelp)
3313 int level = *levelp;
3314 u64 spte = *it.sptep;
3316 if (it.level == level && level > PT_PAGE_TABLE_LEVEL &&
3317 is_nx_huge_page_enabled() &&
3318 is_shadow_present_pte(spte) &&
3319 !is_large_pte(spte)) {
3321 * A small SPTE exists for this pfn, but FNAME(fetch)
3322 * and __direct_map would like to create a large PTE
3323 * instead: just force them to go down another level,
3324 * patching back for them into pfn the next 9 bits of
3327 u64 page_mask = KVM_PAGES_PER_HPAGE(level) - KVM_PAGES_PER_HPAGE(level - 1);
3328 *pfnp |= gfn & page_mask;
3333 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t gpa, int write,
3334 int map_writable, int max_level, kvm_pfn_t pfn,
3335 bool prefault, bool account_disallowed_nx_lpage)
3337 struct kvm_shadow_walk_iterator it;
3338 struct kvm_mmu_page *sp;
3340 gfn_t gfn = gpa >> PAGE_SHIFT;
3341 gfn_t base_gfn = gfn;
3343 if (WARN_ON(!VALID_PAGE(vcpu->arch.mmu->root_hpa)))
3344 return RET_PF_RETRY;
3346 level = kvm_mmu_hugepage_adjust(vcpu, gfn, max_level, &pfn);
3348 trace_kvm_mmu_spte_requested(gpa, level, pfn);
3349 for_each_shadow_entry(vcpu, gpa, it) {
3351 * We cannot overwrite existing page tables with an NX
3352 * large page, as the leaf could be executable.
3354 disallowed_hugepage_adjust(it, gfn, &pfn, &level);
3356 base_gfn = gfn & ~(KVM_PAGES_PER_HPAGE(it.level) - 1);
3357 if (it.level == level)
3360 drop_large_spte(vcpu, it.sptep);
3361 if (!is_shadow_present_pte(*it.sptep)) {
3362 sp = kvm_mmu_get_page(vcpu, base_gfn, it.addr,
3363 it.level - 1, true, ACC_ALL);
3365 link_shadow_page(vcpu, it.sptep, sp);
3366 if (account_disallowed_nx_lpage)
3367 account_huge_nx_page(vcpu->kvm, sp);
3371 ret = mmu_set_spte(vcpu, it.sptep, ACC_ALL,
3372 write, level, base_gfn, pfn, prefault,
3374 direct_pte_prefetch(vcpu, it.sptep);
3375 ++vcpu->stat.pf_fixed;
3379 static void kvm_send_hwpoison_signal(unsigned long address, struct task_struct *tsk)
3381 send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, PAGE_SHIFT, tsk);
3384 static int kvm_handle_bad_page(struct kvm_vcpu *vcpu, gfn_t gfn, kvm_pfn_t pfn)
3387 * Do not cache the mmio info caused by writing the readonly gfn
3388 * into the spte otherwise read access on readonly gfn also can
3389 * caused mmio page fault and treat it as mmio access.
3391 if (pfn == KVM_PFN_ERR_RO_FAULT)
3392 return RET_PF_EMULATE;
3394 if (pfn == KVM_PFN_ERR_HWPOISON) {
3395 kvm_send_hwpoison_signal(kvm_vcpu_gfn_to_hva(vcpu, gfn), current);
3396 return RET_PF_RETRY;
3402 static bool handle_abnormal_pfn(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn,
3403 kvm_pfn_t pfn, unsigned int access,
3406 /* The pfn is invalid, report the error! */
3407 if (unlikely(is_error_pfn(pfn))) {
3408 *ret_val = kvm_handle_bad_page(vcpu, gfn, pfn);
3412 if (unlikely(is_noslot_pfn(pfn)))
3413 vcpu_cache_mmio_info(vcpu, gva, gfn,
3414 access & shadow_mmio_access_mask);
3419 static bool page_fault_can_be_fast(u32 error_code)
3422 * Do not fix the mmio spte with invalid generation number which
3423 * need to be updated by slow page fault path.
3425 if (unlikely(error_code & PFERR_RSVD_MASK))
3428 /* See if the page fault is due to an NX violation */
3429 if (unlikely(((error_code & (PFERR_FETCH_MASK | PFERR_PRESENT_MASK))
3430 == (PFERR_FETCH_MASK | PFERR_PRESENT_MASK))))
3434 * #PF can be fast if:
3435 * 1. The shadow page table entry is not present, which could mean that
3436 * the fault is potentially caused by access tracking (if enabled).
3437 * 2. The shadow page table entry is present and the fault
3438 * is caused by write-protect, that means we just need change the W
3439 * bit of the spte which can be done out of mmu-lock.
3441 * However, if access tracking is disabled we know that a non-present
3442 * page must be a genuine page fault where we have to create a new SPTE.
3443 * So, if access tracking is disabled, we return true only for write
3444 * accesses to a present page.
3447 return shadow_acc_track_mask != 0 ||
3448 ((error_code & (PFERR_WRITE_MASK | PFERR_PRESENT_MASK))
3449 == (PFERR_WRITE_MASK | PFERR_PRESENT_MASK));
3453 * Returns true if the SPTE was fixed successfully. Otherwise,
3454 * someone else modified the SPTE from its original value.
3457 fast_pf_fix_direct_spte(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
3458 u64 *sptep, u64 old_spte, u64 new_spte)
3462 WARN_ON(!sp->role.direct);
3465 * Theoretically we could also set dirty bit (and flush TLB) here in
3466 * order to eliminate unnecessary PML logging. See comments in
3467 * set_spte. But fast_page_fault is very unlikely to happen with PML
3468 * enabled, so we do not do this. This might result in the same GPA
3469 * to be logged in PML buffer again when the write really happens, and
3470 * eventually to be called by mark_page_dirty twice. But it's also no
3471 * harm. This also avoids the TLB flush needed after setting dirty bit
3472 * so non-PML cases won't be impacted.
3474 * Compare with set_spte where instead shadow_dirty_mask is set.
3476 if (cmpxchg64(sptep, old_spte, new_spte) != old_spte)
3479 if (is_writable_pte(new_spte) && !is_writable_pte(old_spte)) {
3481 * The gfn of direct spte is stable since it is
3482 * calculated by sp->gfn.
3484 gfn = kvm_mmu_page_get_gfn(sp, sptep - sp->spt);
3485 kvm_vcpu_mark_page_dirty(vcpu, gfn);
3491 static bool is_access_allowed(u32 fault_err_code, u64 spte)
3493 if (fault_err_code & PFERR_FETCH_MASK)
3494 return is_executable_pte(spte);
3496 if (fault_err_code & PFERR_WRITE_MASK)
3497 return is_writable_pte(spte);
3499 /* Fault was on Read access */
3500 return spte & PT_PRESENT_MASK;
3505 * - true: let the vcpu to access on the same address again.
3506 * - false: let the real page fault path to fix it.
3508 static bool fast_page_fault(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa,
3511 struct kvm_shadow_walk_iterator iterator;
3512 struct kvm_mmu_page *sp;
3513 bool fault_handled = false;
3515 uint retry_count = 0;
3517 if (!page_fault_can_be_fast(error_code))
3520 walk_shadow_page_lockless_begin(vcpu);
3525 for_each_shadow_entry_lockless(vcpu, cr2_or_gpa, iterator, spte)
3526 if (!is_shadow_present_pte(spte))
3529 sp = page_header(__pa(iterator.sptep));
3530 if (!is_last_spte(spte, sp->role.level))
3534 * Check whether the memory access that caused the fault would
3535 * still cause it if it were to be performed right now. If not,
3536 * then this is a spurious fault caused by TLB lazily flushed,
3537 * or some other CPU has already fixed the PTE after the
3538 * current CPU took the fault.
3540 * Need not check the access of upper level table entries since
3541 * they are always ACC_ALL.
3543 if (is_access_allowed(error_code, spte)) {
3544 fault_handled = true;
3550 if (is_access_track_spte(spte))
3551 new_spte = restore_acc_track_spte(new_spte);
3554 * Currently, to simplify the code, write-protection can
3555 * be removed in the fast path only if the SPTE was
3556 * write-protected for dirty-logging or access tracking.
3558 if ((error_code & PFERR_WRITE_MASK) &&
3559 spte_can_locklessly_be_made_writable(spte)) {
3560 new_spte |= PT_WRITABLE_MASK;
3563 * Do not fix write-permission on the large spte. Since
3564 * we only dirty the first page into the dirty-bitmap in
3565 * fast_pf_fix_direct_spte(), other pages are missed
3566 * if its slot has dirty logging enabled.
3568 * Instead, we let the slow page fault path create a
3569 * normal spte to fix the access.
3571 * See the comments in kvm_arch_commit_memory_region().
3573 if (sp->role.level > PT_PAGE_TABLE_LEVEL)
3577 /* Verify that the fault can be handled in the fast path */
3578 if (new_spte == spte ||
3579 !is_access_allowed(error_code, new_spte))
3583 * Currently, fast page fault only works for direct mapping
3584 * since the gfn is not stable for indirect shadow page. See
3585 * Documentation/virt/kvm/locking.txt to get more detail.
3587 fault_handled = fast_pf_fix_direct_spte(vcpu, sp,
3588 iterator.sptep, spte,
3593 if (++retry_count > 4) {
3594 printk_once(KERN_WARNING
3595 "kvm: Fast #PF retrying more than 4 times.\n");
3601 trace_fast_page_fault(vcpu, cr2_or_gpa, error_code, iterator.sptep,
3602 spte, fault_handled);
3603 walk_shadow_page_lockless_end(vcpu);
3605 return fault_handled;
3608 static void mmu_free_root_page(struct kvm *kvm, hpa_t *root_hpa,
3609 struct list_head *invalid_list)
3611 struct kvm_mmu_page *sp;
3613 if (!VALID_PAGE(*root_hpa))
3616 sp = page_header(*root_hpa & PT64_BASE_ADDR_MASK);
3618 if (!sp->root_count && sp->role.invalid)
3619 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
3621 *root_hpa = INVALID_PAGE;
3624 /* roots_to_free must be some combination of the KVM_MMU_ROOT_* flags */
3625 void kvm_mmu_free_roots(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
3626 ulong roots_to_free)
3629 LIST_HEAD(invalid_list);
3630 bool free_active_root = roots_to_free & KVM_MMU_ROOT_CURRENT;
3632 BUILD_BUG_ON(KVM_MMU_NUM_PREV_ROOTS >= BITS_PER_LONG);
3634 /* Before acquiring the MMU lock, see if we need to do any real work. */
3635 if (!(free_active_root && VALID_PAGE(mmu->root_hpa))) {
3636 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
3637 if ((roots_to_free & KVM_MMU_ROOT_PREVIOUS(i)) &&
3638 VALID_PAGE(mmu->prev_roots[i].hpa))
3641 if (i == KVM_MMU_NUM_PREV_ROOTS)
3645 spin_lock(&vcpu->kvm->mmu_lock);
3647 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
3648 if (roots_to_free & KVM_MMU_ROOT_PREVIOUS(i))
3649 mmu_free_root_page(vcpu->kvm, &mmu->prev_roots[i].hpa,
3652 if (free_active_root) {
3653 if (mmu->shadow_root_level >= PT64_ROOT_4LEVEL &&
3654 (mmu->root_level >= PT64_ROOT_4LEVEL || mmu->direct_map)) {
3655 mmu_free_root_page(vcpu->kvm, &mmu->root_hpa,
3658 for (i = 0; i < 4; ++i)
3659 if (mmu->pae_root[i] != 0)
3660 mmu_free_root_page(vcpu->kvm,
3663 mmu->root_hpa = INVALID_PAGE;
3668 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
3669 spin_unlock(&vcpu->kvm->mmu_lock);
3671 EXPORT_SYMBOL_GPL(kvm_mmu_free_roots);
3673 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
3677 if (!kvm_is_visible_gfn(vcpu->kvm, root_gfn)) {
3678 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
3685 static int mmu_alloc_direct_roots(struct kvm_vcpu *vcpu)
3687 struct kvm_mmu_page *sp;
3690 if (vcpu->arch.mmu->shadow_root_level >= PT64_ROOT_4LEVEL) {
3691 spin_lock(&vcpu->kvm->mmu_lock);
3692 if(make_mmu_pages_available(vcpu) < 0) {
3693 spin_unlock(&vcpu->kvm->mmu_lock);
3696 sp = kvm_mmu_get_page(vcpu, 0, 0,
3697 vcpu->arch.mmu->shadow_root_level, 1, ACC_ALL);
3699 spin_unlock(&vcpu->kvm->mmu_lock);
3700 vcpu->arch.mmu->root_hpa = __pa(sp->spt);
3701 } else if (vcpu->arch.mmu->shadow_root_level == PT32E_ROOT_LEVEL) {
3702 for (i = 0; i < 4; ++i) {
3703 hpa_t root = vcpu->arch.mmu->pae_root[i];
3705 MMU_WARN_ON(VALID_PAGE(root));
3706 spin_lock(&vcpu->kvm->mmu_lock);
3707 if (make_mmu_pages_available(vcpu) < 0) {
3708 spin_unlock(&vcpu->kvm->mmu_lock);
3711 sp = kvm_mmu_get_page(vcpu, i << (30 - PAGE_SHIFT),
3712 i << 30, PT32_ROOT_LEVEL, 1, ACC_ALL);
3713 root = __pa(sp->spt);
3715 spin_unlock(&vcpu->kvm->mmu_lock);
3716 vcpu->arch.mmu->pae_root[i] = root | PT_PRESENT_MASK;
3718 vcpu->arch.mmu->root_hpa = __pa(vcpu->arch.mmu->pae_root);
3722 /* root_cr3 is ignored for direct MMUs. */
3723 vcpu->arch.mmu->root_cr3 = 0;
3728 static int mmu_alloc_shadow_roots(struct kvm_vcpu *vcpu)
3730 struct kvm_mmu_page *sp;
3732 gfn_t root_gfn, root_cr3;
3735 root_cr3 = vcpu->arch.mmu->get_guest_pgd(vcpu);
3736 root_gfn = root_cr3 >> PAGE_SHIFT;
3738 if (mmu_check_root(vcpu, root_gfn))
3742 * Do we shadow a long mode page table? If so we need to
3743 * write-protect the guests page table root.
3745 if (vcpu->arch.mmu->root_level >= PT64_ROOT_4LEVEL) {
3746 hpa_t root = vcpu->arch.mmu->root_hpa;
3748 MMU_WARN_ON(VALID_PAGE(root));
3750 spin_lock(&vcpu->kvm->mmu_lock);
3751 if (make_mmu_pages_available(vcpu) < 0) {
3752 spin_unlock(&vcpu->kvm->mmu_lock);
3755 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
3756 vcpu->arch.mmu->shadow_root_level, 0, ACC_ALL);
3757 root = __pa(sp->spt);
3759 spin_unlock(&vcpu->kvm->mmu_lock);
3760 vcpu->arch.mmu->root_hpa = root;
3765 * We shadow a 32 bit page table. This may be a legacy 2-level
3766 * or a PAE 3-level page table. In either case we need to be aware that
3767 * the shadow page table may be a PAE or a long mode page table.
3769 pm_mask = PT_PRESENT_MASK;
3770 if (vcpu->arch.mmu->shadow_root_level == PT64_ROOT_4LEVEL)
3771 pm_mask |= PT_ACCESSED_MASK | PT_WRITABLE_MASK | PT_USER_MASK;
3773 for (i = 0; i < 4; ++i) {
3774 hpa_t root = vcpu->arch.mmu->pae_root[i];
3776 MMU_WARN_ON(VALID_PAGE(root));
3777 if (vcpu->arch.mmu->root_level == PT32E_ROOT_LEVEL) {
3778 pdptr = vcpu->arch.mmu->get_pdptr(vcpu, i);
3779 if (!(pdptr & PT_PRESENT_MASK)) {
3780 vcpu->arch.mmu->pae_root[i] = 0;
3783 root_gfn = pdptr >> PAGE_SHIFT;
3784 if (mmu_check_root(vcpu, root_gfn))
3787 spin_lock(&vcpu->kvm->mmu_lock);
3788 if (make_mmu_pages_available(vcpu) < 0) {
3789 spin_unlock(&vcpu->kvm->mmu_lock);
3792 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30, PT32_ROOT_LEVEL,
3794 root = __pa(sp->spt);
3796 spin_unlock(&vcpu->kvm->mmu_lock);
3798 vcpu->arch.mmu->pae_root[i] = root | pm_mask;
3800 vcpu->arch.mmu->root_hpa = __pa(vcpu->arch.mmu->pae_root);
3803 * If we shadow a 32 bit page table with a long mode page
3804 * table we enter this path.
3806 if (vcpu->arch.mmu->shadow_root_level == PT64_ROOT_4LEVEL) {
3807 if (vcpu->arch.mmu->lm_root == NULL) {
3809 * The additional page necessary for this is only
3810 * allocated on demand.
3815 lm_root = (void*)get_zeroed_page(GFP_KERNEL_ACCOUNT);
3816 if (lm_root == NULL)
3819 lm_root[0] = __pa(vcpu->arch.mmu->pae_root) | pm_mask;
3821 vcpu->arch.mmu->lm_root = lm_root;
3824 vcpu->arch.mmu->root_hpa = __pa(vcpu->arch.mmu->lm_root);
3828 vcpu->arch.mmu->root_cr3 = root_cr3;
3833 static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
3835 if (vcpu->arch.mmu->direct_map)
3836 return mmu_alloc_direct_roots(vcpu);
3838 return mmu_alloc_shadow_roots(vcpu);
3841 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
3844 struct kvm_mmu_page *sp;
3846 if (vcpu->arch.mmu->direct_map)
3849 if (!VALID_PAGE(vcpu->arch.mmu->root_hpa))
3852 vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY);
3854 if (vcpu->arch.mmu->root_level >= PT64_ROOT_4LEVEL) {
3855 hpa_t root = vcpu->arch.mmu->root_hpa;
3856 sp = page_header(root);
3859 * Even if another CPU was marking the SP as unsync-ed
3860 * simultaneously, any guest page table changes are not
3861 * guaranteed to be visible anyway until this VCPU issues a TLB
3862 * flush strictly after those changes are made. We only need to
3863 * ensure that the other CPU sets these flags before any actual
3864 * changes to the page tables are made. The comments in
3865 * mmu_need_write_protect() describe what could go wrong if this
3866 * requirement isn't satisfied.
3868 if (!smp_load_acquire(&sp->unsync) &&
3869 !smp_load_acquire(&sp->unsync_children))
3872 spin_lock(&vcpu->kvm->mmu_lock);
3873 kvm_mmu_audit(vcpu, AUDIT_PRE_SYNC);
3875 mmu_sync_children(vcpu, sp);
3877 kvm_mmu_audit(vcpu, AUDIT_POST_SYNC);
3878 spin_unlock(&vcpu->kvm->mmu_lock);
3882 spin_lock(&vcpu->kvm->mmu_lock);
3883 kvm_mmu_audit(vcpu, AUDIT_PRE_SYNC);
3885 for (i = 0; i < 4; ++i) {
3886 hpa_t root = vcpu->arch.mmu->pae_root[i];
3888 if (root && VALID_PAGE(root)) {
3889 root &= PT64_BASE_ADDR_MASK;
3890 sp = page_header(root);
3891 mmu_sync_children(vcpu, sp);
3895 kvm_mmu_audit(vcpu, AUDIT_POST_SYNC);
3896 spin_unlock(&vcpu->kvm->mmu_lock);
3898 EXPORT_SYMBOL_GPL(kvm_mmu_sync_roots);
3900 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gpa_t vaddr,
3901 u32 access, struct x86_exception *exception)
3904 exception->error_code = 0;
3908 static gpa_t nonpaging_gva_to_gpa_nested(struct kvm_vcpu *vcpu, gpa_t vaddr,
3910 struct x86_exception *exception)
3913 exception->error_code = 0;
3914 return vcpu->arch.nested_mmu.translate_gpa(vcpu, vaddr, access, exception);
3918 __is_rsvd_bits_set(struct rsvd_bits_validate *rsvd_check, u64 pte, int level)
3920 int bit7 = (pte >> 7) & 1;
3922 return pte & rsvd_check->rsvd_bits_mask[bit7][level-1];
3925 static bool __is_bad_mt_xwr(struct rsvd_bits_validate *rsvd_check, u64 pte)
3927 return rsvd_check->bad_mt_xwr & BIT_ULL(pte & 0x3f);
3930 static bool mmio_info_in_cache(struct kvm_vcpu *vcpu, u64 addr, bool direct)
3933 * A nested guest cannot use the MMIO cache if it is using nested
3934 * page tables, because cr2 is a nGPA while the cache stores GPAs.
3936 if (mmu_is_nested(vcpu))
3940 return vcpu_match_mmio_gpa(vcpu, addr);
3942 return vcpu_match_mmio_gva(vcpu, addr);
3945 /* return true if reserved bit is detected on spte. */
3947 walk_shadow_page_get_mmio_spte(struct kvm_vcpu *vcpu, u64 addr, u64 *sptep)
3949 struct kvm_shadow_walk_iterator iterator;
3950 u64 sptes[PT64_ROOT_MAX_LEVEL], spte = 0ull;
3951 struct rsvd_bits_validate *rsvd_check;
3953 bool reserved = false;
3955 rsvd_check = &vcpu->arch.mmu->shadow_zero_check;
3957 walk_shadow_page_lockless_begin(vcpu);
3959 for (shadow_walk_init(&iterator, vcpu, addr),
3960 leaf = root = iterator.level;
3961 shadow_walk_okay(&iterator);
3962 __shadow_walk_next(&iterator, spte)) {
3963 spte = mmu_spte_get_lockless(iterator.sptep);
3965 sptes[leaf - 1] = spte;
3968 if (!is_shadow_present_pte(spte))
3972 * Use a bitwise-OR instead of a logical-OR to aggregate the
3973 * reserved bit and EPT's invalid memtype/XWR checks to avoid
3974 * adding a Jcc in the loop.
3976 reserved |= __is_bad_mt_xwr(rsvd_check, spte) |
3977 __is_rsvd_bits_set(rsvd_check, spte, iterator.level);
3980 walk_shadow_page_lockless_end(vcpu);
3983 pr_err("%s: detect reserved bits on spte, addr 0x%llx, dump hierarchy:\n",
3985 while (root > leaf) {
3986 pr_err("------ spte 0x%llx level %d.\n",
3987 sptes[root - 1], root);
3996 static int handle_mmio_page_fault(struct kvm_vcpu *vcpu, u64 addr, bool direct)
4001 if (mmio_info_in_cache(vcpu, addr, direct))
4002 return RET_PF_EMULATE;
4004 reserved = walk_shadow_page_get_mmio_spte(vcpu, addr, &spte);
4005 if (WARN_ON(reserved))
4008 if (is_mmio_spte(spte)) {
4009 gfn_t gfn = get_mmio_spte_gfn(spte);
4010 unsigned int access = get_mmio_spte_access(spte);
4012 if (!check_mmio_spte(vcpu, spte))
4013 return RET_PF_INVALID;
4018 trace_handle_mmio_page_fault(addr, gfn, access);
4019 vcpu_cache_mmio_info(vcpu, addr, gfn, access);
4020 return RET_PF_EMULATE;
4024 * If the page table is zapped by other cpus, let CPU fault again on
4027 return RET_PF_RETRY;
4030 static bool page_fault_handle_page_track(struct kvm_vcpu *vcpu,
4031 u32 error_code, gfn_t gfn)
4033 if (unlikely(error_code & PFERR_RSVD_MASK))
4036 if (!(error_code & PFERR_PRESENT_MASK) ||
4037 !(error_code & PFERR_WRITE_MASK))
4041 * guest is writing the page which is write tracked which can
4042 * not be fixed by page fault handler.
4044 if (kvm_page_track_is_active(vcpu, gfn, KVM_PAGE_TRACK_WRITE))
4050 static void shadow_page_table_clear_flood(struct kvm_vcpu *vcpu, gva_t addr)
4052 struct kvm_shadow_walk_iterator iterator;
4055 walk_shadow_page_lockless_begin(vcpu);
4056 for_each_shadow_entry_lockless(vcpu, addr, iterator, spte) {
4057 clear_sp_write_flooding_count(iterator.sptep);
4058 if (!is_shadow_present_pte(spte))
4061 walk_shadow_page_lockless_end(vcpu);
4064 static int kvm_arch_setup_async_pf(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa,
4067 struct kvm_arch_async_pf arch;
4069 arch.token = (vcpu->arch.apf.id++ << 12) | vcpu->vcpu_id;
4071 arch.direct_map = vcpu->arch.mmu->direct_map;
4072 arch.cr3 = vcpu->arch.mmu->get_guest_pgd(vcpu);
4074 return kvm_setup_async_pf(vcpu, cr2_or_gpa,
4075 kvm_vcpu_gfn_to_hva(vcpu, gfn), &arch);
4078 static bool try_async_pf(struct kvm_vcpu *vcpu, bool prefault, gfn_t gfn,
4079 gpa_t cr2_or_gpa, kvm_pfn_t *pfn, bool write,
4082 struct kvm_memory_slot *slot;
4086 * Don't expose private memslots to L2.
4088 if (is_guest_mode(vcpu) && !kvm_is_visible_gfn(vcpu->kvm, gfn)) {
4089 *pfn = KVM_PFN_NOSLOT;
4093 slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
4095 *pfn = __gfn_to_pfn_memslot(slot, gfn, false, &async, write, writable);
4097 return false; /* *pfn has correct page already */
4099 if (!prefault && kvm_can_do_async_pf(vcpu)) {
4100 trace_kvm_try_async_get_page(cr2_or_gpa, gfn);
4101 if (kvm_find_async_pf_gfn(vcpu, gfn)) {
4102 trace_kvm_async_pf_doublefault(cr2_or_gpa, gfn);
4103 kvm_make_request(KVM_REQ_APF_HALT, vcpu);
4105 } else if (kvm_arch_setup_async_pf(vcpu, cr2_or_gpa, gfn))
4109 *pfn = __gfn_to_pfn_memslot(slot, gfn, false, NULL, write, writable);
4113 static int direct_page_fault(struct kvm_vcpu *vcpu, gpa_t gpa, u32 error_code,
4114 bool prefault, int max_level, bool is_tdp)
4116 bool write = error_code & PFERR_WRITE_MASK;
4117 bool exec = error_code & PFERR_FETCH_MASK;
4118 bool lpage_disallowed = exec && is_nx_huge_page_enabled();
4121 gfn_t gfn = gpa >> PAGE_SHIFT;
4122 unsigned long mmu_seq;
4126 if (page_fault_handle_page_track(vcpu, error_code, gfn))
4127 return RET_PF_EMULATE;
4129 r = mmu_topup_memory_caches(vcpu);
4133 if (lpage_disallowed)
4134 max_level = PT_PAGE_TABLE_LEVEL;
4136 if (fast_page_fault(vcpu, gpa, error_code))
4137 return RET_PF_RETRY;
4139 mmu_seq = vcpu->kvm->mmu_notifier_seq;
4142 if (try_async_pf(vcpu, prefault, gfn, gpa, &pfn, write, &map_writable))
4143 return RET_PF_RETRY;
4145 if (handle_abnormal_pfn(vcpu, is_tdp ? 0 : gpa, gfn, pfn, ACC_ALL, &r))
4149 spin_lock(&vcpu->kvm->mmu_lock);
4150 if (mmu_notifier_retry(vcpu->kvm, mmu_seq))
4152 if (make_mmu_pages_available(vcpu) < 0)
4154 r = __direct_map(vcpu, gpa, write, map_writable, max_level, pfn,
4155 prefault, is_tdp && lpage_disallowed);
4158 spin_unlock(&vcpu->kvm->mmu_lock);
4159 kvm_release_pfn_clean(pfn);
4163 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gpa_t gpa,
4164 u32 error_code, bool prefault)
4166 pgprintk("%s: gva %lx error %x\n", __func__, gpa, error_code);
4168 /* This path builds a PAE pagetable, we can map 2mb pages at maximum. */
4169 return direct_page_fault(vcpu, gpa & PAGE_MASK, error_code, prefault,
4170 PT_DIRECTORY_LEVEL, false);
4173 int kvm_handle_page_fault(struct kvm_vcpu *vcpu, u64 error_code,
4174 u64 fault_address, char *insn, int insn_len)
4178 #ifndef CONFIG_X86_64
4179 /* A 64-bit CR2 should be impossible on 32-bit KVM. */
4180 if (WARN_ON_ONCE(fault_address >> 32))
4184 vcpu->arch.l1tf_flush_l1d = true;
4185 switch (vcpu->arch.apf.host_apf_reason) {
4187 trace_kvm_page_fault(fault_address, error_code);
4189 if (kvm_event_needs_reinjection(vcpu))
4190 kvm_mmu_unprotect_page_virt(vcpu, fault_address);
4191 r = kvm_mmu_page_fault(vcpu, fault_address, error_code, insn,
4194 case KVM_PV_REASON_PAGE_NOT_PRESENT:
4195 vcpu->arch.apf.host_apf_reason = 0;
4196 local_irq_disable();
4197 kvm_async_pf_task_wait(fault_address, 0);
4200 case KVM_PV_REASON_PAGE_READY:
4201 vcpu->arch.apf.host_apf_reason = 0;
4202 local_irq_disable();
4203 kvm_async_pf_task_wake(fault_address);
4209 EXPORT_SYMBOL_GPL(kvm_handle_page_fault);
4211 int kvm_tdp_page_fault(struct kvm_vcpu *vcpu, gpa_t gpa, u32 error_code,
4216 for (max_level = PT_MAX_HUGEPAGE_LEVEL;
4217 max_level > PT_PAGE_TABLE_LEVEL;
4219 int page_num = KVM_PAGES_PER_HPAGE(max_level);
4220 gfn_t base = (gpa >> PAGE_SHIFT) & ~(page_num - 1);
4222 if (kvm_mtrr_check_gfn_range_consistency(vcpu, base, page_num))
4226 return direct_page_fault(vcpu, gpa, error_code, prefault,
4230 static void nonpaging_init_context(struct kvm_vcpu *vcpu,
4231 struct kvm_mmu *context)
4233 context->page_fault = nonpaging_page_fault;
4234 context->gva_to_gpa = nonpaging_gva_to_gpa;
4235 context->sync_page = nonpaging_sync_page;
4236 context->invlpg = NULL;
4237 context->update_pte = nonpaging_update_pte;
4238 context->root_level = 0;
4239 context->shadow_root_level = PT32E_ROOT_LEVEL;
4240 context->direct_map = true;
4241 context->nx = false;
4244 static inline bool is_root_usable(struct kvm_mmu_root_info *root, gpa_t cr3,
4245 union kvm_mmu_page_role role)
4247 return (role.direct || cr3 == root->cr3) &&
4248 VALID_PAGE(root->hpa) && page_header(root->hpa) &&
4249 role.word == page_header(root->hpa)->role.word;
4253 * Find out if a previously cached root matching the new CR3/role is available.
4254 * The current root is also inserted into the cache.
4255 * If a matching root was found, it is assigned to kvm_mmu->root_hpa and true is
4257 * Otherwise, the LRU root from the cache is assigned to kvm_mmu->root_hpa and
4258 * false is returned. This root should now be freed by the caller.
4260 static bool cached_root_available(struct kvm_vcpu *vcpu, gpa_t new_cr3,
4261 union kvm_mmu_page_role new_role)
4264 struct kvm_mmu_root_info root;
4265 struct kvm_mmu *mmu = vcpu->arch.mmu;
4267 root.cr3 = mmu->root_cr3;
4268 root.hpa = mmu->root_hpa;
4270 if (is_root_usable(&root, new_cr3, new_role))
4273 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) {
4274 swap(root, mmu->prev_roots[i]);
4276 if (is_root_usable(&root, new_cr3, new_role))
4280 mmu->root_hpa = root.hpa;
4281 mmu->root_cr3 = root.cr3;
4283 return i < KVM_MMU_NUM_PREV_ROOTS;
4286 static bool fast_cr3_switch(struct kvm_vcpu *vcpu, gpa_t new_cr3,
4287 union kvm_mmu_page_role new_role,
4288 bool skip_tlb_flush)
4290 struct kvm_mmu *mmu = vcpu->arch.mmu;
4293 * For now, limit the fast switch to 64-bit hosts+VMs in order to avoid
4294 * having to deal with PDPTEs. We may add support for 32-bit hosts/VMs
4295 * later if necessary.
4297 if (mmu->shadow_root_level >= PT64_ROOT_4LEVEL &&
4298 mmu->root_level >= PT64_ROOT_4LEVEL) {
4299 if (mmu_check_root(vcpu, new_cr3 >> PAGE_SHIFT))
4302 if (cached_root_available(vcpu, new_cr3, new_role)) {
4304 * It is possible that the cached previous root page is
4305 * obsolete because of a change in the MMU generation
4306 * number. However, changing the generation number is
4307 * accompanied by KVM_REQ_MMU_RELOAD, which will free
4308 * the root set here and allocate a new one.
4310 kvm_make_request(KVM_REQ_LOAD_MMU_PGD, vcpu);
4311 if (!skip_tlb_flush) {
4312 kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
4313 kvm_make_request(KVM_REQ_TLB_FLUSH, vcpu);
4317 * The last MMIO access's GVA and GPA are cached in the
4318 * VCPU. When switching to a new CR3, that GVA->GPA
4319 * mapping may no longer be valid. So clear any cached
4320 * MMIO info even when we don't need to sync the shadow
4323 vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY);
4325 __clear_sp_write_flooding_count(
4326 page_header(mmu->root_hpa));
4335 static void __kvm_mmu_new_cr3(struct kvm_vcpu *vcpu, gpa_t new_cr3,
4336 union kvm_mmu_page_role new_role,
4337 bool skip_tlb_flush)
4339 if (!fast_cr3_switch(vcpu, new_cr3, new_role, skip_tlb_flush))
4340 kvm_mmu_free_roots(vcpu, vcpu->arch.mmu,
4341 KVM_MMU_ROOT_CURRENT);
4344 void kvm_mmu_new_cr3(struct kvm_vcpu *vcpu, gpa_t new_cr3, bool skip_tlb_flush)
4346 __kvm_mmu_new_cr3(vcpu, new_cr3, kvm_mmu_calc_root_page_role(vcpu),
4349 EXPORT_SYMBOL_GPL(kvm_mmu_new_cr3);
4351 static unsigned long get_cr3(struct kvm_vcpu *vcpu)
4353 return kvm_read_cr3(vcpu);
4356 static bool sync_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, gfn_t gfn,
4357 unsigned int access, int *nr_present)
4359 if (unlikely(is_mmio_spte(*sptep))) {
4360 if (gfn != get_mmio_spte_gfn(*sptep)) {
4361 mmu_spte_clear_no_track(sptep);
4366 mark_mmio_spte(vcpu, sptep, gfn, access);
4373 static inline bool is_last_gpte(struct kvm_mmu *mmu,
4374 unsigned level, unsigned gpte)
4377 * The RHS has bit 7 set iff level < mmu->last_nonleaf_level.
4378 * If it is clear, there are no large pages at this level, so clear
4379 * PT_PAGE_SIZE_MASK in gpte if that is the case.
4381 gpte &= level - mmu->last_nonleaf_level;
4384 * PT_PAGE_TABLE_LEVEL always terminates. The RHS has bit 7 set
4385 * iff level <= PT_PAGE_TABLE_LEVEL, which for our purpose means
4386 * level == PT_PAGE_TABLE_LEVEL; set PT_PAGE_SIZE_MASK in gpte then.
4388 gpte |= level - PT_PAGE_TABLE_LEVEL - 1;
4390 return gpte & PT_PAGE_SIZE_MASK;
4393 #define PTTYPE_EPT 18 /* arbitrary */
4394 #define PTTYPE PTTYPE_EPT
4395 #include "paging_tmpl.h"
4399 #include "paging_tmpl.h"
4403 #include "paging_tmpl.h"
4407 __reset_rsvds_bits_mask(struct kvm_vcpu *vcpu,
4408 struct rsvd_bits_validate *rsvd_check,
4409 int maxphyaddr, int level, bool nx, bool gbpages,
4412 u64 exb_bit_rsvd = 0;
4413 u64 gbpages_bit_rsvd = 0;
4414 u64 nonleaf_bit8_rsvd = 0;
4416 rsvd_check->bad_mt_xwr = 0;
4419 exb_bit_rsvd = rsvd_bits(63, 63);
4421 gbpages_bit_rsvd = rsvd_bits(7, 7);
4424 * Non-leaf PML4Es and PDPEs reserve bit 8 (which would be the G bit for
4425 * leaf entries) on AMD CPUs only.
4428 nonleaf_bit8_rsvd = rsvd_bits(8, 8);
4431 case PT32_ROOT_LEVEL:
4432 /* no rsvd bits for 2 level 4K page table entries */
4433 rsvd_check->rsvd_bits_mask[0][1] = 0;
4434 rsvd_check->rsvd_bits_mask[0][0] = 0;
4435 rsvd_check->rsvd_bits_mask[1][0] =
4436 rsvd_check->rsvd_bits_mask[0][0];
4439 rsvd_check->rsvd_bits_mask[1][1] = 0;
4443 if (is_cpuid_PSE36())
4444 /* 36bits PSE 4MB page */
4445 rsvd_check->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
4447 /* 32 bits PSE 4MB page */
4448 rsvd_check->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
4450 case PT32E_ROOT_LEVEL:
4451 rsvd_check->rsvd_bits_mask[0][2] =
4452 rsvd_bits(maxphyaddr, 63) |
4453 rsvd_bits(5, 8) | rsvd_bits(1, 2); /* PDPTE */
4454 rsvd_check->rsvd_bits_mask[0][1] = exb_bit_rsvd |
4455 rsvd_bits(maxphyaddr, 62); /* PDE */
4456 rsvd_check->rsvd_bits_mask[0][0] = exb_bit_rsvd |
4457 rsvd_bits(maxphyaddr, 62); /* PTE */
4458 rsvd_check->rsvd_bits_mask[1][1] = exb_bit_rsvd |
4459 rsvd_bits(maxphyaddr, 62) |
4460 rsvd_bits(13, 20); /* large page */
4461 rsvd_check->rsvd_bits_mask[1][0] =
4462 rsvd_check->rsvd_bits_mask[0][0];
4464 case PT64_ROOT_5LEVEL:
4465 rsvd_check->rsvd_bits_mask[0][4] = exb_bit_rsvd |
4466 nonleaf_bit8_rsvd | rsvd_bits(7, 7) |
4467 rsvd_bits(maxphyaddr, 51);
4468 rsvd_check->rsvd_bits_mask[1][4] =
4469 rsvd_check->rsvd_bits_mask[0][4];
4471 case PT64_ROOT_4LEVEL:
4472 rsvd_check->rsvd_bits_mask[0][3] = exb_bit_rsvd |
4473 nonleaf_bit8_rsvd | rsvd_bits(7, 7) |
4474 rsvd_bits(maxphyaddr, 51);
4475 rsvd_check->rsvd_bits_mask[0][2] = exb_bit_rsvd |
4476 nonleaf_bit8_rsvd | gbpages_bit_rsvd |
4477 rsvd_bits(maxphyaddr, 51);
4478 rsvd_check->rsvd_bits_mask[0][1] = exb_bit_rsvd |
4479 rsvd_bits(maxphyaddr, 51);
4480 rsvd_check->rsvd_bits_mask[0][0] = exb_bit_rsvd |
4481 rsvd_bits(maxphyaddr, 51);
4482 rsvd_check->rsvd_bits_mask[1][3] =
4483 rsvd_check->rsvd_bits_mask[0][3];
4484 rsvd_check->rsvd_bits_mask[1][2] = exb_bit_rsvd |
4485 gbpages_bit_rsvd | rsvd_bits(maxphyaddr, 51) |
4487 rsvd_check->rsvd_bits_mask[1][1] = exb_bit_rsvd |
4488 rsvd_bits(maxphyaddr, 51) |
4489 rsvd_bits(13, 20); /* large page */
4490 rsvd_check->rsvd_bits_mask[1][0] =
4491 rsvd_check->rsvd_bits_mask[0][0];
4496 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu,
4497 struct kvm_mmu *context)
4499 __reset_rsvds_bits_mask(vcpu, &context->guest_rsvd_check,
4500 cpuid_maxphyaddr(vcpu), context->root_level,
4502 guest_cpuid_has(vcpu, X86_FEATURE_GBPAGES),
4504 guest_cpuid_is_amd_or_hygon(vcpu));
4508 __reset_rsvds_bits_mask_ept(struct rsvd_bits_validate *rsvd_check,
4509 int maxphyaddr, bool execonly)
4513 rsvd_check->rsvd_bits_mask[0][4] =
4514 rsvd_bits(maxphyaddr, 51) | rsvd_bits(3, 7);
4515 rsvd_check->rsvd_bits_mask[0][3] =
4516 rsvd_bits(maxphyaddr, 51) | rsvd_bits(3, 7);
4517 rsvd_check->rsvd_bits_mask[0][2] =
4518 rsvd_bits(maxphyaddr, 51) | rsvd_bits(3, 6);
4519 rsvd_check->rsvd_bits_mask[0][1] =
4520 rsvd_bits(maxphyaddr, 51) | rsvd_bits(3, 6);
4521 rsvd_check->rsvd_bits_mask[0][0] = rsvd_bits(maxphyaddr, 51);
4524 rsvd_check->rsvd_bits_mask[1][4] = rsvd_check->rsvd_bits_mask[0][4];
4525 rsvd_check->rsvd_bits_mask[1][3] = rsvd_check->rsvd_bits_mask[0][3];
4526 rsvd_check->rsvd_bits_mask[1][2] =
4527 rsvd_bits(maxphyaddr, 51) | rsvd_bits(12, 29);
4528 rsvd_check->rsvd_bits_mask[1][1] =
4529 rsvd_bits(maxphyaddr, 51) | rsvd_bits(12, 20);
4530 rsvd_check->rsvd_bits_mask[1][0] = rsvd_check->rsvd_bits_mask[0][0];
4532 bad_mt_xwr = 0xFFull << (2 * 8); /* bits 3..5 must not be 2 */
4533 bad_mt_xwr |= 0xFFull << (3 * 8); /* bits 3..5 must not be 3 */
4534 bad_mt_xwr |= 0xFFull << (7 * 8); /* bits 3..5 must not be 7 */
4535 bad_mt_xwr |= REPEAT_BYTE(1ull << 2); /* bits 0..2 must not be 010 */
4536 bad_mt_xwr |= REPEAT_BYTE(1ull << 6); /* bits 0..2 must not be 110 */
4538 /* bits 0..2 must not be 100 unless VMX capabilities allow it */
4539 bad_mt_xwr |= REPEAT_BYTE(1ull << 4);
4541 rsvd_check->bad_mt_xwr = bad_mt_xwr;
4544 static void reset_rsvds_bits_mask_ept(struct kvm_vcpu *vcpu,
4545 struct kvm_mmu *context, bool execonly)
4547 __reset_rsvds_bits_mask_ept(&context->guest_rsvd_check,
4548 cpuid_maxphyaddr(vcpu), execonly);
4552 * the page table on host is the shadow page table for the page
4553 * table in guest or amd nested guest, its mmu features completely
4554 * follow the features in guest.
4557 reset_shadow_zero_bits_mask(struct kvm_vcpu *vcpu, struct kvm_mmu *context)
4559 bool uses_nx = context->nx ||
4560 context->mmu_role.base.smep_andnot_wp;
4561 struct rsvd_bits_validate *shadow_zero_check;
4565 * Passing "true" to the last argument is okay; it adds a check
4566 * on bit 8 of the SPTEs which KVM doesn't use anyway.
4568 shadow_zero_check = &context->shadow_zero_check;
4569 __reset_rsvds_bits_mask(vcpu, shadow_zero_check,
4571 context->shadow_root_level, uses_nx,
4572 guest_cpuid_has(vcpu, X86_FEATURE_GBPAGES),
4573 is_pse(vcpu), true);
4575 if (!shadow_me_mask)
4578 for (i = context->shadow_root_level; --i >= 0;) {
4579 shadow_zero_check->rsvd_bits_mask[0][i] &= ~shadow_me_mask;
4580 shadow_zero_check->rsvd_bits_mask[1][i] &= ~shadow_me_mask;
4584 EXPORT_SYMBOL_GPL(reset_shadow_zero_bits_mask);
4586 static inline bool boot_cpu_is_amd(void)
4588 WARN_ON_ONCE(!tdp_enabled);
4589 return shadow_x_mask == 0;
4593 * the direct page table on host, use as much mmu features as
4594 * possible, however, kvm currently does not do execution-protection.
4597 reset_tdp_shadow_zero_bits_mask(struct kvm_vcpu *vcpu,
4598 struct kvm_mmu *context)
4600 struct rsvd_bits_validate *shadow_zero_check;
4603 shadow_zero_check = &context->shadow_zero_check;
4605 if (boot_cpu_is_amd())
4606 __reset_rsvds_bits_mask(vcpu, shadow_zero_check,
4608 context->shadow_root_level, false,
4609 boot_cpu_has(X86_FEATURE_GBPAGES),
4612 __reset_rsvds_bits_mask_ept(shadow_zero_check,
4616 if (!shadow_me_mask)
4619 for (i = context->shadow_root_level; --i >= 0;) {
4620 shadow_zero_check->rsvd_bits_mask[0][i] &= ~shadow_me_mask;
4621 shadow_zero_check->rsvd_bits_mask[1][i] &= ~shadow_me_mask;
4626 * as the comments in reset_shadow_zero_bits_mask() except it
4627 * is the shadow page table for intel nested guest.
4630 reset_ept_shadow_zero_bits_mask(struct kvm_vcpu *vcpu,
4631 struct kvm_mmu *context, bool execonly)
4633 __reset_rsvds_bits_mask_ept(&context->shadow_zero_check,
4634 shadow_phys_bits, execonly);
4637 #define BYTE_MASK(access) \
4638 ((1 & (access) ? 2 : 0) | \
4639 (2 & (access) ? 4 : 0) | \
4640 (3 & (access) ? 8 : 0) | \
4641 (4 & (access) ? 16 : 0) | \
4642 (5 & (access) ? 32 : 0) | \
4643 (6 & (access) ? 64 : 0) | \
4644 (7 & (access) ? 128 : 0))
4647 static void update_permission_bitmask(struct kvm_vcpu *vcpu,
4648 struct kvm_mmu *mmu, bool ept)
4652 const u8 x = BYTE_MASK(ACC_EXEC_MASK);
4653 const u8 w = BYTE_MASK(ACC_WRITE_MASK);
4654 const u8 u = BYTE_MASK(ACC_USER_MASK);
4656 bool cr4_smep = kvm_read_cr4_bits(vcpu, X86_CR4_SMEP) != 0;
4657 bool cr4_smap = kvm_read_cr4_bits(vcpu, X86_CR4_SMAP) != 0;
4658 bool cr0_wp = is_write_protection(vcpu);
4660 for (byte = 0; byte < ARRAY_SIZE(mmu->permissions); ++byte) {
4661 unsigned pfec = byte << 1;
4664 * Each "*f" variable has a 1 bit for each UWX value
4665 * that causes a fault with the given PFEC.
4668 /* Faults from writes to non-writable pages */
4669 u8 wf = (pfec & PFERR_WRITE_MASK) ? (u8)~w : 0;
4670 /* Faults from user mode accesses to supervisor pages */
4671 u8 uf = (pfec & PFERR_USER_MASK) ? (u8)~u : 0;
4672 /* Faults from fetches of non-executable pages*/
4673 u8 ff = (pfec & PFERR_FETCH_MASK) ? (u8)~x : 0;
4674 /* Faults from kernel mode fetches of user pages */
4676 /* Faults from kernel mode accesses of user pages */
4680 /* Faults from kernel mode accesses to user pages */
4681 u8 kf = (pfec & PFERR_USER_MASK) ? 0 : u;
4683 /* Not really needed: !nx will cause pte.nx to fault */
4687 /* Allow supervisor writes if !cr0.wp */
4689 wf = (pfec & PFERR_USER_MASK) ? wf : 0;
4691 /* Disallow supervisor fetches of user code if cr4.smep */
4693 smepf = (pfec & PFERR_FETCH_MASK) ? kf : 0;
4696 * SMAP:kernel-mode data accesses from user-mode
4697 * mappings should fault. A fault is considered
4698 * as a SMAP violation if all of the following
4699 * conditions are true:
4700 * - X86_CR4_SMAP is set in CR4
4701 * - A user page is accessed
4702 * - The access is not a fetch
4703 * - Page fault in kernel mode
4704 * - if CPL = 3 or X86_EFLAGS_AC is clear
4706 * Here, we cover the first three conditions.
4707 * The fourth is computed dynamically in permission_fault();
4708 * PFERR_RSVD_MASK bit will be set in PFEC if the access is
4709 * *not* subject to SMAP restrictions.
4712 smapf = (pfec & (PFERR_RSVD_MASK|PFERR_FETCH_MASK)) ? 0 : kf;
4715 mmu->permissions[byte] = ff | uf | wf | smepf | smapf;
4720 * PKU is an additional mechanism by which the paging controls access to
4721 * user-mode addresses based on the value in the PKRU register. Protection
4722 * key violations are reported through a bit in the page fault error code.
4723 * Unlike other bits of the error code, the PK bit is not known at the
4724 * call site of e.g. gva_to_gpa; it must be computed directly in
4725 * permission_fault based on two bits of PKRU, on some machine state (CR4,
4726 * CR0, EFER, CPL), and on other bits of the error code and the page tables.
4728 * In particular the following conditions come from the error code, the
4729 * page tables and the machine state:
4730 * - PK is always zero unless CR4.PKE=1 and EFER.LMA=1
4731 * - PK is always zero if RSVD=1 (reserved bit set) or F=1 (instruction fetch)
4732 * - PK is always zero if U=0 in the page tables
4733 * - PKRU.WD is ignored if CR0.WP=0 and the access is a supervisor access.
4735 * The PKRU bitmask caches the result of these four conditions. The error
4736 * code (minus the P bit) and the page table's U bit form an index into the
4737 * PKRU bitmask. Two bits of the PKRU bitmask are then extracted and ANDed
4738 * with the two bits of the PKRU register corresponding to the protection key.
4739 * For the first three conditions above the bits will be 00, thus masking
4740 * away both AD and WD. For all reads or if the last condition holds, WD
4741 * only will be masked away.
4743 static void update_pkru_bitmask(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
4754 /* PKEY is enabled only if CR4.PKE and EFER.LMA are both set. */
4755 if (!kvm_read_cr4_bits(vcpu, X86_CR4_PKE) || !is_long_mode(vcpu)) {
4760 wp = is_write_protection(vcpu);
4762 for (bit = 0; bit < ARRAY_SIZE(mmu->permissions); ++bit) {
4763 unsigned pfec, pkey_bits;
4764 bool check_pkey, check_write, ff, uf, wf, pte_user;
4767 ff = pfec & PFERR_FETCH_MASK;
4768 uf = pfec & PFERR_USER_MASK;
4769 wf = pfec & PFERR_WRITE_MASK;
4771 /* PFEC.RSVD is replaced by ACC_USER_MASK. */
4772 pte_user = pfec & PFERR_RSVD_MASK;
4775 * Only need to check the access which is not an
4776 * instruction fetch and is to a user page.
4778 check_pkey = (!ff && pte_user);
4780 * write access is controlled by PKRU if it is a
4781 * user access or CR0.WP = 1.
4783 check_write = check_pkey && wf && (uf || wp);
4785 /* PKRU.AD stops both read and write access. */
4786 pkey_bits = !!check_pkey;
4787 /* PKRU.WD stops write access. */
4788 pkey_bits |= (!!check_write) << 1;
4790 mmu->pkru_mask |= (pkey_bits & 3) << pfec;
4794 static void update_last_nonleaf_level(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu)
4796 unsigned root_level = mmu->root_level;
4798 mmu->last_nonleaf_level = root_level;
4799 if (root_level == PT32_ROOT_LEVEL && is_pse(vcpu))
4800 mmu->last_nonleaf_level++;
4803 static void paging64_init_context_common(struct kvm_vcpu *vcpu,
4804 struct kvm_mmu *context,
4807 context->nx = is_nx(vcpu);
4808 context->root_level = level;
4810 reset_rsvds_bits_mask(vcpu, context);
4811 update_permission_bitmask(vcpu, context, false);
4812 update_pkru_bitmask(vcpu, context, false);
4813 update_last_nonleaf_level(vcpu, context);
4815 MMU_WARN_ON(!is_pae(vcpu));
4816 context->page_fault = paging64_page_fault;
4817 context->gva_to_gpa = paging64_gva_to_gpa;
4818 context->sync_page = paging64_sync_page;
4819 context->invlpg = paging64_invlpg;
4820 context->update_pte = paging64_update_pte;
4821 context->shadow_root_level = level;
4822 context->direct_map = false;
4825 static void paging64_init_context(struct kvm_vcpu *vcpu,
4826 struct kvm_mmu *context)
4828 int root_level = is_la57_mode(vcpu) ?
4829 PT64_ROOT_5LEVEL : PT64_ROOT_4LEVEL;
4831 paging64_init_context_common(vcpu, context, root_level);
4834 static void paging32_init_context(struct kvm_vcpu *vcpu,
4835 struct kvm_mmu *context)
4837 context->nx = false;
4838 context->root_level = PT32_ROOT_LEVEL;
4840 reset_rsvds_bits_mask(vcpu, context);
4841 update_permission_bitmask(vcpu, context, false);
4842 update_pkru_bitmask(vcpu, context, false);
4843 update_last_nonleaf_level(vcpu, context);
4845 context->page_fault = paging32_page_fault;
4846 context->gva_to_gpa = paging32_gva_to_gpa;
4847 context->sync_page = paging32_sync_page;
4848 context->invlpg = paging32_invlpg;
4849 context->update_pte = paging32_update_pte;
4850 context->shadow_root_level = PT32E_ROOT_LEVEL;
4851 context->direct_map = false;
4854 static void paging32E_init_context(struct kvm_vcpu *vcpu,
4855 struct kvm_mmu *context)
4857 paging64_init_context_common(vcpu, context, PT32E_ROOT_LEVEL);
4860 static union kvm_mmu_extended_role kvm_calc_mmu_role_ext(struct kvm_vcpu *vcpu)
4862 union kvm_mmu_extended_role ext = {0};
4864 ext.cr0_pg = !!is_paging(vcpu);
4865 ext.cr4_pae = !!is_pae(vcpu);
4866 ext.cr4_smep = !!kvm_read_cr4_bits(vcpu, X86_CR4_SMEP);
4867 ext.cr4_smap = !!kvm_read_cr4_bits(vcpu, X86_CR4_SMAP);
4868 ext.cr4_pse = !!is_pse(vcpu);
4869 ext.cr4_pke = !!kvm_read_cr4_bits(vcpu, X86_CR4_PKE);
4870 ext.maxphyaddr = cpuid_maxphyaddr(vcpu);
4877 static union kvm_mmu_role kvm_calc_mmu_role_common(struct kvm_vcpu *vcpu,
4880 union kvm_mmu_role role = {0};
4882 role.base.access = ACC_ALL;
4883 role.base.nxe = !!is_nx(vcpu);
4884 role.base.cr0_wp = is_write_protection(vcpu);
4885 role.base.smm = is_smm(vcpu);
4886 role.base.guest_mode = is_guest_mode(vcpu);
4891 role.ext = kvm_calc_mmu_role_ext(vcpu);
4896 static union kvm_mmu_role
4897 kvm_calc_tdp_mmu_root_page_role(struct kvm_vcpu *vcpu, bool base_only)
4899 union kvm_mmu_role role = kvm_calc_mmu_role_common(vcpu, base_only);
4901 role.base.ad_disabled = (shadow_accessed_mask == 0);
4902 role.base.level = kvm_x86_ops.get_tdp_level(vcpu);
4903 role.base.direct = true;
4904 role.base.gpte_is_8_bytes = true;
4909 static void init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
4911 struct kvm_mmu *context = vcpu->arch.mmu;
4912 union kvm_mmu_role new_role =
4913 kvm_calc_tdp_mmu_root_page_role(vcpu, false);
4915 if (new_role.as_u64 == context->mmu_role.as_u64)
4918 context->mmu_role.as_u64 = new_role.as_u64;
4919 context->page_fault = kvm_tdp_page_fault;
4920 context->sync_page = nonpaging_sync_page;
4921 context->invlpg = NULL;
4922 context->update_pte = nonpaging_update_pte;
4923 context->shadow_root_level = kvm_x86_ops.get_tdp_level(vcpu);
4924 context->direct_map = true;
4925 context->get_guest_pgd = get_cr3;
4926 context->get_pdptr = kvm_pdptr_read;
4927 context->inject_page_fault = kvm_inject_page_fault;
4929 if (!is_paging(vcpu)) {
4930 context->nx = false;
4931 context->gva_to_gpa = nonpaging_gva_to_gpa;
4932 context->root_level = 0;
4933 } else if (is_long_mode(vcpu)) {
4934 context->nx = is_nx(vcpu);
4935 context->root_level = is_la57_mode(vcpu) ?
4936 PT64_ROOT_5LEVEL : PT64_ROOT_4LEVEL;
4937 reset_rsvds_bits_mask(vcpu, context);
4938 context->gva_to_gpa = paging64_gva_to_gpa;
4939 } else if (is_pae(vcpu)) {
4940 context->nx = is_nx(vcpu);
4941 context->root_level = PT32E_ROOT_LEVEL;
4942 reset_rsvds_bits_mask(vcpu, context);
4943 context->gva_to_gpa = paging64_gva_to_gpa;
4945 context->nx = false;
4946 context->root_level = PT32_ROOT_LEVEL;
4947 reset_rsvds_bits_mask(vcpu, context);
4948 context->gva_to_gpa = paging32_gva_to_gpa;
4951 update_permission_bitmask(vcpu, context, false);
4952 update_pkru_bitmask(vcpu, context, false);
4953 update_last_nonleaf_level(vcpu, context);
4954 reset_tdp_shadow_zero_bits_mask(vcpu, context);
4957 static union kvm_mmu_role
4958 kvm_calc_shadow_mmu_root_page_role(struct kvm_vcpu *vcpu, bool base_only)
4960 union kvm_mmu_role role = kvm_calc_mmu_role_common(vcpu, base_only);
4962 role.base.smep_andnot_wp = role.ext.cr4_smep &&
4963 !is_write_protection(vcpu);
4964 role.base.smap_andnot_wp = role.ext.cr4_smap &&
4965 !is_write_protection(vcpu);
4966 role.base.direct = !is_paging(vcpu);
4967 role.base.gpte_is_8_bytes = !!is_pae(vcpu);
4969 if (!is_long_mode(vcpu))
4970 role.base.level = PT32E_ROOT_LEVEL;
4971 else if (is_la57_mode(vcpu))
4972 role.base.level = PT64_ROOT_5LEVEL;
4974 role.base.level = PT64_ROOT_4LEVEL;
4979 void kvm_init_shadow_mmu(struct kvm_vcpu *vcpu)
4981 struct kvm_mmu *context = vcpu->arch.mmu;
4982 union kvm_mmu_role new_role =
4983 kvm_calc_shadow_mmu_root_page_role(vcpu, false);
4985 if (new_role.as_u64 == context->mmu_role.as_u64)
4988 if (!is_paging(vcpu))
4989 nonpaging_init_context(vcpu, context);
4990 else if (is_long_mode(vcpu))
4991 paging64_init_context(vcpu, context);
4992 else if (is_pae(vcpu))
4993 paging32E_init_context(vcpu, context);
4995 paging32_init_context(vcpu, context);
4997 context->mmu_role.as_u64 = new_role.as_u64;
4998 reset_shadow_zero_bits_mask(vcpu, context);
5000 EXPORT_SYMBOL_GPL(kvm_init_shadow_mmu);
5002 static union kvm_mmu_role
5003 kvm_calc_shadow_ept_root_page_role(struct kvm_vcpu *vcpu, bool accessed_dirty,
5004 bool execonly, u8 level)
5006 union kvm_mmu_role role = {0};
5008 /* SMM flag is inherited from root_mmu */
5009 role.base.smm = vcpu->arch.root_mmu.mmu_role.base.smm;
5011 role.base.level = level;
5012 role.base.gpte_is_8_bytes = true;
5013 role.base.direct = false;
5014 role.base.ad_disabled = !accessed_dirty;
5015 role.base.guest_mode = true;
5016 role.base.access = ACC_ALL;
5019 * WP=1 and NOT_WP=1 is an impossible combination, use WP and the
5020 * SMAP variation to denote shadow EPT entries.
5022 role.base.cr0_wp = true;
5023 role.base.smap_andnot_wp = true;
5025 role.ext = kvm_calc_mmu_role_ext(vcpu);
5026 role.ext.execonly = execonly;
5031 void kvm_init_shadow_ept_mmu(struct kvm_vcpu *vcpu, bool execonly,
5032 bool accessed_dirty, gpa_t new_eptp)
5034 struct kvm_mmu *context = vcpu->arch.mmu;
5035 u8 level = vmx_eptp_page_walk_level(new_eptp);
5036 union kvm_mmu_role new_role =
5037 kvm_calc_shadow_ept_root_page_role(vcpu, accessed_dirty,
5040 __kvm_mmu_new_cr3(vcpu, new_eptp, new_role.base, false);
5042 if (new_role.as_u64 == context->mmu_role.as_u64)
5045 context->shadow_root_level = level;
5048 context->ept_ad = accessed_dirty;
5049 context->page_fault = ept_page_fault;
5050 context->gva_to_gpa = ept_gva_to_gpa;
5051 context->sync_page = ept_sync_page;
5052 context->invlpg = ept_invlpg;
5053 context->update_pte = ept_update_pte;
5054 context->root_level = level;
5055 context->direct_map = false;
5056 context->mmu_role.as_u64 = new_role.as_u64;
5058 update_permission_bitmask(vcpu, context, true);
5059 update_pkru_bitmask(vcpu, context, true);
5060 update_last_nonleaf_level(vcpu, context);
5061 reset_rsvds_bits_mask_ept(vcpu, context, execonly);
5062 reset_ept_shadow_zero_bits_mask(vcpu, context, execonly);
5064 EXPORT_SYMBOL_GPL(kvm_init_shadow_ept_mmu);
5066 static void init_kvm_softmmu(struct kvm_vcpu *vcpu)
5068 struct kvm_mmu *context = vcpu->arch.mmu;
5070 kvm_init_shadow_mmu(vcpu);
5071 context->get_guest_pgd = get_cr3;
5072 context->get_pdptr = kvm_pdptr_read;
5073 context->inject_page_fault = kvm_inject_page_fault;
5076 static void init_kvm_nested_mmu(struct kvm_vcpu *vcpu)
5078 union kvm_mmu_role new_role = kvm_calc_mmu_role_common(vcpu, false);
5079 struct kvm_mmu *g_context = &vcpu->arch.nested_mmu;
5081 if (new_role.as_u64 == g_context->mmu_role.as_u64)
5084 g_context->mmu_role.as_u64 = new_role.as_u64;
5085 g_context->get_guest_pgd = get_cr3;
5086 g_context->get_pdptr = kvm_pdptr_read;
5087 g_context->inject_page_fault = kvm_inject_page_fault;
5090 * L2 page tables are never shadowed, so there is no need to sync
5093 g_context->invlpg = NULL;
5096 * Note that arch.mmu->gva_to_gpa translates l2_gpa to l1_gpa using
5097 * L1's nested page tables (e.g. EPT12). The nested translation
5098 * of l2_gva to l1_gpa is done by arch.nested_mmu.gva_to_gpa using
5099 * L2's page tables as the first level of translation and L1's
5100 * nested page tables as the second level of translation. Basically
5101 * the gva_to_gpa functions between mmu and nested_mmu are swapped.
5103 if (!is_paging(vcpu)) {
5104 g_context->nx = false;
5105 g_context->root_level = 0;
5106 g_context->gva_to_gpa = nonpaging_gva_to_gpa_nested;
5107 } else if (is_long_mode(vcpu)) {
5108 g_context->nx = is_nx(vcpu);
5109 g_context->root_level = is_la57_mode(vcpu) ?
5110 PT64_ROOT_5LEVEL : PT64_ROOT_4LEVEL;
5111 reset_rsvds_bits_mask(vcpu, g_context);
5112 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
5113 } else if (is_pae(vcpu)) {
5114 g_context->nx = is_nx(vcpu);
5115 g_context->root_level = PT32E_ROOT_LEVEL;
5116 reset_rsvds_bits_mask(vcpu, g_context);
5117 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
5119 g_context->nx = false;
5120 g_context->root_level = PT32_ROOT_LEVEL;
5121 reset_rsvds_bits_mask(vcpu, g_context);
5122 g_context->gva_to_gpa = paging32_gva_to_gpa_nested;
5125 update_permission_bitmask(vcpu, g_context, false);
5126 update_pkru_bitmask(vcpu, g_context, false);
5127 update_last_nonleaf_level(vcpu, g_context);
5130 void kvm_init_mmu(struct kvm_vcpu *vcpu, bool reset_roots)
5135 vcpu->arch.mmu->root_hpa = INVALID_PAGE;
5137 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
5138 vcpu->arch.mmu->prev_roots[i] = KVM_MMU_ROOT_INFO_INVALID;
5141 if (mmu_is_nested(vcpu))
5142 init_kvm_nested_mmu(vcpu);
5143 else if (tdp_enabled)
5144 init_kvm_tdp_mmu(vcpu);
5146 init_kvm_softmmu(vcpu);
5148 EXPORT_SYMBOL_GPL(kvm_init_mmu);
5150 static union kvm_mmu_page_role
5151 kvm_mmu_calc_root_page_role(struct kvm_vcpu *vcpu)
5153 union kvm_mmu_role role;
5156 role = kvm_calc_tdp_mmu_root_page_role(vcpu, true);
5158 role = kvm_calc_shadow_mmu_root_page_role(vcpu, true);
5163 void kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
5165 kvm_mmu_unload(vcpu);
5166 kvm_init_mmu(vcpu, true);
5168 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
5170 int kvm_mmu_load(struct kvm_vcpu *vcpu)
5174 r = mmu_topup_memory_caches(vcpu);
5177 r = mmu_alloc_roots(vcpu);
5178 kvm_mmu_sync_roots(vcpu);
5181 kvm_mmu_load_pgd(vcpu);
5182 kvm_x86_ops.tlb_flush(vcpu);
5186 EXPORT_SYMBOL_GPL(kvm_mmu_load);
5188 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
5190 kvm_mmu_free_roots(vcpu, &vcpu->arch.root_mmu, KVM_MMU_ROOTS_ALL);
5191 WARN_ON(VALID_PAGE(vcpu->arch.root_mmu.root_hpa));
5192 kvm_mmu_free_roots(vcpu, &vcpu->arch.guest_mmu, KVM_MMU_ROOTS_ALL);
5193 WARN_ON(VALID_PAGE(vcpu->arch.guest_mmu.root_hpa));
5195 EXPORT_SYMBOL_GPL(kvm_mmu_unload);
5197 static void mmu_pte_write_new_pte(struct kvm_vcpu *vcpu,
5198 struct kvm_mmu_page *sp, u64 *spte,
5201 if (sp->role.level != PT_PAGE_TABLE_LEVEL) {
5202 ++vcpu->kvm->stat.mmu_pde_zapped;
5206 ++vcpu->kvm->stat.mmu_pte_updated;
5207 vcpu->arch.mmu->update_pte(vcpu, sp, spte, new);
5210 static bool need_remote_flush(u64 old, u64 new)
5212 if (!is_shadow_present_pte(old))
5214 if (!is_shadow_present_pte(new))
5216 if ((old ^ new) & PT64_BASE_ADDR_MASK)
5218 old ^= shadow_nx_mask;
5219 new ^= shadow_nx_mask;
5220 return (old & ~new & PT64_PERM_MASK) != 0;
5223 static u64 mmu_pte_write_fetch_gpte(struct kvm_vcpu *vcpu, gpa_t *gpa,
5230 * Assume that the pte write on a page table of the same type
5231 * as the current vcpu paging mode since we update the sptes only
5232 * when they have the same mode.
5234 if (is_pae(vcpu) && *bytes == 4) {
5235 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
5240 if (*bytes == 4 || *bytes == 8) {
5241 r = kvm_vcpu_read_guest_atomic(vcpu, *gpa, &gentry, *bytes);
5250 * If we're seeing too many writes to a page, it may no longer be a page table,
5251 * or we may be forking, in which case it is better to unmap the page.
5253 static bool detect_write_flooding(struct kvm_mmu_page *sp)
5256 * Skip write-flooding detected for the sp whose level is 1, because
5257 * it can become unsync, then the guest page is not write-protected.
5259 if (sp->role.level == PT_PAGE_TABLE_LEVEL)
5262 atomic_inc(&sp->write_flooding_count);
5263 return atomic_read(&sp->write_flooding_count) >= 3;
5267 * Misaligned accesses are too much trouble to fix up; also, they usually
5268 * indicate a page is not used as a page table.
5270 static bool detect_write_misaligned(struct kvm_mmu_page *sp, gpa_t gpa,
5273 unsigned offset, pte_size, misaligned;
5275 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
5276 gpa, bytes, sp->role.word);
5278 offset = offset_in_page(gpa);
5279 pte_size = sp->role.gpte_is_8_bytes ? 8 : 4;
5282 * Sometimes, the OS only writes the last one bytes to update status
5283 * bits, for example, in linux, andb instruction is used in clear_bit().
5285 if (!(offset & (pte_size - 1)) && bytes == 1)
5288 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
5289 misaligned |= bytes < 4;
5294 static u64 *get_written_sptes(struct kvm_mmu_page *sp, gpa_t gpa, int *nspte)
5296 unsigned page_offset, quadrant;
5300 page_offset = offset_in_page(gpa);
5301 level = sp->role.level;
5303 if (!sp->role.gpte_is_8_bytes) {
5304 page_offset <<= 1; /* 32->64 */
5306 * A 32-bit pde maps 4MB while the shadow pdes map
5307 * only 2MB. So we need to double the offset again
5308 * and zap two pdes instead of one.
5310 if (level == PT32_ROOT_LEVEL) {
5311 page_offset &= ~7; /* kill rounding error */
5315 quadrant = page_offset >> PAGE_SHIFT;
5316 page_offset &= ~PAGE_MASK;
5317 if (quadrant != sp->role.quadrant)
5321 spte = &sp->spt[page_offset / sizeof(*spte)];
5326 * Ignore various flags when determining if a SPTE can be immediately
5327 * overwritten for the current MMU.
5328 * - level: explicitly checked in mmu_pte_write_new_pte(), and will never
5329 * match the current MMU role, as MMU's level tracks the root level.
5330 * - access: updated based on the new guest PTE
5331 * - quadrant: handled by get_written_sptes()
5332 * - invalid: always false (loop only walks valid shadow pages)
5334 static const union kvm_mmu_page_role role_ign = {
5341 static void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
5342 const u8 *new, int bytes,
5343 struct kvm_page_track_notifier_node *node)
5345 gfn_t gfn = gpa >> PAGE_SHIFT;
5346 struct kvm_mmu_page *sp;
5347 LIST_HEAD(invalid_list);
5348 u64 entry, gentry, *spte;
5350 bool remote_flush, local_flush;
5353 * If we don't have indirect shadow pages, it means no page is
5354 * write-protected, so we can exit simply.
5356 if (!READ_ONCE(vcpu->kvm->arch.indirect_shadow_pages))
5359 remote_flush = local_flush = false;
5361 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
5364 * No need to care whether allocation memory is successful
5365 * or not since pte prefetch is skiped if it does not have
5366 * enough objects in the cache.
5368 mmu_topup_memory_caches(vcpu);
5370 spin_lock(&vcpu->kvm->mmu_lock);
5372 gentry = mmu_pte_write_fetch_gpte(vcpu, &gpa, &bytes);
5374 ++vcpu->kvm->stat.mmu_pte_write;
5375 kvm_mmu_audit(vcpu, AUDIT_PRE_PTE_WRITE);
5377 for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn) {
5378 if (detect_write_misaligned(sp, gpa, bytes) ||
5379 detect_write_flooding(sp)) {
5380 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
5381 ++vcpu->kvm->stat.mmu_flooded;
5385 spte = get_written_sptes(sp, gpa, &npte);
5391 u32 base_role = vcpu->arch.mmu->mmu_role.base.word;
5394 mmu_page_zap_pte(vcpu->kvm, sp, spte);
5396 !((sp->role.word ^ base_role) & ~role_ign.word) &&
5398 mmu_pte_write_new_pte(vcpu, sp, spte, &gentry);
5399 if (need_remote_flush(entry, *spte))
5400 remote_flush = true;
5404 kvm_mmu_flush_or_zap(vcpu, &invalid_list, remote_flush, local_flush);
5405 kvm_mmu_audit(vcpu, AUDIT_POST_PTE_WRITE);
5406 spin_unlock(&vcpu->kvm->mmu_lock);
5409 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
5414 if (vcpu->arch.mmu->direct_map)
5417 gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
5419 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
5423 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt);
5425 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa, u64 error_code,
5426 void *insn, int insn_len)
5428 int r, emulation_type = EMULTYPE_PF;
5429 bool direct = vcpu->arch.mmu->direct_map;
5431 if (WARN_ON(!VALID_PAGE(vcpu->arch.mmu->root_hpa)))
5432 return RET_PF_RETRY;
5435 if (unlikely(error_code & PFERR_RSVD_MASK)) {
5436 r = handle_mmio_page_fault(vcpu, cr2_or_gpa, direct);
5437 if (r == RET_PF_EMULATE)
5441 if (r == RET_PF_INVALID) {
5442 r = kvm_mmu_do_page_fault(vcpu, cr2_or_gpa,
5443 lower_32_bits(error_code), false);
5444 WARN_ON(r == RET_PF_INVALID);
5447 if (r == RET_PF_RETRY)
5453 * Before emulating the instruction, check if the error code
5454 * was due to a RO violation while translating the guest page.
5455 * This can occur when using nested virtualization with nested
5456 * paging in both guests. If true, we simply unprotect the page
5457 * and resume the guest.
5459 if (vcpu->arch.mmu->direct_map &&
5460 (error_code & PFERR_NESTED_GUEST_PAGE) == PFERR_NESTED_GUEST_PAGE) {
5461 kvm_mmu_unprotect_page(vcpu->kvm, gpa_to_gfn(cr2_or_gpa));
5466 * vcpu->arch.mmu.page_fault returned RET_PF_EMULATE, but we can still
5467 * optimistically try to just unprotect the page and let the processor
5468 * re-execute the instruction that caused the page fault. Do not allow
5469 * retrying MMIO emulation, as it's not only pointless but could also
5470 * cause us to enter an infinite loop because the processor will keep
5471 * faulting on the non-existent MMIO address. Retrying an instruction
5472 * from a nested guest is also pointless and dangerous as we are only
5473 * explicitly shadowing L1's page tables, i.e. unprotecting something
5474 * for L1 isn't going to magically fix whatever issue cause L2 to fail.
5476 if (!mmio_info_in_cache(vcpu, cr2_or_gpa, direct) && !is_guest_mode(vcpu))
5477 emulation_type |= EMULTYPE_ALLOW_RETRY_PF;
5480 * On AMD platforms, under certain conditions insn_len may be zero on #NPF.
5481 * This can happen if a guest gets a page-fault on data access but the HW
5482 * table walker is not able to read the instruction page (e.g instruction
5483 * page is not present in memory). In those cases we simply restart the
5484 * guest, with the exception of AMD Erratum 1096 which is unrecoverable.
5486 if (unlikely(insn && !insn_len)) {
5487 if (!kvm_x86_ops.need_emulation_on_page_fault(vcpu))
5491 return x86_emulate_instruction(vcpu, cr2_or_gpa, emulation_type, insn,
5494 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
5496 void kvm_mmu_invalidate_gva(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
5497 gva_t gva, hpa_t root_hpa)
5501 /* It's actually a GPA for vcpu->arch.guest_mmu. */
5502 if (mmu != &vcpu->arch.guest_mmu) {
5503 /* INVLPG on a non-canonical address is a NOP according to the SDM. */
5504 if (is_noncanonical_address(gva, vcpu))
5507 kvm_x86_ops.tlb_flush_gva(vcpu, gva);
5513 if (root_hpa == INVALID_PAGE) {
5514 mmu->invlpg(vcpu, gva, mmu->root_hpa);
5517 * INVLPG is required to invalidate any global mappings for the VA,
5518 * irrespective of PCID. Since it would take us roughly similar amount
5519 * of work to determine whether any of the prev_root mappings of the VA
5520 * is marked global, or to just sync it blindly, so we might as well
5521 * just always sync it.
5523 * Mappings not reachable via the current cr3 or the prev_roots will be
5524 * synced when switching to that cr3, so nothing needs to be done here
5527 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
5528 if (VALID_PAGE(mmu->prev_roots[i].hpa))
5529 mmu->invlpg(vcpu, gva, mmu->prev_roots[i].hpa);
5531 mmu->invlpg(vcpu, gva, root_hpa);
5534 EXPORT_SYMBOL_GPL(kvm_mmu_invalidate_gva);
5536 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
5538 kvm_mmu_invalidate_gva(vcpu, vcpu->arch.mmu, gva, INVALID_PAGE);
5539 ++vcpu->stat.invlpg;
5541 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
5544 void kvm_mmu_invpcid_gva(struct kvm_vcpu *vcpu, gva_t gva, unsigned long pcid)
5546 struct kvm_mmu *mmu = vcpu->arch.mmu;
5547 bool tlb_flush = false;
5550 if (pcid == kvm_get_active_pcid(vcpu)) {
5551 mmu->invlpg(vcpu, gva, mmu->root_hpa);
5555 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) {
5556 if (VALID_PAGE(mmu->prev_roots[i].hpa) &&
5557 pcid == kvm_get_pcid(vcpu, mmu->prev_roots[i].cr3)) {
5558 mmu->invlpg(vcpu, gva, mmu->prev_roots[i].hpa);
5564 kvm_x86_ops.tlb_flush_gva(vcpu, gva);
5566 ++vcpu->stat.invlpg;
5569 * Mappings not reachable via the current cr3 or the prev_roots will be
5570 * synced when switching to that cr3, so nothing needs to be done here
5574 EXPORT_SYMBOL_GPL(kvm_mmu_invpcid_gva);
5576 void kvm_configure_mmu(bool enable_tdp, int tdp_page_level)
5578 tdp_enabled = enable_tdp;
5581 * max_page_level reflects the capabilities of KVM's MMU irrespective
5582 * of kernel support, e.g. KVM may be capable of using 1GB pages when
5583 * the kernel is not. But, KVM never creates a page size greater than
5584 * what is used by the kernel for any given HVA, i.e. the kernel's
5585 * capabilities are ultimately consulted by kvm_mmu_hugepage_adjust().
5588 max_page_level = tdp_page_level;
5589 else if (boot_cpu_has(X86_FEATURE_GBPAGES))
5590 max_page_level = PT_PDPE_LEVEL;
5592 max_page_level = PT_DIRECTORY_LEVEL;
5594 EXPORT_SYMBOL_GPL(kvm_configure_mmu);
5596 /* The return value indicates if tlb flush on all vcpus is needed. */
5597 typedef bool (*slot_level_handler) (struct kvm *kvm, struct kvm_rmap_head *rmap_head);
5599 /* The caller should hold mmu-lock before calling this function. */
5600 static __always_inline bool
5601 slot_handle_level_range(struct kvm *kvm, struct kvm_memory_slot *memslot,
5602 slot_level_handler fn, int start_level, int end_level,
5603 gfn_t start_gfn, gfn_t end_gfn, bool lock_flush_tlb)
5605 struct slot_rmap_walk_iterator iterator;
5608 for_each_slot_rmap_range(memslot, start_level, end_level, start_gfn,
5609 end_gfn, &iterator) {
5611 flush |= fn(kvm, iterator.rmap);
5613 if (need_resched() || spin_needbreak(&kvm->mmu_lock)) {
5614 if (flush && lock_flush_tlb) {
5615 kvm_flush_remote_tlbs_with_address(kvm,
5617 iterator.gfn - start_gfn + 1);
5620 cond_resched_lock(&kvm->mmu_lock);
5624 if (flush && lock_flush_tlb) {
5625 kvm_flush_remote_tlbs_with_address(kvm, start_gfn,
5626 end_gfn - start_gfn + 1);
5633 static __always_inline bool
5634 slot_handle_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
5635 slot_level_handler fn, int start_level, int end_level,
5636 bool lock_flush_tlb)
5638 return slot_handle_level_range(kvm, memslot, fn, start_level,
5639 end_level, memslot->base_gfn,
5640 memslot->base_gfn + memslot->npages - 1,
5644 static __always_inline bool
5645 slot_handle_all_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
5646 slot_level_handler fn, bool lock_flush_tlb)
5648 return slot_handle_level(kvm, memslot, fn, PT_PAGE_TABLE_LEVEL,
5649 PT_MAX_HUGEPAGE_LEVEL, lock_flush_tlb);
5652 static __always_inline bool
5653 slot_handle_large_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
5654 slot_level_handler fn, bool lock_flush_tlb)
5656 return slot_handle_level(kvm, memslot, fn, PT_PAGE_TABLE_LEVEL + 1,
5657 PT_MAX_HUGEPAGE_LEVEL, lock_flush_tlb);
5660 static __always_inline bool
5661 slot_handle_leaf(struct kvm *kvm, struct kvm_memory_slot *memslot,
5662 slot_level_handler fn, bool lock_flush_tlb)
5664 return slot_handle_level(kvm, memslot, fn, PT_PAGE_TABLE_LEVEL,
5665 PT_PAGE_TABLE_LEVEL, lock_flush_tlb);
5668 static void free_mmu_pages(struct kvm_mmu *mmu)
5670 free_page((unsigned long)mmu->pae_root);
5671 free_page((unsigned long)mmu->lm_root);
5674 static int alloc_mmu_pages(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu)
5680 * When using PAE paging, the four PDPTEs are treated as 'root' pages,
5681 * while the PDP table is a per-vCPU construct that's allocated at MMU
5682 * creation. When emulating 32-bit mode, cr3 is only 32 bits even on
5683 * x86_64. Therefore we need to allocate the PDP table in the first
5684 * 4GB of memory, which happens to fit the DMA32 zone. Except for
5685 * SVM's 32-bit NPT support, TDP paging doesn't use PAE paging and can
5686 * skip allocating the PDP table.
5688 if (tdp_enabled && kvm_x86_ops.get_tdp_level(vcpu) > PT32E_ROOT_LEVEL)
5691 page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_DMA32);
5695 mmu->pae_root = page_address(page);
5696 for (i = 0; i < 4; ++i)
5697 mmu->pae_root[i] = INVALID_PAGE;
5702 int kvm_mmu_create(struct kvm_vcpu *vcpu)
5707 vcpu->arch.mmu = &vcpu->arch.root_mmu;
5708 vcpu->arch.walk_mmu = &vcpu->arch.root_mmu;
5710 vcpu->arch.root_mmu.root_hpa = INVALID_PAGE;
5711 vcpu->arch.root_mmu.root_cr3 = 0;
5712 vcpu->arch.root_mmu.translate_gpa = translate_gpa;
5713 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
5714 vcpu->arch.root_mmu.prev_roots[i] = KVM_MMU_ROOT_INFO_INVALID;
5716 vcpu->arch.guest_mmu.root_hpa = INVALID_PAGE;
5717 vcpu->arch.guest_mmu.root_cr3 = 0;
5718 vcpu->arch.guest_mmu.translate_gpa = translate_gpa;
5719 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
5720 vcpu->arch.guest_mmu.prev_roots[i] = KVM_MMU_ROOT_INFO_INVALID;
5722 vcpu->arch.nested_mmu.translate_gpa = translate_nested_gpa;
5724 ret = alloc_mmu_pages(vcpu, &vcpu->arch.guest_mmu);
5728 ret = alloc_mmu_pages(vcpu, &vcpu->arch.root_mmu);
5730 goto fail_allocate_root;
5734 free_mmu_pages(&vcpu->arch.guest_mmu);
5738 #define BATCH_ZAP_PAGES 10
5739 static void kvm_zap_obsolete_pages(struct kvm *kvm)
5741 struct kvm_mmu_page *sp, *node;
5742 int nr_zapped, batch = 0;
5745 list_for_each_entry_safe_reverse(sp, node,
5746 &kvm->arch.active_mmu_pages, link) {
5748 * No obsolete valid page exists before a newly created page
5749 * since active_mmu_pages is a FIFO list.
5751 if (!is_obsolete_sp(kvm, sp))
5755 * Skip invalid pages with a non-zero root count, zapping pages
5756 * with a non-zero root count will never succeed, i.e. the page
5757 * will get thrown back on active_mmu_pages and we'll get stuck
5758 * in an infinite loop.
5760 if (sp->role.invalid && sp->root_count)
5764 * No need to flush the TLB since we're only zapping shadow
5765 * pages with an obsolete generation number and all vCPUS have
5766 * loaded a new root, i.e. the shadow pages being zapped cannot
5767 * be in active use by the guest.
5769 if (batch >= BATCH_ZAP_PAGES &&
5770 cond_resched_lock(&kvm->mmu_lock)) {
5775 if (__kvm_mmu_prepare_zap_page(kvm, sp,
5776 &kvm->arch.zapped_obsolete_pages, &nr_zapped)) {
5783 * Trigger a remote TLB flush before freeing the page tables to ensure
5784 * KVM is not in the middle of a lockless shadow page table walk, which
5785 * may reference the pages.
5787 kvm_mmu_commit_zap_page(kvm, &kvm->arch.zapped_obsolete_pages);
5791 * Fast invalidate all shadow pages and use lock-break technique
5792 * to zap obsolete pages.
5794 * It's required when memslot is being deleted or VM is being
5795 * destroyed, in these cases, we should ensure that KVM MMU does
5796 * not use any resource of the being-deleted slot or all slots
5797 * after calling the function.
5799 static void kvm_mmu_zap_all_fast(struct kvm *kvm)
5801 lockdep_assert_held(&kvm->slots_lock);
5803 spin_lock(&kvm->mmu_lock);
5804 trace_kvm_mmu_zap_all_fast(kvm);
5807 * Toggle mmu_valid_gen between '0' and '1'. Because slots_lock is
5808 * held for the entire duration of zapping obsolete pages, it's
5809 * impossible for there to be multiple invalid generations associated
5810 * with *valid* shadow pages at any given time, i.e. there is exactly
5811 * one valid generation and (at most) one invalid generation.
5813 kvm->arch.mmu_valid_gen = kvm->arch.mmu_valid_gen ? 0 : 1;
5816 * Notify all vcpus to reload its shadow page table and flush TLB.
5817 * Then all vcpus will switch to new shadow page table with the new
5820 * Note: we need to do this under the protection of mmu_lock,
5821 * otherwise, vcpu would purge shadow page but miss tlb flush.
5823 kvm_reload_remote_mmus(kvm);
5825 kvm_zap_obsolete_pages(kvm);
5826 spin_unlock(&kvm->mmu_lock);
5829 static bool kvm_has_zapped_obsolete_pages(struct kvm *kvm)
5831 return unlikely(!list_empty_careful(&kvm->arch.zapped_obsolete_pages));
5834 static void kvm_mmu_invalidate_zap_pages_in_memslot(struct kvm *kvm,
5835 struct kvm_memory_slot *slot,
5836 struct kvm_page_track_notifier_node *node)
5838 kvm_mmu_zap_all_fast(kvm);
5841 void kvm_mmu_init_vm(struct kvm *kvm)
5843 struct kvm_page_track_notifier_node *node = &kvm->arch.mmu_sp_tracker;
5845 node->track_write = kvm_mmu_pte_write;
5846 node->track_flush_slot = kvm_mmu_invalidate_zap_pages_in_memslot;
5847 kvm_page_track_register_notifier(kvm, node);
5850 void kvm_mmu_uninit_vm(struct kvm *kvm)
5852 struct kvm_page_track_notifier_node *node = &kvm->arch.mmu_sp_tracker;
5854 kvm_page_track_unregister_notifier(kvm, node);
5857 void kvm_zap_gfn_range(struct kvm *kvm, gfn_t gfn_start, gfn_t gfn_end)
5859 struct kvm_memslots *slots;
5860 struct kvm_memory_slot *memslot;
5863 spin_lock(&kvm->mmu_lock);
5864 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
5865 slots = __kvm_memslots(kvm, i);
5866 kvm_for_each_memslot(memslot, slots) {
5869 start = max(gfn_start, memslot->base_gfn);
5870 end = min(gfn_end, memslot->base_gfn + memslot->npages);
5874 slot_handle_level_range(kvm, memslot, kvm_zap_rmapp,
5875 PT_PAGE_TABLE_LEVEL, PT_MAX_HUGEPAGE_LEVEL,
5876 start, end - 1, true);
5880 spin_unlock(&kvm->mmu_lock);
5883 static bool slot_rmap_write_protect(struct kvm *kvm,
5884 struct kvm_rmap_head *rmap_head)
5886 return __rmap_write_protect(kvm, rmap_head, false);
5889 void kvm_mmu_slot_remove_write_access(struct kvm *kvm,
5890 struct kvm_memory_slot *memslot,
5895 spin_lock(&kvm->mmu_lock);
5896 flush = slot_handle_level(kvm, memslot, slot_rmap_write_protect,
5897 start_level, PT_MAX_HUGEPAGE_LEVEL, false);
5898 spin_unlock(&kvm->mmu_lock);
5901 * We can flush all the TLBs out of the mmu lock without TLB
5902 * corruption since we just change the spte from writable to
5903 * readonly so that we only need to care the case of changing
5904 * spte from present to present (changing the spte from present
5905 * to nonpresent will flush all the TLBs immediately), in other
5906 * words, the only case we care is mmu_spte_update() where we
5907 * have checked SPTE_HOST_WRITEABLE | SPTE_MMU_WRITEABLE
5908 * instead of PT_WRITABLE_MASK, that means it does not depend
5909 * on PT_WRITABLE_MASK anymore.
5912 kvm_arch_flush_remote_tlbs_memslot(kvm, memslot);
5915 static bool kvm_mmu_zap_collapsible_spte(struct kvm *kvm,
5916 struct kvm_rmap_head *rmap_head)
5919 struct rmap_iterator iter;
5920 int need_tlb_flush = 0;
5922 struct kvm_mmu_page *sp;
5925 for_each_rmap_spte(rmap_head, &iter, sptep) {
5926 sp = page_header(__pa(sptep));
5927 pfn = spte_to_pfn(*sptep);
5930 * We cannot do huge page mapping for indirect shadow pages,
5931 * which are found on the last rmap (level = 1) when not using
5932 * tdp; such shadow pages are synced with the page table in
5933 * the guest, and the guest page table is using 4K page size
5934 * mapping if the indirect sp has level = 1.
5936 if (sp->role.direct && !kvm_is_reserved_pfn(pfn) &&
5937 (kvm_is_zone_device_pfn(pfn) ||
5938 PageCompound(pfn_to_page(pfn)))) {
5939 pte_list_remove(rmap_head, sptep);
5941 if (kvm_available_flush_tlb_with_range())
5942 kvm_flush_remote_tlbs_with_address(kvm, sp->gfn,
5943 KVM_PAGES_PER_HPAGE(sp->role.level));
5951 return need_tlb_flush;
5954 void kvm_mmu_zap_collapsible_sptes(struct kvm *kvm,
5955 const struct kvm_memory_slot *memslot)
5957 /* FIXME: const-ify all uses of struct kvm_memory_slot. */
5958 spin_lock(&kvm->mmu_lock);
5959 slot_handle_leaf(kvm, (struct kvm_memory_slot *)memslot,
5960 kvm_mmu_zap_collapsible_spte, true);
5961 spin_unlock(&kvm->mmu_lock);
5964 void kvm_arch_flush_remote_tlbs_memslot(struct kvm *kvm,
5965 struct kvm_memory_slot *memslot)
5968 * All current use cases for flushing the TLBs for a specific memslot
5969 * are related to dirty logging, and do the TLB flush out of mmu_lock.
5970 * The interaction between the various operations on memslot must be
5971 * serialized by slots_locks to ensure the TLB flush from one operation
5972 * is observed by any other operation on the same memslot.
5974 lockdep_assert_held(&kvm->slots_lock);
5975 kvm_flush_remote_tlbs_with_address(kvm, memslot->base_gfn,
5979 void kvm_mmu_slot_leaf_clear_dirty(struct kvm *kvm,
5980 struct kvm_memory_slot *memslot)
5984 spin_lock(&kvm->mmu_lock);
5985 flush = slot_handle_leaf(kvm, memslot, __rmap_clear_dirty, false);
5986 spin_unlock(&kvm->mmu_lock);
5989 * It's also safe to flush TLBs out of mmu lock here as currently this
5990 * function is only used for dirty logging, in which case flushing TLB
5991 * out of mmu lock also guarantees no dirty pages will be lost in
5995 kvm_arch_flush_remote_tlbs_memslot(kvm, memslot);
5997 EXPORT_SYMBOL_GPL(kvm_mmu_slot_leaf_clear_dirty);
5999 void kvm_mmu_slot_largepage_remove_write_access(struct kvm *kvm,
6000 struct kvm_memory_slot *memslot)
6004 spin_lock(&kvm->mmu_lock);
6005 flush = slot_handle_large_level(kvm, memslot, slot_rmap_write_protect,
6007 spin_unlock(&kvm->mmu_lock);
6010 kvm_arch_flush_remote_tlbs_memslot(kvm, memslot);
6012 EXPORT_SYMBOL_GPL(kvm_mmu_slot_largepage_remove_write_access);
6014 void kvm_mmu_slot_set_dirty(struct kvm *kvm,
6015 struct kvm_memory_slot *memslot)
6019 spin_lock(&kvm->mmu_lock);
6020 flush = slot_handle_all_level(kvm, memslot, __rmap_set_dirty, false);
6021 spin_unlock(&kvm->mmu_lock);
6024 kvm_arch_flush_remote_tlbs_memslot(kvm, memslot);
6026 EXPORT_SYMBOL_GPL(kvm_mmu_slot_set_dirty);
6028 void kvm_mmu_zap_all(struct kvm *kvm)
6030 struct kvm_mmu_page *sp, *node;
6031 LIST_HEAD(invalid_list);
6034 spin_lock(&kvm->mmu_lock);
6036 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link) {
6037 if (sp->role.invalid && sp->root_count)
6039 if (__kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list, &ign))
6041 if (cond_resched_lock(&kvm->mmu_lock))
6045 kvm_mmu_commit_zap_page(kvm, &invalid_list);
6046 spin_unlock(&kvm->mmu_lock);
6049 void kvm_mmu_invalidate_mmio_sptes(struct kvm *kvm, u64 gen)
6051 WARN_ON(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS);
6053 gen &= MMIO_SPTE_GEN_MASK;
6056 * Generation numbers are incremented in multiples of the number of
6057 * address spaces in order to provide unique generations across all
6058 * address spaces. Strip what is effectively the address space
6059 * modifier prior to checking for a wrap of the MMIO generation so
6060 * that a wrap in any address space is detected.
6062 gen &= ~((u64)KVM_ADDRESS_SPACE_NUM - 1);
6065 * The very rare case: if the MMIO generation number has wrapped,
6066 * zap all shadow pages.
6068 if (unlikely(gen == 0)) {
6069 kvm_debug_ratelimited("kvm: zapping shadow pages for mmio generation wraparound\n");
6070 kvm_mmu_zap_all_fast(kvm);
6074 static unsigned long
6075 mmu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
6078 int nr_to_scan = sc->nr_to_scan;
6079 unsigned long freed = 0;
6081 mutex_lock(&kvm_lock);
6083 list_for_each_entry(kvm, &vm_list, vm_list) {
6085 LIST_HEAD(invalid_list);
6088 * Never scan more than sc->nr_to_scan VM instances.
6089 * Will not hit this condition practically since we do not try
6090 * to shrink more than one VM and it is very unlikely to see
6091 * !n_used_mmu_pages so many times.
6096 * n_used_mmu_pages is accessed without holding kvm->mmu_lock
6097 * here. We may skip a VM instance errorneosly, but we do not
6098 * want to shrink a VM that only started to populate its MMU
6101 if (!kvm->arch.n_used_mmu_pages &&
6102 !kvm_has_zapped_obsolete_pages(kvm))
6105 idx = srcu_read_lock(&kvm->srcu);
6106 spin_lock(&kvm->mmu_lock);
6108 if (kvm_has_zapped_obsolete_pages(kvm)) {
6109 kvm_mmu_commit_zap_page(kvm,
6110 &kvm->arch.zapped_obsolete_pages);
6114 if (prepare_zap_oldest_mmu_page(kvm, &invalid_list))
6116 kvm_mmu_commit_zap_page(kvm, &invalid_list);
6119 spin_unlock(&kvm->mmu_lock);
6120 srcu_read_unlock(&kvm->srcu, idx);
6123 * unfair on small ones
6124 * per-vm shrinkers cry out
6125 * sadness comes quickly
6127 list_move_tail(&kvm->vm_list, &vm_list);
6131 mutex_unlock(&kvm_lock);
6135 static unsigned long
6136 mmu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
6138 return percpu_counter_read_positive(&kvm_total_used_mmu_pages);
6141 static struct shrinker mmu_shrinker = {
6142 .count_objects = mmu_shrink_count,
6143 .scan_objects = mmu_shrink_scan,
6144 .seeks = DEFAULT_SEEKS * 10,
6147 static void mmu_destroy_caches(void)
6149 kmem_cache_destroy(pte_list_desc_cache);
6150 kmem_cache_destroy(mmu_page_header_cache);
6153 static void kvm_set_mmio_spte_mask(void)
6158 * Set the reserved bits and the present bit of an paging-structure
6159 * entry to generate page fault with PFER.RSV = 1.
6163 * Mask the uppermost physical address bit, which would be reserved as
6164 * long as the supported physical address width is less than 52.
6168 /* Set the present bit. */
6172 * If reserved bit is not supported, clear the present bit to disable
6175 if (shadow_phys_bits == 52)
6178 kvm_mmu_set_mmio_spte_mask(mask, mask, ACC_WRITE_MASK | ACC_USER_MASK);
6181 static bool get_nx_auto_mode(void)
6183 /* Return true when CPU has the bug, and mitigations are ON */
6184 return boot_cpu_has_bug(X86_BUG_ITLB_MULTIHIT) && !cpu_mitigations_off();
6187 static void __set_nx_huge_pages(bool val)
6189 nx_huge_pages = itlb_multihit_kvm_mitigation = val;
6192 static int set_nx_huge_pages(const char *val, const struct kernel_param *kp)
6194 bool old_val = nx_huge_pages;
6197 /* In "auto" mode deploy workaround only if CPU has the bug. */
6198 if (sysfs_streq(val, "off"))
6200 else if (sysfs_streq(val, "force"))
6202 else if (sysfs_streq(val, "auto"))
6203 new_val = get_nx_auto_mode();
6204 else if (strtobool(val, &new_val) < 0)
6207 __set_nx_huge_pages(new_val);
6209 if (new_val != old_val) {
6212 mutex_lock(&kvm_lock);
6214 list_for_each_entry(kvm, &vm_list, vm_list) {
6215 mutex_lock(&kvm->slots_lock);
6216 kvm_mmu_zap_all_fast(kvm);
6217 mutex_unlock(&kvm->slots_lock);
6219 wake_up_process(kvm->arch.nx_lpage_recovery_thread);
6221 mutex_unlock(&kvm_lock);
6227 int kvm_mmu_module_init(void)
6231 if (nx_huge_pages == -1)
6232 __set_nx_huge_pages(get_nx_auto_mode());
6235 * MMU roles use union aliasing which is, generally speaking, an
6236 * undefined behavior. However, we supposedly know how compilers behave
6237 * and the current status quo is unlikely to change. Guardians below are
6238 * supposed to let us know if the assumption becomes false.
6240 BUILD_BUG_ON(sizeof(union kvm_mmu_page_role) != sizeof(u32));
6241 BUILD_BUG_ON(sizeof(union kvm_mmu_extended_role) != sizeof(u32));
6242 BUILD_BUG_ON(sizeof(union kvm_mmu_role) != sizeof(u64));
6244 kvm_mmu_reset_all_pte_masks();
6246 kvm_set_mmio_spte_mask();
6248 pte_list_desc_cache = kmem_cache_create("pte_list_desc",
6249 sizeof(struct pte_list_desc),
6250 0, SLAB_ACCOUNT, NULL);
6251 if (!pte_list_desc_cache)
6254 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
6255 sizeof(struct kvm_mmu_page),
6256 0, SLAB_ACCOUNT, NULL);
6257 if (!mmu_page_header_cache)
6260 if (percpu_counter_init(&kvm_total_used_mmu_pages, 0, GFP_KERNEL))
6263 ret = register_shrinker(&mmu_shrinker);
6270 mmu_destroy_caches();
6275 * Calculate mmu pages needed for kvm.
6277 unsigned long kvm_mmu_calculate_default_mmu_pages(struct kvm *kvm)
6279 unsigned long nr_mmu_pages;
6280 unsigned long nr_pages = 0;
6281 struct kvm_memslots *slots;
6282 struct kvm_memory_slot *memslot;
6285 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
6286 slots = __kvm_memslots(kvm, i);
6288 kvm_for_each_memslot(memslot, slots)
6289 nr_pages += memslot->npages;
6292 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
6293 nr_mmu_pages = max(nr_mmu_pages, KVM_MIN_ALLOC_MMU_PAGES);
6295 return nr_mmu_pages;
6298 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
6300 kvm_mmu_unload(vcpu);
6301 free_mmu_pages(&vcpu->arch.root_mmu);
6302 free_mmu_pages(&vcpu->arch.guest_mmu);
6303 mmu_free_memory_caches(vcpu);
6306 void kvm_mmu_module_exit(void)
6308 mmu_destroy_caches();
6309 percpu_counter_destroy(&kvm_total_used_mmu_pages);
6310 unregister_shrinker(&mmu_shrinker);
6311 mmu_audit_disable();
6314 static int set_nx_huge_pages_recovery_ratio(const char *val, const struct kernel_param *kp)
6316 unsigned int old_val;
6319 old_val = nx_huge_pages_recovery_ratio;
6320 err = param_set_uint(val, kp);
6324 if (READ_ONCE(nx_huge_pages) &&
6325 !old_val && nx_huge_pages_recovery_ratio) {
6328 mutex_lock(&kvm_lock);
6330 list_for_each_entry(kvm, &vm_list, vm_list)
6331 wake_up_process(kvm->arch.nx_lpage_recovery_thread);
6333 mutex_unlock(&kvm_lock);
6339 static void kvm_recover_nx_lpages(struct kvm *kvm)
6342 struct kvm_mmu_page *sp;
6344 LIST_HEAD(invalid_list);
6347 rcu_idx = srcu_read_lock(&kvm->srcu);
6348 spin_lock(&kvm->mmu_lock);
6350 ratio = READ_ONCE(nx_huge_pages_recovery_ratio);
6351 to_zap = ratio ? DIV_ROUND_UP(kvm->stat.nx_lpage_splits, ratio) : 0;
6352 while (to_zap && !list_empty(&kvm->arch.lpage_disallowed_mmu_pages)) {
6354 * We use a separate list instead of just using active_mmu_pages
6355 * because the number of lpage_disallowed pages is expected to
6356 * be relatively small compared to the total.
6358 sp = list_first_entry(&kvm->arch.lpage_disallowed_mmu_pages,
6359 struct kvm_mmu_page,
6360 lpage_disallowed_link);
6361 WARN_ON_ONCE(!sp->lpage_disallowed);
6362 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
6363 WARN_ON_ONCE(sp->lpage_disallowed);
6365 if (!--to_zap || need_resched() || spin_needbreak(&kvm->mmu_lock)) {
6366 kvm_mmu_commit_zap_page(kvm, &invalid_list);
6368 cond_resched_lock(&kvm->mmu_lock);
6372 spin_unlock(&kvm->mmu_lock);
6373 srcu_read_unlock(&kvm->srcu, rcu_idx);
6376 static long get_nx_lpage_recovery_timeout(u64 start_time)
6378 return READ_ONCE(nx_huge_pages) && READ_ONCE(nx_huge_pages_recovery_ratio)
6379 ? start_time + 60 * HZ - get_jiffies_64()
6380 : MAX_SCHEDULE_TIMEOUT;
6383 static int kvm_nx_lpage_recovery_worker(struct kvm *kvm, uintptr_t data)
6386 long remaining_time;
6389 start_time = get_jiffies_64();
6390 remaining_time = get_nx_lpage_recovery_timeout(start_time);
6392 set_current_state(TASK_INTERRUPTIBLE);
6393 while (!kthread_should_stop() && remaining_time > 0) {
6394 schedule_timeout(remaining_time);
6395 remaining_time = get_nx_lpage_recovery_timeout(start_time);
6396 set_current_state(TASK_INTERRUPTIBLE);
6399 set_current_state(TASK_RUNNING);
6401 if (kthread_should_stop())
6404 kvm_recover_nx_lpages(kvm);
6408 int kvm_mmu_post_init_vm(struct kvm *kvm)
6412 err = kvm_vm_create_worker_thread(kvm, kvm_nx_lpage_recovery_worker, 0,
6413 "kvm-nx-lpage-recovery",
6414 &kvm->arch.nx_lpage_recovery_thread);
6416 kthread_unpark(kvm->arch.nx_lpage_recovery_thread);
6421 void kvm_mmu_pre_destroy_vm(struct kvm *kvm)
6423 if (kvm->arch.nx_lpage_recovery_thread)
6424 kthread_stop(kvm->arch.nx_lpage_recovery_thread);