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 "mmu_internal.h"
24 #include "kvm_cache_regs.h"
25 #include "kvm_emulate.h"
29 #include <linux/kvm_host.h>
30 #include <linux/types.h>
31 #include <linux/string.h>
33 #include <linux/highmem.h>
34 #include <linux/moduleparam.h>
35 #include <linux/export.h>
36 #include <linux/swap.h>
37 #include <linux/hugetlb.h>
38 #include <linux/compiler.h>
39 #include <linux/srcu.h>
40 #include <linux/slab.h>
41 #include <linux/sched/signal.h>
42 #include <linux/uaccess.h>
43 #include <linux/hash.h>
44 #include <linux/kern_levels.h>
45 #include <linux/kthread.h>
48 #include <asm/memtype.h>
49 #include <asm/cmpxchg.h>
51 #include <asm/set_memory.h>
53 #include <asm/kvm_page_track.h>
56 extern bool itlb_multihit_kvm_mitigation;
58 int __read_mostly nx_huge_pages = -1;
59 #ifdef CONFIG_PREEMPT_RT
60 /* Recovery can cause latency spikes, disable it for PREEMPT_RT. */
61 static uint __read_mostly nx_huge_pages_recovery_ratio = 0;
63 static uint __read_mostly nx_huge_pages_recovery_ratio = 60;
66 static int set_nx_huge_pages(const char *val, const struct kernel_param *kp);
67 static int set_nx_huge_pages_recovery_ratio(const char *val, const struct kernel_param *kp);
69 static const struct kernel_param_ops nx_huge_pages_ops = {
70 .set = set_nx_huge_pages,
71 .get = param_get_bool,
74 static const struct kernel_param_ops nx_huge_pages_recovery_ratio_ops = {
75 .set = set_nx_huge_pages_recovery_ratio,
76 .get = param_get_uint,
79 module_param_cb(nx_huge_pages, &nx_huge_pages_ops, &nx_huge_pages, 0644);
80 __MODULE_PARM_TYPE(nx_huge_pages, "bool");
81 module_param_cb(nx_huge_pages_recovery_ratio, &nx_huge_pages_recovery_ratio_ops,
82 &nx_huge_pages_recovery_ratio, 0644);
83 __MODULE_PARM_TYPE(nx_huge_pages_recovery_ratio, "uint");
85 static bool __read_mostly force_flush_and_sync_on_reuse;
86 module_param_named(flush_on_reuse, force_flush_and_sync_on_reuse, bool, 0644);
89 * When setting this variable to true it enables Two-Dimensional-Paging
90 * where the hardware walks 2 page tables:
91 * 1. the guest-virtual to guest-physical
92 * 2. while doing 1. it walks guest-physical to host-physical
93 * If the hardware supports that we don't need to do shadow paging.
95 bool tdp_enabled = false;
97 static int max_huge_page_level __read_mostly;
98 static int max_tdp_level __read_mostly;
101 AUDIT_PRE_PAGE_FAULT,
102 AUDIT_POST_PAGE_FAULT,
104 AUDIT_POST_PTE_WRITE,
111 module_param(dbg, bool, 0644);
114 #define PTE_PREFETCH_NUM 8
116 #define PT32_LEVEL_BITS 10
118 #define PT32_LEVEL_SHIFT(level) \
119 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
121 #define PT32_LVL_OFFSET_MASK(level) \
122 (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
123 * PT32_LEVEL_BITS))) - 1))
125 #define PT32_INDEX(address, level)\
126 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
129 #define PT32_BASE_ADDR_MASK PAGE_MASK
130 #define PT32_DIR_BASE_ADDR_MASK \
131 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
132 #define PT32_LVL_ADDR_MASK(level) \
133 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
134 * PT32_LEVEL_BITS))) - 1))
136 #include <trace/events/kvm.h>
138 /* make pte_list_desc fit well in cache line */
139 #define PTE_LIST_EXT 3
141 struct pte_list_desc {
142 u64 *sptes[PTE_LIST_EXT];
143 struct pte_list_desc *more;
146 struct kvm_shadow_walk_iterator {
154 #define for_each_shadow_entry_using_root(_vcpu, _root, _addr, _walker) \
155 for (shadow_walk_init_using_root(&(_walker), (_vcpu), \
157 shadow_walk_okay(&(_walker)); \
158 shadow_walk_next(&(_walker)))
160 #define for_each_shadow_entry(_vcpu, _addr, _walker) \
161 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
162 shadow_walk_okay(&(_walker)); \
163 shadow_walk_next(&(_walker)))
165 #define for_each_shadow_entry_lockless(_vcpu, _addr, _walker, spte) \
166 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
167 shadow_walk_okay(&(_walker)) && \
168 ({ spte = mmu_spte_get_lockless(_walker.sptep); 1; }); \
169 __shadow_walk_next(&(_walker), spte))
171 static struct kmem_cache *pte_list_desc_cache;
172 struct kmem_cache *mmu_page_header_cache;
173 static struct percpu_counter kvm_total_used_mmu_pages;
175 static void mmu_spte_set(u64 *sptep, u64 spte);
176 static union kvm_mmu_page_role
177 kvm_mmu_calc_root_page_role(struct kvm_vcpu *vcpu);
179 struct kvm_mmu_role_regs {
180 const unsigned long cr0;
181 const unsigned long cr4;
185 #define CREATE_TRACE_POINTS
186 #include "mmutrace.h"
189 * Yes, lot's of underscores. They're a hint that you probably shouldn't be
190 * reading from the role_regs. Once the mmu_role is constructed, it becomes
191 * the single source of truth for the MMU's state.
193 #define BUILD_MMU_ROLE_REGS_ACCESSOR(reg, name, flag) \
194 static inline bool ____is_##reg##_##name(struct kvm_mmu_role_regs *regs)\
196 return !!(regs->reg & flag); \
198 BUILD_MMU_ROLE_REGS_ACCESSOR(cr0, pg, X86_CR0_PG);
199 BUILD_MMU_ROLE_REGS_ACCESSOR(cr0, wp, X86_CR0_WP);
200 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, pse, X86_CR4_PSE);
201 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, pae, X86_CR4_PAE);
202 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, smep, X86_CR4_SMEP);
203 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, smap, X86_CR4_SMAP);
204 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, pke, X86_CR4_PKE);
205 BUILD_MMU_ROLE_REGS_ACCESSOR(cr4, la57, X86_CR4_LA57);
206 BUILD_MMU_ROLE_REGS_ACCESSOR(efer, nx, EFER_NX);
207 BUILD_MMU_ROLE_REGS_ACCESSOR(efer, lma, EFER_LMA);
210 * The MMU itself (with a valid role) is the single source of truth for the
211 * MMU. Do not use the regs used to build the MMU/role, nor the vCPU. The
212 * regs don't account for dependencies, e.g. clearing CR4 bits if CR0.PG=1,
213 * and the vCPU may be incorrect/irrelevant.
215 #define BUILD_MMU_ROLE_ACCESSOR(base_or_ext, reg, name) \
216 static inline bool is_##reg##_##name(struct kvm_mmu *mmu) \
218 return !!(mmu->mmu_role. base_or_ext . reg##_##name); \
220 BUILD_MMU_ROLE_ACCESSOR(ext, cr0, pg);
221 BUILD_MMU_ROLE_ACCESSOR(base, cr0, wp);
222 BUILD_MMU_ROLE_ACCESSOR(ext, cr4, pse);
223 BUILD_MMU_ROLE_ACCESSOR(ext, cr4, pae);
224 BUILD_MMU_ROLE_ACCESSOR(ext, cr4, smep);
225 BUILD_MMU_ROLE_ACCESSOR(ext, cr4, smap);
226 BUILD_MMU_ROLE_ACCESSOR(ext, cr4, pke);
227 BUILD_MMU_ROLE_ACCESSOR(ext, cr4, la57);
228 BUILD_MMU_ROLE_ACCESSOR(base, efer, nx);
230 static struct kvm_mmu_role_regs vcpu_to_role_regs(struct kvm_vcpu *vcpu)
232 struct kvm_mmu_role_regs regs = {
233 .cr0 = kvm_read_cr0_bits(vcpu, KVM_MMU_CR0_ROLE_BITS),
234 .cr4 = kvm_read_cr4_bits(vcpu, KVM_MMU_CR4_ROLE_BITS),
235 .efer = vcpu->arch.efer,
241 static inline bool kvm_available_flush_tlb_with_range(void)
243 return kvm_x86_ops.tlb_remote_flush_with_range;
246 static void kvm_flush_remote_tlbs_with_range(struct kvm *kvm,
247 struct kvm_tlb_range *range)
251 if (range && kvm_x86_ops.tlb_remote_flush_with_range)
252 ret = static_call(kvm_x86_tlb_remote_flush_with_range)(kvm, range);
255 kvm_flush_remote_tlbs(kvm);
258 void kvm_flush_remote_tlbs_with_address(struct kvm *kvm,
259 u64 start_gfn, u64 pages)
261 struct kvm_tlb_range range;
263 range.start_gfn = start_gfn;
266 kvm_flush_remote_tlbs_with_range(kvm, &range);
269 static void mark_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, u64 gfn,
272 u64 spte = make_mmio_spte(vcpu, gfn, access);
274 trace_mark_mmio_spte(sptep, gfn, spte);
275 mmu_spte_set(sptep, spte);
278 static gfn_t get_mmio_spte_gfn(u64 spte)
280 u64 gpa = spte & shadow_nonpresent_or_rsvd_lower_gfn_mask;
282 gpa |= (spte >> SHADOW_NONPRESENT_OR_RSVD_MASK_LEN)
283 & shadow_nonpresent_or_rsvd_mask;
285 return gpa >> PAGE_SHIFT;
288 static unsigned get_mmio_spte_access(u64 spte)
290 return spte & shadow_mmio_access_mask;
293 static bool check_mmio_spte(struct kvm_vcpu *vcpu, u64 spte)
295 u64 kvm_gen, spte_gen, gen;
297 gen = kvm_vcpu_memslots(vcpu)->generation;
298 if (unlikely(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS))
301 kvm_gen = gen & MMIO_SPTE_GEN_MASK;
302 spte_gen = get_mmio_spte_generation(spte);
304 trace_check_mmio_spte(spte, kvm_gen, spte_gen);
305 return likely(kvm_gen == spte_gen);
308 static gpa_t translate_gpa(struct kvm_vcpu *vcpu, gpa_t gpa, u32 access,
309 struct x86_exception *exception)
311 /* Check if guest physical address doesn't exceed guest maximum */
312 if (kvm_vcpu_is_illegal_gpa(vcpu, gpa)) {
313 exception->error_code |= PFERR_RSVD_MASK;
320 static int is_cpuid_PSE36(void)
325 static gfn_t pse36_gfn_delta(u32 gpte)
327 int shift = 32 - PT32_DIR_PSE36_SHIFT - PAGE_SHIFT;
329 return (gpte & PT32_DIR_PSE36_MASK) << shift;
333 static void __set_spte(u64 *sptep, u64 spte)
335 WRITE_ONCE(*sptep, spte);
338 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
340 WRITE_ONCE(*sptep, spte);
343 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
345 return xchg(sptep, spte);
348 static u64 __get_spte_lockless(u64 *sptep)
350 return READ_ONCE(*sptep);
361 static void count_spte_clear(u64 *sptep, u64 spte)
363 struct kvm_mmu_page *sp = sptep_to_sp(sptep);
365 if (is_shadow_present_pte(spte))
368 /* Ensure the spte is completely set before we increase the count */
370 sp->clear_spte_count++;
373 static void __set_spte(u64 *sptep, u64 spte)
375 union split_spte *ssptep, sspte;
377 ssptep = (union split_spte *)sptep;
378 sspte = (union split_spte)spte;
380 ssptep->spte_high = sspte.spte_high;
383 * If we map the spte from nonpresent to present, We should store
384 * the high bits firstly, then set present bit, so cpu can not
385 * fetch this spte while we are setting the spte.
389 WRITE_ONCE(ssptep->spte_low, sspte.spte_low);
392 static void __update_clear_spte_fast(u64 *sptep, u64 spte)
394 union split_spte *ssptep, sspte;
396 ssptep = (union split_spte *)sptep;
397 sspte = (union split_spte)spte;
399 WRITE_ONCE(ssptep->spte_low, sspte.spte_low);
402 * If we map the spte from present to nonpresent, we should clear
403 * present bit firstly to avoid vcpu fetch the old high bits.
407 ssptep->spte_high = sspte.spte_high;
408 count_spte_clear(sptep, spte);
411 static u64 __update_clear_spte_slow(u64 *sptep, u64 spte)
413 union split_spte *ssptep, sspte, orig;
415 ssptep = (union split_spte *)sptep;
416 sspte = (union split_spte)spte;
418 /* xchg acts as a barrier before the setting of the high bits */
419 orig.spte_low = xchg(&ssptep->spte_low, sspte.spte_low);
420 orig.spte_high = ssptep->spte_high;
421 ssptep->spte_high = sspte.spte_high;
422 count_spte_clear(sptep, spte);
428 * The idea using the light way get the spte on x86_32 guest is from
429 * gup_get_pte (mm/gup.c).
431 * An spte tlb flush may be pending, because kvm_set_pte_rmapp
432 * coalesces them and we are running out of the MMU lock. Therefore
433 * we need to protect against in-progress updates of the spte.
435 * Reading the spte while an update is in progress may get the old value
436 * for the high part of the spte. The race is fine for a present->non-present
437 * change (because the high part of the spte is ignored for non-present spte),
438 * but for a present->present change we must reread the spte.
440 * All such changes are done in two steps (present->non-present and
441 * non-present->present), hence it is enough to count the number of
442 * present->non-present updates: if it changed while reading the spte,
443 * we might have hit the race. This is done using clear_spte_count.
445 static u64 __get_spte_lockless(u64 *sptep)
447 struct kvm_mmu_page *sp = sptep_to_sp(sptep);
448 union split_spte spte, *orig = (union split_spte *)sptep;
452 count = sp->clear_spte_count;
455 spte.spte_low = orig->spte_low;
458 spte.spte_high = orig->spte_high;
461 if (unlikely(spte.spte_low != orig->spte_low ||
462 count != sp->clear_spte_count))
469 static bool spte_has_volatile_bits(u64 spte)
471 if (!is_shadow_present_pte(spte))
475 * Always atomically update spte if it can be updated
476 * out of mmu-lock, it can ensure dirty bit is not lost,
477 * also, it can help us to get a stable is_writable_pte()
478 * to ensure tlb flush is not missed.
480 if (spte_can_locklessly_be_made_writable(spte) ||
481 is_access_track_spte(spte))
484 if (spte_ad_enabled(spte)) {
485 if ((spte & shadow_accessed_mask) == 0 ||
486 (is_writable_pte(spte) && (spte & shadow_dirty_mask) == 0))
493 /* Rules for using mmu_spte_set:
494 * Set the sptep from nonpresent to present.
495 * Note: the sptep being assigned *must* be either not present
496 * or in a state where the hardware will not attempt to update
499 static void mmu_spte_set(u64 *sptep, u64 new_spte)
501 WARN_ON(is_shadow_present_pte(*sptep));
502 __set_spte(sptep, new_spte);
506 * Update the SPTE (excluding the PFN), but do not track changes in its
507 * accessed/dirty status.
509 static u64 mmu_spte_update_no_track(u64 *sptep, u64 new_spte)
511 u64 old_spte = *sptep;
513 WARN_ON(!is_shadow_present_pte(new_spte));
515 if (!is_shadow_present_pte(old_spte)) {
516 mmu_spte_set(sptep, new_spte);
520 if (!spte_has_volatile_bits(old_spte))
521 __update_clear_spte_fast(sptep, new_spte);
523 old_spte = __update_clear_spte_slow(sptep, new_spte);
525 WARN_ON(spte_to_pfn(old_spte) != spte_to_pfn(new_spte));
530 /* Rules for using mmu_spte_update:
531 * Update the state bits, it means the mapped pfn is not changed.
533 * Whenever we overwrite a writable spte with a read-only one we
534 * should flush remote TLBs. Otherwise rmap_write_protect
535 * will find a read-only spte, even though the writable spte
536 * might be cached on a CPU's TLB, the return value indicates this
539 * Returns true if the TLB needs to be flushed
541 static bool mmu_spte_update(u64 *sptep, u64 new_spte)
544 u64 old_spte = mmu_spte_update_no_track(sptep, new_spte);
546 if (!is_shadow_present_pte(old_spte))
550 * For the spte updated out of mmu-lock is safe, since
551 * we always atomically update it, see the comments in
552 * spte_has_volatile_bits().
554 if (spte_can_locklessly_be_made_writable(old_spte) &&
555 !is_writable_pte(new_spte))
559 * Flush TLB when accessed/dirty states are changed in the page tables,
560 * to guarantee consistency between TLB and page tables.
563 if (is_accessed_spte(old_spte) && !is_accessed_spte(new_spte)) {
565 kvm_set_pfn_accessed(spte_to_pfn(old_spte));
568 if (is_dirty_spte(old_spte) && !is_dirty_spte(new_spte)) {
570 kvm_set_pfn_dirty(spte_to_pfn(old_spte));
577 * Rules for using mmu_spte_clear_track_bits:
578 * It sets the sptep from present to nonpresent, and track the
579 * state bits, it is used to clear the last level sptep.
580 * Returns non-zero if the PTE was previously valid.
582 static int mmu_spte_clear_track_bits(u64 *sptep)
585 u64 old_spte = *sptep;
587 if (!spte_has_volatile_bits(old_spte))
588 __update_clear_spte_fast(sptep, 0ull);
590 old_spte = __update_clear_spte_slow(sptep, 0ull);
592 if (!is_shadow_present_pte(old_spte))
595 pfn = spte_to_pfn(old_spte);
598 * KVM does not hold the refcount of the page used by
599 * kvm mmu, before reclaiming the page, we should
600 * unmap it from mmu first.
602 WARN_ON(!kvm_is_reserved_pfn(pfn) && !page_count(pfn_to_page(pfn)));
604 if (is_accessed_spte(old_spte))
605 kvm_set_pfn_accessed(pfn);
607 if (is_dirty_spte(old_spte))
608 kvm_set_pfn_dirty(pfn);
614 * Rules for using mmu_spte_clear_no_track:
615 * Directly clear spte without caring the state bits of sptep,
616 * it is used to set the upper level spte.
618 static void mmu_spte_clear_no_track(u64 *sptep)
620 __update_clear_spte_fast(sptep, 0ull);
623 static u64 mmu_spte_get_lockless(u64 *sptep)
625 return __get_spte_lockless(sptep);
628 /* Restore an acc-track PTE back to a regular PTE */
629 static u64 restore_acc_track_spte(u64 spte)
632 u64 saved_bits = (spte >> SHADOW_ACC_TRACK_SAVED_BITS_SHIFT)
633 & SHADOW_ACC_TRACK_SAVED_BITS_MASK;
635 WARN_ON_ONCE(spte_ad_enabled(spte));
636 WARN_ON_ONCE(!is_access_track_spte(spte));
638 new_spte &= ~shadow_acc_track_mask;
639 new_spte &= ~(SHADOW_ACC_TRACK_SAVED_BITS_MASK <<
640 SHADOW_ACC_TRACK_SAVED_BITS_SHIFT);
641 new_spte |= saved_bits;
646 /* Returns the Accessed status of the PTE and resets it at the same time. */
647 static bool mmu_spte_age(u64 *sptep)
649 u64 spte = mmu_spte_get_lockless(sptep);
651 if (!is_accessed_spte(spte))
654 if (spte_ad_enabled(spte)) {
655 clear_bit((ffs(shadow_accessed_mask) - 1),
656 (unsigned long *)sptep);
659 * Capture the dirty status of the page, so that it doesn't get
660 * lost when the SPTE is marked for access tracking.
662 if (is_writable_pte(spte))
663 kvm_set_pfn_dirty(spte_to_pfn(spte));
665 spte = mark_spte_for_access_track(spte);
666 mmu_spte_update_no_track(sptep, spte);
672 static void walk_shadow_page_lockless_begin(struct kvm_vcpu *vcpu)
675 * Prevent page table teardown by making any free-er wait during
676 * kvm_flush_remote_tlbs() IPI to all active vcpus.
681 * Make sure a following spte read is not reordered ahead of the write
684 smp_store_mb(vcpu->mode, READING_SHADOW_PAGE_TABLES);
687 static void walk_shadow_page_lockless_end(struct kvm_vcpu *vcpu)
690 * Make sure the write to vcpu->mode is not reordered in front of
691 * reads to sptes. If it does, kvm_mmu_commit_zap_page() can see us
692 * OUTSIDE_GUEST_MODE and proceed to free the shadow page table.
694 smp_store_release(&vcpu->mode, OUTSIDE_GUEST_MODE);
698 static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu, bool maybe_indirect)
702 /* 1 rmap, 1 parent PTE per level, and the prefetched rmaps. */
703 r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache,
704 1 + PT64_ROOT_MAX_LEVEL + PTE_PREFETCH_NUM);
707 r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_shadow_page_cache,
708 PT64_ROOT_MAX_LEVEL);
711 if (maybe_indirect) {
712 r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_gfn_array_cache,
713 PT64_ROOT_MAX_LEVEL);
717 return kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_page_header_cache,
718 PT64_ROOT_MAX_LEVEL);
721 static void mmu_free_memory_caches(struct kvm_vcpu *vcpu)
723 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_pte_list_desc_cache);
724 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_shadow_page_cache);
725 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_gfn_array_cache);
726 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_header_cache);
729 static struct pte_list_desc *mmu_alloc_pte_list_desc(struct kvm_vcpu *vcpu)
731 return kvm_mmu_memory_cache_alloc(&vcpu->arch.mmu_pte_list_desc_cache);
734 static void mmu_free_pte_list_desc(struct pte_list_desc *pte_list_desc)
736 kmem_cache_free(pte_list_desc_cache, pte_list_desc);
739 static gfn_t kvm_mmu_page_get_gfn(struct kvm_mmu_page *sp, int index)
741 if (!sp->role.direct)
742 return sp->gfns[index];
744 return sp->gfn + (index << ((sp->role.level - 1) * PT64_LEVEL_BITS));
747 static void kvm_mmu_page_set_gfn(struct kvm_mmu_page *sp, int index, gfn_t gfn)
749 if (!sp->role.direct) {
750 sp->gfns[index] = gfn;
754 if (WARN_ON(gfn != kvm_mmu_page_get_gfn(sp, index)))
755 pr_err_ratelimited("gfn mismatch under direct page %llx "
756 "(expected %llx, got %llx)\n",
758 kvm_mmu_page_get_gfn(sp, index), gfn);
762 * Return the pointer to the large page information for a given gfn,
763 * handling slots that are not large page aligned.
765 static struct kvm_lpage_info *lpage_info_slot(gfn_t gfn,
766 const struct kvm_memory_slot *slot, int level)
770 idx = gfn_to_index(gfn, slot->base_gfn, level);
771 return &slot->arch.lpage_info[level - 2][idx];
774 static void update_gfn_disallow_lpage_count(struct kvm_memory_slot *slot,
775 gfn_t gfn, int count)
777 struct kvm_lpage_info *linfo;
780 for (i = PG_LEVEL_2M; i <= KVM_MAX_HUGEPAGE_LEVEL; ++i) {
781 linfo = lpage_info_slot(gfn, slot, i);
782 linfo->disallow_lpage += count;
783 WARN_ON(linfo->disallow_lpage < 0);
787 void kvm_mmu_gfn_disallow_lpage(struct kvm_memory_slot *slot, gfn_t gfn)
789 update_gfn_disallow_lpage_count(slot, gfn, 1);
792 void kvm_mmu_gfn_allow_lpage(struct kvm_memory_slot *slot, gfn_t gfn)
794 update_gfn_disallow_lpage_count(slot, gfn, -1);
797 static void account_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp)
799 struct kvm_memslots *slots;
800 struct kvm_memory_slot *slot;
803 kvm->arch.indirect_shadow_pages++;
805 slots = kvm_memslots_for_spte_role(kvm, sp->role);
806 slot = __gfn_to_memslot(slots, gfn);
808 /* the non-leaf shadow pages are keeping readonly. */
809 if (sp->role.level > PG_LEVEL_4K)
810 return kvm_slot_page_track_add_page(kvm, slot, gfn,
811 KVM_PAGE_TRACK_WRITE);
813 kvm_mmu_gfn_disallow_lpage(slot, gfn);
816 void account_huge_nx_page(struct kvm *kvm, struct kvm_mmu_page *sp)
818 if (sp->lpage_disallowed)
821 ++kvm->stat.nx_lpage_splits;
822 list_add_tail(&sp->lpage_disallowed_link,
823 &kvm->arch.lpage_disallowed_mmu_pages);
824 sp->lpage_disallowed = true;
827 static void unaccount_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp)
829 struct kvm_memslots *slots;
830 struct kvm_memory_slot *slot;
833 kvm->arch.indirect_shadow_pages--;
835 slots = kvm_memslots_for_spte_role(kvm, sp->role);
836 slot = __gfn_to_memslot(slots, gfn);
837 if (sp->role.level > PG_LEVEL_4K)
838 return kvm_slot_page_track_remove_page(kvm, slot, gfn,
839 KVM_PAGE_TRACK_WRITE);
841 kvm_mmu_gfn_allow_lpage(slot, gfn);
844 void unaccount_huge_nx_page(struct kvm *kvm, struct kvm_mmu_page *sp)
846 --kvm->stat.nx_lpage_splits;
847 sp->lpage_disallowed = false;
848 list_del(&sp->lpage_disallowed_link);
851 static struct kvm_memory_slot *
852 gfn_to_memslot_dirty_bitmap(struct kvm_vcpu *vcpu, gfn_t gfn,
855 struct kvm_memory_slot *slot;
857 slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
858 if (!slot || slot->flags & KVM_MEMSLOT_INVALID)
860 if (no_dirty_log && kvm_slot_dirty_track_enabled(slot))
867 * About rmap_head encoding:
869 * If the bit zero of rmap_head->val is clear, then it points to the only spte
870 * in this rmap chain. Otherwise, (rmap_head->val & ~1) points to a struct
871 * pte_list_desc containing more mappings.
875 * Returns the number of pointers in the rmap chain, not counting the new one.
877 static int pte_list_add(struct kvm_vcpu *vcpu, u64 *spte,
878 struct kvm_rmap_head *rmap_head)
880 struct pte_list_desc *desc;
883 if (!rmap_head->val) {
884 rmap_printk("%p %llx 0->1\n", spte, *spte);
885 rmap_head->val = (unsigned long)spte;
886 } else if (!(rmap_head->val & 1)) {
887 rmap_printk("%p %llx 1->many\n", spte, *spte);
888 desc = mmu_alloc_pte_list_desc(vcpu);
889 desc->sptes[0] = (u64 *)rmap_head->val;
890 desc->sptes[1] = spte;
891 rmap_head->val = (unsigned long)desc | 1;
894 rmap_printk("%p %llx many->many\n", spte, *spte);
895 desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
896 while (desc->sptes[PTE_LIST_EXT-1]) {
897 count += PTE_LIST_EXT;
900 desc->more = mmu_alloc_pte_list_desc(vcpu);
906 for (i = 0; desc->sptes[i]; ++i)
908 desc->sptes[i] = spte;
914 pte_list_desc_remove_entry(struct kvm_rmap_head *rmap_head,
915 struct pte_list_desc *desc, int i,
916 struct pte_list_desc *prev_desc)
920 for (j = PTE_LIST_EXT - 1; !desc->sptes[j] && j > i; --j)
922 desc->sptes[i] = desc->sptes[j];
923 desc->sptes[j] = NULL;
926 if (!prev_desc && !desc->more)
930 prev_desc->more = desc->more;
932 rmap_head->val = (unsigned long)desc->more | 1;
933 mmu_free_pte_list_desc(desc);
936 static void __pte_list_remove(u64 *spte, struct kvm_rmap_head *rmap_head)
938 struct pte_list_desc *desc;
939 struct pte_list_desc *prev_desc;
942 if (!rmap_head->val) {
943 pr_err("%s: %p 0->BUG\n", __func__, spte);
945 } else if (!(rmap_head->val & 1)) {
946 rmap_printk("%p 1->0\n", spte);
947 if ((u64 *)rmap_head->val != spte) {
948 pr_err("%s: %p 1->BUG\n", __func__, spte);
953 rmap_printk("%p many->many\n", spte);
954 desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
957 for (i = 0; i < PTE_LIST_EXT && desc->sptes[i]; ++i) {
958 if (desc->sptes[i] == spte) {
959 pte_list_desc_remove_entry(rmap_head,
967 pr_err("%s: %p many->many\n", __func__, spte);
972 static void pte_list_remove(struct kvm_rmap_head *rmap_head, u64 *sptep)
974 mmu_spte_clear_track_bits(sptep);
975 __pte_list_remove(sptep, rmap_head);
978 static struct kvm_rmap_head *__gfn_to_rmap(gfn_t gfn, int level,
979 struct kvm_memory_slot *slot)
983 idx = gfn_to_index(gfn, slot->base_gfn, level);
984 return &slot->arch.rmap[level - PG_LEVEL_4K][idx];
987 static struct kvm_rmap_head *gfn_to_rmap(struct kvm *kvm, gfn_t gfn,
988 struct kvm_mmu_page *sp)
990 struct kvm_memslots *slots;
991 struct kvm_memory_slot *slot;
993 slots = kvm_memslots_for_spte_role(kvm, sp->role);
994 slot = __gfn_to_memslot(slots, gfn);
995 return __gfn_to_rmap(gfn, sp->role.level, slot);
998 static bool rmap_can_add(struct kvm_vcpu *vcpu)
1000 struct kvm_mmu_memory_cache *mc;
1002 mc = &vcpu->arch.mmu_pte_list_desc_cache;
1003 return kvm_mmu_memory_cache_nr_free_objects(mc);
1006 static int rmap_add(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
1008 struct kvm_mmu_page *sp;
1009 struct kvm_rmap_head *rmap_head;
1011 sp = sptep_to_sp(spte);
1012 kvm_mmu_page_set_gfn(sp, spte - sp->spt, gfn);
1013 rmap_head = gfn_to_rmap(vcpu->kvm, gfn, sp);
1014 return pte_list_add(vcpu, spte, rmap_head);
1017 static void rmap_remove(struct kvm *kvm, u64 *spte)
1019 struct kvm_mmu_page *sp;
1021 struct kvm_rmap_head *rmap_head;
1023 sp = sptep_to_sp(spte);
1024 gfn = kvm_mmu_page_get_gfn(sp, spte - sp->spt);
1025 rmap_head = gfn_to_rmap(kvm, gfn, sp);
1026 __pte_list_remove(spte, rmap_head);
1030 * Used by the following functions to iterate through the sptes linked by a
1031 * rmap. All fields are private and not assumed to be used outside.
1033 struct rmap_iterator {
1034 /* private fields */
1035 struct pte_list_desc *desc; /* holds the sptep if not NULL */
1036 int pos; /* index of the sptep */
1040 * Iteration must be started by this function. This should also be used after
1041 * removing/dropping sptes from the rmap link because in such cases the
1042 * information in the iterator may not be valid.
1044 * Returns sptep if found, NULL otherwise.
1046 static u64 *rmap_get_first(struct kvm_rmap_head *rmap_head,
1047 struct rmap_iterator *iter)
1051 if (!rmap_head->val)
1054 if (!(rmap_head->val & 1)) {
1056 sptep = (u64 *)rmap_head->val;
1060 iter->desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
1062 sptep = iter->desc->sptes[iter->pos];
1064 BUG_ON(!is_shadow_present_pte(*sptep));
1069 * Must be used with a valid iterator: e.g. after rmap_get_first().
1071 * Returns sptep if found, NULL otherwise.
1073 static u64 *rmap_get_next(struct rmap_iterator *iter)
1078 if (iter->pos < PTE_LIST_EXT - 1) {
1080 sptep = iter->desc->sptes[iter->pos];
1085 iter->desc = iter->desc->more;
1089 /* desc->sptes[0] cannot be NULL */
1090 sptep = iter->desc->sptes[iter->pos];
1097 BUG_ON(!is_shadow_present_pte(*sptep));
1101 #define for_each_rmap_spte(_rmap_head_, _iter_, _spte_) \
1102 for (_spte_ = rmap_get_first(_rmap_head_, _iter_); \
1103 _spte_; _spte_ = rmap_get_next(_iter_))
1105 static void drop_spte(struct kvm *kvm, u64 *sptep)
1107 if (mmu_spte_clear_track_bits(sptep))
1108 rmap_remove(kvm, sptep);
1112 static bool __drop_large_spte(struct kvm *kvm, u64 *sptep)
1114 if (is_large_pte(*sptep)) {
1115 WARN_ON(sptep_to_sp(sptep)->role.level == PG_LEVEL_4K);
1116 drop_spte(kvm, sptep);
1124 static void drop_large_spte(struct kvm_vcpu *vcpu, u64 *sptep)
1126 if (__drop_large_spte(vcpu->kvm, sptep)) {
1127 struct kvm_mmu_page *sp = sptep_to_sp(sptep);
1129 kvm_flush_remote_tlbs_with_address(vcpu->kvm, sp->gfn,
1130 KVM_PAGES_PER_HPAGE(sp->role.level));
1135 * Write-protect on the specified @sptep, @pt_protect indicates whether
1136 * spte write-protection is caused by protecting shadow page table.
1138 * Note: write protection is difference between dirty logging and spte
1140 * - for dirty logging, the spte can be set to writable at anytime if
1141 * its dirty bitmap is properly set.
1142 * - for spte protection, the spte can be writable only after unsync-ing
1145 * Return true if tlb need be flushed.
1147 static bool spte_write_protect(u64 *sptep, bool pt_protect)
1151 if (!is_writable_pte(spte) &&
1152 !(pt_protect && spte_can_locklessly_be_made_writable(spte)))
1155 rmap_printk("spte %p %llx\n", sptep, *sptep);
1158 spte &= ~shadow_mmu_writable_mask;
1159 spte = spte & ~PT_WRITABLE_MASK;
1161 return mmu_spte_update(sptep, spte);
1164 static bool __rmap_write_protect(struct kvm *kvm,
1165 struct kvm_rmap_head *rmap_head,
1169 struct rmap_iterator iter;
1172 for_each_rmap_spte(rmap_head, &iter, sptep)
1173 flush |= spte_write_protect(sptep, pt_protect);
1178 static bool spte_clear_dirty(u64 *sptep)
1182 rmap_printk("spte %p %llx\n", sptep, *sptep);
1184 MMU_WARN_ON(!spte_ad_enabled(spte));
1185 spte &= ~shadow_dirty_mask;
1186 return mmu_spte_update(sptep, spte);
1189 static bool spte_wrprot_for_clear_dirty(u64 *sptep)
1191 bool was_writable = test_and_clear_bit(PT_WRITABLE_SHIFT,
1192 (unsigned long *)sptep);
1193 if (was_writable && !spte_ad_enabled(*sptep))
1194 kvm_set_pfn_dirty(spte_to_pfn(*sptep));
1196 return was_writable;
1200 * Gets the GFN ready for another round of dirty logging by clearing the
1201 * - D bit on ad-enabled SPTEs, and
1202 * - W bit on ad-disabled SPTEs.
1203 * Returns true iff any D or W bits were cleared.
1205 static bool __rmap_clear_dirty(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1206 struct kvm_memory_slot *slot)
1209 struct rmap_iterator iter;
1212 for_each_rmap_spte(rmap_head, &iter, sptep)
1213 if (spte_ad_need_write_protect(*sptep))
1214 flush |= spte_wrprot_for_clear_dirty(sptep);
1216 flush |= spte_clear_dirty(sptep);
1222 * kvm_mmu_write_protect_pt_masked - write protect selected PT level pages
1223 * @kvm: kvm instance
1224 * @slot: slot to protect
1225 * @gfn_offset: start of the BITS_PER_LONG pages we care about
1226 * @mask: indicates which pages we should protect
1228 * Used when we do not need to care about huge page mappings.
1230 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1231 struct kvm_memory_slot *slot,
1232 gfn_t gfn_offset, unsigned long mask)
1234 struct kvm_rmap_head *rmap_head;
1236 if (is_tdp_mmu_enabled(kvm))
1237 kvm_tdp_mmu_clear_dirty_pt_masked(kvm, slot,
1238 slot->base_gfn + gfn_offset, mask, true);
1240 if (!kvm_memslots_have_rmaps(kvm))
1244 rmap_head = __gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask),
1246 __rmap_write_protect(kvm, rmap_head, false);
1248 /* clear the first set bit */
1254 * kvm_mmu_clear_dirty_pt_masked - clear MMU D-bit for PT level pages, or write
1255 * protect the page if the D-bit isn't supported.
1256 * @kvm: kvm instance
1257 * @slot: slot to clear D-bit
1258 * @gfn_offset: start of the BITS_PER_LONG pages we care about
1259 * @mask: indicates which pages we should clear D-bit
1261 * Used for PML to re-log the dirty GPAs after userspace querying dirty_bitmap.
1263 static void kvm_mmu_clear_dirty_pt_masked(struct kvm *kvm,
1264 struct kvm_memory_slot *slot,
1265 gfn_t gfn_offset, unsigned long mask)
1267 struct kvm_rmap_head *rmap_head;
1269 if (is_tdp_mmu_enabled(kvm))
1270 kvm_tdp_mmu_clear_dirty_pt_masked(kvm, slot,
1271 slot->base_gfn + gfn_offset, mask, false);
1273 if (!kvm_memslots_have_rmaps(kvm))
1277 rmap_head = __gfn_to_rmap(slot->base_gfn + gfn_offset + __ffs(mask),
1279 __rmap_clear_dirty(kvm, rmap_head, slot);
1281 /* clear the first set bit */
1287 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1290 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1291 * enable dirty logging for them.
1293 * We need to care about huge page mappings: e.g. during dirty logging we may
1294 * have such mappings.
1296 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1297 struct kvm_memory_slot *slot,
1298 gfn_t gfn_offset, unsigned long mask)
1301 * Huge pages are NOT write protected when we start dirty logging in
1302 * initially-all-set mode; must write protect them here so that they
1303 * are split to 4K on the first write.
1305 * The gfn_offset is guaranteed to be aligned to 64, but the base_gfn
1306 * of memslot has no such restriction, so the range can cross two large
1309 if (kvm_dirty_log_manual_protect_and_init_set(kvm)) {
1310 gfn_t start = slot->base_gfn + gfn_offset + __ffs(mask);
1311 gfn_t end = slot->base_gfn + gfn_offset + __fls(mask);
1313 kvm_mmu_slot_gfn_write_protect(kvm, slot, start, PG_LEVEL_2M);
1315 /* Cross two large pages? */
1316 if (ALIGN(start << PAGE_SHIFT, PMD_SIZE) !=
1317 ALIGN(end << PAGE_SHIFT, PMD_SIZE))
1318 kvm_mmu_slot_gfn_write_protect(kvm, slot, end,
1322 /* Now handle 4K PTEs. */
1323 if (kvm_x86_ops.cpu_dirty_log_size)
1324 kvm_mmu_clear_dirty_pt_masked(kvm, slot, gfn_offset, mask);
1326 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1329 int kvm_cpu_dirty_log_size(void)
1331 return kvm_x86_ops.cpu_dirty_log_size;
1334 bool kvm_mmu_slot_gfn_write_protect(struct kvm *kvm,
1335 struct kvm_memory_slot *slot, u64 gfn,
1338 struct kvm_rmap_head *rmap_head;
1340 bool write_protected = false;
1342 if (kvm_memslots_have_rmaps(kvm)) {
1343 for (i = min_level; i <= KVM_MAX_HUGEPAGE_LEVEL; ++i) {
1344 rmap_head = __gfn_to_rmap(gfn, i, slot);
1345 write_protected |= __rmap_write_protect(kvm, rmap_head, true);
1349 if (is_tdp_mmu_enabled(kvm))
1351 kvm_tdp_mmu_write_protect_gfn(kvm, slot, gfn, min_level);
1353 return write_protected;
1356 static bool rmap_write_protect(struct kvm_vcpu *vcpu, u64 gfn)
1358 struct kvm_memory_slot *slot;
1360 slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1361 return kvm_mmu_slot_gfn_write_protect(vcpu->kvm, slot, gfn, PG_LEVEL_4K);
1364 static bool kvm_zap_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1365 struct kvm_memory_slot *slot)
1368 struct rmap_iterator iter;
1371 while ((sptep = rmap_get_first(rmap_head, &iter))) {
1372 rmap_printk("spte %p %llx.\n", sptep, *sptep);
1374 pte_list_remove(rmap_head, sptep);
1381 static bool kvm_unmap_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1382 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1385 return kvm_zap_rmapp(kvm, rmap_head, slot);
1388 static bool kvm_set_pte_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1389 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1393 struct rmap_iterator iter;
1398 WARN_ON(pte_huge(pte));
1399 new_pfn = pte_pfn(pte);
1402 for_each_rmap_spte(rmap_head, &iter, sptep) {
1403 rmap_printk("spte %p %llx gfn %llx (%d)\n",
1404 sptep, *sptep, gfn, level);
1408 if (pte_write(pte)) {
1409 pte_list_remove(rmap_head, sptep);
1412 new_spte = kvm_mmu_changed_pte_notifier_make_spte(
1415 mmu_spte_clear_track_bits(sptep);
1416 mmu_spte_set(sptep, new_spte);
1420 if (need_flush && kvm_available_flush_tlb_with_range()) {
1421 kvm_flush_remote_tlbs_with_address(kvm, gfn, 1);
1428 struct slot_rmap_walk_iterator {
1430 struct kvm_memory_slot *slot;
1436 /* output fields. */
1438 struct kvm_rmap_head *rmap;
1441 /* private field. */
1442 struct kvm_rmap_head *end_rmap;
1446 rmap_walk_init_level(struct slot_rmap_walk_iterator *iterator, int level)
1448 iterator->level = level;
1449 iterator->gfn = iterator->start_gfn;
1450 iterator->rmap = __gfn_to_rmap(iterator->gfn, level, iterator->slot);
1451 iterator->end_rmap = __gfn_to_rmap(iterator->end_gfn, level,
1456 slot_rmap_walk_init(struct slot_rmap_walk_iterator *iterator,
1457 struct kvm_memory_slot *slot, int start_level,
1458 int end_level, gfn_t start_gfn, gfn_t end_gfn)
1460 iterator->slot = slot;
1461 iterator->start_level = start_level;
1462 iterator->end_level = end_level;
1463 iterator->start_gfn = start_gfn;
1464 iterator->end_gfn = end_gfn;
1466 rmap_walk_init_level(iterator, iterator->start_level);
1469 static bool slot_rmap_walk_okay(struct slot_rmap_walk_iterator *iterator)
1471 return !!iterator->rmap;
1474 static void slot_rmap_walk_next(struct slot_rmap_walk_iterator *iterator)
1476 if (++iterator->rmap <= iterator->end_rmap) {
1477 iterator->gfn += (1UL << KVM_HPAGE_GFN_SHIFT(iterator->level));
1481 if (++iterator->level > iterator->end_level) {
1482 iterator->rmap = NULL;
1486 rmap_walk_init_level(iterator, iterator->level);
1489 #define for_each_slot_rmap_range(_slot_, _start_level_, _end_level_, \
1490 _start_gfn, _end_gfn, _iter_) \
1491 for (slot_rmap_walk_init(_iter_, _slot_, _start_level_, \
1492 _end_level_, _start_gfn, _end_gfn); \
1493 slot_rmap_walk_okay(_iter_); \
1494 slot_rmap_walk_next(_iter_))
1496 typedef bool (*rmap_handler_t)(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1497 struct kvm_memory_slot *slot, gfn_t gfn,
1498 int level, pte_t pte);
1500 static __always_inline bool kvm_handle_gfn_range(struct kvm *kvm,
1501 struct kvm_gfn_range *range,
1502 rmap_handler_t handler)
1504 struct slot_rmap_walk_iterator iterator;
1507 for_each_slot_rmap_range(range->slot, PG_LEVEL_4K, KVM_MAX_HUGEPAGE_LEVEL,
1508 range->start, range->end - 1, &iterator)
1509 ret |= handler(kvm, iterator.rmap, range->slot, iterator.gfn,
1510 iterator.level, range->pte);
1515 bool kvm_unmap_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range)
1519 if (kvm_memslots_have_rmaps(kvm))
1520 flush = kvm_handle_gfn_range(kvm, range, kvm_unmap_rmapp);
1522 if (is_tdp_mmu_enabled(kvm))
1523 flush |= kvm_tdp_mmu_unmap_gfn_range(kvm, range, flush);
1528 bool kvm_set_spte_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1532 if (kvm_memslots_have_rmaps(kvm))
1533 flush = kvm_handle_gfn_range(kvm, range, kvm_set_pte_rmapp);
1535 if (is_tdp_mmu_enabled(kvm))
1536 flush |= kvm_tdp_mmu_set_spte_gfn(kvm, range);
1541 static bool kvm_age_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1542 struct kvm_memory_slot *slot, gfn_t gfn, int level,
1546 struct rmap_iterator iter;
1549 for_each_rmap_spte(rmap_head, &iter, sptep)
1550 young |= mmu_spte_age(sptep);
1555 static bool kvm_test_age_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
1556 struct kvm_memory_slot *slot, gfn_t gfn,
1557 int level, pte_t unused)
1560 struct rmap_iterator iter;
1562 for_each_rmap_spte(rmap_head, &iter, sptep)
1563 if (is_accessed_spte(*sptep))
1568 #define RMAP_RECYCLE_THRESHOLD 1000
1570 static void rmap_recycle(struct kvm_vcpu *vcpu, u64 *spte, gfn_t gfn)
1572 struct kvm_rmap_head *rmap_head;
1573 struct kvm_mmu_page *sp;
1575 sp = sptep_to_sp(spte);
1577 rmap_head = gfn_to_rmap(vcpu->kvm, gfn, sp);
1579 kvm_unmap_rmapp(vcpu->kvm, rmap_head, NULL, gfn, sp->role.level, __pte(0));
1580 kvm_flush_remote_tlbs_with_address(vcpu->kvm, sp->gfn,
1581 KVM_PAGES_PER_HPAGE(sp->role.level));
1584 bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1588 if (kvm_memslots_have_rmaps(kvm))
1589 young = kvm_handle_gfn_range(kvm, range, kvm_age_rmapp);
1591 if (is_tdp_mmu_enabled(kvm))
1592 young |= kvm_tdp_mmu_age_gfn_range(kvm, range);
1597 bool kvm_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1601 if (kvm_memslots_have_rmaps(kvm))
1602 young = kvm_handle_gfn_range(kvm, range, kvm_test_age_rmapp);
1604 if (is_tdp_mmu_enabled(kvm))
1605 young |= kvm_tdp_mmu_test_age_gfn(kvm, range);
1611 static int is_empty_shadow_page(u64 *spt)
1616 for (pos = spt, end = pos + PAGE_SIZE / sizeof(u64); pos != end; pos++)
1617 if (is_shadow_present_pte(*pos)) {
1618 printk(KERN_ERR "%s: %p %llx\n", __func__,
1627 * This value is the sum of all of the kvm instances's
1628 * kvm->arch.n_used_mmu_pages values. We need a global,
1629 * aggregate version in order to make the slab shrinker
1632 static inline void kvm_mod_used_mmu_pages(struct kvm *kvm, unsigned long nr)
1634 kvm->arch.n_used_mmu_pages += nr;
1635 percpu_counter_add(&kvm_total_used_mmu_pages, nr);
1638 static void kvm_mmu_free_page(struct kvm_mmu_page *sp)
1640 MMU_WARN_ON(!is_empty_shadow_page(sp->spt));
1641 hlist_del(&sp->hash_link);
1642 list_del(&sp->link);
1643 free_page((unsigned long)sp->spt);
1644 if (!sp->role.direct)
1645 free_page((unsigned long)sp->gfns);
1646 kmem_cache_free(mmu_page_header_cache, sp);
1649 static unsigned kvm_page_table_hashfn(gfn_t gfn)
1651 return hash_64(gfn, KVM_MMU_HASH_SHIFT);
1654 static void mmu_page_add_parent_pte(struct kvm_vcpu *vcpu,
1655 struct kvm_mmu_page *sp, u64 *parent_pte)
1660 pte_list_add(vcpu, parent_pte, &sp->parent_ptes);
1663 static void mmu_page_remove_parent_pte(struct kvm_mmu_page *sp,
1666 __pte_list_remove(parent_pte, &sp->parent_ptes);
1669 static void drop_parent_pte(struct kvm_mmu_page *sp,
1672 mmu_page_remove_parent_pte(sp, parent_pte);
1673 mmu_spte_clear_no_track(parent_pte);
1676 static struct kvm_mmu_page *kvm_mmu_alloc_page(struct kvm_vcpu *vcpu, int direct)
1678 struct kvm_mmu_page *sp;
1680 sp = kvm_mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache);
1681 sp->spt = kvm_mmu_memory_cache_alloc(&vcpu->arch.mmu_shadow_page_cache);
1683 sp->gfns = kvm_mmu_memory_cache_alloc(&vcpu->arch.mmu_gfn_array_cache);
1684 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
1687 * active_mmu_pages must be a FIFO list, as kvm_zap_obsolete_pages()
1688 * depends on valid pages being added to the head of the list. See
1689 * comments in kvm_zap_obsolete_pages().
1691 sp->mmu_valid_gen = vcpu->kvm->arch.mmu_valid_gen;
1692 list_add(&sp->link, &vcpu->kvm->arch.active_mmu_pages);
1693 kvm_mod_used_mmu_pages(vcpu->kvm, +1);
1697 static void mark_unsync(u64 *spte);
1698 static void kvm_mmu_mark_parents_unsync(struct kvm_mmu_page *sp)
1701 struct rmap_iterator iter;
1703 for_each_rmap_spte(&sp->parent_ptes, &iter, sptep) {
1708 static void mark_unsync(u64 *spte)
1710 struct kvm_mmu_page *sp;
1713 sp = sptep_to_sp(spte);
1714 index = spte - sp->spt;
1715 if (__test_and_set_bit(index, sp->unsync_child_bitmap))
1717 if (sp->unsync_children++)
1719 kvm_mmu_mark_parents_unsync(sp);
1722 static int nonpaging_sync_page(struct kvm_vcpu *vcpu,
1723 struct kvm_mmu_page *sp)
1728 #define KVM_PAGE_ARRAY_NR 16
1730 struct kvm_mmu_pages {
1731 struct mmu_page_and_offset {
1732 struct kvm_mmu_page *sp;
1734 } page[KVM_PAGE_ARRAY_NR];
1738 static int mmu_pages_add(struct kvm_mmu_pages *pvec, struct kvm_mmu_page *sp,
1744 for (i=0; i < pvec->nr; i++)
1745 if (pvec->page[i].sp == sp)
1748 pvec->page[pvec->nr].sp = sp;
1749 pvec->page[pvec->nr].idx = idx;
1751 return (pvec->nr == KVM_PAGE_ARRAY_NR);
1754 static inline void clear_unsync_child_bit(struct kvm_mmu_page *sp, int idx)
1756 --sp->unsync_children;
1757 WARN_ON((int)sp->unsync_children < 0);
1758 __clear_bit(idx, sp->unsync_child_bitmap);
1761 static int __mmu_unsync_walk(struct kvm_mmu_page *sp,
1762 struct kvm_mmu_pages *pvec)
1764 int i, ret, nr_unsync_leaf = 0;
1766 for_each_set_bit(i, sp->unsync_child_bitmap, 512) {
1767 struct kvm_mmu_page *child;
1768 u64 ent = sp->spt[i];
1770 if (!is_shadow_present_pte(ent) || is_large_pte(ent)) {
1771 clear_unsync_child_bit(sp, i);
1775 child = to_shadow_page(ent & PT64_BASE_ADDR_MASK);
1777 if (child->unsync_children) {
1778 if (mmu_pages_add(pvec, child, i))
1781 ret = __mmu_unsync_walk(child, pvec);
1783 clear_unsync_child_bit(sp, i);
1785 } else if (ret > 0) {
1786 nr_unsync_leaf += ret;
1789 } else if (child->unsync) {
1791 if (mmu_pages_add(pvec, child, i))
1794 clear_unsync_child_bit(sp, i);
1797 return nr_unsync_leaf;
1800 #define INVALID_INDEX (-1)
1802 static int mmu_unsync_walk(struct kvm_mmu_page *sp,
1803 struct kvm_mmu_pages *pvec)
1806 if (!sp->unsync_children)
1809 mmu_pages_add(pvec, sp, INVALID_INDEX);
1810 return __mmu_unsync_walk(sp, pvec);
1813 static void kvm_unlink_unsync_page(struct kvm *kvm, struct kvm_mmu_page *sp)
1815 WARN_ON(!sp->unsync);
1816 trace_kvm_mmu_sync_page(sp);
1818 --kvm->stat.mmu_unsync;
1821 static bool kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
1822 struct list_head *invalid_list);
1823 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
1824 struct list_head *invalid_list);
1826 #define for_each_valid_sp(_kvm, _sp, _list) \
1827 hlist_for_each_entry(_sp, _list, hash_link) \
1828 if (is_obsolete_sp((_kvm), (_sp))) { \
1831 #define for_each_gfn_indirect_valid_sp(_kvm, _sp, _gfn) \
1832 for_each_valid_sp(_kvm, _sp, \
1833 &(_kvm)->arch.mmu_page_hash[kvm_page_table_hashfn(_gfn)]) \
1834 if ((_sp)->gfn != (_gfn) || (_sp)->role.direct) {} else
1836 static bool kvm_sync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
1837 struct list_head *invalid_list)
1839 if (vcpu->arch.mmu->sync_page(vcpu, sp) == 0) {
1840 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, invalid_list);
1847 static bool kvm_mmu_remote_flush_or_zap(struct kvm *kvm,
1848 struct list_head *invalid_list,
1851 if (!remote_flush && list_empty(invalid_list))
1854 if (!list_empty(invalid_list))
1855 kvm_mmu_commit_zap_page(kvm, invalid_list);
1857 kvm_flush_remote_tlbs(kvm);
1861 static void kvm_mmu_flush_or_zap(struct kvm_vcpu *vcpu,
1862 struct list_head *invalid_list,
1863 bool remote_flush, bool local_flush)
1865 if (kvm_mmu_remote_flush_or_zap(vcpu->kvm, invalid_list, remote_flush))
1869 kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu);
1872 #ifdef CONFIG_KVM_MMU_AUDIT
1873 #include "mmu_audit.c"
1875 static void kvm_mmu_audit(struct kvm_vcpu *vcpu, int point) { }
1876 static void mmu_audit_disable(void) { }
1879 static bool is_obsolete_sp(struct kvm *kvm, struct kvm_mmu_page *sp)
1881 return sp->role.invalid ||
1882 unlikely(sp->mmu_valid_gen != kvm->arch.mmu_valid_gen);
1885 struct mmu_page_path {
1886 struct kvm_mmu_page *parent[PT64_ROOT_MAX_LEVEL];
1887 unsigned int idx[PT64_ROOT_MAX_LEVEL];
1890 #define for_each_sp(pvec, sp, parents, i) \
1891 for (i = mmu_pages_first(&pvec, &parents); \
1892 i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \
1893 i = mmu_pages_next(&pvec, &parents, i))
1895 static int mmu_pages_next(struct kvm_mmu_pages *pvec,
1896 struct mmu_page_path *parents,
1901 for (n = i+1; n < pvec->nr; n++) {
1902 struct kvm_mmu_page *sp = pvec->page[n].sp;
1903 unsigned idx = pvec->page[n].idx;
1904 int level = sp->role.level;
1906 parents->idx[level-1] = idx;
1907 if (level == PG_LEVEL_4K)
1910 parents->parent[level-2] = sp;
1916 static int mmu_pages_first(struct kvm_mmu_pages *pvec,
1917 struct mmu_page_path *parents)
1919 struct kvm_mmu_page *sp;
1925 WARN_ON(pvec->page[0].idx != INVALID_INDEX);
1927 sp = pvec->page[0].sp;
1928 level = sp->role.level;
1929 WARN_ON(level == PG_LEVEL_4K);
1931 parents->parent[level-2] = sp;
1933 /* Also set up a sentinel. Further entries in pvec are all
1934 * children of sp, so this element is never overwritten.
1936 parents->parent[level-1] = NULL;
1937 return mmu_pages_next(pvec, parents, 0);
1940 static void mmu_pages_clear_parents(struct mmu_page_path *parents)
1942 struct kvm_mmu_page *sp;
1943 unsigned int level = 0;
1946 unsigned int idx = parents->idx[level];
1947 sp = parents->parent[level];
1951 WARN_ON(idx == INVALID_INDEX);
1952 clear_unsync_child_bit(sp, idx);
1954 } while (!sp->unsync_children);
1957 static void mmu_sync_children(struct kvm_vcpu *vcpu,
1958 struct kvm_mmu_page *parent)
1961 struct kvm_mmu_page *sp;
1962 struct mmu_page_path parents;
1963 struct kvm_mmu_pages pages;
1964 LIST_HEAD(invalid_list);
1967 while (mmu_unsync_walk(parent, &pages)) {
1968 bool protected = false;
1970 for_each_sp(pages, sp, parents, i)
1971 protected |= rmap_write_protect(vcpu, sp->gfn);
1974 kvm_flush_remote_tlbs(vcpu->kvm);
1978 for_each_sp(pages, sp, parents, i) {
1979 kvm_unlink_unsync_page(vcpu->kvm, sp);
1980 flush |= kvm_sync_page(vcpu, sp, &invalid_list);
1981 mmu_pages_clear_parents(&parents);
1983 if (need_resched() || rwlock_needbreak(&vcpu->kvm->mmu_lock)) {
1984 kvm_mmu_flush_or_zap(vcpu, &invalid_list, false, flush);
1985 cond_resched_rwlock_write(&vcpu->kvm->mmu_lock);
1990 kvm_mmu_flush_or_zap(vcpu, &invalid_list, false, flush);
1993 static void __clear_sp_write_flooding_count(struct kvm_mmu_page *sp)
1995 atomic_set(&sp->write_flooding_count, 0);
1998 static void clear_sp_write_flooding_count(u64 *spte)
2000 __clear_sp_write_flooding_count(sptep_to_sp(spte));
2003 static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
2008 unsigned int access)
2010 bool direct_mmu = vcpu->arch.mmu->direct_map;
2011 union kvm_mmu_page_role role;
2012 struct hlist_head *sp_list;
2014 struct kvm_mmu_page *sp;
2016 LIST_HEAD(invalid_list);
2018 role = vcpu->arch.mmu->mmu_role.base;
2020 role.direct = direct;
2022 role.gpte_is_8_bytes = true;
2023 role.access = access;
2024 if (!direct_mmu && vcpu->arch.mmu->root_level <= PT32_ROOT_LEVEL) {
2025 quadrant = gaddr >> (PAGE_SHIFT + (PT64_PT_BITS * level));
2026 quadrant &= (1 << ((PT32_PT_BITS - PT64_PT_BITS) * level)) - 1;
2027 role.quadrant = quadrant;
2030 sp_list = &vcpu->kvm->arch.mmu_page_hash[kvm_page_table_hashfn(gfn)];
2031 for_each_valid_sp(vcpu->kvm, sp, sp_list) {
2032 if (sp->gfn != gfn) {
2037 if (sp->role.word != role.word) {
2039 * If the guest is creating an upper-level page, zap
2040 * unsync pages for the same gfn. While it's possible
2041 * the guest is using recursive page tables, in all
2042 * likelihood the guest has stopped using the unsync
2043 * page and is installing a completely unrelated page.
2044 * Unsync pages must not be left as is, because the new
2045 * upper-level page will be write-protected.
2047 if (level > PG_LEVEL_4K && sp->unsync)
2048 kvm_mmu_prepare_zap_page(vcpu->kvm, sp,
2054 goto trace_get_page;
2058 * The page is good, but is stale. kvm_sync_page does
2059 * get the latest guest state, but (unlike mmu_unsync_children)
2060 * it doesn't write-protect the page or mark it synchronized!
2061 * This way the validity of the mapping is ensured, but the
2062 * overhead of write protection is not incurred until the
2063 * guest invalidates the TLB mapping. This allows multiple
2064 * SPs for a single gfn to be unsync.
2066 * If the sync fails, the page is zapped. If so, break
2067 * in order to rebuild it.
2069 if (!kvm_sync_page(vcpu, sp, &invalid_list))
2072 WARN_ON(!list_empty(&invalid_list));
2073 kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu);
2076 if (sp->unsync_children)
2077 kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
2079 __clear_sp_write_flooding_count(sp);
2082 trace_kvm_mmu_get_page(sp, false);
2086 ++vcpu->kvm->stat.mmu_cache_miss;
2088 sp = kvm_mmu_alloc_page(vcpu, direct);
2092 hlist_add_head(&sp->hash_link, sp_list);
2094 account_shadowed(vcpu->kvm, sp);
2095 if (level == PG_LEVEL_4K && rmap_write_protect(vcpu, gfn))
2096 kvm_flush_remote_tlbs_with_address(vcpu->kvm, gfn, 1);
2098 trace_kvm_mmu_get_page(sp, true);
2100 kvm_mmu_commit_zap_page(vcpu->kvm, &invalid_list);
2102 if (collisions > vcpu->kvm->stat.max_mmu_page_hash_collisions)
2103 vcpu->kvm->stat.max_mmu_page_hash_collisions = collisions;
2107 static void shadow_walk_init_using_root(struct kvm_shadow_walk_iterator *iterator,
2108 struct kvm_vcpu *vcpu, hpa_t root,
2111 iterator->addr = addr;
2112 iterator->shadow_addr = root;
2113 iterator->level = vcpu->arch.mmu->shadow_root_level;
2115 if (iterator->level == PT64_ROOT_4LEVEL &&
2116 vcpu->arch.mmu->root_level < PT64_ROOT_4LEVEL &&
2117 !vcpu->arch.mmu->direct_map)
2120 if (iterator->level == PT32E_ROOT_LEVEL) {
2122 * prev_root is currently only used for 64-bit hosts. So only
2123 * the active root_hpa is valid here.
2125 BUG_ON(root != vcpu->arch.mmu->root_hpa);
2127 iterator->shadow_addr
2128 = vcpu->arch.mmu->pae_root[(addr >> 30) & 3];
2129 iterator->shadow_addr &= PT64_BASE_ADDR_MASK;
2131 if (!iterator->shadow_addr)
2132 iterator->level = 0;
2136 static void shadow_walk_init(struct kvm_shadow_walk_iterator *iterator,
2137 struct kvm_vcpu *vcpu, u64 addr)
2139 shadow_walk_init_using_root(iterator, vcpu, vcpu->arch.mmu->root_hpa,
2143 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator *iterator)
2145 if (iterator->level < PG_LEVEL_4K)
2148 iterator->index = SHADOW_PT_INDEX(iterator->addr, iterator->level);
2149 iterator->sptep = ((u64 *)__va(iterator->shadow_addr)) + iterator->index;
2153 static void __shadow_walk_next(struct kvm_shadow_walk_iterator *iterator,
2156 if (is_last_spte(spte, iterator->level)) {
2157 iterator->level = 0;
2161 iterator->shadow_addr = spte & PT64_BASE_ADDR_MASK;
2165 static void shadow_walk_next(struct kvm_shadow_walk_iterator *iterator)
2167 __shadow_walk_next(iterator, *iterator->sptep);
2170 static void link_shadow_page(struct kvm_vcpu *vcpu, u64 *sptep,
2171 struct kvm_mmu_page *sp)
2175 BUILD_BUG_ON(VMX_EPT_WRITABLE_MASK != PT_WRITABLE_MASK);
2177 spte = make_nonleaf_spte(sp->spt, sp_ad_disabled(sp));
2179 mmu_spte_set(sptep, spte);
2181 mmu_page_add_parent_pte(vcpu, sp, sptep);
2183 if (sp->unsync_children || sp->unsync)
2187 static void validate_direct_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2188 unsigned direct_access)
2190 if (is_shadow_present_pte(*sptep) && !is_large_pte(*sptep)) {
2191 struct kvm_mmu_page *child;
2194 * For the direct sp, if the guest pte's dirty bit
2195 * changed form clean to dirty, it will corrupt the
2196 * sp's access: allow writable in the read-only sp,
2197 * so we should update the spte at this point to get
2198 * a new sp with the correct access.
2200 child = to_shadow_page(*sptep & PT64_BASE_ADDR_MASK);
2201 if (child->role.access == direct_access)
2204 drop_parent_pte(child, sptep);
2205 kvm_flush_remote_tlbs_with_address(vcpu->kvm, child->gfn, 1);
2209 /* Returns the number of zapped non-leaf child shadow pages. */
2210 static int mmu_page_zap_pte(struct kvm *kvm, struct kvm_mmu_page *sp,
2211 u64 *spte, struct list_head *invalid_list)
2214 struct kvm_mmu_page *child;
2217 if (is_shadow_present_pte(pte)) {
2218 if (is_last_spte(pte, sp->role.level)) {
2219 drop_spte(kvm, spte);
2220 if (is_large_pte(pte))
2223 child = to_shadow_page(pte & PT64_BASE_ADDR_MASK);
2224 drop_parent_pte(child, spte);
2227 * Recursively zap nested TDP SPs, parentless SPs are
2228 * unlikely to be used again in the near future. This
2229 * avoids retaining a large number of stale nested SPs.
2231 if (tdp_enabled && invalid_list &&
2232 child->role.guest_mode && !child->parent_ptes.val)
2233 return kvm_mmu_prepare_zap_page(kvm, child,
2236 } else if (is_mmio_spte(pte)) {
2237 mmu_spte_clear_no_track(spte);
2242 static int kvm_mmu_page_unlink_children(struct kvm *kvm,
2243 struct kvm_mmu_page *sp,
2244 struct list_head *invalid_list)
2249 for (i = 0; i < PT64_ENT_PER_PAGE; ++i)
2250 zapped += mmu_page_zap_pte(kvm, sp, sp->spt + i, invalid_list);
2255 static void kvm_mmu_unlink_parents(struct kvm *kvm, struct kvm_mmu_page *sp)
2258 struct rmap_iterator iter;
2260 while ((sptep = rmap_get_first(&sp->parent_ptes, &iter)))
2261 drop_parent_pte(sp, sptep);
2264 static int mmu_zap_unsync_children(struct kvm *kvm,
2265 struct kvm_mmu_page *parent,
2266 struct list_head *invalid_list)
2269 struct mmu_page_path parents;
2270 struct kvm_mmu_pages pages;
2272 if (parent->role.level == PG_LEVEL_4K)
2275 while (mmu_unsync_walk(parent, &pages)) {
2276 struct kvm_mmu_page *sp;
2278 for_each_sp(pages, sp, parents, i) {
2279 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
2280 mmu_pages_clear_parents(&parents);
2288 static bool __kvm_mmu_prepare_zap_page(struct kvm *kvm,
2289 struct kvm_mmu_page *sp,
2290 struct list_head *invalid_list,
2295 trace_kvm_mmu_prepare_zap_page(sp);
2296 ++kvm->stat.mmu_shadow_zapped;
2297 *nr_zapped = mmu_zap_unsync_children(kvm, sp, invalid_list);
2298 *nr_zapped += kvm_mmu_page_unlink_children(kvm, sp, invalid_list);
2299 kvm_mmu_unlink_parents(kvm, sp);
2301 /* Zapping children means active_mmu_pages has become unstable. */
2302 list_unstable = *nr_zapped;
2304 if (!sp->role.invalid && !sp->role.direct)
2305 unaccount_shadowed(kvm, sp);
2308 kvm_unlink_unsync_page(kvm, sp);
2309 if (!sp->root_count) {
2314 * Already invalid pages (previously active roots) are not on
2315 * the active page list. See list_del() in the "else" case of
2318 if (sp->role.invalid)
2319 list_add(&sp->link, invalid_list);
2321 list_move(&sp->link, invalid_list);
2322 kvm_mod_used_mmu_pages(kvm, -1);
2325 * Remove the active root from the active page list, the root
2326 * will be explicitly freed when the root_count hits zero.
2328 list_del(&sp->link);
2331 * Obsolete pages cannot be used on any vCPUs, see the comment
2332 * in kvm_mmu_zap_all_fast(). Note, is_obsolete_sp() also
2333 * treats invalid shadow pages as being obsolete.
2335 if (!is_obsolete_sp(kvm, sp))
2336 kvm_reload_remote_mmus(kvm);
2339 if (sp->lpage_disallowed)
2340 unaccount_huge_nx_page(kvm, sp);
2342 sp->role.invalid = 1;
2343 return list_unstable;
2346 static bool kvm_mmu_prepare_zap_page(struct kvm *kvm, struct kvm_mmu_page *sp,
2347 struct list_head *invalid_list)
2351 __kvm_mmu_prepare_zap_page(kvm, sp, invalid_list, &nr_zapped);
2355 static void kvm_mmu_commit_zap_page(struct kvm *kvm,
2356 struct list_head *invalid_list)
2358 struct kvm_mmu_page *sp, *nsp;
2360 if (list_empty(invalid_list))
2364 * We need to make sure everyone sees our modifications to
2365 * the page tables and see changes to vcpu->mode here. The barrier
2366 * in the kvm_flush_remote_tlbs() achieves this. This pairs
2367 * with vcpu_enter_guest and walk_shadow_page_lockless_begin/end.
2369 * In addition, kvm_flush_remote_tlbs waits for all vcpus to exit
2370 * guest mode and/or lockless shadow page table walks.
2372 kvm_flush_remote_tlbs(kvm);
2374 list_for_each_entry_safe(sp, nsp, invalid_list, link) {
2375 WARN_ON(!sp->role.invalid || sp->root_count);
2376 kvm_mmu_free_page(sp);
2380 static unsigned long kvm_mmu_zap_oldest_mmu_pages(struct kvm *kvm,
2381 unsigned long nr_to_zap)
2383 unsigned long total_zapped = 0;
2384 struct kvm_mmu_page *sp, *tmp;
2385 LIST_HEAD(invalid_list);
2389 if (list_empty(&kvm->arch.active_mmu_pages))
2393 list_for_each_entry_safe_reverse(sp, tmp, &kvm->arch.active_mmu_pages, link) {
2395 * Don't zap active root pages, the page itself can't be freed
2396 * and zapping it will just force vCPUs to realloc and reload.
2401 unstable = __kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list,
2403 total_zapped += nr_zapped;
2404 if (total_zapped >= nr_to_zap)
2411 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2413 kvm->stat.mmu_recycled += total_zapped;
2414 return total_zapped;
2417 static inline unsigned long kvm_mmu_available_pages(struct kvm *kvm)
2419 if (kvm->arch.n_max_mmu_pages > kvm->arch.n_used_mmu_pages)
2420 return kvm->arch.n_max_mmu_pages -
2421 kvm->arch.n_used_mmu_pages;
2426 static int make_mmu_pages_available(struct kvm_vcpu *vcpu)
2428 unsigned long avail = kvm_mmu_available_pages(vcpu->kvm);
2430 if (likely(avail >= KVM_MIN_FREE_MMU_PAGES))
2433 kvm_mmu_zap_oldest_mmu_pages(vcpu->kvm, KVM_REFILL_PAGES - avail);
2436 * Note, this check is intentionally soft, it only guarantees that one
2437 * page is available, while the caller may end up allocating as many as
2438 * four pages, e.g. for PAE roots or for 5-level paging. Temporarily
2439 * exceeding the (arbitrary by default) limit will not harm the host,
2440 * being too agressive may unnecessarily kill the guest, and getting an
2441 * exact count is far more trouble than it's worth, especially in the
2444 if (!kvm_mmu_available_pages(vcpu->kvm))
2450 * Changing the number of mmu pages allocated to the vm
2451 * Note: if goal_nr_mmu_pages is too small, you will get dead lock
2453 void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned long goal_nr_mmu_pages)
2455 write_lock(&kvm->mmu_lock);
2457 if (kvm->arch.n_used_mmu_pages > goal_nr_mmu_pages) {
2458 kvm_mmu_zap_oldest_mmu_pages(kvm, kvm->arch.n_used_mmu_pages -
2461 goal_nr_mmu_pages = kvm->arch.n_used_mmu_pages;
2464 kvm->arch.n_max_mmu_pages = goal_nr_mmu_pages;
2466 write_unlock(&kvm->mmu_lock);
2469 int kvm_mmu_unprotect_page(struct kvm *kvm, gfn_t gfn)
2471 struct kvm_mmu_page *sp;
2472 LIST_HEAD(invalid_list);
2475 pgprintk("%s: looking for gfn %llx\n", __func__, gfn);
2477 write_lock(&kvm->mmu_lock);
2478 for_each_gfn_indirect_valid_sp(kvm, sp, gfn) {
2479 pgprintk("%s: gfn %llx role %x\n", __func__, gfn,
2482 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
2484 kvm_mmu_commit_zap_page(kvm, &invalid_list);
2485 write_unlock(&kvm->mmu_lock);
2490 static int kvm_mmu_unprotect_page_virt(struct kvm_vcpu *vcpu, gva_t gva)
2495 if (vcpu->arch.mmu->direct_map)
2498 gpa = kvm_mmu_gva_to_gpa_read(vcpu, gva, NULL);
2500 r = kvm_mmu_unprotect_page(vcpu->kvm, gpa >> PAGE_SHIFT);
2505 static void kvm_unsync_page(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
2507 trace_kvm_mmu_unsync_page(sp);
2508 ++vcpu->kvm->stat.mmu_unsync;
2511 kvm_mmu_mark_parents_unsync(sp);
2515 * Attempt to unsync any shadow pages that can be reached by the specified gfn,
2516 * KVM is creating a writable mapping for said gfn. Returns 0 if all pages
2517 * were marked unsync (or if there is no shadow page), -EPERM if the SPTE must
2518 * be write-protected.
2520 int mmu_try_to_unsync_pages(struct kvm_vcpu *vcpu, gfn_t gfn, bool can_unsync)
2522 struct kvm_mmu_page *sp;
2525 * Force write-protection if the page is being tracked. Note, the page
2526 * track machinery is used to write-protect upper-level shadow pages,
2527 * i.e. this guards the role.level == 4K assertion below!
2529 if (kvm_page_track_is_active(vcpu, gfn, KVM_PAGE_TRACK_WRITE))
2533 * The page is not write-tracked, mark existing shadow pages unsync
2534 * unless KVM is synchronizing an unsync SP (can_unsync = false). In
2535 * that case, KVM must complete emulation of the guest TLB flush before
2536 * allowing shadow pages to become unsync (writable by the guest).
2538 for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn) {
2545 WARN_ON(sp->role.level != PG_LEVEL_4K);
2546 kvm_unsync_page(vcpu, sp);
2550 * We need to ensure that the marking of unsync pages is visible
2551 * before the SPTE is updated to allow writes because
2552 * kvm_mmu_sync_roots() checks the unsync flags without holding
2553 * the MMU lock and so can race with this. If the SPTE was updated
2554 * before the page had been marked as unsync-ed, something like the
2555 * following could happen:
2558 * ---------------------------------------------------------------------
2559 * 1.2 Host updates SPTE
2561 * 2.1 Guest writes a GPTE for GVA X.
2562 * (GPTE being in the guest page table shadowed
2563 * by the SP from CPU 1.)
2564 * This reads SPTE during the page table walk.
2565 * Since SPTE.W is read as 1, there is no
2568 * 2.2 Guest issues TLB flush.
2569 * That causes a VM Exit.
2571 * 2.3 Walking of unsync pages sees sp->unsync is
2572 * false and skips the page.
2574 * 2.4 Guest accesses GVA X.
2575 * Since the mapping in the SP was not updated,
2576 * so the old mapping for GVA X incorrectly
2580 * (sp->unsync = true)
2582 * The write barrier below ensures that 1.1 happens before 1.2 and thus
2583 * the situation in 2.4 does not arise. The implicit barrier in 2.2
2584 * pairs with this write barrier.
2591 static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2592 unsigned int pte_access, int level,
2593 gfn_t gfn, kvm_pfn_t pfn, bool speculative,
2594 bool can_unsync, bool host_writable)
2597 struct kvm_mmu_page *sp;
2600 sp = sptep_to_sp(sptep);
2602 ret = make_spte(vcpu, pte_access, level, gfn, pfn, *sptep, speculative,
2603 can_unsync, host_writable, sp_ad_disabled(sp), &spte);
2605 if (spte & PT_WRITABLE_MASK)
2606 kvm_vcpu_mark_page_dirty(vcpu, gfn);
2609 ret |= SET_SPTE_SPURIOUS;
2610 else if (mmu_spte_update(sptep, spte))
2611 ret |= SET_SPTE_NEED_REMOTE_TLB_FLUSH;
2615 static int mmu_set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
2616 unsigned int pte_access, bool write_fault, int level,
2617 gfn_t gfn, kvm_pfn_t pfn, bool speculative,
2620 int was_rmapped = 0;
2623 int ret = RET_PF_FIXED;
2626 pgprintk("%s: spte %llx write_fault %d gfn %llx\n", __func__,
2627 *sptep, write_fault, gfn);
2629 if (unlikely(is_noslot_pfn(pfn))) {
2630 mark_mmio_spte(vcpu, sptep, gfn, pte_access);
2631 return RET_PF_EMULATE;
2634 if (is_shadow_present_pte(*sptep)) {
2636 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
2637 * the parent of the now unreachable PTE.
2639 if (level > PG_LEVEL_4K && !is_large_pte(*sptep)) {
2640 struct kvm_mmu_page *child;
2643 child = to_shadow_page(pte & PT64_BASE_ADDR_MASK);
2644 drop_parent_pte(child, sptep);
2646 } else if (pfn != spte_to_pfn(*sptep)) {
2647 pgprintk("hfn old %llx new %llx\n",
2648 spte_to_pfn(*sptep), pfn);
2649 drop_spte(vcpu->kvm, sptep);
2655 set_spte_ret = set_spte(vcpu, sptep, pte_access, level, gfn, pfn,
2656 speculative, true, host_writable);
2657 if (set_spte_ret & SET_SPTE_WRITE_PROTECTED_PT) {
2659 ret = RET_PF_EMULATE;
2660 kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu);
2663 if (set_spte_ret & SET_SPTE_NEED_REMOTE_TLB_FLUSH || flush)
2664 kvm_flush_remote_tlbs_with_address(vcpu->kvm, gfn,
2665 KVM_PAGES_PER_HPAGE(level));
2668 * The fault is fully spurious if and only if the new SPTE and old SPTE
2669 * are identical, and emulation is not required.
2671 if ((set_spte_ret & SET_SPTE_SPURIOUS) && ret == RET_PF_FIXED) {
2672 WARN_ON_ONCE(!was_rmapped);
2673 return RET_PF_SPURIOUS;
2676 pgprintk("%s: setting spte %llx\n", __func__, *sptep);
2677 trace_kvm_mmu_set_spte(level, gfn, sptep);
2678 if (!was_rmapped && is_large_pte(*sptep))
2679 ++vcpu->kvm->stat.lpages;
2681 if (is_shadow_present_pte(*sptep)) {
2683 rmap_count = rmap_add(vcpu, sptep, gfn);
2684 if (rmap_count > RMAP_RECYCLE_THRESHOLD)
2685 rmap_recycle(vcpu, sptep, gfn);
2692 static kvm_pfn_t pte_prefetch_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn,
2695 struct kvm_memory_slot *slot;
2697 slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, no_dirty_log);
2699 return KVM_PFN_ERR_FAULT;
2701 return gfn_to_pfn_memslot_atomic(slot, gfn);
2704 static int direct_pte_prefetch_many(struct kvm_vcpu *vcpu,
2705 struct kvm_mmu_page *sp,
2706 u64 *start, u64 *end)
2708 struct page *pages[PTE_PREFETCH_NUM];
2709 struct kvm_memory_slot *slot;
2710 unsigned int access = sp->role.access;
2714 gfn = kvm_mmu_page_get_gfn(sp, start - sp->spt);
2715 slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, access & ACC_WRITE_MASK);
2719 ret = gfn_to_page_many_atomic(slot, gfn, pages, end - start);
2723 for (i = 0; i < ret; i++, gfn++, start++) {
2724 mmu_set_spte(vcpu, start, access, false, sp->role.level, gfn,
2725 page_to_pfn(pages[i]), true, true);
2732 static void __direct_pte_prefetch(struct kvm_vcpu *vcpu,
2733 struct kvm_mmu_page *sp, u64 *sptep)
2735 u64 *spte, *start = NULL;
2738 WARN_ON(!sp->role.direct);
2740 i = (sptep - sp->spt) & ~(PTE_PREFETCH_NUM - 1);
2743 for (i = 0; i < PTE_PREFETCH_NUM; i++, spte++) {
2744 if (is_shadow_present_pte(*spte) || spte == sptep) {
2747 if (direct_pte_prefetch_many(vcpu, sp, start, spte) < 0)
2755 static void direct_pte_prefetch(struct kvm_vcpu *vcpu, u64 *sptep)
2757 struct kvm_mmu_page *sp;
2759 sp = sptep_to_sp(sptep);
2762 * Without accessed bits, there's no way to distinguish between
2763 * actually accessed translations and prefetched, so disable pte
2764 * prefetch if accessed bits aren't available.
2766 if (sp_ad_disabled(sp))
2769 if (sp->role.level > PG_LEVEL_4K)
2773 * If addresses are being invalidated, skip prefetching to avoid
2774 * accidentally prefetching those addresses.
2776 if (unlikely(vcpu->kvm->mmu_notifier_count))
2779 __direct_pte_prefetch(vcpu, sp, sptep);
2782 static int host_pfn_mapping_level(struct kvm *kvm, gfn_t gfn, kvm_pfn_t pfn,
2783 const struct kvm_memory_slot *slot)
2789 if (!PageCompound(pfn_to_page(pfn)) && !kvm_is_zone_device_pfn(pfn))
2793 * Note, using the already-retrieved memslot and __gfn_to_hva_memslot()
2794 * is not solely for performance, it's also necessary to avoid the
2795 * "writable" check in __gfn_to_hva_many(), which will always fail on
2796 * read-only memslots due to gfn_to_hva() assuming writes. Earlier
2797 * page fault steps have already verified the guest isn't writing a
2798 * read-only memslot.
2800 hva = __gfn_to_hva_memslot(slot, gfn);
2802 pte = lookup_address_in_mm(kvm->mm, hva, &level);
2809 int kvm_mmu_max_mapping_level(struct kvm *kvm,
2810 const struct kvm_memory_slot *slot, gfn_t gfn,
2811 kvm_pfn_t pfn, int max_level)
2813 struct kvm_lpage_info *linfo;
2815 max_level = min(max_level, max_huge_page_level);
2816 for ( ; max_level > PG_LEVEL_4K; max_level--) {
2817 linfo = lpage_info_slot(gfn, slot, max_level);
2818 if (!linfo->disallow_lpage)
2822 if (max_level == PG_LEVEL_4K)
2825 return host_pfn_mapping_level(kvm, gfn, pfn, slot);
2828 int kvm_mmu_hugepage_adjust(struct kvm_vcpu *vcpu, gfn_t gfn,
2829 int max_level, kvm_pfn_t *pfnp,
2830 bool huge_page_disallowed, int *req_level)
2832 struct kvm_memory_slot *slot;
2833 kvm_pfn_t pfn = *pfnp;
2837 *req_level = PG_LEVEL_4K;
2839 if (unlikely(max_level == PG_LEVEL_4K))
2842 if (is_error_noslot_pfn(pfn) || kvm_is_reserved_pfn(pfn))
2845 slot = gfn_to_memslot_dirty_bitmap(vcpu, gfn, true);
2849 level = kvm_mmu_max_mapping_level(vcpu->kvm, slot, gfn, pfn, max_level);
2850 if (level == PG_LEVEL_4K)
2853 *req_level = level = min(level, max_level);
2856 * Enforce the iTLB multihit workaround after capturing the requested
2857 * level, which will be used to do precise, accurate accounting.
2859 if (huge_page_disallowed)
2863 * mmu_notifier_retry() was successful and mmu_lock is held, so
2864 * the pmd can't be split from under us.
2866 mask = KVM_PAGES_PER_HPAGE(level) - 1;
2867 VM_BUG_ON((gfn & mask) != (pfn & mask));
2868 *pfnp = pfn & ~mask;
2873 void disallowed_hugepage_adjust(u64 spte, gfn_t gfn, int cur_level,
2874 kvm_pfn_t *pfnp, int *goal_levelp)
2876 int level = *goal_levelp;
2878 if (cur_level == level && level > PG_LEVEL_4K &&
2879 is_shadow_present_pte(spte) &&
2880 !is_large_pte(spte)) {
2882 * A small SPTE exists for this pfn, but FNAME(fetch)
2883 * and __direct_map would like to create a large PTE
2884 * instead: just force them to go down another level,
2885 * patching back for them into pfn the next 9 bits of
2888 u64 page_mask = KVM_PAGES_PER_HPAGE(level) -
2889 KVM_PAGES_PER_HPAGE(level - 1);
2890 *pfnp |= gfn & page_mask;
2895 static int __direct_map(struct kvm_vcpu *vcpu, gpa_t gpa, u32 error_code,
2896 int map_writable, int max_level, kvm_pfn_t pfn,
2897 bool prefault, bool is_tdp)
2899 bool nx_huge_page_workaround_enabled = is_nx_huge_page_enabled();
2900 bool write = error_code & PFERR_WRITE_MASK;
2901 bool exec = error_code & PFERR_FETCH_MASK;
2902 bool huge_page_disallowed = exec && nx_huge_page_workaround_enabled;
2903 struct kvm_shadow_walk_iterator it;
2904 struct kvm_mmu_page *sp;
2905 int level, req_level, ret;
2906 gfn_t gfn = gpa >> PAGE_SHIFT;
2907 gfn_t base_gfn = gfn;
2909 level = kvm_mmu_hugepage_adjust(vcpu, gfn, max_level, &pfn,
2910 huge_page_disallowed, &req_level);
2912 trace_kvm_mmu_spte_requested(gpa, level, pfn);
2913 for_each_shadow_entry(vcpu, gpa, it) {
2915 * We cannot overwrite existing page tables with an NX
2916 * large page, as the leaf could be executable.
2918 if (nx_huge_page_workaround_enabled)
2919 disallowed_hugepage_adjust(*it.sptep, gfn, it.level,
2922 base_gfn = gfn & ~(KVM_PAGES_PER_HPAGE(it.level) - 1);
2923 if (it.level == level)
2926 drop_large_spte(vcpu, it.sptep);
2927 if (!is_shadow_present_pte(*it.sptep)) {
2928 sp = kvm_mmu_get_page(vcpu, base_gfn, it.addr,
2929 it.level - 1, true, ACC_ALL);
2931 link_shadow_page(vcpu, it.sptep, sp);
2932 if (is_tdp && huge_page_disallowed &&
2933 req_level >= it.level)
2934 account_huge_nx_page(vcpu->kvm, sp);
2938 ret = mmu_set_spte(vcpu, it.sptep, ACC_ALL,
2939 write, level, base_gfn, pfn, prefault,
2941 if (ret == RET_PF_SPURIOUS)
2944 direct_pte_prefetch(vcpu, it.sptep);
2945 ++vcpu->stat.pf_fixed;
2949 static void kvm_send_hwpoison_signal(unsigned long address, struct task_struct *tsk)
2951 send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, PAGE_SHIFT, tsk);
2954 static int kvm_handle_bad_page(struct kvm_vcpu *vcpu, gfn_t gfn, kvm_pfn_t pfn)
2957 * Do not cache the mmio info caused by writing the readonly gfn
2958 * into the spte otherwise read access on readonly gfn also can
2959 * caused mmio page fault and treat it as mmio access.
2961 if (pfn == KVM_PFN_ERR_RO_FAULT)
2962 return RET_PF_EMULATE;
2964 if (pfn == KVM_PFN_ERR_HWPOISON) {
2965 kvm_send_hwpoison_signal(kvm_vcpu_gfn_to_hva(vcpu, gfn), current);
2966 return RET_PF_RETRY;
2972 static bool handle_abnormal_pfn(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn,
2973 kvm_pfn_t pfn, unsigned int access,
2976 /* The pfn is invalid, report the error! */
2977 if (unlikely(is_error_pfn(pfn))) {
2978 *ret_val = kvm_handle_bad_page(vcpu, gfn, pfn);
2982 if (unlikely(is_noslot_pfn(pfn))) {
2983 vcpu_cache_mmio_info(vcpu, gva, gfn,
2984 access & shadow_mmio_access_mask);
2986 * If MMIO caching is disabled, emulate immediately without
2987 * touching the shadow page tables as attempting to install an
2988 * MMIO SPTE will just be an expensive nop.
2990 if (unlikely(!shadow_mmio_value)) {
2991 *ret_val = RET_PF_EMULATE;
2999 static bool page_fault_can_be_fast(u32 error_code)
3002 * Do not fix the mmio spte with invalid generation number which
3003 * need to be updated by slow page fault path.
3005 if (unlikely(error_code & PFERR_RSVD_MASK))
3008 /* See if the page fault is due to an NX violation */
3009 if (unlikely(((error_code & (PFERR_FETCH_MASK | PFERR_PRESENT_MASK))
3010 == (PFERR_FETCH_MASK | PFERR_PRESENT_MASK))))
3014 * #PF can be fast if:
3015 * 1. The shadow page table entry is not present, which could mean that
3016 * the fault is potentially caused by access tracking (if enabled).
3017 * 2. The shadow page table entry is present and the fault
3018 * is caused by write-protect, that means we just need change the W
3019 * bit of the spte which can be done out of mmu-lock.
3021 * However, if access tracking is disabled we know that a non-present
3022 * page must be a genuine page fault where we have to create a new SPTE.
3023 * So, if access tracking is disabled, we return true only for write
3024 * accesses to a present page.
3027 return shadow_acc_track_mask != 0 ||
3028 ((error_code & (PFERR_WRITE_MASK | PFERR_PRESENT_MASK))
3029 == (PFERR_WRITE_MASK | PFERR_PRESENT_MASK));
3033 * Returns true if the SPTE was fixed successfully. Otherwise,
3034 * someone else modified the SPTE from its original value.
3037 fast_pf_fix_direct_spte(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
3038 u64 *sptep, u64 old_spte, u64 new_spte)
3042 WARN_ON(!sp->role.direct);
3045 * Theoretically we could also set dirty bit (and flush TLB) here in
3046 * order to eliminate unnecessary PML logging. See comments in
3047 * set_spte. But fast_page_fault is very unlikely to happen with PML
3048 * enabled, so we do not do this. This might result in the same GPA
3049 * to be logged in PML buffer again when the write really happens, and
3050 * eventually to be called by mark_page_dirty twice. But it's also no
3051 * harm. This also avoids the TLB flush needed after setting dirty bit
3052 * so non-PML cases won't be impacted.
3054 * Compare with set_spte where instead shadow_dirty_mask is set.
3056 if (cmpxchg64(sptep, old_spte, new_spte) != old_spte)
3059 if (is_writable_pte(new_spte) && !is_writable_pte(old_spte)) {
3061 * The gfn of direct spte is stable since it is
3062 * calculated by sp->gfn.
3064 gfn = kvm_mmu_page_get_gfn(sp, sptep - sp->spt);
3065 kvm_vcpu_mark_page_dirty(vcpu, gfn);
3071 static bool is_access_allowed(u32 fault_err_code, u64 spte)
3073 if (fault_err_code & PFERR_FETCH_MASK)
3074 return is_executable_pte(spte);
3076 if (fault_err_code & PFERR_WRITE_MASK)
3077 return is_writable_pte(spte);
3079 /* Fault was on Read access */
3080 return spte & PT_PRESENT_MASK;
3084 * Returns one of RET_PF_INVALID, RET_PF_FIXED or RET_PF_SPURIOUS.
3086 static int fast_page_fault(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa,
3089 struct kvm_shadow_walk_iterator iterator;
3090 struct kvm_mmu_page *sp;
3091 int ret = RET_PF_INVALID;
3093 uint retry_count = 0;
3095 if (!page_fault_can_be_fast(error_code))
3098 walk_shadow_page_lockless_begin(vcpu);
3103 for_each_shadow_entry_lockless(vcpu, cr2_or_gpa, iterator, spte)
3104 if (!is_shadow_present_pte(spte))
3107 if (!is_shadow_present_pte(spte))
3110 sp = sptep_to_sp(iterator.sptep);
3111 if (!is_last_spte(spte, sp->role.level))
3115 * Check whether the memory access that caused the fault would
3116 * still cause it if it were to be performed right now. If not,
3117 * then this is a spurious fault caused by TLB lazily flushed,
3118 * or some other CPU has already fixed the PTE after the
3119 * current CPU took the fault.
3121 * Need not check the access of upper level table entries since
3122 * they are always ACC_ALL.
3124 if (is_access_allowed(error_code, spte)) {
3125 ret = RET_PF_SPURIOUS;
3131 if (is_access_track_spte(spte))
3132 new_spte = restore_acc_track_spte(new_spte);
3135 * Currently, to simplify the code, write-protection can
3136 * be removed in the fast path only if the SPTE was
3137 * write-protected for dirty-logging or access tracking.
3139 if ((error_code & PFERR_WRITE_MASK) &&
3140 spte_can_locklessly_be_made_writable(spte)) {
3141 new_spte |= PT_WRITABLE_MASK;
3144 * Do not fix write-permission on the large spte. Since
3145 * we only dirty the first page into the dirty-bitmap in
3146 * fast_pf_fix_direct_spte(), other pages are missed
3147 * if its slot has dirty logging enabled.
3149 * Instead, we let the slow page fault path create a
3150 * normal spte to fix the access.
3152 * See the comments in kvm_arch_commit_memory_region().
3154 if (sp->role.level > PG_LEVEL_4K)
3158 /* Verify that the fault can be handled in the fast path */
3159 if (new_spte == spte ||
3160 !is_access_allowed(error_code, new_spte))
3164 * Currently, fast page fault only works for direct mapping
3165 * since the gfn is not stable for indirect shadow page. See
3166 * Documentation/virt/kvm/locking.rst to get more detail.
3168 if (fast_pf_fix_direct_spte(vcpu, sp, iterator.sptep, spte,
3174 if (++retry_count > 4) {
3175 printk_once(KERN_WARNING
3176 "kvm: Fast #PF retrying more than 4 times.\n");
3182 trace_fast_page_fault(vcpu, cr2_or_gpa, error_code, iterator.sptep,
3184 walk_shadow_page_lockless_end(vcpu);
3189 static void mmu_free_root_page(struct kvm *kvm, hpa_t *root_hpa,
3190 struct list_head *invalid_list)
3192 struct kvm_mmu_page *sp;
3194 if (!VALID_PAGE(*root_hpa))
3197 sp = to_shadow_page(*root_hpa & PT64_BASE_ADDR_MASK);
3199 if (is_tdp_mmu_page(sp))
3200 kvm_tdp_mmu_put_root(kvm, sp, false);
3201 else if (!--sp->root_count && sp->role.invalid)
3202 kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
3204 *root_hpa = INVALID_PAGE;
3207 /* roots_to_free must be some combination of the KVM_MMU_ROOT_* flags */
3208 void kvm_mmu_free_roots(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
3209 ulong roots_to_free)
3211 struct kvm *kvm = vcpu->kvm;
3213 LIST_HEAD(invalid_list);
3214 bool free_active_root = roots_to_free & KVM_MMU_ROOT_CURRENT;
3216 BUILD_BUG_ON(KVM_MMU_NUM_PREV_ROOTS >= BITS_PER_LONG);
3218 /* Before acquiring the MMU lock, see if we need to do any real work. */
3219 if (!(free_active_root && VALID_PAGE(mmu->root_hpa))) {
3220 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
3221 if ((roots_to_free & KVM_MMU_ROOT_PREVIOUS(i)) &&
3222 VALID_PAGE(mmu->prev_roots[i].hpa))
3225 if (i == KVM_MMU_NUM_PREV_ROOTS)
3229 write_lock(&kvm->mmu_lock);
3231 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
3232 if (roots_to_free & KVM_MMU_ROOT_PREVIOUS(i))
3233 mmu_free_root_page(kvm, &mmu->prev_roots[i].hpa,
3236 if (free_active_root) {
3237 if (mmu->shadow_root_level >= PT64_ROOT_4LEVEL &&
3238 (mmu->root_level >= PT64_ROOT_4LEVEL || mmu->direct_map)) {
3239 mmu_free_root_page(kvm, &mmu->root_hpa, &invalid_list);
3240 } else if (mmu->pae_root) {
3241 for (i = 0; i < 4; ++i) {
3242 if (!IS_VALID_PAE_ROOT(mmu->pae_root[i]))
3245 mmu_free_root_page(kvm, &mmu->pae_root[i],
3247 mmu->pae_root[i] = INVALID_PAE_ROOT;
3250 mmu->root_hpa = INVALID_PAGE;
3254 kvm_mmu_commit_zap_page(kvm, &invalid_list);
3255 write_unlock(&kvm->mmu_lock);
3257 EXPORT_SYMBOL_GPL(kvm_mmu_free_roots);
3259 void kvm_mmu_free_guest_mode_roots(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu)
3261 unsigned long roots_to_free = 0;
3266 * This should not be called while L2 is active, L2 can't invalidate
3267 * _only_ its own roots, e.g. INVVPID unconditionally exits.
3269 WARN_ON_ONCE(mmu->mmu_role.base.guest_mode);
3271 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) {
3272 root_hpa = mmu->prev_roots[i].hpa;
3273 if (!VALID_PAGE(root_hpa))
3276 if (!to_shadow_page(root_hpa) ||
3277 to_shadow_page(root_hpa)->role.guest_mode)
3278 roots_to_free |= KVM_MMU_ROOT_PREVIOUS(i);
3281 kvm_mmu_free_roots(vcpu, mmu, roots_to_free);
3283 EXPORT_SYMBOL_GPL(kvm_mmu_free_guest_mode_roots);
3286 static int mmu_check_root(struct kvm_vcpu *vcpu, gfn_t root_gfn)
3290 if (!kvm_vcpu_is_visible_gfn(vcpu, root_gfn)) {
3291 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
3298 static hpa_t mmu_alloc_root(struct kvm_vcpu *vcpu, gfn_t gfn, gva_t gva,
3299 u8 level, bool direct)
3301 struct kvm_mmu_page *sp;
3303 sp = kvm_mmu_get_page(vcpu, gfn, gva, level, direct, ACC_ALL);
3306 return __pa(sp->spt);
3309 static int mmu_alloc_direct_roots(struct kvm_vcpu *vcpu)
3311 struct kvm_mmu *mmu = vcpu->arch.mmu;
3312 u8 shadow_root_level = mmu->shadow_root_level;
3317 write_lock(&vcpu->kvm->mmu_lock);
3318 r = make_mmu_pages_available(vcpu);
3322 if (is_tdp_mmu_enabled(vcpu->kvm)) {
3323 root = kvm_tdp_mmu_get_vcpu_root_hpa(vcpu);
3324 mmu->root_hpa = root;
3325 } else if (shadow_root_level >= PT64_ROOT_4LEVEL) {
3326 root = mmu_alloc_root(vcpu, 0, 0, shadow_root_level, true);
3327 mmu->root_hpa = root;
3328 } else if (shadow_root_level == PT32E_ROOT_LEVEL) {
3329 if (WARN_ON_ONCE(!mmu->pae_root)) {
3334 for (i = 0; i < 4; ++i) {
3335 WARN_ON_ONCE(IS_VALID_PAE_ROOT(mmu->pae_root[i]));
3337 root = mmu_alloc_root(vcpu, i << (30 - PAGE_SHIFT),
3338 i << 30, PT32_ROOT_LEVEL, true);
3339 mmu->pae_root[i] = root | PT_PRESENT_MASK |
3342 mmu->root_hpa = __pa(mmu->pae_root);
3344 WARN_ONCE(1, "Bad TDP root level = %d\n", shadow_root_level);
3349 /* root_pgd is ignored for direct MMUs. */
3352 write_unlock(&vcpu->kvm->mmu_lock);
3356 static int mmu_alloc_shadow_roots(struct kvm_vcpu *vcpu)
3358 struct kvm_mmu *mmu = vcpu->arch.mmu;
3359 u64 pdptrs[4], pm_mask;
3360 gfn_t root_gfn, root_pgd;
3365 root_pgd = mmu->get_guest_pgd(vcpu);
3366 root_gfn = root_pgd >> PAGE_SHIFT;
3368 if (mmu_check_root(vcpu, root_gfn))
3372 * On SVM, reading PDPTRs might access guest memory, which might fault
3373 * and thus might sleep. Grab the PDPTRs before acquiring mmu_lock.
3375 if (mmu->root_level == PT32E_ROOT_LEVEL) {
3376 for (i = 0; i < 4; ++i) {
3377 pdptrs[i] = mmu->get_pdptr(vcpu, i);
3378 if (!(pdptrs[i] & PT_PRESENT_MASK))
3381 if (mmu_check_root(vcpu, pdptrs[i] >> PAGE_SHIFT))
3386 r = alloc_all_memslots_rmaps(vcpu->kvm);
3390 write_lock(&vcpu->kvm->mmu_lock);
3391 r = make_mmu_pages_available(vcpu);
3396 * Do we shadow a long mode page table? If so we need to
3397 * write-protect the guests page table root.
3399 if (mmu->root_level >= PT64_ROOT_4LEVEL) {
3400 root = mmu_alloc_root(vcpu, root_gfn, 0,
3401 mmu->shadow_root_level, false);
3402 mmu->root_hpa = root;
3406 if (WARN_ON_ONCE(!mmu->pae_root)) {
3412 * We shadow a 32 bit page table. This may be a legacy 2-level
3413 * or a PAE 3-level page table. In either case we need to be aware that
3414 * the shadow page table may be a PAE or a long mode page table.
3416 pm_mask = PT_PRESENT_MASK | shadow_me_mask;
3417 if (mmu->shadow_root_level == PT64_ROOT_4LEVEL) {
3418 pm_mask |= PT_ACCESSED_MASK | PT_WRITABLE_MASK | PT_USER_MASK;
3420 if (WARN_ON_ONCE(!mmu->pml4_root)) {
3425 mmu->pml4_root[0] = __pa(mmu->pae_root) | pm_mask;
3428 for (i = 0; i < 4; ++i) {
3429 WARN_ON_ONCE(IS_VALID_PAE_ROOT(mmu->pae_root[i]));
3431 if (mmu->root_level == PT32E_ROOT_LEVEL) {
3432 if (!(pdptrs[i] & PT_PRESENT_MASK)) {
3433 mmu->pae_root[i] = INVALID_PAE_ROOT;
3436 root_gfn = pdptrs[i] >> PAGE_SHIFT;
3439 root = mmu_alloc_root(vcpu, root_gfn, i << 30,
3440 PT32_ROOT_LEVEL, false);
3441 mmu->pae_root[i] = root | pm_mask;
3444 if (mmu->shadow_root_level == PT64_ROOT_4LEVEL)
3445 mmu->root_hpa = __pa(mmu->pml4_root);
3447 mmu->root_hpa = __pa(mmu->pae_root);
3450 mmu->root_pgd = root_pgd;
3452 write_unlock(&vcpu->kvm->mmu_lock);
3457 static int mmu_alloc_special_roots(struct kvm_vcpu *vcpu)
3459 struct kvm_mmu *mmu = vcpu->arch.mmu;
3460 u64 *pml4_root, *pae_root;
3463 * When shadowing 32-bit or PAE NPT with 64-bit NPT, the PML4 and PDP
3464 * tables are allocated and initialized at root creation as there is no
3465 * equivalent level in the guest's NPT to shadow. Allocate the tables
3466 * on demand, as running a 32-bit L1 VMM on 64-bit KVM is very rare.
3468 if (mmu->direct_map || mmu->root_level >= PT64_ROOT_4LEVEL ||
3469 mmu->shadow_root_level < PT64_ROOT_4LEVEL)
3473 * This mess only works with 4-level paging and needs to be updated to
3474 * work with 5-level paging.
3476 if (WARN_ON_ONCE(mmu->shadow_root_level != PT64_ROOT_4LEVEL))
3479 if (mmu->pae_root && mmu->pml4_root)
3483 * The special roots should always be allocated in concert. Yell and
3484 * bail if KVM ends up in a state where only one of the roots is valid.
3486 if (WARN_ON_ONCE(!tdp_enabled || mmu->pae_root || mmu->pml4_root))
3490 * Unlike 32-bit NPT, the PDP table doesn't need to be in low mem, and
3491 * doesn't need to be decrypted.
3493 pae_root = (void *)get_zeroed_page(GFP_KERNEL_ACCOUNT);
3497 pml4_root = (void *)get_zeroed_page(GFP_KERNEL_ACCOUNT);
3499 free_page((unsigned long)pae_root);
3503 mmu->pae_root = pae_root;
3504 mmu->pml4_root = pml4_root;
3509 void kvm_mmu_sync_roots(struct kvm_vcpu *vcpu)
3512 struct kvm_mmu_page *sp;
3514 if (vcpu->arch.mmu->direct_map)
3517 if (!VALID_PAGE(vcpu->arch.mmu->root_hpa))
3520 vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY);
3522 if (vcpu->arch.mmu->root_level >= PT64_ROOT_4LEVEL) {
3523 hpa_t root = vcpu->arch.mmu->root_hpa;
3524 sp = to_shadow_page(root);
3527 * Even if another CPU was marking the SP as unsync-ed
3528 * simultaneously, any guest page table changes are not
3529 * guaranteed to be visible anyway until this VCPU issues a TLB
3530 * flush strictly after those changes are made. We only need to
3531 * ensure that the other CPU sets these flags before any actual
3532 * changes to the page tables are made. The comments in
3533 * mmu_try_to_unsync_pages() describe what could go wrong if
3534 * this requirement isn't satisfied.
3536 if (!smp_load_acquire(&sp->unsync) &&
3537 !smp_load_acquire(&sp->unsync_children))
3540 write_lock(&vcpu->kvm->mmu_lock);
3541 kvm_mmu_audit(vcpu, AUDIT_PRE_SYNC);
3543 mmu_sync_children(vcpu, sp);
3545 kvm_mmu_audit(vcpu, AUDIT_POST_SYNC);
3546 write_unlock(&vcpu->kvm->mmu_lock);
3550 write_lock(&vcpu->kvm->mmu_lock);
3551 kvm_mmu_audit(vcpu, AUDIT_PRE_SYNC);
3553 for (i = 0; i < 4; ++i) {
3554 hpa_t root = vcpu->arch.mmu->pae_root[i];
3556 if (IS_VALID_PAE_ROOT(root)) {
3557 root &= PT64_BASE_ADDR_MASK;
3558 sp = to_shadow_page(root);
3559 mmu_sync_children(vcpu, sp);
3563 kvm_mmu_audit(vcpu, AUDIT_POST_SYNC);
3564 write_unlock(&vcpu->kvm->mmu_lock);
3567 static gpa_t nonpaging_gva_to_gpa(struct kvm_vcpu *vcpu, gpa_t vaddr,
3568 u32 access, struct x86_exception *exception)
3571 exception->error_code = 0;
3575 static gpa_t nonpaging_gva_to_gpa_nested(struct kvm_vcpu *vcpu, gpa_t vaddr,
3577 struct x86_exception *exception)
3580 exception->error_code = 0;
3581 return vcpu->arch.nested_mmu.translate_gpa(vcpu, vaddr, access, exception);
3585 __is_rsvd_bits_set(struct rsvd_bits_validate *rsvd_check, u64 pte, int level)
3587 int bit7 = (pte >> 7) & 1;
3589 return pte & rsvd_check->rsvd_bits_mask[bit7][level-1];
3592 static bool __is_bad_mt_xwr(struct rsvd_bits_validate *rsvd_check, u64 pte)
3594 return rsvd_check->bad_mt_xwr & BIT_ULL(pte & 0x3f);
3597 static bool mmio_info_in_cache(struct kvm_vcpu *vcpu, u64 addr, bool direct)
3600 * A nested guest cannot use the MMIO cache if it is using nested
3601 * page tables, because cr2 is a nGPA while the cache stores GPAs.
3603 if (mmu_is_nested(vcpu))
3607 return vcpu_match_mmio_gpa(vcpu, addr);
3609 return vcpu_match_mmio_gva(vcpu, addr);
3613 * Return the level of the lowest level SPTE added to sptes.
3614 * That SPTE may be non-present.
3616 static int get_walk(struct kvm_vcpu *vcpu, u64 addr, u64 *sptes, int *root_level)
3618 struct kvm_shadow_walk_iterator iterator;
3622 walk_shadow_page_lockless_begin(vcpu);
3624 for (shadow_walk_init(&iterator, vcpu, addr),
3625 *root_level = iterator.level;
3626 shadow_walk_okay(&iterator);
3627 __shadow_walk_next(&iterator, spte)) {
3628 leaf = iterator.level;
3629 spte = mmu_spte_get_lockless(iterator.sptep);
3633 if (!is_shadow_present_pte(spte))
3637 walk_shadow_page_lockless_end(vcpu);
3642 /* return true if reserved bit(s) are detected on a valid, non-MMIO SPTE. */
3643 static bool get_mmio_spte(struct kvm_vcpu *vcpu, u64 addr, u64 *sptep)
3645 u64 sptes[PT64_ROOT_MAX_LEVEL + 1];
3646 struct rsvd_bits_validate *rsvd_check;
3647 int root, leaf, level;
3648 bool reserved = false;
3650 if (is_tdp_mmu(vcpu->arch.mmu))
3651 leaf = kvm_tdp_mmu_get_walk(vcpu, addr, sptes, &root);
3653 leaf = get_walk(vcpu, addr, sptes, &root);
3655 if (unlikely(leaf < 0)) {
3660 *sptep = sptes[leaf];
3663 * Skip reserved bits checks on the terminal leaf if it's not a valid
3664 * SPTE. Note, this also (intentionally) skips MMIO SPTEs, which, by
3665 * design, always have reserved bits set. The purpose of the checks is
3666 * to detect reserved bits on non-MMIO SPTEs. i.e. buggy SPTEs.
3668 if (!is_shadow_present_pte(sptes[leaf]))
3671 rsvd_check = &vcpu->arch.mmu->shadow_zero_check;
3673 for (level = root; level >= leaf; level--)
3675 * Use a bitwise-OR instead of a logical-OR to aggregate the
3676 * reserved bit and EPT's invalid memtype/XWR checks to avoid
3677 * adding a Jcc in the loop.
3679 reserved |= __is_bad_mt_xwr(rsvd_check, sptes[level]) |
3680 __is_rsvd_bits_set(rsvd_check, sptes[level], level);
3683 pr_err("%s: reserved bits set on MMU-present spte, addr 0x%llx, hierarchy:\n",
3685 for (level = root; level >= leaf; level--)
3686 pr_err("------ spte = 0x%llx level = %d, rsvd bits = 0x%llx",
3687 sptes[level], level,
3688 rsvd_check->rsvd_bits_mask[(sptes[level] >> 7) & 1][level-1]);
3694 static int handle_mmio_page_fault(struct kvm_vcpu *vcpu, u64 addr, bool direct)
3699 if (mmio_info_in_cache(vcpu, addr, direct))
3700 return RET_PF_EMULATE;
3702 reserved = get_mmio_spte(vcpu, addr, &spte);
3703 if (WARN_ON(reserved))
3706 if (is_mmio_spte(spte)) {
3707 gfn_t gfn = get_mmio_spte_gfn(spte);
3708 unsigned int access = get_mmio_spte_access(spte);
3710 if (!check_mmio_spte(vcpu, spte))
3711 return RET_PF_INVALID;
3716 trace_handle_mmio_page_fault(addr, gfn, access);
3717 vcpu_cache_mmio_info(vcpu, addr, gfn, access);
3718 return RET_PF_EMULATE;
3722 * If the page table is zapped by other cpus, let CPU fault again on
3725 return RET_PF_RETRY;
3728 static bool page_fault_handle_page_track(struct kvm_vcpu *vcpu,
3729 u32 error_code, gfn_t gfn)
3731 if (unlikely(error_code & PFERR_RSVD_MASK))
3734 if (!(error_code & PFERR_PRESENT_MASK) ||
3735 !(error_code & PFERR_WRITE_MASK))
3739 * guest is writing the page which is write tracked which can
3740 * not be fixed by page fault handler.
3742 if (kvm_page_track_is_active(vcpu, gfn, KVM_PAGE_TRACK_WRITE))
3748 static void shadow_page_table_clear_flood(struct kvm_vcpu *vcpu, gva_t addr)
3750 struct kvm_shadow_walk_iterator iterator;
3753 walk_shadow_page_lockless_begin(vcpu);
3754 for_each_shadow_entry_lockless(vcpu, addr, iterator, spte) {
3755 clear_sp_write_flooding_count(iterator.sptep);
3756 if (!is_shadow_present_pte(spte))
3759 walk_shadow_page_lockless_end(vcpu);
3762 static bool kvm_arch_setup_async_pf(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa,
3765 struct kvm_arch_async_pf arch;
3767 arch.token = (vcpu->arch.apf.id++ << 12) | vcpu->vcpu_id;
3769 arch.direct_map = vcpu->arch.mmu->direct_map;
3770 arch.cr3 = vcpu->arch.mmu->get_guest_pgd(vcpu);
3772 return kvm_setup_async_pf(vcpu, cr2_or_gpa,
3773 kvm_vcpu_gfn_to_hva(vcpu, gfn), &arch);
3776 static bool try_async_pf(struct kvm_vcpu *vcpu, bool prefault, gfn_t gfn,
3777 gpa_t cr2_or_gpa, kvm_pfn_t *pfn, hva_t *hva,
3778 bool write, bool *writable)
3780 struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
3784 * Retry the page fault if the gfn hit a memslot that is being deleted
3785 * or moved. This ensures any existing SPTEs for the old memslot will
3786 * be zapped before KVM inserts a new MMIO SPTE for the gfn.
3788 if (slot && (slot->flags & KVM_MEMSLOT_INVALID))
3791 /* Don't expose private memslots to L2. */
3792 if (is_guest_mode(vcpu) && !kvm_is_visible_memslot(slot)) {
3793 *pfn = KVM_PFN_NOSLOT;
3799 *pfn = __gfn_to_pfn_memslot(slot, gfn, false, &async,
3800 write, writable, hva);
3802 return false; /* *pfn has correct page already */
3804 if (!prefault && kvm_can_do_async_pf(vcpu)) {
3805 trace_kvm_try_async_get_page(cr2_or_gpa, gfn);
3806 if (kvm_find_async_pf_gfn(vcpu, gfn)) {
3807 trace_kvm_async_pf_doublefault(cr2_or_gpa, gfn);
3808 kvm_make_request(KVM_REQ_APF_HALT, vcpu);
3810 } else if (kvm_arch_setup_async_pf(vcpu, cr2_or_gpa, gfn))
3814 *pfn = __gfn_to_pfn_memslot(slot, gfn, false, NULL,
3815 write, writable, hva);
3819 static int direct_page_fault(struct kvm_vcpu *vcpu, gpa_t gpa, u32 error_code,
3820 bool prefault, int max_level, bool is_tdp)
3822 bool is_tdp_mmu_fault = is_tdp_mmu(vcpu->arch.mmu);
3823 bool write = error_code & PFERR_WRITE_MASK;
3826 gfn_t gfn = gpa >> PAGE_SHIFT;
3827 unsigned long mmu_seq;
3832 if (page_fault_handle_page_track(vcpu, error_code, gfn))
3833 return RET_PF_EMULATE;
3835 if (!is_tdp_mmu_fault) {
3836 r = fast_page_fault(vcpu, gpa, error_code);
3837 if (r != RET_PF_INVALID)
3841 r = mmu_topup_memory_caches(vcpu, false);
3845 mmu_seq = vcpu->kvm->mmu_notifier_seq;
3848 if (try_async_pf(vcpu, prefault, gfn, gpa, &pfn, &hva,
3849 write, &map_writable))
3850 return RET_PF_RETRY;
3852 if (handle_abnormal_pfn(vcpu, is_tdp ? 0 : gpa, gfn, pfn, ACC_ALL, &r))
3857 if (is_tdp_mmu_fault)
3858 read_lock(&vcpu->kvm->mmu_lock);
3860 write_lock(&vcpu->kvm->mmu_lock);
3862 if (!is_noslot_pfn(pfn) && mmu_notifier_retry_hva(vcpu->kvm, mmu_seq, hva))
3864 r = make_mmu_pages_available(vcpu);
3868 if (is_tdp_mmu_fault)
3869 r = kvm_tdp_mmu_map(vcpu, gpa, error_code, map_writable, max_level,
3872 r = __direct_map(vcpu, gpa, error_code, map_writable, max_level, pfn,
3876 if (is_tdp_mmu_fault)
3877 read_unlock(&vcpu->kvm->mmu_lock);
3879 write_unlock(&vcpu->kvm->mmu_lock);
3880 kvm_release_pfn_clean(pfn);
3884 static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gpa_t gpa,
3885 u32 error_code, bool prefault)
3887 pgprintk("%s: gva %lx error %x\n", __func__, gpa, error_code);
3889 /* This path builds a PAE pagetable, we can map 2mb pages at maximum. */
3890 return direct_page_fault(vcpu, gpa & PAGE_MASK, error_code, prefault,
3891 PG_LEVEL_2M, false);
3894 int kvm_handle_page_fault(struct kvm_vcpu *vcpu, u64 error_code,
3895 u64 fault_address, char *insn, int insn_len)
3898 u32 flags = vcpu->arch.apf.host_apf_flags;
3900 #ifndef CONFIG_X86_64
3901 /* A 64-bit CR2 should be impossible on 32-bit KVM. */
3902 if (WARN_ON_ONCE(fault_address >> 32))
3906 vcpu->arch.l1tf_flush_l1d = true;
3908 trace_kvm_page_fault(fault_address, error_code);
3910 if (kvm_event_needs_reinjection(vcpu))
3911 kvm_mmu_unprotect_page_virt(vcpu, fault_address);
3912 r = kvm_mmu_page_fault(vcpu, fault_address, error_code, insn,
3914 } else if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) {
3915 vcpu->arch.apf.host_apf_flags = 0;
3916 local_irq_disable();
3917 kvm_async_pf_task_wait_schedule(fault_address);
3920 WARN_ONCE(1, "Unexpected host async PF flags: %x\n", flags);
3925 EXPORT_SYMBOL_GPL(kvm_handle_page_fault);
3927 int kvm_tdp_page_fault(struct kvm_vcpu *vcpu, gpa_t gpa, u32 error_code,
3932 for (max_level = KVM_MAX_HUGEPAGE_LEVEL;
3933 max_level > PG_LEVEL_4K;
3935 int page_num = KVM_PAGES_PER_HPAGE(max_level);
3936 gfn_t base = (gpa >> PAGE_SHIFT) & ~(page_num - 1);
3938 if (kvm_mtrr_check_gfn_range_consistency(vcpu, base, page_num))
3942 return direct_page_fault(vcpu, gpa, error_code, prefault,
3946 static void nonpaging_init_context(struct kvm_mmu *context)
3948 context->page_fault = nonpaging_page_fault;
3949 context->gva_to_gpa = nonpaging_gva_to_gpa;
3950 context->sync_page = nonpaging_sync_page;
3951 context->invlpg = NULL;
3952 context->root_level = 0;
3953 context->direct_map = true;
3956 static inline bool is_root_usable(struct kvm_mmu_root_info *root, gpa_t pgd,
3957 union kvm_mmu_page_role role)
3959 return (role.direct || pgd == root->pgd) &&
3960 VALID_PAGE(root->hpa) && to_shadow_page(root->hpa) &&
3961 role.word == to_shadow_page(root->hpa)->role.word;
3965 * Find out if a previously cached root matching the new pgd/role is available.
3966 * The current root is also inserted into the cache.
3967 * If a matching root was found, it is assigned to kvm_mmu->root_hpa and true is
3969 * Otherwise, the LRU root from the cache is assigned to kvm_mmu->root_hpa and
3970 * false is returned. This root should now be freed by the caller.
3972 static bool cached_root_available(struct kvm_vcpu *vcpu, gpa_t new_pgd,
3973 union kvm_mmu_page_role new_role)
3976 struct kvm_mmu_root_info root;
3977 struct kvm_mmu *mmu = vcpu->arch.mmu;
3979 root.pgd = mmu->root_pgd;
3980 root.hpa = mmu->root_hpa;
3982 if (is_root_usable(&root, new_pgd, new_role))
3985 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) {
3986 swap(root, mmu->prev_roots[i]);
3988 if (is_root_usable(&root, new_pgd, new_role))
3992 mmu->root_hpa = root.hpa;
3993 mmu->root_pgd = root.pgd;
3995 return i < KVM_MMU_NUM_PREV_ROOTS;
3998 static bool fast_pgd_switch(struct kvm_vcpu *vcpu, gpa_t new_pgd,
3999 union kvm_mmu_page_role new_role)
4001 struct kvm_mmu *mmu = vcpu->arch.mmu;
4004 * For now, limit the fast switch to 64-bit hosts+VMs in order to avoid
4005 * having to deal with PDPTEs. We may add support for 32-bit hosts/VMs
4006 * later if necessary.
4008 if (mmu->shadow_root_level >= PT64_ROOT_4LEVEL &&
4009 mmu->root_level >= PT64_ROOT_4LEVEL)
4010 return cached_root_available(vcpu, new_pgd, new_role);
4015 static void __kvm_mmu_new_pgd(struct kvm_vcpu *vcpu, gpa_t new_pgd,
4016 union kvm_mmu_page_role new_role)
4018 if (!fast_pgd_switch(vcpu, new_pgd, new_role)) {
4019 kvm_mmu_free_roots(vcpu, vcpu->arch.mmu, KVM_MMU_ROOT_CURRENT);
4024 * It's possible that the cached previous root page is obsolete because
4025 * of a change in the MMU generation number. However, changing the
4026 * generation number is accompanied by KVM_REQ_MMU_RELOAD, which will
4027 * free the root set here and allocate a new one.
4029 kvm_make_request(KVM_REQ_LOAD_MMU_PGD, vcpu);
4031 if (force_flush_and_sync_on_reuse) {
4032 kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
4033 kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu);
4037 * The last MMIO access's GVA and GPA are cached in the VCPU. When
4038 * switching to a new CR3, that GVA->GPA mapping may no longer be
4039 * valid. So clear any cached MMIO info even when we don't need to sync
4040 * the shadow page tables.
4042 vcpu_clear_mmio_info(vcpu, MMIO_GVA_ANY);
4045 * If this is a direct root page, it doesn't have a write flooding
4046 * count. Otherwise, clear the write flooding count.
4048 if (!new_role.direct)
4049 __clear_sp_write_flooding_count(
4050 to_shadow_page(vcpu->arch.mmu->root_hpa));
4053 void kvm_mmu_new_pgd(struct kvm_vcpu *vcpu, gpa_t new_pgd)
4055 __kvm_mmu_new_pgd(vcpu, new_pgd, kvm_mmu_calc_root_page_role(vcpu));
4057 EXPORT_SYMBOL_GPL(kvm_mmu_new_pgd);
4059 static unsigned long get_cr3(struct kvm_vcpu *vcpu)
4061 return kvm_read_cr3(vcpu);
4064 static bool sync_mmio_spte(struct kvm_vcpu *vcpu, u64 *sptep, gfn_t gfn,
4065 unsigned int access, int *nr_present)
4067 if (unlikely(is_mmio_spte(*sptep))) {
4068 if (gfn != get_mmio_spte_gfn(*sptep)) {
4069 mmu_spte_clear_no_track(sptep);
4074 mark_mmio_spte(vcpu, sptep, gfn, access);
4081 static inline bool is_last_gpte(struct kvm_mmu *mmu,
4082 unsigned level, unsigned gpte)
4085 * The RHS has bit 7 set iff level < mmu->last_nonleaf_level.
4086 * If it is clear, there are no large pages at this level, so clear
4087 * PT_PAGE_SIZE_MASK in gpte if that is the case.
4089 gpte &= level - mmu->last_nonleaf_level;
4092 * PG_LEVEL_4K always terminates. The RHS has bit 7 set
4093 * iff level <= PG_LEVEL_4K, which for our purpose means
4094 * level == PG_LEVEL_4K; set PT_PAGE_SIZE_MASK in gpte then.
4096 gpte |= level - PG_LEVEL_4K - 1;
4098 return gpte & PT_PAGE_SIZE_MASK;
4101 #define PTTYPE_EPT 18 /* arbitrary */
4102 #define PTTYPE PTTYPE_EPT
4103 #include "paging_tmpl.h"
4107 #include "paging_tmpl.h"
4111 #include "paging_tmpl.h"
4115 __reset_rsvds_bits_mask(struct rsvd_bits_validate *rsvd_check,
4116 u64 pa_bits_rsvd, int level, bool nx, bool gbpages,
4119 u64 gbpages_bit_rsvd = 0;
4120 u64 nonleaf_bit8_rsvd = 0;
4123 rsvd_check->bad_mt_xwr = 0;
4126 gbpages_bit_rsvd = rsvd_bits(7, 7);
4128 if (level == PT32E_ROOT_LEVEL)
4129 high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 62);
4131 high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 51);
4133 /* Note, NX doesn't exist in PDPTEs, this is handled below. */
4135 high_bits_rsvd |= rsvd_bits(63, 63);
4138 * Non-leaf PML4Es and PDPEs reserve bit 8 (which would be the G bit for
4139 * leaf entries) on AMD CPUs only.
4142 nonleaf_bit8_rsvd = rsvd_bits(8, 8);
4145 case PT32_ROOT_LEVEL:
4146 /* no rsvd bits for 2 level 4K page table entries */
4147 rsvd_check->rsvd_bits_mask[0][1] = 0;
4148 rsvd_check->rsvd_bits_mask[0][0] = 0;
4149 rsvd_check->rsvd_bits_mask[1][0] =
4150 rsvd_check->rsvd_bits_mask[0][0];
4153 rsvd_check->rsvd_bits_mask[1][1] = 0;
4157 if (is_cpuid_PSE36())
4158 /* 36bits PSE 4MB page */
4159 rsvd_check->rsvd_bits_mask[1][1] = rsvd_bits(17, 21);
4161 /* 32 bits PSE 4MB page */
4162 rsvd_check->rsvd_bits_mask[1][1] = rsvd_bits(13, 21);
4164 case PT32E_ROOT_LEVEL:
4165 rsvd_check->rsvd_bits_mask[0][2] = rsvd_bits(63, 63) |
4168 rsvd_bits(1, 2); /* PDPTE */
4169 rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd; /* PDE */
4170 rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd; /* PTE */
4171 rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd |
4172 rsvd_bits(13, 20); /* large page */
4173 rsvd_check->rsvd_bits_mask[1][0] =
4174 rsvd_check->rsvd_bits_mask[0][0];
4176 case PT64_ROOT_5LEVEL:
4177 rsvd_check->rsvd_bits_mask[0][4] = high_bits_rsvd |
4180 rsvd_check->rsvd_bits_mask[1][4] =
4181 rsvd_check->rsvd_bits_mask[0][4];
4183 case PT64_ROOT_4LEVEL:
4184 rsvd_check->rsvd_bits_mask[0][3] = high_bits_rsvd |
4187 rsvd_check->rsvd_bits_mask[0][2] = high_bits_rsvd |
4189 rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd;
4190 rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd;
4191 rsvd_check->rsvd_bits_mask[1][3] =
4192 rsvd_check->rsvd_bits_mask[0][3];
4193 rsvd_check->rsvd_bits_mask[1][2] = high_bits_rsvd |
4196 rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd |
4197 rsvd_bits(13, 20); /* large page */
4198 rsvd_check->rsvd_bits_mask[1][0] =
4199 rsvd_check->rsvd_bits_mask[0][0];
4204 static void reset_rsvds_bits_mask(struct kvm_vcpu *vcpu,
4205 struct kvm_mmu *context)
4207 __reset_rsvds_bits_mask(&context->guest_rsvd_check,
4208 vcpu->arch.reserved_gpa_bits,
4209 context->root_level, is_efer_nx(context),
4210 guest_cpuid_has(vcpu, X86_FEATURE_GBPAGES),
4211 is_cr4_pse(context),
4212 guest_cpuid_is_amd_or_hygon(vcpu));
4216 __reset_rsvds_bits_mask_ept(struct rsvd_bits_validate *rsvd_check,
4217 u64 pa_bits_rsvd, bool execonly)
4219 u64 high_bits_rsvd = pa_bits_rsvd & rsvd_bits(0, 51);
4222 rsvd_check->rsvd_bits_mask[0][4] = high_bits_rsvd | rsvd_bits(3, 7);
4223 rsvd_check->rsvd_bits_mask[0][3] = high_bits_rsvd | rsvd_bits(3, 7);
4224 rsvd_check->rsvd_bits_mask[0][2] = high_bits_rsvd | rsvd_bits(3, 6);
4225 rsvd_check->rsvd_bits_mask[0][1] = high_bits_rsvd | rsvd_bits(3, 6);
4226 rsvd_check->rsvd_bits_mask[0][0] = high_bits_rsvd;
4229 rsvd_check->rsvd_bits_mask[1][4] = rsvd_check->rsvd_bits_mask[0][4];
4230 rsvd_check->rsvd_bits_mask[1][3] = rsvd_check->rsvd_bits_mask[0][3];
4231 rsvd_check->rsvd_bits_mask[1][2] = high_bits_rsvd | rsvd_bits(12, 29);
4232 rsvd_check->rsvd_bits_mask[1][1] = high_bits_rsvd | rsvd_bits(12, 20);
4233 rsvd_check->rsvd_bits_mask[1][0] = rsvd_check->rsvd_bits_mask[0][0];
4235 bad_mt_xwr = 0xFFull << (2 * 8); /* bits 3..5 must not be 2 */
4236 bad_mt_xwr |= 0xFFull << (3 * 8); /* bits 3..5 must not be 3 */
4237 bad_mt_xwr |= 0xFFull << (7 * 8); /* bits 3..5 must not be 7 */
4238 bad_mt_xwr |= REPEAT_BYTE(1ull << 2); /* bits 0..2 must not be 010 */
4239 bad_mt_xwr |= REPEAT_BYTE(1ull << 6); /* bits 0..2 must not be 110 */
4241 /* bits 0..2 must not be 100 unless VMX capabilities allow it */
4242 bad_mt_xwr |= REPEAT_BYTE(1ull << 4);
4244 rsvd_check->bad_mt_xwr = bad_mt_xwr;
4247 static void reset_rsvds_bits_mask_ept(struct kvm_vcpu *vcpu,
4248 struct kvm_mmu *context, bool execonly)
4250 __reset_rsvds_bits_mask_ept(&context->guest_rsvd_check,
4251 vcpu->arch.reserved_gpa_bits, execonly);
4254 static inline u64 reserved_hpa_bits(void)
4256 return rsvd_bits(shadow_phys_bits, 63);
4260 * the page table on host is the shadow page table for the page
4261 * table in guest or amd nested guest, its mmu features completely
4262 * follow the features in guest.
4264 static void reset_shadow_zero_bits_mask(struct kvm_vcpu *vcpu,
4265 struct kvm_mmu *context)
4268 * KVM uses NX when TDP is disabled to handle a variety of scenarios,
4269 * notably for huge SPTEs if iTLB multi-hit mitigation is enabled and
4270 * to generate correct permissions for CR0.WP=0/CR4.SMEP=1/EFER.NX=0.
4271 * The iTLB multi-hit workaround can be toggled at any time, so assume
4272 * NX can be used by any non-nested shadow MMU to avoid having to reset
4273 * MMU contexts. Note, KVM forces EFER.NX=1 when TDP is disabled.
4275 bool uses_nx = is_efer_nx(context) || !tdp_enabled;
4277 /* @amd adds a check on bit of SPTEs, which KVM shouldn't use anyways. */
4279 /* KVM doesn't use 2-level page tables for the shadow MMU. */
4280 bool is_pse = false;
4281 struct rsvd_bits_validate *shadow_zero_check;
4284 WARN_ON_ONCE(context->shadow_root_level < PT32E_ROOT_LEVEL);
4286 shadow_zero_check = &context->shadow_zero_check;
4287 __reset_rsvds_bits_mask(shadow_zero_check, reserved_hpa_bits(),
4288 context->shadow_root_level, uses_nx,
4289 guest_cpuid_has(vcpu, X86_FEATURE_GBPAGES),
4292 if (!shadow_me_mask)
4295 for (i = context->shadow_root_level; --i >= 0;) {
4296 shadow_zero_check->rsvd_bits_mask[0][i] &= ~shadow_me_mask;
4297 shadow_zero_check->rsvd_bits_mask[1][i] &= ~shadow_me_mask;
4302 static inline bool boot_cpu_is_amd(void)
4304 WARN_ON_ONCE(!tdp_enabled);
4305 return shadow_x_mask == 0;
4309 * the direct page table on host, use as much mmu features as
4310 * possible, however, kvm currently does not do execution-protection.
4313 reset_tdp_shadow_zero_bits_mask(struct kvm_vcpu *vcpu,
4314 struct kvm_mmu *context)
4316 struct rsvd_bits_validate *shadow_zero_check;
4319 shadow_zero_check = &context->shadow_zero_check;
4321 if (boot_cpu_is_amd())
4322 __reset_rsvds_bits_mask(shadow_zero_check, reserved_hpa_bits(),
4323 context->shadow_root_level, false,
4324 boot_cpu_has(X86_FEATURE_GBPAGES),
4327 __reset_rsvds_bits_mask_ept(shadow_zero_check,
4328 reserved_hpa_bits(), false);
4330 if (!shadow_me_mask)
4333 for (i = context->shadow_root_level; --i >= 0;) {
4334 shadow_zero_check->rsvd_bits_mask[0][i] &= ~shadow_me_mask;
4335 shadow_zero_check->rsvd_bits_mask[1][i] &= ~shadow_me_mask;
4340 * as the comments in reset_shadow_zero_bits_mask() except it
4341 * is the shadow page table for intel nested guest.
4344 reset_ept_shadow_zero_bits_mask(struct kvm_vcpu *vcpu,
4345 struct kvm_mmu *context, bool execonly)
4347 __reset_rsvds_bits_mask_ept(&context->shadow_zero_check,
4348 reserved_hpa_bits(), execonly);
4351 #define BYTE_MASK(access) \
4352 ((1 & (access) ? 2 : 0) | \
4353 (2 & (access) ? 4 : 0) | \
4354 (3 & (access) ? 8 : 0) | \
4355 (4 & (access) ? 16 : 0) | \
4356 (5 & (access) ? 32 : 0) | \
4357 (6 & (access) ? 64 : 0) | \
4358 (7 & (access) ? 128 : 0))
4361 static void update_permission_bitmask(struct kvm_mmu *mmu, bool ept)
4365 const u8 x = BYTE_MASK(ACC_EXEC_MASK);
4366 const u8 w = BYTE_MASK(ACC_WRITE_MASK);
4367 const u8 u = BYTE_MASK(ACC_USER_MASK);
4369 bool cr4_smep = is_cr4_smep(mmu);
4370 bool cr4_smap = is_cr4_smap(mmu);
4371 bool cr0_wp = is_cr0_wp(mmu);
4372 bool efer_nx = is_efer_nx(mmu);
4374 for (byte = 0; byte < ARRAY_SIZE(mmu->permissions); ++byte) {
4375 unsigned pfec = byte << 1;
4378 * Each "*f" variable has a 1 bit for each UWX value
4379 * that causes a fault with the given PFEC.
4382 /* Faults from writes to non-writable pages */
4383 u8 wf = (pfec & PFERR_WRITE_MASK) ? (u8)~w : 0;
4384 /* Faults from user mode accesses to supervisor pages */
4385 u8 uf = (pfec & PFERR_USER_MASK) ? (u8)~u : 0;
4386 /* Faults from fetches of non-executable pages*/
4387 u8 ff = (pfec & PFERR_FETCH_MASK) ? (u8)~x : 0;
4388 /* Faults from kernel mode fetches of user pages */
4390 /* Faults from kernel mode accesses of user pages */
4394 /* Faults from kernel mode accesses to user pages */
4395 u8 kf = (pfec & PFERR_USER_MASK) ? 0 : u;
4397 /* Not really needed: !nx will cause pte.nx to fault */
4401 /* Allow supervisor writes if !cr0.wp */
4403 wf = (pfec & PFERR_USER_MASK) ? wf : 0;
4405 /* Disallow supervisor fetches of user code if cr4.smep */
4407 smepf = (pfec & PFERR_FETCH_MASK) ? kf : 0;
4410 * SMAP:kernel-mode data accesses from user-mode
4411 * mappings should fault. A fault is considered
4412 * as a SMAP violation if all of the following
4413 * conditions are true:
4414 * - X86_CR4_SMAP is set in CR4
4415 * - A user page is accessed
4416 * - The access is not a fetch
4417 * - Page fault in kernel mode
4418 * - if CPL = 3 or X86_EFLAGS_AC is clear
4420 * Here, we cover the first three conditions.
4421 * The fourth is computed dynamically in permission_fault();
4422 * PFERR_RSVD_MASK bit will be set in PFEC if the access is
4423 * *not* subject to SMAP restrictions.
4426 smapf = (pfec & (PFERR_RSVD_MASK|PFERR_FETCH_MASK)) ? 0 : kf;
4429 mmu->permissions[byte] = ff | uf | wf | smepf | smapf;
4434 * PKU is an additional mechanism by which the paging controls access to
4435 * user-mode addresses based on the value in the PKRU register. Protection
4436 * key violations are reported through a bit in the page fault error code.
4437 * Unlike other bits of the error code, the PK bit is not known at the
4438 * call site of e.g. gva_to_gpa; it must be computed directly in
4439 * permission_fault based on two bits of PKRU, on some machine state (CR4,
4440 * CR0, EFER, CPL), and on other bits of the error code and the page tables.
4442 * In particular the following conditions come from the error code, the
4443 * page tables and the machine state:
4444 * - PK is always zero unless CR4.PKE=1 and EFER.LMA=1
4445 * - PK is always zero if RSVD=1 (reserved bit set) or F=1 (instruction fetch)
4446 * - PK is always zero if U=0 in the page tables
4447 * - PKRU.WD is ignored if CR0.WP=0 and the access is a supervisor access.
4449 * The PKRU bitmask caches the result of these four conditions. The error
4450 * code (minus the P bit) and the page table's U bit form an index into the
4451 * PKRU bitmask. Two bits of the PKRU bitmask are then extracted and ANDed
4452 * with the two bits of the PKRU register corresponding to the protection key.
4453 * For the first three conditions above the bits will be 00, thus masking
4454 * away both AD and WD. For all reads or if the last condition holds, WD
4455 * only will be masked away.
4457 static void update_pkru_bitmask(struct kvm_mmu *mmu)
4462 if (!is_cr4_pke(mmu)) {
4467 wp = is_cr0_wp(mmu);
4469 for (bit = 0; bit < ARRAY_SIZE(mmu->permissions); ++bit) {
4470 unsigned pfec, pkey_bits;
4471 bool check_pkey, check_write, ff, uf, wf, pte_user;
4474 ff = pfec & PFERR_FETCH_MASK;
4475 uf = pfec & PFERR_USER_MASK;
4476 wf = pfec & PFERR_WRITE_MASK;
4478 /* PFEC.RSVD is replaced by ACC_USER_MASK. */
4479 pte_user = pfec & PFERR_RSVD_MASK;
4482 * Only need to check the access which is not an
4483 * instruction fetch and is to a user page.
4485 check_pkey = (!ff && pte_user);
4487 * write access is controlled by PKRU if it is a
4488 * user access or CR0.WP = 1.
4490 check_write = check_pkey && wf && (uf || wp);
4492 /* PKRU.AD stops both read and write access. */
4493 pkey_bits = !!check_pkey;
4494 /* PKRU.WD stops write access. */
4495 pkey_bits |= (!!check_write) << 1;
4497 mmu->pkru_mask |= (pkey_bits & 3) << pfec;
4501 static void update_last_nonleaf_level(struct kvm_mmu *mmu)
4503 unsigned root_level = mmu->root_level;
4505 mmu->last_nonleaf_level = root_level;
4506 if (root_level == PT32_ROOT_LEVEL && is_cr4_pse(mmu))
4507 mmu->last_nonleaf_level++;
4510 static void reset_guest_paging_metadata(struct kvm_vcpu *vcpu,
4511 struct kvm_mmu *mmu)
4513 if (!is_cr0_pg(mmu))
4516 reset_rsvds_bits_mask(vcpu, mmu);
4517 update_permission_bitmask(mmu, false);
4518 update_pkru_bitmask(mmu);
4519 update_last_nonleaf_level(mmu);
4522 static void paging64_init_context_common(struct kvm_mmu *context,
4525 context->root_level = root_level;
4527 WARN_ON_ONCE(!is_cr4_pae(context));
4528 context->page_fault = paging64_page_fault;
4529 context->gva_to_gpa = paging64_gva_to_gpa;
4530 context->sync_page = paging64_sync_page;
4531 context->invlpg = paging64_invlpg;
4532 context->direct_map = false;
4535 static void paging64_init_context(struct kvm_mmu *context,
4536 struct kvm_mmu_role_regs *regs)
4538 int root_level = ____is_cr4_la57(regs) ? PT64_ROOT_5LEVEL :
4541 paging64_init_context_common(context, root_level);
4544 static void paging32_init_context(struct kvm_mmu *context)
4546 context->root_level = PT32_ROOT_LEVEL;
4547 context->page_fault = paging32_page_fault;
4548 context->gva_to_gpa = paging32_gva_to_gpa;
4549 context->sync_page = paging32_sync_page;
4550 context->invlpg = paging32_invlpg;
4551 context->direct_map = false;
4554 static void paging32E_init_context(struct kvm_mmu *context)
4556 paging64_init_context_common(context, PT32E_ROOT_LEVEL);
4559 static union kvm_mmu_extended_role kvm_calc_mmu_role_ext(struct kvm_vcpu *vcpu,
4560 struct kvm_mmu_role_regs *regs)
4562 union kvm_mmu_extended_role ext = {0};
4564 if (____is_cr0_pg(regs)) {
4566 ext.cr4_pae = ____is_cr4_pae(regs);
4567 ext.cr4_smep = ____is_cr4_smep(regs);
4568 ext.cr4_smap = ____is_cr4_smap(regs);
4569 ext.cr4_pse = ____is_cr4_pse(regs);
4571 /* PKEY and LA57 are active iff long mode is active. */
4572 ext.cr4_pke = ____is_efer_lma(regs) && ____is_cr4_pke(regs);
4573 ext.cr4_la57 = ____is_efer_lma(regs) && ____is_cr4_la57(regs);
4581 static union kvm_mmu_role kvm_calc_mmu_role_common(struct kvm_vcpu *vcpu,
4582 struct kvm_mmu_role_regs *regs,
4585 union kvm_mmu_role role = {0};
4587 role.base.access = ACC_ALL;
4588 if (____is_cr0_pg(regs)) {
4589 role.base.efer_nx = ____is_efer_nx(regs);
4590 role.base.cr0_wp = ____is_cr0_wp(regs);
4592 role.base.smm = is_smm(vcpu);
4593 role.base.guest_mode = is_guest_mode(vcpu);
4598 role.ext = kvm_calc_mmu_role_ext(vcpu, regs);
4603 static inline int kvm_mmu_get_tdp_level(struct kvm_vcpu *vcpu)
4605 /* Use 5-level TDP if and only if it's useful/necessary. */
4606 if (max_tdp_level == 5 && cpuid_maxphyaddr(vcpu) <= 48)
4609 return max_tdp_level;
4612 static union kvm_mmu_role
4613 kvm_calc_tdp_mmu_root_page_role(struct kvm_vcpu *vcpu,
4614 struct kvm_mmu_role_regs *regs, bool base_only)
4616 union kvm_mmu_role role = kvm_calc_mmu_role_common(vcpu, regs, base_only);
4618 role.base.ad_disabled = (shadow_accessed_mask == 0);
4619 role.base.level = kvm_mmu_get_tdp_level(vcpu);
4620 role.base.direct = true;
4621 role.base.gpte_is_8_bytes = true;
4626 static void init_kvm_tdp_mmu(struct kvm_vcpu *vcpu)
4628 struct kvm_mmu *context = &vcpu->arch.root_mmu;
4629 struct kvm_mmu_role_regs regs = vcpu_to_role_regs(vcpu);
4630 union kvm_mmu_role new_role =
4631 kvm_calc_tdp_mmu_root_page_role(vcpu, ®s, false);
4633 if (new_role.as_u64 == context->mmu_role.as_u64)
4636 context->mmu_role.as_u64 = new_role.as_u64;
4637 context->page_fault = kvm_tdp_page_fault;
4638 context->sync_page = nonpaging_sync_page;
4639 context->invlpg = NULL;
4640 context->shadow_root_level = kvm_mmu_get_tdp_level(vcpu);
4641 context->direct_map = true;
4642 context->get_guest_pgd = get_cr3;
4643 context->get_pdptr = kvm_pdptr_read;
4644 context->inject_page_fault = kvm_inject_page_fault;
4646 if (!is_paging(vcpu)) {
4647 context->gva_to_gpa = nonpaging_gva_to_gpa;
4648 context->root_level = 0;
4649 } else if (is_long_mode(vcpu)) {
4650 context->root_level = ____is_cr4_la57(®s) ?
4651 PT64_ROOT_5LEVEL : PT64_ROOT_4LEVEL;
4652 context->gva_to_gpa = paging64_gva_to_gpa;
4653 } else if (is_pae(vcpu)) {
4654 context->root_level = PT32E_ROOT_LEVEL;
4655 context->gva_to_gpa = paging64_gva_to_gpa;
4657 context->root_level = PT32_ROOT_LEVEL;
4658 context->gva_to_gpa = paging32_gva_to_gpa;
4661 reset_guest_paging_metadata(vcpu, context);
4662 reset_tdp_shadow_zero_bits_mask(vcpu, context);
4665 static union kvm_mmu_role
4666 kvm_calc_shadow_root_page_role_common(struct kvm_vcpu *vcpu,
4667 struct kvm_mmu_role_regs *regs, bool base_only)
4669 union kvm_mmu_role role = kvm_calc_mmu_role_common(vcpu, regs, base_only);
4671 role.base.smep_andnot_wp = role.ext.cr4_smep && !____is_cr0_wp(regs);
4672 role.base.smap_andnot_wp = role.ext.cr4_smap && !____is_cr0_wp(regs);
4673 role.base.gpte_is_8_bytes = ____is_cr0_pg(regs) && ____is_cr4_pae(regs);
4678 static union kvm_mmu_role
4679 kvm_calc_shadow_mmu_root_page_role(struct kvm_vcpu *vcpu,
4680 struct kvm_mmu_role_regs *regs, bool base_only)
4682 union kvm_mmu_role role =
4683 kvm_calc_shadow_root_page_role_common(vcpu, regs, base_only);
4685 role.base.direct = !____is_cr0_pg(regs);
4687 if (!____is_efer_lma(regs))
4688 role.base.level = PT32E_ROOT_LEVEL;
4689 else if (____is_cr4_la57(regs))
4690 role.base.level = PT64_ROOT_5LEVEL;
4692 role.base.level = PT64_ROOT_4LEVEL;
4697 static void shadow_mmu_init_context(struct kvm_vcpu *vcpu, struct kvm_mmu *context,
4698 struct kvm_mmu_role_regs *regs,
4699 union kvm_mmu_role new_role)
4701 if (new_role.as_u64 == context->mmu_role.as_u64)
4704 context->mmu_role.as_u64 = new_role.as_u64;
4706 if (!____is_cr0_pg(regs))
4707 nonpaging_init_context(context);
4708 else if (____is_efer_lma(regs))
4709 paging64_init_context(context, regs);
4710 else if (____is_cr4_pae(regs))
4711 paging32E_init_context(context);
4713 paging32_init_context(context);
4715 reset_guest_paging_metadata(vcpu, context);
4716 context->shadow_root_level = new_role.base.level;
4718 reset_shadow_zero_bits_mask(vcpu, context);
4721 static void kvm_init_shadow_mmu(struct kvm_vcpu *vcpu,
4722 struct kvm_mmu_role_regs *regs)
4724 struct kvm_mmu *context = &vcpu->arch.root_mmu;
4725 union kvm_mmu_role new_role =
4726 kvm_calc_shadow_mmu_root_page_role(vcpu, regs, false);
4728 shadow_mmu_init_context(vcpu, context, regs, new_role);
4731 static union kvm_mmu_role
4732 kvm_calc_shadow_npt_root_page_role(struct kvm_vcpu *vcpu,
4733 struct kvm_mmu_role_regs *regs)
4735 union kvm_mmu_role role =
4736 kvm_calc_shadow_root_page_role_common(vcpu, regs, false);
4738 role.base.direct = false;
4739 role.base.level = kvm_mmu_get_tdp_level(vcpu);
4744 void kvm_init_shadow_npt_mmu(struct kvm_vcpu *vcpu, unsigned long cr0,
4745 unsigned long cr4, u64 efer, gpa_t nested_cr3)
4747 struct kvm_mmu *context = &vcpu->arch.guest_mmu;
4748 struct kvm_mmu_role_regs regs = {
4753 union kvm_mmu_role new_role;
4755 new_role = kvm_calc_shadow_npt_root_page_role(vcpu, ®s);
4757 __kvm_mmu_new_pgd(vcpu, nested_cr3, new_role.base);
4759 shadow_mmu_init_context(vcpu, context, ®s, new_role);
4762 * Redo the shadow bits, the reset done by shadow_mmu_init_context()
4763 * (above) may use the wrong shadow_root_level.
4765 reset_shadow_zero_bits_mask(vcpu, context);
4767 EXPORT_SYMBOL_GPL(kvm_init_shadow_npt_mmu);
4769 static union kvm_mmu_role
4770 kvm_calc_shadow_ept_root_page_role(struct kvm_vcpu *vcpu, bool accessed_dirty,
4771 bool execonly, u8 level)
4773 union kvm_mmu_role role = {0};
4775 /* SMM flag is inherited from root_mmu */
4776 role.base.smm = vcpu->arch.root_mmu.mmu_role.base.smm;
4778 role.base.level = level;
4779 role.base.gpte_is_8_bytes = true;
4780 role.base.direct = false;
4781 role.base.ad_disabled = !accessed_dirty;
4782 role.base.guest_mode = true;
4783 role.base.access = ACC_ALL;
4785 /* EPT, and thus nested EPT, does not consume CR0, CR4, nor EFER. */
4787 role.ext.execonly = execonly;
4793 void kvm_init_shadow_ept_mmu(struct kvm_vcpu *vcpu, bool execonly,
4794 bool accessed_dirty, gpa_t new_eptp)
4796 struct kvm_mmu *context = &vcpu->arch.guest_mmu;
4797 u8 level = vmx_eptp_page_walk_level(new_eptp);
4798 union kvm_mmu_role new_role =
4799 kvm_calc_shadow_ept_root_page_role(vcpu, accessed_dirty,
4802 __kvm_mmu_new_pgd(vcpu, new_eptp, new_role.base);
4804 if (new_role.as_u64 == context->mmu_role.as_u64)
4807 context->mmu_role.as_u64 = new_role.as_u64;
4809 context->shadow_root_level = level;
4811 context->ept_ad = accessed_dirty;
4812 context->page_fault = ept_page_fault;
4813 context->gva_to_gpa = ept_gva_to_gpa;
4814 context->sync_page = ept_sync_page;
4815 context->invlpg = ept_invlpg;
4816 context->root_level = level;
4817 context->direct_map = false;
4819 update_permission_bitmask(context, true);
4820 update_last_nonleaf_level(context);
4821 update_pkru_bitmask(context);
4822 reset_rsvds_bits_mask_ept(vcpu, context, execonly);
4823 reset_ept_shadow_zero_bits_mask(vcpu, context, execonly);
4825 EXPORT_SYMBOL_GPL(kvm_init_shadow_ept_mmu);
4827 static void init_kvm_softmmu(struct kvm_vcpu *vcpu)
4829 struct kvm_mmu *context = &vcpu->arch.root_mmu;
4830 struct kvm_mmu_role_regs regs = vcpu_to_role_regs(vcpu);
4832 kvm_init_shadow_mmu(vcpu, ®s);
4834 context->get_guest_pgd = get_cr3;
4835 context->get_pdptr = kvm_pdptr_read;
4836 context->inject_page_fault = kvm_inject_page_fault;
4839 static union kvm_mmu_role
4840 kvm_calc_nested_mmu_role(struct kvm_vcpu *vcpu, struct kvm_mmu_role_regs *regs)
4842 union kvm_mmu_role role;
4844 role = kvm_calc_shadow_root_page_role_common(vcpu, regs, false);
4847 * Nested MMUs are used only for walking L2's gva->gpa, they never have
4848 * shadow pages of their own and so "direct" has no meaning. Set it
4849 * to "true" to try to detect bogus usage of the nested MMU.
4851 role.base.direct = true;
4853 if (!____is_cr0_pg(regs))
4854 role.base.level = 0;
4855 else if (____is_efer_lma(regs))
4856 role.base.level = ____is_cr4_la57(regs) ? PT64_ROOT_5LEVEL :
4858 else if (____is_cr4_pae(regs))
4859 role.base.level = PT32E_ROOT_LEVEL;
4861 role.base.level = PT32_ROOT_LEVEL;
4866 static void init_kvm_nested_mmu(struct kvm_vcpu *vcpu)
4868 struct kvm_mmu_role_regs regs = vcpu_to_role_regs(vcpu);
4869 union kvm_mmu_role new_role = kvm_calc_nested_mmu_role(vcpu, ®s);
4870 struct kvm_mmu *g_context = &vcpu->arch.nested_mmu;
4872 if (new_role.as_u64 == g_context->mmu_role.as_u64)
4875 g_context->mmu_role.as_u64 = new_role.as_u64;
4876 g_context->get_guest_pgd = get_cr3;
4877 g_context->get_pdptr = kvm_pdptr_read;
4878 g_context->inject_page_fault = kvm_inject_page_fault;
4879 g_context->root_level = new_role.base.level;
4882 * L2 page tables are never shadowed, so there is no need to sync
4885 g_context->invlpg = NULL;
4888 * Note that arch.mmu->gva_to_gpa translates l2_gpa to l1_gpa using
4889 * L1's nested page tables (e.g. EPT12). The nested translation
4890 * of l2_gva to l1_gpa is done by arch.nested_mmu.gva_to_gpa using
4891 * L2's page tables as the first level of translation and L1's
4892 * nested page tables as the second level of translation. Basically
4893 * the gva_to_gpa functions between mmu and nested_mmu are swapped.
4895 if (!is_paging(vcpu))
4896 g_context->gva_to_gpa = nonpaging_gva_to_gpa_nested;
4897 else if (is_long_mode(vcpu))
4898 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
4899 else if (is_pae(vcpu))
4900 g_context->gva_to_gpa = paging64_gva_to_gpa_nested;
4902 g_context->gva_to_gpa = paging32_gva_to_gpa_nested;
4904 reset_guest_paging_metadata(vcpu, g_context);
4907 void kvm_init_mmu(struct kvm_vcpu *vcpu)
4909 if (mmu_is_nested(vcpu))
4910 init_kvm_nested_mmu(vcpu);
4911 else if (tdp_enabled)
4912 init_kvm_tdp_mmu(vcpu);
4914 init_kvm_softmmu(vcpu);
4916 EXPORT_SYMBOL_GPL(kvm_init_mmu);
4918 static union kvm_mmu_page_role
4919 kvm_mmu_calc_root_page_role(struct kvm_vcpu *vcpu)
4921 struct kvm_mmu_role_regs regs = vcpu_to_role_regs(vcpu);
4922 union kvm_mmu_role role;
4925 role = kvm_calc_tdp_mmu_root_page_role(vcpu, ®s, true);
4927 role = kvm_calc_shadow_mmu_root_page_role(vcpu, ®s, true);
4932 void kvm_mmu_after_set_cpuid(struct kvm_vcpu *vcpu)
4935 * Invalidate all MMU roles to force them to reinitialize as CPUID
4936 * information is factored into reserved bit calculations.
4938 vcpu->arch.root_mmu.mmu_role.ext.valid = 0;
4939 vcpu->arch.guest_mmu.mmu_role.ext.valid = 0;
4940 vcpu->arch.nested_mmu.mmu_role.ext.valid = 0;
4941 kvm_mmu_reset_context(vcpu);
4944 * KVM does not correctly handle changing guest CPUID after KVM_RUN, as
4945 * MAXPHYADDR, GBPAGES support, AMD reserved bit behavior, etc.. aren't
4946 * tracked in kvm_mmu_page_role. As a result, KVM may miss guest page
4947 * faults due to reusing SPs/SPTEs. Alert userspace, but otherwise
4948 * sweep the problem under the rug.
4950 * KVM's horrific CPUID ABI makes the problem all but impossible to
4951 * solve, as correctly handling multiple vCPU models (with respect to
4952 * paging and physical address properties) in a single VM would require
4953 * tracking all relevant CPUID information in kvm_mmu_page_role. That
4954 * is very undesirable as it would double the memory requirements for
4955 * gfn_track (see struct kvm_mmu_page_role comments), and in practice
4956 * no sane VMM mucks with the core vCPU model on the fly.
4958 if (vcpu->arch.last_vmentry_cpu != -1) {
4959 pr_warn_ratelimited("KVM: KVM_SET_CPUID{,2} after KVM_RUN may cause guest instability\n");
4960 pr_warn_ratelimited("KVM: KVM_SET_CPUID{,2} will fail after KVM_RUN starting with Linux 5.16\n");
4964 void kvm_mmu_reset_context(struct kvm_vcpu *vcpu)
4966 kvm_mmu_unload(vcpu);
4969 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context);
4971 int kvm_mmu_load(struct kvm_vcpu *vcpu)
4975 r = mmu_topup_memory_caches(vcpu, !vcpu->arch.mmu->direct_map);
4978 r = mmu_alloc_special_roots(vcpu);
4981 if (vcpu->arch.mmu->direct_map)
4982 r = mmu_alloc_direct_roots(vcpu);
4984 r = mmu_alloc_shadow_roots(vcpu);
4988 kvm_mmu_sync_roots(vcpu);
4990 kvm_mmu_load_pgd(vcpu);
4991 static_call(kvm_x86_tlb_flush_current)(vcpu);
4996 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
4998 kvm_mmu_free_roots(vcpu, &vcpu->arch.root_mmu, KVM_MMU_ROOTS_ALL);
4999 WARN_ON(VALID_PAGE(vcpu->arch.root_mmu.root_hpa));
5000 kvm_mmu_free_roots(vcpu, &vcpu->arch.guest_mmu, KVM_MMU_ROOTS_ALL);
5001 WARN_ON(VALID_PAGE(vcpu->arch.guest_mmu.root_hpa));
5004 static bool need_remote_flush(u64 old, u64 new)
5006 if (!is_shadow_present_pte(old))
5008 if (!is_shadow_present_pte(new))
5010 if ((old ^ new) & PT64_BASE_ADDR_MASK)
5012 old ^= shadow_nx_mask;
5013 new ^= shadow_nx_mask;
5014 return (old & ~new & PT64_PERM_MASK) != 0;
5017 static u64 mmu_pte_write_fetch_gpte(struct kvm_vcpu *vcpu, gpa_t *gpa,
5024 * Assume that the pte write on a page table of the same type
5025 * as the current vcpu paging mode since we update the sptes only
5026 * when they have the same mode.
5028 if (is_pae(vcpu) && *bytes == 4) {
5029 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
5034 if (*bytes == 4 || *bytes == 8) {
5035 r = kvm_vcpu_read_guest_atomic(vcpu, *gpa, &gentry, *bytes);
5044 * If we're seeing too many writes to a page, it may no longer be a page table,
5045 * or we may be forking, in which case it is better to unmap the page.
5047 static bool detect_write_flooding(struct kvm_mmu_page *sp)
5050 * Skip write-flooding detected for the sp whose level is 1, because
5051 * it can become unsync, then the guest page is not write-protected.
5053 if (sp->role.level == PG_LEVEL_4K)
5056 atomic_inc(&sp->write_flooding_count);
5057 return atomic_read(&sp->write_flooding_count) >= 3;
5061 * Misaligned accesses are too much trouble to fix up; also, they usually
5062 * indicate a page is not used as a page table.
5064 static bool detect_write_misaligned(struct kvm_mmu_page *sp, gpa_t gpa,
5067 unsigned offset, pte_size, misaligned;
5069 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
5070 gpa, bytes, sp->role.word);
5072 offset = offset_in_page(gpa);
5073 pte_size = sp->role.gpte_is_8_bytes ? 8 : 4;
5076 * Sometimes, the OS only writes the last one bytes to update status
5077 * bits, for example, in linux, andb instruction is used in clear_bit().
5079 if (!(offset & (pte_size - 1)) && bytes == 1)
5082 misaligned = (offset ^ (offset + bytes - 1)) & ~(pte_size - 1);
5083 misaligned |= bytes < 4;
5088 static u64 *get_written_sptes(struct kvm_mmu_page *sp, gpa_t gpa, int *nspte)
5090 unsigned page_offset, quadrant;
5094 page_offset = offset_in_page(gpa);
5095 level = sp->role.level;
5097 if (!sp->role.gpte_is_8_bytes) {
5098 page_offset <<= 1; /* 32->64 */
5100 * A 32-bit pde maps 4MB while the shadow pdes map
5101 * only 2MB. So we need to double the offset again
5102 * and zap two pdes instead of one.
5104 if (level == PT32_ROOT_LEVEL) {
5105 page_offset &= ~7; /* kill rounding error */
5109 quadrant = page_offset >> PAGE_SHIFT;
5110 page_offset &= ~PAGE_MASK;
5111 if (quadrant != sp->role.quadrant)
5115 spte = &sp->spt[page_offset / sizeof(*spte)];
5119 static void kvm_mmu_pte_write(struct kvm_vcpu *vcpu, gpa_t gpa,
5120 const u8 *new, int bytes,
5121 struct kvm_page_track_notifier_node *node)
5123 gfn_t gfn = gpa >> PAGE_SHIFT;
5124 struct kvm_mmu_page *sp;
5125 LIST_HEAD(invalid_list);
5126 u64 entry, gentry, *spte;
5128 bool remote_flush, local_flush;
5131 * If we don't have indirect shadow pages, it means no page is
5132 * write-protected, so we can exit simply.
5134 if (!READ_ONCE(vcpu->kvm->arch.indirect_shadow_pages))
5137 remote_flush = local_flush = false;
5139 pgprintk("%s: gpa %llx bytes %d\n", __func__, gpa, bytes);
5142 * No need to care whether allocation memory is successful
5143 * or not since pte prefetch is skipped if it does not have
5144 * enough objects in the cache.
5146 mmu_topup_memory_caches(vcpu, true);
5148 write_lock(&vcpu->kvm->mmu_lock);
5150 gentry = mmu_pte_write_fetch_gpte(vcpu, &gpa, &bytes);
5152 ++vcpu->kvm->stat.mmu_pte_write;
5153 kvm_mmu_audit(vcpu, AUDIT_PRE_PTE_WRITE);
5155 for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn) {
5156 if (detect_write_misaligned(sp, gpa, bytes) ||
5157 detect_write_flooding(sp)) {
5158 kvm_mmu_prepare_zap_page(vcpu->kvm, sp, &invalid_list);
5159 ++vcpu->kvm->stat.mmu_flooded;
5163 spte = get_written_sptes(sp, gpa, &npte);
5170 mmu_page_zap_pte(vcpu->kvm, sp, spte, NULL);
5171 if (gentry && sp->role.level != PG_LEVEL_4K)
5172 ++vcpu->kvm->stat.mmu_pde_zapped;
5173 if (need_remote_flush(entry, *spte))
5174 remote_flush = true;
5178 kvm_mmu_flush_or_zap(vcpu, &invalid_list, remote_flush, local_flush);
5179 kvm_mmu_audit(vcpu, AUDIT_POST_PTE_WRITE);
5180 write_unlock(&vcpu->kvm->mmu_lock);
5183 int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa, u64 error_code,
5184 void *insn, int insn_len)
5186 int r, emulation_type = EMULTYPE_PF;
5187 bool direct = vcpu->arch.mmu->direct_map;
5189 if (WARN_ON(!VALID_PAGE(vcpu->arch.mmu->root_hpa)))
5190 return RET_PF_RETRY;
5193 if (unlikely(error_code & PFERR_RSVD_MASK)) {
5194 r = handle_mmio_page_fault(vcpu, cr2_or_gpa, direct);
5195 if (r == RET_PF_EMULATE)
5199 if (r == RET_PF_INVALID) {
5200 r = kvm_mmu_do_page_fault(vcpu, cr2_or_gpa,
5201 lower_32_bits(error_code), false);
5202 if (WARN_ON_ONCE(r == RET_PF_INVALID))
5208 if (r != RET_PF_EMULATE)
5212 * Before emulating the instruction, check if the error code
5213 * was due to a RO violation while translating the guest page.
5214 * This can occur when using nested virtualization with nested
5215 * paging in both guests. If true, we simply unprotect the page
5216 * and resume the guest.
5218 if (vcpu->arch.mmu->direct_map &&
5219 (error_code & PFERR_NESTED_GUEST_PAGE) == PFERR_NESTED_GUEST_PAGE) {
5220 kvm_mmu_unprotect_page(vcpu->kvm, gpa_to_gfn(cr2_or_gpa));
5225 * vcpu->arch.mmu.page_fault returned RET_PF_EMULATE, but we can still
5226 * optimistically try to just unprotect the page and let the processor
5227 * re-execute the instruction that caused the page fault. Do not allow
5228 * retrying MMIO emulation, as it's not only pointless but could also
5229 * cause us to enter an infinite loop because the processor will keep
5230 * faulting on the non-existent MMIO address. Retrying an instruction
5231 * from a nested guest is also pointless and dangerous as we are only
5232 * explicitly shadowing L1's page tables, i.e. unprotecting something
5233 * for L1 isn't going to magically fix whatever issue cause L2 to fail.
5235 if (!mmio_info_in_cache(vcpu, cr2_or_gpa, direct) && !is_guest_mode(vcpu))
5236 emulation_type |= EMULTYPE_ALLOW_RETRY_PF;
5238 return x86_emulate_instruction(vcpu, cr2_or_gpa, emulation_type, insn,
5241 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault);
5243 void kvm_mmu_invalidate_gva(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
5244 gva_t gva, hpa_t root_hpa)
5248 /* It's actually a GPA for vcpu->arch.guest_mmu. */
5249 if (mmu != &vcpu->arch.guest_mmu) {
5250 /* INVLPG on a non-canonical address is a NOP according to the SDM. */
5251 if (is_noncanonical_address(gva, vcpu))
5254 static_call(kvm_x86_tlb_flush_gva)(vcpu, gva);
5260 if (root_hpa == INVALID_PAGE) {
5261 mmu->invlpg(vcpu, gva, mmu->root_hpa);
5264 * INVLPG is required to invalidate any global mappings for the VA,
5265 * irrespective of PCID. Since it would take us roughly similar amount
5266 * of work to determine whether any of the prev_root mappings of the VA
5267 * is marked global, or to just sync it blindly, so we might as well
5268 * just always sync it.
5270 * Mappings not reachable via the current cr3 or the prev_roots will be
5271 * synced when switching to that cr3, so nothing needs to be done here
5274 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
5275 if (VALID_PAGE(mmu->prev_roots[i].hpa))
5276 mmu->invlpg(vcpu, gva, mmu->prev_roots[i].hpa);
5278 mmu->invlpg(vcpu, gva, root_hpa);
5282 void kvm_mmu_invlpg(struct kvm_vcpu *vcpu, gva_t gva)
5284 kvm_mmu_invalidate_gva(vcpu, vcpu->arch.mmu, gva, INVALID_PAGE);
5285 ++vcpu->stat.invlpg;
5287 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg);
5290 void kvm_mmu_invpcid_gva(struct kvm_vcpu *vcpu, gva_t gva, unsigned long pcid)
5292 struct kvm_mmu *mmu = vcpu->arch.mmu;
5293 bool tlb_flush = false;
5296 if (pcid == kvm_get_active_pcid(vcpu)) {
5297 mmu->invlpg(vcpu, gva, mmu->root_hpa);
5301 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++) {
5302 if (VALID_PAGE(mmu->prev_roots[i].hpa) &&
5303 pcid == kvm_get_pcid(vcpu, mmu->prev_roots[i].pgd)) {
5304 mmu->invlpg(vcpu, gva, mmu->prev_roots[i].hpa);
5310 static_call(kvm_x86_tlb_flush_gva)(vcpu, gva);
5312 ++vcpu->stat.invlpg;
5315 * Mappings not reachable via the current cr3 or the prev_roots will be
5316 * synced when switching to that cr3, so nothing needs to be done here
5321 void kvm_configure_mmu(bool enable_tdp, int tdp_max_root_level,
5322 int tdp_huge_page_level)
5324 tdp_enabled = enable_tdp;
5325 max_tdp_level = tdp_max_root_level;
5328 * max_huge_page_level reflects KVM's MMU capabilities irrespective
5329 * of kernel support, e.g. KVM may be capable of using 1GB pages when
5330 * the kernel is not. But, KVM never creates a page size greater than
5331 * what is used by the kernel for any given HVA, i.e. the kernel's
5332 * capabilities are ultimately consulted by kvm_mmu_hugepage_adjust().
5335 max_huge_page_level = tdp_huge_page_level;
5336 else if (boot_cpu_has(X86_FEATURE_GBPAGES))
5337 max_huge_page_level = PG_LEVEL_1G;
5339 max_huge_page_level = PG_LEVEL_2M;
5341 EXPORT_SYMBOL_GPL(kvm_configure_mmu);
5343 /* The return value indicates if tlb flush on all vcpus is needed. */
5344 typedef bool (*slot_level_handler) (struct kvm *kvm, struct kvm_rmap_head *rmap_head,
5345 struct kvm_memory_slot *slot);
5347 /* The caller should hold mmu-lock before calling this function. */
5348 static __always_inline bool
5349 slot_handle_level_range(struct kvm *kvm, struct kvm_memory_slot *memslot,
5350 slot_level_handler fn, int start_level, int end_level,
5351 gfn_t start_gfn, gfn_t end_gfn, bool flush_on_yield,
5354 struct slot_rmap_walk_iterator iterator;
5356 for_each_slot_rmap_range(memslot, start_level, end_level, start_gfn,
5357 end_gfn, &iterator) {
5359 flush |= fn(kvm, iterator.rmap, memslot);
5361 if (need_resched() || rwlock_needbreak(&kvm->mmu_lock)) {
5362 if (flush && flush_on_yield) {
5363 kvm_flush_remote_tlbs_with_address(kvm,
5365 iterator.gfn - start_gfn + 1);
5368 cond_resched_rwlock_write(&kvm->mmu_lock);
5375 static __always_inline bool
5376 slot_handle_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
5377 slot_level_handler fn, int start_level, int end_level,
5378 bool flush_on_yield)
5380 return slot_handle_level_range(kvm, memslot, fn, start_level,
5381 end_level, memslot->base_gfn,
5382 memslot->base_gfn + memslot->npages - 1,
5383 flush_on_yield, false);
5386 static __always_inline bool
5387 slot_handle_leaf(struct kvm *kvm, struct kvm_memory_slot *memslot,
5388 slot_level_handler fn, bool flush_on_yield)
5390 return slot_handle_level(kvm, memslot, fn, PG_LEVEL_4K,
5391 PG_LEVEL_4K, flush_on_yield);
5394 static void free_mmu_pages(struct kvm_mmu *mmu)
5396 if (!tdp_enabled && mmu->pae_root)
5397 set_memory_encrypted((unsigned long)mmu->pae_root, 1);
5398 free_page((unsigned long)mmu->pae_root);
5399 free_page((unsigned long)mmu->pml4_root);
5402 static int __kvm_mmu_create(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu)
5407 mmu->root_hpa = INVALID_PAGE;
5409 mmu->translate_gpa = translate_gpa;
5410 for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
5411 mmu->prev_roots[i] = KVM_MMU_ROOT_INFO_INVALID;
5414 * When using PAE paging, the four PDPTEs are treated as 'root' pages,
5415 * while the PDP table is a per-vCPU construct that's allocated at MMU
5416 * creation. When emulating 32-bit mode, cr3 is only 32 bits even on
5417 * x86_64. Therefore we need to allocate the PDP table in the first
5418 * 4GB of memory, which happens to fit the DMA32 zone. TDP paging
5419 * generally doesn't use PAE paging and can skip allocating the PDP
5420 * table. The main exception, handled here, is SVM's 32-bit NPT. The
5421 * other exception is for shadowing L1's 32-bit or PAE NPT on 64-bit
5422 * KVM; that horror is handled on-demand by mmu_alloc_shadow_roots().
5424 if (tdp_enabled && kvm_mmu_get_tdp_level(vcpu) > PT32E_ROOT_LEVEL)
5427 page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_DMA32);
5431 mmu->pae_root = page_address(page);
5434 * CR3 is only 32 bits when PAE paging is used, thus it's impossible to
5435 * get the CPU to treat the PDPTEs as encrypted. Decrypt the page so
5436 * that KVM's writes and the CPU's reads get along. Note, this is
5437 * only necessary when using shadow paging, as 64-bit NPT can get at
5438 * the C-bit even when shadowing 32-bit NPT, and SME isn't supported
5439 * by 32-bit kernels (when KVM itself uses 32-bit NPT).
5442 set_memory_decrypted((unsigned long)mmu->pae_root, 1);
5444 WARN_ON_ONCE(shadow_me_mask);
5446 for (i = 0; i < 4; ++i)
5447 mmu->pae_root[i] = INVALID_PAE_ROOT;
5452 int kvm_mmu_create(struct kvm_vcpu *vcpu)
5456 vcpu->arch.mmu_pte_list_desc_cache.kmem_cache = pte_list_desc_cache;
5457 vcpu->arch.mmu_pte_list_desc_cache.gfp_zero = __GFP_ZERO;
5459 vcpu->arch.mmu_page_header_cache.kmem_cache = mmu_page_header_cache;
5460 vcpu->arch.mmu_page_header_cache.gfp_zero = __GFP_ZERO;
5462 vcpu->arch.mmu_shadow_page_cache.gfp_zero = __GFP_ZERO;
5464 vcpu->arch.mmu = &vcpu->arch.root_mmu;
5465 vcpu->arch.walk_mmu = &vcpu->arch.root_mmu;
5467 vcpu->arch.nested_mmu.translate_gpa = translate_nested_gpa;
5469 ret = __kvm_mmu_create(vcpu, &vcpu->arch.guest_mmu);
5473 ret = __kvm_mmu_create(vcpu, &vcpu->arch.root_mmu);
5475 goto fail_allocate_root;
5479 free_mmu_pages(&vcpu->arch.guest_mmu);
5483 #define BATCH_ZAP_PAGES 10
5484 static void kvm_zap_obsolete_pages(struct kvm *kvm)
5486 struct kvm_mmu_page *sp, *node;
5487 int nr_zapped, batch = 0;
5490 list_for_each_entry_safe_reverse(sp, node,
5491 &kvm->arch.active_mmu_pages, link) {
5493 * No obsolete valid page exists before a newly created page
5494 * since active_mmu_pages is a FIFO list.
5496 if (!is_obsolete_sp(kvm, sp))
5500 * Invalid pages should never land back on the list of active
5501 * pages. Skip the bogus page, otherwise we'll get stuck in an
5502 * infinite loop if the page gets put back on the list (again).
5504 if (WARN_ON(sp->role.invalid))
5508 * No need to flush the TLB since we're only zapping shadow
5509 * pages with an obsolete generation number and all vCPUS have
5510 * loaded a new root, i.e. the shadow pages being zapped cannot
5511 * be in active use by the guest.
5513 if (batch >= BATCH_ZAP_PAGES &&
5514 cond_resched_rwlock_write(&kvm->mmu_lock)) {
5519 if (__kvm_mmu_prepare_zap_page(kvm, sp,
5520 &kvm->arch.zapped_obsolete_pages, &nr_zapped)) {
5527 * Trigger a remote TLB flush before freeing the page tables to ensure
5528 * KVM is not in the middle of a lockless shadow page table walk, which
5529 * may reference the pages.
5531 kvm_mmu_commit_zap_page(kvm, &kvm->arch.zapped_obsolete_pages);
5535 * Fast invalidate all shadow pages and use lock-break technique
5536 * to zap obsolete pages.
5538 * It's required when memslot is being deleted or VM is being
5539 * destroyed, in these cases, we should ensure that KVM MMU does
5540 * not use any resource of the being-deleted slot or all slots
5541 * after calling the function.
5543 static void kvm_mmu_zap_all_fast(struct kvm *kvm)
5545 lockdep_assert_held(&kvm->slots_lock);
5547 write_lock(&kvm->mmu_lock);
5548 trace_kvm_mmu_zap_all_fast(kvm);
5551 * Toggle mmu_valid_gen between '0' and '1'. Because slots_lock is
5552 * held for the entire duration of zapping obsolete pages, it's
5553 * impossible for there to be multiple invalid generations associated
5554 * with *valid* shadow pages at any given time, i.e. there is exactly
5555 * one valid generation and (at most) one invalid generation.
5557 kvm->arch.mmu_valid_gen = kvm->arch.mmu_valid_gen ? 0 : 1;
5559 /* In order to ensure all threads see this change when
5560 * handling the MMU reload signal, this must happen in the
5561 * same critical section as kvm_reload_remote_mmus, and
5562 * before kvm_zap_obsolete_pages as kvm_zap_obsolete_pages
5563 * could drop the MMU lock and yield.
5565 if (is_tdp_mmu_enabled(kvm))
5566 kvm_tdp_mmu_invalidate_all_roots(kvm);
5569 * Notify all vcpus to reload its shadow page table and flush TLB.
5570 * Then all vcpus will switch to new shadow page table with the new
5573 * Note: we need to do this under the protection of mmu_lock,
5574 * otherwise, vcpu would purge shadow page but miss tlb flush.
5576 kvm_reload_remote_mmus(kvm);
5578 kvm_zap_obsolete_pages(kvm);
5580 write_unlock(&kvm->mmu_lock);
5582 if (is_tdp_mmu_enabled(kvm)) {
5583 read_lock(&kvm->mmu_lock);
5584 kvm_tdp_mmu_zap_invalidated_roots(kvm);
5585 read_unlock(&kvm->mmu_lock);
5589 static bool kvm_has_zapped_obsolete_pages(struct kvm *kvm)
5591 return unlikely(!list_empty_careful(&kvm->arch.zapped_obsolete_pages));
5594 static void kvm_mmu_invalidate_zap_pages_in_memslot(struct kvm *kvm,
5595 struct kvm_memory_slot *slot,
5596 struct kvm_page_track_notifier_node *node)
5598 kvm_mmu_zap_all_fast(kvm);
5601 void kvm_mmu_init_vm(struct kvm *kvm)
5603 struct kvm_page_track_notifier_node *node = &kvm->arch.mmu_sp_tracker;
5605 if (!kvm_mmu_init_tdp_mmu(kvm))
5607 * No smp_load/store wrappers needed here as we are in
5608 * VM init and there cannot be any memslots / other threads
5609 * accessing this struct kvm yet.
5611 kvm->arch.memslots_have_rmaps = true;
5613 node->track_write = kvm_mmu_pte_write;
5614 node->track_flush_slot = kvm_mmu_invalidate_zap_pages_in_memslot;
5615 kvm_page_track_register_notifier(kvm, node);
5618 void kvm_mmu_uninit_vm(struct kvm *kvm)
5620 struct kvm_page_track_notifier_node *node = &kvm->arch.mmu_sp_tracker;
5622 kvm_page_track_unregister_notifier(kvm, node);
5624 kvm_mmu_uninit_tdp_mmu(kvm);
5627 void kvm_zap_gfn_range(struct kvm *kvm, gfn_t gfn_start, gfn_t gfn_end)
5629 struct kvm_memslots *slots;
5630 struct kvm_memory_slot *memslot;
5634 if (kvm_memslots_have_rmaps(kvm)) {
5635 write_lock(&kvm->mmu_lock);
5636 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
5637 slots = __kvm_memslots(kvm, i);
5638 kvm_for_each_memslot(memslot, slots) {
5641 start = max(gfn_start, memslot->base_gfn);
5642 end = min(gfn_end, memslot->base_gfn + memslot->npages);
5646 flush = slot_handle_level_range(kvm, memslot,
5647 kvm_zap_rmapp, PG_LEVEL_4K,
5648 KVM_MAX_HUGEPAGE_LEVEL, start,
5649 end - 1, true, flush);
5653 kvm_flush_remote_tlbs_with_address(kvm, gfn_start, gfn_end);
5654 write_unlock(&kvm->mmu_lock);
5657 if (is_tdp_mmu_enabled(kvm)) {
5660 read_lock(&kvm->mmu_lock);
5661 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++)
5662 flush = kvm_tdp_mmu_zap_gfn_range(kvm, i, gfn_start,
5663 gfn_end, flush, true);
5665 kvm_flush_remote_tlbs_with_address(kvm, gfn_start,
5668 read_unlock(&kvm->mmu_lock);
5672 static bool slot_rmap_write_protect(struct kvm *kvm,
5673 struct kvm_rmap_head *rmap_head,
5674 struct kvm_memory_slot *slot)
5676 return __rmap_write_protect(kvm, rmap_head, false);
5679 void kvm_mmu_slot_remove_write_access(struct kvm *kvm,
5680 struct kvm_memory_slot *memslot,
5685 if (kvm_memslots_have_rmaps(kvm)) {
5686 write_lock(&kvm->mmu_lock);
5687 flush = slot_handle_level(kvm, memslot, slot_rmap_write_protect,
5688 start_level, KVM_MAX_HUGEPAGE_LEVEL,
5690 write_unlock(&kvm->mmu_lock);
5693 if (is_tdp_mmu_enabled(kvm)) {
5694 read_lock(&kvm->mmu_lock);
5695 flush |= kvm_tdp_mmu_wrprot_slot(kvm, memslot, start_level);
5696 read_unlock(&kvm->mmu_lock);
5700 * We can flush all the TLBs out of the mmu lock without TLB
5701 * corruption since we just change the spte from writable to
5702 * readonly so that we only need to care the case of changing
5703 * spte from present to present (changing the spte from present
5704 * to nonpresent will flush all the TLBs immediately), in other
5705 * words, the only case we care is mmu_spte_update() where we
5706 * have checked Host-writable | MMU-writable instead of
5707 * PT_WRITABLE_MASK, that means it does not depend on PT_WRITABLE_MASK
5711 kvm_arch_flush_remote_tlbs_memslot(kvm, memslot);
5714 static bool kvm_mmu_zap_collapsible_spte(struct kvm *kvm,
5715 struct kvm_rmap_head *rmap_head,
5716 struct kvm_memory_slot *slot)
5719 struct rmap_iterator iter;
5720 int need_tlb_flush = 0;
5722 struct kvm_mmu_page *sp;
5725 for_each_rmap_spte(rmap_head, &iter, sptep) {
5726 sp = sptep_to_sp(sptep);
5727 pfn = spte_to_pfn(*sptep);
5730 * We cannot do huge page mapping for indirect shadow pages,
5731 * which are found on the last rmap (level = 1) when not using
5732 * tdp; such shadow pages are synced with the page table in
5733 * the guest, and the guest page table is using 4K page size
5734 * mapping if the indirect sp has level = 1.
5736 if (sp->role.direct && !kvm_is_reserved_pfn(pfn) &&
5737 sp->role.level < kvm_mmu_max_mapping_level(kvm, slot, sp->gfn,
5738 pfn, PG_LEVEL_NUM)) {
5739 pte_list_remove(rmap_head, sptep);
5741 if (kvm_available_flush_tlb_with_range())
5742 kvm_flush_remote_tlbs_with_address(kvm, sp->gfn,
5743 KVM_PAGES_PER_HPAGE(sp->role.level));
5751 return need_tlb_flush;
5754 void kvm_mmu_zap_collapsible_sptes(struct kvm *kvm,
5755 const struct kvm_memory_slot *memslot)
5757 /* FIXME: const-ify all uses of struct kvm_memory_slot. */
5758 struct kvm_memory_slot *slot = (struct kvm_memory_slot *)memslot;
5761 if (kvm_memslots_have_rmaps(kvm)) {
5762 write_lock(&kvm->mmu_lock);
5763 flush = slot_handle_leaf(kvm, slot, kvm_mmu_zap_collapsible_spte, true);
5765 kvm_arch_flush_remote_tlbs_memslot(kvm, slot);
5766 write_unlock(&kvm->mmu_lock);
5769 if (is_tdp_mmu_enabled(kvm)) {
5770 read_lock(&kvm->mmu_lock);
5771 flush = kvm_tdp_mmu_zap_collapsible_sptes(kvm, slot, flush);
5773 kvm_arch_flush_remote_tlbs_memslot(kvm, slot);
5774 read_unlock(&kvm->mmu_lock);
5778 void kvm_arch_flush_remote_tlbs_memslot(struct kvm *kvm,
5779 const struct kvm_memory_slot *memslot)
5782 * All current use cases for flushing the TLBs for a specific memslot
5783 * related to dirty logging, and many do the TLB flush out of mmu_lock.
5784 * The interaction between the various operations on memslot must be
5785 * serialized by slots_locks to ensure the TLB flush from one operation
5786 * is observed by any other operation on the same memslot.
5788 lockdep_assert_held(&kvm->slots_lock);
5789 kvm_flush_remote_tlbs_with_address(kvm, memslot->base_gfn,
5793 void kvm_mmu_slot_leaf_clear_dirty(struct kvm *kvm,
5794 struct kvm_memory_slot *memslot)
5798 if (kvm_memslots_have_rmaps(kvm)) {
5799 write_lock(&kvm->mmu_lock);
5800 flush = slot_handle_leaf(kvm, memslot, __rmap_clear_dirty,
5802 write_unlock(&kvm->mmu_lock);
5805 if (is_tdp_mmu_enabled(kvm)) {
5806 read_lock(&kvm->mmu_lock);
5807 flush |= kvm_tdp_mmu_clear_dirty_slot(kvm, memslot);
5808 read_unlock(&kvm->mmu_lock);
5812 * It's also safe to flush TLBs out of mmu lock here as currently this
5813 * function is only used for dirty logging, in which case flushing TLB
5814 * out of mmu lock also guarantees no dirty pages will be lost in
5818 kvm_arch_flush_remote_tlbs_memslot(kvm, memslot);
5821 void kvm_mmu_zap_all(struct kvm *kvm)
5823 struct kvm_mmu_page *sp, *node;
5824 LIST_HEAD(invalid_list);
5827 write_lock(&kvm->mmu_lock);
5829 list_for_each_entry_safe(sp, node, &kvm->arch.active_mmu_pages, link) {
5830 if (WARN_ON(sp->role.invalid))
5832 if (__kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list, &ign))
5834 if (cond_resched_rwlock_write(&kvm->mmu_lock))
5838 kvm_mmu_commit_zap_page(kvm, &invalid_list);
5840 if (is_tdp_mmu_enabled(kvm))
5841 kvm_tdp_mmu_zap_all(kvm);
5843 write_unlock(&kvm->mmu_lock);
5846 void kvm_mmu_invalidate_mmio_sptes(struct kvm *kvm, u64 gen)
5848 WARN_ON(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS);
5850 gen &= MMIO_SPTE_GEN_MASK;
5853 * Generation numbers are incremented in multiples of the number of
5854 * address spaces in order to provide unique generations across all
5855 * address spaces. Strip what is effectively the address space
5856 * modifier prior to checking for a wrap of the MMIO generation so
5857 * that a wrap in any address space is detected.
5859 gen &= ~((u64)KVM_ADDRESS_SPACE_NUM - 1);
5862 * The very rare case: if the MMIO generation number has wrapped,
5863 * zap all shadow pages.
5865 if (unlikely(gen == 0)) {
5866 kvm_debug_ratelimited("kvm: zapping shadow pages for mmio generation wraparound\n");
5867 kvm_mmu_zap_all_fast(kvm);
5871 static unsigned long
5872 mmu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
5875 int nr_to_scan = sc->nr_to_scan;
5876 unsigned long freed = 0;
5878 mutex_lock(&kvm_lock);
5880 list_for_each_entry(kvm, &vm_list, vm_list) {
5882 LIST_HEAD(invalid_list);
5885 * Never scan more than sc->nr_to_scan VM instances.
5886 * Will not hit this condition practically since we do not try
5887 * to shrink more than one VM and it is very unlikely to see
5888 * !n_used_mmu_pages so many times.
5893 * n_used_mmu_pages is accessed without holding kvm->mmu_lock
5894 * here. We may skip a VM instance errorneosly, but we do not
5895 * want to shrink a VM that only started to populate its MMU
5898 if (!kvm->arch.n_used_mmu_pages &&
5899 !kvm_has_zapped_obsolete_pages(kvm))
5902 idx = srcu_read_lock(&kvm->srcu);
5903 write_lock(&kvm->mmu_lock);
5905 if (kvm_has_zapped_obsolete_pages(kvm)) {
5906 kvm_mmu_commit_zap_page(kvm,
5907 &kvm->arch.zapped_obsolete_pages);
5911 freed = kvm_mmu_zap_oldest_mmu_pages(kvm, sc->nr_to_scan);
5914 write_unlock(&kvm->mmu_lock);
5915 srcu_read_unlock(&kvm->srcu, idx);
5918 * unfair on small ones
5919 * per-vm shrinkers cry out
5920 * sadness comes quickly
5922 list_move_tail(&kvm->vm_list, &vm_list);
5926 mutex_unlock(&kvm_lock);
5930 static unsigned long
5931 mmu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
5933 return percpu_counter_read_positive(&kvm_total_used_mmu_pages);
5936 static struct shrinker mmu_shrinker = {
5937 .count_objects = mmu_shrink_count,
5938 .scan_objects = mmu_shrink_scan,
5939 .seeks = DEFAULT_SEEKS * 10,
5942 static void mmu_destroy_caches(void)
5944 kmem_cache_destroy(pte_list_desc_cache);
5945 kmem_cache_destroy(mmu_page_header_cache);
5948 static bool get_nx_auto_mode(void)
5950 /* Return true when CPU has the bug, and mitigations are ON */
5951 return boot_cpu_has_bug(X86_BUG_ITLB_MULTIHIT) && !cpu_mitigations_off();
5954 static void __set_nx_huge_pages(bool val)
5956 nx_huge_pages = itlb_multihit_kvm_mitigation = val;
5959 static int set_nx_huge_pages(const char *val, const struct kernel_param *kp)
5961 bool old_val = nx_huge_pages;
5964 /* In "auto" mode deploy workaround only if CPU has the bug. */
5965 if (sysfs_streq(val, "off"))
5967 else if (sysfs_streq(val, "force"))
5969 else if (sysfs_streq(val, "auto"))
5970 new_val = get_nx_auto_mode();
5971 else if (strtobool(val, &new_val) < 0)
5974 __set_nx_huge_pages(new_val);
5976 if (new_val != old_val) {
5979 mutex_lock(&kvm_lock);
5981 list_for_each_entry(kvm, &vm_list, vm_list) {
5982 mutex_lock(&kvm->slots_lock);
5983 kvm_mmu_zap_all_fast(kvm);
5984 mutex_unlock(&kvm->slots_lock);
5986 wake_up_process(kvm->arch.nx_lpage_recovery_thread);
5988 mutex_unlock(&kvm_lock);
5994 int kvm_mmu_module_init(void)
5998 if (nx_huge_pages == -1)
5999 __set_nx_huge_pages(get_nx_auto_mode());
6002 * MMU roles use union aliasing which is, generally speaking, an
6003 * undefined behavior. However, we supposedly know how compilers behave
6004 * and the current status quo is unlikely to change. Guardians below are
6005 * supposed to let us know if the assumption becomes false.
6007 BUILD_BUG_ON(sizeof(union kvm_mmu_page_role) != sizeof(u32));
6008 BUILD_BUG_ON(sizeof(union kvm_mmu_extended_role) != sizeof(u32));
6009 BUILD_BUG_ON(sizeof(union kvm_mmu_role) != sizeof(u64));
6011 kvm_mmu_reset_all_pte_masks();
6013 pte_list_desc_cache = kmem_cache_create("pte_list_desc",
6014 sizeof(struct pte_list_desc),
6015 0, SLAB_ACCOUNT, NULL);
6016 if (!pte_list_desc_cache)
6019 mmu_page_header_cache = kmem_cache_create("kvm_mmu_page_header",
6020 sizeof(struct kvm_mmu_page),
6021 0, SLAB_ACCOUNT, NULL);
6022 if (!mmu_page_header_cache)
6025 if (percpu_counter_init(&kvm_total_used_mmu_pages, 0, GFP_KERNEL))
6028 ret = register_shrinker(&mmu_shrinker);
6035 mmu_destroy_caches();
6040 * Calculate mmu pages needed for kvm.
6042 unsigned long kvm_mmu_calculate_default_mmu_pages(struct kvm *kvm)
6044 unsigned long nr_mmu_pages;
6045 unsigned long nr_pages = 0;
6046 struct kvm_memslots *slots;
6047 struct kvm_memory_slot *memslot;
6050 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
6051 slots = __kvm_memslots(kvm, i);
6053 kvm_for_each_memslot(memslot, slots)
6054 nr_pages += memslot->npages;
6057 nr_mmu_pages = nr_pages * KVM_PERMILLE_MMU_PAGES / 1000;
6058 nr_mmu_pages = max(nr_mmu_pages, KVM_MIN_ALLOC_MMU_PAGES);
6060 return nr_mmu_pages;
6063 void kvm_mmu_destroy(struct kvm_vcpu *vcpu)
6065 kvm_mmu_unload(vcpu);
6066 free_mmu_pages(&vcpu->arch.root_mmu);
6067 free_mmu_pages(&vcpu->arch.guest_mmu);
6068 mmu_free_memory_caches(vcpu);
6071 void kvm_mmu_module_exit(void)
6073 mmu_destroy_caches();
6074 percpu_counter_destroy(&kvm_total_used_mmu_pages);
6075 unregister_shrinker(&mmu_shrinker);
6076 mmu_audit_disable();
6079 static int set_nx_huge_pages_recovery_ratio(const char *val, const struct kernel_param *kp)
6081 unsigned int old_val;
6084 old_val = nx_huge_pages_recovery_ratio;
6085 err = param_set_uint(val, kp);
6089 if (READ_ONCE(nx_huge_pages) &&
6090 !old_val && nx_huge_pages_recovery_ratio) {
6093 mutex_lock(&kvm_lock);
6095 list_for_each_entry(kvm, &vm_list, vm_list)
6096 wake_up_process(kvm->arch.nx_lpage_recovery_thread);
6098 mutex_unlock(&kvm_lock);
6104 static void kvm_recover_nx_lpages(struct kvm *kvm)
6106 unsigned long nx_lpage_splits = kvm->stat.nx_lpage_splits;
6108 struct kvm_mmu_page *sp;
6110 LIST_HEAD(invalid_list);
6114 rcu_idx = srcu_read_lock(&kvm->srcu);
6115 write_lock(&kvm->mmu_lock);
6117 ratio = READ_ONCE(nx_huge_pages_recovery_ratio);
6118 to_zap = ratio ? DIV_ROUND_UP(nx_lpage_splits, ratio) : 0;
6119 for ( ; to_zap; --to_zap) {
6120 if (list_empty(&kvm->arch.lpage_disallowed_mmu_pages))
6124 * We use a separate list instead of just using active_mmu_pages
6125 * because the number of lpage_disallowed pages is expected to
6126 * be relatively small compared to the total.
6128 sp = list_first_entry(&kvm->arch.lpage_disallowed_mmu_pages,
6129 struct kvm_mmu_page,
6130 lpage_disallowed_link);
6131 WARN_ON_ONCE(!sp->lpage_disallowed);
6132 if (is_tdp_mmu_page(sp)) {
6133 flush |= kvm_tdp_mmu_zap_sp(kvm, sp);
6135 kvm_mmu_prepare_zap_page(kvm, sp, &invalid_list);
6136 WARN_ON_ONCE(sp->lpage_disallowed);
6139 if (need_resched() || rwlock_needbreak(&kvm->mmu_lock)) {
6140 kvm_mmu_remote_flush_or_zap(kvm, &invalid_list, flush);
6141 cond_resched_rwlock_write(&kvm->mmu_lock);
6145 kvm_mmu_remote_flush_or_zap(kvm, &invalid_list, flush);
6147 write_unlock(&kvm->mmu_lock);
6148 srcu_read_unlock(&kvm->srcu, rcu_idx);
6151 static long get_nx_lpage_recovery_timeout(u64 start_time)
6153 return READ_ONCE(nx_huge_pages) && READ_ONCE(nx_huge_pages_recovery_ratio)
6154 ? start_time + 60 * HZ - get_jiffies_64()
6155 : MAX_SCHEDULE_TIMEOUT;
6158 static int kvm_nx_lpage_recovery_worker(struct kvm *kvm, uintptr_t data)
6161 long remaining_time;
6164 start_time = get_jiffies_64();
6165 remaining_time = get_nx_lpage_recovery_timeout(start_time);
6167 set_current_state(TASK_INTERRUPTIBLE);
6168 while (!kthread_should_stop() && remaining_time > 0) {
6169 schedule_timeout(remaining_time);
6170 remaining_time = get_nx_lpage_recovery_timeout(start_time);
6171 set_current_state(TASK_INTERRUPTIBLE);
6174 set_current_state(TASK_RUNNING);
6176 if (kthread_should_stop())
6179 kvm_recover_nx_lpages(kvm);
6183 int kvm_mmu_post_init_vm(struct kvm *kvm)
6187 err = kvm_vm_create_worker_thread(kvm, kvm_nx_lpage_recovery_worker, 0,
6188 "kvm-nx-lpage-recovery",
6189 &kvm->arch.nx_lpage_recovery_thread);
6191 kthread_unpark(kvm->arch.nx_lpage_recovery_thread);
6196 void kvm_mmu_pre_destroy_vm(struct kvm *kvm)
6198 if (kvm->arch.nx_lpage_recovery_thread)
6199 kthread_stop(kvm->arch.nx_lpage_recovery_thread);