1 // SPDX-License-Identifier: GPL-2.0
4 #include "mmu_internal.h"
10 #include <asm/cmpxchg.h>
11 #include <trace/events/kvm.h>
13 static bool __read_mostly tdp_mmu_enabled = false;
14 module_param_named(tdp_mmu, tdp_mmu_enabled, bool, 0644);
16 /* Initializes the TDP MMU for the VM, if enabled. */
17 void kvm_mmu_init_tdp_mmu(struct kvm *kvm)
19 if (!tdp_enabled || !READ_ONCE(tdp_mmu_enabled))
22 /* This should not be changed for the lifetime of the VM. */
23 kvm->arch.tdp_mmu_enabled = true;
25 INIT_LIST_HEAD(&kvm->arch.tdp_mmu_roots);
26 spin_lock_init(&kvm->arch.tdp_mmu_pages_lock);
27 INIT_LIST_HEAD(&kvm->arch.tdp_mmu_pages);
30 void kvm_mmu_uninit_tdp_mmu(struct kvm *kvm)
32 if (!kvm->arch.tdp_mmu_enabled)
35 WARN_ON(!list_empty(&kvm->arch.tdp_mmu_roots));
38 * Ensure that all the outstanding RCU callbacks to free shadow pages
39 * can run before the VM is torn down.
44 static void tdp_mmu_put_root(struct kvm *kvm, struct kvm_mmu_page *root)
46 if (kvm_mmu_put_root(kvm, root))
47 kvm_tdp_mmu_free_root(kvm, root);
50 static inline bool tdp_mmu_next_root_valid(struct kvm *kvm,
51 struct kvm_mmu_page *root)
53 lockdep_assert_held_write(&kvm->mmu_lock);
55 if (list_entry_is_head(root, &kvm->arch.tdp_mmu_roots, link))
58 kvm_mmu_get_root(kvm, root);
63 static inline struct kvm_mmu_page *tdp_mmu_next_root(struct kvm *kvm,
64 struct kvm_mmu_page *root)
66 struct kvm_mmu_page *next_root;
68 next_root = list_next_entry(root, link);
69 tdp_mmu_put_root(kvm, root);
74 * Note: this iterator gets and puts references to the roots it iterates over.
75 * This makes it safe to release the MMU lock and yield within the loop, but
76 * if exiting the loop early, the caller must drop the reference to the most
77 * recent root. (Unless keeping a live reference is desirable.)
79 #define for_each_tdp_mmu_root_yield_safe(_kvm, _root) \
80 for (_root = list_first_entry(&_kvm->arch.tdp_mmu_roots, \
81 typeof(*_root), link); \
82 tdp_mmu_next_root_valid(_kvm, _root); \
83 _root = tdp_mmu_next_root(_kvm, _root))
85 #define for_each_tdp_mmu_root(_kvm, _root) \
86 list_for_each_entry(_root, &_kvm->arch.tdp_mmu_roots, link)
88 static bool zap_gfn_range(struct kvm *kvm, struct kvm_mmu_page *root,
89 gfn_t start, gfn_t end, bool can_yield, bool flush);
91 void kvm_tdp_mmu_free_root(struct kvm *kvm, struct kvm_mmu_page *root)
93 gfn_t max_gfn = 1ULL << (shadow_phys_bits - PAGE_SHIFT);
95 lockdep_assert_held_write(&kvm->mmu_lock);
97 WARN_ON(root->root_count);
98 WARN_ON(!root->tdp_mmu_page);
100 list_del(&root->link);
102 zap_gfn_range(kvm, root, 0, max_gfn, false, false);
104 free_page((unsigned long)root->spt);
105 kmem_cache_free(mmu_page_header_cache, root);
108 static union kvm_mmu_page_role page_role_for_level(struct kvm_vcpu *vcpu,
111 union kvm_mmu_page_role role;
113 role = vcpu->arch.mmu->mmu_role.base;
116 role.gpte_is_8_bytes = true;
117 role.access = ACC_ALL;
122 static struct kvm_mmu_page *alloc_tdp_mmu_page(struct kvm_vcpu *vcpu, gfn_t gfn,
125 struct kvm_mmu_page *sp;
127 sp = kvm_mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache);
128 sp->spt = kvm_mmu_memory_cache_alloc(&vcpu->arch.mmu_shadow_page_cache);
129 set_page_private(virt_to_page(sp->spt), (unsigned long)sp);
131 sp->role.word = page_role_for_level(vcpu, level).word;
133 sp->tdp_mmu_page = true;
135 trace_kvm_mmu_get_page(sp, true);
140 hpa_t kvm_tdp_mmu_get_vcpu_root_hpa(struct kvm_vcpu *vcpu)
142 union kvm_mmu_page_role role;
143 struct kvm *kvm = vcpu->kvm;
144 struct kvm_mmu_page *root;
146 lockdep_assert_held_write(&kvm->mmu_lock);
148 role = page_role_for_level(vcpu, vcpu->arch.mmu->shadow_root_level);
150 /* Check for an existing root before allocating a new one. */
151 for_each_tdp_mmu_root(kvm, root) {
152 if (root->role.word == role.word) {
153 kvm_mmu_get_root(kvm, root);
158 root = alloc_tdp_mmu_page(vcpu, 0, vcpu->arch.mmu->shadow_root_level);
159 root->root_count = 1;
161 list_add(&root->link, &kvm->arch.tdp_mmu_roots);
164 return __pa(root->spt);
167 static void tdp_mmu_free_sp(struct kvm_mmu_page *sp)
169 free_page((unsigned long)sp->spt);
170 kmem_cache_free(mmu_page_header_cache, sp);
174 * This is called through call_rcu in order to free TDP page table memory
175 * safely with respect to other kernel threads that may be operating on
177 * By only accessing TDP MMU page table memory in an RCU read critical
178 * section, and freeing it after a grace period, lockless access to that
179 * memory won't use it after it is freed.
181 static void tdp_mmu_free_sp_rcu_callback(struct rcu_head *head)
183 struct kvm_mmu_page *sp = container_of(head, struct kvm_mmu_page,
189 static void handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn,
190 u64 old_spte, u64 new_spte, int level,
193 static void handle_changed_spte_acc_track(u64 old_spte, u64 new_spte, int level)
195 if (!is_shadow_present_pte(old_spte) || !is_last_spte(old_spte, level))
198 if (is_accessed_spte(old_spte) &&
199 (!is_shadow_present_pte(new_spte) || !is_accessed_spte(new_spte) ||
200 spte_to_pfn(old_spte) != spte_to_pfn(new_spte)))
201 kvm_set_pfn_accessed(spte_to_pfn(old_spte));
204 static void handle_changed_spte_dirty_log(struct kvm *kvm, int as_id, gfn_t gfn,
205 u64 old_spte, u64 new_spte, int level)
208 struct kvm_memory_slot *slot;
210 if (level > PG_LEVEL_4K)
213 pfn_changed = spte_to_pfn(old_spte) != spte_to_pfn(new_spte);
215 if ((!is_writable_pte(old_spte) || pfn_changed) &&
216 is_writable_pte(new_spte)) {
217 slot = __gfn_to_memslot(__kvm_memslots(kvm, as_id), gfn);
218 mark_page_dirty_in_slot(kvm, slot, gfn);
223 * tdp_mmu_link_page - Add a new page to the list of pages used by the TDP MMU
227 * @shared: This operation may not be running under the exclusive use of
228 * the MMU lock and the operation must synchronize with other
229 * threads that might be adding or removing pages.
230 * @account_nx: This page replaces a NX large page and should be marked for
233 static void tdp_mmu_link_page(struct kvm *kvm, struct kvm_mmu_page *sp,
234 bool shared, bool account_nx)
237 spin_lock(&kvm->arch.tdp_mmu_pages_lock);
239 lockdep_assert_held_write(&kvm->mmu_lock);
241 list_add(&sp->link, &kvm->arch.tdp_mmu_pages);
243 account_huge_nx_page(kvm, sp);
246 spin_unlock(&kvm->arch.tdp_mmu_pages_lock);
250 * tdp_mmu_unlink_page - Remove page from the list of pages used by the TDP MMU
253 * @sp: the page to be removed
254 * @shared: This operation may not be running under the exclusive use of
255 * the MMU lock and the operation must synchronize with other
256 * threads that might be adding or removing pages.
258 static void tdp_mmu_unlink_page(struct kvm *kvm, struct kvm_mmu_page *sp,
262 spin_lock(&kvm->arch.tdp_mmu_pages_lock);
264 lockdep_assert_held_write(&kvm->mmu_lock);
267 if (sp->lpage_disallowed)
268 unaccount_huge_nx_page(kvm, sp);
271 spin_unlock(&kvm->arch.tdp_mmu_pages_lock);
275 * handle_removed_tdp_mmu_page - handle a pt removed from the TDP structure
278 * @pt: the page removed from the paging structure
279 * @shared: This operation may not be running under the exclusive use
280 * of the MMU lock and the operation must synchronize with other
281 * threads that might be modifying SPTEs.
283 * Given a page table that has been removed from the TDP paging structure,
284 * iterates through the page table to clear SPTEs and free child page tables.
286 * Note that pt is passed in as a tdp_ptep_t, but it does not need RCU
287 * protection. Since this thread removed it from the paging structure,
288 * this thread will be responsible for ensuring the page is freed. Hence the
289 * early rcu_dereferences in the function.
291 static void handle_removed_tdp_mmu_page(struct kvm *kvm, tdp_ptep_t pt,
294 struct kvm_mmu_page *sp = sptep_to_sp(rcu_dereference(pt));
295 int level = sp->role.level;
296 gfn_t base_gfn = sp->gfn;
302 trace_kvm_mmu_prepare_zap_page(sp);
304 tdp_mmu_unlink_page(kvm, sp, shared);
306 for (i = 0; i < PT64_ENT_PER_PAGE; i++) {
307 sptep = rcu_dereference(pt) + i;
308 gfn = base_gfn + (i * KVM_PAGES_PER_HPAGE(level - 1));
312 * Set the SPTE to a nonpresent value that other
313 * threads will not overwrite. If the SPTE was
314 * already marked as removed then another thread
315 * handling a page fault could overwrite it, so
316 * set the SPTE until it is set from some other
317 * value to the removed SPTE value.
320 old_child_spte = xchg(sptep, REMOVED_SPTE);
321 if (!is_removed_spte(old_child_spte))
327 * If the SPTE is not MMU-present, there is no backing
328 * page associated with the SPTE and so no side effects
329 * that need to be recorded, and exclusive ownership of
330 * mmu_lock ensures the SPTE can't be made present.
331 * Note, zapping MMIO SPTEs is also unnecessary as they
332 * are guarded by the memslots generation, not by being
335 old_child_spte = READ_ONCE(*sptep);
336 if (!is_shadow_present_pte(old_child_spte))
340 * Marking the SPTE as a removed SPTE is not
341 * strictly necessary here as the MMU lock will
342 * stop other threads from concurrently modifying
343 * this SPTE. Using the removed SPTE value keeps
344 * the two branches consistent and simplifies
347 WRITE_ONCE(*sptep, REMOVED_SPTE);
349 handle_changed_spte(kvm, kvm_mmu_page_as_id(sp), gfn,
350 old_child_spte, REMOVED_SPTE, level - 1,
354 kvm_flush_remote_tlbs_with_address(kvm, gfn,
355 KVM_PAGES_PER_HPAGE(level));
357 call_rcu(&sp->rcu_head, tdp_mmu_free_sp_rcu_callback);
361 * handle_changed_spte - handle bookkeeping associated with an SPTE change
363 * @as_id: the address space of the paging structure the SPTE was a part of
364 * @gfn: the base GFN that was mapped by the SPTE
365 * @old_spte: The value of the SPTE before the change
366 * @new_spte: The value of the SPTE after the change
367 * @level: the level of the PT the SPTE is part of in the paging structure
368 * @shared: This operation may not be running under the exclusive use of
369 * the MMU lock and the operation must synchronize with other
370 * threads that might be modifying SPTEs.
372 * Handle bookkeeping that might result from the modification of a SPTE.
373 * This function must be called for all TDP SPTE modifications.
375 static void __handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn,
376 u64 old_spte, u64 new_spte, int level,
379 bool was_present = is_shadow_present_pte(old_spte);
380 bool is_present = is_shadow_present_pte(new_spte);
381 bool was_leaf = was_present && is_last_spte(old_spte, level);
382 bool is_leaf = is_present && is_last_spte(new_spte, level);
383 bool pfn_changed = spte_to_pfn(old_spte) != spte_to_pfn(new_spte);
385 WARN_ON(level > PT64_ROOT_MAX_LEVEL);
386 WARN_ON(level < PG_LEVEL_4K);
387 WARN_ON(gfn & (KVM_PAGES_PER_HPAGE(level) - 1));
390 * If this warning were to trigger it would indicate that there was a
391 * missing MMU notifier or a race with some notifier handler.
392 * A present, leaf SPTE should never be directly replaced with another
393 * present leaf SPTE pointing to a differnt PFN. A notifier handler
394 * should be zapping the SPTE before the main MM's page table is
395 * changed, or the SPTE should be zeroed, and the TLBs flushed by the
396 * thread before replacement.
398 if (was_leaf && is_leaf && pfn_changed) {
399 pr_err("Invalid SPTE change: cannot replace a present leaf\n"
400 "SPTE with another present leaf SPTE mapping a\n"
402 "as_id: %d gfn: %llx old_spte: %llx new_spte: %llx level: %d",
403 as_id, gfn, old_spte, new_spte, level);
406 * Crash the host to prevent error propagation and guest data
412 if (old_spte == new_spte)
415 trace_kvm_tdp_mmu_spte_changed(as_id, gfn, level, old_spte, new_spte);
418 * The only times a SPTE should be changed from a non-present to
419 * non-present state is when an MMIO entry is installed/modified/
420 * removed. In that case, there is nothing to do here.
422 if (!was_present && !is_present) {
424 * If this change does not involve a MMIO SPTE or removed SPTE,
425 * it is unexpected. Log the change, though it should not
426 * impact the guest since both the former and current SPTEs
429 if (WARN_ON(!is_mmio_spte(old_spte) &&
430 !is_mmio_spte(new_spte) &&
431 !is_removed_spte(new_spte)))
432 pr_err("Unexpected SPTE change! Nonpresent SPTEs\n"
433 "should not be replaced with another,\n"
434 "different nonpresent SPTE, unless one or both\n"
435 "are MMIO SPTEs, or the new SPTE is\n"
436 "a temporary removed SPTE.\n"
437 "as_id: %d gfn: %llx old_spte: %llx new_spte: %llx level: %d",
438 as_id, gfn, old_spte, new_spte, level);
443 if (was_leaf && is_dirty_spte(old_spte) &&
444 (!is_present || !is_dirty_spte(new_spte) || pfn_changed))
445 kvm_set_pfn_dirty(spte_to_pfn(old_spte));
448 * Recursively handle child PTs if the change removed a subtree from
449 * the paging structure.
451 if (was_present && !was_leaf && (pfn_changed || !is_present))
452 handle_removed_tdp_mmu_page(kvm,
453 spte_to_child_pt(old_spte, level), shared);
456 static void handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn,
457 u64 old_spte, u64 new_spte, int level,
460 __handle_changed_spte(kvm, as_id, gfn, old_spte, new_spte, level,
462 handle_changed_spte_acc_track(old_spte, new_spte, level);
463 handle_changed_spte_dirty_log(kvm, as_id, gfn, old_spte,
468 * tdp_mmu_set_spte_atomic - Set a TDP MMU SPTE atomically and handle the
469 * associated bookkeeping
472 * @iter: a tdp_iter instance currently on the SPTE that should be set
473 * @new_spte: The value the SPTE should be set to
474 * Returns: true if the SPTE was set, false if it was not. If false is returned,
475 * this function will have no side-effects.
477 static inline bool tdp_mmu_set_spte_atomic(struct kvm *kvm,
478 struct tdp_iter *iter,
481 lockdep_assert_held_read(&kvm->mmu_lock);
484 * Do not change removed SPTEs. Only the thread that froze the SPTE
487 if (is_removed_spte(iter->old_spte))
490 if (cmpxchg64(rcu_dereference(iter->sptep), iter->old_spte,
491 new_spte) != iter->old_spte)
494 handle_changed_spte(kvm, iter->as_id, iter->gfn, iter->old_spte,
495 new_spte, iter->level, true);
500 static inline bool tdp_mmu_zap_spte_atomic(struct kvm *kvm,
501 struct tdp_iter *iter)
504 * Freeze the SPTE by setting it to a special,
505 * non-present value. This will stop other threads from
506 * immediately installing a present entry in its place
507 * before the TLBs are flushed.
509 if (!tdp_mmu_set_spte_atomic(kvm, iter, REMOVED_SPTE))
512 kvm_flush_remote_tlbs_with_address(kvm, iter->gfn,
513 KVM_PAGES_PER_HPAGE(iter->level));
516 * No other thread can overwrite the removed SPTE as they
517 * must either wait on the MMU lock or use
518 * tdp_mmu_set_spte_atomic which will not overrite the
519 * special removed SPTE value. No bookkeeping is needed
520 * here since the SPTE is going from non-present
523 WRITE_ONCE(*rcu_dereference(iter->sptep), 0);
530 * __tdp_mmu_set_spte - Set a TDP MMU SPTE and handle the associated bookkeeping
532 * @iter: a tdp_iter instance currently on the SPTE that should be set
533 * @new_spte: The value the SPTE should be set to
534 * @record_acc_track: Notify the MM subsystem of changes to the accessed state
535 * of the page. Should be set unless handling an MMU
536 * notifier for access tracking. Leaving record_acc_track
537 * unset in that case prevents page accesses from being
539 * @record_dirty_log: Record the page as dirty in the dirty bitmap if
540 * appropriate for the change being made. Should be set
541 * unless performing certain dirty logging operations.
542 * Leaving record_dirty_log unset in that case prevents page
543 * writes from being double counted.
545 static inline void __tdp_mmu_set_spte(struct kvm *kvm, struct tdp_iter *iter,
546 u64 new_spte, bool record_acc_track,
547 bool record_dirty_log)
549 lockdep_assert_held_write(&kvm->mmu_lock);
552 * No thread should be using this function to set SPTEs to the
553 * temporary removed SPTE value.
554 * If operating under the MMU lock in read mode, tdp_mmu_set_spte_atomic
555 * should be used. If operating under the MMU lock in write mode, the
556 * use of the removed SPTE should not be necessary.
558 WARN_ON(is_removed_spte(iter->old_spte));
560 WRITE_ONCE(*rcu_dereference(iter->sptep), new_spte);
562 __handle_changed_spte(kvm, iter->as_id, iter->gfn, iter->old_spte,
563 new_spte, iter->level, false);
564 if (record_acc_track)
565 handle_changed_spte_acc_track(iter->old_spte, new_spte,
567 if (record_dirty_log)
568 handle_changed_spte_dirty_log(kvm, iter->as_id, iter->gfn,
569 iter->old_spte, new_spte,
573 static inline void tdp_mmu_set_spte(struct kvm *kvm, struct tdp_iter *iter,
576 __tdp_mmu_set_spte(kvm, iter, new_spte, true, true);
579 static inline void tdp_mmu_set_spte_no_acc_track(struct kvm *kvm,
580 struct tdp_iter *iter,
583 __tdp_mmu_set_spte(kvm, iter, new_spte, false, true);
586 static inline void tdp_mmu_set_spte_no_dirty_log(struct kvm *kvm,
587 struct tdp_iter *iter,
590 __tdp_mmu_set_spte(kvm, iter, new_spte, true, false);
593 #define tdp_root_for_each_pte(_iter, _root, _start, _end) \
594 for_each_tdp_pte(_iter, _root->spt, _root->role.level, _start, _end)
596 #define tdp_root_for_each_leaf_pte(_iter, _root, _start, _end) \
597 tdp_root_for_each_pte(_iter, _root, _start, _end) \
598 if (!is_shadow_present_pte(_iter.old_spte) || \
599 !is_last_spte(_iter.old_spte, _iter.level)) \
603 #define tdp_mmu_for_each_pte(_iter, _mmu, _start, _end) \
604 for_each_tdp_pte(_iter, __va(_mmu->root_hpa), \
605 _mmu->shadow_root_level, _start, _end)
608 * Yield if the MMU lock is contended or this thread needs to return control
611 * If this function should yield and flush is set, it will perform a remote
612 * TLB flush before yielding.
614 * If this function yields, it will also reset the tdp_iter's walk over the
615 * paging structure and the calling function should skip to the next
616 * iteration to allow the iterator to continue its traversal from the
617 * paging structure root.
619 * Return true if this function yielded and the iterator's traversal was reset.
620 * Return false if a yield was not needed.
622 static inline bool tdp_mmu_iter_cond_resched(struct kvm *kvm,
623 struct tdp_iter *iter, bool flush)
625 /* Ensure forward progress has been made before yielding. */
626 if (iter->next_last_level_gfn == iter->yielded_gfn)
629 if (need_resched() || rwlock_needbreak(&kvm->mmu_lock)) {
633 kvm_flush_remote_tlbs(kvm);
635 cond_resched_rwlock_write(&kvm->mmu_lock);
638 WARN_ON(iter->gfn > iter->next_last_level_gfn);
640 tdp_iter_restart(iter);
649 * Tears down the mappings for the range of gfns, [start, end), and frees the
650 * non-root pages mapping GFNs strictly within that range. Returns true if
651 * SPTEs have been cleared and a TLB flush is needed before releasing the
653 * If can_yield is true, will release the MMU lock and reschedule if the
654 * scheduler needs the CPU or there is contention on the MMU lock. If this
655 * function cannot yield, it will not release the MMU lock or reschedule and
656 * the caller must ensure it does not supply too large a GFN range, or the
657 * operation can cause a soft lockup. Note, in some use cases a flush may be
658 * required by prior actions. Ensure the pending flush is performed prior to
661 static bool zap_gfn_range(struct kvm *kvm, struct kvm_mmu_page *root,
662 gfn_t start, gfn_t end, bool can_yield, bool flush)
664 struct tdp_iter iter;
668 tdp_root_for_each_pte(iter, root, start, end) {
670 tdp_mmu_iter_cond_resched(kvm, &iter, flush)) {
675 if (!is_shadow_present_pte(iter.old_spte))
679 * If this is a non-last-level SPTE that covers a larger range
680 * than should be zapped, continue, and zap the mappings at a
683 if ((iter.gfn < start ||
684 iter.gfn + KVM_PAGES_PER_HPAGE(iter.level) > end) &&
685 !is_last_spte(iter.old_spte, iter.level))
688 tdp_mmu_set_spte(kvm, &iter, 0);
697 * Tears down the mappings for the range of gfns, [start, end), and frees the
698 * non-root pages mapping GFNs strictly within that range. Returns true if
699 * SPTEs have been cleared and a TLB flush is needed before releasing the
702 bool __kvm_tdp_mmu_zap_gfn_range(struct kvm *kvm, int as_id, gfn_t start,
703 gfn_t end, bool can_yield, bool flush)
705 struct kvm_mmu_page *root;
707 for_each_tdp_mmu_root_yield_safe(kvm, root) {
708 if (kvm_mmu_page_as_id(root) != as_id)
710 flush = zap_gfn_range(kvm, root, start, end, can_yield, flush);
716 void kvm_tdp_mmu_zap_all(struct kvm *kvm)
718 gfn_t max_gfn = 1ULL << (shadow_phys_bits - PAGE_SHIFT);
722 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++)
723 flush = kvm_tdp_mmu_zap_gfn_range(kvm, i, 0, max_gfn, flush);
726 kvm_flush_remote_tlbs(kvm);
730 * Installs a last-level SPTE to handle a TDP page fault.
731 * (NPT/EPT violation/misconfiguration)
733 static int tdp_mmu_map_handle_target_level(struct kvm_vcpu *vcpu, int write,
735 struct tdp_iter *iter,
736 kvm_pfn_t pfn, bool prefault)
740 int make_spte_ret = 0;
742 if (unlikely(is_noslot_pfn(pfn)))
743 new_spte = make_mmio_spte(vcpu, iter->gfn, ACC_ALL);
745 make_spte_ret = make_spte(vcpu, ACC_ALL, iter->level, iter->gfn,
746 pfn, iter->old_spte, prefault, true,
747 map_writable, !shadow_accessed_mask,
750 if (new_spte == iter->old_spte)
751 ret = RET_PF_SPURIOUS;
752 else if (!tdp_mmu_set_spte_atomic(vcpu->kvm, iter, new_spte))
756 * If the page fault was caused by a write but the page is write
757 * protected, emulation is needed. If the emulation was skipped,
758 * the vCPU would have the same fault again.
760 if (make_spte_ret & SET_SPTE_WRITE_PROTECTED_PT) {
762 ret = RET_PF_EMULATE;
763 kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu);
766 /* If a MMIO SPTE is installed, the MMIO will need to be emulated. */
767 if (unlikely(is_mmio_spte(new_spte))) {
768 trace_mark_mmio_spte(rcu_dereference(iter->sptep), iter->gfn,
770 ret = RET_PF_EMULATE;
772 trace_kvm_mmu_set_spte(iter->level, iter->gfn,
773 rcu_dereference(iter->sptep));
777 vcpu->stat.pf_fixed++;
783 * Handle a TDP page fault (NPT/EPT violation/misconfiguration) by installing
784 * page tables and SPTEs to translate the faulting guest physical address.
786 int kvm_tdp_mmu_map(struct kvm_vcpu *vcpu, gpa_t gpa, u32 error_code,
787 int map_writable, int max_level, kvm_pfn_t pfn,
790 bool nx_huge_page_workaround_enabled = is_nx_huge_page_enabled();
791 bool write = error_code & PFERR_WRITE_MASK;
792 bool exec = error_code & PFERR_FETCH_MASK;
793 bool huge_page_disallowed = exec && nx_huge_page_workaround_enabled;
794 struct kvm_mmu *mmu = vcpu->arch.mmu;
795 struct tdp_iter iter;
796 struct kvm_mmu_page *sp;
800 gfn_t gfn = gpa >> PAGE_SHIFT;
804 if (WARN_ON(!VALID_PAGE(vcpu->arch.mmu->root_hpa)))
806 if (WARN_ON(!is_tdp_mmu_root(vcpu->kvm, vcpu->arch.mmu->root_hpa)))
809 level = kvm_mmu_hugepage_adjust(vcpu, gfn, max_level, &pfn,
810 huge_page_disallowed, &req_level);
812 trace_kvm_mmu_spte_requested(gpa, level, pfn);
816 tdp_mmu_for_each_pte(iter, mmu, gfn, gfn + 1) {
817 if (nx_huge_page_workaround_enabled)
818 disallowed_hugepage_adjust(iter.old_spte, gfn,
819 iter.level, &pfn, &level);
821 if (iter.level == level)
825 * If there is an SPTE mapping a large page at a higher level
826 * than the target, that SPTE must be cleared and replaced
827 * with a non-leaf SPTE.
829 if (is_shadow_present_pte(iter.old_spte) &&
830 is_large_pte(iter.old_spte)) {
831 if (!tdp_mmu_zap_spte_atomic(vcpu->kvm, &iter))
835 * The iter must explicitly re-read the spte here
836 * because the new value informs the !present
839 iter.old_spte = READ_ONCE(*rcu_dereference(iter.sptep));
842 if (!is_shadow_present_pte(iter.old_spte)) {
843 sp = alloc_tdp_mmu_page(vcpu, iter.gfn, iter.level);
846 new_spte = make_nonleaf_spte(child_pt,
847 !shadow_accessed_mask);
849 if (tdp_mmu_set_spte_atomic(vcpu->kvm, &iter,
851 tdp_mmu_link_page(vcpu->kvm, sp, true,
852 huge_page_disallowed &&
853 req_level >= iter.level);
855 trace_kvm_mmu_get_page(sp, true);
863 if (iter.level != level) {
868 ret = tdp_mmu_map_handle_target_level(vcpu, write, map_writable, &iter,
875 typedef int (*tdp_handler_t)(struct kvm *kvm, struct kvm_memory_slot *slot,
876 struct kvm_mmu_page *root, gfn_t start, gfn_t end,
879 static __always_inline int kvm_tdp_mmu_handle_hva_range(struct kvm *kvm,
883 tdp_handler_t handler)
885 struct kvm_memslots *slots;
886 struct kvm_memory_slot *memslot;
887 struct kvm_mmu_page *root;
891 for_each_tdp_mmu_root_yield_safe(kvm, root) {
892 as_id = kvm_mmu_page_as_id(root);
893 slots = __kvm_memslots(kvm, as_id);
894 kvm_for_each_memslot(memslot, slots) {
895 unsigned long hva_start, hva_end;
896 gfn_t gfn_start, gfn_end;
898 hva_start = max(start, memslot->userspace_addr);
899 hva_end = min(end, memslot->userspace_addr +
900 (memslot->npages << PAGE_SHIFT));
901 if (hva_start >= hva_end)
904 * {gfn(page) | page intersects with [hva_start, hva_end)} =
905 * {gfn_start, gfn_start+1, ..., gfn_end-1}.
907 gfn_start = hva_to_gfn_memslot(hva_start, memslot);
908 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
910 ret |= handler(kvm, memslot, root, gfn_start,
918 static __always_inline int kvm_tdp_mmu_handle_hva(struct kvm *kvm,
921 tdp_handler_t handler)
923 return kvm_tdp_mmu_handle_hva_range(kvm, addr, addr + 1, data, handler);
926 static int zap_gfn_range_hva_wrapper(struct kvm *kvm,
927 struct kvm_memory_slot *slot,
928 struct kvm_mmu_page *root, gfn_t start,
929 gfn_t end, unsigned long unused)
931 return zap_gfn_range(kvm, root, start, end, false, false);
934 int kvm_tdp_mmu_zap_hva_range(struct kvm *kvm, unsigned long start,
937 return kvm_tdp_mmu_handle_hva_range(kvm, start, end, 0,
938 zap_gfn_range_hva_wrapper);
942 * Mark the SPTEs range of GFNs [start, end) unaccessed and return non-zero
943 * if any of the GFNs in the range have been accessed.
945 static int age_gfn_range(struct kvm *kvm, struct kvm_memory_slot *slot,
946 struct kvm_mmu_page *root, gfn_t start, gfn_t end,
947 unsigned long unused)
949 struct tdp_iter iter;
955 tdp_root_for_each_leaf_pte(iter, root, start, end) {
957 * If we have a non-accessed entry we don't need to change the
960 if (!is_accessed_spte(iter.old_spte))
963 new_spte = iter.old_spte;
965 if (spte_ad_enabled(new_spte)) {
966 clear_bit((ffs(shadow_accessed_mask) - 1),
967 (unsigned long *)&new_spte);
970 * Capture the dirty status of the page, so that it doesn't get
971 * lost when the SPTE is marked for access tracking.
973 if (is_writable_pte(new_spte))
974 kvm_set_pfn_dirty(spte_to_pfn(new_spte));
976 new_spte = mark_spte_for_access_track(new_spte);
978 new_spte &= ~shadow_dirty_mask;
980 tdp_mmu_set_spte_no_acc_track(kvm, &iter, new_spte);
983 trace_kvm_age_page(iter.gfn, iter.level, slot, young);
991 int kvm_tdp_mmu_age_hva_range(struct kvm *kvm, unsigned long start,
994 return kvm_tdp_mmu_handle_hva_range(kvm, start, end, 0,
998 static int test_age_gfn(struct kvm *kvm, struct kvm_memory_slot *slot,
999 struct kvm_mmu_page *root, gfn_t gfn, gfn_t end,
1000 unsigned long unused)
1002 struct tdp_iter iter;
1004 tdp_root_for_each_leaf_pte(iter, root, gfn, end)
1005 if (is_accessed_spte(iter.old_spte))
1011 int kvm_tdp_mmu_test_age_hva(struct kvm *kvm, unsigned long hva)
1013 return kvm_tdp_mmu_handle_hva(kvm, hva, 0, test_age_gfn);
1017 * Handle the changed_pte MMU notifier for the TDP MMU.
1018 * data is a pointer to the new pte_t mapping the HVA specified by the MMU
1020 * Returns non-zero if a flush is needed before releasing the MMU lock.
1022 static int set_tdp_spte(struct kvm *kvm, struct kvm_memory_slot *slot,
1023 struct kvm_mmu_page *root, gfn_t gfn, gfn_t end,
1026 struct tdp_iter iter;
1027 pte_t *ptep = (pte_t *)data;
1034 WARN_ON(pte_huge(*ptep) || (gfn + 1) != end);
1036 new_pfn = pte_pfn(*ptep);
1038 tdp_root_for_each_pte(iter, root, gfn, gfn + 1) {
1039 if (iter.level != PG_LEVEL_4K)
1042 if (!is_shadow_present_pte(iter.old_spte))
1046 * Note, when changing a read-only SPTE, it's not strictly
1047 * necessary to zero the SPTE before setting the new PFN, but
1048 * doing so preserves the invariant that the PFN of a present
1049 * leaf SPTE can never change. See __handle_changed_spte().
1051 tdp_mmu_set_spte(kvm, &iter, 0);
1053 if (!pte_write(*ptep)) {
1054 new_spte = kvm_mmu_changed_pte_notifier_make_spte(
1055 iter.old_spte, new_pfn);
1057 tdp_mmu_set_spte(kvm, &iter, new_spte);
1064 kvm_flush_remote_tlbs_with_address(kvm, gfn, 1);
1071 int kvm_tdp_mmu_set_spte_hva(struct kvm *kvm, unsigned long address,
1074 return kvm_tdp_mmu_handle_hva(kvm, address, (unsigned long)host_ptep,
1079 * Remove write access from all the SPTEs mapping GFNs [start, end). If
1080 * skip_4k is set, SPTEs that map 4k pages, will not be write-protected.
1081 * Returns true if an SPTE has been changed and the TLBs need to be flushed.
1083 static bool wrprot_gfn_range(struct kvm *kvm, struct kvm_mmu_page *root,
1084 gfn_t start, gfn_t end, int min_level)
1086 struct tdp_iter iter;
1088 bool spte_set = false;
1092 BUG_ON(min_level > KVM_MAX_HUGEPAGE_LEVEL);
1094 for_each_tdp_pte_min_level(iter, root->spt, root->role.level,
1095 min_level, start, end) {
1096 if (tdp_mmu_iter_cond_resched(kvm, &iter, false))
1099 if (!is_shadow_present_pte(iter.old_spte) ||
1100 !is_last_spte(iter.old_spte, iter.level) ||
1101 !(iter.old_spte & PT_WRITABLE_MASK))
1104 new_spte = iter.old_spte & ~PT_WRITABLE_MASK;
1106 tdp_mmu_set_spte_no_dirty_log(kvm, &iter, new_spte);
1115 * Remove write access from all the SPTEs mapping GFNs in the memslot. Will
1116 * only affect leaf SPTEs down to min_level.
1117 * Returns true if an SPTE has been changed and the TLBs need to be flushed.
1119 bool kvm_tdp_mmu_wrprot_slot(struct kvm *kvm, struct kvm_memory_slot *slot,
1122 struct kvm_mmu_page *root;
1124 bool spte_set = false;
1126 for_each_tdp_mmu_root_yield_safe(kvm, root) {
1127 root_as_id = kvm_mmu_page_as_id(root);
1128 if (root_as_id != slot->as_id)
1131 spte_set |= wrprot_gfn_range(kvm, root, slot->base_gfn,
1132 slot->base_gfn + slot->npages, min_level);
1139 * Clear the dirty status of all the SPTEs mapping GFNs in the memslot. If
1140 * AD bits are enabled, this will involve clearing the dirty bit on each SPTE.
1141 * If AD bits are not enabled, this will require clearing the writable bit on
1142 * each SPTE. Returns true if an SPTE has been changed and the TLBs need to
1145 static bool clear_dirty_gfn_range(struct kvm *kvm, struct kvm_mmu_page *root,
1146 gfn_t start, gfn_t end)
1148 struct tdp_iter iter;
1150 bool spte_set = false;
1154 tdp_root_for_each_leaf_pte(iter, root, start, end) {
1155 if (tdp_mmu_iter_cond_resched(kvm, &iter, false))
1158 if (spte_ad_need_write_protect(iter.old_spte)) {
1159 if (is_writable_pte(iter.old_spte))
1160 new_spte = iter.old_spte & ~PT_WRITABLE_MASK;
1164 if (iter.old_spte & shadow_dirty_mask)
1165 new_spte = iter.old_spte & ~shadow_dirty_mask;
1170 tdp_mmu_set_spte_no_dirty_log(kvm, &iter, new_spte);
1179 * Clear the dirty status of all the SPTEs mapping GFNs in the memslot. If
1180 * AD bits are enabled, this will involve clearing the dirty bit on each SPTE.
1181 * If AD bits are not enabled, this will require clearing the writable bit on
1182 * each SPTE. Returns true if an SPTE has been changed and the TLBs need to
1185 bool kvm_tdp_mmu_clear_dirty_slot(struct kvm *kvm, struct kvm_memory_slot *slot)
1187 struct kvm_mmu_page *root;
1189 bool spte_set = false;
1191 for_each_tdp_mmu_root_yield_safe(kvm, root) {
1192 root_as_id = kvm_mmu_page_as_id(root);
1193 if (root_as_id != slot->as_id)
1196 spte_set |= clear_dirty_gfn_range(kvm, root, slot->base_gfn,
1197 slot->base_gfn + slot->npages);
1204 * Clears the dirty status of all the 4k SPTEs mapping GFNs for which a bit is
1205 * set in mask, starting at gfn. The given memslot is expected to contain all
1206 * the GFNs represented by set bits in the mask. If AD bits are enabled,
1207 * clearing the dirty status will involve clearing the dirty bit on each SPTE
1208 * or, if AD bits are not enabled, clearing the writable bit on each SPTE.
1210 static void clear_dirty_pt_masked(struct kvm *kvm, struct kvm_mmu_page *root,
1211 gfn_t gfn, unsigned long mask, bool wrprot)
1213 struct tdp_iter iter;
1218 tdp_root_for_each_leaf_pte(iter, root, gfn + __ffs(mask),
1219 gfn + BITS_PER_LONG) {
1223 if (iter.level > PG_LEVEL_4K ||
1224 !(mask & (1UL << (iter.gfn - gfn))))
1227 mask &= ~(1UL << (iter.gfn - gfn));
1229 if (wrprot || spte_ad_need_write_protect(iter.old_spte)) {
1230 if (is_writable_pte(iter.old_spte))
1231 new_spte = iter.old_spte & ~PT_WRITABLE_MASK;
1235 if (iter.old_spte & shadow_dirty_mask)
1236 new_spte = iter.old_spte & ~shadow_dirty_mask;
1241 tdp_mmu_set_spte_no_dirty_log(kvm, &iter, new_spte);
1248 * Clears the dirty status of all the 4k SPTEs mapping GFNs for which a bit is
1249 * set in mask, starting at gfn. The given memslot is expected to contain all
1250 * the GFNs represented by set bits in the mask. If AD bits are enabled,
1251 * clearing the dirty status will involve clearing the dirty bit on each SPTE
1252 * or, if AD bits are not enabled, clearing the writable bit on each SPTE.
1254 void kvm_tdp_mmu_clear_dirty_pt_masked(struct kvm *kvm,
1255 struct kvm_memory_slot *slot,
1256 gfn_t gfn, unsigned long mask,
1259 struct kvm_mmu_page *root;
1262 lockdep_assert_held_write(&kvm->mmu_lock);
1263 for_each_tdp_mmu_root(kvm, root) {
1264 root_as_id = kvm_mmu_page_as_id(root);
1265 if (root_as_id != slot->as_id)
1268 clear_dirty_pt_masked(kvm, root, gfn, mask, wrprot);
1273 * Clear leaf entries which could be replaced by large mappings, for
1274 * GFNs within the slot.
1276 static bool zap_collapsible_spte_range(struct kvm *kvm,
1277 struct kvm_mmu_page *root,
1278 struct kvm_memory_slot *slot,
1281 gfn_t start = slot->base_gfn;
1282 gfn_t end = start + slot->npages;
1283 struct tdp_iter iter;
1288 tdp_root_for_each_pte(iter, root, start, end) {
1289 if (tdp_mmu_iter_cond_resched(kvm, &iter, flush)) {
1294 if (!is_shadow_present_pte(iter.old_spte) ||
1295 !is_last_spte(iter.old_spte, iter.level))
1298 pfn = spte_to_pfn(iter.old_spte);
1299 if (kvm_is_reserved_pfn(pfn) ||
1300 iter.level >= kvm_mmu_max_mapping_level(kvm, slot, iter.gfn,
1304 tdp_mmu_set_spte(kvm, &iter, 0);
1315 * Clear non-leaf entries (and free associated page tables) which could
1316 * be replaced by large mappings, for GFNs within the slot.
1318 bool kvm_tdp_mmu_zap_collapsible_sptes(struct kvm *kvm,
1319 struct kvm_memory_slot *slot, bool flush)
1321 struct kvm_mmu_page *root;
1324 for_each_tdp_mmu_root_yield_safe(kvm, root) {
1325 root_as_id = kvm_mmu_page_as_id(root);
1326 if (root_as_id != slot->as_id)
1329 flush = zap_collapsible_spte_range(kvm, root, slot, flush);
1336 * Removes write access on the last level SPTE mapping this GFN and unsets the
1337 * MMU-writable bit to ensure future writes continue to be intercepted.
1338 * Returns true if an SPTE was set and a TLB flush is needed.
1340 static bool write_protect_gfn(struct kvm *kvm, struct kvm_mmu_page *root,
1343 struct tdp_iter iter;
1345 bool spte_set = false;
1349 tdp_root_for_each_leaf_pte(iter, root, gfn, gfn + 1) {
1350 if (!is_writable_pte(iter.old_spte))
1353 new_spte = iter.old_spte &
1354 ~(PT_WRITABLE_MASK | shadow_mmu_writable_mask);
1356 tdp_mmu_set_spte(kvm, &iter, new_spte);
1366 * Removes write access on the last level SPTE mapping this GFN and unsets the
1367 * MMU-writable bit to ensure future writes continue to be intercepted.
1368 * Returns true if an SPTE was set and a TLB flush is needed.
1370 bool kvm_tdp_mmu_write_protect_gfn(struct kvm *kvm,
1371 struct kvm_memory_slot *slot, gfn_t gfn)
1373 struct kvm_mmu_page *root;
1375 bool spte_set = false;
1377 lockdep_assert_held_write(&kvm->mmu_lock);
1378 for_each_tdp_mmu_root(kvm, root) {
1379 root_as_id = kvm_mmu_page_as_id(root);
1380 if (root_as_id != slot->as_id)
1383 spte_set |= write_protect_gfn(kvm, root, gfn);
1389 * Return the level of the lowest level SPTE added to sptes.
1390 * That SPTE may be non-present.
1392 int kvm_tdp_mmu_get_walk(struct kvm_vcpu *vcpu, u64 addr, u64 *sptes,
1395 struct tdp_iter iter;
1396 struct kvm_mmu *mmu = vcpu->arch.mmu;
1397 gfn_t gfn = addr >> PAGE_SHIFT;
1400 *root_level = vcpu->arch.mmu->shadow_root_level;
1404 tdp_mmu_for_each_pte(iter, mmu, gfn, gfn + 1) {
1406 sptes[leaf] = iter.old_spte;