1 // SPDX-License-Identifier: GPL-2.0-only
3 * Kernel-based Virtual Machine driver for Linux
5 * This module enables kernel and guest-mode vCPU access to guest physical
6 * memory with suitable invalidation mechanisms.
8 * Copyright © 2021 Amazon.com, Inc. or its affiliates.
11 * David Woodhouse <dwmw2@infradead.org>
14 #include <linux/kvm_host.h>
15 #include <linux/kvm.h>
16 #include <linux/highmem.h>
17 #include <linux/module.h>
18 #include <linux/errno.h>
23 * MMU notifier 'invalidate_range_start' hook.
25 void gfn_to_pfn_cache_invalidate_start(struct kvm *kvm, unsigned long start,
26 unsigned long end, bool may_block)
28 DECLARE_BITMAP(vcpu_bitmap, KVM_MAX_VCPUS);
29 struct gfn_to_pfn_cache *gpc;
30 bool evict_vcpus = false;
32 spin_lock(&kvm->gpc_lock);
33 list_for_each_entry(gpc, &kvm->gpc_list, list) {
34 write_lock_irq(&gpc->lock);
36 /* Only a single page so no need to care about length */
37 if (gpc->valid && !is_error_noslot_pfn(gpc->pfn) &&
38 gpc->uhva >= start && gpc->uhva < end) {
42 * If a guest vCPU could be using the physical address,
43 * it needs to be forced out of guest mode.
45 if (gpc->usage & KVM_GUEST_USES_PFN) {
48 bitmap_zero(vcpu_bitmap, KVM_MAX_VCPUS);
50 __set_bit(gpc->vcpu->vcpu_idx, vcpu_bitmap);
53 write_unlock_irq(&gpc->lock);
55 spin_unlock(&kvm->gpc_lock);
59 * KVM needs to ensure the vCPU is fully out of guest context
60 * before allowing the invalidation to continue.
62 unsigned int req = KVM_REQ_OUTSIDE_GUEST_MODE;
66 * If the OOM reaper is active, then all vCPUs should have
67 * been stopped already, so perform the request without
68 * KVM_REQUEST_WAIT and be sad if any needed to be IPI'd.
71 req &= ~KVM_REQUEST_WAIT;
73 called = kvm_make_vcpus_request_mask(kvm, req, vcpu_bitmap);
75 WARN_ON_ONCE(called && !may_block);
79 bool kvm_gfn_to_pfn_cache_check(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,
80 gpa_t gpa, unsigned long len)
82 struct kvm_memslots *slots = kvm_memslots(kvm);
87 if ((gpa & ~PAGE_MASK) + len > PAGE_SIZE)
90 if (gpc->gpa != gpa || gpc->generation != slots->generation ||
91 kvm_is_error_hva(gpc->uhva))
99 EXPORT_SYMBOL_GPL(kvm_gfn_to_pfn_cache_check);
101 static void gpc_unmap_khva(struct kvm *kvm, kvm_pfn_t pfn, void *khva)
103 /* Unmap the old pfn/page if it was mapped before. */
104 if (!is_error_noslot_pfn(pfn) && khva) {
106 kunmap(pfn_to_page(pfn));
107 #ifdef CONFIG_HAS_IOMEM
114 static inline bool mmu_notifier_retry_cache(struct kvm *kvm, unsigned long mmu_seq)
117 * mn_active_invalidate_count acts for all intents and purposes
118 * like mmu_invalidate_in_progress here; but the latter cannot
119 * be used here because the invalidation of caches in the
120 * mmu_notifier event occurs _before_ mmu_invalidate_in_progress
123 * Note, it does not matter that mn_active_invalidate_count
124 * is not protected by gpc->lock. It is guaranteed to
125 * be elevated before the mmu_notifier acquires gpc->lock, and
126 * isn't dropped until after mmu_invalidate_seq is updated.
128 if (kvm->mn_active_invalidate_count)
132 * Ensure mn_active_invalidate_count is read before
133 * mmu_invalidate_seq. This pairs with the smp_wmb() in
134 * mmu_notifier_invalidate_range_end() to guarantee either the
135 * old (non-zero) value of mn_active_invalidate_count or the
136 * new (incremented) value of mmu_invalidate_seq is observed.
139 return kvm->mmu_invalidate_seq != mmu_seq;
142 static kvm_pfn_t hva_to_pfn_retry(struct kvm *kvm, struct gfn_to_pfn_cache *gpc)
144 /* Note, the new page offset may be different than the old! */
145 void *old_khva = gpc->khva - offset_in_page(gpc->khva);
146 kvm_pfn_t new_pfn = KVM_PFN_ERR_FAULT;
147 void *new_khva = NULL;
148 unsigned long mmu_seq;
150 lockdep_assert_held(&gpc->refresh_lock);
152 lockdep_assert_held_write(&gpc->lock);
155 * Invalidate the cache prior to dropping gpc->lock, the gpa=>uhva
156 * assets have already been updated and so a concurrent check() from a
157 * different task may not fail the gpa/uhva/generation checks.
162 mmu_seq = kvm->mmu_invalidate_seq;
165 write_unlock_irq(&gpc->lock);
168 * If the previous iteration "failed" due to an mmu_notifier
169 * event, release the pfn and unmap the kernel virtual address
170 * from the previous attempt. Unmapping might sleep, so this
171 * needs to be done after dropping the lock. Opportunistically
172 * check for resched while the lock isn't held.
174 if (new_pfn != KVM_PFN_ERR_FAULT) {
176 * Keep the mapping if the previous iteration reused
177 * the existing mapping and didn't create a new one.
179 if (new_khva != old_khva)
180 gpc_unmap_khva(kvm, new_pfn, new_khva);
182 kvm_release_pfn_clean(new_pfn);
187 /* We always request a writeable mapping */
188 new_pfn = hva_to_pfn(gpc->uhva, false, NULL, true, NULL);
189 if (is_error_noslot_pfn(new_pfn))
193 * Obtain a new kernel mapping if KVM itself will access the
194 * pfn. Note, kmap() and memremap() can both sleep, so this
195 * too must be done outside of gpc->lock!
197 if (gpc->usage & KVM_HOST_USES_PFN) {
198 if (new_pfn == gpc->pfn) {
200 } else if (pfn_valid(new_pfn)) {
201 new_khva = kmap(pfn_to_page(new_pfn));
202 #ifdef CONFIG_HAS_IOMEM
204 new_khva = memremap(pfn_to_hpa(new_pfn), PAGE_SIZE, MEMREMAP_WB);
208 kvm_release_pfn_clean(new_pfn);
213 write_lock_irq(&gpc->lock);
216 * Other tasks must wait for _this_ refresh to complete before
217 * attempting to refresh.
219 WARN_ON_ONCE(gpc->valid);
220 } while (mmu_notifier_retry_cache(kvm, mmu_seq));
224 gpc->khva = new_khva + (gpc->gpa & ~PAGE_MASK);
227 * Put the reference to the _new_ pfn. The pfn is now tracked by the
228 * cache and can be safely migrated, swapped, etc... as the cache will
229 * invalidate any mappings in response to relevant mmu_notifier events.
231 kvm_release_pfn_clean(new_pfn);
236 write_lock_irq(&gpc->lock);
241 int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,
242 gpa_t gpa, unsigned long len)
244 struct kvm_memslots *slots = kvm_memslots(kvm);
245 unsigned long page_offset = gpa & ~PAGE_MASK;
246 bool unmap_old = false;
247 unsigned long old_uhva;
253 * If must fit within a single page. The 'len' argument is
254 * only to enforce that.
256 if (page_offset + len > PAGE_SIZE)
260 * If another task is refreshing the cache, wait for it to complete.
261 * There is no guarantee that concurrent refreshes will see the same
262 * gpa, memslots generation, etc..., so they must be fully serialized.
264 mutex_lock(&gpc->refresh_lock);
266 write_lock_irq(&gpc->lock);
274 old_khva = gpc->khva - offset_in_page(gpc->khva);
275 old_uhva = gpc->uhva;
277 /* If the userspace HVA is invalid, refresh that first */
278 if (gpc->gpa != gpa || gpc->generation != slots->generation ||
279 kvm_is_error_hva(gpc->uhva)) {
280 gfn_t gfn = gpa_to_gfn(gpa);
283 gpc->generation = slots->generation;
284 gpc->memslot = __gfn_to_memslot(slots, gfn);
285 gpc->uhva = gfn_to_hva_memslot(gpc->memslot, gfn);
287 if (kvm_is_error_hva(gpc->uhva)) {
294 * If the userspace HVA changed or the PFN was already invalid,
295 * drop the lock and do the HVA to PFN lookup again.
297 if (!gpc->valid || old_uhva != gpc->uhva) {
298 ret = hva_to_pfn_retry(kvm, gpc);
301 * If the HVA→PFN mapping was already valid, don't unmap it.
302 * But do update gpc->khva because the offset within the page
305 gpc->khva = old_khva + page_offset;
306 old_pfn = KVM_PFN_ERR_FAULT;
313 * Invalidate the cache and purge the pfn/khva if the refresh failed.
314 * Some/all of the uhva, gpa, and memslot generation info may still be
315 * valid, leave it as is.
319 gpc->pfn = KVM_PFN_ERR_FAULT;
323 /* Detect a pfn change before dropping the lock! */
324 unmap_old = (old_pfn != gpc->pfn);
327 write_unlock_irq(&gpc->lock);
329 mutex_unlock(&gpc->refresh_lock);
332 gpc_unmap_khva(kvm, old_pfn, old_khva);
336 EXPORT_SYMBOL_GPL(kvm_gfn_to_pfn_cache_refresh);
338 void kvm_gfn_to_pfn_cache_unmap(struct kvm *kvm, struct gfn_to_pfn_cache *gpc)
343 mutex_lock(&gpc->refresh_lock);
344 write_lock_irq(&gpc->lock);
348 old_khva = gpc->khva - offset_in_page(gpc->khva);
352 * We can leave the GPA → uHVA map cache intact but the PFN
353 * lookup will need to be redone even for the same page.
356 gpc->pfn = KVM_PFN_ERR_FAULT;
358 write_unlock_irq(&gpc->lock);
359 mutex_unlock(&gpc->refresh_lock);
361 gpc_unmap_khva(kvm, old_pfn, old_khva);
363 EXPORT_SYMBOL_GPL(kvm_gfn_to_pfn_cache_unmap);
365 void kvm_gpc_init(struct gfn_to_pfn_cache *gpc)
367 rwlock_init(&gpc->lock);
368 mutex_init(&gpc->refresh_lock);
370 EXPORT_SYMBOL_GPL(kvm_gpc_init);
372 int kvm_gpc_activate(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,
373 struct kvm_vcpu *vcpu, enum pfn_cache_usage usage,
374 gpa_t gpa, unsigned long len)
376 WARN_ON_ONCE(!usage || (usage & KVM_GUEST_AND_HOST_USE_PFN) != usage);
380 gpc->pfn = KVM_PFN_ERR_FAULT;
381 gpc->uhva = KVM_HVA_ERR_BAD;
386 spin_lock(&kvm->gpc_lock);
387 list_add(&gpc->list, &kvm->gpc_list);
388 spin_unlock(&kvm->gpc_lock);
391 * Activate the cache after adding it to the list, a concurrent
392 * refresh must not establish a mapping until the cache is
393 * reachable by mmu_notifier events.
395 write_lock_irq(&gpc->lock);
397 write_unlock_irq(&gpc->lock);
399 return kvm_gfn_to_pfn_cache_refresh(kvm, gpc, gpa, len);
401 EXPORT_SYMBOL_GPL(kvm_gpc_activate);
403 void kvm_gpc_deactivate(struct kvm *kvm, struct gfn_to_pfn_cache *gpc)
407 * Deactivate the cache before removing it from the list, KVM
408 * must stall mmu_notifier events until all users go away, i.e.
409 * until gpc->lock is dropped and refresh is guaranteed to fail.
411 write_lock_irq(&gpc->lock);
413 write_unlock_irq(&gpc->lock);
415 spin_lock(&kvm->gpc_lock);
416 list_del(&gpc->list);
417 spin_unlock(&kvm->gpc_lock);
419 kvm_gfn_to_pfn_cache_unmap(kvm, gpc);
422 EXPORT_SYMBOL_GPL(kvm_gpc_deactivate);