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,
28 struct gfn_to_pfn_cache *gpc;
30 spin_lock(&kvm->gpc_lock);
31 list_for_each_entry(gpc, &kvm->gpc_list, list) {
32 read_lock_irq(&gpc->lock);
34 /* Only a single page so no need to care about length */
35 if (gpc->valid && !is_error_noslot_pfn(gpc->pfn) &&
36 gpc->uhva >= start && gpc->uhva < end) {
37 read_unlock_irq(&gpc->lock);
40 * There is a small window here where the cache could
41 * be modified, and invalidation would no longer be
42 * necessary. Hence check again whether invalidation
43 * is still necessary once the write lock has been
47 write_lock_irq(&gpc->lock);
48 if (gpc->valid && !is_error_noslot_pfn(gpc->pfn) &&
49 gpc->uhva >= start && gpc->uhva < end)
51 write_unlock_irq(&gpc->lock);
55 read_unlock_irq(&gpc->lock);
57 spin_unlock(&kvm->gpc_lock);
60 static bool kvm_gpc_is_valid_len(gpa_t gpa, unsigned long uhva,
63 unsigned long offset = kvm_is_error_gpa(gpa) ? offset_in_page(uhva) :
67 * The cached access must fit within a single page. The 'len' argument
68 * to activate() and refresh() exists only to enforce that.
70 return offset + len <= PAGE_SIZE;
73 bool kvm_gpc_check(struct gfn_to_pfn_cache *gpc, unsigned long len)
75 struct kvm_memslots *slots = kvm_memslots(gpc->kvm);
81 * If the page was cached from a memslot, make sure the memslots have
82 * not been re-configured.
84 if (!kvm_is_error_gpa(gpc->gpa) && gpc->generation != slots->generation)
87 if (kvm_is_error_hva(gpc->uhva))
90 if (!kvm_gpc_is_valid_len(gpc->gpa, gpc->uhva, len))
99 static void *gpc_map(kvm_pfn_t pfn)
102 return kmap(pfn_to_page(pfn));
104 #ifdef CONFIG_HAS_IOMEM
105 return memremap(pfn_to_hpa(pfn), PAGE_SIZE, MEMREMAP_WB);
111 static void gpc_unmap(kvm_pfn_t pfn, void *khva)
113 /* Unmap the old pfn/page if it was mapped before. */
114 if (is_error_noslot_pfn(pfn) || !khva)
117 if (pfn_valid(pfn)) {
118 kunmap(pfn_to_page(pfn));
122 #ifdef CONFIG_HAS_IOMEM
127 static inline bool mmu_notifier_retry_cache(struct kvm *kvm, unsigned long mmu_seq)
130 * mn_active_invalidate_count acts for all intents and purposes
131 * like mmu_invalidate_in_progress here; but the latter cannot
132 * be used here because the invalidation of caches in the
133 * mmu_notifier event occurs _before_ mmu_invalidate_in_progress
136 * Note, it does not matter that mn_active_invalidate_count
137 * is not protected by gpc->lock. It is guaranteed to
138 * be elevated before the mmu_notifier acquires gpc->lock, and
139 * isn't dropped until after mmu_invalidate_seq is updated.
141 if (kvm->mn_active_invalidate_count)
145 * Ensure mn_active_invalidate_count is read before
146 * mmu_invalidate_seq. This pairs with the smp_wmb() in
147 * mmu_notifier_invalidate_range_end() to guarantee either the
148 * old (non-zero) value of mn_active_invalidate_count or the
149 * new (incremented) value of mmu_invalidate_seq is observed.
152 return kvm->mmu_invalidate_seq != mmu_seq;
155 static kvm_pfn_t hva_to_pfn_retry(struct gfn_to_pfn_cache *gpc)
157 /* Note, the new page offset may be different than the old! */
158 void *old_khva = (void *)PAGE_ALIGN_DOWN((uintptr_t)gpc->khva);
159 kvm_pfn_t new_pfn = KVM_PFN_ERR_FAULT;
160 void *new_khva = NULL;
161 unsigned long mmu_seq;
163 lockdep_assert_held(&gpc->refresh_lock);
165 lockdep_assert_held_write(&gpc->lock);
168 * Invalidate the cache prior to dropping gpc->lock, the gpa=>uhva
169 * assets have already been updated and so a concurrent check() from a
170 * different task may not fail the gpa/uhva/generation checks.
175 mmu_seq = gpc->kvm->mmu_invalidate_seq;
178 write_unlock_irq(&gpc->lock);
181 * If the previous iteration "failed" due to an mmu_notifier
182 * event, release the pfn and unmap the kernel virtual address
183 * from the previous attempt. Unmapping might sleep, so this
184 * needs to be done after dropping the lock. Opportunistically
185 * check for resched while the lock isn't held.
187 if (new_pfn != KVM_PFN_ERR_FAULT) {
189 * Keep the mapping if the previous iteration reused
190 * the existing mapping and didn't create a new one.
192 if (new_khva != old_khva)
193 gpc_unmap(new_pfn, new_khva);
195 kvm_release_pfn_clean(new_pfn);
200 /* We always request a writeable mapping */
201 new_pfn = hva_to_pfn(gpc->uhva, false, false, NULL, true, NULL);
202 if (is_error_noslot_pfn(new_pfn))
206 * Obtain a new kernel mapping if KVM itself will access the
207 * pfn. Note, kmap() and memremap() can both sleep, so this
208 * too must be done outside of gpc->lock!
210 if (new_pfn == gpc->pfn)
213 new_khva = gpc_map(new_pfn);
216 kvm_release_pfn_clean(new_pfn);
220 write_lock_irq(&gpc->lock);
223 * Other tasks must wait for _this_ refresh to complete before
224 * attempting to refresh.
226 WARN_ON_ONCE(gpc->valid);
227 } while (mmu_notifier_retry_cache(gpc->kvm, mmu_seq));
231 gpc->khva = new_khva + offset_in_page(gpc->uhva);
234 * Put the reference to the _new_ pfn. The pfn is now tracked by the
235 * cache and can be safely migrated, swapped, etc... as the cache will
236 * invalidate any mappings in response to relevant mmu_notifier events.
238 kvm_release_pfn_clean(new_pfn);
243 write_lock_irq(&gpc->lock);
248 static int __kvm_gpc_refresh(struct gfn_to_pfn_cache *gpc, gpa_t gpa, unsigned long uhva)
250 unsigned long page_offset;
251 bool unmap_old = false;
252 unsigned long old_uhva;
254 bool hva_change = false;
258 /* Either gpa or uhva must be valid, but not both */
259 if (WARN_ON_ONCE(kvm_is_error_gpa(gpa) == kvm_is_error_hva(uhva)))
262 lockdep_assert_held(&gpc->refresh_lock);
264 write_lock_irq(&gpc->lock);
272 old_khva = (void *)PAGE_ALIGN_DOWN((uintptr_t)gpc->khva);
273 old_uhva = PAGE_ALIGN_DOWN(gpc->uhva);
275 if (kvm_is_error_gpa(gpa)) {
276 page_offset = offset_in_page(uhva);
278 gpc->gpa = INVALID_GPA;
280 gpc->uhva = PAGE_ALIGN_DOWN(uhva);
282 if (gpc->uhva != old_uhva)
285 struct kvm_memslots *slots = kvm_memslots(gpc->kvm);
287 page_offset = offset_in_page(gpa);
289 if (gpc->gpa != gpa || gpc->generation != slots->generation ||
290 kvm_is_error_hva(gpc->uhva)) {
291 gfn_t gfn = gpa_to_gfn(gpa);
294 gpc->generation = slots->generation;
295 gpc->memslot = __gfn_to_memslot(slots, gfn);
296 gpc->uhva = gfn_to_hva_memslot(gpc->memslot, gfn);
298 if (kvm_is_error_hva(gpc->uhva)) {
304 * Even if the GPA and/or the memslot generation changed, the
305 * HVA may still be the same.
307 if (gpc->uhva != old_uhva)
310 gpc->uhva = old_uhva;
314 /* Note: the offset must be correct before calling hva_to_pfn_retry() */
315 gpc->uhva += page_offset;
318 * If the userspace HVA changed or the PFN was already invalid,
319 * drop the lock and do the HVA to PFN lookup again.
321 if (!gpc->valid || hva_change) {
322 ret = hva_to_pfn_retry(gpc);
325 * If the HVA→PFN mapping was already valid, don't unmap it.
326 * But do update gpc->khva because the offset within the page
329 gpc->khva = old_khva + page_offset;
336 * Invalidate the cache and purge the pfn/khva if the refresh failed.
337 * Some/all of the uhva, gpa, and memslot generation info may still be
338 * valid, leave it as is.
342 gpc->pfn = KVM_PFN_ERR_FAULT;
346 /* Detect a pfn change before dropping the lock! */
347 unmap_old = (old_pfn != gpc->pfn);
350 write_unlock_irq(&gpc->lock);
353 gpc_unmap(old_pfn, old_khva);
358 int kvm_gpc_refresh(struct gfn_to_pfn_cache *gpc, unsigned long len)
362 guard(mutex)(&gpc->refresh_lock);
364 if (!kvm_gpc_is_valid_len(gpc->gpa, gpc->uhva, len))
368 * If the GPA is valid then ignore the HVA, as a cache can be GPA-based
369 * or HVA-based, not both. For GPA-based caches, the HVA will be
370 * recomputed during refresh if necessary.
372 uhva = kvm_is_error_gpa(gpc->gpa) ? gpc->uhva : KVM_HVA_ERR_BAD;
374 return __kvm_gpc_refresh(gpc, gpc->gpa, uhva);
377 void kvm_gpc_init(struct gfn_to_pfn_cache *gpc, struct kvm *kvm)
379 rwlock_init(&gpc->lock);
380 mutex_init(&gpc->refresh_lock);
383 gpc->pfn = KVM_PFN_ERR_FAULT;
384 gpc->gpa = INVALID_GPA;
385 gpc->uhva = KVM_HVA_ERR_BAD;
386 gpc->active = gpc->valid = false;
389 static int __kvm_gpc_activate(struct gfn_to_pfn_cache *gpc, gpa_t gpa, unsigned long uhva,
392 struct kvm *kvm = gpc->kvm;
394 if (!kvm_gpc_is_valid_len(gpa, uhva, len))
397 guard(mutex)(&gpc->refresh_lock);
400 if (KVM_BUG_ON(gpc->valid, kvm))
403 spin_lock(&kvm->gpc_lock);
404 list_add(&gpc->list, &kvm->gpc_list);
405 spin_unlock(&kvm->gpc_lock);
408 * Activate the cache after adding it to the list, a concurrent
409 * refresh must not establish a mapping until the cache is
410 * reachable by mmu_notifier events.
412 write_lock_irq(&gpc->lock);
414 write_unlock_irq(&gpc->lock);
416 return __kvm_gpc_refresh(gpc, gpa, uhva);
419 int kvm_gpc_activate(struct gfn_to_pfn_cache *gpc, gpa_t gpa, unsigned long len)
422 * Explicitly disallow INVALID_GPA so that the magic value can be used
423 * by KVM to differentiate between GPA-based and HVA-based caches.
425 if (WARN_ON_ONCE(kvm_is_error_gpa(gpa)))
428 return __kvm_gpc_activate(gpc, gpa, KVM_HVA_ERR_BAD, len);
431 int kvm_gpc_activate_hva(struct gfn_to_pfn_cache *gpc, unsigned long uhva, unsigned long len)
433 return __kvm_gpc_activate(gpc, INVALID_GPA, uhva, len);
436 void kvm_gpc_deactivate(struct gfn_to_pfn_cache *gpc)
438 struct kvm *kvm = gpc->kvm;
442 guard(mutex)(&gpc->refresh_lock);
446 * Deactivate the cache before removing it from the list, KVM
447 * must stall mmu_notifier events until all users go away, i.e.
448 * until gpc->lock is dropped and refresh is guaranteed to fail.
450 write_lock_irq(&gpc->lock);
455 * Leave the GPA => uHVA cache intact, it's protected by the
456 * memslot generation. The PFN lookup needs to be redone every
457 * time as mmu_notifier protection is lost when the cache is
458 * removed from the VM's gpc_list.
460 old_khva = gpc->khva - offset_in_page(gpc->khva);
464 gpc->pfn = KVM_PFN_ERR_FAULT;
465 write_unlock_irq(&gpc->lock);
467 spin_lock(&kvm->gpc_lock);
468 list_del(&gpc->list);
469 spin_unlock(&kvm->gpc_lock);
471 gpc_unmap(old_pfn, old_khva);