virt/kvm/pfncache.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Kernel-based Virtual Machine driver for Linux
   4  *
   5  * This module enables kernel and guest-mode vCPU access to guest physical
   6  * memory with suitable invalidation mechanisms.
   7  *
   8  * Copyright © 2021 Amazon.com, Inc. or its affiliates.
   9  *
  10  * Authors:
  11  *   David Woodhouse <dwmw2@infradead.org>
  12  */
  13
  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>
  19
  20 #include "kvm_mm.h"
  21
  22 /*
  23  * MMU notifier 'invalidate_range_start' hook.
  24  */
  25 void gfn_to_pfn_cache_invalidate_start(struct kvm *kvm, unsigned long start,
  26                                        unsigned long end, bool may_block)
  27 {
  28         DECLARE_BITMAP(vcpu_bitmap, KVM_MAX_VCPUS);
  29         struct gfn_to_pfn_cache *gpc;
  30         bool evict_vcpus = false;
  31
  32         spin_lock(&kvm->gpc_lock);
  33         list_for_each_entry(gpc, &kvm->gpc_list, list) {
  34                 write_lock_irq(&gpc->lock);
  35
  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) {
  39                         gpc->valid = false;
  40
  41                         /*
  42                          * If a guest vCPU could be using the physical address,
  43                          * it needs to be forced out of guest mode.
  44                          */
  45                         if (gpc->usage & KVM_GUEST_USES_PFN) {
  46                                 if (!evict_vcpus) {
  47                                         evict_vcpus = true;
  48                                         bitmap_zero(vcpu_bitmap, KVM_MAX_VCPUS);
  49                                 }
  50                                 __set_bit(gpc->vcpu->vcpu_idx, vcpu_bitmap);
  51                         }
  52                 }
  53                 write_unlock_irq(&gpc->lock);
  54         }
  55         spin_unlock(&kvm->gpc_lock);
  56
  57         if (evict_vcpus) {
  58                 /*
  59                  * KVM needs to ensure the vCPU is fully out of guest context
  60                  * before allowing the invalidation to continue.
  61                  */
  62                 unsigned int req = KVM_REQ_OUTSIDE_GUEST_MODE;
  63                 bool called;
  64
  65                 /*
  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.
  69                  */
  70                 if (!may_block)
  71                         req &= ~KVM_REQUEST_WAIT;
  72
  73                 called = kvm_make_vcpus_request_mask(kvm, req, vcpu_bitmap);
  74
  75                 WARN_ON_ONCE(called && !may_block);
  76         }
  77 }
  78
  79 bool kvm_gfn_to_pfn_cache_check(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,
  80                                 gpa_t gpa, unsigned long len)
  81 {
  82         struct kvm_memslots *slots = kvm_memslots(kvm);
  83
  84         if ((gpa & ~PAGE_MASK) + len > PAGE_SIZE)
  85                 return false;
  86
  87         if (gpc->gpa != gpa || gpc->generation != slots->generation ||
  88             kvm_is_error_hva(gpc->uhva))
  89                 return false;
  90
  91         if (!gpc->valid)
  92                 return false;
  93
  94         return true;
  95 }
  96 EXPORT_SYMBOL_GPL(kvm_gfn_to_pfn_cache_check);
  97
  98 static void gpc_unmap_khva(struct kvm *kvm, kvm_pfn_t pfn, void *khva)
  99 {
 100         /* Unmap the old pfn/page if it was mapped before. */
 101         if (!is_error_noslot_pfn(pfn) && khva) {
 102                 if (pfn_valid(pfn))
 103                         kunmap(pfn_to_page(pfn));
 104 #ifdef CONFIG_HAS_IOMEM
 105                 else
 106                         memunmap(khva);
 107 #endif
 108         }
 109 }
 110
 111 static inline bool mmu_notifier_retry_cache(struct kvm *kvm, unsigned long mmu_seq)
 112 {
 113         /*
 114          * mn_active_invalidate_count acts for all intents and purposes
 115          * like mmu_notifier_count here; but the latter cannot be used
 116          * here because the invalidation of caches in the mmu_notifier
 117          * event occurs _before_ mmu_notifier_count is elevated.
 118          *
 119          * Note, it does not matter that mn_active_invalidate_count
 120          * is not protected by gpc->lock.  It is guaranteed to
 121          * be elevated before the mmu_notifier acquires gpc->lock, and
 122          * isn't dropped until after mmu_notifier_seq is updated.
 123          */
 124         if (kvm->mn_active_invalidate_count)
 125                 return true;
 126
 127         /*
 128          * Ensure mn_active_invalidate_count is read before
 129          * mmu_notifier_seq.  This pairs with the smp_wmb() in
 130          * mmu_notifier_invalidate_range_end() to guarantee either the
 131          * old (non-zero) value of mn_active_invalidate_count or the
 132          * new (incremented) value of mmu_notifier_seq is observed.
 133          */
 134         smp_rmb();
 135         return kvm->mmu_notifier_seq != mmu_seq;
 136 }
 137
 138 static kvm_pfn_t hva_to_pfn_retry(struct kvm *kvm, struct gfn_to_pfn_cache *gpc)
 139 {
 140         /* Note, the new page offset may be different than the old! */
 141         void *old_khva = gpc->khva - offset_in_page(gpc->khva);
 142         kvm_pfn_t new_pfn = KVM_PFN_ERR_FAULT;
 143         void *new_khva = NULL;
 144         unsigned long mmu_seq;
 145
 146         lockdep_assert_held(&gpc->refresh_lock);
 147
 148         lockdep_assert_held_write(&gpc->lock);
 149
 150         /*
 151          * Invalidate the cache prior to dropping gpc->lock, the gpa=>uhva
 152          * assets have already been updated and so a concurrent check() from a
 153          * different task may not fail the gpa/uhva/generation checks.
 154          */
 155         gpc->valid = false;
 156
 157         do {
 158                 mmu_seq = kvm->mmu_notifier_seq;
 159                 smp_rmb();
 160
 161                 write_unlock_irq(&gpc->lock);
 162
 163                 /*
 164                  * If the previous iteration "failed" due to an mmu_notifier
 165                  * event, release the pfn and unmap the kernel virtual address
 166                  * from the previous attempt.  Unmapping might sleep, so this
 167                  * needs to be done after dropping the lock.  Opportunistically
 168                  * check for resched while the lock isn't held.
 169                  */
 170                 if (new_pfn != KVM_PFN_ERR_FAULT) {
 171                         /*
 172                          * Keep the mapping if the previous iteration reused
 173                          * the existing mapping and didn't create a new one.
 174                          */
 175                         if (new_khva != old_khva)
 176                                 gpc_unmap_khva(kvm, new_pfn, new_khva);
 177
 178                         kvm_release_pfn_clean(new_pfn);
 179
 180                         cond_resched();
 181                 }
 182
 183                 /* We always request a writeable mapping */
 184                 new_pfn = hva_to_pfn(gpc->uhva, false, NULL, true, NULL);
 185                 if (is_error_noslot_pfn(new_pfn))
 186                         goto out_error;
 187
 188                 /*
 189                  * Obtain a new kernel mapping if KVM itself will access the
 190                  * pfn.  Note, kmap() and memremap() can both sleep, so this
 191                  * too must be done outside of gpc->lock!
 192                  */
 193                 if (gpc->usage & KVM_HOST_USES_PFN) {
 194                         if (new_pfn == gpc->pfn) {
 195                                 new_khva = old_khva;
 196                         } else if (pfn_valid(new_pfn)) {
 197                                 new_khva = kmap(pfn_to_page(new_pfn));
 198 #ifdef CONFIG_HAS_IOMEM
 199                         } else {
 200                                 new_khva = memremap(pfn_to_hpa(new_pfn), PAGE_SIZE, MEMREMAP_WB);
 201 #endif
 202                         }
 203                         if (!new_khva) {
 204                                 kvm_release_pfn_clean(new_pfn);
 205                                 goto out_error;
 206                         }
 207                 }
 208
 209                 write_lock_irq(&gpc->lock);
 210
 211                 /*
 212                  * Other tasks must wait for _this_ refresh to complete before
 213                  * attempting to refresh.
 214                  */
 215                 WARN_ON_ONCE(gpc->valid);
 216         } while (mmu_notifier_retry_cache(kvm, mmu_seq));
 217
 218         gpc->valid = true;
 219         gpc->pfn = new_pfn;
 220         gpc->khva = new_khva + (gpc->gpa & ~PAGE_MASK);
 221
 222         /*
 223          * Put the reference to the _new_ pfn.  The pfn is now tracked by the
 224          * cache and can be safely migrated, swapped, etc... as the cache will
 225          * invalidate any mappings in response to relevant mmu_notifier events.
 226          */
 227         kvm_release_pfn_clean(new_pfn);
 228
 229         return 0;
 230
 231 out_error:
 232         write_lock_irq(&gpc->lock);
 233
 234         return -EFAULT;
 235 }
 236
 237 int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,
 238                                  gpa_t gpa, unsigned long len)
 239 {
 240         struct kvm_memslots *slots = kvm_memslots(kvm);
 241         unsigned long page_offset = gpa & ~PAGE_MASK;
 242         kvm_pfn_t old_pfn, new_pfn;
 243         unsigned long old_uhva;
 244         void *old_khva;
 245         int ret = 0;
 246
 247         /*
 248          * If must fit within a single page. The 'len' argument is
 249          * only to enforce that.
 250          */
 251         if (page_offset + len > PAGE_SIZE)
 252                 return -EINVAL;
 253
 254         /*
 255          * If another task is refreshing the cache, wait for it to complete.
 256          * There is no guarantee that concurrent refreshes will see the same
 257          * gpa, memslots generation, etc..., so they must be fully serialized.
 258          */
 259         mutex_lock(&gpc->refresh_lock);
 260
 261         write_lock_irq(&gpc->lock);
 262
 263         old_pfn = gpc->pfn;
 264         old_khva = gpc->khva - offset_in_page(gpc->khva);
 265         old_uhva = gpc->uhva;
 266
 267         /* If the userspace HVA is invalid, refresh that first */
 268         if (gpc->gpa != gpa || gpc->generation != slots->generation ||
 269             kvm_is_error_hva(gpc->uhva)) {
 270                 gfn_t gfn = gpa_to_gfn(gpa);
 271
 272                 gpc->gpa = gpa;
 273                 gpc->generation = slots->generation;
 274                 gpc->memslot = __gfn_to_memslot(slots, gfn);
 275                 gpc->uhva = gfn_to_hva_memslot(gpc->memslot, gfn);
 276
 277                 if (kvm_is_error_hva(gpc->uhva)) {
 278                         ret = -EFAULT;
 279                         goto out;
 280                 }
 281         }
 282
 283         /*
 284          * If the userspace HVA changed or the PFN was already invalid,
 285          * drop the lock and do the HVA to PFN lookup again.
 286          */
 287         if (!gpc->valid || old_uhva != gpc->uhva) {
 288                 ret = hva_to_pfn_retry(kvm, gpc);
 289         } else {
 290                 /* If the HVA→PFN mapping was already valid, don't unmap it. */
 291                 old_pfn = KVM_PFN_ERR_FAULT;
 292                 old_khva = NULL;
 293         }
 294
 295  out:
 296         /*
 297          * Invalidate the cache and purge the pfn/khva if the refresh failed.
 298          * Some/all of the uhva, gpa, and memslot generation info may still be
 299          * valid, leave it as is.
 300          */
 301         if (ret) {
 302                 gpc->valid = false;
 303                 gpc->pfn = KVM_PFN_ERR_FAULT;
 304                 gpc->khva = NULL;
 305         }
 306
 307         /* Snapshot the new pfn before dropping the lock! */
 308         new_pfn = gpc->pfn;
 309
 310         write_unlock_irq(&gpc->lock);
 311
 312         mutex_unlock(&gpc->refresh_lock);
 313
 314         if (old_pfn != new_pfn)
 315                 gpc_unmap_khva(kvm, old_pfn, old_khva);
 316
 317         return ret;
 318 }
 319 EXPORT_SYMBOL_GPL(kvm_gfn_to_pfn_cache_refresh);
 320
 321 void kvm_gfn_to_pfn_cache_unmap(struct kvm *kvm, struct gfn_to_pfn_cache *gpc)
 322 {
 323         void *old_khva;
 324         kvm_pfn_t old_pfn;
 325
 326         mutex_lock(&gpc->refresh_lock);
 327         write_lock_irq(&gpc->lock);
 328
 329         gpc->valid = false;
 330
 331         old_khva = gpc->khva - offset_in_page(gpc->khva);
 332         old_pfn = gpc->pfn;
 333
 334         /*
 335          * We can leave the GPA → uHVA map cache intact but the PFN
 336          * lookup will need to be redone even for the same page.
 337          */
 338         gpc->khva = NULL;
 339         gpc->pfn = KVM_PFN_ERR_FAULT;
 340
 341         write_unlock_irq(&gpc->lock);
 342         mutex_unlock(&gpc->refresh_lock);
 343
 344         gpc_unmap_khva(kvm, old_pfn, old_khva);
 345 }
 346 EXPORT_SYMBOL_GPL(kvm_gfn_to_pfn_cache_unmap);
 347
 348
 349 int kvm_gfn_to_pfn_cache_init(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,
 350                               struct kvm_vcpu *vcpu, enum pfn_cache_usage usage,
 351                               gpa_t gpa, unsigned long len)
 352 {
 353         WARN_ON_ONCE(!usage || (usage & KVM_GUEST_AND_HOST_USE_PFN) != usage);
 354
 355         if (!gpc->active) {
 356                 rwlock_init(&gpc->lock);
 357                 mutex_init(&gpc->refresh_lock);
 358
 359                 gpc->khva = NULL;
 360                 gpc->pfn = KVM_PFN_ERR_FAULT;
 361                 gpc->uhva = KVM_HVA_ERR_BAD;
 362                 gpc->vcpu = vcpu;
 363                 gpc->usage = usage;
 364                 gpc->valid = false;
 365                 gpc->active = true;
 366
 367                 spin_lock(&kvm->gpc_lock);
 368                 list_add(&gpc->list, &kvm->gpc_list);
 369                 spin_unlock(&kvm->gpc_lock);
 370         }
 371         return kvm_gfn_to_pfn_cache_refresh(kvm, gpc, gpa, len);
 372 }
 373 EXPORT_SYMBOL_GPL(kvm_gfn_to_pfn_cache_init);
 374
 375 void kvm_gfn_to_pfn_cache_destroy(struct kvm *kvm, struct gfn_to_pfn_cache *gpc)
 376 {
 377         if (gpc->active) {
 378                 spin_lock(&kvm->gpc_lock);
 379                 list_del(&gpc->list);
 380                 spin_unlock(&kvm->gpc_lock);
 381
 382                 kvm_gfn_to_pfn_cache_unmap(kvm, gpc);
 383                 gpc->active = false;
 384         }
 385 }
 386 EXPORT_SYMBOL_GPL(kvm_gfn_to_pfn_cache_destroy);