mm/kfence/core.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * KFENCE guarded object allocator and fault handling.
   4  *
   5  * Copyright (C) 2020, Google LLC.
   6  */
   7
   8 #define pr_fmt(fmt) "kfence: " fmt
   9
  10 #include <linux/atomic.h>
  11 #include <linux/bug.h>
  12 #include <linux/debugfs.h>
  13 #include <linux/irq_work.h>
  14 #include <linux/kcsan-checks.h>
  15 #include <linux/kfence.h>
  16 #include <linux/kmemleak.h>
  17 #include <linux/list.h>
  18 #include <linux/lockdep.h>
  19 #include <linux/memblock.h>
  20 #include <linux/moduleparam.h>
  21 #include <linux/random.h>
  22 #include <linux/rcupdate.h>
  23 #include <linux/sched/clock.h>
  24 #include <linux/sched/sysctl.h>
  25 #include <linux/seq_file.h>
  26 #include <linux/slab.h>
  27 #include <linux/spinlock.h>
  28 #include <linux/string.h>
  29
  30 #include <asm/kfence.h>
  31
  32 #include "kfence.h"
  33
  34 /* Disables KFENCE on the first warning assuming an irrecoverable error. */
  35 #define KFENCE_WARN_ON(cond)                                                   \
  36         ({                                                                     \
  37                 const bool __cond = WARN_ON(cond);                             \
  38                 if (unlikely(__cond))                                          \
  39                         WRITE_ONCE(kfence_enabled, false);                     \
  40                 __cond;                                                        \
  41         })
  42
  43 /* === Data ================================================================= */
  44
  45 static bool kfence_enabled __read_mostly;
  46
  47 static unsigned long kfence_sample_interval __read_mostly = CONFIG_KFENCE_SAMPLE_INTERVAL;
  48
  49 #ifdef MODULE_PARAM_PREFIX
  50 #undef MODULE_PARAM_PREFIX
  51 #endif
  52 #define MODULE_PARAM_PREFIX "kfence."
  53
  54 static int param_set_sample_interval(const char *val, const struct kernel_param *kp)
  55 {
  56         unsigned long num;
  57         int ret = kstrtoul(val, 0, &num);
  58
  59         if (ret < 0)
  60                 return ret;
  61
  62         if (!num) /* Using 0 to indicate KFENCE is disabled. */
  63                 WRITE_ONCE(kfence_enabled, false);
  64         else if (!READ_ONCE(kfence_enabled) && system_state != SYSTEM_BOOTING)
  65                 return -EINVAL; /* Cannot (re-)enable KFENCE on-the-fly. */
  66
  67         *((unsigned long *)kp->arg) = num;
  68         return 0;
  69 }
  70
  71 static int param_get_sample_interval(char *buffer, const struct kernel_param *kp)
  72 {
  73         if (!READ_ONCE(kfence_enabled))
  74                 return sprintf(buffer, "0\n");
  75
  76         return param_get_ulong(buffer, kp);
  77 }
  78
  79 static const struct kernel_param_ops sample_interval_param_ops = {
  80         .set = param_set_sample_interval,
  81         .get = param_get_sample_interval,
  82 };
  83 module_param_cb(sample_interval, &sample_interval_param_ops, &kfence_sample_interval, 0600);
  84
  85 /* The pool of pages used for guard pages and objects. */
  86 char *__kfence_pool __ro_after_init;
  87 EXPORT_SYMBOL(__kfence_pool); /* Export for test modules. */
  88
  89 /*
  90  * Per-object metadata, with one-to-one mapping of object metadata to
  91  * backing pages (in __kfence_pool).
  92  */
  93 static_assert(CONFIG_KFENCE_NUM_OBJECTS > 0);
  94 struct kfence_metadata kfence_metadata[CONFIG_KFENCE_NUM_OBJECTS];
  95
  96 /* Freelist with available objects. */
  97 static struct list_head kfence_freelist = LIST_HEAD_INIT(kfence_freelist);
  98 static DEFINE_RAW_SPINLOCK(kfence_freelist_lock); /* Lock protecting freelist. */
  99
 100 #ifdef CONFIG_KFENCE_STATIC_KEYS
 101 /* The static key to set up a KFENCE allocation. */
 102 DEFINE_STATIC_KEY_FALSE(kfence_allocation_key);
 103 #endif
 104
 105 /* Gates the allocation, ensuring only one succeeds in a given period. */
 106 atomic_t kfence_allocation_gate = ATOMIC_INIT(1);
 107
 108 /* Statistics counters for debugfs. */
 109 enum kfence_counter_id {
 110         KFENCE_COUNTER_ALLOCATED,
 111         KFENCE_COUNTER_ALLOCS,
 112         KFENCE_COUNTER_FREES,
 113         KFENCE_COUNTER_ZOMBIES,
 114         KFENCE_COUNTER_BUGS,
 115         KFENCE_COUNTER_COUNT,
 116 };
 117 static atomic_long_t counters[KFENCE_COUNTER_COUNT];
 118 static const char *const counter_names[] = {
 119         [KFENCE_COUNTER_ALLOCATED]      = "currently allocated",
 120         [KFENCE_COUNTER_ALLOCS]         = "total allocations",
 121         [KFENCE_COUNTER_FREES]          = "total frees",
 122         [KFENCE_COUNTER_ZOMBIES]        = "zombie allocations",
 123         [KFENCE_COUNTER_BUGS]           = "total bugs",
 124 };
 125 static_assert(ARRAY_SIZE(counter_names) == KFENCE_COUNTER_COUNT);
 126
 127 /* === Internals ============================================================ */
 128
 129 static bool kfence_protect(unsigned long addr)
 130 {
 131         return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), true));
 132 }
 133
 134 static bool kfence_unprotect(unsigned long addr)
 135 {
 136         return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), false));
 137 }
 138
 139 static inline struct kfence_metadata *addr_to_metadata(unsigned long addr)
 140 {
 141         long index;
 142
 143         /* The checks do not affect performance; only called from slow-paths. */
 144
 145         if (!is_kfence_address((void *)addr))
 146                 return NULL;
 147
 148         /*
 149          * May be an invalid index if called with an address at the edge of
 150          * __kfence_pool, in which case we would report an "invalid access"
 151          * error.
 152          */
 153         index = (addr - (unsigned long)__kfence_pool) / (PAGE_SIZE * 2) - 1;
 154         if (index < 0 || index >= CONFIG_KFENCE_NUM_OBJECTS)
 155                 return NULL;
 156
 157         return &kfence_metadata[index];
 158 }
 159
 160 static inline unsigned long metadata_to_pageaddr(const struct kfence_metadata *meta)
 161 {
 162         unsigned long offset = (meta - kfence_metadata + 1) * PAGE_SIZE * 2;
 163         unsigned long pageaddr = (unsigned long)&__kfence_pool[offset];
 164
 165         /* The checks do not affect performance; only called from slow-paths. */
 166
 167         /* Only call with a pointer into kfence_metadata. */
 168         if (KFENCE_WARN_ON(meta < kfence_metadata ||
 169                            meta >= kfence_metadata + CONFIG_KFENCE_NUM_OBJECTS))
 170                 return 0;
 171
 172         /*
 173          * This metadata object only ever maps to 1 page; verify that the stored
 174          * address is in the expected range.
 175          */
 176         if (KFENCE_WARN_ON(ALIGN_DOWN(meta->addr, PAGE_SIZE) != pageaddr))
 177                 return 0;
 178
 179         return pageaddr;
 180 }
 181
 182 /*
 183  * Update the object's metadata state, including updating the alloc/free stacks
 184  * depending on the state transition.
 185  */
 186 static noinline void metadata_update_state(struct kfence_metadata *meta,
 187                                            enum kfence_object_state next)
 188 {
 189         struct kfence_track *track =
 190                 next == KFENCE_OBJECT_FREED ? &meta->free_track : &meta->alloc_track;
 191
 192         lockdep_assert_held(&meta->lock);
 193
 194         /*
 195          * Skip over 1 (this) functions; noinline ensures we do not accidentally
 196          * skip over the caller by never inlining.
 197          */
 198         track->num_stack_entries = stack_trace_save(track->stack_entries, KFENCE_STACK_DEPTH, 1);
 199         track->pid = task_pid_nr(current);
 200         track->cpu = raw_smp_processor_id();
 201         track->ts_nsec = local_clock(); /* Same source as printk timestamps. */
 202
 203         /*
 204          * Pairs with READ_ONCE() in
 205          *      kfence_shutdown_cache(),
 206          *      kfence_handle_page_fault().
 207          */
 208         WRITE_ONCE(meta->state, next);
 209 }
 210
 211 /* Write canary byte to @addr. */
 212 static inline bool set_canary_byte(u8 *addr)
 213 {
 214         *addr = KFENCE_CANARY_PATTERN(addr);
 215         return true;
 216 }
 217
 218 /* Check canary byte at @addr. */
 219 static inline bool check_canary_byte(u8 *addr)
 220 {
 221         if (likely(*addr == KFENCE_CANARY_PATTERN(addr)))
 222                 return true;
 223
 224         atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
 225         kfence_report_error((unsigned long)addr, false, NULL, addr_to_metadata((unsigned long)addr),
 226                             KFENCE_ERROR_CORRUPTION);
 227         return false;
 228 }
 229
 230 /* __always_inline this to ensure we won't do an indirect call to fn. */
 231 static __always_inline void for_each_canary(const struct kfence_metadata *meta, bool (*fn)(u8 *))
 232 {
 233         const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE);
 234         unsigned long addr;
 235
 236         lockdep_assert_held(&meta->lock);
 237
 238         /*
 239          * We'll iterate over each canary byte per-side until fn() returns
 240          * false. However, we'll still iterate over the canary bytes to the
 241          * right of the object even if there was an error in the canary bytes to
 242          * the left of the object. Specifically, if check_canary_byte()
 243          * generates an error, showing both sides might give more clues as to
 244          * what the error is about when displaying which bytes were corrupted.
 245          */
 246
 247         /* Apply to left of object. */
 248         for (addr = pageaddr; addr < meta->addr; addr++) {
 249                 if (!fn((u8 *)addr))
 250                         break;
 251         }
 252
 253         /* Apply to right of object. */
 254         for (addr = meta->addr + meta->size; addr < pageaddr + PAGE_SIZE; addr++) {
 255                 if (!fn((u8 *)addr))
 256                         break;
 257         }
 258 }
 259
 260 static void *kfence_guarded_alloc(struct kmem_cache *cache, size_t size, gfp_t gfp)
 261 {
 262         struct kfence_metadata *meta = NULL;
 263         unsigned long flags;
 264         struct page *page;
 265         void *addr;
 266
 267         /* Try to obtain a free object. */
 268         raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
 269         if (!list_empty(&kfence_freelist)) {
 270                 meta = list_entry(kfence_freelist.next, struct kfence_metadata, list);
 271                 list_del_init(&meta->list);
 272         }
 273         raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
 274         if (!meta)
 275                 return NULL;
 276
 277         if (unlikely(!raw_spin_trylock_irqsave(&meta->lock, flags))) {
 278                 /*
 279                  * This is extremely unlikely -- we are reporting on a
 280                  * use-after-free, which locked meta->lock, and the reporting
 281                  * code via printk calls kmalloc() which ends up in
 282                  * kfence_alloc() and tries to grab the same object that we're
 283                  * reporting on. While it has never been observed, lockdep does
 284                  * report that there is a possibility of deadlock. Fix it by
 285                  * using trylock and bailing out gracefully.
 286                  */
 287                 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
 288                 /* Put the object back on the freelist. */
 289                 list_add_tail(&meta->list, &kfence_freelist);
 290                 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
 291
 292                 return NULL;
 293         }
 294
 295         meta->addr = metadata_to_pageaddr(meta);
 296         /* Unprotect if we're reusing this page. */
 297         if (meta->state == KFENCE_OBJECT_FREED)
 298                 kfence_unprotect(meta->addr);
 299
 300         /*
 301          * Note: for allocations made before RNG initialization, will always
 302          * return zero. We still benefit from enabling KFENCE as early as
 303          * possible, even when the RNG is not yet available, as this will allow
 304          * KFENCE to detect bugs due to earlier allocations. The only downside
 305          * is that the out-of-bounds accesses detected are deterministic for
 306          * such allocations.
 307          */
 308         if (prandom_u32_max(2)) {
 309                 /* Allocate on the "right" side, re-calculate address. */
 310                 meta->addr += PAGE_SIZE - size;
 311                 meta->addr = ALIGN_DOWN(meta->addr, cache->align);
 312         }
 313
 314         addr = (void *)meta->addr;
 315
 316         /* Update remaining metadata. */
 317         metadata_update_state(meta, KFENCE_OBJECT_ALLOCATED);
 318         /* Pairs with READ_ONCE() in kfence_shutdown_cache(). */
 319         WRITE_ONCE(meta->cache, cache);
 320         meta->size = size;
 321         for_each_canary(meta, set_canary_byte);
 322
 323         /* Set required struct page fields. */
 324         page = virt_to_page(meta->addr);
 325         page->slab_cache = cache;
 326         if (IS_ENABLED(CONFIG_SLUB))
 327                 page->objects = 1;
 328         if (IS_ENABLED(CONFIG_SLAB))
 329                 page->s_mem = addr;
 330
 331         raw_spin_unlock_irqrestore(&meta->lock, flags);
 332
 333         /* Memory initialization. */
 334
 335         /*
 336          * We check slab_want_init_on_alloc() ourselves, rather than letting
 337          * SL*B do the initialization, as otherwise we might overwrite KFENCE's
 338          * redzone.
 339          */
 340         if (unlikely(slab_want_init_on_alloc(gfp, cache)))
 341                 memzero_explicit(addr, size);
 342         if (cache->ctor)
 343                 cache->ctor(addr);
 344
 345         if (CONFIG_KFENCE_STRESS_TEST_FAULTS && !prandom_u32_max(CONFIG_KFENCE_STRESS_TEST_FAULTS))
 346                 kfence_protect(meta->addr); /* Random "faults" by protecting the object. */
 347
 348         atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCATED]);
 349         atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCS]);
 350
 351         return addr;
 352 }
 353
 354 static void kfence_guarded_free(void *addr, struct kfence_metadata *meta, bool zombie)
 355 {
 356         struct kcsan_scoped_access assert_page_exclusive;
 357         unsigned long flags;
 358
 359         raw_spin_lock_irqsave(&meta->lock, flags);
 360
 361         if (meta->state != KFENCE_OBJECT_ALLOCATED || meta->addr != (unsigned long)addr) {
 362                 /* Invalid or double-free, bail out. */
 363                 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
 364                 kfence_report_error((unsigned long)addr, false, NULL, meta,
 365                                     KFENCE_ERROR_INVALID_FREE);
 366                 raw_spin_unlock_irqrestore(&meta->lock, flags);
 367                 return;
 368         }
 369
 370         /* Detect racy use-after-free, or incorrect reallocation of this page by KFENCE. */
 371         kcsan_begin_scoped_access((void *)ALIGN_DOWN((unsigned long)addr, PAGE_SIZE), PAGE_SIZE,
 372                                   KCSAN_ACCESS_SCOPED | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT,
 373                                   &assert_page_exclusive);
 374
 375         if (CONFIG_KFENCE_STRESS_TEST_FAULTS)
 376                 kfence_unprotect((unsigned long)addr); /* To check canary bytes. */
 377
 378         /* Restore page protection if there was an OOB access. */
 379         if (meta->unprotected_page) {
 380                 memzero_explicit((void *)ALIGN_DOWN(meta->unprotected_page, PAGE_SIZE), PAGE_SIZE);
 381                 kfence_protect(meta->unprotected_page);
 382                 meta->unprotected_page = 0;
 383         }
 384
 385         /* Check canary bytes for memory corruption. */
 386         for_each_canary(meta, check_canary_byte);
 387
 388         /*
 389          * Clear memory if init-on-free is set. While we protect the page, the
 390          * data is still there, and after a use-after-free is detected, we
 391          * unprotect the page, so the data is still accessible.
 392          */
 393         if (!zombie && unlikely(slab_want_init_on_free(meta->cache)))
 394                 memzero_explicit(addr, meta->size);
 395
 396         /* Mark the object as freed. */
 397         metadata_update_state(meta, KFENCE_OBJECT_FREED);
 398
 399         raw_spin_unlock_irqrestore(&meta->lock, flags);
 400
 401         /* Protect to detect use-after-frees. */
 402         kfence_protect((unsigned long)addr);
 403
 404         kcsan_end_scoped_access(&assert_page_exclusive);
 405         if (!zombie) {
 406                 /* Add it to the tail of the freelist for reuse. */
 407                 raw_spin_lock_irqsave(&kfence_freelist_lock, flags);
 408                 KFENCE_WARN_ON(!list_empty(&meta->list));
 409                 list_add_tail(&meta->list, &kfence_freelist);
 410                 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags);
 411
 412                 atomic_long_dec(&counters[KFENCE_COUNTER_ALLOCATED]);
 413                 atomic_long_inc(&counters[KFENCE_COUNTER_FREES]);
 414         } else {
 415                 /* See kfence_shutdown_cache(). */
 416                 atomic_long_inc(&counters[KFENCE_COUNTER_ZOMBIES]);
 417         }
 418 }
 419
 420 static void rcu_guarded_free(struct rcu_head *h)
 421 {
 422         struct kfence_metadata *meta = container_of(h, struct kfence_metadata, rcu_head);
 423
 424         kfence_guarded_free((void *)meta->addr, meta, false);
 425 }
 426
 427 static bool __init kfence_init_pool(void)
 428 {
 429         unsigned long addr = (unsigned long)__kfence_pool;
 430         struct page *pages;
 431         int i;
 432
 433         if (!__kfence_pool)
 434                 return false;
 435
 436         if (!arch_kfence_init_pool())
 437                 goto err;
 438
 439         pages = virt_to_page(addr);
 440
 441         /*
 442          * Set up object pages: they must have PG_slab set, to avoid freeing
 443          * these as real pages.
 444          *
 445          * We also want to avoid inserting kfence_free() in the kfree()
 446          * fast-path in SLUB, and therefore need to ensure kfree() correctly
 447          * enters __slab_free() slow-path.
 448          */
 449         for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) {
 450                 if (!i || (i % 2))
 451                         continue;
 452
 453                 /* Verify we do not have a compound head page. */
 454                 if (WARN_ON(compound_head(&pages[i]) != &pages[i]))
 455                         goto err;
 456
 457                 __SetPageSlab(&pages[i]);
 458         }
 459
 460         /*
 461          * Protect the first 2 pages. The first page is mostly unnecessary, and
 462          * merely serves as an extended guard page. However, adding one
 463          * additional page in the beginning gives us an even number of pages,
 464          * which simplifies the mapping of address to metadata index.
 465          */
 466         for (i = 0; i < 2; i++) {
 467                 if (unlikely(!kfence_protect(addr)))
 468                         goto err;
 469
 470                 addr += PAGE_SIZE;
 471         }
 472
 473         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 474                 struct kfence_metadata *meta = &kfence_metadata[i];
 475
 476                 /* Initialize metadata. */
 477                 INIT_LIST_HEAD(&meta->list);
 478                 raw_spin_lock_init(&meta->lock);
 479                 meta->state = KFENCE_OBJECT_UNUSED;
 480                 meta->addr = addr; /* Initialize for validation in metadata_to_pageaddr(). */
 481                 list_add_tail(&meta->list, &kfence_freelist);
 482
 483                 /* Protect the right redzone. */
 484                 if (unlikely(!kfence_protect(addr + PAGE_SIZE)))
 485                         goto err;
 486
 487                 addr += 2 * PAGE_SIZE;
 488         }
 489
 490         /*
 491          * The pool is live and will never be deallocated from this point on.
 492          * Remove the pool object from the kmemleak object tree, as it would
 493          * otherwise overlap with allocations returned by kfence_alloc(), which
 494          * are registered with kmemleak through the slab post-alloc hook.
 495          */
 496         kmemleak_free(__kfence_pool);
 497
 498         return true;
 499
 500 err:
 501         /*
 502          * Only release unprotected pages, and do not try to go back and change
 503          * page attributes due to risk of failing to do so as well. If changing
 504          * page attributes for some pages fails, it is very likely that it also
 505          * fails for the first page, and therefore expect addr==__kfence_pool in
 506          * most failure cases.
 507          */
 508         memblock_free_late(__pa(addr), KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool));
 509         __kfence_pool = NULL;
 510         return false;
 511 }
 512
 513 /* === DebugFS Interface ==================================================== */
 514
 515 static int stats_show(struct seq_file *seq, void *v)
 516 {
 517         int i;
 518
 519         seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled));
 520         for (i = 0; i < KFENCE_COUNTER_COUNT; i++)
 521                 seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i]));
 522
 523         return 0;
 524 }
 525 DEFINE_SHOW_ATTRIBUTE(stats);
 526
 527 /*
 528  * debugfs seq_file operations for /sys/kernel/debug/kfence/objects.
 529  * start_object() and next_object() return the object index + 1, because NULL is used
 530  * to stop iteration.
 531  */
 532 static void *start_object(struct seq_file *seq, loff_t *pos)
 533 {
 534         if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
 535                 return (void *)((long)*pos + 1);
 536         return NULL;
 537 }
 538
 539 static void stop_object(struct seq_file *seq, void *v)
 540 {
 541 }
 542
 543 static void *next_object(struct seq_file *seq, void *v, loff_t *pos)
 544 {
 545         ++*pos;
 546         if (*pos < CONFIG_KFENCE_NUM_OBJECTS)
 547                 return (void *)((long)*pos + 1);
 548         return NULL;
 549 }
 550
 551 static int show_object(struct seq_file *seq, void *v)
 552 {
 553         struct kfence_metadata *meta = &kfence_metadata[(long)v - 1];
 554         unsigned long flags;
 555
 556         raw_spin_lock_irqsave(&meta->lock, flags);
 557         kfence_print_object(seq, meta);
 558         raw_spin_unlock_irqrestore(&meta->lock, flags);
 559         seq_puts(seq, "---------------------------------\n");
 560
 561         return 0;
 562 }
 563
 564 static const struct seq_operations object_seqops = {
 565         .start = start_object,
 566         .next = next_object,
 567         .stop = stop_object,
 568         .show = show_object,
 569 };
 570
 571 static int open_objects(struct inode *inode, struct file *file)
 572 {
 573         return seq_open(file, &object_seqops);
 574 }
 575
 576 static const struct file_operations objects_fops = {
 577         .open = open_objects,
 578         .read = seq_read,
 579         .llseek = seq_lseek,
 580 };
 581
 582 static int __init kfence_debugfs_init(void)
 583 {
 584         struct dentry *kfence_dir = debugfs_create_dir("kfence", NULL);
 585
 586         debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops);
 587         debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops);
 588         return 0;
 589 }
 590
 591 late_initcall(kfence_debugfs_init);
 592
 593 /* === Allocation Gate Timer ================================================ */
 594
 595 #ifdef CONFIG_KFENCE_STATIC_KEYS
 596 /* Wait queue to wake up allocation-gate timer task. */
 597 static DECLARE_WAIT_QUEUE_HEAD(allocation_wait);
 598
 599 static void wake_up_kfence_timer(struct irq_work *work)
 600 {
 601         wake_up(&allocation_wait);
 602 }
 603 static DEFINE_IRQ_WORK(wake_up_kfence_timer_work, wake_up_kfence_timer);
 604 #endif
 605
 606 /*
 607  * Set up delayed work, which will enable and disable the static key. We need to
 608  * use a work queue (rather than a simple timer), since enabling and disabling a
 609  * static key cannot be done from an interrupt.
 610  *
 611  * Note: Toggling a static branch currently causes IPIs, and here we'll end up
 612  * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with
 613  * more aggressive sampling intervals), we could get away with a variant that
 614  * avoids IPIs, at the cost of not immediately capturing allocations if the
 615  * instructions remain cached.
 616  */
 617 static struct delayed_work kfence_timer;
 618 static void toggle_allocation_gate(struct work_struct *work)
 619 {
 620         if (!READ_ONCE(kfence_enabled))
 621                 return;
 622
 623         atomic_set(&kfence_allocation_gate, 0);
 624 #ifdef CONFIG_KFENCE_STATIC_KEYS
 625         /* Enable static key, and await allocation to happen. */
 626         static_branch_enable(&kfence_allocation_key);
 627
 628         if (sysctl_hung_task_timeout_secs) {
 629                 /*
 630                  * During low activity with no allocations we might wait a
 631                  * while; let's avoid the hung task warning.
 632                  */
 633                 wait_event_idle_timeout(allocation_wait, atomic_read(&kfence_allocation_gate),
 634                                         sysctl_hung_task_timeout_secs * HZ / 2);
 635         } else {
 636                 wait_event_idle(allocation_wait, atomic_read(&kfence_allocation_gate));
 637         }
 638
 639         /* Disable static key and reset timer. */
 640         static_branch_disable(&kfence_allocation_key);
 641 #endif
 642         queue_delayed_work(system_unbound_wq, &kfence_timer,
 643                            msecs_to_jiffies(kfence_sample_interval));
 644 }
 645 static DECLARE_DELAYED_WORK(kfence_timer, toggle_allocation_gate);
 646
 647 /* === Public interface ===================================================== */
 648
 649 void __init kfence_alloc_pool(void)
 650 {
 651         if (!kfence_sample_interval)
 652                 return;
 653
 654         __kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
 655
 656         if (!__kfence_pool)
 657                 pr_err("failed to allocate pool\n");
 658 }
 659
 660 void __init kfence_init(void)
 661 {
 662         /* Setting kfence_sample_interval to 0 on boot disables KFENCE. */
 663         if (!kfence_sample_interval)
 664                 return;
 665
 666         if (!kfence_init_pool()) {
 667                 pr_err("%s failed\n", __func__);
 668                 return;
 669         }
 670
 671         WRITE_ONCE(kfence_enabled, true);
 672         queue_delayed_work(system_unbound_wq, &kfence_timer, 0);
 673         pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE,
 674                 CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool,
 675                 (void *)(__kfence_pool + KFENCE_POOL_SIZE));
 676 }
 677
 678 void kfence_shutdown_cache(struct kmem_cache *s)
 679 {
 680         unsigned long flags;
 681         struct kfence_metadata *meta;
 682         int i;
 683
 684         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 685                 bool in_use;
 686
 687                 meta = &kfence_metadata[i];
 688
 689                 /*
 690                  * If we observe some inconsistent cache and state pair where we
 691                  * should have returned false here, cache destruction is racing
 692                  * with either kmem_cache_alloc() or kmem_cache_free(). Taking
 693                  * the lock will not help, as different critical section
 694                  * serialization will have the same outcome.
 695                  */
 696                 if (READ_ONCE(meta->cache) != s ||
 697                     READ_ONCE(meta->state) != KFENCE_OBJECT_ALLOCATED)
 698                         continue;
 699
 700                 raw_spin_lock_irqsave(&meta->lock, flags);
 701                 in_use = meta->cache == s && meta->state == KFENCE_OBJECT_ALLOCATED;
 702                 raw_spin_unlock_irqrestore(&meta->lock, flags);
 703
 704                 if (in_use) {
 705                         /*
 706                          * This cache still has allocations, and we should not
 707                          * release them back into the freelist so they can still
 708                          * safely be used and retain the kernel's default
 709                          * behaviour of keeping the allocations alive (leak the
 710                          * cache); however, they effectively become "zombie
 711                          * allocations" as the KFENCE objects are the only ones
 712                          * still in use and the owning cache is being destroyed.
 713                          *
 714                          * We mark them freed, so that any subsequent use shows
 715                          * more useful error messages that will include stack
 716                          * traces of the user of the object, the original
 717                          * allocation, and caller to shutdown_cache().
 718                          */
 719                         kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true);
 720                 }
 721         }
 722
 723         for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) {
 724                 meta = &kfence_metadata[i];
 725
 726                 /* See above. */
 727                 if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED)
 728                         continue;
 729
 730                 raw_spin_lock_irqsave(&meta->lock, flags);
 731                 if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED)
 732                         meta->cache = NULL;
 733                 raw_spin_unlock_irqrestore(&meta->lock, flags);
 734         }
 735 }
 736
 737 void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags)
 738 {
 739         /*
 740          * Perform size check before switching kfence_allocation_gate, so that
 741          * we don't disable KFENCE without making an allocation.
 742          */
 743         if (size > PAGE_SIZE)
 744                 return NULL;
 745
 746         /*
 747          * Skip allocations from non-default zones, including DMA. We cannot
 748          * guarantee that pages in the KFENCE pool will have the requested
 749          * properties (e.g. reside in DMAable memory).
 750          */
 751         if ((flags & GFP_ZONEMASK) ||
 752             (s->flags & (SLAB_CACHE_DMA | SLAB_CACHE_DMA32)))
 753                 return NULL;
 754
 755         /*
 756          * allocation_gate only needs to become non-zero, so it doesn't make
 757          * sense to continue writing to it and pay the associated contention
 758          * cost, in case we have a large number of concurrent allocations.
 759          */
 760         if (atomic_read(&kfence_allocation_gate) || atomic_inc_return(&kfence_allocation_gate) > 1)
 761                 return NULL;
 762 #ifdef CONFIG_KFENCE_STATIC_KEYS
 763         /*
 764          * waitqueue_active() is fully ordered after the update of
 765          * kfence_allocation_gate per atomic_inc_return().
 766          */
 767         if (waitqueue_active(&allocation_wait)) {
 768                 /*
 769                  * Calling wake_up() here may deadlock when allocations happen
 770                  * from within timer code. Use an irq_work to defer it.
 771                  */
 772                 irq_work_queue(&wake_up_kfence_timer_work);
 773         }
 774 #endif
 775
 776         if (!READ_ONCE(kfence_enabled))
 777                 return NULL;
 778
 779         return kfence_guarded_alloc(s, size, flags);
 780 }
 781
 782 size_t kfence_ksize(const void *addr)
 783 {
 784         const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
 785
 786         /*
 787          * Read locklessly -- if there is a race with __kfence_alloc(), this is
 788          * either a use-after-free or invalid access.
 789          */
 790         return meta ? meta->size : 0;
 791 }
 792
 793 void *kfence_object_start(const void *addr)
 794 {
 795         const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
 796
 797         /*
 798          * Read locklessly -- if there is a race with __kfence_alloc(), this is
 799          * either a use-after-free or invalid access.
 800          */
 801         return meta ? (void *)meta->addr : NULL;
 802 }
 803
 804 void __kfence_free(void *addr)
 805 {
 806         struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr);
 807
 808         /*
 809          * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing
 810          * the object, as the object page may be recycled for other-typed
 811          * objects once it has been freed. meta->cache may be NULL if the cache
 812          * was destroyed.
 813          */
 814         if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU)))
 815                 call_rcu(&meta->rcu_head, rcu_guarded_free);
 816         else
 817                 kfence_guarded_free(addr, meta, false);
 818 }
 819
 820 bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs)
 821 {
 822         const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE;
 823         struct kfence_metadata *to_report = NULL;
 824         enum kfence_error_type error_type;
 825         unsigned long flags;
 826
 827         if (!is_kfence_address((void *)addr))
 828                 return false;
 829
 830         if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */
 831                 return kfence_unprotect(addr); /* ... unprotect and proceed. */
 832
 833         atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]);
 834
 835         if (page_index % 2) {
 836                 /* This is a redzone, report a buffer overflow. */
 837                 struct kfence_metadata *meta;
 838                 int distance = 0;
 839
 840                 meta = addr_to_metadata(addr - PAGE_SIZE);
 841                 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
 842                         to_report = meta;
 843                         /* Data race ok; distance calculation approximate. */
 844                         distance = addr - data_race(meta->addr + meta->size);
 845                 }
 846
 847                 meta = addr_to_metadata(addr + PAGE_SIZE);
 848                 if (meta && READ_ONCE(meta->state) == KFENCE_OBJECT_ALLOCATED) {
 849                         /* Data race ok; distance calculation approximate. */
 850                         if (!to_report || distance > data_race(meta->addr) - addr)
 851                                 to_report = meta;
 852                 }
 853
 854                 if (!to_report)
 855                         goto out;
 856
 857                 raw_spin_lock_irqsave(&to_report->lock, flags);
 858                 to_report->unprotected_page = addr;
 859                 error_type = KFENCE_ERROR_OOB;
 860
 861                 /*
 862                  * If the object was freed before we took the look we can still
 863                  * report this as an OOB -- the report will simply show the
 864                  * stacktrace of the free as well.
 865                  */
 866         } else {
 867                 to_report = addr_to_metadata(addr);
 868                 if (!to_report)
 869                         goto out;
 870
 871                 raw_spin_lock_irqsave(&to_report->lock, flags);
 872                 error_type = KFENCE_ERROR_UAF;
 873                 /*
 874                  * We may race with __kfence_alloc(), and it is possible that a
 875                  * freed object may be reallocated. We simply report this as a
 876                  * use-after-free, with the stack trace showing the place where
 877                  * the object was re-allocated.
 878                  */
 879         }
 880
 881 out:
 882         if (to_report) {
 883                 kfence_report_error(addr, is_write, regs, to_report, error_type);
 884                 raw_spin_unlock_irqrestore(&to_report->lock, flags);
 885         } else {
 886                 /* This may be a UAF or OOB access, but we can't be sure. */
 887                 kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID);
 888         }
 889
 890         return kfence_unprotect(addr); /* Unprotect and let access proceed. */
 891 }