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