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