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