src/rseq-mempool.c

   1 // SPDX-License-Identifier: MIT
   2 // SPDX-FileCopyrightText: 2024 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   3
   4 #include <rseq/mempool.h>
   5 #include <sys/mman.h>
   6 #include <assert.h>
   7 #include <string.h>
   8 #include <pthread.h>
   9 #include <unistd.h>
  10 #include <stdlib.h>
  11 #include <rseq/compiler.h>
  12 #include <errno.h>
  13 #include <stdint.h>
  14 #include <stdbool.h>
  15 #include <stdio.h>
  16
  17 #ifdef HAVE_LIBNUMA
  18 # include <numa.h>
  19 # include <numaif.h>
  20 #endif
  21
  22 #include "rseq-utils.h"
  23 #include <rseq/rseq.h>
  24
  25 /*
  26  * rseq-mempool.c: rseq CPU-Local Storage (CLS) memory allocator.
  27  *
  28  * The rseq per-CPU memory allocator allows the application the request
  29  * memory pools of CPU-Local memory each of containing objects of a
  30  * given size (rounded to next power of 2), reserving a given virtual
  31  * address size per CPU, for a given maximum number of CPUs.
  32  *
  33  * The per-CPU memory allocator is analogous to TLS (Thread-Local
  34  * Storage) memory: TLS is Thread-Local Storage, whereas the per-CPU
  35  * memory allocator provides CPU-Local Storage.
  36  */
  37
  38 #define POOL_SET_NR_ENTRIES     RSEQ_BITS_PER_LONG
  39
  40 /*
  41  * Smallest allocation should hold enough space for a free list pointer.
  42  */
  43 #if RSEQ_BITS_PER_LONG == 64
  44 # define POOL_SET_MIN_ENTRY     3       /* Smallest item_len=8 */
  45 #else
  46 # define POOL_SET_MIN_ENTRY     2       /* Smallest item_len=4 */
  47 #endif
  48
  49 #define BIT_PER_ULONG           (8 * sizeof(unsigned long))
  50
  51 #define MOVE_PAGES_BATCH_SIZE   4096
  52
  53 #define RANGE_HEADER_OFFSET     sizeof(struct rseq_mempool_range)
  54
  55 struct free_list_node;
  56
  57 struct free_list_node {
  58         struct free_list_node *next;
  59 };
  60
  61 enum mempool_type {
  62         MEMPOOL_TYPE_GLOBAL = 0,        /* Default */
  63         MEMPOOL_TYPE_PERCPU = 1,
  64 };
  65
  66 struct rseq_mempool_attr {
  67         bool mmap_set;
  68         void *(*mmap_func)(void *priv, size_t len);
  69         int (*munmap_func)(void *priv, void *ptr, size_t len);
  70         void *mmap_priv;
  71
  72         bool init_set;
  73         int (*init_func)(void *priv, void *addr, size_t len, int cpu);
  74         void *init_priv;
  75
  76         bool robust_set;
  77
  78         enum mempool_type type;
  79         size_t stride;
  80         int max_nr_cpus;
  81
  82         unsigned long max_nr_ranges;
  83
  84         bool poison_set;
  85         uintptr_t poison;
  86 };
  87
  88 struct rseq_mempool_range;
  89
  90 struct rseq_mempool_range {
  91         struct rseq_mempool_range *next;        /* Linked list of ranges. */
  92         struct rseq_mempool *pool;              /* Backward reference to container pool. */
  93         void *header;
  94         void *base;
  95         size_t next_unused;
  96         /* Track alloc/free. */
  97         unsigned long *alloc_bitmap;
  98 };
  99
 100 struct rseq_mempool {
 101         /* Head of ranges linked-list. */
 102         struct rseq_mempool_range *range_list;
 103         unsigned long nr_ranges;
 104
 105         size_t item_len;
 106         int item_order;
 107
 108         /*
 109          * The free list chains freed items on the CPU 0 address range.
 110          * We should rethink this decision if false sharing between
 111          * malloc/free from other CPUs and data accesses from CPU 0
 112          * becomes an issue. This is a NULL-terminated singly-linked
 113          * list.
 114          */
 115         struct free_list_node *free_list_head;
 116
 117         /* This lock protects allocation/free within the pool. */
 118         pthread_mutex_t lock;
 119
 120         struct rseq_mempool_attr attr;
 121         char *name;
 122 };
 123
 124 /*
 125  * Pool set entries are indexed by item_len rounded to the next power of
 126  * 2. A pool set can contain NULL pool entries, in which case the next
 127  * large enough entry will be used for allocation.
 128  */
 129 struct rseq_mempool_set {
 130         /* This lock protects add vs malloc/zmalloc within the pool set. */
 131         pthread_mutex_t lock;
 132         struct rseq_mempool *entries[POOL_SET_NR_ENTRIES];
 133 };
 134
 135 static
 136 const char *get_pool_name(const struct rseq_mempool *pool)
 137 {
 138         return pool->name ? : "<anonymous>";
 139 }
 140
 141 static
 142 void *__rseq_pool_range_percpu_ptr(struct rseq_mempool_range *range, int cpu,
 143                 uintptr_t item_offset, size_t stride)
 144 {
 145         return range->base + (stride * cpu) + item_offset;
 146 }
 147
 148 static
 149 void rseq_percpu_zero_item(struct rseq_mempool *pool,
 150                 struct rseq_mempool_range *range, uintptr_t item_offset)
 151 {
 152         int i;
 153
 154         for (i = 0; i < pool->attr.max_nr_cpus; i++) {
 155                 char *p = __rseq_pool_range_percpu_ptr(range, i,
 156                                 item_offset, pool->attr.stride);
 157                 memset(p, 0, pool->item_len);
 158         }
 159 }
 160
 161 static
 162 void rseq_percpu_poison_item(struct rseq_mempool *pool,
 163                 struct rseq_mempool_range *range, uintptr_t item_offset)
 164 {
 165         uintptr_t poison = pool->attr.poison;
 166         int i;
 167
 168         for (i = 0; i < pool->attr.max_nr_cpus; i++) {
 169                 char *p = __rseq_pool_range_percpu_ptr(range, i,
 170                                 item_offset, pool->attr.stride);
 171                 size_t offset;
 172
 173                 for (offset = 0; offset < pool->item_len; offset += sizeof(uintptr_t))
 174                         *((uintptr_t *) (p + offset)) = poison;
 175         }
 176 }
 177
 178 /* Always inline for __builtin_return_address(0). */
 179 static inline __attribute__((always_inline))
 180 void rseq_percpu_check_poison_item(struct rseq_mempool *pool,
 181                 struct rseq_mempool_range *range, uintptr_t item_offset)
 182 {
 183         uintptr_t poison = pool->attr.poison;
 184         int i;
 185
 186         if (!pool->attr.robust_set || !pool->attr.poison_set)
 187                 return;
 188         for (i = 0; i < pool->attr.max_nr_cpus; i++) {
 189                 char *p = __rseq_pool_range_percpu_ptr(range, i,
 190                                 item_offset, pool->attr.stride);
 191                 size_t offset;
 192
 193                 for (offset = 0; offset < pool->item_len; offset += sizeof(uintptr_t)) {
 194                         uintptr_t v;
 195
 196                         /* Skip poison check for free-list pointer. */
 197                         if (i == 0 && offset == 0)
 198                                 continue;
 199                         v = *((uintptr_t *) (p + offset));
 200                         if (v != poison) {
 201                                 fprintf(stderr, "%s: Poison corruption detected (0x%lx) for pool: \"%s\" (%p), item offset: %zu, caller: %p.\n",
 202                                         __func__, (unsigned long) v, get_pool_name(pool), pool, item_offset, (void *) __builtin_return_address(0));
 203                                 abort();
 204                         }
 205                 }
 206         }
 207 }
 208
 209 #ifdef HAVE_LIBNUMA
 210 int rseq_mempool_range_init_numa(void *addr, size_t len, int cpu, int numa_flags)
 211 {
 212         unsigned long nr_pages, page_len;
 213         int status[MOVE_PAGES_BATCH_SIZE];
 214         int nodes[MOVE_PAGES_BATCH_SIZE];
 215         void *pages[MOVE_PAGES_BATCH_SIZE];
 216         long ret;
 217
 218         if (!numa_flags) {
 219                 errno = EINVAL;
 220                 return -1;
 221         }
 222         page_len = rseq_get_page_len();
 223         nr_pages = len >> rseq_get_count_order_ulong(page_len);
 224
 225         nodes[0] = numa_node_of_cpu(cpu);
 226         if (nodes[0] < 0)
 227                 return -1;
 228
 229         for (size_t k = 1; k < RSEQ_ARRAY_SIZE(nodes); ++k) {
 230                 nodes[k] = nodes[0];
 231         }
 232
 233         for (unsigned long page = 0; page < nr_pages;) {
 234
 235                 size_t max_k = RSEQ_ARRAY_SIZE(pages);
 236                 size_t left = nr_pages - page;
 237
 238                 if (left < max_k) {
 239                         max_k = left;
 240                 }
 241
 242                 for (size_t k = 0; k < max_k; ++k, ++page) {
 243                         pages[k] = addr + (page * page_len);
 244                         status[k] = -EPERM;
 245                 }
 246
 247                 ret = move_pages(0, max_k, pages, nodes, status, numa_flags);
 248
 249                 if (ret < 0)
 250                         return ret;
 251
 252                 if (ret > 0) {
 253                         fprintf(stderr, "%lu pages were not migrated\n", ret);
 254                         for (size_t k = 0; k < max_k; ++k) {
 255                                 if (status[k] < 0)
 256                                         fprintf(stderr,
 257                                                 "Error while moving page %p to numa node %d: %u\n",
 258                                                 pages[k], nodes[k], -status[k]);
 259                         }
 260                 }
 261         }
 262         return 0;
 263 }
 264 #else
 265 int rseq_mempool_range_init_numa(void *addr __attribute__((unused)),
 266                 size_t len __attribute__((unused)),
 267                 int cpu __attribute__((unused)),
 268                 int numa_flags __attribute__((unused)))
 269 {
 270         errno = ENOSYS;
 271         return -1;
 272 }
 273 #endif
 274
 275 static
 276 void *default_mmap_func(void *priv __attribute__((unused)), size_t len)
 277 {
 278         void *base;
 279
 280         base = mmap(NULL, len, PROT_READ | PROT_WRITE,
 281                         MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
 282         if (base == MAP_FAILED)
 283                 return NULL;
 284         return base;
 285 }
 286
 287 static
 288 int default_munmap_func(void *priv __attribute__((unused)), void *ptr, size_t len)
 289 {
 290         return munmap(ptr, len);
 291 }
 292
 293 static
 294 int create_alloc_bitmap(struct rseq_mempool *pool, struct rseq_mempool_range *range)
 295 {
 296         size_t count;
 297
 298         count = ((pool->attr.stride >> pool->item_order) + BIT_PER_ULONG - 1) / BIT_PER_ULONG;
 299
 300         /*
 301          * Not being able to create the validation bitmap is an error
 302          * that needs to be reported.
 303          */
 304         range->alloc_bitmap = calloc(count, sizeof(unsigned long));
 305         if (!range->alloc_bitmap)
 306                 return -1;
 307         return 0;
 308 }
 309
 310 static
 311 bool addr_in_pool(const struct rseq_mempool *pool, void *addr)
 312 {
 313         struct rseq_mempool_range *range;
 314
 315         for (range = pool->range_list; range; range = range->next) {
 316                 if (addr >= range->base && addr < range->base + range->next_unused)
 317                         return true;
 318         }
 319         return false;
 320 }
 321
 322 /* Always inline for __builtin_return_address(0). */
 323 static inline __attribute__((always_inline))
 324 void check_free_list(const struct rseq_mempool *pool)
 325 {
 326         size_t total_item = 0, total_never_allocated = 0, total_freed = 0,
 327                 max_list_traversal = 0, traversal_iteration = 0;
 328         struct rseq_mempool_range *range;
 329
 330         if (!pool->attr.robust_set)
 331                 return;
 332
 333         for (range = pool->range_list; range; range = range->next) {
 334                 total_item += pool->attr.stride >> pool->item_order;
 335                 total_never_allocated += (pool->attr.stride - range->next_unused) >> pool->item_order;
 336         }
 337         max_list_traversal = total_item - total_never_allocated;
 338
 339         for (struct free_list_node *node = pool->free_list_head, *prev = NULL;
 340              node;
 341              prev = node,
 342              node = node->next) {
 343
 344                 void *node_addr = node;
 345
 346                 if (traversal_iteration >= max_list_traversal) {
 347                         fprintf(stderr, "%s: Corrupted free-list; Possibly infinite loop in pool \"%s\" (%p), caller %p.\n",
 348                                 __func__, get_pool_name(pool), pool, __builtin_return_address(0));
 349                         abort();
 350                 }
 351
 352                 /* Node is out of range. */
 353                 if (!addr_in_pool(pool, node_addr)) {
 354                         if (prev)
 355                                 fprintf(stderr, "%s: Corrupted free-list node %p -> [out-of-range %p] in pool \"%s\" (%p), caller %p.\n",
 356                                         __func__, prev, node, get_pool_name(pool), pool, __builtin_return_address(0));
 357                         else
 358                                 fprintf(stderr, "%s: Corrupted free-list node [out-of-range %p] in pool \"%s\" (%p), caller %p.\n",
 359                                         __func__, node, get_pool_name(pool), pool, __builtin_return_address(0));
 360                         abort();
 361                 }
 362
 363                 traversal_iteration++;
 364                 total_freed++;
 365         }
 366
 367         if (total_never_allocated + total_freed != total_item) {
 368                 fprintf(stderr, "%s: Corrupted free-list in pool \"%s\" (%p); total-item: %zu total-never-used: %zu total-freed: %zu, caller %p.\n",
 369                         __func__, get_pool_name(pool), pool, total_item, total_never_allocated, total_freed, __builtin_return_address(0));
 370                 abort();
 371         }
 372 }
 373
 374 /* Always inline for __builtin_return_address(0). */
 375 static inline __attribute__((always_inline))
 376 void destroy_alloc_bitmap(struct rseq_mempool *pool, struct rseq_mempool_range *range)
 377 {
 378         unsigned long *bitmap = range->alloc_bitmap;
 379         size_t count, total_leaks = 0;
 380
 381         if (!bitmap)
 382                 return;
 383
 384         count = ((pool->attr.stride >> pool->item_order) + BIT_PER_ULONG - 1) / BIT_PER_ULONG;
 385
 386         /* Assert that all items in the pool were freed. */
 387         for (size_t k = 0; k < count; ++k)
 388                 total_leaks += rseq_hweight_ulong(bitmap[k]);
 389         if (total_leaks) {
 390                 fprintf(stderr, "%s: Pool \"%s\" (%p) has %zu leaked items on destroy, caller: %p.\n",
 391                         __func__, get_pool_name(pool), pool, total_leaks, (void *) __builtin_return_address(0));
 392                 abort();
 393         }
 394
 395         free(bitmap);
 396 }
 397
 398 /* Always inline for __builtin_return_address(0). */
 399 static inline __attribute__((always_inline))
 400 int rseq_mempool_range_destroy(struct rseq_mempool *pool,
 401                 struct rseq_mempool_range *range)
 402 {
 403         destroy_alloc_bitmap(pool, range);
 404         /* range is a header located one page before the aligned mapping. */
 405         return pool->attr.munmap_func(pool->attr.mmap_priv, range->header,
 406                         (pool->attr.stride * pool->attr.max_nr_cpus) + rseq_get_page_len());
 407 }
 408
 409 /*
 410  * Allocate a memory mapping aligned on @alignment, with an optional
 411  * @pre_header before the mapping.
 412  */
 413 static
 414 void *aligned_mmap_anonymous(struct rseq_mempool *pool,
 415                 size_t page_size, size_t len, size_t alignment,
 416                 void **pre_header, size_t pre_header_len)
 417 {
 418         size_t minimum_page_count, page_count, extra, total_allocate = 0;
 419         int page_order;
 420         void *ptr;
 421
 422         if (len < page_size || alignment < page_size ||
 423                         !is_pow2(alignment) || (len & (alignment - 1))) {
 424                 errno = EINVAL;
 425                 return NULL;
 426         }
 427         page_order = rseq_get_count_order_ulong(page_size);
 428         if (page_order < 0) {
 429                 errno = EINVAL;
 430                 return NULL;
 431         }
 432         if (pre_header_len && (pre_header_len & (page_size - 1))) {
 433                 errno = EINVAL;
 434                 return NULL;
 435         }
 436
 437         minimum_page_count = (pre_header_len + len) >> page_order;
 438         page_count = (pre_header_len + len + alignment - page_size) >> page_order;
 439
 440         assert(page_count >= minimum_page_count);
 441
 442         ptr = pool->attr.mmap_func(pool->attr.mmap_priv, page_count << page_order);
 443         if (!ptr)
 444                 goto alloc_error;
 445
 446         total_allocate = page_count << page_order;
 447
 448         if (!(((uintptr_t) ptr + pre_header_len) & (alignment - 1))) {
 449                 /* Pointer is already aligned. ptr points to pre_header. */
 450                 goto out;
 451         }
 452
 453         /* Unmap extra before. */
 454         extra = offset_align((uintptr_t) ptr + pre_header_len, alignment);
 455         assert(!(extra & (page_size - 1)));
 456         if (pool->attr.munmap_func(pool->attr.mmap_priv, ptr, extra)) {
 457                 perror("munmap");
 458                 abort();
 459         }
 460         total_allocate -= extra;
 461         ptr += extra;   /* ptr points to pre_header */
 462         page_count -= extra >> page_order;
 463 out:
 464         assert(page_count >= minimum_page_count);
 465
 466         if (page_count > minimum_page_count) {
 467                 void *extra_ptr;
 468
 469                 /* Unmap extra after. */
 470                 extra_ptr = ptr + (minimum_page_count << page_order);
 471                 extra = (page_count - minimum_page_count) << page_order;
 472                 if (pool->attr.munmap_func(pool->attr.mmap_priv, extra_ptr, extra)) {
 473                         perror("munmap");
 474                         abort();
 475                 }
 476                 total_allocate -= extra;
 477         }
 478
 479         assert(!(((uintptr_t)ptr + pre_header_len) & (alignment - 1)));
 480         assert(total_allocate == len + pre_header_len);
 481
 482 alloc_error:
 483         if (ptr) {
 484                 if (pre_header)
 485                         *pre_header = ptr;
 486                 ptr += pre_header_len;
 487         }
 488         return ptr;
 489 }
 490
 491 static
 492 struct rseq_mempool_range *rseq_mempool_range_create(struct rseq_mempool *pool)
 493 {
 494         struct rseq_mempool_range *range;
 495         unsigned long page_size;
 496         void *header;
 497         void *base;
 498
 499         if (pool->attr.max_nr_ranges &&
 500                         pool->nr_ranges >= pool->attr.max_nr_ranges) {
 501                 errno = ENOMEM;
 502                 return NULL;
 503         }
 504         page_size = rseq_get_page_len();
 505
 506         base = aligned_mmap_anonymous(pool, page_size,
 507                         pool->attr.stride * pool->attr.max_nr_cpus,
 508                         pool->attr.stride,
 509                         &header, page_size);
 510         if (!base)
 511                 return NULL;
 512         range = (struct rseq_mempool_range *) (base - RANGE_HEADER_OFFSET);
 513         range->pool = pool;
 514         range->base = base;
 515         range->header = header;
 516         if (pool->attr.robust_set) {
 517                 if (create_alloc_bitmap(pool, range))
 518                         goto error_alloc;
 519         }
 520         if (pool->attr.init_set) {
 521                 switch (pool->attr.type) {
 522                 case MEMPOOL_TYPE_GLOBAL:
 523                         if (pool->attr.init_func(pool->attr.init_priv,
 524                                         base, pool->attr.stride, -1)) {
 525                                 goto error_alloc;
 526                         }
 527                         break;
 528                 case MEMPOOL_TYPE_PERCPU:
 529                 {
 530                         int cpu;
 531                         for (cpu = 0; cpu < pool->attr.max_nr_cpus; cpu++) {
 532                                 if (pool->attr.init_func(pool->attr.init_priv,
 533                                                 base + (pool->attr.stride * cpu),
 534                                                 pool->attr.stride, cpu)) {
 535                                         goto error_alloc;
 536                                 }
 537                         }
 538                         break;
 539                 }
 540                 default:
 541                         abort();
 542                 }
 543         }
 544         pool->nr_ranges++;
 545         return range;
 546
 547 error_alloc:
 548         (void) rseq_mempool_range_destroy(pool, range);
 549         return NULL;
 550 }
 551
 552 int rseq_mempool_destroy(struct rseq_mempool *pool)
 553 {
 554         struct rseq_mempool_range *range, *next_range;
 555         int ret = 0;
 556
 557         if (!pool)
 558                 return 0;
 559         check_free_list(pool);
 560         /* Iteration safe against removal. */
 561         for (range = pool->range_list; range && (next_range = range->next, 1); range = next_range) {
 562                 if (rseq_mempool_range_destroy(pool, range))
 563                         goto end;
 564                 /* Update list head to keep list coherent in case of partial failure. */
 565                 pool->range_list = next_range;
 566         }
 567         pthread_mutex_destroy(&pool->lock);
 568         free(pool->name);
 569         memset(pool, 0, sizeof(*pool));
 570 end:
 571         return ret;
 572 }
 573
 574 struct rseq_mempool *rseq_mempool_create(const char *pool_name,
 575                 size_t item_len, const struct rseq_mempool_attr *_attr)
 576 {
 577         struct rseq_mempool *pool;
 578         struct rseq_mempool_attr attr = {};
 579         int order;
 580
 581         /* Make sure each item is large enough to contain free list pointers. */
 582         if (item_len < sizeof(void *))
 583                 item_len = sizeof(void *);
 584
 585         /* Align item_len on next power of two. */
 586         order = rseq_get_count_order_ulong(item_len);
 587         if (order < 0) {
 588                 errno = EINVAL;
 589                 return NULL;
 590         }
 591         item_len = 1UL << order;
 592
 593         if (_attr)
 594                 memcpy(&attr, _attr, sizeof(attr));
 595         if (!attr.mmap_set) {
 596                 attr.mmap_func = default_mmap_func;
 597                 attr.munmap_func = default_munmap_func;
 598                 attr.mmap_priv = NULL;
 599         }
 600
 601         switch (attr.type) {
 602         case MEMPOOL_TYPE_PERCPU:
 603                 if (attr.max_nr_cpus < 0) {
 604                         errno = EINVAL;
 605                         return NULL;
 606                 }
 607                 if (attr.max_nr_cpus == 0) {
 608                         /* Auto-detect */
 609                         attr.max_nr_cpus = rseq_get_max_nr_cpus();
 610                         if (attr.max_nr_cpus == 0) {
 611                                 errno = EINVAL;
 612                                 return NULL;
 613                         }
 614                 }
 615                 break;
 616         case MEMPOOL_TYPE_GLOBAL:
 617                 /* Use a 1-cpu pool for global mempool type. */
 618                 attr.max_nr_cpus = 1;
 619                 break;
 620         }
 621         if (!attr.stride)
 622                 attr.stride = RSEQ_MEMPOOL_STRIDE;      /* Use default */
 623         if (item_len > attr.stride || attr.stride < (size_t) rseq_get_page_len() ||
 624                         !is_pow2(attr.stride)) {
 625                 errno = EINVAL;
 626                 return NULL;
 627         }
 628
 629         pool = calloc(1, sizeof(struct rseq_mempool));
 630         if (!pool)
 631                 return NULL;
 632
 633         memcpy(&pool->attr, &attr, sizeof(attr));
 634         pthread_mutex_init(&pool->lock, NULL);
 635         pool->item_len = item_len;
 636         pool->item_order = order;
 637
 638         pool->range_list = rseq_mempool_range_create(pool);
 639         if (!pool->range_list)
 640                 goto error_alloc;
 641
 642         if (pool_name) {
 643                 pool->name = strdup(pool_name);
 644                 if (!pool->name)
 645                         goto error_alloc;
 646         }
 647         return pool;
 648
 649 error_alloc:
 650         rseq_mempool_destroy(pool);
 651         errno = ENOMEM;
 652         return NULL;
 653 }
 654
 655 /* Always inline for __builtin_return_address(0). */
 656 static inline __attribute__((always_inline))
 657 void set_alloc_slot(struct rseq_mempool *pool, struct rseq_mempool_range *range, size_t item_offset)
 658 {
 659         unsigned long *bitmap = range->alloc_bitmap;
 660         size_t item_index = item_offset >> pool->item_order;
 661         unsigned long mask;
 662         size_t k;
 663
 664         if (!bitmap)
 665                 return;
 666
 667         k = item_index / BIT_PER_ULONG;
 668         mask = 1ULL << (item_index % BIT_PER_ULONG);
 669
 670         /* Print error if bit is already set. */
 671         if (bitmap[k] & mask) {
 672                 fprintf(stderr, "%s: Allocator corruption detected for pool: \"%s\" (%p), item offset: %zu, caller: %p.\n",
 673                         __func__, get_pool_name(pool), pool, item_offset, (void *) __builtin_return_address(0));
 674                 abort();
 675         }
 676         bitmap[k] |= mask;
 677 }
 678
 679 static
 680 void __rseq_percpu *__rseq_percpu_malloc(struct rseq_mempool *pool, bool zeroed)
 681 {
 682         struct rseq_mempool_range *range;
 683         struct free_list_node *node;
 684         uintptr_t item_offset;
 685         void __rseq_percpu *addr;
 686
 687         pthread_mutex_lock(&pool->lock);
 688         /* Get first entry from free list. */
 689         node = pool->free_list_head;
 690         if (node != NULL) {
 691                 uintptr_t ptr = (uintptr_t) node;
 692                 void *range_base = (void *) (ptr & (~(pool->attr.stride - 1)));
 693
 694                 range = (struct rseq_mempool_range *) (range_base - RANGE_HEADER_OFFSET);
 695                 /* Remove node from free list (update head). */
 696                 pool->free_list_head = node->next;
 697                 item_offset = (uintptr_t) ((void *) node - range_base);
 698                 rseq_percpu_check_poison_item(pool, range, item_offset);
 699                 addr = (void __rseq_percpu *) node;
 700                 goto end;
 701         }
 702         /*
 703          * If the most recent range (first in list) does not have any
 704          * room left, create a new range and prepend it to the list
 705          * head.
 706          */
 707         range = pool->range_list;
 708         if (range->next_unused + pool->item_len > pool->attr.stride) {
 709                 range = rseq_mempool_range_create(pool);
 710                 if (!range) {
 711                         errno = ENOMEM;
 712                         addr = NULL;
 713                         goto end;
 714                 }
 715                 /* Add range to head of list. */
 716                 range->next = pool->range_list;
 717                 pool->range_list = range;
 718         }
 719         /* First range in list has room left. */
 720         item_offset = range->next_unused;
 721         addr = (void __rseq_percpu *) (range->base + item_offset);
 722         range->next_unused += pool->item_len;
 723 end:
 724         if (addr)
 725                 set_alloc_slot(pool, range, item_offset);
 726         pthread_mutex_unlock(&pool->lock);
 727         if (zeroed && addr)
 728                 rseq_percpu_zero_item(pool, range, item_offset);
 729         return addr;
 730 }
 731
 732 void __rseq_percpu *rseq_mempool_percpu_malloc(struct rseq_mempool *pool)
 733 {
 734         return __rseq_percpu_malloc(pool, false);
 735 }
 736
 737 void __rseq_percpu *rseq_mempool_percpu_zmalloc(struct rseq_mempool *pool)
 738 {
 739         return __rseq_percpu_malloc(pool, true);
 740 }
 741
 742 /* Always inline for __builtin_return_address(0). */
 743 static inline __attribute__((always_inline))
 744 void clear_alloc_slot(struct rseq_mempool *pool, struct rseq_mempool_range *range, size_t item_offset)
 745 {
 746         unsigned long *bitmap = range->alloc_bitmap;
 747         size_t item_index = item_offset >> pool->item_order;
 748         unsigned long mask;
 749         size_t k;
 750
 751         if (!bitmap)
 752                 return;
 753
 754         k = item_index / BIT_PER_ULONG;
 755         mask = 1ULL << (item_index % BIT_PER_ULONG);
 756
 757         /* Print error if bit is not set. */
 758         if (!(bitmap[k] & mask)) {
 759                 fprintf(stderr, "%s: Double-free detected for pool: \"%s\" (%p), item offset: %zu, caller: %p.\n",
 760                         __func__, get_pool_name(pool), pool, item_offset,
 761                         (void *) __builtin_return_address(0));
 762                 abort();
 763         }
 764         bitmap[k] &= ~mask;
 765 }
 766
 767 void librseq_mempool_percpu_free(void __rseq_percpu *_ptr, size_t stride)
 768 {
 769         uintptr_t ptr = (uintptr_t) _ptr;
 770         void *range_base = (void *) (ptr & (~(stride - 1)));
 771         struct rseq_mempool_range *range = (struct rseq_mempool_range *) (range_base - RANGE_HEADER_OFFSET);
 772         struct rseq_mempool *pool = range->pool;
 773         uintptr_t item_offset = ptr & (stride - 1);
 774         struct free_list_node *head, *item;
 775
 776         pthread_mutex_lock(&pool->lock);
 777         clear_alloc_slot(pool, range, item_offset);
 778         /* Add ptr to head of free list */
 779         head = pool->free_list_head;
 780         if (pool->attr.poison_set)
 781                 rseq_percpu_poison_item(pool, range, item_offset);
 782         /* Free-list is in CPU 0 range. */
 783         item = (struct free_list_node *) ptr;
 784         /*
 785          * Setting the next pointer will overwrite the first uintptr_t
 786          * poison for CPU 0.
 787          */
 788         item->next = head;
 789         pool->free_list_head = item;
 790         pthread_mutex_unlock(&pool->lock);
 791 }
 792
 793 struct rseq_mempool_set *rseq_mempool_set_create(void)
 794 {
 795         struct rseq_mempool_set *pool_set;
 796
 797         pool_set = calloc(1, sizeof(struct rseq_mempool_set));
 798         if (!pool_set)
 799                 return NULL;
 800         pthread_mutex_init(&pool_set->lock, NULL);
 801         return pool_set;
 802 }
 803
 804 int rseq_mempool_set_destroy(struct rseq_mempool_set *pool_set)
 805 {
 806         int order, ret;
 807
 808         for (order = POOL_SET_MIN_ENTRY; order < POOL_SET_NR_ENTRIES; order++) {
 809                 struct rseq_mempool *pool = pool_set->entries[order];
 810
 811                 if (!pool)
 812                         continue;
 813                 ret = rseq_mempool_destroy(pool);
 814                 if (ret)
 815                         return ret;
 816                 pool_set->entries[order] = NULL;
 817         }
 818         pthread_mutex_destroy(&pool_set->lock);
 819         free(pool_set);
 820         return 0;
 821 }
 822
 823 /* Ownership of pool is handed over to pool set on success. */
 824 int rseq_mempool_set_add_pool(struct rseq_mempool_set *pool_set, struct rseq_mempool *pool)
 825 {
 826         size_t item_order = pool->item_order;
 827         int ret = 0;
 828
 829         pthread_mutex_lock(&pool_set->lock);
 830         if (pool_set->entries[item_order]) {
 831                 errno = EBUSY;
 832                 ret = -1;
 833                 goto end;
 834         }
 835         pool_set->entries[pool->item_order] = pool;
 836 end:
 837         pthread_mutex_unlock(&pool_set->lock);
 838         return ret;
 839 }
 840
 841 static
 842 void __rseq_percpu *__rseq_mempool_set_malloc(struct rseq_mempool_set *pool_set, size_t len, bool zeroed)
 843 {
 844         int order, min_order = POOL_SET_MIN_ENTRY;
 845         struct rseq_mempool *pool;
 846         void __rseq_percpu *addr;
 847
 848         order = rseq_get_count_order_ulong(len);
 849         if (order > POOL_SET_MIN_ENTRY)
 850                 min_order = order;
 851 again:
 852         pthread_mutex_lock(&pool_set->lock);
 853         /* First smallest present pool where @len fits. */
 854         for (order = min_order; order < POOL_SET_NR_ENTRIES; order++) {
 855                 pool = pool_set->entries[order];
 856
 857                 if (!pool)
 858                         continue;
 859                 if (pool->item_len >= len)
 860                         goto found;
 861         }
 862         pool = NULL;
 863 found:
 864         pthread_mutex_unlock(&pool_set->lock);
 865         if (pool) {
 866                 addr = __rseq_percpu_malloc(pool, zeroed);
 867                 if (addr == NULL && errno == ENOMEM) {
 868                         /*
 869                          * If the allocation failed, try again with a
 870                          * larger pool.
 871                          */
 872                         min_order = order + 1;
 873                         goto again;
 874                 }
 875         } else {
 876                 /* Not found. */
 877                 errno = ENOMEM;
 878                 addr = NULL;
 879         }
 880         return addr;
 881 }
 882
 883 void __rseq_percpu *rseq_mempool_set_percpu_malloc(struct rseq_mempool_set *pool_set, size_t len)
 884 {
 885         return __rseq_mempool_set_malloc(pool_set, len, false);
 886 }
 887
 888 void __rseq_percpu *rseq_mempool_set_percpu_zmalloc(struct rseq_mempool_set *pool_set, size_t len)
 889 {
 890         return __rseq_mempool_set_malloc(pool_set, len, true);
 891 }
 892
 893 struct rseq_mempool_attr *rseq_mempool_attr_create(void)
 894 {
 895         return calloc(1, sizeof(struct rseq_mempool_attr));
 896 }
 897
 898 void rseq_mempool_attr_destroy(struct rseq_mempool_attr *attr)
 899 {
 900         free(attr);
 901 }
 902
 903 int rseq_mempool_attr_set_mmap(struct rseq_mempool_attr *attr,
 904                 void *(*mmap_func)(void *priv, size_t len),
 905                 int (*munmap_func)(void *priv, void *ptr, size_t len),
 906                 void *mmap_priv)
 907 {
 908         if (!attr) {
 909                 errno = EINVAL;
 910                 return -1;
 911         }
 912         attr->mmap_set = true;
 913         attr->mmap_func = mmap_func;
 914         attr->munmap_func = munmap_func;
 915         attr->mmap_priv = mmap_priv;
 916         return 0;
 917 }
 918
 919 int rseq_mempool_attr_set_init(struct rseq_mempool_attr *attr,
 920                 int (*init_func)(void *priv, void *addr, size_t len, int cpu),
 921                 void *init_priv)
 922 {
 923         if (!attr) {
 924                 errno = EINVAL;
 925                 return -1;
 926         }
 927         attr->init_set = true;
 928         attr->init_func = init_func;
 929         attr->init_priv = init_priv;
 930         return 0;
 931 }
 932
 933 int rseq_mempool_attr_set_robust(struct rseq_mempool_attr *attr)
 934 {
 935         if (!attr) {
 936                 errno = EINVAL;
 937                 return -1;
 938         }
 939         attr->robust_set = true;
 940         return 0;
 941 }
 942
 943 int rseq_mempool_attr_set_percpu(struct rseq_mempool_attr *attr,
 944                 size_t stride, int max_nr_cpus)
 945 {
 946         if (!attr) {
 947                 errno = EINVAL;
 948                 return -1;
 949         }
 950         attr->type = MEMPOOL_TYPE_PERCPU;
 951         attr->stride = stride;
 952         attr->max_nr_cpus = max_nr_cpus;
 953         return 0;
 954 }
 955
 956 int rseq_mempool_attr_set_global(struct rseq_mempool_attr *attr,
 957                 size_t stride)
 958 {
 959         if (!attr) {
 960                 errno = EINVAL;
 961                 return -1;
 962         }
 963         attr->type = MEMPOOL_TYPE_GLOBAL;
 964         attr->stride = stride;
 965         attr->max_nr_cpus = 0;
 966         return 0;
 967 }
 968
 969 int rseq_mempool_attr_set_max_nr_ranges(struct rseq_mempool_attr *attr,
 970                 unsigned long max_nr_ranges)
 971 {
 972         if (!attr) {
 973                 errno = EINVAL;
 974                 return -1;
 975         }
 976         attr->max_nr_ranges = max_nr_ranges;
 977         return 0;
 978 }
 979
 980 int rseq_mempool_attr_set_poison(struct rseq_mempool_attr *attr,
 981                 uintptr_t poison)
 982 {
 983         if (!attr) {
 984                 errno = EINVAL;
 985                 return -1;
 986         }
 987         attr->poison_set = true;
 988         attr->poison = poison;
 989         return 0;
 990 }
 991
 992 int rseq_mempool_get_max_nr_cpus(struct rseq_mempool *mempool)
 993 {
 994         if (!mempool || mempool->attr.type != MEMPOOL_TYPE_PERCPU) {
 995                 errno = EINVAL;
 996                 return -1;
 997         }
 998         return mempool->attr.max_nr_cpus;
 999 }