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