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 #include <fcntl.h>
  17
  18 #ifdef HAVE_LIBNUMA
  19 # include <numa.h>
  20 # include <numaif.h>
  21 #endif
  22
  23 #include "rseq-utils.h"
  24 #include <rseq/rseq.h>
  25
  26 /*
  27  * rseq-mempool.c: rseq CPU-Local Storage (CLS) memory allocator.
  28  *
  29  * The rseq per-CPU memory allocator allows the application the request
  30  * memory pools of CPU-Local memory each of containing objects of a
  31  * given size (rounded to next power of 2), reserving a given virtual
  32  * address size per CPU, for a given maximum number of CPUs.
  33  *
  34  * The per-CPU memory allocator is analogous to TLS (Thread-Local
  35  * Storage) memory: TLS is Thread-Local Storage, whereas the per-CPU
  36  * memory allocator provides CPU-Local Storage.
  37  */
  38
  39 #define POOL_SET_NR_ENTRIES     RSEQ_BITS_PER_LONG
  40
  41 /*
  42  * Smallest allocation should hold enough space for a free list pointer.
  43  */
  44 #if RSEQ_BITS_PER_LONG == 64
  45 # define POOL_SET_MIN_ENTRY     3       /* Smallest item_len=8 */
  46 #else
  47 # define POOL_SET_MIN_ENTRY     2       /* Smallest item_len=4 */
  48 #endif
  49
  50 #define BIT_PER_ULONG           (8 * sizeof(unsigned long))
  51
  52 #define MOVE_PAGES_BATCH_SIZE   4096
  53
  54 #define RANGE_HEADER_OFFSET     sizeof(struct rseq_mempool_range)
  55
  56 #if RSEQ_BITS_PER_LONG == 64
  57 # define DEFAULT_PRIVATE_POISON_VALUE   0x5555555555555555ULL
  58 #else
  59 # define DEFAULT_PRIVATE_POISON_VALUE   0x55555555UL
  60 #endif
  61
  62 struct free_list_node;
  63
  64 struct free_list_node {
  65         struct free_list_node *next;
  66 };
  67
  68 enum mempool_type {
  69         MEMPOOL_TYPE_GLOBAL = 0,        /* Default */
  70         MEMPOOL_TYPE_PERCPU = 1,
  71 };
  72
  73 struct rseq_mempool_attr {
  74         bool init_set;
  75         int (*init_func)(void *priv, void *addr, size_t len, int cpu);
  76         void *init_priv;
  77
  78         bool robust_set;
  79
  80         enum mempool_type type;
  81         size_t stride;
  82         int max_nr_cpus;
  83
  84         unsigned long max_nr_ranges;
  85
  86         bool poison_set;
  87         uintptr_t poison;
  88
  89         enum rseq_mempool_populate_policy populate_policy;
  90 };
  91
  92 struct rseq_mempool_range;
  93
  94 struct rseq_mempool_range {
  95         struct rseq_mempool_range *next;        /* Linked list of ranges. */
  96         struct rseq_mempool *pool;              /* Backward reference to container pool. */
  97
  98         /*
  99          * Memory layout of a mempool range:
 100          * - Header page (contains struct rseq_mempool_range at the very end),
 101          * - Base of the per-cpu data, starting with CPU 0.
 102          *   Aliases with free-list for non-robust populate all pool.
 103          * - CPU 1,
 104          * ...
 105          * - CPU max_nr_cpus - 1
 106          * - init values (unpopulated for RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL).
 107          *   Aliases with free-list for non-robust populate none pool.
 108          * - free list (for robust pool).
 109          *
 110          * The free list aliases the CPU 0 memory area for non-robust
 111          * populate all pools. It aliases with init values for
 112          * non-robust populate none pools. It is located immediately
 113          * after the init values for robust pools.
 114          */
 115         void *header;
 116         void *base;
 117         /*
 118          * The init values contains malloc_init/zmalloc values.
 119          * Pointer is NULL for RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL.
 120          */
 121         void *init;
 122         size_t next_unused;
 123
 124         /* Pool range mmap/munmap */
 125         void *mmap_addr;
 126         size_t mmap_len;
 127
 128         /* Track alloc/free. */
 129         unsigned long *alloc_bitmap;
 130 };
 131
 132 struct rseq_mempool {
 133         /* Head of ranges linked-list. */
 134         struct rseq_mempool_range *range_list;
 135         unsigned long nr_ranges;
 136
 137         size_t item_len;
 138         int item_order;
 139
 140         /*
 141          * The free list chains freed items on the CPU 0 address range.
 142          * We should rethink this decision if false sharing between
 143          * malloc/free from other CPUs and data accesses from CPU 0
 144          * becomes an issue. This is a NULL-terminated singly-linked
 145          * list.
 146          */
 147         struct free_list_node *free_list_head;
 148
 149         /* This lock protects allocation/free within the pool. */
 150         pthread_mutex_t lock;
 151
 152         struct rseq_mempool_attr attr;
 153         char *name;
 154 };
 155
 156 /*
 157  * Pool set entries are indexed by item_len rounded to the next power of
 158  * 2. A pool set can contain NULL pool entries, in which case the next
 159  * large enough entry will be used for allocation.
 160  */
 161 struct rseq_mempool_set {
 162         /* This lock protects add vs malloc/zmalloc within the pool set. */
 163         pthread_mutex_t lock;
 164         struct rseq_mempool *entries[POOL_SET_NR_ENTRIES];
 165 };
 166
 167 static
 168 const char *get_pool_name(const struct rseq_mempool *pool)
 169 {
 170         return pool->name ? : "<anonymous>";
 171 }
 172
 173 static
 174 void *__rseq_pool_range_percpu_ptr(const struct rseq_mempool_range *range, int cpu,
 175                 uintptr_t item_offset, size_t stride)
 176 {
 177         return range->base + (stride * cpu) + item_offset;
 178 }
 179
 180 static
 181 void *__rseq_pool_range_init_ptr(const struct rseq_mempool_range *range,
 182                 uintptr_t item_offset)
 183 {
 184         if (!range->init)
 185                 return NULL;
 186         return range->init + item_offset;
 187 }
 188
 189 static
 190 void __rseq_percpu *__rseq_free_list_to_percpu_ptr(const struct rseq_mempool *pool,
 191                 struct free_list_node *node)
 192 {
 193         void __rseq_percpu *p = (void __rseq_percpu *) node;
 194
 195         if (pool->attr.robust_set) {
 196                 /* Skip cpus. */
 197                 p -= pool->attr.max_nr_cpus * pool->attr.stride;
 198                 /* Skip init values */
 199                 if (pool->attr.populate_policy != RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL)
 200                         p -= pool->attr.stride;
 201
 202         } else {
 203                 /* Populate none free list is in init values */
 204                 if (pool->attr.populate_policy != RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL)
 205                         p -= pool->attr.max_nr_cpus * pool->attr.stride;
 206         }
 207         return p;
 208 }
 209
 210 static
 211 struct free_list_node *__rseq_percpu_to_free_list_ptr(const struct rseq_mempool *pool,
 212                 void __rseq_percpu *p)
 213 {
 214         if (pool->attr.robust_set) {
 215                 /* Skip cpus. */
 216                 p += pool->attr.max_nr_cpus * pool->attr.stride;
 217                 /* Skip init values */
 218                 if (pool->attr.populate_policy != RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL)
 219                         p += pool->attr.stride;
 220
 221         } else {
 222                 /* Populate none free list is in init values */
 223                 if (pool->attr.populate_policy != RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL)
 224                         p += pool->attr.max_nr_cpus * pool->attr.stride;
 225         }
 226         return (struct free_list_node *) p;
 227 }
 228
 229 static
 230 intptr_t rseq_cmp_item(void *p, size_t item_len, intptr_t cmp_value, intptr_t *unexpected_value)
 231 {
 232         size_t offset;
 233         intptr_t res = 0;
 234
 235         for (offset = 0; offset < item_len; offset += sizeof(uintptr_t)) {
 236                 intptr_t v = *((intptr_t *) (p + offset));
 237
 238                 if ((res = v - cmp_value) != 0) {
 239                         if (unexpected_value)
 240                                 *unexpected_value = v;
 241                         break;
 242                 }
 243         }
 244         return res;
 245 }
 246
 247 static
 248 void rseq_percpu_zero_item(struct rseq_mempool *pool,
 249                 struct rseq_mempool_range *range, uintptr_t item_offset)
 250 {
 251         char *init_p = NULL;
 252         int i;
 253
 254         init_p = __rseq_pool_range_init_ptr(range, item_offset);
 255         if (init_p)
 256                 bzero(init_p, pool->item_len);
 257         for (i = 0; i < pool->attr.max_nr_cpus; i++) {
 258                 char *p = __rseq_pool_range_percpu_ptr(range, i,
 259                                 item_offset, pool->attr.stride);
 260
 261                 /*
 262                  * If item is already zeroed, either because the
 263                  * init range update has propagated or because the
 264                  * content is already zeroed (e.g. zero page), don't
 265                  * write to the page. This eliminates useless COW over
 266                  * the zero page just for overwriting it with zeroes.
 267                  *
 268                  * This means zmalloc() in populate all policy pool do
 269                  * not trigger COW for CPUs which are not actively
 270                  * writing to the pool. This is however not the case for
 271                  * malloc_init() in populate-all pools if it populates
 272                  * non-zero content.
 273                  */
 274                 if (!rseq_cmp_item(p, pool->item_len, 0, NULL))
 275                         continue;
 276                 bzero(p, pool->item_len);
 277         }
 278 }
 279
 280 static
 281 void rseq_percpu_init_item(struct rseq_mempool *pool,
 282                 struct rseq_mempool_range *range, uintptr_t item_offset,
 283                 void *init_ptr, size_t init_len)
 284 {
 285         char *init_p = NULL;
 286         int i;
 287
 288         init_p = __rseq_pool_range_init_ptr(range, item_offset);
 289         if (init_p)
 290                 memcpy(init_p, init_ptr, init_len);
 291         for (i = 0; i < pool->attr.max_nr_cpus; i++) {
 292                 char *p = __rseq_pool_range_percpu_ptr(range, i,
 293                                 item_offset, pool->attr.stride);
 294
 295                 /*
 296                  * If the update propagated through a shared mapping,
 297                  * or the item already has the correct content, skip
 298                  * writing it into the cpu item to eliminate useless
 299                  * COW of the page.
 300                  */
 301                 if (!memcmp(init_ptr, p, init_len))
 302                         continue;
 303                 memcpy(p, init_ptr, init_len);
 304         }
 305 }
 306
 307 static
 308 void rseq_poison_item(void *p, size_t item_len, uintptr_t poison)
 309 {
 310         size_t offset;
 311
 312         for (offset = 0; offset < item_len; offset += sizeof(uintptr_t))
 313                 *((uintptr_t *) (p + offset)) = poison;
 314 }
 315
 316 static
 317 void rseq_percpu_poison_item(struct rseq_mempool *pool,
 318                 struct rseq_mempool_range *range, uintptr_t item_offset)
 319 {
 320         uintptr_t poison = pool->attr.poison;
 321         char *init_p = NULL;
 322         int i;
 323
 324         init_p = __rseq_pool_range_init_ptr(range, item_offset);
 325         if (init_p)
 326                 rseq_poison_item(init_p, pool->item_len, poison);
 327         for (i = 0; i < pool->attr.max_nr_cpus; i++) {
 328                 char *p = __rseq_pool_range_percpu_ptr(range, i,
 329                                 item_offset, pool->attr.stride);
 330
 331                 /*
 332                  * If the update propagated through a shared mapping,
 333                  * or the item already has the correct content, skip
 334                  * writing it into the cpu item to eliminate useless
 335                  * COW of the page.
 336                  *
 337                  * It is recommended to use zero as poison value for
 338                  * populate-all pools to eliminate COW due to writing
 339                  * poison to unused CPU memory.
 340                  */
 341                 if (rseq_cmp_item(p, pool->item_len, poison, NULL) == 0)
 342                         continue;
 343                 rseq_poison_item(p, pool->item_len, poison);
 344         }
 345 }
 346
 347 /* Always inline for __builtin_return_address(0). */
 348 static inline __attribute__((always_inline))
 349 void rseq_check_poison_item(const struct rseq_mempool *pool, uintptr_t item_offset,
 350                 void *p, size_t item_len, uintptr_t poison)
 351 {
 352         intptr_t unexpected_value;
 353
 354         if (rseq_cmp_item(p, item_len, poison, &unexpected_value) == 0)
 355                 return;
 356
 357         fprintf(stderr, "%s: Poison corruption detected (0x%lx) for pool: \"%s\" (%p), item offset: %zu, caller: %p.\n",
 358                 __func__, (unsigned long) unexpected_value, get_pool_name(pool), pool, item_offset, (void *) __builtin_return_address(0));
 359         abort();
 360 }
 361
 362 /* Always inline for __builtin_return_address(0). */
 363 static inline __attribute__((always_inline))
 364 void rseq_percpu_check_poison_item(const struct rseq_mempool *pool,
 365                 const struct rseq_mempool_range *range, uintptr_t item_offset)
 366 {
 367         uintptr_t poison = pool->attr.poison;
 368         char *init_p;
 369         int i;
 370
 371         if (!pool->attr.robust_set)
 372                 return;
 373         init_p = __rseq_pool_range_init_ptr(range, item_offset);
 374         if (init_p)
 375                 rseq_check_poison_item(pool, item_offset, init_p, pool->item_len, poison);
 376         for (i = 0; i < pool->attr.max_nr_cpus; i++) {
 377                 char *p = __rseq_pool_range_percpu_ptr(range, i,
 378                                 item_offset, pool->attr.stride);
 379                 rseq_check_poison_item(pool, item_offset, p, pool->item_len, poison);
 380         }
 381 }
 382
 383 #ifdef HAVE_LIBNUMA
 384 int rseq_mempool_range_init_numa(void *addr, size_t len, int cpu, int numa_flags)
 385 {
 386         unsigned long nr_pages, page_len;
 387         int status[MOVE_PAGES_BATCH_SIZE];
 388         int nodes[MOVE_PAGES_BATCH_SIZE];
 389         void *pages[MOVE_PAGES_BATCH_SIZE];
 390         long ret;
 391
 392         if (!numa_flags) {
 393                 errno = EINVAL;
 394                 return -1;
 395         }
 396         page_len = rseq_get_page_len();
 397         nr_pages = len >> rseq_get_count_order_ulong(page_len);
 398
 399         nodes[0] = numa_node_of_cpu(cpu);
 400         if (nodes[0] < 0)
 401                 return -1;
 402
 403         for (size_t k = 1; k < RSEQ_ARRAY_SIZE(nodes); ++k) {
 404                 nodes[k] = nodes[0];
 405         }
 406
 407         for (unsigned long page = 0; page < nr_pages;) {
 408
 409                 size_t max_k = RSEQ_ARRAY_SIZE(pages);
 410                 size_t left = nr_pages - page;
 411
 412                 if (left < max_k) {
 413                         max_k = left;
 414                 }
 415
 416                 for (size_t k = 0; k < max_k; ++k, ++page) {
 417                         pages[k] = addr + (page * page_len);
 418                         status[k] = -EPERM;
 419                 }
 420
 421                 ret = move_pages(0, max_k, pages, nodes, status, numa_flags);
 422
 423                 if (ret < 0)
 424                         return ret;
 425
 426                 if (ret > 0) {
 427                         fprintf(stderr, "%lu pages were not migrated\n", ret);
 428                         for (size_t k = 0; k < max_k; ++k) {
 429                                 if (status[k] < 0)
 430                                         fprintf(stderr,
 431                                                 "Error while moving page %p to numa node %d: %u\n",
 432                                                 pages[k], nodes[k], -status[k]);
 433                         }
 434                 }
 435         }
 436         return 0;
 437 }
 438 #else
 439 int rseq_mempool_range_init_numa(void *addr __attribute__((unused)),
 440                 size_t len __attribute__((unused)),
 441                 int cpu __attribute__((unused)),
 442                 int numa_flags __attribute__((unused)))
 443 {
 444         errno = ENOSYS;
 445         return -1;
 446 }
 447 #endif
 448
 449 static
 450 int create_alloc_bitmap(struct rseq_mempool *pool, struct rseq_mempool_range *range)
 451 {
 452         size_t count;
 453
 454         count = ((pool->attr.stride >> pool->item_order) + BIT_PER_ULONG - 1) / BIT_PER_ULONG;
 455
 456         /*
 457          * Not being able to create the validation bitmap is an error
 458          * that needs to be reported.
 459          */
 460         range->alloc_bitmap = calloc(count, sizeof(unsigned long));
 461         if (!range->alloc_bitmap)
 462                 return -1;
 463         return 0;
 464 }
 465
 466 static
 467 bool percpu_addr_in_pool(const struct rseq_mempool *pool, void __rseq_percpu *_addr)
 468 {
 469         struct rseq_mempool_range *range;
 470         void *addr = (void *) _addr;
 471
 472         for (range = pool->range_list; range; range = range->next) {
 473                 if (addr >= range->base && addr < range->base + range->next_unused)
 474                         return true;
 475         }
 476         return false;
 477 }
 478
 479 /* Always inline for __builtin_return_address(0). */
 480 static inline __attribute__((always_inline))
 481 void check_free_list(const struct rseq_mempool *pool)
 482 {
 483         size_t total_item = 0, total_never_allocated = 0, total_freed = 0,
 484                 max_list_traversal = 0, traversal_iteration = 0;
 485         struct rseq_mempool_range *range;
 486
 487         if (!pool->attr.robust_set)
 488                 return;
 489
 490         for (range = pool->range_list; range; range = range->next) {
 491                 total_item += pool->attr.stride >> pool->item_order;
 492                 total_never_allocated += (pool->attr.stride - range->next_unused) >> pool->item_order;
 493         }
 494         max_list_traversal = total_item - total_never_allocated;
 495
 496         for (struct free_list_node *node = pool->free_list_head, *prev = NULL;
 497              node;
 498              prev = node,
 499              node = node->next) {
 500
 501                 if (traversal_iteration >= max_list_traversal) {
 502                         fprintf(stderr, "%s: Corrupted free-list; Possibly infinite loop in pool \"%s\" (%p), caller %p.\n",
 503                                 __func__, get_pool_name(pool), pool, __builtin_return_address(0));
 504                         abort();
 505                 }
 506
 507                 /* Node is out of range. */
 508                 if (!percpu_addr_in_pool(pool, __rseq_free_list_to_percpu_ptr(pool, node))) {
 509                         if (prev)
 510                                 fprintf(stderr, "%s: Corrupted free-list node %p -> [out-of-range %p] in pool \"%s\" (%p), caller %p.\n",
 511                                         __func__, prev, node, get_pool_name(pool), pool, __builtin_return_address(0));
 512                         else
 513                                 fprintf(stderr, "%s: Corrupted free-list node [out-of-range %p] in pool \"%s\" (%p), caller %p.\n",
 514                                         __func__, node, get_pool_name(pool), pool, __builtin_return_address(0));
 515                         abort();
 516                 }
 517
 518                 traversal_iteration++;
 519                 total_freed++;
 520         }
 521
 522         if (total_never_allocated + total_freed != total_item) {
 523                 fprintf(stderr, "%s: Corrupted free-list in pool \"%s\" (%p); total-item: %zu total-never-used: %zu total-freed: %zu, caller %p.\n",
 524                         __func__, get_pool_name(pool), pool, total_item, total_never_allocated, total_freed, __builtin_return_address(0));
 525                 abort();
 526         }
 527 }
 528
 529 /* Always inline for __builtin_return_address(0). */
 530 static inline __attribute__((always_inline))
 531 void check_range_poison(const struct rseq_mempool *pool,
 532                 const struct rseq_mempool_range *range)
 533 {
 534         size_t item_offset;
 535
 536         for (item_offset = 0; item_offset < range->next_unused;
 537                         item_offset += pool->item_len)
 538                 rseq_percpu_check_poison_item(pool, range, item_offset);
 539 }
 540
 541 /* Always inline for __builtin_return_address(0). */
 542 static inline __attribute__((always_inline))
 543 void check_pool_poison(const struct rseq_mempool *pool)
 544 {
 545         struct rseq_mempool_range *range;
 546
 547         if (!pool->attr.robust_set)
 548                 return;
 549         for (range = pool->range_list; range; range = range->next)
 550                 check_range_poison(pool, range);
 551 }
 552
 553 /* Always inline for __builtin_return_address(0). */
 554 static inline __attribute__((always_inline))
 555 void destroy_alloc_bitmap(struct rseq_mempool *pool, struct rseq_mempool_range *range)
 556 {
 557         unsigned long *bitmap = range->alloc_bitmap;
 558         size_t count, total_leaks = 0;
 559
 560         if (!bitmap)
 561                 return;
 562
 563         count = ((pool->attr.stride >> pool->item_order) + BIT_PER_ULONG - 1) / BIT_PER_ULONG;
 564
 565         /* Assert that all items in the pool were freed. */
 566         for (size_t k = 0; k < count; ++k)
 567                 total_leaks += rseq_hweight_ulong(bitmap[k]);
 568         if (total_leaks) {
 569                 fprintf(stderr, "%s: Pool \"%s\" (%p) has %zu leaked items on destroy, caller: %p.\n",
 570                         __func__, get_pool_name(pool), pool, total_leaks, (void *) __builtin_return_address(0));
 571                 abort();
 572         }
 573
 574         free(bitmap);
 575         range->alloc_bitmap = NULL;
 576 }
 577
 578 /* Always inline for __builtin_return_address(0). */
 579 static inline __attribute__((always_inline))
 580 int rseq_mempool_range_destroy(struct rseq_mempool *pool,
 581                 struct rseq_mempool_range *range)
 582 {
 583         destroy_alloc_bitmap(pool, range);
 584
 585         /* range is a header located one page before the aligned mapping. */
 586         return munmap(range->mmap_addr, range->mmap_len);
 587 }
 588
 589 /*
 590  * Allocate a memory mapping aligned on @alignment, with an optional
 591  * @pre_header before the mapping.
 592  */
 593 static
 594 void *aligned_mmap_anonymous(size_t page_size, size_t len, size_t alignment,
 595                 void **pre_header, size_t pre_header_len)
 596 {
 597         size_t minimum_page_count, page_count, extra, total_allocate = 0;
 598         int page_order;
 599         void *ptr;
 600
 601         if (len < page_size || alignment < page_size ||
 602                         !is_pow2(alignment) || (len & (alignment - 1))) {
 603                 errno = EINVAL;
 604                 return NULL;
 605         }
 606         page_order = rseq_get_count_order_ulong(page_size);
 607         if (page_order < 0) {
 608                 errno = EINVAL;
 609                 return NULL;
 610         }
 611         if (pre_header_len && (pre_header_len & (page_size - 1))) {
 612                 errno = EINVAL;
 613                 return NULL;
 614         }
 615
 616         minimum_page_count = (pre_header_len + len) >> page_order;
 617         page_count = (pre_header_len + len + alignment - page_size) >> page_order;
 618
 619         assert(page_count >= minimum_page_count);
 620
 621         ptr = mmap(NULL, page_count << page_order, PROT_READ | PROT_WRITE,
 622                         MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
 623         if (ptr == MAP_FAILED) {
 624                 ptr = NULL;
 625                 goto alloc_error;
 626         }
 627
 628         total_allocate = page_count << page_order;
 629
 630         if (!(((uintptr_t) ptr + pre_header_len) & (alignment - 1))) {
 631                 /* Pointer is already aligned. ptr points to pre_header. */
 632                 goto out;
 633         }
 634
 635         /* Unmap extra before. */
 636         extra = offset_align((uintptr_t) ptr + pre_header_len, alignment);
 637         assert(!(extra & (page_size - 1)));
 638         if (munmap(ptr, extra)) {
 639                 perror("munmap");
 640                 abort();
 641         }
 642         total_allocate -= extra;
 643         ptr += extra;   /* ptr points to pre_header */
 644         page_count -= extra >> page_order;
 645 out:
 646         assert(page_count >= minimum_page_count);
 647
 648         if (page_count > minimum_page_count) {
 649                 void *extra_ptr;
 650
 651                 /* Unmap extra after. */
 652                 extra_ptr = ptr + (minimum_page_count << page_order);
 653                 extra = (page_count - minimum_page_count) << page_order;
 654                 if (munmap(extra_ptr, extra)) {
 655                         perror("munmap");
 656                         abort();
 657                 }
 658                 total_allocate -= extra;
 659         }
 660
 661         assert(!(((uintptr_t)ptr + pre_header_len) & (alignment - 1)));
 662         assert(total_allocate == len + pre_header_len);
 663
 664 alloc_error:
 665         if (ptr) {
 666                 if (pre_header)
 667                         *pre_header = ptr;
 668                 ptr += pre_header_len;
 669         }
 670         return ptr;
 671 }
 672
 673 static
 674 int rseq_memfd_create_init(const char *poolname, size_t init_len)
 675 {
 676         int fd;
 677         char buf[249];          /* Limit is 249 bytes. */
 678         const char *name;
 679
 680         if (poolname) {
 681                 snprintf(buf, sizeof(buf), "%s:rseq-mempool", poolname);
 682                 name = buf;
 683         } else {
 684                 name = "<anonymous>:rseq-mempool";
 685         }
 686
 687         fd = memfd_create(name, MFD_CLOEXEC);
 688         if (fd < 0) {
 689                 perror("memfd_create");
 690                 goto end;
 691         }
 692         if (ftruncate(fd, (off_t) init_len)) {
 693                 if (close(fd))
 694                         perror("close");
 695                 fd = -1;
 696                 goto end;
 697         }
 698 end:
 699         return fd;
 700 }
 701
 702 static
 703 void rseq_memfd_close(int fd)
 704 {
 705         if (fd < 0)
 706                 return;
 707         if (close(fd))
 708                 perror("close");
 709 }
 710
 711 static
 712 struct rseq_mempool_range *rseq_mempool_range_create(struct rseq_mempool *pool)
 713 {
 714         struct rseq_mempool_range *range;
 715         unsigned long page_size;
 716         void *header;
 717         void *base;
 718         size_t range_len;       /* Range len excludes header. */
 719         int memfd = -1;
 720
 721         if (pool->attr.max_nr_ranges &&
 722                         pool->nr_ranges >= pool->attr.max_nr_ranges) {
 723                 errno = ENOMEM;
 724                 return NULL;
 725         }
 726         page_size = rseq_get_page_len();
 727
 728         range_len = pool->attr.stride * pool->attr.max_nr_cpus;
 729         if (pool->attr.populate_policy != RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL)
 730                 range_len += pool->attr.stride; /* init values */
 731         if (pool->attr.robust_set)
 732                 range_len += pool->attr.stride; /* free list */
 733         base = aligned_mmap_anonymous(page_size, range_len,
 734                         pool->attr.stride, &header, page_size);
 735         if (!base)
 736                 return NULL;
 737         range = (struct rseq_mempool_range *) (base - RANGE_HEADER_OFFSET);
 738         range->pool = pool;
 739         range->header = header;
 740         range->base = base;
 741         range->mmap_addr = header;
 742         range->mmap_len = page_size + range_len;
 743
 744         if (pool->attr.populate_policy != RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL) {
 745                 range->init = base + (pool->attr.stride * pool->attr.max_nr_cpus);
 746                 /* Populate init values pages from memfd */
 747                 memfd = rseq_memfd_create_init(pool->name, pool->attr.stride);
 748                 if (memfd < 0)
 749                         goto error_alloc;
 750                 if (mmap(range->init, pool->attr.stride, PROT_READ | PROT_WRITE,
 751                                 MAP_SHARED | MAP_FIXED, memfd, 0) != (void *) range->init) {
 752                         goto error_alloc;
 753                 }
 754                 /*
 755                  * Make sure the init values shared mapping is not
 756                  * shared with the children processes across fork.
 757                  */
 758                 if (madvise(range->init, pool->attr.stride, MADV_DONTFORK))
 759                         goto error_alloc;
 760                 assert(pool->attr.type == MEMPOOL_TYPE_PERCPU);
 761                 /*
 762                  * Map per-cpu memory as private COW mappings of init values.
 763                  */
 764                 {
 765                         int cpu;
 766
 767                         for (cpu = 0; cpu < pool->attr.max_nr_cpus; cpu++) {
 768                                 void *p = base + (pool->attr.stride * cpu);
 769                                 size_t len = pool->attr.stride;
 770
 771                                 if (mmap(p, len, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_FIXED,
 772                                                 memfd, 0) != (void *) p) {
 773                                         goto error_alloc;
 774                                 }
 775                         }
 776                 }
 777                 rseq_memfd_close(memfd);
 778                 memfd = -1;
 779         }
 780
 781         if (pool->attr.robust_set) {
 782                 if (create_alloc_bitmap(pool, range))
 783                         goto error_alloc;
 784         }
 785         if (pool->attr.init_set) {
 786                 switch (pool->attr.type) {
 787                 case MEMPOOL_TYPE_GLOBAL:
 788                         if (pool->attr.init_func(pool->attr.init_priv,
 789                                         base, pool->attr.stride, -1)) {
 790                                 goto error_alloc;
 791                         }
 792                         break;
 793                 case MEMPOOL_TYPE_PERCPU:
 794                 {
 795                         int cpu;
 796                         for (cpu = 0; cpu < pool->attr.max_nr_cpus; cpu++) {
 797                                 if (pool->attr.init_func(pool->attr.init_priv,
 798                                                 base + (pool->attr.stride * cpu),
 799                                                 pool->attr.stride, cpu)) {
 800                                         goto error_alloc;
 801                                 }
 802                         }
 803                         break;
 804                 }
 805                 default:
 806                         abort();
 807                 }
 808         }
 809         pool->nr_ranges++;
 810         return range;
 811
 812 error_alloc:
 813         rseq_memfd_close(memfd);
 814         (void) rseq_mempool_range_destroy(pool, range);
 815         return NULL;
 816 }
 817
 818 int rseq_mempool_destroy(struct rseq_mempool *pool)
 819 {
 820         struct rseq_mempool_range *range, *next_range;
 821         int ret = 0;
 822
 823         if (!pool)
 824                 return 0;
 825         check_free_list(pool);
 826         check_pool_poison(pool);
 827         /* Iteration safe against removal. */
 828         for (range = pool->range_list; range && (next_range = range->next, 1); range = next_range) {
 829                 if (rseq_mempool_range_destroy(pool, range))
 830                         goto end;
 831                 /* Update list head to keep list coherent in case of partial failure. */
 832                 pool->range_list = next_range;
 833         }
 834         pthread_mutex_destroy(&pool->lock);
 835         free(pool->name);
 836         free(pool);
 837 end:
 838         return ret;
 839 }
 840
 841 struct rseq_mempool *rseq_mempool_create(const char *pool_name,
 842                 size_t item_len, const struct rseq_mempool_attr *_attr)
 843 {
 844         struct rseq_mempool *pool;
 845         struct rseq_mempool_attr attr = {};
 846         int order;
 847
 848         /* Make sure each item is large enough to contain free list pointers. */
 849         if (item_len < sizeof(void *))
 850                 item_len = sizeof(void *);
 851
 852         /* Align item_len on next power of two. */
 853         order = rseq_get_count_order_ulong(item_len);
 854         if (order < 0) {
 855                 errno = EINVAL;
 856                 return NULL;
 857         }
 858         item_len = 1UL << order;
 859
 860         if (_attr)
 861                 memcpy(&attr, _attr, sizeof(attr));
 862
 863         switch (attr.type) {
 864         case MEMPOOL_TYPE_PERCPU:
 865                 if (attr.max_nr_cpus < 0) {
 866                         errno = EINVAL;
 867                         return NULL;
 868                 }
 869                 if (attr.max_nr_cpus == 0) {
 870                         /* Auto-detect */
 871                         attr.max_nr_cpus = rseq_get_max_nr_cpus();
 872                         if (attr.max_nr_cpus == 0) {
 873                                 errno = EINVAL;
 874                                 return NULL;
 875                         }
 876                 }
 877                 break;
 878         case MEMPOOL_TYPE_GLOBAL:
 879                 /* Override populate policy for global type. */
 880                 if (attr.populate_policy == RSEQ_MEMPOOL_POPULATE_PRIVATE_NONE)
 881                         attr.populate_policy = RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL;
 882                 /* Use a 1-cpu pool for global mempool type. */
 883                 attr.max_nr_cpus = 1;
 884                 break;
 885         }
 886         if (!attr.stride)
 887                 attr.stride = RSEQ_MEMPOOL_STRIDE;      /* Use default */
 888         if (attr.robust_set && !attr.poison_set) {
 889                 attr.poison_set = true;
 890                 attr.poison = DEFAULT_PRIVATE_POISON_VALUE;
 891         }
 892         if (item_len > attr.stride || attr.stride < (size_t) rseq_get_page_len() ||
 893                         !is_pow2(attr.stride)) {
 894                 errno = EINVAL;
 895                 return NULL;
 896         }
 897
 898         pool = calloc(1, sizeof(struct rseq_mempool));
 899         if (!pool)
 900                 return NULL;
 901
 902         memcpy(&pool->attr, &attr, sizeof(attr));
 903         pthread_mutex_init(&pool->lock, NULL);
 904         pool->item_len = item_len;
 905         pool->item_order = order;
 906
 907         pool->range_list = rseq_mempool_range_create(pool);
 908         if (!pool->range_list)
 909                 goto error_alloc;
 910
 911         if (pool_name) {
 912                 pool->name = strdup(pool_name);
 913                 if (!pool->name)
 914                         goto error_alloc;
 915         }
 916         return pool;
 917
 918 error_alloc:
 919         rseq_mempool_destroy(pool);
 920         errno = ENOMEM;
 921         return NULL;
 922 }
 923
 924 /* Always inline for __builtin_return_address(0). */
 925 static inline __attribute__((always_inline))
 926 void set_alloc_slot(struct rseq_mempool *pool, struct rseq_mempool_range *range, size_t item_offset)
 927 {
 928         unsigned long *bitmap = range->alloc_bitmap;
 929         size_t item_index = item_offset >> pool->item_order;
 930         unsigned long mask;
 931         size_t k;
 932
 933         if (!bitmap)
 934                 return;
 935
 936         k = item_index / BIT_PER_ULONG;
 937         mask = 1ULL << (item_index % BIT_PER_ULONG);
 938
 939         /* Print error if bit is already set. */
 940         if (bitmap[k] & mask) {
 941                 fprintf(stderr, "%s: Allocator corruption detected for pool: \"%s\" (%p), item offset: %zu, caller: %p.\n",
 942                         __func__, get_pool_name(pool), pool, item_offset, (void *) __builtin_return_address(0));
 943                 abort();
 944         }
 945         bitmap[k] |= mask;
 946 }
 947
 948 static
 949 void __rseq_percpu *__rseq_percpu_malloc(struct rseq_mempool *pool,
 950                 bool zeroed, void *init_ptr, size_t init_len)
 951 {
 952         struct rseq_mempool_range *range;
 953         struct free_list_node *node;
 954         uintptr_t item_offset;
 955         void __rseq_percpu *addr;
 956
 957         if (init_len > pool->item_len) {
 958                 errno = EINVAL;
 959                 return NULL;
 960         }
 961         pthread_mutex_lock(&pool->lock);
 962         /* Get first entry from free list. */
 963         node = pool->free_list_head;
 964         if (node != NULL) {
 965                 void *range_base, *ptr;
 966
 967                 ptr = __rseq_free_list_to_percpu_ptr(pool, node);
 968                 range_base = (void *) ((uintptr_t) ptr & (~(pool->attr.stride - 1)));
 969                 range = (struct rseq_mempool_range *) (range_base - RANGE_HEADER_OFFSET);
 970                 /* Remove node from free list (update head). */
 971                 pool->free_list_head = node->next;
 972                 item_offset = (uintptr_t) (ptr - range_base);
 973                 rseq_percpu_check_poison_item(pool, range, item_offset);
 974                 addr = __rseq_free_list_to_percpu_ptr(pool, node);
 975                 goto end;
 976         }
 977         /*
 978          * If the most recent range (first in list) does not have any
 979          * room left, create a new range and prepend it to the list
 980          * head.
 981          */
 982         range = pool->range_list;
 983         if (range->next_unused + pool->item_len > pool->attr.stride) {
 984                 range = rseq_mempool_range_create(pool);
 985                 if (!range) {
 986                         errno = ENOMEM;
 987                         addr = NULL;
 988                         goto end;
 989                 }
 990                 /* Add range to head of list. */
 991                 range->next = pool->range_list;
 992                 pool->range_list = range;
 993         }
 994         /* First range in list has room left. */
 995         item_offset = range->next_unused;
 996         addr = (void __rseq_percpu *) (range->base + item_offset);
 997         range->next_unused += pool->item_len;
 998 end:
 999         if (addr)
1000                 set_alloc_slot(pool, range, item_offset);
1001         pthread_mutex_unlock(&pool->lock);
1002         if (addr) {
1003                 if (zeroed)
1004                         rseq_percpu_zero_item(pool, range, item_offset);
1005                 else if (init_ptr) {
1006                         rseq_percpu_init_item(pool, range, item_offset,
1007                                         init_ptr, init_len);
1008                 }
1009         }
1010         return addr;
1011 }
1012
1013 void __rseq_percpu *rseq_mempool_percpu_malloc(struct rseq_mempool *pool)
1014 {
1015         return __rseq_percpu_malloc(pool, false, NULL, 0);
1016 }
1017
1018 void __rseq_percpu *rseq_mempool_percpu_zmalloc(struct rseq_mempool *pool)
1019 {
1020         return __rseq_percpu_malloc(pool, true, NULL, 0);
1021 }
1022
1023 void __rseq_percpu *rseq_mempool_percpu_malloc_init(struct rseq_mempool *pool,
1024                 void *init_ptr, size_t len)
1025 {
1026         return __rseq_percpu_malloc(pool, false, init_ptr, len);
1027 }
1028
1029 /* Always inline for __builtin_return_address(0). */
1030 static inline __attribute__((always_inline))
1031 void clear_alloc_slot(struct rseq_mempool *pool, struct rseq_mempool_range *range, size_t item_offset)
1032 {
1033         unsigned long *bitmap = range->alloc_bitmap;
1034         size_t item_index = item_offset >> pool->item_order;
1035         unsigned long mask;
1036         size_t k;
1037
1038         if (!bitmap)
1039                 return;
1040
1041         k = item_index / BIT_PER_ULONG;
1042         mask = 1ULL << (item_index % BIT_PER_ULONG);
1043
1044         /* Print error if bit is not set. */
1045         if (!(bitmap[k] & mask)) {
1046                 fprintf(stderr, "%s: Double-free detected for pool: \"%s\" (%p), item offset: %zu, caller: %p.\n",
1047                         __func__, get_pool_name(pool), pool, item_offset,
1048                         (void *) __builtin_return_address(0));
1049                 abort();
1050         }
1051         bitmap[k] &= ~mask;
1052 }
1053
1054 void librseq_mempool_percpu_free(void __rseq_percpu *_ptr, size_t stride)
1055 {
1056         uintptr_t ptr = (uintptr_t) _ptr;
1057         void *range_base = (void *) (ptr & (~(stride - 1)));
1058         struct rseq_mempool_range *range = (struct rseq_mempool_range *) (range_base - RANGE_HEADER_OFFSET);
1059         struct rseq_mempool *pool = range->pool;
1060         uintptr_t item_offset = ptr & (stride - 1);
1061         struct free_list_node *head, *item;
1062
1063         pthread_mutex_lock(&pool->lock);
1064         clear_alloc_slot(pool, range, item_offset);
1065         /* Add ptr to head of free list */
1066         head = pool->free_list_head;
1067         if (pool->attr.poison_set)
1068                 rseq_percpu_poison_item(pool, range, item_offset);
1069         item = __rseq_percpu_to_free_list_ptr(pool, _ptr);
1070         /*
1071          * Setting the next pointer will overwrite the first uintptr_t
1072          * poison for either CPU 0 (populate all) or init data (populate
1073          * none).
1074          */
1075         item->next = head;
1076         pool->free_list_head = item;
1077         pthread_mutex_unlock(&pool->lock);
1078 }
1079
1080 struct rseq_mempool_set *rseq_mempool_set_create(void)
1081 {
1082         struct rseq_mempool_set *pool_set;
1083
1084         pool_set = calloc(1, sizeof(struct rseq_mempool_set));
1085         if (!pool_set)
1086                 return NULL;
1087         pthread_mutex_init(&pool_set->lock, NULL);
1088         return pool_set;
1089 }
1090
1091 int rseq_mempool_set_destroy(struct rseq_mempool_set *pool_set)
1092 {
1093         int order, ret;
1094
1095         for (order = POOL_SET_MIN_ENTRY; order < POOL_SET_NR_ENTRIES; order++) {
1096                 struct rseq_mempool *pool = pool_set->entries[order];
1097
1098                 if (!pool)
1099                         continue;
1100                 ret = rseq_mempool_destroy(pool);
1101                 if (ret)
1102                         return ret;
1103                 pool_set->entries[order] = NULL;
1104         }
1105         pthread_mutex_destroy(&pool_set->lock);
1106         free(pool_set);
1107         return 0;
1108 }
1109
1110 /* Ownership of pool is handed over to pool set on success. */
1111 int rseq_mempool_set_add_pool(struct rseq_mempool_set *pool_set, struct rseq_mempool *pool)
1112 {
1113         size_t item_order = pool->item_order;
1114         int ret = 0;
1115
1116         pthread_mutex_lock(&pool_set->lock);
1117         if (pool_set->entries[item_order]) {
1118                 errno = EBUSY;
1119                 ret = -1;
1120                 goto end;
1121         }
1122         pool_set->entries[pool->item_order] = pool;
1123 end:
1124         pthread_mutex_unlock(&pool_set->lock);
1125         return ret;
1126 }
1127
1128 static
1129 void __rseq_percpu *__rseq_mempool_set_malloc(struct rseq_mempool_set *pool_set,
1130                 void *init_ptr, size_t len, bool zeroed)
1131 {
1132         int order, min_order = POOL_SET_MIN_ENTRY;
1133         struct rseq_mempool *pool;
1134         void __rseq_percpu *addr;
1135
1136         order = rseq_get_count_order_ulong(len);
1137         if (order > POOL_SET_MIN_ENTRY)
1138                 min_order = order;
1139 again:
1140         pthread_mutex_lock(&pool_set->lock);
1141         /* First smallest present pool where @len fits. */
1142         for (order = min_order; order < POOL_SET_NR_ENTRIES; order++) {
1143                 pool = pool_set->entries[order];
1144
1145                 if (!pool)
1146                         continue;
1147                 if (pool->item_len >= len)
1148                         goto found;
1149         }
1150         pool = NULL;
1151 found:
1152         pthread_mutex_unlock(&pool_set->lock);
1153         if (pool) {
1154                 addr = __rseq_percpu_malloc(pool, zeroed, init_ptr, len);
1155                 if (addr == NULL && errno == ENOMEM) {
1156                         /*
1157                          * If the allocation failed, try again with a
1158                          * larger pool.
1159                          */
1160                         min_order = order + 1;
1161                         goto again;
1162                 }
1163         } else {
1164                 /* Not found. */
1165                 errno = ENOMEM;
1166                 addr = NULL;
1167         }
1168         return addr;
1169 }
1170
1171 void __rseq_percpu *rseq_mempool_set_percpu_malloc(struct rseq_mempool_set *pool_set, size_t len)
1172 {
1173         return __rseq_mempool_set_malloc(pool_set, NULL, len, false);
1174 }
1175
1176 void __rseq_percpu *rseq_mempool_set_percpu_zmalloc(struct rseq_mempool_set *pool_set, size_t len)
1177 {
1178         return __rseq_mempool_set_malloc(pool_set, NULL, len, true);
1179 }
1180
1181 void __rseq_percpu *rseq_mempool_set_percpu_malloc_init(struct rseq_mempool_set *pool_set,
1182                 void *init_ptr, size_t len)
1183 {
1184         return __rseq_mempool_set_malloc(pool_set, init_ptr, len, true);
1185 }
1186
1187 struct rseq_mempool_attr *rseq_mempool_attr_create(void)
1188 {
1189         return calloc(1, sizeof(struct rseq_mempool_attr));
1190 }
1191
1192 void rseq_mempool_attr_destroy(struct rseq_mempool_attr *attr)
1193 {
1194         free(attr);
1195 }
1196
1197 int rseq_mempool_attr_set_init(struct rseq_mempool_attr *attr,
1198                 int (*init_func)(void *priv, void *addr, size_t len, int cpu),
1199                 void *init_priv)
1200 {
1201         if (!attr) {
1202                 errno = EINVAL;
1203                 return -1;
1204         }
1205         attr->init_set = true;
1206         attr->init_func = init_func;
1207         attr->init_priv = init_priv;
1208         return 0;
1209 }
1210
1211 int rseq_mempool_attr_set_robust(struct rseq_mempool_attr *attr)
1212 {
1213         if (!attr) {
1214                 errno = EINVAL;
1215                 return -1;
1216         }
1217         attr->robust_set = true;
1218         return 0;
1219 }
1220
1221 int rseq_mempool_attr_set_percpu(struct rseq_mempool_attr *attr,
1222                 size_t stride, int max_nr_cpus)
1223 {
1224         if (!attr) {
1225                 errno = EINVAL;
1226                 return -1;
1227         }
1228         attr->type = MEMPOOL_TYPE_PERCPU;
1229         attr->stride = stride;
1230         attr->max_nr_cpus = max_nr_cpus;
1231         return 0;
1232 }
1233
1234 int rseq_mempool_attr_set_global(struct rseq_mempool_attr *attr,
1235                 size_t stride)
1236 {
1237         if (!attr) {
1238                 errno = EINVAL;
1239                 return -1;
1240         }
1241         attr->type = MEMPOOL_TYPE_GLOBAL;
1242         attr->stride = stride;
1243         attr->max_nr_cpus = 0;
1244         return 0;
1245 }
1246
1247 int rseq_mempool_attr_set_max_nr_ranges(struct rseq_mempool_attr *attr,
1248                 unsigned long max_nr_ranges)
1249 {
1250         if (!attr) {
1251                 errno = EINVAL;
1252                 return -1;
1253         }
1254         attr->max_nr_ranges = max_nr_ranges;
1255         return 0;
1256 }
1257
1258 int rseq_mempool_attr_set_poison(struct rseq_mempool_attr *attr,
1259                 uintptr_t poison)
1260 {
1261         if (!attr) {
1262                 errno = EINVAL;
1263                 return -1;
1264         }
1265         attr->poison_set = true;
1266         attr->poison = poison;
1267         return 0;
1268 }
1269
1270 int rseq_mempool_attr_set_populate_policy(struct rseq_mempool_attr *attr,
1271                 enum rseq_mempool_populate_policy policy)
1272 {
1273         if (!attr) {
1274                 errno = EINVAL;
1275                 return -1;
1276         }
1277         attr->populate_policy = policy;
1278         return 0;
1279 }
1280
1281 int rseq_mempool_get_max_nr_cpus(struct rseq_mempool *mempool)
1282 {
1283         if (!mempool || mempool->attr.type != MEMPOOL_TYPE_PERCPU) {
1284                 errno = EINVAL;
1285                 return -1;
1286         }
1287         return mempool->attr.max_nr_cpus;
1288 }