include/rseq/mempool.h

   1 /* SPDX-License-Identifier: MIT */
   2 /* SPDX-FileCopyrightText: 2024 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> */
   3
   4 #ifndef _RSEQ_MEMPOOL_H
   5 #define _RSEQ_MEMPOOL_H
   6
   7 #include <stddef.h>
   8 #include <sys/types.h>
   9 #include <sys/mman.h>
  10
  11 /*
  12  * rseq/mempool.h: rseq CPU-Local Storage (CLS) memory allocator.
  13  *
  14  * The rseq per-CPU memory allocator allows the application the request
  15  * memory pools of CPU-Local memory each of containing objects of a
  16  * given size (rounded to next power of 2), reserving a given virtual
  17  * address size per CPU, for a given maximum number of CPUs.
  18  *
  19  * The per-CPU memory allocator is analogous to TLS (Thread-Local
  20  * Storage) memory: TLS is Thread-Local Storage, whereas the per-CPU
  21  * memory allocator provides CPU-Local Storage.
  22  */
  23
  24 #ifdef __cplusplus
  25 extern "C" {
  26 #endif
  27
  28 /*
  29  * Tag pointers returned by:
  30  * - rseq_percpu_malloc(),
  31  * - rseq_percpu_zmalloc(),
  32  * - rseq_percpu_pool_set_malloc(),
  33  * - rseq_percpu_pool_set_zmalloc().
  34  *
  35  * and passed as parameter to:
  36  * - rseq_percpu_ptr(),
  37  * - rseq_percpu_free().
  38  *
  39  * with __rseq_percpu for use by static analyzers.
  40  */
  41 #define __rseq_percpu
  42
  43 struct rseq_pool_attr;
  44 struct rseq_percpu_pool;
  45
  46 /*
  47  * rseq_percpu_pool_create: Create a per-cpu memory pool.
  48  *
  49  * Create a per-cpu memory pool for items of size @item_len (rounded to
  50  * next power of two). The reserved allocation size is @percpu_len, and
  51  * the maximum CPU value expected is (@max_nr_cpus - 1).
  52  *
  53  * The @attr pointer used to specify the pool attributes. If NULL, use a
  54  * default attribute values. The @attr can be destroyed immediately
  55  * after rseq_percpu_pool_create() returns. The caller keeps ownership
  56  * of @attr.
  57  *
  58  * The argument @pool_name can be used to given a name to the pool for
  59  * debugging purposes. It can be NULL if no name is given.
  60  *
  61  * Returns a pointer to the created percpu pool. Return NULL on error,
  62  * with errno set accordingly:
  63  *   EINVAL: Invalid argument.
  64  *   ENOMEM: Not enough resources (memory or pool indexes) available to
  65  *           allocate pool.
  66  *
  67  * In addition, if the attr mmap callback fails, NULL is returned and
  68  * errno is propagated from the callback. The default callback can
  69  * return errno=ENOMEM.
  70  *
  71  * This API is MT-safe.
  72  */
  73 struct rseq_percpu_pool *rseq_percpu_pool_create(const char *pool_name,
  74                 size_t item_len, size_t percpu_len, int max_nr_cpus,
  75                 const struct rseq_pool_attr *attr);
  76
  77 /*
  78  * rseq_percpu_pool_destroy: Destroy a per-cpu memory pool.
  79  *
  80  * Destroy a per-cpu memory pool, unmapping its memory and removing the
  81  * pool entry from the global index. No pointers allocated from the
  82  * pool should be used when it is destroyed. This includes rseq_percpu_ptr().
  83  *
  84  * Argument @pool is a pointer to the per-cpu pool to destroy.
  85  *
  86  * Return values: 0 on success, -1 on error, with errno set accordingly:
  87  *   ENOENT: Trying to free a pool which was not allocated.
  88  *
  89  * If the munmap_func callback fails, -1 is returned and errno is
  90  * propagated from the callback. The default callback can return
  91  * errno=EINVAL.
  92  *
  93  * This API is MT-safe.
  94  */
  95 int rseq_percpu_pool_destroy(struct rseq_percpu_pool *pool);
  96
  97 /*
  98  * rseq_percpu_malloc: Allocate memory from a per-cpu pool.
  99  *
 100  * Allocate an item from a per-cpu @pool. The allocation will reserve
 101  * an item of the size specified by @item_len (rounded to next power of
 102  * two) at pool creation. This effectively reserves space for this item
 103  * on all CPUs.
 104  *
 105  * On success, return a "__rseq_percpu" encoded pointer to the pool
 106  * item. This encoded pointer is meant to be passed to rseq_percpu_ptr()
 107  * to be decoded to a valid address before being accessed.
 108  *
 109  * Return NULL (errno=ENOMEM) if there is not enough space left in the
 110  * pool to allocate an item.
 111  *
 112  * This API is MT-safe.
 113  */
 114 void __rseq_percpu *rseq_percpu_malloc(struct rseq_percpu_pool *pool);
 115
 116 /*
 117  * rseq_percpu_zmalloc: Allocated zero-initialized memory from a per-cpu pool.
 118  *
 119  * Allocate memory for an item within the pool, and zero-initialize its
 120  * memory on all CPUs. See rseq_percpu_malloc for details.
 121  *
 122  * This API is MT-safe.
 123  */
 124 void __rseq_percpu *rseq_percpu_zmalloc(struct rseq_percpu_pool *pool);
 125
 126 /*
 127  * rseq_percpu_free: Free memory from a per-cpu pool.
 128  *
 129  * Free an item pointed to by @ptr from its per-cpu pool.
 130  *
 131  * The @ptr argument is a __rseq_percpu encoded pointer returned by
 132  * either:
 133  *
 134  * - rseq_percpu_malloc(),
 135  * - rseq_percpu_zmalloc(),
 136  * - rseq_percpu_pool_set_malloc(),
 137  * - rseq_percpu_pool_set_zmalloc().
 138  *
 139  * This API is MT-safe.
 140  */
 141 void rseq_percpu_free(void __rseq_percpu *ptr);
 142
 143 /*
 144  * rseq_percpu_ptr: Decode a per-cpu pointer.
 145  *
 146  * Decode a per-cpu pointer @ptr to get the associated pointer for the
 147  * given @cpu. The @ptr argument is a __rseq_percpu encoded pointer
 148  * returned by either:
 149  *
 150  * - rseq_percpu_malloc(),
 151  * - rseq_percpu_zmalloc(),
 152  * - rseq_percpu_pool_set_malloc(),
 153  * - rseq_percpu_pool_set_zmalloc().
 154  *
 155  * The __rseq_percpu pointer can be decoded with rseq_percpu_ptr() even
 156  * after it has been freed, as long as its associated pool has not been
 157  * destroyed. However, memory pointed to by the decoded pointer should
 158  * not be accessed after the __rseq_percpu pointer has been freed.
 159  *
 160  * The macro rseq_percpu_ptr() preserves the type of the @ptr parameter
 161  * for the returned pointer, but removes the __rseq_percpu annotation.
 162  *
 163  * This API is MT-safe.
 164  */
 165 void *__rseq_percpu_ptr(void __rseq_percpu *ptr, int cpu);
 166 #define rseq_percpu_ptr(ptr, cpu)       ((__typeof__(*(ptr)) *) __rseq_percpu_ptr(ptr, cpu))
 167
 168 /*
 169  * rseq_percpu_pool_set_create: Create a pool set.
 170  *
 171  * Create a set of pools. Its purpose is to offer a memory allocator API
 172  * for variable-length items (e.g. variable length strings). When
 173  * created, the pool set has no pool. Pools can be created and added to
 174  * the set. One common approach would be to create pools for each
 175  * relevant power of two allocation size useful for the application.
 176  * Only one pool can be added to the pool set for each power of two
 177  * allocation size.
 178  *
 179  * Returns a pool set pointer on success, else returns NULL with
 180  * errno=ENOMEM (out of memory).
 181  *
 182  * This API is MT-safe.
 183  */
 184 struct rseq_percpu_pool_set *rseq_percpu_pool_set_create(void);
 185
 186 /*
 187  * rseq_percpu_pool_set_destroy: Destroy a pool set.
 188  *
 189  * Destroy a pool set and its associated resources. The pools that were
 190  * added to the pool set are destroyed as well.
 191  *
 192  * Returns 0 on success, -1 on failure (or partial failure), with errno
 193  * set by rseq_percpu_pool_destroy(). Using a pool set after destroy
 194  * failure is undefined.
 195  *
 196  * This API is MT-safe.
 197  */
 198 int rseq_percpu_pool_set_destroy(struct rseq_percpu_pool_set *pool_set);
 199
 200 /*
 201  * rseq_percpu_pool_set_add_pool: Add a pool to a pool set.
 202  *
 203  * Add a @pool to the @pool_set. On success, its ownership is handed
 204  * over to the pool set, so the caller should not destroy it explicitly.
 205  * Only one pool can be added to the pool set for each power of two
 206  * allocation size.
 207  *
 208  * Returns 0 on success, -1 on error with the following errno:
 209  * - EBUSY: A pool already exists in the pool set for this power of two
 210  *          allocation size.
 211  *
 212  * This API is MT-safe.
 213  */
 214 int rseq_percpu_pool_set_add_pool(struct rseq_percpu_pool_set *pool_set,
 215                 struct rseq_percpu_pool *pool);
 216
 217 /*
 218  * rseq_percpu_pool_set_malloc: Allocate memory from a per-cpu pool set.
 219  *
 220  * Allocate an item from a per-cpu @pool. The allocation will reserve
 221  * an item of the size specified by @len (rounded to next power of
 222  * two). This effectively reserves space for this item on all CPUs.
 223  *
 224  * The space reservation will search for the smallest pool within
 225  * @pool_set which respects the following conditions:
 226  *
 227  * - it has an item size large enough to fit @len,
 228  * - it has space available.
 229  *
 230  * On success, return a "__rseq_percpu" encoded pointer to the pool
 231  * item. This encoded pointer is meant to be passed to rseq_percpu_ptr()
 232  * to be decoded to a valid address before being accessed.
 233  *
 234  * Return NULL (errno=ENOMEM) if there is not enough space left in the
 235  * pool to allocate an item.
 236  *
 237  * This API is MT-safe.
 238  */
 239 void __rseq_percpu *rseq_percpu_pool_set_malloc(struct rseq_percpu_pool_set *pool_set, size_t len);
 240
 241 /*
 242  * rseq_percpu_pool_set_zmalloc: Allocated zero-initialized memory from a per-cpu pool set.
 243  *
 244  * Allocate memory for an item within the pool, and zero-initialize its
 245  * memory on all CPUs. See rseq_percpu_pool_set_malloc for details.
 246  *
 247  * This API is MT-safe.
 248  */
 249 void __rseq_percpu *rseq_percpu_pool_set_zmalloc(struct rseq_percpu_pool_set *pool_set, size_t len);
 250
 251 /*
 252  * rseq_percpu_pool_init_numa: Move pages to the NUMA node associated to their CPU topology.
 253  *
 254  * For pages allocated within @pool, invoke move_pages(2) with the given
 255  * @numa_flags to move the pages to the NUMA node associated to their
 256  * CPU topology.
 257  *
 258  * Argument @numa_flags are passed to move_pages(2). The expected flags are:
 259  *   MPOL_MF_MOVE:     move process-private pages to cpu-specific numa nodes.
 260  *   MPOL_MF_MOVE_ALL: move shared pages to cpu-specific numa nodes
 261  *                     (requires CAP_SYS_NICE).
 262  *
 263  * Returns 0 on success, else return -1 with errno set by move_pages(2).
 264  */
 265 int rseq_percpu_pool_init_numa(struct rseq_percpu_pool *pool, int numa_flags);
 266
 267 /*
 268  * rseq_pool_attr_create: Create a pool attribute structure.
 269  */
 270 struct rseq_pool_attr *rseq_pool_attr_create(void);
 271
 272 /*
 273  * rseq_pool_attr_destroy: Destroy a pool attribute structure.
 274  */
 275 void rseq_pool_attr_destroy(struct rseq_pool_attr *attr);
 276
 277 /*
 278  * rseq_pool_attr_set_mmap: Set pool attribute structure mmap functions.
 279  *
 280  * The @mmap_func callback used to map the memory for the pool.
 281  *
 282  * The @munmap_func callback used to unmap the memory when the pool
 283  * is destroyed.
 284  *
 285  * The @mmap_priv argument is a private data pointer passed to both
 286  * @mmap_func and @munmap_func callbacks.
 287  *
 288  * Returns 0 on success, -1 with errno=EINVAL if arguments are invalid.
 289  */
 290 int rseq_pool_attr_set_mmap(struct rseq_pool_attr *attr,
 291                 void *(*mmap_func)(void *priv, size_t len),
 292                 int (*munmap_func)(void *priv, void *ptr, size_t len),
 293                 void *mmap_priv);
 294
 295 /*
 296  * rseq_pool_attr_set_robust: Set pool robust attribute.
 297  *
 298  * The robust pool attribute enables runtime validation of the pool:
 299  *
 300  *   - Check for double-free of pointers.
 301  *
 302  *   - Detect memory leaks on pool destruction.
 303   *
 304  *   - Detect free-list corruption on pool destruction.
 305  *
 306  * There is a marginal runtime overhead on malloc/free operations.
 307  *
 308  * The memory overhead is (pool->percpu_len / pool->item_len) / CHAR_BIT
 309  * bytes, over the lifetime of the pool.
 310  *
 311  * Returns 0 on success, -1 with errno=EINVAL if arguments are invalid.
 312  */
 313 int rseq_pool_attr_set_robust(struct rseq_pool_attr *attr);
 314
 315 #ifdef __cplusplus
 316 }
 317 #endif
 318
 319 #endif /* _RSEQ_MEMPOOL_H */