Commit | Line | Data |
---|---|---|
ef6695f1 MD |
1 | // SPDX-License-Identifier: MIT |
2 | // SPDX-FileCopyrightText: 2024 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
3 | ||
34337fec | 4 | #include <rseq/mempool.h> |
ef6695f1 MD |
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> | |
367e559c | 15 | #include <stdio.h> |
a5694a4d | 16 | #include <fcntl.h> |
367e559c MD |
17 | |
18 | #ifdef HAVE_LIBNUMA | |
19 | # include <numa.h> | |
20 | # include <numaif.h> | |
21 | #endif | |
ef6695f1 | 22 | |
34337fec | 23 | #include "rseq-utils.h" |
47c725dd | 24 | #include <rseq/rseq.h> |
19be9217 | 25 | |
ef6695f1 | 26 | /* |
b73b0c25 | 27 | * rseq-mempool.c: rseq CPU-Local Storage (CLS) memory allocator. |
ef6695f1 | 28 | * |
8ab16a24 MD |
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 | |
8aa1462d MD |
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. | |
8ab16a24 MD |
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. | |
ef6695f1 MD |
37 | */ |
38 | ||
3236da62 | 39 | #define POOL_SET_NR_ENTRIES RSEQ_BITS_PER_LONG |
ef6695f1 | 40 | |
72b100a1 MD |
41 | /* |
42 | * Smallest allocation should hold enough space for a free list pointer. | |
43 | */ | |
ef6695f1 MD |
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 | ||
0fdf7a4c OD |
50 | #define BIT_PER_ULONG (8 * sizeof(unsigned long)) |
51 | ||
57d8b586 OD |
52 | #define MOVE_PAGES_BATCH_SIZE 4096 |
53 | ||
0ba2a93e | 54 | #define RANGE_HEADER_OFFSET sizeof(struct rseq_mempool_range) |
4aa3220c | 55 | |
3975084e | 56 | #if RSEQ_BITS_PER_LONG == 64 |
dbccd436 | 57 | # define DEFAULT_PRIVATE_POISON_VALUE 0x5555555555555555ULL |
3975084e | 58 | #else |
dbccd436 | 59 | # define DEFAULT_PRIVATE_POISON_VALUE 0x55555555UL |
3975084e MD |
60 | #endif |
61 | ||
ef6695f1 MD |
62 | struct free_list_node; |
63 | ||
64 | struct free_list_node { | |
65 | struct free_list_node *next; | |
66 | }; | |
67 | ||
cb475906 | 68 | enum mempool_type { |
89b7e681 MD |
69 | MEMPOOL_TYPE_GLOBAL = 0, /* Default */ |
70 | MEMPOOL_TYPE_PERCPU = 1, | |
cb475906 MD |
71 | }; |
72 | ||
0ba2a93e | 73 | struct rseq_mempool_attr { |
135811f2 | 74 | bool init_set; |
6e329183 | 75 | int (*init_func)(void *priv, void *addr, size_t len, int cpu); |
135811f2 MD |
76 | void *init_priv; |
77 | ||
d6acc8aa | 78 | bool robust_set; |
cb475906 MD |
79 | |
80 | enum mempool_type type; | |
81 | size_t stride; | |
82 | int max_nr_cpus; | |
e11a02d7 MD |
83 | |
84 | unsigned long max_nr_ranges; | |
455e090e MD |
85 | |
86 | bool poison_set; | |
87 | uintptr_t poison; | |
a5694a4d MD |
88 | |
89 | enum rseq_mempool_populate_policy populate_policy; | |
9bd07c29 MD |
90 | }; |
91 | ||
0ba2a93e | 92 | struct rseq_mempool_range; |
b73b0c25 | 93 | |
0ba2a93e | 94 | struct rseq_mempool_range { |
9d986353 MD |
95 | struct rseq_mempool_range *next; /* Linked list of ranges. */ |
96 | struct rseq_mempool *pool; /* Backward reference to container pool. */ | |
a5694a4d MD |
97 | |
98 | /* | |
99 | * Memory layout of a mempool range: | |
100 | * - Header page (contains struct rseq_mempool_range at the very end), | |
c0de0012 MD |
101 | * - Base of the per-cpu data, starting with CPU 0. |
102 | * Aliases with free-list for non-robust populate all pool. | |
a5694a4d MD |
103 | * - CPU 1, |
104 | * ... | |
105 | * - CPU max_nr_cpus - 1 | |
4e8ae59d | 106 | * - init values (unpopulated for RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL). |
c0de0012 MD |
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. | |
a5694a4d | 114 | */ |
4aa3220c | 115 | void *header; |
ef6695f1 | 116 | void *base; |
a5694a4d MD |
117 | /* |
118 | * The init values contains malloc_init/zmalloc values. | |
4e8ae59d | 119 | * Pointer is NULL for RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL. |
a5694a4d MD |
120 | */ |
121 | void *init; | |
b73b0c25 | 122 | size_t next_unused; |
fa6a0fb3 MD |
123 | |
124 | /* Pool range mmap/munmap */ | |
125 | void *mmap_addr; | |
126 | size_t mmap_len; | |
127 | ||
b73b0c25 MD |
128 | /* Track alloc/free. */ |
129 | unsigned long *alloc_bitmap; | |
130 | }; | |
131 | ||
0ba2a93e | 132 | struct rseq_mempool { |
9d986353 MD |
133 | /* Head of ranges linked-list. */ |
134 | struct rseq_mempool_range *range_list; | |
135 | unsigned long nr_ranges; | |
b73b0c25 | 136 | |
ef6695f1 | 137 | size_t item_len; |
ef6695f1 | 138 | int item_order; |
ef6695f1 MD |
139 | |
140 | /* | |
8ab16a24 | 141 | * The free list chains freed items on the CPU 0 address range. |
ef6695f1 | 142 | * We should rethink this decision if false sharing between |
8ab16a24 | 143 | * malloc/free from other CPUs and data accesses from CPU 0 |
ef6695f1 MD |
144 | * becomes an issue. This is a NULL-terminated singly-linked |
145 | * list. | |
146 | */ | |
147 | struct free_list_node *free_list_head; | |
b73b0c25 | 148 | |
ef6695f1 MD |
149 | /* This lock protects allocation/free within the pool. */ |
150 | pthread_mutex_t lock; | |
9bd07c29 | 151 | |
0ba2a93e | 152 | struct rseq_mempool_attr attr; |
ca452fee | 153 | char *name; |
ef6695f1 MD |
154 | }; |
155 | ||
ef6695f1 MD |
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 | */ | |
0ba2a93e | 161 | struct rseq_mempool_set { |
ef6695f1 MD |
162 | /* This lock protects add vs malloc/zmalloc within the pool set. */ |
163 | pthread_mutex_t lock; | |
0ba2a93e | 164 | struct rseq_mempool *entries[POOL_SET_NR_ENTRIES]; |
ef6695f1 MD |
165 | }; |
166 | ||
86617384 MD |
167 | static |
168 | const char *get_pool_name(const struct rseq_mempool *pool) | |
169 | { | |
170 | return pool->name ? : "<anonymous>"; | |
171 | } | |
172 | ||
367e559c | 173 | static |
6fbf1fb6 | 174 | void *__rseq_pool_range_percpu_ptr(const struct rseq_mempool_range *range, int cpu, |
f2981623 | 175 | uintptr_t item_offset, size_t stride) |
367e559c | 176 | { |
15b63c9f | 177 | return range->base + (stride * cpu) + item_offset; |
367e559c MD |
178 | } |
179 | ||
a5694a4d MD |
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 | ||
c0de0012 MD |
195 | if (pool->attr.robust_set) { |
196 | /* Skip cpus. */ | |
a5694a4d | 197 | p -= pool->attr.max_nr_cpus * pool->attr.stride; |
c0de0012 | 198 | /* Skip init values */ |
4e8ae59d | 199 | if (pool->attr.populate_policy != RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL) |
c0de0012 MD |
200 | p -= pool->attr.stride; |
201 | ||
202 | } else { | |
203 | /* Populate none free list is in init values */ | |
4e8ae59d | 204 | if (pool->attr.populate_policy != RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL) |
c0de0012 MD |
205 | p -= pool->attr.max_nr_cpus * pool->attr.stride; |
206 | } | |
a5694a4d MD |
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 | { | |
c0de0012 MD |
214 | if (pool->attr.robust_set) { |
215 | /* Skip cpus. */ | |
a5694a4d | 216 | p += pool->attr.max_nr_cpus * pool->attr.stride; |
c0de0012 | 217 | /* Skip init values */ |
4e8ae59d | 218 | if (pool->attr.populate_policy != RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL) |
c0de0012 MD |
219 | p += pool->attr.stride; |
220 | ||
221 | } else { | |
222 | /* Populate none free list is in init values */ | |
4e8ae59d | 223 | if (pool->attr.populate_policy != RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL) |
c0de0012 MD |
224 | p += pool->attr.max_nr_cpus * pool->attr.stride; |
225 | } | |
a5694a4d MD |
226 | return (struct free_list_node *) p; |
227 | } | |
228 | ||
229 | static | |
14af0aa2 | 230 | intptr_t rseq_cmp_item(void *p, size_t item_len, intptr_t cmp_value, intptr_t *unexpected_value) |
a5694a4d | 231 | { |
14af0aa2 MD |
232 | size_t offset; |
233 | intptr_t res = 0; | |
a5694a4d | 234 | |
14af0aa2 MD |
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; | |
a5694a4d | 241 | break; |
14af0aa2 MD |
242 | } |
243 | } | |
a5694a4d MD |
244 | return res; |
245 | } | |
246 | ||
367e559c | 247 | static |
15b63c9f MD |
248 | void rseq_percpu_zero_item(struct rseq_mempool *pool, |
249 | struct rseq_mempool_range *range, uintptr_t item_offset) | |
367e559c | 250 | { |
a5694a4d | 251 | char *init_p = NULL; |
367e559c MD |
252 | int i; |
253 | ||
a5694a4d MD |
254 | init_p = __rseq_pool_range_init_ptr(range, item_offset); |
255 | if (init_p) | |
644298bb | 256 | bzero(init_p, pool->item_len); |
cb475906 | 257 | for (i = 0; i < pool->attr.max_nr_cpus; i++) { |
15b63c9f | 258 | char *p = __rseq_pool_range_percpu_ptr(range, i, |
cb475906 | 259 | item_offset, pool->attr.stride); |
a5694a4d | 260 | |
1b658191 MD |
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 | */ | |
14af0aa2 | 274 | if (!rseq_cmp_item(p, pool->item_len, 0, NULL)) |
a5694a4d | 275 | continue; |
644298bb | 276 | bzero(p, pool->item_len); |
367e559c MD |
277 | } |
278 | } | |
279 | ||
6ff43d9a MD |
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 | { | |
a5694a4d | 285 | char *init_p = NULL; |
6ff43d9a MD |
286 | int i; |
287 | ||
a5694a4d MD |
288 | init_p = __rseq_pool_range_init_ptr(range, item_offset); |
289 | if (init_p) | |
290 | memcpy(init_p, init_ptr, init_len); | |
6ff43d9a MD |
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); | |
a5694a4d | 294 | |
1b658191 MD |
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)) | |
a5694a4d | 302 | continue; |
6ff43d9a MD |
303 | memcpy(p, init_ptr, init_len); |
304 | } | |
305 | } | |
306 | ||
a5694a4d MD |
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 | ||
455e090e MD |
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; | |
a5694a4d | 321 | char *init_p = NULL; |
455e090e MD |
322 | int i; |
323 | ||
a5694a4d MD |
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); | |
455e090e MD |
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); | |
455e090e | 330 | |
1b658191 MD |
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 | */ | |
14af0aa2 | 341 | if (rseq_cmp_item(p, pool->item_len, poison, NULL) == 0) |
a5694a4d MD |
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, | |
c0de0012 | 350 | void *p, size_t item_len, uintptr_t poison) |
a5694a4d | 351 | { |
1b658191 | 352 | intptr_t unexpected_value; |
a5694a4d | 353 | |
14af0aa2 | 354 | if (rseq_cmp_item(p, item_len, poison, &unexpected_value) == 0) |
1b658191 | 355 | return; |
a5694a4d | 356 | |
1b658191 MD |
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(); | |
86617384 MD |
360 | } |
361 | ||
362 | /* Always inline for __builtin_return_address(0). */ | |
363 | static inline __attribute__((always_inline)) | |
6fbf1fb6 MD |
364 | void rseq_percpu_check_poison_item(const struct rseq_mempool *pool, |
365 | const struct rseq_mempool_range *range, uintptr_t item_offset) | |
86617384 MD |
366 | { |
367 | uintptr_t poison = pool->attr.poison; | |
a5694a4d | 368 | char *init_p; |
86617384 MD |
369 | int i; |
370 | ||
3975084e | 371 | if (!pool->attr.robust_set) |
86617384 | 372 | return; |
a5694a4d MD |
373 | init_p = __rseq_pool_range_init_ptr(range, item_offset); |
374 | if (init_p) | |
c0de0012 | 375 | rseq_check_poison_item(pool, item_offset, init_p, pool->item_len, poison); |
86617384 MD |
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); | |
c0de0012 | 379 | rseq_check_poison_item(pool, item_offset, p, pool->item_len, poison); |
455e090e MD |
380 | } |
381 | } | |
382 | ||
15b63c9f | 383 | #ifdef HAVE_LIBNUMA |
c6fd3981 | 384 | int rseq_mempool_range_init_numa(void *addr, size_t len, int cpu, int numa_flags) |
367e559c | 385 | { |
f2981623 | 386 | unsigned long nr_pages, page_len; |
c6fd3981 MD |
387 | int status[MOVE_PAGES_BATCH_SIZE]; |
388 | int nodes[MOVE_PAGES_BATCH_SIZE]; | |
389 | void *pages[MOVE_PAGES_BATCH_SIZE]; | |
f2981623 | 390 | long ret; |
367e559c | 391 | |
c6fd3981 MD |
392 | if (!numa_flags) { |
393 | errno = EINVAL; | |
394 | return -1; | |
395 | } | |
367e559c | 396 | page_len = rseq_get_page_len(); |
c6fd3981 | 397 | nr_pages = len >> rseq_get_count_order_ulong(page_len); |
57d8b586 | 398 | |
c6fd3981 MD |
399 | nodes[0] = numa_node_of_cpu(cpu); |
400 | if (nodes[0] < 0) | |
401 | return -1; | |
57d8b586 | 402 | |
c6fd3981 MD |
403 | for (size_t k = 1; k < RSEQ_ARRAY_SIZE(nodes); ++k) { |
404 | nodes[k] = nodes[0]; | |
405 | } | |
57d8b586 | 406 | |
c6fd3981 | 407 | for (unsigned long page = 0; page < nr_pages;) { |
57d8b586 | 408 | |
c6fd3981 MD |
409 | size_t max_k = RSEQ_ARRAY_SIZE(pages); |
410 | size_t left = nr_pages - page; | |
57d8b586 | 411 | |
c6fd3981 MD |
412 | if (left < max_k) { |
413 | max_k = left; | |
414 | } | |
57d8b586 | 415 | |
c6fd3981 MD |
416 | for (size_t k = 0; k < max_k; ++k, ++page) { |
417 | pages[k] = addr + (page * page_len); | |
418 | status[k] = -EPERM; | |
367e559c | 419 | } |
b73b0c25 | 420 | |
c6fd3981 MD |
421 | ret = move_pages(0, max_k, pages, nodes, status, numa_flags); |
422 | ||
423 | if (ret < 0) | |
b73b0c25 | 424 | return ret; |
c6fd3981 MD |
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 | } | |
b73b0c25 MD |
435 | } |
436 | return 0; | |
437 | } | |
367e559c | 438 | #else |
c6fd3981 MD |
439 | int rseq_mempool_range_init_numa(void *addr __attribute__((unused)), |
440 | size_t len __attribute__((unused)), | |
441 | int cpu __attribute__((unused)), | |
367e559c MD |
442 | int numa_flags __attribute__((unused))) |
443 | { | |
c6fd3981 MD |
444 | errno = ENOSYS; |
445 | return -1; | |
367e559c MD |
446 | } |
447 | #endif | |
448 | ||
0fdf7a4c | 449 | static |
0ba2a93e | 450 | int create_alloc_bitmap(struct rseq_mempool *pool, struct rseq_mempool_range *range) |
0fdf7a4c OD |
451 | { |
452 | size_t count; | |
453 | ||
cb475906 | 454 | count = ((pool->attr.stride >> pool->item_order) + BIT_PER_ULONG - 1) / BIT_PER_ULONG; |
0fdf7a4c OD |
455 | |
456 | /* | |
9649c7ee MD |
457 | * Not being able to create the validation bitmap is an error |
458 | * that needs to be reported. | |
0fdf7a4c | 459 | */ |
b73b0c25 MD |
460 | range->alloc_bitmap = calloc(count, sizeof(unsigned long)); |
461 | if (!range->alloc_bitmap) | |
9649c7ee MD |
462 | return -1; |
463 | return 0; | |
0fdf7a4c OD |
464 | } |
465 | ||
b73b0c25 | 466 | static |
a5694a4d | 467 | bool percpu_addr_in_pool(const struct rseq_mempool *pool, void __rseq_percpu *_addr) |
b73b0c25 | 468 | { |
0ba2a93e | 469 | struct rseq_mempool_range *range; |
a5694a4d | 470 | void *addr = (void *) _addr; |
b73b0c25 | 471 | |
9d986353 | 472 | for (range = pool->range_list; range; range = range->next) { |
b73b0c25 MD |
473 | if (addr >= range->base && addr < range->base + range->next_unused) |
474 | return true; | |
475 | } | |
476 | return false; | |
477 | } | |
478 | ||
a9ec6111 OD |
479 | /* Always inline for __builtin_return_address(0). */ |
480 | static inline __attribute__((always_inline)) | |
0ba2a93e | 481 | void check_free_list(const struct rseq_mempool *pool) |
a9ec6111 | 482 | { |
b73b0c25 MD |
483 | size_t total_item = 0, total_never_allocated = 0, total_freed = 0, |
484 | max_list_traversal = 0, traversal_iteration = 0; | |
0ba2a93e | 485 | struct rseq_mempool_range *range; |
b73b0c25 MD |
486 | |
487 | if (!pool->attr.robust_set) | |
488 | return; | |
489 | ||
9d986353 | 490 | for (range = pool->range_list; range; range = range->next) { |
cb475906 MD |
491 | total_item += pool->attr.stride >> pool->item_order; |
492 | total_never_allocated += (pool->attr.stride - range->next_unused) >> pool->item_order; | |
b73b0c25 MD |
493 | } |
494 | max_list_traversal = total_item - total_never_allocated; | |
a9ec6111 OD |
495 | |
496 | for (struct free_list_node *node = pool->free_list_head, *prev = NULL; | |
497 | node; | |
498 | prev = node, | |
499 | node = node->next) { | |
500 | ||
a9ec6111 | 501 | if (traversal_iteration >= max_list_traversal) { |
ca452fee MD |
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)); | |
a9ec6111 OD |
504 | abort(); |
505 | } | |
506 | ||
507 | /* Node is out of range. */ | |
a5694a4d | 508 | if (!percpu_addr_in_pool(pool, __rseq_free_list_to_percpu_ptr(pool, node))) { |
a9ec6111 | 509 | if (prev) |
ca452fee MD |
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)); | |
a9ec6111 | 512 | else |
ca452fee MD |
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)); | |
a9ec6111 OD |
515 | abort(); |
516 | } | |
517 | ||
b73b0c25 MD |
518 | traversal_iteration++; |
519 | total_freed++; | |
a9ec6111 OD |
520 | } |
521 | ||
522 | if (total_never_allocated + total_freed != total_item) { | |
ca452fee MD |
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)); | |
a9ec6111 OD |
525 | abort(); |
526 | } | |
a9ec6111 OD |
527 | } |
528 | ||
6fbf1fb6 MD |
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 | ||
3975084e | 547 | if (!pool->attr.robust_set) |
6fbf1fb6 MD |
548 | return; |
549 | for (range = pool->range_list; range; range = range->next) | |
550 | check_range_poison(pool, range); | |
551 | } | |
552 | ||
e7cbbc10 MD |
553 | /* Always inline for __builtin_return_address(0). */ |
554 | static inline __attribute__((always_inline)) | |
0ba2a93e | 555 | void destroy_alloc_bitmap(struct rseq_mempool *pool, struct rseq_mempool_range *range) |
0fdf7a4c | 556 | { |
b73b0c25 | 557 | unsigned long *bitmap = range->alloc_bitmap; |
9649c7ee | 558 | size_t count, total_leaks = 0; |
0fdf7a4c | 559 | |
9649c7ee | 560 | if (!bitmap) |
0fdf7a4c | 561 | return; |
0fdf7a4c | 562 | |
cb475906 | 563 | count = ((pool->attr.stride >> pool->item_order) + BIT_PER_ULONG - 1) / BIT_PER_ULONG; |
0fdf7a4c OD |
564 | |
565 | /* Assert that all items in the pool were freed. */ | |
9649c7ee MD |
566 | for (size_t k = 0; k < count; ++k) |
567 | total_leaks += rseq_hweight_ulong(bitmap[k]); | |
568 | if (total_leaks) { | |
ca452fee MD |
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)); | |
9649c7ee | 571 | abort(); |
0fdf7a4c OD |
572 | } |
573 | ||
574 | free(bitmap); | |
a5694a4d | 575 | range->alloc_bitmap = NULL; |
0fdf7a4c OD |
576 | } |
577 | ||
b73b0c25 MD |
578 | /* Always inline for __builtin_return_address(0). */ |
579 | static inline __attribute__((always_inline)) | |
0ba2a93e MD |
580 | int rseq_mempool_range_destroy(struct rseq_mempool *pool, |
581 | struct rseq_mempool_range *range) | |
b73b0c25 MD |
582 | { |
583 | destroy_alloc_bitmap(pool, range); | |
a5694a4d | 584 | |
5c99f3d6 | 585 | /* range is a header located one page before the aligned mapping. */ |
5cd72fc7 | 586 | return munmap(range->mmap_addr, range->mmap_len); |
5c99f3d6 MD |
587 | } |
588 | ||
589 | /* | |
590 | * Allocate a memory mapping aligned on @alignment, with an optional | |
591 | * @pre_header before the mapping. | |
592 | */ | |
593 | static | |
5cd72fc7 | 594 | void *aligned_mmap_anonymous(size_t page_size, size_t len, size_t alignment, |
5c99f3d6 MD |
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 || | |
b72b2d9e | 602 | !is_pow2(alignment) || (len & (alignment - 1))) { |
5c99f3d6 MD |
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 | ||
5cd72fc7 MD |
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; | |
5c99f3d6 | 625 | goto alloc_error; |
5cd72fc7 | 626 | } |
5c99f3d6 MD |
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))); | |
5cd72fc7 | 638 | if (munmap(ptr, extra)) { |
5c99f3d6 MD |
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; | |
5cd72fc7 | 654 | if (munmap(extra_ptr, extra)) { |
5c99f3d6 MD |
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; | |
b73b0c25 MD |
671 | } |
672 | ||
a5694a4d | 673 | static |
cc0413ab | 674 | int rseq_memfd_create_init(const char *poolname, size_t init_len) |
a5694a4d | 675 | { |
a10c1c93 | 676 | int fd; |
cc0413ab OD |
677 | char buf[249]; /* Limit is 249 bytes. */ |
678 | const char *name; | |
a10c1c93 | 679 | |
cc0413ab OD |
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); | |
a10c1c93 MD |
688 | if (fd < 0) { |
689 | perror("memfd_create"); | |
690 | goto end; | |
a5694a4d | 691 | } |
a10c1c93 | 692 | if (ftruncate(fd, (off_t) init_len)) { |
025165ad MD |
693 | if (close(fd)) |
694 | perror("close"); | |
a10c1c93 MD |
695 | fd = -1; |
696 | goto end; | |
697 | } | |
698 | end: | |
699 | return fd; | |
700 | } | |
701 | ||
702 | static | |
703 | void rseq_memfd_close(int fd) | |
704 | { | |
025165ad MD |
705 | if (fd < 0) |
706 | return; | |
a10c1c93 MD |
707 | if (close(fd)) |
708 | perror("close"); | |
a5694a4d MD |
709 | } |
710 | ||
b73b0c25 | 711 | static |
0ba2a93e | 712 | struct rseq_mempool_range *rseq_mempool_range_create(struct rseq_mempool *pool) |
b73b0c25 | 713 | { |
0ba2a93e | 714 | struct rseq_mempool_range *range; |
5c99f3d6 | 715 | unsigned long page_size; |
4aa3220c | 716 | void *header; |
b73b0c25 | 717 | void *base; |
a5694a4d | 718 | size_t range_len; /* Range len excludes header. */ |
025165ad | 719 | int memfd = -1; |
b73b0c25 | 720 | |
e11a02d7 MD |
721 | if (pool->attr.max_nr_ranges && |
722 | pool->nr_ranges >= pool->attr.max_nr_ranges) { | |
9d986353 MD |
723 | errno = ENOMEM; |
724 | return NULL; | |
725 | } | |
5c99f3d6 | 726 | page_size = rseq_get_page_len(); |
b73b0c25 | 727 | |
a5694a4d | 728 | range_len = pool->attr.stride * pool->attr.max_nr_cpus; |
4e8ae59d | 729 | if (pool->attr.populate_policy != RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL) |
a5694a4d | 730 | range_len += pool->attr.stride; /* init values */ |
c0de0012 MD |
731 | if (pool->attr.robust_set) |
732 | range_len += pool->attr.stride; /* free list */ | |
5cd72fc7 MD |
733 | base = aligned_mmap_anonymous(page_size, range_len, |
734 | pool->attr.stride, &header, page_size); | |
b73b0c25 | 735 | if (!base) |
5c99f3d6 | 736 | return NULL; |
0ba2a93e | 737 | range = (struct rseq_mempool_range *) (base - RANGE_HEADER_OFFSET); |
5c99f3d6 | 738 | range->pool = pool; |
4aa3220c | 739 | range->header = header; |
a5694a4d | 740 | range->base = base; |
fa6a0fb3 | 741 | range->mmap_addr = header; |
a5694a4d MD |
742 | range->mmap_len = page_size + range_len; |
743 | ||
4e8ae59d | 744 | if (pool->attr.populate_policy != RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL) { |
a5694a4d MD |
745 | range->init = base + (pool->attr.stride * pool->attr.max_nr_cpus); |
746 | /* Populate init values pages from memfd */ | |
cc0413ab | 747 | memfd = rseq_memfd_create_init(pool->name, pool->attr.stride); |
a10c1c93 | 748 | if (memfd < 0) |
a5694a4d MD |
749 | goto error_alloc; |
750 | if (mmap(range->init, pool->attr.stride, PROT_READ | PROT_WRITE, | |
a10c1c93 | 751 | MAP_SHARED | MAP_FIXED, memfd, 0) != (void *) range->init) { |
a5694a4d MD |
752 | goto error_alloc; |
753 | } | |
754 | assert(pool->attr.type == MEMPOOL_TYPE_PERCPU); | |
755 | /* | |
756 | * Map per-cpu memory as private COW mappings of init values. | |
757 | */ | |
758 | { | |
759 | int cpu; | |
760 | ||
761 | for (cpu = 0; cpu < pool->attr.max_nr_cpus; cpu++) { | |
762 | void *p = base + (pool->attr.stride * cpu); | |
763 | size_t len = pool->attr.stride; | |
764 | ||
765 | if (mmap(p, len, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_FIXED, | |
a10c1c93 | 766 | memfd, 0) != (void *) p) { |
a5694a4d MD |
767 | goto error_alloc; |
768 | } | |
769 | } | |
770 | } | |
bf7b01a3 MD |
771 | /* |
772 | * The init values shared mapping should not be shared | |
773 | * with the children processes across fork. Prevent the | |
774 | * whole mapping from being used across fork. | |
775 | */ | |
776 | if (madvise(range->mmap_addr, pool->mmap_len, MADV_DONTFORK)) | |
777 | goto error_alloc; | |
a10c1c93 | 778 | rseq_memfd_close(memfd); |
025165ad | 779 | memfd = -1; |
a5694a4d MD |
780 | } |
781 | ||
b73b0c25 MD |
782 | if (pool->attr.robust_set) { |
783 | if (create_alloc_bitmap(pool, range)) | |
784 | goto error_alloc; | |
785 | } | |
135811f2 | 786 | if (pool->attr.init_set) { |
374c2773 MD |
787 | switch (pool->attr.type) { |
788 | case MEMPOOL_TYPE_GLOBAL: | |
6e329183 | 789 | if (pool->attr.init_func(pool->attr.init_priv, |
374c2773 | 790 | base, pool->attr.stride, -1)) { |
6e329183 MD |
791 | goto error_alloc; |
792 | } | |
374c2773 MD |
793 | break; |
794 | case MEMPOOL_TYPE_PERCPU: | |
795 | { | |
796 | int cpu; | |
797 | for (cpu = 0; cpu < pool->attr.max_nr_cpus; cpu++) { | |
798 | if (pool->attr.init_func(pool->attr.init_priv, | |
799 | base + (pool->attr.stride * cpu), | |
800 | pool->attr.stride, cpu)) { | |
801 | goto error_alloc; | |
802 | } | |
803 | } | |
804 | break; | |
805 | } | |
806 | default: | |
807 | abort(); | |
135811f2 MD |
808 | } |
809 | } | |
9d986353 | 810 | pool->nr_ranges++; |
b73b0c25 MD |
811 | return range; |
812 | ||
813 | error_alloc: | |
025165ad | 814 | rseq_memfd_close(memfd); |
0ba2a93e | 815 | (void) rseq_mempool_range_destroy(pool, range); |
b73b0c25 MD |
816 | return NULL; |
817 | } | |
818 | ||
0ba2a93e | 819 | int rseq_mempool_destroy(struct rseq_mempool *pool) |
9649c7ee | 820 | { |
0ba2a93e | 821 | struct rseq_mempool_range *range, *next_range; |
b73b0c25 | 822 | int ret = 0; |
9649c7ee | 823 | |
f510ddc5 MD |
824 | if (!pool) |
825 | return 0; | |
b73b0c25 | 826 | check_free_list(pool); |
6fbf1fb6 | 827 | check_pool_poison(pool); |
b73b0c25 | 828 | /* Iteration safe against removal. */ |
9d986353 | 829 | for (range = pool->range_list; range && (next_range = range->next, 1); range = next_range) { |
0ba2a93e | 830 | if (rseq_mempool_range_destroy(pool, range)) |
b73b0c25 MD |
831 | goto end; |
832 | /* Update list head to keep list coherent in case of partial failure. */ | |
9d986353 | 833 | pool->range_list = next_range; |
b73b0c25 | 834 | } |
9649c7ee | 835 | pthread_mutex_destroy(&pool->lock); |
ca452fee | 836 | free(pool->name); |
eb8db04d | 837 | free(pool); |
9649c7ee | 838 | end: |
b73b0c25 | 839 | return ret; |
9649c7ee MD |
840 | } |
841 | ||
0ba2a93e | 842 | struct rseq_mempool *rseq_mempool_create(const char *pool_name, |
cb475906 | 843 | size_t item_len, const struct rseq_mempool_attr *_attr) |
ef6695f1 | 844 | { |
0ba2a93e MD |
845 | struct rseq_mempool *pool; |
846 | struct rseq_mempool_attr attr = {}; | |
ef6695f1 | 847 | int order; |
ef6695f1 MD |
848 | |
849 | /* Make sure each item is large enough to contain free list pointers. */ | |
850 | if (item_len < sizeof(void *)) | |
851 | item_len = sizeof(void *); | |
852 | ||
853 | /* Align item_len on next power of two. */ | |
19be9217 | 854 | order = rseq_get_count_order_ulong(item_len); |
ef6695f1 MD |
855 | if (order < 0) { |
856 | errno = EINVAL; | |
857 | return NULL; | |
858 | } | |
859 | item_len = 1UL << order; | |
860 | ||
a82006d0 MD |
861 | if (_attr) |
862 | memcpy(&attr, _attr, sizeof(attr)); | |
a82006d0 | 863 | |
cb475906 MD |
864 | switch (attr.type) { |
865 | case MEMPOOL_TYPE_PERCPU: | |
866 | if (attr.max_nr_cpus < 0) { | |
867 | errno = EINVAL; | |
868 | return NULL; | |
869 | } | |
870 | if (attr.max_nr_cpus == 0) { | |
871 | /* Auto-detect */ | |
47c725dd | 872 | attr.max_nr_cpus = rseq_get_max_nr_cpus(); |
cb475906 MD |
873 | if (attr.max_nr_cpus == 0) { |
874 | errno = EINVAL; | |
875 | return NULL; | |
876 | } | |
877 | } | |
878 | break; | |
879 | case MEMPOOL_TYPE_GLOBAL: | |
a5694a4d | 880 | /* Override populate policy for global type. */ |
4e8ae59d MD |
881 | if (attr.populate_policy == RSEQ_MEMPOOL_POPULATE_PRIVATE_NONE) |
882 | attr.populate_policy = RSEQ_MEMPOOL_POPULATE_PRIVATE_ALL; | |
89b7e681 MD |
883 | /* Use a 1-cpu pool for global mempool type. */ |
884 | attr.max_nr_cpus = 1; | |
cb475906 MD |
885 | break; |
886 | } | |
887 | if (!attr.stride) | |
888 | attr.stride = RSEQ_MEMPOOL_STRIDE; /* Use default */ | |
3975084e MD |
889 | if (attr.robust_set && !attr.poison_set) { |
890 | attr.poison_set = true; | |
dbccd436 | 891 | attr.poison = DEFAULT_PRIVATE_POISON_VALUE; |
3975084e | 892 | } |
cb475906 MD |
893 | if (item_len > attr.stride || attr.stride < (size_t) rseq_get_page_len() || |
894 | !is_pow2(attr.stride)) { | |
895 | errno = EINVAL; | |
896 | return NULL; | |
897 | } | |
898 | ||
0ba2a93e | 899 | pool = calloc(1, sizeof(struct rseq_mempool)); |
bc510b60 MD |
900 | if (!pool) |
901 | return NULL; | |
ef6695f1 | 902 | |
b73b0c25 | 903 | memcpy(&pool->attr, &attr, sizeof(attr)); |
ef6695f1 | 904 | pthread_mutex_init(&pool->lock, NULL); |
ef6695f1 MD |
905 | pool->item_len = item_len; |
906 | pool->item_order = order; | |
b73b0c25 | 907 | |
9d986353 MD |
908 | pool->range_list = rseq_mempool_range_create(pool); |
909 | if (!pool->range_list) | |
b73b0c25 | 910 | goto error_alloc; |
0fdf7a4c | 911 | |
ca452fee MD |
912 | if (pool_name) { |
913 | pool->name = strdup(pool_name); | |
914 | if (!pool->name) | |
915 | goto error_alloc; | |
916 | } | |
ef6695f1 | 917 | return pool; |
ef6695f1 | 918 | |
9649c7ee | 919 | error_alloc: |
0ba2a93e | 920 | rseq_mempool_destroy(pool); |
9649c7ee MD |
921 | errno = ENOMEM; |
922 | return NULL; | |
ef6695f1 MD |
923 | } |
924 | ||
e7cbbc10 MD |
925 | /* Always inline for __builtin_return_address(0). */ |
926 | static inline __attribute__((always_inline)) | |
9d986353 | 927 | void set_alloc_slot(struct rseq_mempool *pool, struct rseq_mempool_range *range, size_t item_offset) |
0fdf7a4c | 928 | { |
9d986353 | 929 | unsigned long *bitmap = range->alloc_bitmap; |
9649c7ee | 930 | size_t item_index = item_offset >> pool->item_order; |
0fdf7a4c OD |
931 | unsigned long mask; |
932 | size_t k; | |
933 | ||
9649c7ee | 934 | if (!bitmap) |
0fdf7a4c | 935 | return; |
0fdf7a4c | 936 | |
9649c7ee | 937 | k = item_index / BIT_PER_ULONG; |
0fdf7a4c OD |
938 | mask = 1ULL << (item_index % BIT_PER_ULONG); |
939 | ||
9649c7ee MD |
940 | /* Print error if bit is already set. */ |
941 | if (bitmap[k] & mask) { | |
ca452fee MD |
942 | fprintf(stderr, "%s: Allocator corruption detected for pool: \"%s\" (%p), item offset: %zu, caller: %p.\n", |
943 | __func__, get_pool_name(pool), pool, item_offset, (void *) __builtin_return_address(0)); | |
9649c7ee MD |
944 | abort(); |
945 | } | |
0fdf7a4c OD |
946 | bitmap[k] |= mask; |
947 | } | |
948 | ||
ef6695f1 | 949 | static |
6ff43d9a MD |
950 | void __rseq_percpu *__rseq_percpu_malloc(struct rseq_mempool *pool, |
951 | bool zeroed, void *init_ptr, size_t init_len) | |
ef6695f1 | 952 | { |
9d986353 | 953 | struct rseq_mempool_range *range; |
ef6695f1 MD |
954 | struct free_list_node *node; |
955 | uintptr_t item_offset; | |
d24ee051 | 956 | void __rseq_percpu *addr; |
ef6695f1 | 957 | |
6ff43d9a MD |
958 | if (init_len > pool->item_len) { |
959 | errno = EINVAL; | |
960 | return NULL; | |
961 | } | |
ef6695f1 MD |
962 | pthread_mutex_lock(&pool->lock); |
963 | /* Get first entry from free list. */ | |
964 | node = pool->free_list_head; | |
965 | if (node != NULL) { | |
a5694a4d | 966 | void *range_base, *ptr; |
9d986353 | 967 | |
a5694a4d MD |
968 | ptr = __rseq_free_list_to_percpu_ptr(pool, node); |
969 | range_base = (void *) ((uintptr_t) ptr & (~(pool->attr.stride - 1))); | |
9d986353 | 970 | range = (struct rseq_mempool_range *) (range_base - RANGE_HEADER_OFFSET); |
ef6695f1 MD |
971 | /* Remove node from free list (update head). */ |
972 | pool->free_list_head = node->next; | |
a5694a4d | 973 | item_offset = (uintptr_t) (ptr - range_base); |
86617384 | 974 | rseq_percpu_check_poison_item(pool, range, item_offset); |
a5694a4d | 975 | addr = __rseq_free_list_to_percpu_ptr(pool, node); |
ef6695f1 MD |
976 | goto end; |
977 | } | |
9d986353 MD |
978 | /* |
979 | * If the most recent range (first in list) does not have any | |
980 | * room left, create a new range and prepend it to the list | |
981 | * head. | |
982 | */ | |
983 | range = pool->range_list; | |
984 | if (range->next_unused + pool->item_len > pool->attr.stride) { | |
985 | range = rseq_mempool_range_create(pool); | |
986 | if (!range) { | |
987 | errno = ENOMEM; | |
988 | addr = NULL; | |
989 | goto end; | |
990 | } | |
991 | /* Add range to head of list. */ | |
992 | range->next = pool->range_list; | |
993 | pool->range_list = range; | |
ef6695f1 | 994 | } |
9d986353 MD |
995 | /* First range in list has room left. */ |
996 | item_offset = range->next_unused; | |
997 | addr = (void __rseq_percpu *) (range->base + item_offset); | |
998 | range->next_unused += pool->item_len; | |
ef6695f1 | 999 | end: |
8f28507f | 1000 | if (addr) |
9d986353 | 1001 | set_alloc_slot(pool, range, item_offset); |
ef6695f1 | 1002 | pthread_mutex_unlock(&pool->lock); |
6ff43d9a MD |
1003 | if (addr) { |
1004 | if (zeroed) | |
1005 | rseq_percpu_zero_item(pool, range, item_offset); | |
1006 | else if (init_ptr) { | |
1007 | rseq_percpu_init_item(pool, range, item_offset, | |
1008 | init_ptr, init_len); | |
1009 | } | |
1010 | } | |
ef6695f1 MD |
1011 | return addr; |
1012 | } | |
1013 | ||
15da5c27 | 1014 | void __rseq_percpu *rseq_mempool_percpu_malloc(struct rseq_mempool *pool) |
ef6695f1 | 1015 | { |
6ff43d9a | 1016 | return __rseq_percpu_malloc(pool, false, NULL, 0); |
ef6695f1 MD |
1017 | } |
1018 | ||
15da5c27 | 1019 | void __rseq_percpu *rseq_mempool_percpu_zmalloc(struct rseq_mempool *pool) |
ef6695f1 | 1020 | { |
6ff43d9a MD |
1021 | return __rseq_percpu_malloc(pool, true, NULL, 0); |
1022 | } | |
1023 | ||
1024 | void __rseq_percpu *rseq_mempool_percpu_malloc_init(struct rseq_mempool *pool, | |
1025 | void *init_ptr, size_t len) | |
1026 | { | |
1027 | return __rseq_percpu_malloc(pool, false, init_ptr, len); | |
ef6695f1 MD |
1028 | } |
1029 | ||
e7cbbc10 MD |
1030 | /* Always inline for __builtin_return_address(0). */ |
1031 | static inline __attribute__((always_inline)) | |
9d986353 | 1032 | void clear_alloc_slot(struct rseq_mempool *pool, struct rseq_mempool_range *range, size_t item_offset) |
0fdf7a4c | 1033 | { |
9d986353 | 1034 | unsigned long *bitmap = range->alloc_bitmap; |
9649c7ee | 1035 | size_t item_index = item_offset >> pool->item_order; |
0fdf7a4c OD |
1036 | unsigned long mask; |
1037 | size_t k; | |
1038 | ||
9649c7ee | 1039 | if (!bitmap) |
0fdf7a4c | 1040 | return; |
0fdf7a4c | 1041 | |
9649c7ee MD |
1042 | k = item_index / BIT_PER_ULONG; |
1043 | mask = 1ULL << (item_index % BIT_PER_ULONG); | |
0fdf7a4c | 1044 | |
9649c7ee MD |
1045 | /* Print error if bit is not set. */ |
1046 | if (!(bitmap[k] & mask)) { | |
ca452fee MD |
1047 | fprintf(stderr, "%s: Double-free detected for pool: \"%s\" (%p), item offset: %zu, caller: %p.\n", |
1048 | __func__, get_pool_name(pool), pool, item_offset, | |
1049 | (void *) __builtin_return_address(0)); | |
9649c7ee MD |
1050 | abort(); |
1051 | } | |
0fdf7a4c OD |
1052 | bitmap[k] &= ~mask; |
1053 | } | |
1054 | ||
cb475906 | 1055 | void librseq_mempool_percpu_free(void __rseq_percpu *_ptr, size_t stride) |
ef6695f1 MD |
1056 | { |
1057 | uintptr_t ptr = (uintptr_t) _ptr; | |
cb475906 | 1058 | void *range_base = (void *) (ptr & (~(stride - 1))); |
0ba2a93e MD |
1059 | struct rseq_mempool_range *range = (struct rseq_mempool_range *) (range_base - RANGE_HEADER_OFFSET); |
1060 | struct rseq_mempool *pool = range->pool; | |
cb475906 | 1061 | uintptr_t item_offset = ptr & (stride - 1); |
ef6695f1 MD |
1062 | struct free_list_node *head, *item; |
1063 | ||
1064 | pthread_mutex_lock(&pool->lock); | |
9d986353 | 1065 | clear_alloc_slot(pool, range, item_offset); |
ef6695f1 MD |
1066 | /* Add ptr to head of free list */ |
1067 | head = pool->free_list_head; | |
455e090e MD |
1068 | if (pool->attr.poison_set) |
1069 | rseq_percpu_poison_item(pool, range, item_offset); | |
a5694a4d | 1070 | item = __rseq_percpu_to_free_list_ptr(pool, _ptr); |
455e090e MD |
1071 | /* |
1072 | * Setting the next pointer will overwrite the first uintptr_t | |
a5694a4d MD |
1073 | * poison for either CPU 0 (populate all) or init data (populate |
1074 | * none). | |
455e090e | 1075 | */ |
ef6695f1 MD |
1076 | item->next = head; |
1077 | pool->free_list_head = item; | |
1078 | pthread_mutex_unlock(&pool->lock); | |
1079 | } | |
1080 | ||
0ba2a93e | 1081 | struct rseq_mempool_set *rseq_mempool_set_create(void) |
ef6695f1 | 1082 | { |
0ba2a93e | 1083 | struct rseq_mempool_set *pool_set; |
ef6695f1 | 1084 | |
0ba2a93e | 1085 | pool_set = calloc(1, sizeof(struct rseq_mempool_set)); |
ef6695f1 MD |
1086 | if (!pool_set) |
1087 | return NULL; | |
1088 | pthread_mutex_init(&pool_set->lock, NULL); | |
1089 | return pool_set; | |
1090 | } | |
1091 | ||
0ba2a93e | 1092 | int rseq_mempool_set_destroy(struct rseq_mempool_set *pool_set) |
ef6695f1 MD |
1093 | { |
1094 | int order, ret; | |
1095 | ||
1096 | for (order = POOL_SET_MIN_ENTRY; order < POOL_SET_NR_ENTRIES; order++) { | |
0ba2a93e | 1097 | struct rseq_mempool *pool = pool_set->entries[order]; |
ef6695f1 MD |
1098 | |
1099 | if (!pool) | |
1100 | continue; | |
0ba2a93e | 1101 | ret = rseq_mempool_destroy(pool); |
ef6695f1 MD |
1102 | if (ret) |
1103 | return ret; | |
1104 | pool_set->entries[order] = NULL; | |
1105 | } | |
1106 | pthread_mutex_destroy(&pool_set->lock); | |
1107 | free(pool_set); | |
1108 | return 0; | |
1109 | } | |
1110 | ||
1111 | /* Ownership of pool is handed over to pool set on success. */ | |
0ba2a93e | 1112 | int rseq_mempool_set_add_pool(struct rseq_mempool_set *pool_set, struct rseq_mempool *pool) |
ef6695f1 MD |
1113 | { |
1114 | size_t item_order = pool->item_order; | |
1115 | int ret = 0; | |
1116 | ||
1117 | pthread_mutex_lock(&pool_set->lock); | |
1118 | if (pool_set->entries[item_order]) { | |
1119 | errno = EBUSY; | |
1120 | ret = -1; | |
1121 | goto end; | |
1122 | } | |
1123 | pool_set->entries[pool->item_order] = pool; | |
1124 | end: | |
1125 | pthread_mutex_unlock(&pool_set->lock); | |
1126 | return ret; | |
1127 | } | |
1128 | ||
1129 | static | |
6ff43d9a MD |
1130 | void __rseq_percpu *__rseq_mempool_set_malloc(struct rseq_mempool_set *pool_set, |
1131 | void *init_ptr, size_t len, bool zeroed) | |
ef6695f1 MD |
1132 | { |
1133 | int order, min_order = POOL_SET_MIN_ENTRY; | |
0ba2a93e | 1134 | struct rseq_mempool *pool; |
d24ee051 | 1135 | void __rseq_percpu *addr; |
ef6695f1 | 1136 | |
d06f5cf5 MD |
1137 | order = rseq_get_count_order_ulong(len); |
1138 | if (order > POOL_SET_MIN_ENTRY) | |
1139 | min_order = order; | |
ef6695f1 MD |
1140 | again: |
1141 | pthread_mutex_lock(&pool_set->lock); | |
1142 | /* First smallest present pool where @len fits. */ | |
1143 | for (order = min_order; order < POOL_SET_NR_ENTRIES; order++) { | |
1144 | pool = pool_set->entries[order]; | |
1145 | ||
1146 | if (!pool) | |
1147 | continue; | |
1148 | if (pool->item_len >= len) | |
1149 | goto found; | |
1150 | } | |
1151 | pool = NULL; | |
1152 | found: | |
1153 | pthread_mutex_unlock(&pool_set->lock); | |
1154 | if (pool) { | |
6ff43d9a | 1155 | addr = __rseq_percpu_malloc(pool, zeroed, init_ptr, len); |
ef6695f1 MD |
1156 | if (addr == NULL && errno == ENOMEM) { |
1157 | /* | |
1158 | * If the allocation failed, try again with a | |
1159 | * larger pool. | |
1160 | */ | |
1161 | min_order = order + 1; | |
1162 | goto again; | |
1163 | } | |
1164 | } else { | |
1165 | /* Not found. */ | |
1166 | errno = ENOMEM; | |
1167 | addr = NULL; | |
1168 | } | |
1169 | return addr; | |
1170 | } | |
1171 | ||
15da5c27 | 1172 | void __rseq_percpu *rseq_mempool_set_percpu_malloc(struct rseq_mempool_set *pool_set, size_t len) |
ef6695f1 | 1173 | { |
6ff43d9a | 1174 | return __rseq_mempool_set_malloc(pool_set, NULL, len, false); |
ef6695f1 MD |
1175 | } |
1176 | ||
15da5c27 | 1177 | void __rseq_percpu *rseq_mempool_set_percpu_zmalloc(struct rseq_mempool_set *pool_set, size_t len) |
ef6695f1 | 1178 | { |
6ff43d9a MD |
1179 | return __rseq_mempool_set_malloc(pool_set, NULL, len, true); |
1180 | } | |
1181 | ||
1182 | void __rseq_percpu *rseq_mempool_set_percpu_malloc_init(struct rseq_mempool_set *pool_set, | |
1183 | void *init_ptr, size_t len) | |
1184 | { | |
1185 | return __rseq_mempool_set_malloc(pool_set, init_ptr, len, true); | |
ef6695f1 | 1186 | } |
9bd07c29 | 1187 | |
0ba2a93e | 1188 | struct rseq_mempool_attr *rseq_mempool_attr_create(void) |
a82006d0 | 1189 | { |
0ba2a93e | 1190 | return calloc(1, sizeof(struct rseq_mempool_attr)); |
a82006d0 MD |
1191 | } |
1192 | ||
0ba2a93e | 1193 | void rseq_mempool_attr_destroy(struct rseq_mempool_attr *attr) |
a82006d0 MD |
1194 | { |
1195 | free(attr); | |
1196 | } | |
1197 | ||
135811f2 | 1198 | int rseq_mempool_attr_set_init(struct rseq_mempool_attr *attr, |
6e329183 | 1199 | int (*init_func)(void *priv, void *addr, size_t len, int cpu), |
135811f2 MD |
1200 | void *init_priv) |
1201 | { | |
1202 | if (!attr) { | |
1203 | errno = EINVAL; | |
1204 | return -1; | |
1205 | } | |
1206 | attr->init_set = true; | |
1207 | attr->init_func = init_func; | |
1208 | attr->init_priv = init_priv; | |
1209 | return 0; | |
1210 | } | |
1211 | ||
0ba2a93e | 1212 | int rseq_mempool_attr_set_robust(struct rseq_mempool_attr *attr) |
d6acc8aa MD |
1213 | { |
1214 | if (!attr) { | |
1215 | errno = EINVAL; | |
1216 | return -1; | |
1217 | } | |
1218 | attr->robust_set = true; | |
1219 | return 0; | |
1220 | } | |
cb475906 MD |
1221 | |
1222 | int rseq_mempool_attr_set_percpu(struct rseq_mempool_attr *attr, | |
1223 | size_t stride, int max_nr_cpus) | |
1224 | { | |
1225 | if (!attr) { | |
1226 | errno = EINVAL; | |
1227 | return -1; | |
1228 | } | |
1229 | attr->type = MEMPOOL_TYPE_PERCPU; | |
1230 | attr->stride = stride; | |
1231 | attr->max_nr_cpus = max_nr_cpus; | |
1232 | return 0; | |
1233 | } | |
1234 | ||
1235 | int rseq_mempool_attr_set_global(struct rseq_mempool_attr *attr, | |
1236 | size_t stride) | |
1237 | { | |
1238 | if (!attr) { | |
1239 | errno = EINVAL; | |
1240 | return -1; | |
1241 | } | |
1242 | attr->type = MEMPOOL_TYPE_GLOBAL; | |
1243 | attr->stride = stride; | |
89b7e681 | 1244 | attr->max_nr_cpus = 0; |
cb475906 MD |
1245 | return 0; |
1246 | } | |
6037d364 | 1247 | |
e11a02d7 MD |
1248 | int rseq_mempool_attr_set_max_nr_ranges(struct rseq_mempool_attr *attr, |
1249 | unsigned long max_nr_ranges) | |
1250 | { | |
1251 | if (!attr) { | |
1252 | errno = EINVAL; | |
1253 | return -1; | |
1254 | } | |
1255 | attr->max_nr_ranges = max_nr_ranges; | |
1256 | return 0; | |
1257 | } | |
1258 | ||
455e090e MD |
1259 | int rseq_mempool_attr_set_poison(struct rseq_mempool_attr *attr, |
1260 | uintptr_t poison) | |
1261 | { | |
1262 | if (!attr) { | |
1263 | errno = EINVAL; | |
1264 | return -1; | |
1265 | } | |
1266 | attr->poison_set = true; | |
1267 | attr->poison = poison; | |
1268 | return 0; | |
1269 | } | |
1270 | ||
a5694a4d MD |
1271 | int rseq_mempool_attr_set_populate_policy(struct rseq_mempool_attr *attr, |
1272 | enum rseq_mempool_populate_policy policy) | |
1273 | { | |
1274 | if (!attr) { | |
1275 | errno = EINVAL; | |
1276 | return -1; | |
1277 | } | |
1278 | attr->populate_policy = policy; | |
1279 | return 0; | |
1280 | } | |
1281 | ||
6037d364 MD |
1282 | int rseq_mempool_get_max_nr_cpus(struct rseq_mempool *mempool) |
1283 | { | |
1284 | if (!mempool || mempool->attr.type != MEMPOOL_TYPE_PERCPU) { | |
1285 | errno = EINVAL; | |
1286 | return -1; | |
1287 | } | |
1288 | return mempool->attr.max_nr_cpus; | |
1289 | } |