| 1 | /* |
| 2 | * Copyright (C) 2011 David Goulet <david.goulet@polymtl.ca> |
| 3 | * |
| 4 | * SPDX-License-Identifier: GPL-2.0-only |
| 5 | * |
| 6 | */ |
| 7 | |
| 8 | #define _LGPL_SOURCE |
| 9 | #include <assert.h> |
| 10 | #include <fcntl.h> |
| 11 | #include <limits.h> |
| 12 | #include <stdlib.h> |
| 13 | #include <sys/types.h> |
| 14 | #include <sys/stat.h> |
| 15 | #include <unistd.h> |
| 16 | #include <stdbool.h> |
| 17 | |
| 18 | #include <common/error.h> |
| 19 | #include <common/defaults.h> |
| 20 | #include <common/macros.h> |
| 21 | #include <common/utils.h> |
| 22 | |
| 23 | #include "poll.h" |
| 24 | |
| 25 | /* |
| 26 | * Maximum number of fd we can monitor. |
| 27 | * |
| 28 | * For epoll(7), /proc/sys/fs/epoll/max_user_watches (since Linux 2.6.28) will |
| 29 | * be used for the maximum size of the poll set. If this interface is not |
| 30 | * available, according to the manpage, the max_user_watches value is 1/25 (4%) |
| 31 | * of the available low memory divided by the registration cost in bytes which |
| 32 | * is 90 bytes on a 32-bit kernel and 160 bytes on a 64-bit kernel. |
| 33 | * |
| 34 | */ |
| 35 | static unsigned int poll_max_size; |
| 36 | |
| 37 | /* |
| 38 | * Resize the epoll events structure of the new size. |
| 39 | * |
| 40 | * Return 0 on success or else -1 with the current events pointer untouched. |
| 41 | */ |
| 42 | static int resize_poll_event(struct lttng_poll_event *events, |
| 43 | uint32_t new_size) |
| 44 | { |
| 45 | struct epoll_event *ptr; |
| 46 | |
| 47 | assert(events); |
| 48 | |
| 49 | ptr = realloc(events->events, new_size * sizeof(*ptr)); |
| 50 | if (ptr == NULL) { |
| 51 | PERROR("realloc epoll add"); |
| 52 | goto error; |
| 53 | } |
| 54 | if (new_size > events->alloc_size) { |
| 55 | /* Zero newly allocated memory */ |
| 56 | memset(ptr + events->alloc_size, 0, |
| 57 | (new_size - events->alloc_size) * sizeof(*ptr)); |
| 58 | } |
| 59 | events->events = ptr; |
| 60 | events->alloc_size = new_size; |
| 61 | |
| 62 | return 0; |
| 63 | |
| 64 | error: |
| 65 | return -1; |
| 66 | } |
| 67 | |
| 68 | /* |
| 69 | * Create epoll set and allocate returned events structure. |
| 70 | */ |
| 71 | LTTNG_HIDDEN |
| 72 | int compat_epoll_create(struct lttng_poll_event *events, int size, int flags) |
| 73 | { |
| 74 | int ret; |
| 75 | |
| 76 | if (events == NULL || size <= 0) { |
| 77 | goto error; |
| 78 | } |
| 79 | |
| 80 | if (!poll_max_size) { |
| 81 | if (lttng_poll_set_max_size()) { |
| 82 | goto error; |
| 83 | } |
| 84 | } |
| 85 | |
| 86 | /* Don't bust the limit here */ |
| 87 | if (size > poll_max_size) { |
| 88 | size = poll_max_size; |
| 89 | } |
| 90 | |
| 91 | ret = compat_glibc_epoll_create(size, flags); |
| 92 | if (ret < 0) { |
| 93 | /* At this point, every error is fatal */ |
| 94 | PERROR("epoll_create1"); |
| 95 | goto error; |
| 96 | } |
| 97 | |
| 98 | events->epfd = ret; |
| 99 | |
| 100 | /* This *must* be freed by using lttng_poll_free() */ |
| 101 | events->events = zmalloc(size * sizeof(struct epoll_event)); |
| 102 | if (events->events == NULL) { |
| 103 | PERROR("zmalloc epoll set"); |
| 104 | goto error_close; |
| 105 | } |
| 106 | |
| 107 | events->alloc_size = events->init_size = size; |
| 108 | events->nb_fd = 0; |
| 109 | |
| 110 | return 0; |
| 111 | |
| 112 | error_close: |
| 113 | ret = close(events->epfd); |
| 114 | if (ret) { |
| 115 | PERROR("close"); |
| 116 | } |
| 117 | error: |
| 118 | return -1; |
| 119 | } |
| 120 | |
| 121 | /* |
| 122 | * Add a fd to the epoll set with requesting events. |
| 123 | */ |
| 124 | LTTNG_HIDDEN |
| 125 | int compat_epoll_add(struct lttng_poll_event *events, int fd, uint32_t req_events) |
| 126 | { |
| 127 | int ret; |
| 128 | struct epoll_event ev; |
| 129 | |
| 130 | if (events == NULL || events->events == NULL || fd < 0) { |
| 131 | ERR("Bad compat epoll add arguments"); |
| 132 | goto error; |
| 133 | } |
| 134 | |
| 135 | /* |
| 136 | * Zero struct epoll_event to ensure all representations of its |
| 137 | * union are zeroed. |
| 138 | */ |
| 139 | memset(&ev, 0, sizeof(ev)); |
| 140 | ev.events = req_events; |
| 141 | ev.data.fd = fd; |
| 142 | |
| 143 | ret = epoll_ctl(events->epfd, EPOLL_CTL_ADD, fd, &ev); |
| 144 | if (ret < 0) { |
| 145 | switch (errno) { |
| 146 | case EEXIST: |
| 147 | /* If exist, it's OK. */ |
| 148 | goto end; |
| 149 | case ENOSPC: |
| 150 | case EPERM: |
| 151 | /* Print PERROR and goto end not failing. Show must go on. */ |
| 152 | PERROR("epoll_ctl ADD"); |
| 153 | goto end; |
| 154 | default: |
| 155 | PERROR("epoll_ctl ADD fatal"); |
| 156 | goto error; |
| 157 | } |
| 158 | } |
| 159 | |
| 160 | events->nb_fd++; |
| 161 | |
| 162 | end: |
| 163 | return 0; |
| 164 | |
| 165 | error: |
| 166 | return -1; |
| 167 | } |
| 168 | |
| 169 | /* |
| 170 | * Remove a fd from the epoll set. |
| 171 | */ |
| 172 | LTTNG_HIDDEN |
| 173 | int compat_epoll_del(struct lttng_poll_event *events, int fd) |
| 174 | { |
| 175 | int ret; |
| 176 | |
| 177 | if (events == NULL || fd < 0 || events->nb_fd == 0) { |
| 178 | goto error; |
| 179 | } |
| 180 | |
| 181 | ret = epoll_ctl(events->epfd, EPOLL_CTL_DEL, fd, NULL); |
| 182 | if (ret < 0) { |
| 183 | switch (errno) { |
| 184 | case ENOENT: |
| 185 | case EPERM: |
| 186 | /* Print PERROR and goto end not failing. Show must go on. */ |
| 187 | PERROR("epoll_ctl DEL"); |
| 188 | goto end; |
| 189 | default: |
| 190 | PERROR("epoll_ctl DEL fatal"); |
| 191 | goto error; |
| 192 | } |
| 193 | } |
| 194 | |
| 195 | events->nb_fd--; |
| 196 | |
| 197 | end: |
| 198 | return 0; |
| 199 | |
| 200 | error: |
| 201 | return -1; |
| 202 | } |
| 203 | |
| 204 | /* |
| 205 | * Set an fd's events. |
| 206 | */ |
| 207 | LTTNG_HIDDEN |
| 208 | int compat_epoll_mod(struct lttng_poll_event *events, int fd, uint32_t req_events) |
| 209 | { |
| 210 | int ret; |
| 211 | struct epoll_event ev; |
| 212 | |
| 213 | if (events == NULL || fd < 0 || events->nb_fd == 0) { |
| 214 | goto error; |
| 215 | } |
| 216 | |
| 217 | /* |
| 218 | * Zero struct epoll_event to ensure all representations of its |
| 219 | * union are zeroed. |
| 220 | */ |
| 221 | memset(&ev, 0, sizeof(ev)); |
| 222 | ev.events = req_events; |
| 223 | ev.data.fd = fd; |
| 224 | |
| 225 | ret = epoll_ctl(events->epfd, EPOLL_CTL_MOD, fd, &ev); |
| 226 | if (ret < 0) { |
| 227 | switch (errno) { |
| 228 | case ENOENT: |
| 229 | case EPERM: |
| 230 | /* Print PERROR and goto end not failing. Show must go on. */ |
| 231 | PERROR("epoll_ctl MOD"); |
| 232 | goto end; |
| 233 | default: |
| 234 | PERROR("epoll_ctl MOD fatal"); |
| 235 | goto error; |
| 236 | } |
| 237 | } |
| 238 | |
| 239 | end: |
| 240 | return 0; |
| 241 | |
| 242 | error: |
| 243 | return -1; |
| 244 | } |
| 245 | |
| 246 | /* |
| 247 | * Wait on epoll set. This is a blocking call of timeout value. |
| 248 | */ |
| 249 | LTTNG_HIDDEN |
| 250 | int compat_epoll_wait(struct lttng_poll_event *events, int timeout, |
| 251 | bool interruptible) |
| 252 | { |
| 253 | int ret; |
| 254 | uint32_t new_size; |
| 255 | |
| 256 | if (events == NULL || events->events == NULL) { |
| 257 | ERR("Wrong arguments in compat_epoll_wait"); |
| 258 | goto error; |
| 259 | } |
| 260 | |
| 261 | if (events->nb_fd == 0) { |
| 262 | errno = EINVAL; |
| 263 | return -1; |
| 264 | } |
| 265 | |
| 266 | /* |
| 267 | * Resize if needed before waiting. We could either expand the array or |
| 268 | * shrink it down. It's important to note that after this step, we are |
| 269 | * ensured that the events argument of the epoll_wait call will be large |
| 270 | * enough to hold every possible returned events. |
| 271 | */ |
| 272 | new_size = 1U << utils_get_count_order_u32(events->nb_fd); |
| 273 | if (new_size != events->alloc_size && new_size >= events->init_size) { |
| 274 | ret = resize_poll_event(events, new_size); |
| 275 | if (ret < 0) { |
| 276 | /* ENOMEM problem at this point. */ |
| 277 | goto error; |
| 278 | } |
| 279 | } |
| 280 | |
| 281 | do { |
| 282 | ret = epoll_wait(events->epfd, events->events, events->nb_fd, timeout); |
| 283 | } while (!interruptible && ret == -1 && errno == EINTR); |
| 284 | if (ret < 0) { |
| 285 | if (errno != EINTR) { |
| 286 | PERROR("epoll_wait"); |
| 287 | } |
| 288 | goto error; |
| 289 | } |
| 290 | |
| 291 | /* |
| 292 | * Since the returned events are set sequentially in the "events" structure |
| 293 | * we only need to return the epoll_wait value and iterate over it. |
| 294 | */ |
| 295 | return ret; |
| 296 | |
| 297 | error: |
| 298 | return -1; |
| 299 | } |
| 300 | |
| 301 | /* |
| 302 | * Setup poll set maximum size. |
| 303 | */ |
| 304 | LTTNG_HIDDEN |
| 305 | int compat_epoll_set_max_size(void) |
| 306 | { |
| 307 | int ret, fd, retval = 0; |
| 308 | ssize_t size_ret; |
| 309 | char buf[64]; |
| 310 | |
| 311 | fd = open(COMPAT_EPOLL_PROC_PATH, O_RDONLY); |
| 312 | if (fd < 0) { |
| 313 | /* |
| 314 | * Failing on opening [1] is not an error per see. [1] was |
| 315 | * introduced in Linux 2.6.28 but epoll is available since |
| 316 | * 2.5.44. Hence, goto end and set a default value without |
| 317 | * setting an error return value. |
| 318 | * |
| 319 | * [1] /proc/sys/fs/epoll/max_user_watches |
| 320 | */ |
| 321 | retval = 0; |
| 322 | goto end; |
| 323 | } |
| 324 | |
| 325 | size_ret = lttng_read(fd, buf, sizeof(buf)); |
| 326 | /* |
| 327 | * Allow reading a file smaller than buf, but keep space for |
| 328 | * final \0. |
| 329 | */ |
| 330 | if (size_ret < 0 || size_ret >= sizeof(buf)) { |
| 331 | PERROR("read set max size"); |
| 332 | retval = -1; |
| 333 | goto end_read; |
| 334 | } |
| 335 | buf[size_ret] = '\0'; |
| 336 | poll_max_size = atoi(buf); |
| 337 | end_read: |
| 338 | ret = close(fd); |
| 339 | if (ret) { |
| 340 | PERROR("close"); |
| 341 | } |
| 342 | end: |
| 343 | if (!poll_max_size) { |
| 344 | poll_max_size = DEFAULT_POLL_SIZE; |
| 345 | } |
| 346 | DBG("epoll set max size is %d", poll_max_size); |
| 347 | return retval; |
| 348 | } |