| 1 | // SPDX-License-Identifier: GPL-3.0-or-later |
| 2 | |
| 3 | #include "../libnetdata.h" |
| 4 | #include "nd-poll.h" |
| 5 | |
| 6 | #ifndef POLLRDHUP |
| 7 | #define POLLRDHUP 0 |
| 8 | #endif |
| 9 | |
| 10 | #if defined(OS_LINUX) |
| 11 | #include <sys/epoll.h> |
| 12 | |
| 13 | struct fd_info { |
| 14 | uint32_t events; |
| 15 | uint32_t last_served; |
| 16 | const void *data; |
| 17 | }; |
| 18 | |
| 19 | DEFINE_JUDYL_TYPED(POINTERS, struct fd_info *); |
| 20 | |
| 21 | #define MAX_EVENTS_PER_CALL 100 |
| 22 | |
| 23 | // Event poll context |
| 24 | struct nd_poll_t { |
| 25 | int epoll_fd; |
| 26 | |
| 27 | struct epoll_event ev[MAX_EVENTS_PER_CALL]; |
| 28 | size_t last_pos; |
| 29 | size_t used; |
| 30 | |
| 31 | POINTERS_JudyLSet pointers; // Judy array to store user data |
| 32 | |
| 33 | uint32_t nfds; // the number of sockets we have |
| 34 | uint32_t iteration_counter; |
| 35 | }; |
| 36 | |
| 37 | // Initialize the event poll context |
| 38 | nd_poll_t *nd_poll_create() { |
| 39 | nd_poll_t *ndpl = callocz(1, sizeof(nd_poll_t)); |
| 40 | |
| 41 | ndpl->epoll_fd = epoll_create1(0); |
| 42 | if (ndpl->epoll_fd < 0) { |
| 43 | freez(ndpl); |
| 44 | return NULL; |
| 45 | } |
| 46 | |
| 47 | return ndpl; |
| 48 | } |
| 49 | |
| 50 | static inline uint32_t nd_poll_events_to_epoll_events(nd_poll_event_t events) { |
| 51 | uint32_t pevents = EPOLLERR | EPOLLHUP; |
| 52 | if (events & ND_POLL_READ) pevents |= EPOLLIN; |
| 53 | if (events & ND_POLL_WRITE) pevents |= EPOLLOUT; |
| 54 | return pevents; |
| 55 | } |
| 56 | |
| 57 | static inline nd_poll_event_t nd_poll_events_from_epoll_events(uint32_t events) { |
| 58 | nd_poll_event_t nd_poll_events = ND_POLL_NONE; |
| 59 | |
| 60 | if (events & (EPOLLIN|EPOLLPRI|EPOLLRDNORM|EPOLLRDBAND)) |
| 61 | nd_poll_events |= ND_POLL_READ; |
| 62 | |
| 63 | if (events & (EPOLLOUT|EPOLLWRNORM|EPOLLWRBAND)) |
| 64 | nd_poll_events |= ND_POLL_WRITE; |
| 65 | |
| 66 | if (events & EPOLLERR) |
| 67 | nd_poll_events |= ND_POLL_ERROR; |
| 68 | |
| 69 | if (events & (EPOLLHUP|EPOLLRDHUP)) |
| 70 | nd_poll_events |= ND_POLL_HUP; |
| 71 | |
| 72 | return nd_poll_events; |
| 73 | } |
| 74 | |
| 75 | // Add a file descriptor to the event poll |
| 76 | bool nd_poll_add(nd_poll_t *ndpl, int fd, nd_poll_event_t events, const void *data) { |
| 77 | internal_fatal(!data, "nd_poll() does not support NULL data pointers"); |
| 78 | |
| 79 | struct fd_info *fdi = mallocz(sizeof(*fdi)); |
| 80 | fdi->data = data; |
| 81 | fdi->last_served = 0; |
| 82 | fdi->events = nd_poll_events_to_epoll_events(events); |
| 83 | |
| 84 | if(POINTERS_GET(&ndpl->pointers, fd) || !POINTERS_SET(&ndpl->pointers, fd, fdi)) { |
| 85 | freez(fdi); |
| 86 | return false; |
| 87 | } |
| 88 | |
| 89 | struct epoll_event ev = { |
| 90 | .events = fdi->events, |
| 91 | .data.fd = fd, |
| 92 | }; |
| 93 | |
| 94 | bool rc = epoll_ctl(ndpl->epoll_fd, EPOLL_CTL_ADD, fd, &ev) == 0; |
| 95 | if(rc) |
| 96 | ndpl->nfds++; |
| 97 | else { |
| 98 | POINTERS_DEL(&ndpl->pointers, fd); |
| 99 | freez(fdi); |
| 100 | } |
| 101 | |
| 102 | internal_fatal(!rc, "epoll_ctl() failed"); |
| 103 | |
| 104 | return rc; |
| 105 | } |
| 106 | |
| 107 | // Remove a file descriptor from the event poll |
| 108 | bool nd_poll_del(nd_poll_t *ndpl, int fd) { |
| 109 | struct fd_info *fdi = POINTERS_GET(&ndpl->pointers, fd); |
| 110 | if(!fdi) return false; |
| 111 | |
| 112 | POINTERS_DEL(&ndpl->pointers, fd); |
| 113 | freez(fdi); |
| 114 | |
| 115 | ndpl->nfds--; // we can't check for success/failure here, because epoll() removes fds when they are closed |
| 116 | bool rc = epoll_ctl(ndpl->epoll_fd, EPOLL_CTL_DEL, fd, NULL) == 0; |
| 117 | internal_error(!rc, "epoll_ctl() failed (is the socket already closed)"); // this is ok if the socket is already closed |
| 118 | return rc; |
| 119 | } |
| 120 | |
| 121 | // Update an existing file descriptor in the event poll |
| 122 | ALWAYS_INLINE_HOT_FLATTEN |
| 123 | bool nd_poll_upd(nd_poll_t *ndpl, int fd, nd_poll_event_t events) { |
| 124 | struct fd_info *fdi = POINTERS_GET(&ndpl->pointers, fd); |
| 125 | if(!fdi) return false; |
| 126 | |
| 127 | fdi->events = nd_poll_events_to_epoll_events(events); |
| 128 | |
| 129 | struct epoll_event ev = { |
| 130 | .events = fdi->events, |
| 131 | .data.fd = fd, |
| 132 | }; |
| 133 | bool rc = epoll_ctl(ndpl->epoll_fd, EPOLL_CTL_MOD, fd, &ev) == 0; |
| 134 | internal_fatal(!rc, "epoll_ctl() failed"); // this may happen if fd is closed |
| 135 | return rc; |
| 136 | } |
| 137 | |
| 138 | static inline bool nd_poll_get_next_event(nd_poll_t *ndpl, nd_poll_result_t *result) { |
| 139 | while(ndpl->last_pos < ndpl->used) { |
| 140 | struct fd_info *fdi = POINTERS_GET(&ndpl->pointers, ndpl->ev[ndpl->last_pos].data.fd); |
| 141 | |
| 142 | // Skip events that have been invalidated by nd_poll_del() |
| 143 | if(!fdi || !fdi->data) { |
| 144 | ndpl->last_pos++; |
| 145 | continue; |
| 146 | } |
| 147 | |
| 148 | *result = (nd_poll_result_t){ |
| 149 | .events = nd_poll_events_from_epoll_events(ndpl->ev[ndpl->last_pos].events & fdi->events), |
| 150 | .data = fdi->data, |
| 151 | }; |
| 152 | |
| 153 | ndpl->last_pos++; |
| 154 | |
| 155 | if(!result->events) |
| 156 | // nd_poll_upd() may have removed some flags since we got this |
| 157 | continue; |
| 158 | |
| 159 | fdi->last_served = ndpl->iteration_counter; |
| 160 | return true; |
| 161 | } |
| 162 | |
| 163 | return false; |
| 164 | } |
| 165 | |
| 166 | typedef struct { |
| 167 | struct epoll_event event; |
| 168 | uint32_t last_served; |
| 169 | } sortable_event_t; |
| 170 | |
| 171 | static int compare_last_served(const void *a, const void *b) { |
| 172 | const sortable_event_t *ev_a = (const sortable_event_t *)a; |
| 173 | const sortable_event_t *ev_b = (const sortable_event_t *)b; |
| 174 | |
| 175 | if (ev_a->last_served < ev_b->last_served) |
| 176 | return -1; |
| 177 | if (ev_a->last_served > ev_b->last_served) |
| 178 | return 1; |
| 179 | |
| 180 | return 0; |
| 181 | } |
| 182 | |
| 183 | static void sort_events(nd_poll_t *ndpl) { |
| 184 | if(ndpl->used <= 1) return; |
| 185 | |
| 186 | internal_fatal(ndpl->used > MAX_EVENTS_PER_CALL, "ndpl->used exceeds MAX_EVENTS_PER_CALL"); |
| 187 | |
| 188 | sortable_event_t sortable_array[MAX_EVENTS_PER_CALL]; |
| 189 | for (size_t i = 0; i < ndpl->used; ++i) { |
| 190 | struct fd_info *fdi = POINTERS_GET(&ndpl->pointers, ndpl->ev[i].data.fd); |
| 191 | sortable_array[i] = (sortable_event_t){ |
| 192 | .event = ndpl->ev[i], |
| 193 | .last_served = fdi ? fdi->last_served : UINT32_MAX, |
| 194 | }; |
| 195 | } |
| 196 | |
| 197 | qsort(sortable_array, ndpl->used, sizeof(sortable_event_t), compare_last_served); |
| 198 | |
| 199 | // Reorder `ndpl->ev` based on the sorted order |
| 200 | for (size_t i = 0; i < ndpl->used; ++i) |
| 201 | ndpl->ev[i] = sortable_array[i].event; |
| 202 | } |
| 203 | |
| 204 | // Wait for events |
| 205 | ALWAYS_INLINE_HOT_FLATTEN |
| 206 | int nd_poll_wait(nd_poll_t *ndpl, int timeout_ms, nd_poll_result_t *result) { |
| 207 | ndpl->iteration_counter++; |
| 208 | |
| 209 | if(nd_poll_get_next_event(ndpl, result)) |
| 210 | return 1; |
| 211 | |
| 212 | do { |
| 213 | errno_clear(); |
| 214 | ndpl->last_pos = 0; |
| 215 | ndpl->used = 0; |
| 216 | |
| 217 | int n = epoll_wait(ndpl->epoll_fd, &ndpl->ev[0], _countof(ndpl->ev), timeout_ms); |
| 218 | |
| 219 | if(unlikely(n <= 0)) { |
| 220 | if(n == 0) { |
| 221 | result->events = ND_POLL_TIMEOUT; |
| 222 | result->data = NULL; |
| 223 | return 0; |
| 224 | } |
| 225 | |
| 226 | if(errno == EINTR || errno == EAGAIN) |
| 227 | continue; |
| 228 | |
| 229 | result->events = ND_POLL_POLL_FAILED; |
| 230 | result->data = NULL; |
| 231 | return -1; |
| 232 | } |
| 233 | |
| 234 | ndpl->used = n; |
| 235 | ndpl->last_pos = 0; |
| 236 | sort_events(ndpl); |
| 237 | if (nd_poll_get_next_event(ndpl, result)) |
| 238 | return 1; |
| 239 | |
| 240 | internal_fatal(true, "nd_poll_get_next_event() should have 1 event!"); |
| 241 | } while(true); |
| 242 | } |
| 243 | |
| 244 | static void nd_poll_free_callback(Word_t fd __maybe_unused, struct fd_info *fdi, void *data __maybe_unused) { |
| 245 | freez(fdi); |
| 246 | } |
| 247 | |
| 248 | // Destroy the event poll context |
| 249 | void nd_poll_destroy(nd_poll_t *ndpl) { |
| 250 | if (ndpl) { |
| 251 | close(ndpl->epoll_fd); |
| 252 | POINTERS_FREE(&ndpl->pointers, nd_poll_free_callback, NULL); |
| 253 | freez(ndpl); |
| 254 | } |
| 255 | } |
| 256 | #else |
| 257 | |
| 258 | DEFINE_JUDYL_TYPED(POINTERS, const void *); |
| 259 | |
| 260 | struct nd_poll_t { |
| 261 | struct pollfd *fds; // Array of file descriptors |
| 262 | nfds_t nfds; // Number of active file descriptors |
| 263 | nfds_t capacity; // Allocated capacity for `fds` array |
| 264 | nfds_t last_pos; |
| 265 | POINTERS_JudyLSet pointers; // Judy array to store user data |
| 266 | }; |
| 267 | |
| 268 | #define INITIAL_CAPACITY 4 |
| 269 | |
| 270 | // Initialize the event poll context |
| 271 | nd_poll_t *nd_poll_create() { |
| 272 | nd_poll_t *ndpl = callocz(1, sizeof(nd_poll_t)); |
| 273 | ndpl->fds = mallocz(INITIAL_CAPACITY * sizeof(struct pollfd)); |
| 274 | ndpl->nfds = 0; |
| 275 | ndpl->capacity = INITIAL_CAPACITY; |
| 276 | |
| 277 | POINTERS_INIT(&ndpl->pointers); |
| 278 | |
| 279 | return ndpl; |
| 280 | } |
| 281 | |
| 282 | // Ensure capacity for adding new file descriptors |
| 283 | static void ensure_capacity(nd_poll_t *ndpl) { |
| 284 | if (ndpl->nfds < ndpl->capacity) return; |
| 285 | |
| 286 | nfds_t new_capacity = ndpl->capacity * 2; |
| 287 | struct pollfd *new_fds = reallocz(ndpl->fds, new_capacity * sizeof(struct pollfd)); |
| 288 | |
| 289 | ndpl->fds = new_fds; |
| 290 | ndpl->capacity = new_capacity; |
| 291 | } |
| 292 | |
| 293 | static inline short int nd_poll_events_to_poll_events(nd_poll_event_t events) { |
| 294 | short int pevents = POLLERR | POLLHUP | POLLNVAL; |
| 295 | if (events & ND_POLL_READ) pevents |= POLLIN; |
| 296 | if (events & ND_POLL_WRITE) pevents |= POLLOUT; |
| 297 | return pevents; |
| 298 | } |
| 299 | |
| 300 | static inline nd_poll_event_t nd_poll_events_from_poll_revents(short int events) { |
| 301 | nd_poll_event_t nd_poll_events = ND_POLL_NONE; |
| 302 | |
| 303 | if (events & (POLLIN|POLLPRI|POLLRDNORM|POLLRDBAND)) |
| 304 | nd_poll_events |= ND_POLL_READ; |
| 305 | |
| 306 | if (events & (POLLOUT|POLLWRNORM|POLLWRBAND)) |
| 307 | nd_poll_events |= ND_POLL_WRITE; |
| 308 | |
| 309 | if (events & POLLERR) |
| 310 | nd_poll_events |= ND_POLL_ERROR; |
| 311 | |
| 312 | if (events & (POLLHUP|POLLRDHUP)) |
| 313 | nd_poll_events |= ND_POLL_HUP; |
| 314 | |
| 315 | if (events & (POLLNVAL)) |
| 316 | nd_poll_events |= ND_POLL_INVALID; |
| 317 | |
| 318 | return nd_poll_events; |
| 319 | } |
| 320 | |
| 321 | bool nd_poll_add(nd_poll_t *ndpl, int fd, nd_poll_event_t events, const void *data) { |
| 322 | internal_fatal(POINTERS_GET(&ndpl->pointers, fd) != NULL, "File descriptor %d is already served - cannot add", fd); |
| 323 | |
| 324 | if(POINTERS_GET(&ndpl->pointers, fd) || !POINTERS_SET(&ndpl->pointers, fd, data)) |
| 325 | return false; |
| 326 | |
| 327 | ensure_capacity(ndpl); |
| 328 | struct pollfd *pfd = &ndpl->fds[ndpl->nfds++]; |
| 329 | pfd->fd = fd; |
| 330 | pfd->events = nd_poll_events_to_poll_events(events); |
| 331 | pfd->revents = 0; |
| 332 | |
| 333 | return true; |
| 334 | } |
| 335 | |
| 336 | // Remove a file descriptor from the event poll |
| 337 | bool nd_poll_del(nd_poll_t *ndpl, int fd) { |
| 338 | if(!POINTERS_DEL(&ndpl->pointers, fd)) |
| 339 | return false; |
| 340 | |
| 341 | for (nfds_t i = 0; i < ndpl->nfds; i++) { |
| 342 | if (ndpl->fds[i].fd == fd) { |
| 343 | |
| 344 | // Remove the file descriptor by shifting the array |
| 345 | memmove(&ndpl->fds[i], &ndpl->fds[i + 1], (ndpl->nfds - i - 1) * sizeof(struct pollfd)); |
| 346 | ndpl->nfds--; |
| 347 | |
| 348 | if(i < ndpl->last_pos) |
| 349 | ndpl->last_pos--; |
| 350 | |
| 351 | return true; |
| 352 | } |
| 353 | } |
| 354 | |
| 355 | return false; // File descriptor not found |
| 356 | } |
| 357 | |
| 358 | // Update an existing file descriptor in the event poll |
| 359 | ALWAYS_INLINE_HOT_FLATTEN |
| 360 | bool nd_poll_upd(nd_poll_t *ndpl, int fd, nd_poll_event_t events) { |
| 361 | for (nfds_t i = 0; i < ndpl->nfds; i++) { |
| 362 | if (ndpl->fds[i].fd == fd) { |
| 363 | struct pollfd *pfd = &ndpl->fds[i]; |
| 364 | pfd->events = nd_poll_events_to_poll_events(events); |
| 365 | return true; |
| 366 | } |
| 367 | } |
| 368 | |
| 369 | // File descriptor not found |
| 370 | return false; |
| 371 | } |
| 372 | |
| 373 | static inline bool nd_poll_get_next_event(nd_poll_t *ndpl, nd_poll_result_t *result) { |
| 374 | for (nfds_t i = ndpl->last_pos; i < ndpl->nfds; i++) { |
| 375 | if (ndpl->fds[i].revents != 0) { |
| 376 | |
| 377 | result->data = POINTERS_GET(&ndpl->pointers, ndpl->fds[i].fd); |
| 378 | if(!result->data) |
| 379 | continue; |
| 380 | |
| 381 | result->events = nd_poll_events_from_poll_revents(ndpl->fds[i].revents & ndpl->fds[i].events); |
| 382 | if(!result->events) |
| 383 | // nd_poll_upd() may have removed some flags since we got this |
| 384 | continue; |
| 385 | |
| 386 | ndpl->fds[i].revents = 0; |
| 387 | |
| 388 | ndpl->last_pos = i + 1; |
| 389 | return true; |
| 390 | } |
| 391 | } |
| 392 | |
| 393 | ndpl->last_pos = ndpl->nfds; |
| 394 | return false; |
| 395 | } |
| 396 | |
| 397 | // Rotate the fds array to prevent starvation |
| 398 | static inline void rotate_fds(nd_poll_t *ndpl) { |
| 399 | if (ndpl->nfds == 0 || ndpl->nfds == 1) |
| 400 | return; // No rotation needed for empty or single-entry arrays |
| 401 | |
| 402 | struct pollfd first = ndpl->fds[0]; |
| 403 | memmove(&ndpl->fds[0], &ndpl->fds[1], (ndpl->nfds - 1) * sizeof(struct pollfd)); |
| 404 | ndpl->fds[ndpl->nfds - 1] = first; |
| 405 | } |
| 406 | |
| 407 | // Wait for events |
| 408 | ALWAYS_INLINE_HOT_FLATTEN |
| 409 | int nd_poll_wait(nd_poll_t *ndpl, int timeout_ms, nd_poll_result_t *result) { |
| 410 | if (nd_poll_get_next_event(ndpl, result)) |
| 411 | return 1; // Return immediately if there's a pending event |
| 412 | |
| 413 | do { |
| 414 | errno_clear(); |
| 415 | ndpl->last_pos = 0; |
| 416 | rotate_fds(ndpl); // Rotate the array on every wait |
| 417 | int ret = poll(ndpl->fds, ndpl->nfds, timeout_ms); |
| 418 | |
| 419 | if(unlikely(ret <= 0)) { |
| 420 | if(ret == 0) { |
| 421 | result->events = ND_POLL_TIMEOUT; |
| 422 | result->data = NULL; |
| 423 | return 0; |
| 424 | } |
| 425 | |
| 426 | if(errno == EAGAIN || errno == EINTR) |
| 427 | continue; |
| 428 | |
| 429 | result->events = ND_POLL_POLL_FAILED; |
| 430 | result->data = NULL; |
| 431 | return -1; |
| 432 | } |
| 433 | |
| 434 | // Process the next event |
| 435 | if (nd_poll_get_next_event(ndpl, result)) |
| 436 | return 1; |
| 437 | |
| 438 | internal_fatal(true, "nd_poll_get_next_event() should have 1 event!"); |
| 439 | } while (true); |
| 440 | } |
| 441 | |
| 442 | // Destroy the event poll context |
| 443 | void nd_poll_destroy(nd_poll_t *ndpl) { |
| 444 | if (ndpl) { |
| 445 | free(ndpl->fds); |
| 446 | POINTERS_FREE(&ndpl->pointers, NULL, NULL); |
| 447 | freez(ndpl); |
| 448 | } |
| 449 | } |
| 450 | |
| 451 | #endif |