Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 1 | /* |
| 2 | * File descriptors management functions. |
| 3 | * |
| 4 | * Copyright 2000-2006 Willy Tarreau <w@1wt.eu> |
| 5 | * |
| 6 | * This program is free software; you can redistribute it and/or |
| 7 | * modify it under the terms of the GNU General Public License |
| 8 | * as published by the Free Software Foundation; either version |
| 9 | * 2 of the License, or (at your option) any later version. |
| 10 | * |
| 11 | */ |
| 12 | |
| 13 | /* |
| 14 | * FIXME: |
| 15 | * - we still use 'listeners' to check whether we want to stop or not. |
| 16 | * - the various pollers should be moved to other external files, possibly |
| 17 | * dynamic libs. |
| 18 | * - merge event_cli_read() and event_srv_read(). The difference is res_*, |
| 19 | * buffer (at the beginning) and timeouts (at the end). |
| 20 | * => event_tcp_read(). It may be called from event_accept(). |
| 21 | * - extract the connect code from event_srv_write() |
| 22 | * => event_tcp_connect(). It must then call event_write(). |
| 23 | * - merge the remaining event_cli_write() and event_srv_write() |
| 24 | * => single event_tcp_write(). Check buffer, fd_state, res*, and timeouts. |
| 25 | * |
| 26 | */ |
| 27 | |
| 28 | #include <unistd.h> |
| 29 | #include <sys/time.h> |
| 30 | #include <sys/types.h> |
| 31 | |
Willy Tarreau | 2dd0d47 | 2006-06-29 17:53:05 +0200 | [diff] [blame] | 32 | #include <common/compat.h> |
| 33 | #include <common/config.h> |
| 34 | #include <common/time.h> |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 35 | |
| 36 | #include <types/fd.h> |
| 37 | #include <types/global.h> |
| 38 | |
| 39 | #include <proto/polling.h> |
| 40 | #include <proto/task.h> |
| 41 | |
| 42 | struct fdtab *fdtab = NULL; /* array of all the file descriptors */ |
| 43 | int maxfd; /* # of the highest fd + 1 */ |
| 44 | int totalconn; /* total # of terminated sessions */ |
| 45 | int actconn; /* # of active sessions */ |
| 46 | |
| 47 | fd_set *StaticReadEvent, *StaticWriteEvent; |
| 48 | int cfg_polling_mechanism = 0; /* POLL_USE_{SELECT|POLL|EPOLL} */ |
| 49 | |
| 50 | |
| 51 | /****************************** |
| 52 | * pollers |
| 53 | ******************************/ |
| 54 | |
| 55 | |
| 56 | /* |
| 57 | * FIXME: this is dirty, but at the moment, there's no other solution to remove |
| 58 | * the old FDs from outside the loop. Perhaps we should export a global 'poll' |
| 59 | * structure with pointers to functions such as init_fd() and close_fd(), plus |
| 60 | * a private structure with several pointers to places such as below. |
| 61 | */ |
| 62 | |
| 63 | #if defined(ENABLE_EPOLL) |
| 64 | fd_set *PrevReadEvent = NULL, *PrevWriteEvent = NULL; |
| 65 | |
| 66 | #if defined(USE_MY_EPOLL) |
Willy Tarreau | 2dd0d47 | 2006-06-29 17:53:05 +0200 | [diff] [blame] | 67 | #include <errno.h> |
| 68 | #include <sys/syscall.h> |
Willy Tarreau | baaee00 | 2006-06-26 02:48:02 +0200 | [diff] [blame] | 69 | _syscall1 (int, epoll_create, int, size); |
| 70 | _syscall4 (int, epoll_ctl, int, epfd, int, op, int, fd, struct epoll_event *, event); |
| 71 | _syscall4 (int, epoll_wait, int, epfd, struct epoll_event *, events, int, maxevents, int, timeout); |
| 72 | #endif |
| 73 | |
| 74 | /* |
| 75 | * Main epoll() loop. |
| 76 | * does 3 actions : |
| 77 | * 0 (POLL_LOOP_ACTION_INIT) : initializes necessary private structures |
| 78 | * 1 (POLL_LOOP_ACTION_RUN) : runs the loop |
| 79 | * 2 (POLL_LOOP_ACTION_CLEAN) : cleans up |
| 80 | * |
| 81 | * returns 0 if initialization failed, !0 otherwise. |
| 82 | */ |
| 83 | |
| 84 | int epoll_loop(int action) |
| 85 | { |
| 86 | int next_time; |
| 87 | int status; |
| 88 | int fd; |
| 89 | |
| 90 | int fds, count; |
| 91 | int pr, pw, sr, sw; |
| 92 | unsigned rn, ro, wn, wo; /* read new, read old, write new, write old */ |
| 93 | struct epoll_event ev; |
| 94 | |
| 95 | /* private data */ |
| 96 | static struct epoll_event *epoll_events = NULL; |
| 97 | static int epoll_fd; |
| 98 | |
| 99 | if (action == POLL_LOOP_ACTION_INIT) { |
| 100 | epoll_fd = epoll_create(global.maxsock + 1); |
| 101 | if (epoll_fd < 0) |
| 102 | return 0; |
| 103 | else { |
| 104 | epoll_events = (struct epoll_event*) |
| 105 | calloc(1, sizeof(struct epoll_event) * global.maxsock); |
| 106 | PrevReadEvent = (fd_set *) |
| 107 | calloc(1, sizeof(fd_set) * (global.maxsock + FD_SETSIZE - 1) / FD_SETSIZE); |
| 108 | PrevWriteEvent = (fd_set *) |
| 109 | calloc(1, sizeof(fd_set) * (global.maxsock + FD_SETSIZE - 1) / FD_SETSIZE); |
| 110 | } |
| 111 | return 1; |
| 112 | } |
| 113 | else if (action == POLL_LOOP_ACTION_CLEAN) { |
| 114 | if (PrevWriteEvent) free(PrevWriteEvent); |
| 115 | if (PrevReadEvent) free(PrevReadEvent); |
| 116 | if (epoll_events) free(epoll_events); |
| 117 | close(epoll_fd); |
| 118 | epoll_fd = 0; |
| 119 | return 1; |
| 120 | } |
| 121 | |
| 122 | /* OK, it's POLL_LOOP_ACTION_RUN */ |
| 123 | |
| 124 | tv_now(&now); |
| 125 | |
| 126 | while (1) { |
| 127 | next_time = process_runnable_tasks(); |
| 128 | |
| 129 | /* stop when there's no connection left and we don't allow them anymore */ |
| 130 | if (!actconn && listeners == 0) |
| 131 | break; |
| 132 | |
| 133 | for (fds = 0; (fds << INTBITS) < maxfd; fds++) { |
| 134 | |
| 135 | rn = ((int*)StaticReadEvent)[fds]; ro = ((int*)PrevReadEvent)[fds]; |
| 136 | wn = ((int*)StaticWriteEvent)[fds]; wo = ((int*)PrevWriteEvent)[fds]; |
| 137 | |
| 138 | if ((ro^rn) | (wo^wn)) { |
| 139 | for (count = 0, fd = fds << INTBITS; count < (1<<INTBITS) && fd < maxfd; count++, fd++) { |
| 140 | #define FDSETS_ARE_INT_ALIGNED |
| 141 | #ifdef FDSETS_ARE_INT_ALIGNED |
| 142 | |
| 143 | #define WE_REALLY_NOW_THAT_FDSETS_ARE_INTS |
| 144 | #ifdef WE_REALLY_NOW_THAT_FDSETS_ARE_INTS |
| 145 | pr = (ro >> count) & 1; |
| 146 | pw = (wo >> count) & 1; |
| 147 | sr = (rn >> count) & 1; |
| 148 | sw = (wn >> count) & 1; |
| 149 | #else |
| 150 | pr = FD_ISSET(fd&((1<<INTBITS)-1), (typeof(fd_set*))&ro); |
| 151 | pw = FD_ISSET(fd&((1<<INTBITS)-1), (typeof(fd_set*))&wo); |
| 152 | sr = FD_ISSET(fd&((1<<INTBITS)-1), (typeof(fd_set*))&rn); |
| 153 | sw = FD_ISSET(fd&((1<<INTBITS)-1), (typeof(fd_set*))&wn); |
| 154 | #endif |
| 155 | #else |
| 156 | pr = FD_ISSET(fd, PrevReadEvent); |
| 157 | pw = FD_ISSET(fd, PrevWriteEvent); |
| 158 | sr = FD_ISSET(fd, StaticReadEvent); |
| 159 | sw = FD_ISSET(fd, StaticWriteEvent); |
| 160 | #endif |
| 161 | if (!((sr^pr) | (sw^pw))) |
| 162 | continue; |
| 163 | |
| 164 | ev.events = (sr ? EPOLLIN : 0) | (sw ? EPOLLOUT : 0); |
| 165 | ev.data.fd = fd; |
| 166 | |
| 167 | #ifdef EPOLL_CTL_MOD_WORKAROUND |
| 168 | /* I encountered a rarely reproducible problem with |
| 169 | * EPOLL_CTL_MOD where a modified FD (systematically |
| 170 | * the one in epoll_events[0], fd#7) would sometimes |
| 171 | * be set EPOLL_OUT while asked for a read ! This is |
| 172 | * with the 2.4 epoll patch. The workaround is to |
| 173 | * delete then recreate in case of modification. |
| 174 | * This is in 2.4 up to epoll-lt-0.21 but not in 2.6 |
| 175 | * nor RHEL kernels. |
| 176 | */ |
| 177 | |
| 178 | if ((pr | pw) && fdtab[fd].state != FD_STCLOSE) |
| 179 | epoll_ctl(epoll_fd, EPOLL_CTL_DEL, fd, &ev); |
| 180 | |
| 181 | if ((sr | sw)) |
| 182 | epoll_ctl(epoll_fd, EPOLL_CTL_ADD, fd, &ev); |
| 183 | #else |
| 184 | if ((pr | pw)) { |
| 185 | /* the file-descriptor already exists... */ |
| 186 | if ((sr | sw)) { |
| 187 | /* ...and it will still exist */ |
| 188 | if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, fd, &ev) < 0) { |
| 189 | // perror("epoll_ctl(MOD)"); |
| 190 | // exit(1); |
| 191 | } |
| 192 | } else { |
| 193 | /* ...and it will be removed */ |
| 194 | if (fdtab[fd].state != FD_STCLOSE && |
| 195 | epoll_ctl(epoll_fd, EPOLL_CTL_DEL, fd, &ev) < 0) { |
| 196 | // perror("epoll_ctl(DEL)"); |
| 197 | // exit(1); |
| 198 | } |
| 199 | } |
| 200 | } else { |
| 201 | /* the file-descriptor did not exist, let's add it */ |
| 202 | if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, fd, &ev) < 0) { |
| 203 | // perror("epoll_ctl(ADD)"); |
| 204 | // exit(1); |
| 205 | } |
| 206 | } |
| 207 | #endif // EPOLL_CTL_MOD_WORKAROUND |
| 208 | } |
| 209 | ((int*)PrevReadEvent)[fds] = rn; |
| 210 | ((int*)PrevWriteEvent)[fds] = wn; |
| 211 | } |
| 212 | } |
| 213 | |
| 214 | /* now let's wait for events */ |
| 215 | status = epoll_wait(epoll_fd, epoll_events, maxfd, next_time); |
| 216 | tv_now(&now); |
| 217 | |
| 218 | for (count = 0; count < status; count++) { |
| 219 | fd = epoll_events[count].data.fd; |
| 220 | |
| 221 | if (FD_ISSET(fd, StaticReadEvent)) { |
| 222 | if (fdtab[fd].state == FD_STCLOSE) |
| 223 | continue; |
| 224 | if (epoll_events[count].events & ( EPOLLIN | EPOLLERR | EPOLLHUP )) |
| 225 | fdtab[fd].read(fd); |
| 226 | } |
| 227 | |
| 228 | if (FD_ISSET(fd, StaticWriteEvent)) { |
| 229 | if (fdtab[fd].state == FD_STCLOSE) |
| 230 | continue; |
| 231 | if (epoll_events[count].events & ( EPOLLOUT | EPOLLERR | EPOLLHUP )) |
| 232 | fdtab[fd].write(fd); |
| 233 | } |
| 234 | } |
| 235 | } |
| 236 | return 1; |
| 237 | } |
| 238 | #endif |
| 239 | |
| 240 | |
| 241 | |
| 242 | #if defined(ENABLE_POLL) |
| 243 | /* |
| 244 | * Main poll() loop. |
| 245 | * does 3 actions : |
| 246 | * 0 (POLL_LOOP_ACTION_INIT) : initializes necessary private structures |
| 247 | * 1 (POLL_LOOP_ACTION_RUN) : runs the loop |
| 248 | * 2 (POLL_LOOP_ACTION_CLEAN) : cleans up |
| 249 | * |
| 250 | * returns 0 if initialization failed, !0 otherwise. |
| 251 | */ |
| 252 | |
| 253 | int poll_loop(int action) |
| 254 | { |
| 255 | int next_time; |
| 256 | int status; |
| 257 | int fd, nbfd; |
| 258 | |
| 259 | int fds, count; |
| 260 | int sr, sw; |
| 261 | unsigned rn, wn; /* read new, write new */ |
| 262 | |
| 263 | /* private data */ |
| 264 | static struct pollfd *poll_events = NULL; |
| 265 | |
| 266 | if (action == POLL_LOOP_ACTION_INIT) { |
| 267 | poll_events = (struct pollfd*) |
| 268 | calloc(1, sizeof(struct pollfd) * global.maxsock); |
| 269 | return 1; |
| 270 | } |
| 271 | else if (action == POLL_LOOP_ACTION_CLEAN) { |
| 272 | if (poll_events) |
| 273 | free(poll_events); |
| 274 | return 1; |
| 275 | } |
| 276 | |
| 277 | /* OK, it's POLL_LOOP_ACTION_RUN */ |
| 278 | |
| 279 | tv_now(&now); |
| 280 | |
| 281 | while (1) { |
| 282 | next_time = process_runnable_tasks(); |
| 283 | |
| 284 | /* stop when there's no connection left and we don't allow them anymore */ |
| 285 | if (!actconn && listeners == 0) |
| 286 | break; |
| 287 | |
| 288 | nbfd = 0; |
| 289 | for (fds = 0; (fds << INTBITS) < maxfd; fds++) { |
| 290 | |
| 291 | rn = ((int*)StaticReadEvent)[fds]; |
| 292 | wn = ((int*)StaticWriteEvent)[fds]; |
| 293 | |
| 294 | if ((rn|wn)) { |
| 295 | for (count = 0, fd = fds << INTBITS; count < (1<<INTBITS) && fd < maxfd; count++, fd++) { |
| 296 | #define FDSETS_ARE_INT_ALIGNED |
| 297 | #ifdef FDSETS_ARE_INT_ALIGNED |
| 298 | |
| 299 | #define WE_REALLY_NOW_THAT_FDSETS_ARE_INTS |
| 300 | #ifdef WE_REALLY_NOW_THAT_FDSETS_ARE_INTS |
| 301 | sr = (rn >> count) & 1; |
| 302 | sw = (wn >> count) & 1; |
| 303 | #else |
| 304 | sr = FD_ISSET(fd&((1<<INTBITS)-1), (typeof(fd_set*))&rn); |
| 305 | sw = FD_ISSET(fd&((1<<INTBITS)-1), (typeof(fd_set*))&wn); |
| 306 | #endif |
| 307 | #else |
| 308 | sr = FD_ISSET(fd, StaticReadEvent); |
| 309 | sw = FD_ISSET(fd, StaticWriteEvent); |
| 310 | #endif |
| 311 | if ((sr|sw)) { |
| 312 | poll_events[nbfd].fd = fd; |
| 313 | poll_events[nbfd].events = (sr ? POLLIN : 0) | (sw ? POLLOUT : 0); |
| 314 | nbfd++; |
| 315 | } |
| 316 | } |
| 317 | } |
| 318 | } |
| 319 | |
| 320 | /* now let's wait for events */ |
| 321 | status = poll(poll_events, nbfd, next_time); |
| 322 | tv_now(&now); |
| 323 | |
| 324 | for (count = 0; status > 0 && count < nbfd; count++) { |
| 325 | fd = poll_events[count].fd; |
| 326 | |
| 327 | if (!(poll_events[count].revents & ( POLLOUT | POLLIN | POLLERR | POLLHUP ))) |
| 328 | continue; |
| 329 | |
| 330 | /* ok, we found one active fd */ |
| 331 | status--; |
| 332 | |
| 333 | if (FD_ISSET(fd, StaticReadEvent)) { |
| 334 | if (fdtab[fd].state == FD_STCLOSE) |
| 335 | continue; |
| 336 | if (poll_events[count].revents & ( POLLIN | POLLERR | POLLHUP )) |
| 337 | fdtab[fd].read(fd); |
| 338 | } |
| 339 | |
| 340 | if (FD_ISSET(fd, StaticWriteEvent)) { |
| 341 | if (fdtab[fd].state == FD_STCLOSE) |
| 342 | continue; |
| 343 | if (poll_events[count].revents & ( POLLOUT | POLLERR | POLLHUP )) |
| 344 | fdtab[fd].write(fd); |
| 345 | } |
| 346 | } |
| 347 | } |
| 348 | return 1; |
| 349 | } |
| 350 | #endif |
| 351 | |
| 352 | |
| 353 | |
| 354 | /* |
| 355 | * Main select() loop. |
| 356 | * does 3 actions : |
| 357 | * 0 (POLL_LOOP_ACTION_INIT) : initializes necessary private structures |
| 358 | * 1 (POLL_LOOP_ACTION_RUN) : runs the loop |
| 359 | * 2 (POLL_LOOP_ACTION_CLEAN) : cleans up |
| 360 | * |
| 361 | * returns 0 if initialization failed, !0 otherwise. |
| 362 | */ |
| 363 | |
| 364 | |
| 365 | int select_loop(int action) |
| 366 | { |
| 367 | int next_time; |
| 368 | int status; |
| 369 | int fd,i; |
| 370 | struct timeval delta; |
| 371 | int readnotnull, writenotnull; |
| 372 | static fd_set *ReadEvent = NULL, *WriteEvent = NULL; |
| 373 | |
| 374 | if (action == POLL_LOOP_ACTION_INIT) { |
| 375 | ReadEvent = (fd_set *) |
| 376 | calloc(1, sizeof(fd_set) * (global.maxsock + FD_SETSIZE - 1) / FD_SETSIZE); |
| 377 | WriteEvent = (fd_set *) |
| 378 | calloc(1, sizeof(fd_set) * (global.maxsock + FD_SETSIZE - 1) / FD_SETSIZE); |
| 379 | return 1; |
| 380 | } |
| 381 | else if (action == POLL_LOOP_ACTION_CLEAN) { |
| 382 | if (WriteEvent) free(WriteEvent); |
| 383 | if (ReadEvent) free(ReadEvent); |
| 384 | return 1; |
| 385 | } |
| 386 | |
| 387 | /* OK, it's POLL_LOOP_ACTION_RUN */ |
| 388 | |
| 389 | tv_now(&now); |
| 390 | |
| 391 | while (1) { |
| 392 | next_time = process_runnable_tasks(); |
| 393 | |
| 394 | /* stop when there's no connection left and we don't allow them anymore */ |
| 395 | if (!actconn && listeners == 0) |
| 396 | break; |
| 397 | |
| 398 | if (next_time > 0) { /* FIXME */ |
| 399 | /* Convert to timeval */ |
| 400 | /* to avoid eventual select loops due to timer precision */ |
| 401 | next_time += SCHEDULER_RESOLUTION; |
| 402 | delta.tv_sec = next_time / 1000; |
| 403 | delta.tv_usec = (next_time % 1000) * 1000; |
| 404 | } |
| 405 | else if (next_time == 0) { /* allow select to return immediately when needed */ |
| 406 | delta.tv_sec = delta.tv_usec = 0; |
| 407 | } |
| 408 | |
| 409 | |
| 410 | /* let's restore fdset state */ |
| 411 | |
| 412 | readnotnull = 0; writenotnull = 0; |
| 413 | for (i = 0; i < (maxfd + FD_SETSIZE - 1)/(8*sizeof(int)); i++) { |
| 414 | readnotnull |= (*(((int*)ReadEvent)+i) = *(((int*)StaticReadEvent)+i)) != 0; |
| 415 | writenotnull |= (*(((int*)WriteEvent)+i) = *(((int*)StaticWriteEvent)+i)) != 0; |
| 416 | } |
| 417 | |
| 418 | // /* just a verification code, needs to be removed for performance */ |
| 419 | // for (i=0; i<maxfd; i++) { |
| 420 | // if (FD_ISSET(i, ReadEvent) != FD_ISSET(i, StaticReadEvent)) |
| 421 | // abort(); |
| 422 | // if (FD_ISSET(i, WriteEvent) != FD_ISSET(i, StaticWriteEvent)) |
| 423 | // abort(); |
| 424 | // |
| 425 | // } |
| 426 | |
| 427 | status = select(maxfd, |
| 428 | readnotnull ? ReadEvent : NULL, |
| 429 | writenotnull ? WriteEvent : NULL, |
| 430 | NULL, |
| 431 | (next_time >= 0) ? &delta : NULL); |
| 432 | |
| 433 | /* this is an experiment on the separation of the select work */ |
| 434 | // status = (readnotnull ? select(maxfd, ReadEvent, NULL, NULL, (next_time >= 0) ? &delta : NULL) : 0); |
| 435 | // status |= (writenotnull ? select(maxfd, NULL, WriteEvent, NULL, (next_time >= 0) ? &delta : NULL) : 0); |
| 436 | |
| 437 | tv_now(&now); |
| 438 | |
| 439 | if (status > 0) { /* must proceed with events */ |
| 440 | |
| 441 | int fds; |
| 442 | char count; |
| 443 | |
| 444 | for (fds = 0; (fds << INTBITS) < maxfd; fds++) |
| 445 | if ((((int *)(ReadEvent))[fds] | ((int *)(WriteEvent))[fds]) != 0) |
| 446 | for (count = 1<<INTBITS, fd = fds << INTBITS; count && fd < maxfd; count--, fd++) { |
| 447 | |
| 448 | /* if we specify read first, the accepts and zero reads will be |
| 449 | * seen first. Moreover, system buffers will be flushed faster. |
| 450 | */ |
| 451 | if (FD_ISSET(fd, ReadEvent)) { |
| 452 | if (fdtab[fd].state == FD_STCLOSE) |
| 453 | continue; |
| 454 | fdtab[fd].read(fd); |
| 455 | } |
| 456 | |
| 457 | if (FD_ISSET(fd, WriteEvent)) { |
| 458 | if (fdtab[fd].state == FD_STCLOSE) |
| 459 | continue; |
| 460 | fdtab[fd].write(fd); |
| 461 | } |
| 462 | } |
| 463 | } |
| 464 | else { |
| 465 | // fprintf(stderr,"select returned %d, maxfd=%d\n", status, maxfd); |
| 466 | } |
| 467 | } |
| 468 | return 1; |
| 469 | } |
| 470 | |
| 471 | |
| 472 | |
| 473 | /********************* |
| 474 | * generic functions |
| 475 | *********************/ |
| 476 | |
| 477 | |
| 478 | /* Deletes an FD from the fdsets, and recomputes the maxfd limit. |
| 479 | * The file descriptor is also closed. |
| 480 | */ |
| 481 | void fd_delete(int fd) |
| 482 | { |
| 483 | FD_CLR(fd, StaticReadEvent); |
| 484 | FD_CLR(fd, StaticWriteEvent); |
| 485 | #if defined(ENABLE_EPOLL) |
| 486 | if (PrevReadEvent) { |
| 487 | FD_CLR(fd, PrevReadEvent); |
| 488 | FD_CLR(fd, PrevWriteEvent); |
| 489 | } |
| 490 | #endif |
| 491 | |
| 492 | close(fd); |
| 493 | fdtab[fd].state = FD_STCLOSE; |
| 494 | |
| 495 | while ((maxfd-1 >= 0) && (fdtab[maxfd-1].state == FD_STCLOSE)) |
| 496 | maxfd--; |
| 497 | } |
| 498 | |
| 499 | |
| 500 | /* |
| 501 | * Local variables: |
| 502 | * c-indent-level: 8 |
| 503 | * c-basic-offset: 8 |
| 504 | * End: |
| 505 | */ |