]>
Commit | Line | Data |
---|---|---|
718e3744 | 1 | /* Thread management routine |
2 | * Copyright (C) 1998, 2000 Kunihiro Ishiguro <kunihiro@zebra.org> | |
3 | * | |
4 | * This file is part of GNU Zebra. | |
5 | * | |
6 | * GNU Zebra is free software; you can redistribute it and/or modify it | |
7 | * under the terms of the GNU General Public License as published by the | |
8 | * Free Software Foundation; either version 2, or (at your option) any | |
9 | * later version. | |
10 | * | |
11 | * GNU Zebra is distributed in the hope that it will be useful, but | |
12 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
14 | * General Public License for more details. | |
15 | * | |
896014f4 DL |
16 | * You should have received a copy of the GNU General Public License along |
17 | * with this program; see the file COPYING; if not, write to the Free Software | |
18 | * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA | |
718e3744 | 19 | */ |
20 | ||
21 | /* #define DEBUG */ | |
22 | ||
23 | #include <zebra.h> | |
308d14ae | 24 | #include <sys/resource.h> |
718e3744 | 25 | |
26 | #include "thread.h" | |
27 | #include "memory.h" | |
28 | #include "log.h" | |
e04ab74d | 29 | #include "hash.h" |
4becea72 | 30 | #include "pqueue.h" |
e04ab74d | 31 | #include "command.h" |
05c447dd | 32 | #include "sigevent.h" |
3bf2673b | 33 | #include "network.h" |
bd74dc61 | 34 | #include "jhash.h" |
d6be5fb9 | 35 | |
d62a17ae | 36 | DEFINE_MTYPE_STATIC(LIB, THREAD, "Thread") |
4a1ab8e4 | 37 | DEFINE_MTYPE_STATIC(LIB, THREAD_MASTER, "Thread master") |
d62a17ae | 38 | DEFINE_MTYPE_STATIC(LIB, THREAD_STATS, "Thread stats") |
4a1ab8e4 | 39 | |
3b96b781 HT |
40 | #if defined(__APPLE__) |
41 | #include <mach/mach.h> | |
42 | #include <mach/mach_time.h> | |
43 | #endif | |
44 | ||
d62a17ae | 45 | #define AWAKEN(m) \ |
46 | do { \ | |
47 | static unsigned char wakebyte = 0x01; \ | |
48 | write(m->io_pipe[1], &wakebyte, 1); \ | |
49 | } while (0); | |
3bf2673b | 50 | |
62f44022 | 51 | /* control variable for initializer */ |
e0bebc7c | 52 | pthread_once_t init_once = PTHREAD_ONCE_INIT; |
e0bebc7c | 53 | pthread_key_t thread_current; |
6b0655a2 | 54 | |
62f44022 QY |
55 | pthread_mutex_t masters_mtx = PTHREAD_MUTEX_INITIALIZER; |
56 | static struct list *masters; | |
57 | ||
6b0655a2 | 58 | |
62f44022 | 59 | /* CLI start ---------------------------------------------------------------- */ |
d62a17ae | 60 | static unsigned int cpu_record_hash_key(struct cpu_thread_history *a) |
e04ab74d | 61 | { |
996c9314 | 62 | int size = sizeof(&a->func); |
bd74dc61 DS |
63 | |
64 | return jhash(&a->func, size, 0); | |
e04ab74d | 65 | } |
66 | ||
d62a17ae | 67 | static int cpu_record_hash_cmp(const struct cpu_thread_history *a, |
68 | const struct cpu_thread_history *b) | |
e04ab74d | 69 | { |
d62a17ae | 70 | return a->func == b->func; |
e04ab74d | 71 | } |
72 | ||
d62a17ae | 73 | static void *cpu_record_hash_alloc(struct cpu_thread_history *a) |
e04ab74d | 74 | { |
d62a17ae | 75 | struct cpu_thread_history *new; |
76 | new = XCALLOC(MTYPE_THREAD_STATS, sizeof(struct cpu_thread_history)); | |
77 | new->func = a->func; | |
78 | new->funcname = a->funcname; | |
79 | return new; | |
e04ab74d | 80 | } |
81 | ||
d62a17ae | 82 | static void cpu_record_hash_free(void *a) |
228da428 | 83 | { |
d62a17ae | 84 | struct cpu_thread_history *hist = a; |
85 | ||
86 | XFREE(MTYPE_THREAD_STATS, hist); | |
228da428 CC |
87 | } |
88 | ||
d62a17ae | 89 | static void vty_out_cpu_thread_history(struct vty *vty, |
90 | struct cpu_thread_history *a) | |
e04ab74d | 91 | { |
d1667f53 | 92 | vty_out(vty, "%5d %10lu.%03lu %9u %8lu %9lu %8lu %9lu", a->total_active, |
d62a17ae | 93 | a->cpu.total / 1000, a->cpu.total % 1000, a->total_calls, |
94 | a->cpu.total / a->total_calls, a->cpu.max, | |
95 | a->real.total / a->total_calls, a->real.max); | |
96 | vty_out(vty, " %c%c%c%c%c %s\n", | |
97 | a->types & (1 << THREAD_READ) ? 'R' : ' ', | |
98 | a->types & (1 << THREAD_WRITE) ? 'W' : ' ', | |
99 | a->types & (1 << THREAD_TIMER) ? 'T' : ' ', | |
100 | a->types & (1 << THREAD_EVENT) ? 'E' : ' ', | |
101 | a->types & (1 << THREAD_EXECUTE) ? 'X' : ' ', a->funcname); | |
e04ab74d | 102 | } |
103 | ||
d62a17ae | 104 | static void cpu_record_hash_print(struct hash_backet *bucket, void *args[]) |
e04ab74d | 105 | { |
d62a17ae | 106 | struct cpu_thread_history *totals = args[0]; |
107 | struct vty *vty = args[1]; | |
108 | thread_type *filter = args[2]; | |
109 | ||
110 | struct cpu_thread_history *a = bucket->data; | |
111 | ||
112 | if (!(a->types & *filter)) | |
113 | return; | |
114 | vty_out_cpu_thread_history(vty, a); | |
115 | totals->total_active += a->total_active; | |
116 | totals->total_calls += a->total_calls; | |
117 | totals->real.total += a->real.total; | |
118 | if (totals->real.max < a->real.max) | |
119 | totals->real.max = a->real.max; | |
120 | totals->cpu.total += a->cpu.total; | |
121 | if (totals->cpu.max < a->cpu.max) | |
122 | totals->cpu.max = a->cpu.max; | |
e04ab74d | 123 | } |
124 | ||
d62a17ae | 125 | static void cpu_record_print(struct vty *vty, thread_type filter) |
e04ab74d | 126 | { |
d62a17ae | 127 | struct cpu_thread_history tmp; |
128 | void *args[3] = {&tmp, vty, &filter}; | |
129 | struct thread_master *m; | |
130 | struct listnode *ln; | |
131 | ||
132 | memset(&tmp, 0, sizeof tmp); | |
133 | tmp.funcname = "TOTAL"; | |
134 | tmp.types = filter; | |
135 | ||
136 | pthread_mutex_lock(&masters_mtx); | |
137 | { | |
138 | for (ALL_LIST_ELEMENTS_RO(masters, ln, m)) { | |
139 | const char *name = m->name ? m->name : "main"; | |
140 | ||
141 | char underline[strlen(name) + 1]; | |
142 | memset(underline, '-', sizeof(underline)); | |
143 | underline[sizeof(underline)] = '\0'; | |
144 | ||
145 | vty_out(vty, "\n"); | |
146 | vty_out(vty, "Showing statistics for pthread %s\n", | |
147 | name); | |
148 | vty_out(vty, "-------------------------------%s\n", | |
149 | underline); | |
150 | vty_out(vty, "%21s %18s %18s\n", "", | |
151 | "CPU (user+system):", "Real (wall-clock):"); | |
152 | vty_out(vty, | |
153 | "Active Runtime(ms) Invoked Avg uSec Max uSecs"); | |
154 | vty_out(vty, " Avg uSec Max uSecs"); | |
155 | vty_out(vty, " Type Thread\n"); | |
156 | ||
157 | if (m->cpu_record->count) | |
158 | hash_iterate( | |
159 | m->cpu_record, | |
160 | (void (*)(struct hash_backet *, | |
161 | void *))cpu_record_hash_print, | |
162 | args); | |
163 | else | |
164 | vty_out(vty, "No data to display yet.\n"); | |
165 | ||
166 | vty_out(vty, "\n"); | |
167 | } | |
168 | } | |
169 | pthread_mutex_unlock(&masters_mtx); | |
170 | ||
171 | vty_out(vty, "\n"); | |
172 | vty_out(vty, "Total thread statistics\n"); | |
173 | vty_out(vty, "-------------------------\n"); | |
174 | vty_out(vty, "%21s %18s %18s\n", "", | |
175 | "CPU (user+system):", "Real (wall-clock):"); | |
176 | vty_out(vty, "Active Runtime(ms) Invoked Avg uSec Max uSecs"); | |
177 | vty_out(vty, " Avg uSec Max uSecs"); | |
178 | vty_out(vty, " Type Thread\n"); | |
179 | ||
180 | if (tmp.total_calls > 0) | |
181 | vty_out_cpu_thread_history(vty, &tmp); | |
e04ab74d | 182 | } |
183 | ||
d62a17ae | 184 | static void cpu_record_hash_clear(struct hash_backet *bucket, void *args[]) |
e276eb82 | 185 | { |
d62a17ae | 186 | thread_type *filter = args[0]; |
187 | struct hash *cpu_record = args[1]; | |
188 | ||
189 | struct cpu_thread_history *a = bucket->data; | |
62f44022 | 190 | |
d62a17ae | 191 | if (!(a->types & *filter)) |
192 | return; | |
f48f65d2 | 193 | |
d62a17ae | 194 | hash_release(cpu_record, bucket->data); |
e276eb82 PJ |
195 | } |
196 | ||
d62a17ae | 197 | static void cpu_record_clear(thread_type filter) |
e276eb82 | 198 | { |
d62a17ae | 199 | thread_type *tmp = &filter; |
200 | struct thread_master *m; | |
201 | struct listnode *ln; | |
202 | ||
203 | pthread_mutex_lock(&masters_mtx); | |
204 | { | |
205 | for (ALL_LIST_ELEMENTS_RO(masters, ln, m)) { | |
206 | pthread_mutex_lock(&m->mtx); | |
207 | { | |
208 | void *args[2] = {tmp, m->cpu_record}; | |
209 | hash_iterate( | |
210 | m->cpu_record, | |
211 | (void (*)(struct hash_backet *, | |
212 | void *))cpu_record_hash_clear, | |
213 | args); | |
214 | } | |
215 | pthread_mutex_unlock(&m->mtx); | |
216 | } | |
217 | } | |
218 | pthread_mutex_unlock(&masters_mtx); | |
62f44022 QY |
219 | } |
220 | ||
d62a17ae | 221 | static thread_type parse_filter(const char *filterstr) |
62f44022 | 222 | { |
d62a17ae | 223 | int i = 0; |
224 | int filter = 0; | |
225 | ||
226 | while (filterstr[i] != '\0') { | |
227 | switch (filterstr[i]) { | |
228 | case 'r': | |
229 | case 'R': | |
230 | filter |= (1 << THREAD_READ); | |
231 | break; | |
232 | case 'w': | |
233 | case 'W': | |
234 | filter |= (1 << THREAD_WRITE); | |
235 | break; | |
236 | case 't': | |
237 | case 'T': | |
238 | filter |= (1 << THREAD_TIMER); | |
239 | break; | |
240 | case 'e': | |
241 | case 'E': | |
242 | filter |= (1 << THREAD_EVENT); | |
243 | break; | |
244 | case 'x': | |
245 | case 'X': | |
246 | filter |= (1 << THREAD_EXECUTE); | |
247 | break; | |
248 | default: | |
249 | break; | |
250 | } | |
251 | ++i; | |
252 | } | |
253 | return filter; | |
62f44022 QY |
254 | } |
255 | ||
256 | DEFUN (show_thread_cpu, | |
257 | show_thread_cpu_cmd, | |
258 | "show thread cpu [FILTER]", | |
259 | SHOW_STR | |
260 | "Thread information\n" | |
261 | "Thread CPU usage\n" | |
262 | "Display filter (rwtexb)\n") | |
263 | { | |
d62a17ae | 264 | thread_type filter = (thread_type)-1U; |
265 | int idx = 0; | |
266 | ||
267 | if (argv_find(argv, argc, "FILTER", &idx)) { | |
268 | filter = parse_filter(argv[idx]->arg); | |
269 | if (!filter) { | |
270 | vty_out(vty, | |
271 | "Invalid filter \"%s\" specified; must contain at least" | |
272 | "one of 'RWTEXB'\n", | |
273 | argv[idx]->arg); | |
274 | return CMD_WARNING; | |
275 | } | |
276 | } | |
277 | ||
278 | cpu_record_print(vty, filter); | |
279 | return CMD_SUCCESS; | |
e276eb82 PJ |
280 | } |
281 | ||
49d41a26 DS |
282 | DEFUN (clear_thread_cpu, |
283 | clear_thread_cpu_cmd, | |
284 | "clear thread cpu [FILTER]", | |
62f44022 | 285 | "Clear stored data in all pthreads\n" |
49d41a26 DS |
286 | "Thread information\n" |
287 | "Thread CPU usage\n" | |
288 | "Display filter (rwtexb)\n") | |
e276eb82 | 289 | { |
d62a17ae | 290 | thread_type filter = (thread_type)-1U; |
291 | int idx = 0; | |
292 | ||
293 | if (argv_find(argv, argc, "FILTER", &idx)) { | |
294 | filter = parse_filter(argv[idx]->arg); | |
295 | if (!filter) { | |
296 | vty_out(vty, | |
297 | "Invalid filter \"%s\" specified; must contain at least" | |
298 | "one of 'RWTEXB'\n", | |
299 | argv[idx]->arg); | |
300 | return CMD_WARNING; | |
301 | } | |
302 | } | |
303 | ||
304 | cpu_record_clear(filter); | |
305 | return CMD_SUCCESS; | |
e276eb82 | 306 | } |
6b0655a2 | 307 | |
d62a17ae | 308 | void thread_cmd_init(void) |
0b84f294 | 309 | { |
d62a17ae | 310 | install_element(VIEW_NODE, &show_thread_cpu_cmd); |
311 | install_element(ENABLE_NODE, &clear_thread_cpu_cmd); | |
0b84f294 | 312 | } |
62f44022 QY |
313 | /* CLI end ------------------------------------------------------------------ */ |
314 | ||
0b84f294 | 315 | |
d62a17ae | 316 | static int thread_timer_cmp(void *a, void *b) |
4becea72 | 317 | { |
d62a17ae | 318 | struct thread *thread_a = a; |
319 | struct thread *thread_b = b; | |
320 | ||
321 | if (timercmp(&thread_a->u.sands, &thread_b->u.sands, <)) | |
322 | return -1; | |
323 | if (timercmp(&thread_a->u.sands, &thread_b->u.sands, >)) | |
324 | return 1; | |
325 | return 0; | |
4becea72 CF |
326 | } |
327 | ||
d62a17ae | 328 | static void thread_timer_update(void *node, int actual_position) |
4becea72 | 329 | { |
d62a17ae | 330 | struct thread *thread = node; |
4becea72 | 331 | |
d62a17ae | 332 | thread->index = actual_position; |
4becea72 CF |
333 | } |
334 | ||
d62a17ae | 335 | static void cancelreq_del(void *cr) |
63ccb9cb | 336 | { |
d62a17ae | 337 | XFREE(MTYPE_TMP, cr); |
63ccb9cb QY |
338 | } |
339 | ||
e0bebc7c | 340 | /* initializer, only ever called once */ |
d62a17ae | 341 | static void initializer() |
e0bebc7c | 342 | { |
d62a17ae | 343 | pthread_key_create(&thread_current, NULL); |
e0bebc7c QY |
344 | } |
345 | ||
d62a17ae | 346 | struct thread_master *thread_master_create(const char *name) |
718e3744 | 347 | { |
d62a17ae | 348 | struct thread_master *rv; |
349 | struct rlimit limit; | |
350 | ||
351 | pthread_once(&init_once, &initializer); | |
352 | ||
353 | rv = XCALLOC(MTYPE_THREAD_MASTER, sizeof(struct thread_master)); | |
354 | if (rv == NULL) | |
355 | return NULL; | |
356 | ||
357 | /* Initialize master mutex */ | |
358 | pthread_mutex_init(&rv->mtx, NULL); | |
359 | pthread_cond_init(&rv->cancel_cond, NULL); | |
360 | ||
361 | /* Set name */ | |
362 | rv->name = name ? XSTRDUP(MTYPE_THREAD_MASTER, name) : NULL; | |
363 | ||
364 | /* Initialize I/O task data structures */ | |
365 | getrlimit(RLIMIT_NOFILE, &limit); | |
366 | rv->fd_limit = (int)limit.rlim_cur; | |
367 | rv->read = | |
368 | XCALLOC(MTYPE_THREAD, sizeof(struct thread *) * rv->fd_limit); | |
369 | if (rv->read == NULL) { | |
370 | XFREE(MTYPE_THREAD_MASTER, rv); | |
371 | return NULL; | |
372 | } | |
373 | rv->write = | |
374 | XCALLOC(MTYPE_THREAD, sizeof(struct thread *) * rv->fd_limit); | |
375 | if (rv->write == NULL) { | |
376 | XFREE(MTYPE_THREAD, rv->read); | |
377 | XFREE(MTYPE_THREAD_MASTER, rv); | |
378 | return NULL; | |
379 | } | |
380 | ||
bd74dc61 | 381 | rv->cpu_record = hash_create_size( |
996c9314 | 382 | 8, (unsigned int (*)(void *))cpu_record_hash_key, |
bd74dc61 DS |
383 | (int (*)(const void *, const void *))cpu_record_hash_cmp, |
384 | "Thread Hash"); | |
d62a17ae | 385 | |
386 | ||
387 | /* Initialize the timer queues */ | |
388 | rv->timer = pqueue_create(); | |
389 | rv->timer->cmp = thread_timer_cmp; | |
390 | rv->timer->update = thread_timer_update; | |
391 | ||
392 | /* Initialize thread_fetch() settings */ | |
393 | rv->spin = true; | |
394 | rv->handle_signals = true; | |
395 | ||
396 | /* Set pthread owner, should be updated by actual owner */ | |
397 | rv->owner = pthread_self(); | |
398 | rv->cancel_req = list_new(); | |
399 | rv->cancel_req->del = cancelreq_del; | |
400 | rv->canceled = true; | |
401 | ||
402 | /* Initialize pipe poker */ | |
403 | pipe(rv->io_pipe); | |
404 | set_nonblocking(rv->io_pipe[0]); | |
405 | set_nonblocking(rv->io_pipe[1]); | |
406 | ||
407 | /* Initialize data structures for poll() */ | |
408 | rv->handler.pfdsize = rv->fd_limit; | |
409 | rv->handler.pfdcount = 0; | |
410 | rv->handler.pfds = XCALLOC(MTYPE_THREAD_MASTER, | |
411 | sizeof(struct pollfd) * rv->handler.pfdsize); | |
412 | rv->handler.copy = XCALLOC(MTYPE_THREAD_MASTER, | |
413 | sizeof(struct pollfd) * rv->handler.pfdsize); | |
414 | ||
eff09c66 | 415 | /* add to list of threadmasters */ |
d62a17ae | 416 | pthread_mutex_lock(&masters_mtx); |
417 | { | |
eff09c66 QY |
418 | if (!masters) |
419 | masters = list_new(); | |
420 | ||
d62a17ae | 421 | listnode_add(masters, rv); |
422 | } | |
423 | pthread_mutex_unlock(&masters_mtx); | |
424 | ||
425 | return rv; | |
718e3744 | 426 | } |
427 | ||
d8a8a8de QY |
428 | void thread_master_set_name(struct thread_master *master, const char *name) |
429 | { | |
430 | pthread_mutex_lock(&master->mtx); | |
431 | { | |
432 | if (master->name) | |
433 | XFREE(MTYPE_THREAD_MASTER, master->name); | |
434 | master->name = XSTRDUP(MTYPE_THREAD_MASTER, name); | |
435 | } | |
436 | pthread_mutex_unlock(&master->mtx); | |
437 | } | |
438 | ||
718e3744 | 439 | /* Add a new thread to the list. */ |
d62a17ae | 440 | static void thread_list_add(struct thread_list *list, struct thread *thread) |
718e3744 | 441 | { |
d62a17ae | 442 | thread->next = NULL; |
443 | thread->prev = list->tail; | |
444 | if (list->tail) | |
445 | list->tail->next = thread; | |
446 | else | |
447 | list->head = thread; | |
448 | list->tail = thread; | |
449 | list->count++; | |
718e3744 | 450 | } |
451 | ||
718e3744 | 452 | /* Delete a thread from the list. */ |
d62a17ae | 453 | static struct thread *thread_list_delete(struct thread_list *list, |
454 | struct thread *thread) | |
718e3744 | 455 | { |
d62a17ae | 456 | if (thread->next) |
457 | thread->next->prev = thread->prev; | |
458 | else | |
459 | list->tail = thread->prev; | |
460 | if (thread->prev) | |
461 | thread->prev->next = thread->next; | |
462 | else | |
463 | list->head = thread->next; | |
464 | thread->next = thread->prev = NULL; | |
465 | list->count--; | |
466 | return thread; | |
718e3744 | 467 | } |
468 | ||
495f0b13 | 469 | /* Thread list is empty or not. */ |
d62a17ae | 470 | static int thread_empty(struct thread_list *list) |
495f0b13 | 471 | { |
d62a17ae | 472 | return list->head ? 0 : 1; |
495f0b13 DS |
473 | } |
474 | ||
475 | /* Delete top of the list and return it. */ | |
d62a17ae | 476 | static struct thread *thread_trim_head(struct thread_list *list) |
495f0b13 | 477 | { |
d62a17ae | 478 | if (!thread_empty(list)) |
479 | return thread_list_delete(list, list->head); | |
480 | return NULL; | |
495f0b13 DS |
481 | } |
482 | ||
718e3744 | 483 | /* Move thread to unuse list. */ |
d62a17ae | 484 | static void thread_add_unuse(struct thread_master *m, struct thread *thread) |
718e3744 | 485 | { |
d62a17ae | 486 | assert(m != NULL && thread != NULL); |
487 | assert(thread->next == NULL); | |
488 | assert(thread->prev == NULL); | |
489 | thread->ref = NULL; | |
490 | ||
491 | thread->type = THREAD_UNUSED; | |
492 | thread->hist->total_active--; | |
493 | thread_list_add(&m->unuse, thread); | |
718e3744 | 494 | } |
495 | ||
496 | /* Free all unused thread. */ | |
d62a17ae | 497 | static void thread_list_free(struct thread_master *m, struct thread_list *list) |
718e3744 | 498 | { |
d62a17ae | 499 | struct thread *t; |
500 | struct thread *next; | |
501 | ||
502 | for (t = list->head; t; t = next) { | |
503 | next = t->next; | |
504 | XFREE(MTYPE_THREAD, t); | |
505 | list->count--; | |
506 | m->alloc--; | |
507 | } | |
718e3744 | 508 | } |
509 | ||
d62a17ae | 510 | static void thread_array_free(struct thread_master *m, |
511 | struct thread **thread_array) | |
308d14ae | 512 | { |
d62a17ae | 513 | struct thread *t; |
514 | int index; | |
515 | ||
516 | for (index = 0; index < m->fd_limit; ++index) { | |
517 | t = thread_array[index]; | |
518 | if (t) { | |
519 | thread_array[index] = NULL; | |
520 | XFREE(MTYPE_THREAD, t); | |
521 | m->alloc--; | |
522 | } | |
523 | } | |
524 | XFREE(MTYPE_THREAD, thread_array); | |
308d14ae DV |
525 | } |
526 | ||
d62a17ae | 527 | static void thread_queue_free(struct thread_master *m, struct pqueue *queue) |
4becea72 | 528 | { |
d62a17ae | 529 | int i; |
4becea72 | 530 | |
d62a17ae | 531 | for (i = 0; i < queue->size; i++) |
532 | XFREE(MTYPE_THREAD, queue->array[i]); | |
4becea72 | 533 | |
d62a17ae | 534 | m->alloc -= queue->size; |
535 | pqueue_delete(queue); | |
4becea72 CF |
536 | } |
537 | ||
495f0b13 DS |
538 | /* |
539 | * thread_master_free_unused | |
540 | * | |
541 | * As threads are finished with they are put on the | |
542 | * unuse list for later reuse. | |
543 | * If we are shutting down, Free up unused threads | |
544 | * So we can see if we forget to shut anything off | |
545 | */ | |
d62a17ae | 546 | void thread_master_free_unused(struct thread_master *m) |
495f0b13 | 547 | { |
d62a17ae | 548 | pthread_mutex_lock(&m->mtx); |
549 | { | |
550 | struct thread *t; | |
551 | while ((t = thread_trim_head(&m->unuse)) != NULL) { | |
552 | pthread_mutex_destroy(&t->mtx); | |
553 | XFREE(MTYPE_THREAD, t); | |
554 | } | |
555 | } | |
556 | pthread_mutex_unlock(&m->mtx); | |
495f0b13 DS |
557 | } |
558 | ||
718e3744 | 559 | /* Stop thread scheduler. */ |
d62a17ae | 560 | void thread_master_free(struct thread_master *m) |
718e3744 | 561 | { |
d62a17ae | 562 | pthread_mutex_lock(&masters_mtx); |
563 | { | |
564 | listnode_delete(masters, m); | |
eff09c66 | 565 | if (masters->count == 0) { |
acdf5e25 | 566 | list_delete_and_null(&masters); |
eff09c66 | 567 | } |
d62a17ae | 568 | } |
569 | pthread_mutex_unlock(&masters_mtx); | |
570 | ||
571 | thread_array_free(m, m->read); | |
572 | thread_array_free(m, m->write); | |
573 | thread_queue_free(m, m->timer); | |
574 | thread_list_free(m, &m->event); | |
575 | thread_list_free(m, &m->ready); | |
576 | thread_list_free(m, &m->unuse); | |
577 | pthread_mutex_destroy(&m->mtx); | |
33844bbe | 578 | pthread_cond_destroy(&m->cancel_cond); |
d62a17ae | 579 | close(m->io_pipe[0]); |
580 | close(m->io_pipe[1]); | |
affe9e99 | 581 | list_delete_and_null(&m->cancel_req); |
1a0a92ea | 582 | m->cancel_req = NULL; |
d62a17ae | 583 | |
584 | hash_clean(m->cpu_record, cpu_record_hash_free); | |
585 | hash_free(m->cpu_record); | |
586 | m->cpu_record = NULL; | |
587 | ||
4e1000a1 QY |
588 | if (m->name) |
589 | XFREE(MTYPE_THREAD_MASTER, m->name); | |
d62a17ae | 590 | XFREE(MTYPE_THREAD_MASTER, m->handler.pfds); |
591 | XFREE(MTYPE_THREAD_MASTER, m->handler.copy); | |
592 | XFREE(MTYPE_THREAD_MASTER, m); | |
718e3744 | 593 | } |
594 | ||
718e3744 | 595 | /* Return remain time in second. */ |
d62a17ae | 596 | unsigned long thread_timer_remain_second(struct thread *thread) |
718e3744 | 597 | { |
d62a17ae | 598 | int64_t remain; |
1189d95f | 599 | |
d62a17ae | 600 | pthread_mutex_lock(&thread->mtx); |
601 | { | |
602 | remain = monotime_until(&thread->u.sands, NULL) / 1000000LL; | |
603 | } | |
604 | pthread_mutex_unlock(&thread->mtx); | |
1189d95f | 605 | |
d62a17ae | 606 | return remain < 0 ? 0 : remain; |
718e3744 | 607 | } |
608 | ||
9c7753e4 DL |
609 | #define debugargdef const char *funcname, const char *schedfrom, int fromln |
610 | #define debugargpass funcname, schedfrom, fromln | |
e04ab74d | 611 | |
d62a17ae | 612 | struct timeval thread_timer_remain(struct thread *thread) |
6ac44687 | 613 | { |
d62a17ae | 614 | struct timeval remain; |
615 | pthread_mutex_lock(&thread->mtx); | |
616 | { | |
617 | monotime_until(&thread->u.sands, &remain); | |
618 | } | |
619 | pthread_mutex_unlock(&thread->mtx); | |
620 | return remain; | |
6ac44687 CF |
621 | } |
622 | ||
718e3744 | 623 | /* Get new thread. */ |
d7c0a89a | 624 | static struct thread *thread_get(struct thread_master *m, uint8_t type, |
d62a17ae | 625 | int (*func)(struct thread *), void *arg, |
626 | debugargdef) | |
718e3744 | 627 | { |
d62a17ae | 628 | struct thread *thread = thread_trim_head(&m->unuse); |
629 | struct cpu_thread_history tmp; | |
630 | ||
631 | if (!thread) { | |
632 | thread = XCALLOC(MTYPE_THREAD, sizeof(struct thread)); | |
633 | /* mutex only needs to be initialized at struct creation. */ | |
634 | pthread_mutex_init(&thread->mtx, NULL); | |
635 | m->alloc++; | |
636 | } | |
637 | ||
638 | thread->type = type; | |
639 | thread->add_type = type; | |
640 | thread->master = m; | |
641 | thread->arg = arg; | |
642 | thread->index = -1; | |
643 | thread->yield = THREAD_YIELD_TIME_SLOT; /* default */ | |
644 | thread->ref = NULL; | |
645 | ||
646 | /* | |
647 | * So if the passed in funcname is not what we have | |
648 | * stored that means the thread->hist needs to be | |
649 | * updated. We keep the last one around in unused | |
650 | * under the assumption that we are probably | |
651 | * going to immediately allocate the same | |
652 | * type of thread. | |
653 | * This hopefully saves us some serious | |
654 | * hash_get lookups. | |
655 | */ | |
656 | if (thread->funcname != funcname || thread->func != func) { | |
657 | tmp.func = func; | |
658 | tmp.funcname = funcname; | |
659 | thread->hist = | |
660 | hash_get(m->cpu_record, &tmp, | |
661 | (void *(*)(void *))cpu_record_hash_alloc); | |
662 | } | |
663 | thread->hist->total_active++; | |
664 | thread->func = func; | |
665 | thread->funcname = funcname; | |
666 | thread->schedfrom = schedfrom; | |
667 | thread->schedfrom_line = fromln; | |
668 | ||
669 | return thread; | |
718e3744 | 670 | } |
671 | ||
d62a17ae | 672 | static int fd_poll(struct thread_master *m, struct pollfd *pfds, nfds_t pfdsize, |
673 | nfds_t count, const struct timeval *timer_wait) | |
209a72a6 | 674 | { |
d62a17ae | 675 | /* If timer_wait is null here, that means poll() should block |
676 | * indefinitely, | |
677 | * unless the thread_master has overriden it by setting | |
678 | * ->selectpoll_timeout. | |
679 | * If the value is positive, it specifies the maximum number of | |
680 | * milliseconds | |
681 | * to wait. If the timeout is -1, it specifies that we should never wait | |
682 | * and | |
683 | * always return immediately even if no event is detected. If the value | |
684 | * is | |
685 | * zero, the behavior is default. */ | |
686 | int timeout = -1; | |
687 | ||
688 | /* number of file descriptors with events */ | |
689 | int num; | |
690 | ||
691 | if (timer_wait != NULL | |
692 | && m->selectpoll_timeout == 0) // use the default value | |
693 | timeout = (timer_wait->tv_sec * 1000) | |
694 | + (timer_wait->tv_usec / 1000); | |
695 | else if (m->selectpoll_timeout > 0) // use the user's timeout | |
696 | timeout = m->selectpoll_timeout; | |
697 | else if (m->selectpoll_timeout | |
698 | < 0) // effect a poll (return immediately) | |
699 | timeout = 0; | |
700 | ||
701 | /* add poll pipe poker */ | |
702 | assert(count + 1 < pfdsize); | |
703 | pfds[count].fd = m->io_pipe[0]; | |
704 | pfds[count].events = POLLIN; | |
705 | pfds[count].revents = 0x00; | |
706 | ||
707 | num = poll(pfds, count + 1, timeout); | |
708 | ||
709 | unsigned char trash[64]; | |
710 | if (num > 0 && pfds[count].revents != 0 && num--) | |
711 | while (read(m->io_pipe[0], &trash, sizeof(trash)) > 0) | |
712 | ; | |
713 | ||
714 | return num; | |
209a72a6 DS |
715 | } |
716 | ||
718e3744 | 717 | /* Add new read thread. */ |
d62a17ae | 718 | struct thread *funcname_thread_add_read_write(int dir, struct thread_master *m, |
719 | int (*func)(struct thread *), | |
720 | void *arg, int fd, | |
721 | struct thread **t_ptr, | |
722 | debugargdef) | |
718e3744 | 723 | { |
d62a17ae | 724 | struct thread *thread = NULL; |
725 | ||
726 | pthread_mutex_lock(&m->mtx); | |
727 | { | |
728 | if (t_ptr | |
729 | && *t_ptr) // thread is already scheduled; don't reschedule | |
730 | { | |
731 | pthread_mutex_unlock(&m->mtx); | |
732 | return NULL; | |
733 | } | |
734 | ||
735 | /* default to a new pollfd */ | |
736 | nfds_t queuepos = m->handler.pfdcount; | |
737 | ||
738 | /* if we already have a pollfd for our file descriptor, find and | |
739 | * use it */ | |
740 | for (nfds_t i = 0; i < m->handler.pfdcount; i++) | |
741 | if (m->handler.pfds[i].fd == fd) { | |
742 | queuepos = i; | |
743 | break; | |
744 | } | |
745 | ||
746 | /* make sure we have room for this fd + pipe poker fd */ | |
747 | assert(queuepos + 1 < m->handler.pfdsize); | |
748 | ||
749 | thread = thread_get(m, dir, func, arg, debugargpass); | |
750 | ||
751 | m->handler.pfds[queuepos].fd = fd; | |
752 | m->handler.pfds[queuepos].events |= | |
753 | (dir == THREAD_READ ? POLLIN : POLLOUT); | |
754 | ||
755 | if (queuepos == m->handler.pfdcount) | |
756 | m->handler.pfdcount++; | |
757 | ||
758 | if (thread) { | |
759 | pthread_mutex_lock(&thread->mtx); | |
760 | { | |
761 | thread->u.fd = fd; | |
762 | if (dir == THREAD_READ) | |
763 | m->read[thread->u.fd] = thread; | |
764 | else | |
765 | m->write[thread->u.fd] = thread; | |
766 | } | |
767 | pthread_mutex_unlock(&thread->mtx); | |
768 | ||
769 | if (t_ptr) { | |
770 | *t_ptr = thread; | |
771 | thread->ref = t_ptr; | |
772 | } | |
773 | } | |
774 | ||
775 | AWAKEN(m); | |
776 | } | |
777 | pthread_mutex_unlock(&m->mtx); | |
778 | ||
779 | return thread; | |
718e3744 | 780 | } |
781 | ||
56a94b36 | 782 | static struct thread * |
d62a17ae | 783 | funcname_thread_add_timer_timeval(struct thread_master *m, |
784 | int (*func)(struct thread *), int type, | |
785 | void *arg, struct timeval *time_relative, | |
786 | struct thread **t_ptr, debugargdef) | |
718e3744 | 787 | { |
d62a17ae | 788 | struct thread *thread; |
789 | struct pqueue *queue; | |
790 | ||
791 | assert(m != NULL); | |
792 | ||
793 | assert(type == THREAD_TIMER); | |
794 | assert(time_relative); | |
795 | ||
796 | pthread_mutex_lock(&m->mtx); | |
797 | { | |
798 | if (t_ptr | |
799 | && *t_ptr) // thread is already scheduled; don't reschedule | |
800 | { | |
801 | pthread_mutex_unlock(&m->mtx); | |
802 | return NULL; | |
803 | } | |
804 | ||
805 | queue = m->timer; | |
806 | thread = thread_get(m, type, func, arg, debugargpass); | |
807 | ||
808 | pthread_mutex_lock(&thread->mtx); | |
809 | { | |
810 | monotime(&thread->u.sands); | |
811 | timeradd(&thread->u.sands, time_relative, | |
812 | &thread->u.sands); | |
813 | pqueue_enqueue(thread, queue); | |
814 | if (t_ptr) { | |
815 | *t_ptr = thread; | |
816 | thread->ref = t_ptr; | |
817 | } | |
818 | } | |
819 | pthread_mutex_unlock(&thread->mtx); | |
820 | ||
821 | AWAKEN(m); | |
822 | } | |
823 | pthread_mutex_unlock(&m->mtx); | |
824 | ||
825 | return thread; | |
9e867fe6 | 826 | } |
827 | ||
98c91ac6 | 828 | |
829 | /* Add timer event thread. */ | |
d62a17ae | 830 | struct thread *funcname_thread_add_timer(struct thread_master *m, |
831 | int (*func)(struct thread *), | |
832 | void *arg, long timer, | |
833 | struct thread **t_ptr, debugargdef) | |
9e867fe6 | 834 | { |
d62a17ae | 835 | struct timeval trel; |
9e867fe6 | 836 | |
d62a17ae | 837 | assert(m != NULL); |
9e867fe6 | 838 | |
d62a17ae | 839 | trel.tv_sec = timer; |
840 | trel.tv_usec = 0; | |
9e867fe6 | 841 | |
d62a17ae | 842 | return funcname_thread_add_timer_timeval(m, func, THREAD_TIMER, arg, |
843 | &trel, t_ptr, debugargpass); | |
98c91ac6 | 844 | } |
9e867fe6 | 845 | |
98c91ac6 | 846 | /* Add timer event thread with "millisecond" resolution */ |
d62a17ae | 847 | struct thread *funcname_thread_add_timer_msec(struct thread_master *m, |
848 | int (*func)(struct thread *), | |
849 | void *arg, long timer, | |
850 | struct thread **t_ptr, | |
851 | debugargdef) | |
98c91ac6 | 852 | { |
d62a17ae | 853 | struct timeval trel; |
9e867fe6 | 854 | |
d62a17ae | 855 | assert(m != NULL); |
718e3744 | 856 | |
d62a17ae | 857 | trel.tv_sec = timer / 1000; |
858 | trel.tv_usec = 1000 * (timer % 1000); | |
98c91ac6 | 859 | |
d62a17ae | 860 | return funcname_thread_add_timer_timeval(m, func, THREAD_TIMER, arg, |
861 | &trel, t_ptr, debugargpass); | |
a48b4e6d | 862 | } |
863 | ||
d03c4cbd | 864 | /* Add timer event thread with "millisecond" resolution */ |
d62a17ae | 865 | struct thread *funcname_thread_add_timer_tv(struct thread_master *m, |
866 | int (*func)(struct thread *), | |
867 | void *arg, struct timeval *tv, | |
868 | struct thread **t_ptr, debugargdef) | |
d03c4cbd | 869 | { |
d62a17ae | 870 | return funcname_thread_add_timer_timeval(m, func, THREAD_TIMER, arg, tv, |
871 | t_ptr, debugargpass); | |
d03c4cbd DL |
872 | } |
873 | ||
718e3744 | 874 | /* Add simple event thread. */ |
d62a17ae | 875 | struct thread *funcname_thread_add_event(struct thread_master *m, |
876 | int (*func)(struct thread *), | |
877 | void *arg, int val, | |
878 | struct thread **t_ptr, debugargdef) | |
718e3744 | 879 | { |
d62a17ae | 880 | struct thread *thread; |
881 | ||
882 | assert(m != NULL); | |
883 | ||
884 | pthread_mutex_lock(&m->mtx); | |
885 | { | |
886 | if (t_ptr | |
887 | && *t_ptr) // thread is already scheduled; don't reschedule | |
888 | { | |
889 | pthread_mutex_unlock(&m->mtx); | |
890 | return NULL; | |
891 | } | |
892 | ||
893 | thread = thread_get(m, THREAD_EVENT, func, arg, debugargpass); | |
894 | pthread_mutex_lock(&thread->mtx); | |
895 | { | |
896 | thread->u.val = val; | |
897 | thread_list_add(&m->event, thread); | |
898 | } | |
899 | pthread_mutex_unlock(&thread->mtx); | |
900 | ||
901 | if (t_ptr) { | |
902 | *t_ptr = thread; | |
903 | thread->ref = t_ptr; | |
904 | } | |
905 | ||
906 | AWAKEN(m); | |
907 | } | |
908 | pthread_mutex_unlock(&m->mtx); | |
909 | ||
910 | return thread; | |
718e3744 | 911 | } |
912 | ||
63ccb9cb QY |
913 | /* Thread cancellation ------------------------------------------------------ */ |
914 | ||
8797240e QY |
915 | /** |
916 | * NOT's out the .events field of pollfd corresponding to the given file | |
917 | * descriptor. The event to be NOT'd is passed in the 'state' parameter. | |
918 | * | |
919 | * This needs to happen for both copies of pollfd's. See 'thread_fetch' | |
920 | * implementation for details. | |
921 | * | |
922 | * @param master | |
923 | * @param fd | |
924 | * @param state the event to cancel. One or more (OR'd together) of the | |
925 | * following: | |
926 | * - POLLIN | |
927 | * - POLLOUT | |
928 | */ | |
d62a17ae | 929 | static void thread_cancel_rw(struct thread_master *master, int fd, short state) |
0a95a0d0 | 930 | { |
42d74538 QY |
931 | bool found = false; |
932 | ||
d62a17ae | 933 | /* Cancel POLLHUP too just in case some bozo set it */ |
934 | state |= POLLHUP; | |
935 | ||
936 | /* find the index of corresponding pollfd */ | |
937 | nfds_t i; | |
938 | ||
939 | for (i = 0; i < master->handler.pfdcount; i++) | |
42d74538 QY |
940 | if (master->handler.pfds[i].fd == fd) { |
941 | found = true; | |
d62a17ae | 942 | break; |
42d74538 QY |
943 | } |
944 | ||
945 | if (!found) { | |
946 | zlog_debug( | |
947 | "[!] Received cancellation request for nonexistent rw job"); | |
948 | zlog_debug("[!] threadmaster: %s | fd: %d", | |
996c9314 | 949 | master->name ? master->name : "", fd); |
42d74538 QY |
950 | return; |
951 | } | |
d62a17ae | 952 | |
953 | /* NOT out event. */ | |
954 | master->handler.pfds[i].events &= ~(state); | |
955 | ||
956 | /* If all events are canceled, delete / resize the pollfd array. */ | |
957 | if (master->handler.pfds[i].events == 0) { | |
958 | memmove(master->handler.pfds + i, master->handler.pfds + i + 1, | |
959 | (master->handler.pfdcount - i - 1) | |
960 | * sizeof(struct pollfd)); | |
961 | master->handler.pfdcount--; | |
962 | } | |
963 | ||
964 | /* If we have the same pollfd in the copy, perform the same operations, | |
965 | * otherwise return. */ | |
966 | if (i >= master->handler.copycount) | |
967 | return; | |
968 | ||
969 | master->handler.copy[i].events &= ~(state); | |
970 | ||
971 | if (master->handler.copy[i].events == 0) { | |
972 | memmove(master->handler.copy + i, master->handler.copy + i + 1, | |
973 | (master->handler.copycount - i - 1) | |
974 | * sizeof(struct pollfd)); | |
975 | master->handler.copycount--; | |
976 | } | |
0a95a0d0 DS |
977 | } |
978 | ||
1189d95f | 979 | /** |
63ccb9cb | 980 | * Process cancellation requests. |
1189d95f | 981 | * |
63ccb9cb QY |
982 | * This may only be run from the pthread which owns the thread_master. |
983 | * | |
984 | * @param master the thread master to process | |
985 | * @REQUIRE master->mtx | |
1189d95f | 986 | */ |
d62a17ae | 987 | static void do_thread_cancel(struct thread_master *master) |
718e3744 | 988 | { |
d62a17ae | 989 | struct thread_list *list = NULL; |
990 | struct pqueue *queue = NULL; | |
991 | struct thread **thread_array = NULL; | |
992 | struct thread *thread; | |
993 | ||
994 | struct cancel_req *cr; | |
995 | struct listnode *ln; | |
996 | for (ALL_LIST_ELEMENTS_RO(master->cancel_req, ln, cr)) { | |
997 | /* If this is an event object cancellation, linear search | |
998 | * through event | |
999 | * list deleting any events which have the specified argument. | |
1000 | * We also | |
1001 | * need to check every thread in the ready queue. */ | |
1002 | if (cr->eventobj) { | |
1003 | struct thread *t; | |
1004 | thread = master->event.head; | |
1005 | ||
1006 | while (thread) { | |
1007 | t = thread; | |
1008 | thread = t->next; | |
1009 | ||
1010 | if (t->arg == cr->eventobj) { | |
1011 | thread_list_delete(&master->event, t); | |
1012 | if (t->ref) | |
1013 | *t->ref = NULL; | |
1014 | thread_add_unuse(master, t); | |
1015 | } | |
1016 | } | |
1017 | ||
1018 | thread = master->ready.head; | |
1019 | while (thread) { | |
1020 | t = thread; | |
1021 | thread = t->next; | |
1022 | ||
1023 | if (t->arg == cr->eventobj) { | |
1024 | thread_list_delete(&master->ready, t); | |
1025 | if (t->ref) | |
1026 | *t->ref = NULL; | |
1027 | thread_add_unuse(master, t); | |
1028 | } | |
1029 | } | |
1030 | continue; | |
1031 | } | |
1032 | ||
1033 | /* The pointer varies depending on whether the cancellation | |
1034 | * request was | |
1035 | * made asynchronously or not. If it was, we need to check | |
1036 | * whether the | |
1037 | * thread even exists anymore before cancelling it. */ | |
1038 | thread = (cr->thread) ? cr->thread : *cr->threadref; | |
1039 | ||
1040 | if (!thread) | |
1041 | continue; | |
1042 | ||
1043 | /* Determine the appropriate queue to cancel the thread from */ | |
1044 | switch (thread->type) { | |
1045 | case THREAD_READ: | |
1046 | thread_cancel_rw(master, thread->u.fd, POLLIN); | |
1047 | thread_array = master->read; | |
1048 | break; | |
1049 | case THREAD_WRITE: | |
1050 | thread_cancel_rw(master, thread->u.fd, POLLOUT); | |
1051 | thread_array = master->write; | |
1052 | break; | |
1053 | case THREAD_TIMER: | |
1054 | queue = master->timer; | |
1055 | break; | |
1056 | case THREAD_EVENT: | |
1057 | list = &master->event; | |
1058 | break; | |
1059 | case THREAD_READY: | |
1060 | list = &master->ready; | |
1061 | break; | |
1062 | default: | |
1063 | continue; | |
1064 | break; | |
1065 | } | |
1066 | ||
1067 | if (queue) { | |
1068 | assert(thread->index >= 0); | |
522f7f99 DS |
1069 | assert(thread == queue->array[thread->index]); |
1070 | pqueue_remove_at(thread->index, queue); | |
d62a17ae | 1071 | } else if (list) { |
1072 | thread_list_delete(list, thread); | |
1073 | } else if (thread_array) { | |
1074 | thread_array[thread->u.fd] = NULL; | |
1075 | } else { | |
1076 | assert(!"Thread should be either in queue or list or array!"); | |
1077 | } | |
1078 | ||
1079 | if (thread->ref) | |
1080 | *thread->ref = NULL; | |
1081 | ||
1082 | thread_add_unuse(thread->master, thread); | |
1083 | } | |
1084 | ||
1085 | /* Delete and free all cancellation requests */ | |
1086 | list_delete_all_node(master->cancel_req); | |
1087 | ||
1088 | /* Wake up any threads which may be blocked in thread_cancel_async() */ | |
1089 | master->canceled = true; | |
1090 | pthread_cond_broadcast(&master->cancel_cond); | |
718e3744 | 1091 | } |
1092 | ||
63ccb9cb QY |
1093 | /** |
1094 | * Cancel any events which have the specified argument. | |
1095 | * | |
1096 | * MT-Unsafe | |
1097 | * | |
1098 | * @param m the thread_master to cancel from | |
1099 | * @param arg the argument passed when creating the event | |
1100 | */ | |
d62a17ae | 1101 | void thread_cancel_event(struct thread_master *master, void *arg) |
718e3744 | 1102 | { |
d62a17ae | 1103 | assert(master->owner == pthread_self()); |
1104 | ||
1105 | pthread_mutex_lock(&master->mtx); | |
1106 | { | |
1107 | struct cancel_req *cr = | |
1108 | XCALLOC(MTYPE_TMP, sizeof(struct cancel_req)); | |
1109 | cr->eventobj = arg; | |
1110 | listnode_add(master->cancel_req, cr); | |
1111 | do_thread_cancel(master); | |
1112 | } | |
1113 | pthread_mutex_unlock(&master->mtx); | |
63ccb9cb | 1114 | } |
1189d95f | 1115 | |
63ccb9cb QY |
1116 | /** |
1117 | * Cancel a specific task. | |
1118 | * | |
1119 | * MT-Unsafe | |
1120 | * | |
1121 | * @param thread task to cancel | |
1122 | */ | |
d62a17ae | 1123 | void thread_cancel(struct thread *thread) |
63ccb9cb | 1124 | { |
d62a17ae | 1125 | assert(thread->master->owner == pthread_self()); |
1126 | ||
1127 | pthread_mutex_lock(&thread->master->mtx); | |
1128 | { | |
1129 | struct cancel_req *cr = | |
1130 | XCALLOC(MTYPE_TMP, sizeof(struct cancel_req)); | |
1131 | cr->thread = thread; | |
1132 | listnode_add(thread->master->cancel_req, cr); | |
1133 | do_thread_cancel(thread->master); | |
1134 | } | |
1135 | pthread_mutex_unlock(&thread->master->mtx); | |
63ccb9cb | 1136 | } |
1189d95f | 1137 | |
63ccb9cb QY |
1138 | /** |
1139 | * Asynchronous cancellation. | |
1140 | * | |
8797240e QY |
1141 | * Called with either a struct thread ** or void * to an event argument, |
1142 | * this function posts the correct cancellation request and blocks until it is | |
1143 | * serviced. | |
63ccb9cb QY |
1144 | * |
1145 | * If the thread is currently running, execution blocks until it completes. | |
1146 | * | |
8797240e QY |
1147 | * The last two parameters are mutually exclusive, i.e. if you pass one the |
1148 | * other must be NULL. | |
1149 | * | |
1150 | * When the cancellation procedure executes on the target thread_master, the | |
1151 | * thread * provided is checked for nullity. If it is null, the thread is | |
1152 | * assumed to no longer exist and the cancellation request is a no-op. Thus | |
1153 | * users of this API must pass a back-reference when scheduling the original | |
1154 | * task. | |
1155 | * | |
63ccb9cb QY |
1156 | * MT-Safe |
1157 | * | |
8797240e QY |
1158 | * @param master the thread master with the relevant event / task |
1159 | * @param thread pointer to thread to cancel | |
1160 | * @param eventobj the event | |
63ccb9cb | 1161 | */ |
d62a17ae | 1162 | void thread_cancel_async(struct thread_master *master, struct thread **thread, |
1163 | void *eventobj) | |
63ccb9cb | 1164 | { |
d62a17ae | 1165 | assert(!(thread && eventobj) && (thread || eventobj)); |
1166 | assert(master->owner != pthread_self()); | |
1167 | ||
1168 | pthread_mutex_lock(&master->mtx); | |
1169 | { | |
1170 | master->canceled = false; | |
1171 | ||
1172 | if (thread) { | |
1173 | struct cancel_req *cr = | |
1174 | XCALLOC(MTYPE_TMP, sizeof(struct cancel_req)); | |
1175 | cr->threadref = thread; | |
1176 | listnode_add(master->cancel_req, cr); | |
1177 | } else if (eventobj) { | |
1178 | struct cancel_req *cr = | |
1179 | XCALLOC(MTYPE_TMP, sizeof(struct cancel_req)); | |
1180 | cr->eventobj = eventobj; | |
1181 | listnode_add(master->cancel_req, cr); | |
1182 | } | |
1183 | AWAKEN(master); | |
1184 | ||
1185 | while (!master->canceled) | |
1186 | pthread_cond_wait(&master->cancel_cond, &master->mtx); | |
1187 | } | |
1188 | pthread_mutex_unlock(&master->mtx); | |
718e3744 | 1189 | } |
63ccb9cb | 1190 | /* ------------------------------------------------------------------------- */ |
718e3744 | 1191 | |
d62a17ae | 1192 | static struct timeval *thread_timer_wait(struct pqueue *queue, |
1193 | struct timeval *timer_val) | |
718e3744 | 1194 | { |
d62a17ae | 1195 | if (queue->size) { |
1196 | struct thread *next_timer = queue->array[0]; | |
1197 | monotime_until(&next_timer->u.sands, timer_val); | |
1198 | return timer_val; | |
1199 | } | |
1200 | return NULL; | |
718e3744 | 1201 | } |
718e3744 | 1202 | |
d62a17ae | 1203 | static struct thread *thread_run(struct thread_master *m, struct thread *thread, |
1204 | struct thread *fetch) | |
718e3744 | 1205 | { |
d62a17ae | 1206 | *fetch = *thread; |
1207 | thread_add_unuse(m, thread); | |
1208 | return fetch; | |
718e3744 | 1209 | } |
1210 | ||
d62a17ae | 1211 | static int thread_process_io_helper(struct thread_master *m, |
1212 | struct thread *thread, short state, int pos) | |
5d4ccd4e | 1213 | { |
d62a17ae | 1214 | struct thread **thread_array; |
1215 | ||
1216 | if (!thread) | |
1217 | return 0; | |
1218 | ||
1219 | if (thread->type == THREAD_READ) | |
1220 | thread_array = m->read; | |
1221 | else | |
1222 | thread_array = m->write; | |
1223 | ||
1224 | thread_array[thread->u.fd] = NULL; | |
1225 | thread_list_add(&m->ready, thread); | |
1226 | thread->type = THREAD_READY; | |
1227 | /* if another pthread scheduled this file descriptor for the event we're | |
1228 | * responding to, no problem; we're getting to it now */ | |
1229 | thread->master->handler.pfds[pos].events &= ~(state); | |
1230 | return 1; | |
5d4ccd4e DS |
1231 | } |
1232 | ||
8797240e QY |
1233 | /** |
1234 | * Process I/O events. | |
1235 | * | |
1236 | * Walks through file descriptor array looking for those pollfds whose .revents | |
1237 | * field has something interesting. Deletes any invalid file descriptors. | |
1238 | * | |
1239 | * @param m the thread master | |
1240 | * @param num the number of active file descriptors (return value of poll()) | |
1241 | */ | |
d62a17ae | 1242 | static void thread_process_io(struct thread_master *m, unsigned int num) |
0a95a0d0 | 1243 | { |
d62a17ae | 1244 | unsigned int ready = 0; |
1245 | struct pollfd *pfds = m->handler.copy; | |
1246 | ||
1247 | for (nfds_t i = 0; i < m->handler.copycount && ready < num; ++i) { | |
1248 | /* no event for current fd? immediately continue */ | |
1249 | if (pfds[i].revents == 0) | |
1250 | continue; | |
1251 | ||
1252 | ready++; | |
1253 | ||
1254 | /* Unless someone has called thread_cancel from another pthread, | |
1255 | * the only | |
1256 | * thing that could have changed in m->handler.pfds while we | |
1257 | * were | |
1258 | * asleep is the .events field in a given pollfd. Barring | |
1259 | * thread_cancel() | |
1260 | * that value should be a superset of the values we have in our | |
1261 | * copy, so | |
1262 | * there's no need to update it. Similarily, barring deletion, | |
1263 | * the fd | |
1264 | * should still be a valid index into the master's pfds. */ | |
1265 | if (pfds[i].revents & (POLLIN | POLLHUP)) | |
1266 | thread_process_io_helper(m, m->read[pfds[i].fd], POLLIN, | |
1267 | i); | |
1268 | if (pfds[i].revents & POLLOUT) | |
1269 | thread_process_io_helper(m, m->write[pfds[i].fd], | |
1270 | POLLOUT, i); | |
1271 | ||
1272 | /* if one of our file descriptors is garbage, remove the same | |
1273 | * from | |
1274 | * both pfds + update sizes and index */ | |
1275 | if (pfds[i].revents & POLLNVAL) { | |
1276 | memmove(m->handler.pfds + i, m->handler.pfds + i + 1, | |
1277 | (m->handler.pfdcount - i - 1) | |
1278 | * sizeof(struct pollfd)); | |
1279 | m->handler.pfdcount--; | |
1280 | ||
1281 | memmove(pfds + i, pfds + i + 1, | |
1282 | (m->handler.copycount - i - 1) | |
1283 | * sizeof(struct pollfd)); | |
1284 | m->handler.copycount--; | |
1285 | ||
1286 | i--; | |
1287 | } | |
1288 | } | |
718e3744 | 1289 | } |
1290 | ||
8b70d0b0 | 1291 | /* Add all timers that have popped to the ready list. */ |
d62a17ae | 1292 | static unsigned int thread_process_timers(struct pqueue *queue, |
1293 | struct timeval *timenow) | |
a48b4e6d | 1294 | { |
d62a17ae | 1295 | struct thread *thread; |
1296 | unsigned int ready = 0; | |
1297 | ||
1298 | while (queue->size) { | |
1299 | thread = queue->array[0]; | |
1300 | if (timercmp(timenow, &thread->u.sands, <)) | |
1301 | return ready; | |
1302 | pqueue_dequeue(queue); | |
1303 | thread->type = THREAD_READY; | |
1304 | thread_list_add(&thread->master->ready, thread); | |
1305 | ready++; | |
1306 | } | |
1307 | return ready; | |
a48b4e6d | 1308 | } |
1309 | ||
2613abe6 | 1310 | /* process a list en masse, e.g. for event thread lists */ |
d62a17ae | 1311 | static unsigned int thread_process(struct thread_list *list) |
2613abe6 | 1312 | { |
d62a17ae | 1313 | struct thread *thread; |
1314 | struct thread *next; | |
1315 | unsigned int ready = 0; | |
1316 | ||
1317 | for (thread = list->head; thread; thread = next) { | |
1318 | next = thread->next; | |
1319 | thread_list_delete(list, thread); | |
1320 | thread->type = THREAD_READY; | |
1321 | thread_list_add(&thread->master->ready, thread); | |
1322 | ready++; | |
1323 | } | |
1324 | return ready; | |
2613abe6 PJ |
1325 | } |
1326 | ||
1327 | ||
718e3744 | 1328 | /* Fetch next ready thread. */ |
d62a17ae | 1329 | struct thread *thread_fetch(struct thread_master *m, struct thread *fetch) |
718e3744 | 1330 | { |
d62a17ae | 1331 | struct thread *thread = NULL; |
1332 | struct timeval now; | |
1333 | struct timeval zerotime = {0, 0}; | |
1334 | struct timeval tv; | |
1335 | struct timeval *tw = NULL; | |
1336 | ||
1337 | int num = 0; | |
1338 | ||
1339 | do { | |
1340 | /* Handle signals if any */ | |
1341 | if (m->handle_signals) | |
1342 | quagga_sigevent_process(); | |
1343 | ||
1344 | pthread_mutex_lock(&m->mtx); | |
1345 | ||
1346 | /* Process any pending cancellation requests */ | |
1347 | do_thread_cancel(m); | |
1348 | ||
e3c9529e QY |
1349 | /* |
1350 | * Attempt to flush ready queue before going into poll(). | |
1351 | * This is performance-critical. Think twice before modifying. | |
1352 | */ | |
1353 | if ((thread = thread_trim_head(&m->ready))) { | |
1354 | fetch = thread_run(m, thread, fetch); | |
1355 | if (fetch->ref) | |
1356 | *fetch->ref = NULL; | |
1357 | pthread_mutex_unlock(&m->mtx); | |
1358 | break; | |
1359 | } | |
1360 | ||
1361 | /* otherwise, tick through scheduling sequence */ | |
1362 | ||
bca37d17 QY |
1363 | /* |
1364 | * Post events to ready queue. This must come before the | |
1365 | * following block since events should occur immediately | |
1366 | */ | |
d62a17ae | 1367 | thread_process(&m->event); |
1368 | ||
bca37d17 QY |
1369 | /* |
1370 | * If there are no tasks on the ready queue, we will poll() | |
1371 | * until a timer expires or we receive I/O, whichever comes | |
1372 | * first. The strategy for doing this is: | |
d62a17ae | 1373 | * |
1374 | * - If there are events pending, set the poll() timeout to zero | |
1375 | * - If there are no events pending, but there are timers | |
1376 | * pending, set the | |
1377 | * timeout to the smallest remaining time on any timer | |
1378 | * - If there are neither timers nor events pending, but there | |
1379 | * are file | |
1380 | * descriptors pending, block indefinitely in poll() | |
1381 | * - If nothing is pending, it's time for the application to die | |
1382 | * | |
1383 | * In every case except the last, we need to hit poll() at least | |
bca37d17 QY |
1384 | * once per loop to avoid starvation by events |
1385 | */ | |
d62a17ae | 1386 | if (m->ready.count == 0) |
1387 | tw = thread_timer_wait(m->timer, &tv); | |
1388 | ||
1389 | if (m->ready.count != 0 || (tw && !timercmp(tw, &zerotime, >))) | |
1390 | tw = &zerotime; | |
1391 | ||
1392 | if (!tw && m->handler.pfdcount == 0) { /* die */ | |
1393 | pthread_mutex_unlock(&m->mtx); | |
1394 | fetch = NULL; | |
1395 | break; | |
1396 | } | |
1397 | ||
bca37d17 QY |
1398 | /* |
1399 | * Copy pollfd array + # active pollfds in it. Not necessary to | |
1400 | * copy the array size as this is fixed. | |
1401 | */ | |
d62a17ae | 1402 | m->handler.copycount = m->handler.pfdcount; |
1403 | memcpy(m->handler.copy, m->handler.pfds, | |
1404 | m->handler.copycount * sizeof(struct pollfd)); | |
1405 | ||
e3c9529e QY |
1406 | pthread_mutex_unlock(&m->mtx); |
1407 | { | |
1408 | num = fd_poll(m, m->handler.copy, m->handler.pfdsize, | |
1409 | m->handler.copycount, tw); | |
1410 | } | |
1411 | pthread_mutex_lock(&m->mtx); | |
d764d2cc | 1412 | |
e3c9529e QY |
1413 | /* Handle any errors received in poll() */ |
1414 | if (num < 0) { | |
1415 | if (errno == EINTR) { | |
d62a17ae | 1416 | pthread_mutex_unlock(&m->mtx); |
e3c9529e QY |
1417 | /* loop around to signal handler */ |
1418 | continue; | |
d62a17ae | 1419 | } |
1420 | ||
e3c9529e QY |
1421 | /* else die */ |
1422 | zlog_warn("poll() error: %s", safe_strerror(errno)); | |
1423 | pthread_mutex_unlock(&m->mtx); | |
1424 | fetch = NULL; | |
1425 | break; | |
bca37d17 | 1426 | } |
d62a17ae | 1427 | |
1428 | /* Post timers to ready queue. */ | |
1429 | monotime(&now); | |
1430 | thread_process_timers(m->timer, &now); | |
1431 | ||
1432 | /* Post I/O to ready queue. */ | |
1433 | if (num > 0) | |
1434 | thread_process_io(m, num); | |
1435 | ||
d62a17ae | 1436 | pthread_mutex_unlock(&m->mtx); |
1437 | ||
1438 | } while (!thread && m->spin); | |
1439 | ||
1440 | return fetch; | |
718e3744 | 1441 | } |
1442 | ||
d62a17ae | 1443 | static unsigned long timeval_elapsed(struct timeval a, struct timeval b) |
62f44022 | 1444 | { |
d62a17ae | 1445 | return (((a.tv_sec - b.tv_sec) * TIMER_SECOND_MICRO) |
1446 | + (a.tv_usec - b.tv_usec)); | |
62f44022 QY |
1447 | } |
1448 | ||
d62a17ae | 1449 | unsigned long thread_consumed_time(RUSAGE_T *now, RUSAGE_T *start, |
1450 | unsigned long *cputime) | |
718e3744 | 1451 | { |
d62a17ae | 1452 | /* This is 'user + sys' time. */ |
1453 | *cputime = timeval_elapsed(now->cpu.ru_utime, start->cpu.ru_utime) | |
1454 | + timeval_elapsed(now->cpu.ru_stime, start->cpu.ru_stime); | |
1455 | return timeval_elapsed(now->real, start->real); | |
8b70d0b0 | 1456 | } |
1457 | ||
50596be0 DS |
1458 | /* We should aim to yield after yield milliseconds, which defaults |
1459 | to THREAD_YIELD_TIME_SLOT . | |
8b70d0b0 | 1460 | Note: we are using real (wall clock) time for this calculation. |
1461 | It could be argued that CPU time may make more sense in certain | |
1462 | contexts. The things to consider are whether the thread may have | |
1463 | blocked (in which case wall time increases, but CPU time does not), | |
1464 | or whether the system is heavily loaded with other processes competing | |
d62a17ae | 1465 | for CPU time. On balance, wall clock time seems to make sense. |
8b70d0b0 | 1466 | Plus it has the added benefit that gettimeofday should be faster |
1467 | than calling getrusage. */ | |
d62a17ae | 1468 | int thread_should_yield(struct thread *thread) |
718e3744 | 1469 | { |
d62a17ae | 1470 | int result; |
1471 | pthread_mutex_lock(&thread->mtx); | |
1472 | { | |
1473 | result = monotime_since(&thread->real, NULL) | |
1474 | > (int64_t)thread->yield; | |
1475 | } | |
1476 | pthread_mutex_unlock(&thread->mtx); | |
1477 | return result; | |
50596be0 DS |
1478 | } |
1479 | ||
d62a17ae | 1480 | void thread_set_yield_time(struct thread *thread, unsigned long yield_time) |
50596be0 | 1481 | { |
d62a17ae | 1482 | pthread_mutex_lock(&thread->mtx); |
1483 | { | |
1484 | thread->yield = yield_time; | |
1485 | } | |
1486 | pthread_mutex_unlock(&thread->mtx); | |
718e3744 | 1487 | } |
1488 | ||
d62a17ae | 1489 | void thread_getrusage(RUSAGE_T *r) |
db9c0df9 | 1490 | { |
d62a17ae | 1491 | monotime(&r->real); |
1492 | getrusage(RUSAGE_SELF, &(r->cpu)); | |
db9c0df9 PJ |
1493 | } |
1494 | ||
718e3744 | 1495 | /* We check thread consumed time. If the system has getrusage, we'll |
8b70d0b0 | 1496 | use that to get in-depth stats on the performance of the thread in addition |
1497 | to wall clock time stats from gettimeofday. */ | |
d62a17ae | 1498 | void thread_call(struct thread *thread) |
718e3744 | 1499 | { |
d62a17ae | 1500 | unsigned long realtime, cputime; |
1501 | RUSAGE_T before, after; | |
cc8b13a0 | 1502 | |
d62a17ae | 1503 | GETRUSAGE(&before); |
1504 | thread->real = before.real; | |
718e3744 | 1505 | |
d62a17ae | 1506 | pthread_setspecific(thread_current, thread); |
1507 | (*thread->func)(thread); | |
1508 | pthread_setspecific(thread_current, NULL); | |
718e3744 | 1509 | |
d62a17ae | 1510 | GETRUSAGE(&after); |
718e3744 | 1511 | |
d62a17ae | 1512 | realtime = thread_consumed_time(&after, &before, &cputime); |
1513 | thread->hist->real.total += realtime; | |
1514 | if (thread->hist->real.max < realtime) | |
1515 | thread->hist->real.max = realtime; | |
1516 | thread->hist->cpu.total += cputime; | |
1517 | if (thread->hist->cpu.max < cputime) | |
1518 | thread->hist->cpu.max = cputime; | |
e04ab74d | 1519 | |
d62a17ae | 1520 | ++(thread->hist->total_calls); |
1521 | thread->hist->types |= (1 << thread->add_type); | |
718e3744 | 1522 | |
924b9229 | 1523 | #ifdef CONSUMED_TIME_CHECK |
d62a17ae | 1524 | if (realtime > CONSUMED_TIME_CHECK) { |
1525 | /* | |
1526 | * We have a CPU Hog on our hands. | |
1527 | * Whinge about it now, so we're aware this is yet another task | |
1528 | * to fix. | |
1529 | */ | |
1530 | zlog_warn( | |
1531 | "SLOW THREAD: task %s (%lx) ran for %lums (cpu time %lums)", | |
1532 | thread->funcname, (unsigned long)thread->func, | |
1533 | realtime / 1000, cputime / 1000); | |
1534 | } | |
924b9229 | 1535 | #endif /* CONSUMED_TIME_CHECK */ |
718e3744 | 1536 | } |
1537 | ||
1538 | /* Execute thread */ | |
d62a17ae | 1539 | void funcname_thread_execute(struct thread_master *m, |
1540 | int (*func)(struct thread *), void *arg, int val, | |
1541 | debugargdef) | |
718e3744 | 1542 | { |
d62a17ae | 1543 | struct cpu_thread_history tmp; |
1544 | struct thread dummy; | |
718e3744 | 1545 | |
d62a17ae | 1546 | memset(&dummy, 0, sizeof(struct thread)); |
718e3744 | 1547 | |
d62a17ae | 1548 | pthread_mutex_init(&dummy.mtx, NULL); |
1549 | dummy.type = THREAD_EVENT; | |
1550 | dummy.add_type = THREAD_EXECUTE; | |
1551 | dummy.master = NULL; | |
1552 | dummy.arg = arg; | |
1553 | dummy.u.val = val; | |
9c7753e4 | 1554 | |
d62a17ae | 1555 | tmp.func = dummy.func = func; |
1556 | tmp.funcname = dummy.funcname = funcname; | |
1557 | dummy.hist = hash_get(m->cpu_record, &tmp, | |
1558 | (void *(*)(void *))cpu_record_hash_alloc); | |
f7c62e11 | 1559 | |
d62a17ae | 1560 | dummy.schedfrom = schedfrom; |
1561 | dummy.schedfrom_line = fromln; | |
9c7753e4 | 1562 | |
d62a17ae | 1563 | thread_call(&dummy); |
718e3744 | 1564 | } |