]>
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" | |
3e41733f | 28 | #include "frrcu.h" |
718e3744 | 29 | #include "log.h" |
e04ab74d | 30 | #include "hash.h" |
31 | #include "command.h" | |
05c447dd | 32 | #include "sigevent.h" |
3bf2673b | 33 | #include "network.h" |
bd74dc61 | 34 | #include "jhash.h" |
fbcac826 | 35 | #include "frratomic.h" |
00dffa8c | 36 | #include "frr_pthread.h" |
9ef9495e | 37 | #include "lib_errors.h" |
912d45a1 | 38 | #include "libfrr_trace.h" |
1a9f340b | 39 | #include "libfrr.h" |
d6be5fb9 | 40 | |
d62a17ae | 41 | DEFINE_MTYPE_STATIC(LIB, THREAD, "Thread") |
4a1ab8e4 | 42 | DEFINE_MTYPE_STATIC(LIB, THREAD_MASTER, "Thread master") |
a6f235f3 | 43 | DEFINE_MTYPE_STATIC(LIB, THREAD_POLL, "Thread Poll Info") |
d62a17ae | 44 | DEFINE_MTYPE_STATIC(LIB, THREAD_STATS, "Thread stats") |
4a1ab8e4 | 45 | |
c284542b DL |
46 | DECLARE_LIST(thread_list, struct thread, threaditem) |
47 | ||
27d29ced DL |
48 | static int thread_timer_cmp(const struct thread *a, const struct thread *b) |
49 | { | |
50 | if (a->u.sands.tv_sec < b->u.sands.tv_sec) | |
51 | return -1; | |
52 | if (a->u.sands.tv_sec > b->u.sands.tv_sec) | |
53 | return 1; | |
54 | if (a->u.sands.tv_usec < b->u.sands.tv_usec) | |
55 | return -1; | |
56 | if (a->u.sands.tv_usec > b->u.sands.tv_usec) | |
57 | return 1; | |
58 | return 0; | |
59 | } | |
60 | ||
61 | DECLARE_HEAP(thread_timer_list, struct thread, timeritem, | |
62 | thread_timer_cmp) | |
63 | ||
3b96b781 HT |
64 | #if defined(__APPLE__) |
65 | #include <mach/mach.h> | |
66 | #include <mach/mach_time.h> | |
67 | #endif | |
68 | ||
d62a17ae | 69 | #define AWAKEN(m) \ |
70 | do { \ | |
2b64873d | 71 | const unsigned char wakebyte = 0x01; \ |
d62a17ae | 72 | write(m->io_pipe[1], &wakebyte, 1); \ |
73 | } while (0); | |
3bf2673b | 74 | |
62f44022 | 75 | /* control variable for initializer */ |
c17faa4b | 76 | static pthread_once_t init_once = PTHREAD_ONCE_INIT; |
e0bebc7c | 77 | pthread_key_t thread_current; |
6b0655a2 | 78 | |
c17faa4b | 79 | static pthread_mutex_t masters_mtx = PTHREAD_MUTEX_INITIALIZER; |
62f44022 QY |
80 | static struct list *masters; |
81 | ||
6655966d | 82 | static void thread_free(struct thread_master *master, struct thread *thread); |
6b0655a2 | 83 | |
62f44022 | 84 | /* CLI start ---------------------------------------------------------------- */ |
d8b87afe | 85 | static unsigned int cpu_record_hash_key(const struct cpu_thread_history *a) |
e04ab74d | 86 | { |
883cc51d | 87 | int size = sizeof(a->func); |
bd74dc61 DS |
88 | |
89 | return jhash(&a->func, size, 0); | |
e04ab74d | 90 | } |
91 | ||
74df8d6d | 92 | static bool cpu_record_hash_cmp(const struct cpu_thread_history *a, |
d62a17ae | 93 | const struct cpu_thread_history *b) |
e04ab74d | 94 | { |
d62a17ae | 95 | return a->func == b->func; |
e04ab74d | 96 | } |
97 | ||
d62a17ae | 98 | static void *cpu_record_hash_alloc(struct cpu_thread_history *a) |
e04ab74d | 99 | { |
d62a17ae | 100 | struct cpu_thread_history *new; |
101 | new = XCALLOC(MTYPE_THREAD_STATS, sizeof(struct cpu_thread_history)); | |
102 | new->func = a->func; | |
103 | new->funcname = a->funcname; | |
104 | return new; | |
e04ab74d | 105 | } |
106 | ||
d62a17ae | 107 | static void cpu_record_hash_free(void *a) |
228da428 | 108 | { |
d62a17ae | 109 | struct cpu_thread_history *hist = a; |
110 | ||
111 | XFREE(MTYPE_THREAD_STATS, hist); | |
228da428 CC |
112 | } |
113 | ||
f75e802d | 114 | #ifndef EXCLUDE_CPU_TIME |
d62a17ae | 115 | static void vty_out_cpu_thread_history(struct vty *vty, |
116 | struct cpu_thread_history *a) | |
e04ab74d | 117 | { |
9a8a7b0e | 118 | vty_out(vty, "%5zu %10zu.%03zu %9zu %8zu %9zu %8zu %9zu", |
72327cf3 MS |
119 | a->total_active, a->cpu.total / 1000, a->cpu.total % 1000, |
120 | a->total_calls, (a->cpu.total / a->total_calls), a->cpu.max, | |
121 | (a->real.total / a->total_calls), a->real.max); | |
84d951d0 | 122 | vty_out(vty, " %c%c%c%c%c %s\n", |
d62a17ae | 123 | a->types & (1 << THREAD_READ) ? 'R' : ' ', |
124 | a->types & (1 << THREAD_WRITE) ? 'W' : ' ', | |
125 | a->types & (1 << THREAD_TIMER) ? 'T' : ' ', | |
126 | a->types & (1 << THREAD_EVENT) ? 'E' : ' ', | |
127 | a->types & (1 << THREAD_EXECUTE) ? 'X' : ' ', a->funcname); | |
e04ab74d | 128 | } |
129 | ||
e3b78da8 | 130 | static void cpu_record_hash_print(struct hash_bucket *bucket, void *args[]) |
e04ab74d | 131 | { |
d62a17ae | 132 | struct cpu_thread_history *totals = args[0]; |
fbcac826 | 133 | struct cpu_thread_history copy; |
d62a17ae | 134 | struct vty *vty = args[1]; |
fbcac826 | 135 | uint8_t *filter = args[2]; |
d62a17ae | 136 | |
137 | struct cpu_thread_history *a = bucket->data; | |
138 | ||
fbcac826 QY |
139 | copy.total_active = |
140 | atomic_load_explicit(&a->total_active, memory_order_seq_cst); | |
141 | copy.total_calls = | |
142 | atomic_load_explicit(&a->total_calls, memory_order_seq_cst); | |
143 | copy.cpu.total = | |
144 | atomic_load_explicit(&a->cpu.total, memory_order_seq_cst); | |
145 | copy.cpu.max = atomic_load_explicit(&a->cpu.max, memory_order_seq_cst); | |
146 | copy.real.total = | |
147 | atomic_load_explicit(&a->real.total, memory_order_seq_cst); | |
148 | copy.real.max = | |
149 | atomic_load_explicit(&a->real.max, memory_order_seq_cst); | |
150 | copy.types = atomic_load_explicit(&a->types, memory_order_seq_cst); | |
151 | copy.funcname = a->funcname; | |
152 | ||
153 | if (!(copy.types & *filter)) | |
d62a17ae | 154 | return; |
fbcac826 QY |
155 | |
156 | vty_out_cpu_thread_history(vty, ©); | |
157 | totals->total_active += copy.total_active; | |
158 | totals->total_calls += copy.total_calls; | |
159 | totals->real.total += copy.real.total; | |
160 | if (totals->real.max < copy.real.max) | |
161 | totals->real.max = copy.real.max; | |
162 | totals->cpu.total += copy.cpu.total; | |
163 | if (totals->cpu.max < copy.cpu.max) | |
164 | totals->cpu.max = copy.cpu.max; | |
e04ab74d | 165 | } |
166 | ||
fbcac826 | 167 | static void cpu_record_print(struct vty *vty, uint8_t filter) |
e04ab74d | 168 | { |
d62a17ae | 169 | struct cpu_thread_history tmp; |
170 | void *args[3] = {&tmp, vty, &filter}; | |
171 | struct thread_master *m; | |
172 | struct listnode *ln; | |
173 | ||
0d6f7fd6 | 174 | memset(&tmp, 0, sizeof(tmp)); |
d62a17ae | 175 | tmp.funcname = "TOTAL"; |
176 | tmp.types = filter; | |
177 | ||
00dffa8c | 178 | frr_with_mutex(&masters_mtx) { |
d62a17ae | 179 | for (ALL_LIST_ELEMENTS_RO(masters, ln, m)) { |
180 | const char *name = m->name ? m->name : "main"; | |
181 | ||
182 | char underline[strlen(name) + 1]; | |
183 | memset(underline, '-', sizeof(underline)); | |
4f113d60 | 184 | underline[sizeof(underline) - 1] = '\0'; |
d62a17ae | 185 | |
186 | vty_out(vty, "\n"); | |
187 | vty_out(vty, "Showing statistics for pthread %s\n", | |
188 | name); | |
189 | vty_out(vty, "-------------------------------%s\n", | |
190 | underline); | |
84d951d0 | 191 | vty_out(vty, "%30s %18s %18s\n", "", |
d62a17ae | 192 | "CPU (user+system):", "Real (wall-clock):"); |
193 | vty_out(vty, | |
194 | "Active Runtime(ms) Invoked Avg uSec Max uSecs"); | |
195 | vty_out(vty, " Avg uSec Max uSecs"); | |
196 | vty_out(vty, " Type Thread\n"); | |
197 | ||
198 | if (m->cpu_record->count) | |
199 | hash_iterate( | |
200 | m->cpu_record, | |
e3b78da8 | 201 | (void (*)(struct hash_bucket *, |
d62a17ae | 202 | void *))cpu_record_hash_print, |
203 | args); | |
204 | else | |
205 | vty_out(vty, "No data to display yet.\n"); | |
206 | ||
207 | vty_out(vty, "\n"); | |
208 | } | |
209 | } | |
d62a17ae | 210 | |
211 | vty_out(vty, "\n"); | |
212 | vty_out(vty, "Total thread statistics\n"); | |
213 | vty_out(vty, "-------------------------\n"); | |
84d951d0 | 214 | vty_out(vty, "%30s %18s %18s\n", "", |
d62a17ae | 215 | "CPU (user+system):", "Real (wall-clock):"); |
216 | vty_out(vty, "Active Runtime(ms) Invoked Avg uSec Max uSecs"); | |
217 | vty_out(vty, " Avg uSec Max uSecs"); | |
218 | vty_out(vty, " Type Thread\n"); | |
219 | ||
220 | if (tmp.total_calls > 0) | |
221 | vty_out_cpu_thread_history(vty, &tmp); | |
e04ab74d | 222 | } |
f75e802d | 223 | #endif |
e04ab74d | 224 | |
e3b78da8 | 225 | static void cpu_record_hash_clear(struct hash_bucket *bucket, void *args[]) |
e276eb82 | 226 | { |
fbcac826 | 227 | uint8_t *filter = args[0]; |
d62a17ae | 228 | struct hash *cpu_record = args[1]; |
229 | ||
230 | struct cpu_thread_history *a = bucket->data; | |
62f44022 | 231 | |
d62a17ae | 232 | if (!(a->types & *filter)) |
233 | return; | |
f48f65d2 | 234 | |
d62a17ae | 235 | hash_release(cpu_record, bucket->data); |
e276eb82 PJ |
236 | } |
237 | ||
fbcac826 | 238 | static void cpu_record_clear(uint8_t filter) |
e276eb82 | 239 | { |
fbcac826 | 240 | uint8_t *tmp = &filter; |
d62a17ae | 241 | struct thread_master *m; |
242 | struct listnode *ln; | |
243 | ||
00dffa8c | 244 | frr_with_mutex(&masters_mtx) { |
d62a17ae | 245 | for (ALL_LIST_ELEMENTS_RO(masters, ln, m)) { |
00dffa8c | 246 | frr_with_mutex(&m->mtx) { |
d62a17ae | 247 | void *args[2] = {tmp, m->cpu_record}; |
248 | hash_iterate( | |
249 | m->cpu_record, | |
e3b78da8 | 250 | (void (*)(struct hash_bucket *, |
d62a17ae | 251 | void *))cpu_record_hash_clear, |
252 | args); | |
253 | } | |
d62a17ae | 254 | } |
255 | } | |
62f44022 QY |
256 | } |
257 | ||
fbcac826 | 258 | static uint8_t parse_filter(const char *filterstr) |
62f44022 | 259 | { |
d62a17ae | 260 | int i = 0; |
261 | int filter = 0; | |
262 | ||
263 | while (filterstr[i] != '\0') { | |
264 | switch (filterstr[i]) { | |
265 | case 'r': | |
266 | case 'R': | |
267 | filter |= (1 << THREAD_READ); | |
268 | break; | |
269 | case 'w': | |
270 | case 'W': | |
271 | filter |= (1 << THREAD_WRITE); | |
272 | break; | |
273 | case 't': | |
274 | case 'T': | |
275 | filter |= (1 << THREAD_TIMER); | |
276 | break; | |
277 | case 'e': | |
278 | case 'E': | |
279 | filter |= (1 << THREAD_EVENT); | |
280 | break; | |
281 | case 'x': | |
282 | case 'X': | |
283 | filter |= (1 << THREAD_EXECUTE); | |
284 | break; | |
285 | default: | |
286 | break; | |
287 | } | |
288 | ++i; | |
289 | } | |
290 | return filter; | |
62f44022 QY |
291 | } |
292 | ||
f75e802d | 293 | #ifndef EXCLUDE_CPU_TIME |
62f44022 QY |
294 | DEFUN (show_thread_cpu, |
295 | show_thread_cpu_cmd, | |
296 | "show thread cpu [FILTER]", | |
297 | SHOW_STR | |
298 | "Thread information\n" | |
299 | "Thread CPU usage\n" | |
61fa0b97 | 300 | "Display filter (rwtex)\n") |
62f44022 | 301 | { |
fbcac826 | 302 | uint8_t filter = (uint8_t)-1U; |
d62a17ae | 303 | int idx = 0; |
304 | ||
305 | if (argv_find(argv, argc, "FILTER", &idx)) { | |
306 | filter = parse_filter(argv[idx]->arg); | |
307 | if (!filter) { | |
308 | vty_out(vty, | |
3efd0893 | 309 | "Invalid filter \"%s\" specified; must contain at leastone of 'RWTEXB'\n", |
d62a17ae | 310 | argv[idx]->arg); |
311 | return CMD_WARNING; | |
312 | } | |
313 | } | |
314 | ||
315 | cpu_record_print(vty, filter); | |
316 | return CMD_SUCCESS; | |
e276eb82 | 317 | } |
f75e802d | 318 | #endif |
e276eb82 | 319 | |
8872626b DS |
320 | static void show_thread_poll_helper(struct vty *vty, struct thread_master *m) |
321 | { | |
322 | const char *name = m->name ? m->name : "main"; | |
323 | char underline[strlen(name) + 1]; | |
a0b36ae6 | 324 | struct thread *thread; |
8872626b DS |
325 | uint32_t i; |
326 | ||
327 | memset(underline, '-', sizeof(underline)); | |
328 | underline[sizeof(underline) - 1] = '\0'; | |
329 | ||
330 | vty_out(vty, "\nShowing poll FD's for %s\n", name); | |
331 | vty_out(vty, "----------------------%s\n", underline); | |
6c19478a DS |
332 | vty_out(vty, "Count: %u/%d\n", (uint32_t)m->handler.pfdcount, |
333 | m->fd_limit); | |
a0b36ae6 DS |
334 | for (i = 0; i < m->handler.pfdcount; i++) { |
335 | vty_out(vty, "\t%6d fd:%6d events:%2d revents:%2d\t\t", i, | |
336 | m->handler.pfds[i].fd, m->handler.pfds[i].events, | |
8872626b | 337 | m->handler.pfds[i].revents); |
a0b36ae6 DS |
338 | |
339 | if (m->handler.pfds[i].events & POLLIN) { | |
340 | thread = m->read[m->handler.pfds[i].fd]; | |
341 | ||
342 | if (!thread) | |
343 | vty_out(vty, "ERROR "); | |
344 | else | |
60a3efec | 345 | vty_out(vty, "%s ", thread->xref->funcname); |
a0b36ae6 DS |
346 | } else |
347 | vty_out(vty, " "); | |
348 | ||
349 | if (m->handler.pfds[i].events & POLLOUT) { | |
350 | thread = m->write[m->handler.pfds[i].fd]; | |
351 | ||
352 | if (!thread) | |
353 | vty_out(vty, "ERROR\n"); | |
354 | else | |
60a3efec | 355 | vty_out(vty, "%s\n", thread->xref->funcname); |
a0b36ae6 DS |
356 | } else |
357 | vty_out(vty, "\n"); | |
358 | } | |
8872626b DS |
359 | } |
360 | ||
361 | DEFUN (show_thread_poll, | |
362 | show_thread_poll_cmd, | |
363 | "show thread poll", | |
364 | SHOW_STR | |
365 | "Thread information\n" | |
366 | "Show poll FD's and information\n") | |
367 | { | |
368 | struct listnode *node; | |
369 | struct thread_master *m; | |
370 | ||
00dffa8c | 371 | frr_with_mutex(&masters_mtx) { |
8872626b DS |
372 | for (ALL_LIST_ELEMENTS_RO(masters, node, m)) { |
373 | show_thread_poll_helper(vty, m); | |
374 | } | |
375 | } | |
8872626b DS |
376 | |
377 | return CMD_SUCCESS; | |
378 | } | |
379 | ||
380 | ||
49d41a26 DS |
381 | DEFUN (clear_thread_cpu, |
382 | clear_thread_cpu_cmd, | |
383 | "clear thread cpu [FILTER]", | |
62f44022 | 384 | "Clear stored data in all pthreads\n" |
49d41a26 DS |
385 | "Thread information\n" |
386 | "Thread CPU usage\n" | |
387 | "Display filter (rwtexb)\n") | |
e276eb82 | 388 | { |
fbcac826 | 389 | uint8_t filter = (uint8_t)-1U; |
d62a17ae | 390 | int idx = 0; |
391 | ||
392 | if (argv_find(argv, argc, "FILTER", &idx)) { | |
393 | filter = parse_filter(argv[idx]->arg); | |
394 | if (!filter) { | |
395 | vty_out(vty, | |
3efd0893 | 396 | "Invalid filter \"%s\" specified; must contain at leastone of 'RWTEXB'\n", |
d62a17ae | 397 | argv[idx]->arg); |
398 | return CMD_WARNING; | |
399 | } | |
400 | } | |
401 | ||
402 | cpu_record_clear(filter); | |
403 | return CMD_SUCCESS; | |
e276eb82 | 404 | } |
6b0655a2 | 405 | |
d62a17ae | 406 | void thread_cmd_init(void) |
0b84f294 | 407 | { |
f75e802d | 408 | #ifndef EXCLUDE_CPU_TIME |
d62a17ae | 409 | install_element(VIEW_NODE, &show_thread_cpu_cmd); |
f75e802d | 410 | #endif |
8872626b | 411 | install_element(VIEW_NODE, &show_thread_poll_cmd); |
d62a17ae | 412 | install_element(ENABLE_NODE, &clear_thread_cpu_cmd); |
0b84f294 | 413 | } |
62f44022 QY |
414 | /* CLI end ------------------------------------------------------------------ */ |
415 | ||
0b84f294 | 416 | |
d62a17ae | 417 | static void cancelreq_del(void *cr) |
63ccb9cb | 418 | { |
d62a17ae | 419 | XFREE(MTYPE_TMP, cr); |
63ccb9cb QY |
420 | } |
421 | ||
e0bebc7c | 422 | /* initializer, only ever called once */ |
4d762f26 | 423 | static void initializer(void) |
e0bebc7c | 424 | { |
d62a17ae | 425 | pthread_key_create(&thread_current, NULL); |
e0bebc7c QY |
426 | } |
427 | ||
d62a17ae | 428 | struct thread_master *thread_master_create(const char *name) |
718e3744 | 429 | { |
d62a17ae | 430 | struct thread_master *rv; |
431 | struct rlimit limit; | |
432 | ||
433 | pthread_once(&init_once, &initializer); | |
434 | ||
435 | rv = XCALLOC(MTYPE_THREAD_MASTER, sizeof(struct thread_master)); | |
d62a17ae | 436 | |
437 | /* Initialize master mutex */ | |
438 | pthread_mutex_init(&rv->mtx, NULL); | |
439 | pthread_cond_init(&rv->cancel_cond, NULL); | |
440 | ||
441 | /* Set name */ | |
7ffcd8bd QY |
442 | name = name ? name : "default"; |
443 | rv->name = XSTRDUP(MTYPE_THREAD_MASTER, name); | |
d62a17ae | 444 | |
445 | /* Initialize I/O task data structures */ | |
1a9f340b MS |
446 | |
447 | /* Use configured limit if present, ulimit otherwise. */ | |
448 | rv->fd_limit = frr_get_fd_limit(); | |
449 | if (rv->fd_limit == 0) { | |
450 | getrlimit(RLIMIT_NOFILE, &limit); | |
451 | rv->fd_limit = (int)limit.rlim_cur; | |
452 | } | |
453 | ||
a6f235f3 DS |
454 | rv->read = XCALLOC(MTYPE_THREAD_POLL, |
455 | sizeof(struct thread *) * rv->fd_limit); | |
456 | ||
457 | rv->write = XCALLOC(MTYPE_THREAD_POLL, | |
458 | sizeof(struct thread *) * rv->fd_limit); | |
d62a17ae | 459 | |
7ffcd8bd QY |
460 | char tmhashname[strlen(name) + 32]; |
461 | snprintf(tmhashname, sizeof(tmhashname), "%s - threadmaster event hash", | |
462 | name); | |
bd74dc61 | 463 | rv->cpu_record = hash_create_size( |
d8b87afe | 464 | 8, (unsigned int (*)(const void *))cpu_record_hash_key, |
74df8d6d | 465 | (bool (*)(const void *, const void *))cpu_record_hash_cmp, |
7ffcd8bd | 466 | tmhashname); |
d62a17ae | 467 | |
c284542b DL |
468 | thread_list_init(&rv->event); |
469 | thread_list_init(&rv->ready); | |
470 | thread_list_init(&rv->unuse); | |
27d29ced | 471 | thread_timer_list_init(&rv->timer); |
d62a17ae | 472 | |
473 | /* Initialize thread_fetch() settings */ | |
474 | rv->spin = true; | |
475 | rv->handle_signals = true; | |
476 | ||
477 | /* Set pthread owner, should be updated by actual owner */ | |
478 | rv->owner = pthread_self(); | |
479 | rv->cancel_req = list_new(); | |
480 | rv->cancel_req->del = cancelreq_del; | |
481 | rv->canceled = true; | |
482 | ||
483 | /* Initialize pipe poker */ | |
484 | pipe(rv->io_pipe); | |
485 | set_nonblocking(rv->io_pipe[0]); | |
486 | set_nonblocking(rv->io_pipe[1]); | |
487 | ||
488 | /* Initialize data structures for poll() */ | |
489 | rv->handler.pfdsize = rv->fd_limit; | |
490 | rv->handler.pfdcount = 0; | |
491 | rv->handler.pfds = XCALLOC(MTYPE_THREAD_MASTER, | |
492 | sizeof(struct pollfd) * rv->handler.pfdsize); | |
493 | rv->handler.copy = XCALLOC(MTYPE_THREAD_MASTER, | |
494 | sizeof(struct pollfd) * rv->handler.pfdsize); | |
495 | ||
eff09c66 | 496 | /* add to list of threadmasters */ |
00dffa8c | 497 | frr_with_mutex(&masters_mtx) { |
eff09c66 QY |
498 | if (!masters) |
499 | masters = list_new(); | |
500 | ||
d62a17ae | 501 | listnode_add(masters, rv); |
502 | } | |
d62a17ae | 503 | |
504 | return rv; | |
718e3744 | 505 | } |
506 | ||
d8a8a8de QY |
507 | void thread_master_set_name(struct thread_master *master, const char *name) |
508 | { | |
00dffa8c | 509 | frr_with_mutex(&master->mtx) { |
0a22ddfb | 510 | XFREE(MTYPE_THREAD_MASTER, master->name); |
d8a8a8de QY |
511 | master->name = XSTRDUP(MTYPE_THREAD_MASTER, name); |
512 | } | |
d8a8a8de QY |
513 | } |
514 | ||
6ed04aa2 DS |
515 | #define THREAD_UNUSED_DEPTH 10 |
516 | ||
718e3744 | 517 | /* Move thread to unuse list. */ |
d62a17ae | 518 | static void thread_add_unuse(struct thread_master *m, struct thread *thread) |
718e3744 | 519 | { |
6655966d RZ |
520 | pthread_mutex_t mtxc = thread->mtx; |
521 | ||
d62a17ae | 522 | assert(m != NULL && thread != NULL); |
d62a17ae | 523 | |
d62a17ae | 524 | thread->hist->total_active--; |
6ed04aa2 DS |
525 | memset(thread, 0, sizeof(struct thread)); |
526 | thread->type = THREAD_UNUSED; | |
527 | ||
6655966d RZ |
528 | /* Restore the thread mutex context. */ |
529 | thread->mtx = mtxc; | |
530 | ||
c284542b DL |
531 | if (thread_list_count(&m->unuse) < THREAD_UNUSED_DEPTH) { |
532 | thread_list_add_tail(&m->unuse, thread); | |
6655966d RZ |
533 | return; |
534 | } | |
535 | ||
536 | thread_free(m, thread); | |
718e3744 | 537 | } |
538 | ||
539 | /* Free all unused thread. */ | |
c284542b DL |
540 | static void thread_list_free(struct thread_master *m, |
541 | struct thread_list_head *list) | |
718e3744 | 542 | { |
d62a17ae | 543 | struct thread *t; |
d62a17ae | 544 | |
c284542b | 545 | while ((t = thread_list_pop(list))) |
6655966d | 546 | thread_free(m, t); |
718e3744 | 547 | } |
548 | ||
d62a17ae | 549 | static void thread_array_free(struct thread_master *m, |
550 | struct thread **thread_array) | |
308d14ae | 551 | { |
d62a17ae | 552 | struct thread *t; |
553 | int index; | |
554 | ||
555 | for (index = 0; index < m->fd_limit; ++index) { | |
556 | t = thread_array[index]; | |
557 | if (t) { | |
558 | thread_array[index] = NULL; | |
6655966d | 559 | thread_free(m, t); |
d62a17ae | 560 | } |
561 | } | |
a6f235f3 | 562 | XFREE(MTYPE_THREAD_POLL, thread_array); |
308d14ae DV |
563 | } |
564 | ||
495f0b13 DS |
565 | /* |
566 | * thread_master_free_unused | |
567 | * | |
568 | * As threads are finished with they are put on the | |
569 | * unuse list for later reuse. | |
570 | * If we are shutting down, Free up unused threads | |
571 | * So we can see if we forget to shut anything off | |
572 | */ | |
d62a17ae | 573 | void thread_master_free_unused(struct thread_master *m) |
495f0b13 | 574 | { |
00dffa8c | 575 | frr_with_mutex(&m->mtx) { |
d62a17ae | 576 | struct thread *t; |
c284542b | 577 | while ((t = thread_list_pop(&m->unuse))) |
6655966d | 578 | thread_free(m, t); |
d62a17ae | 579 | } |
495f0b13 DS |
580 | } |
581 | ||
718e3744 | 582 | /* Stop thread scheduler. */ |
d62a17ae | 583 | void thread_master_free(struct thread_master *m) |
718e3744 | 584 | { |
27d29ced DL |
585 | struct thread *t; |
586 | ||
00dffa8c | 587 | frr_with_mutex(&masters_mtx) { |
d62a17ae | 588 | listnode_delete(masters, m); |
eff09c66 | 589 | if (masters->count == 0) { |
6a154c88 | 590 | list_delete(&masters); |
eff09c66 | 591 | } |
d62a17ae | 592 | } |
d62a17ae | 593 | |
594 | thread_array_free(m, m->read); | |
595 | thread_array_free(m, m->write); | |
27d29ced DL |
596 | while ((t = thread_timer_list_pop(&m->timer))) |
597 | thread_free(m, t); | |
d62a17ae | 598 | thread_list_free(m, &m->event); |
599 | thread_list_free(m, &m->ready); | |
600 | thread_list_free(m, &m->unuse); | |
601 | pthread_mutex_destroy(&m->mtx); | |
33844bbe | 602 | pthread_cond_destroy(&m->cancel_cond); |
d62a17ae | 603 | close(m->io_pipe[0]); |
604 | close(m->io_pipe[1]); | |
6a154c88 | 605 | list_delete(&m->cancel_req); |
1a0a92ea | 606 | m->cancel_req = NULL; |
d62a17ae | 607 | |
608 | hash_clean(m->cpu_record, cpu_record_hash_free); | |
609 | hash_free(m->cpu_record); | |
610 | m->cpu_record = NULL; | |
611 | ||
0a22ddfb | 612 | XFREE(MTYPE_THREAD_MASTER, m->name); |
d62a17ae | 613 | XFREE(MTYPE_THREAD_MASTER, m->handler.pfds); |
614 | XFREE(MTYPE_THREAD_MASTER, m->handler.copy); | |
615 | XFREE(MTYPE_THREAD_MASTER, m); | |
718e3744 | 616 | } |
617 | ||
78ca0342 CF |
618 | /* Return remain time in miliseconds. */ |
619 | unsigned long thread_timer_remain_msec(struct thread *thread) | |
718e3744 | 620 | { |
d62a17ae | 621 | int64_t remain; |
1189d95f | 622 | |
00dffa8c | 623 | frr_with_mutex(&thread->mtx) { |
78ca0342 | 624 | remain = monotime_until(&thread->u.sands, NULL) / 1000LL; |
d62a17ae | 625 | } |
1189d95f | 626 | |
d62a17ae | 627 | return remain < 0 ? 0 : remain; |
718e3744 | 628 | } |
629 | ||
78ca0342 CF |
630 | /* Return remain time in seconds. */ |
631 | unsigned long thread_timer_remain_second(struct thread *thread) | |
632 | { | |
633 | return thread_timer_remain_msec(thread) / 1000LL; | |
634 | } | |
635 | ||
9c7753e4 DL |
636 | #define debugargdef const char *funcname, const char *schedfrom, int fromln |
637 | #define debugargpass funcname, schedfrom, fromln | |
e04ab74d | 638 | |
d62a17ae | 639 | struct timeval thread_timer_remain(struct thread *thread) |
6ac44687 | 640 | { |
d62a17ae | 641 | struct timeval remain; |
00dffa8c | 642 | frr_with_mutex(&thread->mtx) { |
d62a17ae | 643 | monotime_until(&thread->u.sands, &remain); |
644 | } | |
d62a17ae | 645 | return remain; |
6ac44687 CF |
646 | } |
647 | ||
0447957e AK |
648 | static int time_hhmmss(char *buf, int buf_size, long sec) |
649 | { | |
650 | long hh; | |
651 | long mm; | |
652 | int wr; | |
653 | ||
654 | zassert(buf_size >= 8); | |
655 | ||
656 | hh = sec / 3600; | |
657 | sec %= 3600; | |
658 | mm = sec / 60; | |
659 | sec %= 60; | |
660 | ||
661 | wr = snprintf(buf, buf_size, "%02ld:%02ld:%02ld", hh, mm, sec); | |
662 | ||
663 | return wr != 8; | |
664 | } | |
665 | ||
666 | char *thread_timer_to_hhmmss(char *buf, int buf_size, | |
667 | struct thread *t_timer) | |
668 | { | |
669 | if (t_timer) { | |
670 | time_hhmmss(buf, buf_size, | |
671 | thread_timer_remain_second(t_timer)); | |
672 | } else { | |
673 | snprintf(buf, buf_size, "--:--:--"); | |
674 | } | |
675 | return buf; | |
676 | } | |
677 | ||
718e3744 | 678 | /* Get new thread. */ |
d7c0a89a | 679 | static struct thread *thread_get(struct thread_master *m, uint8_t type, |
d62a17ae | 680 | int (*func)(struct thread *), void *arg, |
60a3efec | 681 | const struct xref_threadsched *xref) |
718e3744 | 682 | { |
c284542b | 683 | struct thread *thread = thread_list_pop(&m->unuse); |
d62a17ae | 684 | struct cpu_thread_history tmp; |
685 | ||
686 | if (!thread) { | |
687 | thread = XCALLOC(MTYPE_THREAD, sizeof(struct thread)); | |
688 | /* mutex only needs to be initialized at struct creation. */ | |
689 | pthread_mutex_init(&thread->mtx, NULL); | |
690 | m->alloc++; | |
691 | } | |
692 | ||
693 | thread->type = type; | |
694 | thread->add_type = type; | |
695 | thread->master = m; | |
696 | thread->arg = arg; | |
d62a17ae | 697 | thread->yield = THREAD_YIELD_TIME_SLOT; /* default */ |
698 | thread->ref = NULL; | |
699 | ||
700 | /* | |
701 | * So if the passed in funcname is not what we have | |
702 | * stored that means the thread->hist needs to be | |
703 | * updated. We keep the last one around in unused | |
704 | * under the assumption that we are probably | |
705 | * going to immediately allocate the same | |
706 | * type of thread. | |
707 | * This hopefully saves us some serious | |
708 | * hash_get lookups. | |
709 | */ | |
60a3efec DL |
710 | if ((thread->xref && thread->xref->funcname != xref->funcname) |
711 | || thread->func != func) { | |
d62a17ae | 712 | tmp.func = func; |
60a3efec | 713 | tmp.funcname = xref->funcname; |
d62a17ae | 714 | thread->hist = |
715 | hash_get(m->cpu_record, &tmp, | |
716 | (void *(*)(void *))cpu_record_hash_alloc); | |
717 | } | |
718 | thread->hist->total_active++; | |
719 | thread->func = func; | |
60a3efec | 720 | thread->xref = xref; |
d62a17ae | 721 | |
722 | return thread; | |
718e3744 | 723 | } |
724 | ||
6655966d RZ |
725 | static void thread_free(struct thread_master *master, struct thread *thread) |
726 | { | |
727 | /* Update statistics. */ | |
728 | assert(master->alloc > 0); | |
729 | master->alloc--; | |
730 | ||
731 | /* Free allocated resources. */ | |
732 | pthread_mutex_destroy(&thread->mtx); | |
733 | XFREE(MTYPE_THREAD, thread); | |
734 | } | |
735 | ||
d81ca9a3 MS |
736 | static int fd_poll(struct thread_master *m, const struct timeval *timer_wait, |
737 | bool *eintr_p) | |
209a72a6 | 738 | { |
d81ca9a3 MS |
739 | sigset_t origsigs; |
740 | unsigned char trash[64]; | |
741 | nfds_t count = m->handler.copycount; | |
742 | ||
d279ef57 DS |
743 | /* |
744 | * If timer_wait is null here, that means poll() should block | |
745 | * indefinitely, unless the thread_master has overridden it by setting | |
d62a17ae | 746 | * ->selectpoll_timeout. |
d279ef57 | 747 | * |
d62a17ae | 748 | * If the value is positive, it specifies the maximum number of |
d279ef57 DS |
749 | * milliseconds to wait. If the timeout is -1, it specifies that |
750 | * we should never wait and always return immediately even if no | |
751 | * event is detected. If the value is zero, the behavior is default. | |
752 | */ | |
d62a17ae | 753 | int timeout = -1; |
754 | ||
755 | /* number of file descriptors with events */ | |
756 | int num; | |
757 | ||
758 | if (timer_wait != NULL | |
759 | && m->selectpoll_timeout == 0) // use the default value | |
760 | timeout = (timer_wait->tv_sec * 1000) | |
761 | + (timer_wait->tv_usec / 1000); | |
762 | else if (m->selectpoll_timeout > 0) // use the user's timeout | |
763 | timeout = m->selectpoll_timeout; | |
764 | else if (m->selectpoll_timeout | |
765 | < 0) // effect a poll (return immediately) | |
766 | timeout = 0; | |
767 | ||
0bdeb5e5 | 768 | zlog_tls_buffer_flush(); |
3e41733f DL |
769 | rcu_read_unlock(); |
770 | rcu_assert_read_unlocked(); | |
771 | ||
d62a17ae | 772 | /* add poll pipe poker */ |
d81ca9a3 MS |
773 | assert(count + 1 < m->handler.pfdsize); |
774 | m->handler.copy[count].fd = m->io_pipe[0]; | |
775 | m->handler.copy[count].events = POLLIN; | |
776 | m->handler.copy[count].revents = 0x00; | |
777 | ||
778 | /* We need to deal with a signal-handling race here: we | |
779 | * don't want to miss a crucial signal, such as SIGTERM or SIGINT, | |
780 | * that may arrive just before we enter poll(). We will block the | |
781 | * key signals, then check whether any have arrived - if so, we return | |
782 | * before calling poll(). If not, we'll re-enable the signals | |
783 | * in the ppoll() call. | |
784 | */ | |
785 | ||
786 | sigemptyset(&origsigs); | |
787 | if (m->handle_signals) { | |
788 | /* Main pthread that handles the app signals */ | |
789 | if (frr_sigevent_check(&origsigs)) { | |
790 | /* Signal to process - restore signal mask and return */ | |
791 | pthread_sigmask(SIG_SETMASK, &origsigs, NULL); | |
792 | num = -1; | |
793 | *eintr_p = true; | |
794 | goto done; | |
795 | } | |
796 | } else { | |
797 | /* Don't make any changes for the non-main pthreads */ | |
798 | pthread_sigmask(SIG_SETMASK, NULL, &origsigs); | |
799 | } | |
d62a17ae | 800 | |
d81ca9a3 MS |
801 | #if defined(HAVE_PPOLL) |
802 | struct timespec ts, *tsp; | |
803 | ||
804 | if (timeout >= 0) { | |
805 | ts.tv_sec = timeout / 1000; | |
806 | ts.tv_nsec = (timeout % 1000) * 1000000; | |
807 | tsp = &ts; | |
808 | } else | |
809 | tsp = NULL; | |
810 | ||
811 | num = ppoll(m->handler.copy, count + 1, tsp, &origsigs); | |
812 | pthread_sigmask(SIG_SETMASK, &origsigs, NULL); | |
813 | #else | |
814 | /* Not ideal - there is a race after we restore the signal mask */ | |
815 | pthread_sigmask(SIG_SETMASK, &origsigs, NULL); | |
816 | num = poll(m->handler.copy, count + 1, timeout); | |
817 | #endif | |
d62a17ae | 818 | |
d81ca9a3 MS |
819 | done: |
820 | ||
821 | if (num < 0 && errno == EINTR) | |
822 | *eintr_p = true; | |
823 | ||
824 | if (num > 0 && m->handler.copy[count].revents != 0 && num--) | |
d62a17ae | 825 | while (read(m->io_pipe[0], &trash, sizeof(trash)) > 0) |
826 | ; | |
827 | ||
3e41733f DL |
828 | rcu_read_lock(); |
829 | ||
d62a17ae | 830 | return num; |
209a72a6 DS |
831 | } |
832 | ||
718e3744 | 833 | /* Add new read thread. */ |
60a3efec DL |
834 | struct thread *_thread_add_read_write(const struct xref_threadsched *xref, |
835 | struct thread_master *m, | |
836 | int (*func)(struct thread *), | |
837 | void *arg, int fd, struct thread **t_ptr) | |
718e3744 | 838 | { |
60a3efec | 839 | int dir = xref->thread_type; |
d62a17ae | 840 | struct thread *thread = NULL; |
1ef14bee | 841 | struct thread **thread_array; |
d62a17ae | 842 | |
abf96a87 | 843 | if (dir == THREAD_READ) |
c7bb4f00 QY |
844 | frrtrace(9, frr_libfrr, schedule_read, m, funcname, schedfrom, |
845 | fromln, t_ptr, fd, 0, arg, 0); | |
abf96a87 | 846 | else |
c7bb4f00 QY |
847 | frrtrace(9, frr_libfrr, schedule_write, m, funcname, schedfrom, |
848 | fromln, t_ptr, fd, 0, arg, 0); | |
abf96a87 | 849 | |
9b864cd3 | 850 | assert(fd >= 0 && fd < m->fd_limit); |
00dffa8c DL |
851 | frr_with_mutex(&m->mtx) { |
852 | if (t_ptr && *t_ptr) | |
853 | // thread is already scheduled; don't reschedule | |
854 | break; | |
d62a17ae | 855 | |
856 | /* default to a new pollfd */ | |
857 | nfds_t queuepos = m->handler.pfdcount; | |
858 | ||
1ef14bee DS |
859 | if (dir == THREAD_READ) |
860 | thread_array = m->read; | |
861 | else | |
862 | thread_array = m->write; | |
863 | ||
d62a17ae | 864 | /* if we already have a pollfd for our file descriptor, find and |
865 | * use it */ | |
866 | for (nfds_t i = 0; i < m->handler.pfdcount; i++) | |
867 | if (m->handler.pfds[i].fd == fd) { | |
868 | queuepos = i; | |
1ef14bee DS |
869 | |
870 | #ifdef DEV_BUILD | |
871 | /* | |
872 | * What happens if we have a thread already | |
873 | * created for this event? | |
874 | */ | |
875 | if (thread_array[fd]) | |
876 | assert(!"Thread already scheduled for file descriptor"); | |
877 | #endif | |
d62a17ae | 878 | break; |
879 | } | |
880 | ||
881 | /* make sure we have room for this fd + pipe poker fd */ | |
882 | assert(queuepos + 1 < m->handler.pfdsize); | |
883 | ||
60a3efec | 884 | thread = thread_get(m, dir, func, arg, xref); |
d62a17ae | 885 | |
886 | m->handler.pfds[queuepos].fd = fd; | |
887 | m->handler.pfds[queuepos].events |= | |
888 | (dir == THREAD_READ ? POLLIN : POLLOUT); | |
889 | ||
890 | if (queuepos == m->handler.pfdcount) | |
891 | m->handler.pfdcount++; | |
892 | ||
893 | if (thread) { | |
00dffa8c | 894 | frr_with_mutex(&thread->mtx) { |
d62a17ae | 895 | thread->u.fd = fd; |
1ef14bee | 896 | thread_array[thread->u.fd] = thread; |
d62a17ae | 897 | } |
d62a17ae | 898 | |
899 | if (t_ptr) { | |
900 | *t_ptr = thread; | |
901 | thread->ref = t_ptr; | |
902 | } | |
903 | } | |
904 | ||
905 | AWAKEN(m); | |
906 | } | |
d62a17ae | 907 | |
908 | return thread; | |
718e3744 | 909 | } |
910 | ||
56a94b36 | 911 | static struct thread * |
60a3efec DL |
912 | _thread_add_timer_timeval(const struct xref_threadsched *xref, |
913 | struct thread_master *m, int (*func)(struct thread *), | |
914 | int type, void *arg, struct timeval *time_relative, | |
915 | struct thread **t_ptr) | |
718e3744 | 916 | { |
d62a17ae | 917 | struct thread *thread; |
d62a17ae | 918 | |
919 | assert(m != NULL); | |
920 | ||
921 | assert(type == THREAD_TIMER); | |
922 | assert(time_relative); | |
923 | ||
c7bb4f00 QY |
924 | frrtrace(9, frr_libfrr, schedule_timer, m, funcname, schedfrom, fromln, |
925 | t_ptr, 0, 0, arg, (long)time_relative->tv_sec); | |
abf96a87 | 926 | |
00dffa8c DL |
927 | frr_with_mutex(&m->mtx) { |
928 | if (t_ptr && *t_ptr) | |
d279ef57 | 929 | /* thread is already scheduled; don't reschedule */ |
d62a17ae | 930 | return NULL; |
d62a17ae | 931 | |
60a3efec | 932 | thread = thread_get(m, type, func, arg, xref); |
d62a17ae | 933 | |
00dffa8c | 934 | frr_with_mutex(&thread->mtx) { |
d62a17ae | 935 | monotime(&thread->u.sands); |
936 | timeradd(&thread->u.sands, time_relative, | |
937 | &thread->u.sands); | |
27d29ced | 938 | thread_timer_list_add(&m->timer, thread); |
d62a17ae | 939 | if (t_ptr) { |
940 | *t_ptr = thread; | |
941 | thread->ref = t_ptr; | |
942 | } | |
943 | } | |
d62a17ae | 944 | |
945 | AWAKEN(m); | |
946 | } | |
d62a17ae | 947 | |
948 | return thread; | |
9e867fe6 | 949 | } |
950 | ||
98c91ac6 | 951 | |
952 | /* Add timer event thread. */ | |
60a3efec DL |
953 | struct thread *_thread_add_timer(const struct xref_threadsched *xref, |
954 | struct thread_master *m, | |
955 | int (*func)(struct thread *), | |
956 | void *arg, long timer, struct thread **t_ptr) | |
9e867fe6 | 957 | { |
d62a17ae | 958 | struct timeval trel; |
9e867fe6 | 959 | |
d62a17ae | 960 | assert(m != NULL); |
9e867fe6 | 961 | |
d62a17ae | 962 | trel.tv_sec = timer; |
963 | trel.tv_usec = 0; | |
9e867fe6 | 964 | |
60a3efec DL |
965 | return _thread_add_timer_timeval(xref, m, func, THREAD_TIMER, arg, |
966 | &trel, t_ptr); | |
98c91ac6 | 967 | } |
9e867fe6 | 968 | |
98c91ac6 | 969 | /* Add timer event thread with "millisecond" resolution */ |
60a3efec DL |
970 | struct thread *_thread_add_timer_msec(const struct xref_threadsched *xref, |
971 | struct thread_master *m, | |
972 | int (*func)(struct thread *), | |
973 | void *arg, long timer, | |
974 | struct thread **t_ptr) | |
98c91ac6 | 975 | { |
d62a17ae | 976 | struct timeval trel; |
9e867fe6 | 977 | |
d62a17ae | 978 | assert(m != NULL); |
718e3744 | 979 | |
d62a17ae | 980 | trel.tv_sec = timer / 1000; |
981 | trel.tv_usec = 1000 * (timer % 1000); | |
98c91ac6 | 982 | |
60a3efec DL |
983 | return _thread_add_timer_timeval(xref, m, func, THREAD_TIMER, arg, |
984 | &trel, t_ptr); | |
a48b4e6d | 985 | } |
986 | ||
d03c4cbd | 987 | /* Add timer event thread with "millisecond" resolution */ |
60a3efec DL |
988 | struct thread *_thread_add_timer_tv(const struct xref_threadsched *xref, |
989 | struct thread_master *m, | |
990 | int (*func)(struct thread *), | |
991 | void *arg, struct timeval *tv, | |
992 | struct thread **t_ptr) | |
d03c4cbd | 993 | { |
60a3efec DL |
994 | return _thread_add_timer_timeval(xref, m, func, THREAD_TIMER, arg, tv, |
995 | t_ptr); | |
d03c4cbd DL |
996 | } |
997 | ||
718e3744 | 998 | /* Add simple event thread. */ |
60a3efec DL |
999 | struct thread *_thread_add_event(const struct xref_threadsched *xref, |
1000 | struct thread_master *m, | |
1001 | int (*func)(struct thread *), | |
1002 | void *arg, int val, struct thread **t_ptr) | |
718e3744 | 1003 | { |
00dffa8c | 1004 | struct thread *thread = NULL; |
d62a17ae | 1005 | |
c7bb4f00 QY |
1006 | frrtrace(9, frr_libfrr, schedule_event, m, funcname, schedfrom, fromln, |
1007 | t_ptr, 0, val, arg, 0); | |
abf96a87 | 1008 | |
d62a17ae | 1009 | assert(m != NULL); |
1010 | ||
00dffa8c DL |
1011 | frr_with_mutex(&m->mtx) { |
1012 | if (t_ptr && *t_ptr) | |
d279ef57 | 1013 | /* thread is already scheduled; don't reschedule */ |
00dffa8c | 1014 | break; |
d62a17ae | 1015 | |
60a3efec | 1016 | thread = thread_get(m, THREAD_EVENT, func, arg, xref); |
00dffa8c | 1017 | frr_with_mutex(&thread->mtx) { |
d62a17ae | 1018 | thread->u.val = val; |
c284542b | 1019 | thread_list_add_tail(&m->event, thread); |
d62a17ae | 1020 | } |
d62a17ae | 1021 | |
1022 | if (t_ptr) { | |
1023 | *t_ptr = thread; | |
1024 | thread->ref = t_ptr; | |
1025 | } | |
1026 | ||
1027 | AWAKEN(m); | |
1028 | } | |
d62a17ae | 1029 | |
1030 | return thread; | |
718e3744 | 1031 | } |
1032 | ||
63ccb9cb QY |
1033 | /* Thread cancellation ------------------------------------------------------ */ |
1034 | ||
8797240e QY |
1035 | /** |
1036 | * NOT's out the .events field of pollfd corresponding to the given file | |
1037 | * descriptor. The event to be NOT'd is passed in the 'state' parameter. | |
1038 | * | |
1039 | * This needs to happen for both copies of pollfd's. See 'thread_fetch' | |
1040 | * implementation for details. | |
1041 | * | |
1042 | * @param master | |
1043 | * @param fd | |
1044 | * @param state the event to cancel. One or more (OR'd together) of the | |
1045 | * following: | |
1046 | * - POLLIN | |
1047 | * - POLLOUT | |
1048 | */ | |
d62a17ae | 1049 | static void thread_cancel_rw(struct thread_master *master, int fd, short state) |
0a95a0d0 | 1050 | { |
42d74538 QY |
1051 | bool found = false; |
1052 | ||
d62a17ae | 1053 | /* Cancel POLLHUP too just in case some bozo set it */ |
1054 | state |= POLLHUP; | |
1055 | ||
1056 | /* find the index of corresponding pollfd */ | |
1057 | nfds_t i; | |
1058 | ||
1059 | for (i = 0; i < master->handler.pfdcount; i++) | |
42d74538 QY |
1060 | if (master->handler.pfds[i].fd == fd) { |
1061 | found = true; | |
d62a17ae | 1062 | break; |
42d74538 QY |
1063 | } |
1064 | ||
1065 | if (!found) { | |
1066 | zlog_debug( | |
1067 | "[!] Received cancellation request for nonexistent rw job"); | |
1068 | zlog_debug("[!] threadmaster: %s | fd: %d", | |
996c9314 | 1069 | master->name ? master->name : "", fd); |
42d74538 QY |
1070 | return; |
1071 | } | |
d62a17ae | 1072 | |
1073 | /* NOT out event. */ | |
1074 | master->handler.pfds[i].events &= ~(state); | |
1075 | ||
1076 | /* If all events are canceled, delete / resize the pollfd array. */ | |
1077 | if (master->handler.pfds[i].events == 0) { | |
1078 | memmove(master->handler.pfds + i, master->handler.pfds + i + 1, | |
1079 | (master->handler.pfdcount - i - 1) | |
1080 | * sizeof(struct pollfd)); | |
1081 | master->handler.pfdcount--; | |
e985cda0 S |
1082 | master->handler.pfds[master->handler.pfdcount].fd = 0; |
1083 | master->handler.pfds[master->handler.pfdcount].events = 0; | |
d62a17ae | 1084 | } |
1085 | ||
1086 | /* If we have the same pollfd in the copy, perform the same operations, | |
1087 | * otherwise return. */ | |
1088 | if (i >= master->handler.copycount) | |
1089 | return; | |
1090 | ||
1091 | master->handler.copy[i].events &= ~(state); | |
1092 | ||
1093 | if (master->handler.copy[i].events == 0) { | |
1094 | memmove(master->handler.copy + i, master->handler.copy + i + 1, | |
1095 | (master->handler.copycount - i - 1) | |
1096 | * sizeof(struct pollfd)); | |
1097 | master->handler.copycount--; | |
e985cda0 S |
1098 | master->handler.copy[master->handler.copycount].fd = 0; |
1099 | master->handler.copy[master->handler.copycount].events = 0; | |
d62a17ae | 1100 | } |
0a95a0d0 DS |
1101 | } |
1102 | ||
1189d95f | 1103 | /** |
63ccb9cb | 1104 | * Process cancellation requests. |
1189d95f | 1105 | * |
63ccb9cb QY |
1106 | * This may only be run from the pthread which owns the thread_master. |
1107 | * | |
1108 | * @param master the thread master to process | |
1109 | * @REQUIRE master->mtx | |
1189d95f | 1110 | */ |
d62a17ae | 1111 | static void do_thread_cancel(struct thread_master *master) |
718e3744 | 1112 | { |
c284542b | 1113 | struct thread_list_head *list = NULL; |
d62a17ae | 1114 | struct thread **thread_array = NULL; |
1115 | struct thread *thread; | |
1116 | ||
1117 | struct cancel_req *cr; | |
1118 | struct listnode *ln; | |
1119 | for (ALL_LIST_ELEMENTS_RO(master->cancel_req, ln, cr)) { | |
d279ef57 DS |
1120 | /* |
1121 | * If this is an event object cancellation, linear search | |
1122 | * through event list deleting any events which have the | |
1123 | * specified argument. We also need to check every thread | |
1124 | * in the ready queue. | |
1125 | */ | |
d62a17ae | 1126 | if (cr->eventobj) { |
1127 | struct thread *t; | |
c284542b | 1128 | |
81fddbe7 | 1129 | frr_each_safe(thread_list, &master->event, t) { |
c284542b DL |
1130 | if (t->arg != cr->eventobj) |
1131 | continue; | |
1132 | thread_list_del(&master->event, t); | |
1133 | if (t->ref) | |
1134 | *t->ref = NULL; | |
1135 | thread_add_unuse(master, t); | |
d62a17ae | 1136 | } |
1137 | ||
81fddbe7 | 1138 | frr_each_safe(thread_list, &master->ready, t) { |
c284542b DL |
1139 | if (t->arg != cr->eventobj) |
1140 | continue; | |
1141 | thread_list_del(&master->ready, t); | |
1142 | if (t->ref) | |
1143 | *t->ref = NULL; | |
1144 | thread_add_unuse(master, t); | |
d62a17ae | 1145 | } |
1146 | continue; | |
1147 | } | |
1148 | ||
d279ef57 DS |
1149 | /* |
1150 | * The pointer varies depending on whether the cancellation | |
1151 | * request was made asynchronously or not. If it was, we | |
1152 | * need to check whether the thread even exists anymore | |
1153 | * before cancelling it. | |
1154 | */ | |
d62a17ae | 1155 | thread = (cr->thread) ? cr->thread : *cr->threadref; |
1156 | ||
1157 | if (!thread) | |
1158 | continue; | |
1159 | ||
1160 | /* Determine the appropriate queue to cancel the thread from */ | |
1161 | switch (thread->type) { | |
1162 | case THREAD_READ: | |
1163 | thread_cancel_rw(master, thread->u.fd, POLLIN); | |
1164 | thread_array = master->read; | |
1165 | break; | |
1166 | case THREAD_WRITE: | |
1167 | thread_cancel_rw(master, thread->u.fd, POLLOUT); | |
1168 | thread_array = master->write; | |
1169 | break; | |
1170 | case THREAD_TIMER: | |
27d29ced | 1171 | thread_timer_list_del(&master->timer, thread); |
d62a17ae | 1172 | break; |
1173 | case THREAD_EVENT: | |
1174 | list = &master->event; | |
1175 | break; | |
1176 | case THREAD_READY: | |
1177 | list = &master->ready; | |
1178 | break; | |
1179 | default: | |
1180 | continue; | |
1181 | break; | |
1182 | } | |
1183 | ||
27d29ced | 1184 | if (list) { |
c284542b | 1185 | thread_list_del(list, thread); |
d62a17ae | 1186 | } else if (thread_array) { |
1187 | thread_array[thread->u.fd] = NULL; | |
d62a17ae | 1188 | } |
1189 | ||
1190 | if (thread->ref) | |
1191 | *thread->ref = NULL; | |
1192 | ||
1193 | thread_add_unuse(thread->master, thread); | |
1194 | } | |
1195 | ||
1196 | /* Delete and free all cancellation requests */ | |
41b21bfa MS |
1197 | if (master->cancel_req) |
1198 | list_delete_all_node(master->cancel_req); | |
d62a17ae | 1199 | |
1200 | /* Wake up any threads which may be blocked in thread_cancel_async() */ | |
1201 | master->canceled = true; | |
1202 | pthread_cond_broadcast(&master->cancel_cond); | |
718e3744 | 1203 | } |
1204 | ||
63ccb9cb QY |
1205 | /** |
1206 | * Cancel any events which have the specified argument. | |
1207 | * | |
1208 | * MT-Unsafe | |
1209 | * | |
1210 | * @param m the thread_master to cancel from | |
1211 | * @param arg the argument passed when creating the event | |
1212 | */ | |
d62a17ae | 1213 | void thread_cancel_event(struct thread_master *master, void *arg) |
718e3744 | 1214 | { |
d62a17ae | 1215 | assert(master->owner == pthread_self()); |
1216 | ||
00dffa8c | 1217 | frr_with_mutex(&master->mtx) { |
d62a17ae | 1218 | struct cancel_req *cr = |
1219 | XCALLOC(MTYPE_TMP, sizeof(struct cancel_req)); | |
1220 | cr->eventobj = arg; | |
1221 | listnode_add(master->cancel_req, cr); | |
1222 | do_thread_cancel(master); | |
1223 | } | |
63ccb9cb | 1224 | } |
1189d95f | 1225 | |
63ccb9cb QY |
1226 | /** |
1227 | * Cancel a specific task. | |
1228 | * | |
1229 | * MT-Unsafe | |
1230 | * | |
1231 | * @param thread task to cancel | |
1232 | */ | |
b3d6bc6e | 1233 | void thread_cancel(struct thread **thread) |
63ccb9cb | 1234 | { |
b3d6bc6e MS |
1235 | struct thread_master *master; |
1236 | ||
1237 | if (thread == NULL || *thread == NULL) | |
1238 | return; | |
1239 | ||
1240 | master = (*thread)->master; | |
d62a17ae | 1241 | |
b4d6e855 QY |
1242 | frrtrace(9, frr_libfrr, thread_cancel, master, (*thread)->funcname, |
1243 | (*thread)->schedfrom, (*thread)->schedfrom_line, NULL, (*thread)->u.fd, | |
1244 | (*thread)->u.val, (*thread)->arg, (*thread)->u.sands.tv_sec); | |
abf96a87 | 1245 | |
6ed04aa2 DS |
1246 | assert(master->owner == pthread_self()); |
1247 | ||
00dffa8c | 1248 | frr_with_mutex(&master->mtx) { |
d62a17ae | 1249 | struct cancel_req *cr = |
1250 | XCALLOC(MTYPE_TMP, sizeof(struct cancel_req)); | |
b3d6bc6e | 1251 | cr->thread = *thread; |
6ed04aa2 DS |
1252 | listnode_add(master->cancel_req, cr); |
1253 | do_thread_cancel(master); | |
d62a17ae | 1254 | } |
b3d6bc6e MS |
1255 | |
1256 | *thread = NULL; | |
63ccb9cb | 1257 | } |
1189d95f | 1258 | |
63ccb9cb QY |
1259 | /** |
1260 | * Asynchronous cancellation. | |
1261 | * | |
8797240e QY |
1262 | * Called with either a struct thread ** or void * to an event argument, |
1263 | * this function posts the correct cancellation request and blocks until it is | |
1264 | * serviced. | |
63ccb9cb QY |
1265 | * |
1266 | * If the thread is currently running, execution blocks until it completes. | |
1267 | * | |
8797240e QY |
1268 | * The last two parameters are mutually exclusive, i.e. if you pass one the |
1269 | * other must be NULL. | |
1270 | * | |
1271 | * When the cancellation procedure executes on the target thread_master, the | |
1272 | * thread * provided is checked for nullity. If it is null, the thread is | |
1273 | * assumed to no longer exist and the cancellation request is a no-op. Thus | |
1274 | * users of this API must pass a back-reference when scheduling the original | |
1275 | * task. | |
1276 | * | |
63ccb9cb QY |
1277 | * MT-Safe |
1278 | * | |
8797240e QY |
1279 | * @param master the thread master with the relevant event / task |
1280 | * @param thread pointer to thread to cancel | |
1281 | * @param eventobj the event | |
63ccb9cb | 1282 | */ |
d62a17ae | 1283 | void thread_cancel_async(struct thread_master *master, struct thread **thread, |
1284 | void *eventobj) | |
63ccb9cb | 1285 | { |
d62a17ae | 1286 | assert(!(thread && eventobj) && (thread || eventobj)); |
abf96a87 QY |
1287 | |
1288 | if (thread && *thread) | |
c7bb4f00 QY |
1289 | frrtrace(9, frr_libfrr, thread_cancel_async, master, |
1290 | (*thread)->funcname, (*thread)->schedfrom, | |
1291 | (*thread)->schedfrom_line, NULL, (*thread)->u.fd, | |
1292 | (*thread)->u.val, (*thread)->arg, | |
1293 | (*thread)->u.sands.tv_sec); | |
abf96a87 | 1294 | else |
c7bb4f00 QY |
1295 | frrtrace(9, frr_libfrr, thread_cancel_async, master, NULL, NULL, |
1296 | 0, NULL, 0, 0, eventobj, 0); | |
abf96a87 | 1297 | |
d62a17ae | 1298 | assert(master->owner != pthread_self()); |
1299 | ||
00dffa8c | 1300 | frr_with_mutex(&master->mtx) { |
d62a17ae | 1301 | master->canceled = false; |
1302 | ||
1303 | if (thread) { | |
1304 | struct cancel_req *cr = | |
1305 | XCALLOC(MTYPE_TMP, sizeof(struct cancel_req)); | |
1306 | cr->threadref = thread; | |
1307 | listnode_add(master->cancel_req, cr); | |
1308 | } else if (eventobj) { | |
1309 | struct cancel_req *cr = | |
1310 | XCALLOC(MTYPE_TMP, sizeof(struct cancel_req)); | |
1311 | cr->eventobj = eventobj; | |
1312 | listnode_add(master->cancel_req, cr); | |
1313 | } | |
1314 | AWAKEN(master); | |
1315 | ||
1316 | while (!master->canceled) | |
1317 | pthread_cond_wait(&master->cancel_cond, &master->mtx); | |
1318 | } | |
50478845 MS |
1319 | |
1320 | if (thread) | |
1321 | *thread = NULL; | |
718e3744 | 1322 | } |
63ccb9cb | 1323 | /* ------------------------------------------------------------------------- */ |
718e3744 | 1324 | |
27d29ced | 1325 | static struct timeval *thread_timer_wait(struct thread_timer_list_head *timers, |
d62a17ae | 1326 | struct timeval *timer_val) |
718e3744 | 1327 | { |
27d29ced DL |
1328 | if (!thread_timer_list_count(timers)) |
1329 | return NULL; | |
1330 | ||
1331 | struct thread *next_timer = thread_timer_list_first(timers); | |
1332 | monotime_until(&next_timer->u.sands, timer_val); | |
1333 | return timer_val; | |
718e3744 | 1334 | } |
718e3744 | 1335 | |
d62a17ae | 1336 | static struct thread *thread_run(struct thread_master *m, struct thread *thread, |
1337 | struct thread *fetch) | |
718e3744 | 1338 | { |
d62a17ae | 1339 | *fetch = *thread; |
1340 | thread_add_unuse(m, thread); | |
1341 | return fetch; | |
718e3744 | 1342 | } |
1343 | ||
d62a17ae | 1344 | static int thread_process_io_helper(struct thread_master *m, |
45f3d590 DS |
1345 | struct thread *thread, short state, |
1346 | short actual_state, int pos) | |
5d4ccd4e | 1347 | { |
d62a17ae | 1348 | struct thread **thread_array; |
1349 | ||
45f3d590 DS |
1350 | /* |
1351 | * poll() clears the .events field, but the pollfd array we | |
1352 | * pass to poll() is a copy of the one used to schedule threads. | |
1353 | * We need to synchronize state between the two here by applying | |
1354 | * the same changes poll() made on the copy of the "real" pollfd | |
1355 | * array. | |
1356 | * | |
1357 | * This cleans up a possible infinite loop where we refuse | |
1358 | * to respond to a poll event but poll is insistent that | |
1359 | * we should. | |
1360 | */ | |
1361 | m->handler.pfds[pos].events &= ~(state); | |
1362 | ||
1363 | if (!thread) { | |
1364 | if ((actual_state & (POLLHUP|POLLIN)) != POLLHUP) | |
1365 | flog_err(EC_LIB_NO_THREAD, | |
1366 | "Attempting to process an I/O event but for fd: %d(%d) no thread to handle this!\n", | |
1367 | m->handler.pfds[pos].fd, actual_state); | |
d62a17ae | 1368 | return 0; |
45f3d590 | 1369 | } |
d62a17ae | 1370 | |
1371 | if (thread->type == THREAD_READ) | |
1372 | thread_array = m->read; | |
1373 | else | |
1374 | thread_array = m->write; | |
1375 | ||
1376 | thread_array[thread->u.fd] = NULL; | |
c284542b | 1377 | thread_list_add_tail(&m->ready, thread); |
d62a17ae | 1378 | thread->type = THREAD_READY; |
45f3d590 | 1379 | |
d62a17ae | 1380 | return 1; |
5d4ccd4e DS |
1381 | } |
1382 | ||
8797240e QY |
1383 | /** |
1384 | * Process I/O events. | |
1385 | * | |
1386 | * Walks through file descriptor array looking for those pollfds whose .revents | |
1387 | * field has something interesting. Deletes any invalid file descriptors. | |
1388 | * | |
1389 | * @param m the thread master | |
1390 | * @param num the number of active file descriptors (return value of poll()) | |
1391 | */ | |
d62a17ae | 1392 | static void thread_process_io(struct thread_master *m, unsigned int num) |
0a95a0d0 | 1393 | { |
d62a17ae | 1394 | unsigned int ready = 0; |
1395 | struct pollfd *pfds = m->handler.copy; | |
1396 | ||
1397 | for (nfds_t i = 0; i < m->handler.copycount && ready < num; ++i) { | |
1398 | /* no event for current fd? immediately continue */ | |
1399 | if (pfds[i].revents == 0) | |
1400 | continue; | |
1401 | ||
1402 | ready++; | |
1403 | ||
d279ef57 DS |
1404 | /* |
1405 | * Unless someone has called thread_cancel from another | |
1406 | * pthread, the only thing that could have changed in | |
1407 | * m->handler.pfds while we were asleep is the .events | |
1408 | * field in a given pollfd. Barring thread_cancel() that | |
1409 | * value should be a superset of the values we have in our | |
1410 | * copy, so there's no need to update it. Similarily, | |
1411 | * barring deletion, the fd should still be a valid index | |
1412 | * into the master's pfds. | |
d142453d DS |
1413 | * |
1414 | * We are including POLLERR here to do a READ event | |
1415 | * this is because the read should fail and the | |
1416 | * read function should handle it appropriately | |
d279ef57 | 1417 | */ |
d142453d | 1418 | if (pfds[i].revents & (POLLIN | POLLHUP | POLLERR)) { |
d62a17ae | 1419 | thread_process_io_helper(m, m->read[pfds[i].fd], POLLIN, |
45f3d590 DS |
1420 | pfds[i].revents, i); |
1421 | } | |
d62a17ae | 1422 | if (pfds[i].revents & POLLOUT) |
1423 | thread_process_io_helper(m, m->write[pfds[i].fd], | |
45f3d590 | 1424 | POLLOUT, pfds[i].revents, i); |
d62a17ae | 1425 | |
1426 | /* if one of our file descriptors is garbage, remove the same | |
1427 | * from | |
1428 | * both pfds + update sizes and index */ | |
1429 | if (pfds[i].revents & POLLNVAL) { | |
1430 | memmove(m->handler.pfds + i, m->handler.pfds + i + 1, | |
1431 | (m->handler.pfdcount - i - 1) | |
1432 | * sizeof(struct pollfd)); | |
1433 | m->handler.pfdcount--; | |
e985cda0 S |
1434 | m->handler.pfds[m->handler.pfdcount].fd = 0; |
1435 | m->handler.pfds[m->handler.pfdcount].events = 0; | |
d62a17ae | 1436 | |
1437 | memmove(pfds + i, pfds + i + 1, | |
1438 | (m->handler.copycount - i - 1) | |
1439 | * sizeof(struct pollfd)); | |
1440 | m->handler.copycount--; | |
e985cda0 S |
1441 | m->handler.copy[m->handler.copycount].fd = 0; |
1442 | m->handler.copy[m->handler.copycount].events = 0; | |
d62a17ae | 1443 | |
1444 | i--; | |
1445 | } | |
1446 | } | |
718e3744 | 1447 | } |
1448 | ||
8b70d0b0 | 1449 | /* Add all timers that have popped to the ready list. */ |
27d29ced | 1450 | static unsigned int thread_process_timers(struct thread_timer_list_head *timers, |
d62a17ae | 1451 | struct timeval *timenow) |
a48b4e6d | 1452 | { |
d62a17ae | 1453 | struct thread *thread; |
1454 | unsigned int ready = 0; | |
1455 | ||
27d29ced | 1456 | while ((thread = thread_timer_list_first(timers))) { |
d62a17ae | 1457 | if (timercmp(timenow, &thread->u.sands, <)) |
1458 | return ready; | |
27d29ced | 1459 | thread_timer_list_pop(timers); |
d62a17ae | 1460 | thread->type = THREAD_READY; |
c284542b | 1461 | thread_list_add_tail(&thread->master->ready, thread); |
d62a17ae | 1462 | ready++; |
1463 | } | |
1464 | return ready; | |
a48b4e6d | 1465 | } |
1466 | ||
2613abe6 | 1467 | /* process a list en masse, e.g. for event thread lists */ |
c284542b | 1468 | static unsigned int thread_process(struct thread_list_head *list) |
2613abe6 | 1469 | { |
d62a17ae | 1470 | struct thread *thread; |
d62a17ae | 1471 | unsigned int ready = 0; |
1472 | ||
c284542b | 1473 | while ((thread = thread_list_pop(list))) { |
d62a17ae | 1474 | thread->type = THREAD_READY; |
c284542b | 1475 | thread_list_add_tail(&thread->master->ready, thread); |
d62a17ae | 1476 | ready++; |
1477 | } | |
1478 | return ready; | |
2613abe6 PJ |
1479 | } |
1480 | ||
1481 | ||
718e3744 | 1482 | /* Fetch next ready thread. */ |
d62a17ae | 1483 | struct thread *thread_fetch(struct thread_master *m, struct thread *fetch) |
718e3744 | 1484 | { |
d62a17ae | 1485 | struct thread *thread = NULL; |
1486 | struct timeval now; | |
1487 | struct timeval zerotime = {0, 0}; | |
1488 | struct timeval tv; | |
1489 | struct timeval *tw = NULL; | |
d81ca9a3 | 1490 | bool eintr_p = false; |
d62a17ae | 1491 | int num = 0; |
1492 | ||
1493 | do { | |
1494 | /* Handle signals if any */ | |
1495 | if (m->handle_signals) | |
1496 | quagga_sigevent_process(); | |
1497 | ||
1498 | pthread_mutex_lock(&m->mtx); | |
1499 | ||
1500 | /* Process any pending cancellation requests */ | |
1501 | do_thread_cancel(m); | |
1502 | ||
e3c9529e QY |
1503 | /* |
1504 | * Attempt to flush ready queue before going into poll(). | |
1505 | * This is performance-critical. Think twice before modifying. | |
1506 | */ | |
c284542b | 1507 | if ((thread = thread_list_pop(&m->ready))) { |
e3c9529e QY |
1508 | fetch = thread_run(m, thread, fetch); |
1509 | if (fetch->ref) | |
1510 | *fetch->ref = NULL; | |
1511 | pthread_mutex_unlock(&m->mtx); | |
1512 | break; | |
1513 | } | |
1514 | ||
1515 | /* otherwise, tick through scheduling sequence */ | |
1516 | ||
bca37d17 QY |
1517 | /* |
1518 | * Post events to ready queue. This must come before the | |
1519 | * following block since events should occur immediately | |
1520 | */ | |
d62a17ae | 1521 | thread_process(&m->event); |
1522 | ||
bca37d17 QY |
1523 | /* |
1524 | * If there are no tasks on the ready queue, we will poll() | |
1525 | * until a timer expires or we receive I/O, whichever comes | |
1526 | * first. The strategy for doing this is: | |
d62a17ae | 1527 | * |
1528 | * - If there are events pending, set the poll() timeout to zero | |
1529 | * - If there are no events pending, but there are timers | |
d279ef57 DS |
1530 | * pending, set the timeout to the smallest remaining time on |
1531 | * any timer. | |
d62a17ae | 1532 | * - If there are neither timers nor events pending, but there |
d279ef57 | 1533 | * are file descriptors pending, block indefinitely in poll() |
d62a17ae | 1534 | * - If nothing is pending, it's time for the application to die |
1535 | * | |
1536 | * In every case except the last, we need to hit poll() at least | |
bca37d17 QY |
1537 | * once per loop to avoid starvation by events |
1538 | */ | |
c284542b | 1539 | if (!thread_list_count(&m->ready)) |
27d29ced | 1540 | tw = thread_timer_wait(&m->timer, &tv); |
d62a17ae | 1541 | |
c284542b DL |
1542 | if (thread_list_count(&m->ready) || |
1543 | (tw && !timercmp(tw, &zerotime, >))) | |
d62a17ae | 1544 | tw = &zerotime; |
1545 | ||
1546 | if (!tw && m->handler.pfdcount == 0) { /* die */ | |
1547 | pthread_mutex_unlock(&m->mtx); | |
1548 | fetch = NULL; | |
1549 | break; | |
1550 | } | |
1551 | ||
bca37d17 QY |
1552 | /* |
1553 | * Copy pollfd array + # active pollfds in it. Not necessary to | |
1554 | * copy the array size as this is fixed. | |
1555 | */ | |
d62a17ae | 1556 | m->handler.copycount = m->handler.pfdcount; |
1557 | memcpy(m->handler.copy, m->handler.pfds, | |
1558 | m->handler.copycount * sizeof(struct pollfd)); | |
1559 | ||
e3c9529e QY |
1560 | pthread_mutex_unlock(&m->mtx); |
1561 | { | |
d81ca9a3 MS |
1562 | eintr_p = false; |
1563 | num = fd_poll(m, tw, &eintr_p); | |
e3c9529e QY |
1564 | } |
1565 | pthread_mutex_lock(&m->mtx); | |
d764d2cc | 1566 | |
e3c9529e QY |
1567 | /* Handle any errors received in poll() */ |
1568 | if (num < 0) { | |
d81ca9a3 | 1569 | if (eintr_p) { |
d62a17ae | 1570 | pthread_mutex_unlock(&m->mtx); |
e3c9529e QY |
1571 | /* loop around to signal handler */ |
1572 | continue; | |
d62a17ae | 1573 | } |
1574 | ||
e3c9529e | 1575 | /* else die */ |
450971aa | 1576 | flog_err(EC_LIB_SYSTEM_CALL, "poll() error: %s", |
9ef9495e | 1577 | safe_strerror(errno)); |
e3c9529e QY |
1578 | pthread_mutex_unlock(&m->mtx); |
1579 | fetch = NULL; | |
1580 | break; | |
bca37d17 | 1581 | } |
d62a17ae | 1582 | |
1583 | /* Post timers to ready queue. */ | |
1584 | monotime(&now); | |
27d29ced | 1585 | thread_process_timers(&m->timer, &now); |
d62a17ae | 1586 | |
1587 | /* Post I/O to ready queue. */ | |
1588 | if (num > 0) | |
1589 | thread_process_io(m, num); | |
1590 | ||
d62a17ae | 1591 | pthread_mutex_unlock(&m->mtx); |
1592 | ||
1593 | } while (!thread && m->spin); | |
1594 | ||
1595 | return fetch; | |
718e3744 | 1596 | } |
1597 | ||
d62a17ae | 1598 | static unsigned long timeval_elapsed(struct timeval a, struct timeval b) |
62f44022 | 1599 | { |
d62a17ae | 1600 | return (((a.tv_sec - b.tv_sec) * TIMER_SECOND_MICRO) |
1601 | + (a.tv_usec - b.tv_usec)); | |
62f44022 QY |
1602 | } |
1603 | ||
d62a17ae | 1604 | unsigned long thread_consumed_time(RUSAGE_T *now, RUSAGE_T *start, |
1605 | unsigned long *cputime) | |
718e3744 | 1606 | { |
d62a17ae | 1607 | /* This is 'user + sys' time. */ |
1608 | *cputime = timeval_elapsed(now->cpu.ru_utime, start->cpu.ru_utime) | |
1609 | + timeval_elapsed(now->cpu.ru_stime, start->cpu.ru_stime); | |
1610 | return timeval_elapsed(now->real, start->real); | |
8b70d0b0 | 1611 | } |
1612 | ||
50596be0 DS |
1613 | /* We should aim to yield after yield milliseconds, which defaults |
1614 | to THREAD_YIELD_TIME_SLOT . | |
8b70d0b0 | 1615 | Note: we are using real (wall clock) time for this calculation. |
1616 | It could be argued that CPU time may make more sense in certain | |
1617 | contexts. The things to consider are whether the thread may have | |
1618 | blocked (in which case wall time increases, but CPU time does not), | |
1619 | or whether the system is heavily loaded with other processes competing | |
d62a17ae | 1620 | for CPU time. On balance, wall clock time seems to make sense. |
8b70d0b0 | 1621 | Plus it has the added benefit that gettimeofday should be faster |
1622 | than calling getrusage. */ | |
d62a17ae | 1623 | int thread_should_yield(struct thread *thread) |
718e3744 | 1624 | { |
d62a17ae | 1625 | int result; |
00dffa8c | 1626 | frr_with_mutex(&thread->mtx) { |
d62a17ae | 1627 | result = monotime_since(&thread->real, NULL) |
1628 | > (int64_t)thread->yield; | |
1629 | } | |
d62a17ae | 1630 | return result; |
50596be0 DS |
1631 | } |
1632 | ||
d62a17ae | 1633 | void thread_set_yield_time(struct thread *thread, unsigned long yield_time) |
50596be0 | 1634 | { |
00dffa8c | 1635 | frr_with_mutex(&thread->mtx) { |
d62a17ae | 1636 | thread->yield = yield_time; |
1637 | } | |
718e3744 | 1638 | } |
1639 | ||
d62a17ae | 1640 | void thread_getrusage(RUSAGE_T *r) |
db9c0df9 | 1641 | { |
231db9a6 DS |
1642 | #if defined RUSAGE_THREAD |
1643 | #define FRR_RUSAGE RUSAGE_THREAD | |
1644 | #else | |
1645 | #define FRR_RUSAGE RUSAGE_SELF | |
1646 | #endif | |
d62a17ae | 1647 | monotime(&r->real); |
f75e802d | 1648 | #ifndef EXCLUDE_CPU_TIME |
231db9a6 | 1649 | getrusage(FRR_RUSAGE, &(r->cpu)); |
f75e802d | 1650 | #endif |
db9c0df9 PJ |
1651 | } |
1652 | ||
fbcac826 QY |
1653 | /* |
1654 | * Call a thread. | |
1655 | * | |
1656 | * This function will atomically update the thread's usage history. At present | |
1657 | * this is the only spot where usage history is written. Nevertheless the code | |
1658 | * has been written such that the introduction of writers in the future should | |
1659 | * not need to update it provided the writers atomically perform only the | |
1660 | * operations done here, i.e. updating the total and maximum times. In | |
1661 | * particular, the maximum real and cpu times must be monotonically increasing | |
1662 | * or this code is not correct. | |
1663 | */ | |
d62a17ae | 1664 | void thread_call(struct thread *thread) |
718e3744 | 1665 | { |
f75e802d | 1666 | #ifndef EXCLUDE_CPU_TIME |
fbcac826 QY |
1667 | _Atomic unsigned long realtime, cputime; |
1668 | unsigned long exp; | |
1669 | unsigned long helper; | |
f75e802d | 1670 | #endif |
d62a17ae | 1671 | RUSAGE_T before, after; |
cc8b13a0 | 1672 | |
d62a17ae | 1673 | GETRUSAGE(&before); |
1674 | thread->real = before.real; | |
718e3744 | 1675 | |
c7bb4f00 QY |
1676 | frrtrace(9, frr_libfrr, thread_call, thread->master, thread->funcname, |
1677 | thread->schedfrom, thread->schedfrom_line, NULL, thread->u.fd, | |
1678 | thread->u.val, thread->arg, thread->u.sands.tv_sec); | |
abf96a87 | 1679 | |
d62a17ae | 1680 | pthread_setspecific(thread_current, thread); |
1681 | (*thread->func)(thread); | |
1682 | pthread_setspecific(thread_current, NULL); | |
718e3744 | 1683 | |
d62a17ae | 1684 | GETRUSAGE(&after); |
718e3744 | 1685 | |
f75e802d | 1686 | #ifndef EXCLUDE_CPU_TIME |
fbcac826 QY |
1687 | realtime = thread_consumed_time(&after, &before, &helper); |
1688 | cputime = helper; | |
1689 | ||
1690 | /* update realtime */ | |
1691 | atomic_fetch_add_explicit(&thread->hist->real.total, realtime, | |
1692 | memory_order_seq_cst); | |
1693 | exp = atomic_load_explicit(&thread->hist->real.max, | |
1694 | memory_order_seq_cst); | |
1695 | while (exp < realtime | |
1696 | && !atomic_compare_exchange_weak_explicit( | |
1697 | &thread->hist->real.max, &exp, realtime, | |
1698 | memory_order_seq_cst, memory_order_seq_cst)) | |
1699 | ; | |
1700 | ||
1701 | /* update cputime */ | |
1702 | atomic_fetch_add_explicit(&thread->hist->cpu.total, cputime, | |
1703 | memory_order_seq_cst); | |
1704 | exp = atomic_load_explicit(&thread->hist->cpu.max, | |
1705 | memory_order_seq_cst); | |
1706 | while (exp < cputime | |
1707 | && !atomic_compare_exchange_weak_explicit( | |
1708 | &thread->hist->cpu.max, &exp, cputime, | |
1709 | memory_order_seq_cst, memory_order_seq_cst)) | |
1710 | ; | |
1711 | ||
1712 | atomic_fetch_add_explicit(&thread->hist->total_calls, 1, | |
1713 | memory_order_seq_cst); | |
1714 | atomic_fetch_or_explicit(&thread->hist->types, 1 << thread->add_type, | |
1715 | memory_order_seq_cst); | |
718e3744 | 1716 | |
924b9229 | 1717 | #ifdef CONSUMED_TIME_CHECK |
d62a17ae | 1718 | if (realtime > CONSUMED_TIME_CHECK) { |
1719 | /* | |
1720 | * We have a CPU Hog on our hands. | |
1721 | * Whinge about it now, so we're aware this is yet another task | |
1722 | * to fix. | |
1723 | */ | |
9ef9495e | 1724 | flog_warn( |
450971aa | 1725 | EC_LIB_SLOW_THREAD, |
d62a17ae | 1726 | "SLOW THREAD: task %s (%lx) ran for %lums (cpu time %lums)", |
60a3efec | 1727 | thread->xref->funcname, (unsigned long)thread->func, |
d62a17ae | 1728 | realtime / 1000, cputime / 1000); |
1729 | } | |
924b9229 | 1730 | #endif /* CONSUMED_TIME_CHECK */ |
f75e802d | 1731 | #endif /* Exclude CPU Time */ |
718e3744 | 1732 | } |
1733 | ||
1734 | /* Execute thread */ | |
60a3efec DL |
1735 | void _thread_execute(const struct xref_threadsched *xref, |
1736 | struct thread_master *m, int (*func)(struct thread *), | |
1737 | void *arg, int val) | |
718e3744 | 1738 | { |
c4345fbf | 1739 | struct thread *thread; |
718e3744 | 1740 | |
c4345fbf | 1741 | /* Get or allocate new thread to execute. */ |
00dffa8c | 1742 | frr_with_mutex(&m->mtx) { |
60a3efec | 1743 | thread = thread_get(m, THREAD_EVENT, func, arg, xref); |
9c7753e4 | 1744 | |
c4345fbf | 1745 | /* Set its event value. */ |
00dffa8c | 1746 | frr_with_mutex(&thread->mtx) { |
c4345fbf RZ |
1747 | thread->add_type = THREAD_EXECUTE; |
1748 | thread->u.val = val; | |
1749 | thread->ref = &thread; | |
1750 | } | |
c4345fbf | 1751 | } |
f7c62e11 | 1752 | |
c4345fbf RZ |
1753 | /* Execute thread doing all accounting. */ |
1754 | thread_call(thread); | |
9c7753e4 | 1755 | |
c4345fbf RZ |
1756 | /* Give back or free thread. */ |
1757 | thread_add_unuse(m, thread); | |
718e3744 | 1758 | } |
1543c387 MS |
1759 | |
1760 | /* Debug signal mask - if 'sigs' is NULL, use current effective mask. */ | |
1761 | void debug_signals(const sigset_t *sigs) | |
1762 | { | |
1763 | int i, found; | |
1764 | sigset_t tmpsigs; | |
1765 | char buf[300]; | |
1766 | ||
1767 | /* | |
1768 | * We're only looking at the non-realtime signals here, so we need | |
1769 | * some limit value. Platform differences mean at some point we just | |
1770 | * need to pick a reasonable value. | |
1771 | */ | |
1772 | #if defined SIGRTMIN | |
1773 | # define LAST_SIGNAL SIGRTMIN | |
1774 | #else | |
1775 | # define LAST_SIGNAL 32 | |
1776 | #endif | |
1777 | ||
1778 | ||
1779 | if (sigs == NULL) { | |
1780 | sigemptyset(&tmpsigs); | |
1781 | pthread_sigmask(SIG_BLOCK, NULL, &tmpsigs); | |
1782 | sigs = &tmpsigs; | |
1783 | } | |
1784 | ||
1785 | found = 0; | |
1786 | buf[0] = '\0'; | |
1787 | ||
1788 | for (i = 0; i < LAST_SIGNAL; i++) { | |
1789 | char tmp[20]; | |
1790 | ||
1791 | if (sigismember(sigs, i) > 0) { | |
1792 | if (found > 0) | |
1793 | strlcat(buf, ",", sizeof(buf)); | |
1794 | snprintf(tmp, sizeof(tmp), "%d", i); | |
1795 | strlcat(buf, tmp, sizeof(buf)); | |
1796 | found++; | |
1797 | } | |
1798 | } | |
1799 | ||
1800 | if (found == 0) | |
1801 | snprintf(buf, sizeof(buf), "<none>"); | |
1802 | ||
1803 | zlog_debug("%s: %s", __func__, buf); | |
1804 | } |