]>
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 | ||
a5824290 | 61 | DECLARE_HEAP(thread_timer_list, struct thread, timeritem, thread_timer_cmp) |
27d29ced | 62 | |
3b96b781 HT |
63 | #if defined(__APPLE__) |
64 | #include <mach/mach.h> | |
65 | #include <mach/mach_time.h> | |
66 | #endif | |
67 | ||
d62a17ae | 68 | #define AWAKEN(m) \ |
69 | do { \ | |
2b64873d | 70 | const unsigned char wakebyte = 0x01; \ |
d62a17ae | 71 | write(m->io_pipe[1], &wakebyte, 1); \ |
72 | } while (0); | |
3bf2673b | 73 | |
62f44022 | 74 | /* control variable for initializer */ |
c17faa4b | 75 | static pthread_once_t init_once = PTHREAD_ONCE_INIT; |
e0bebc7c | 76 | pthread_key_t thread_current; |
6b0655a2 | 77 | |
c17faa4b | 78 | static pthread_mutex_t masters_mtx = PTHREAD_MUTEX_INITIALIZER; |
62f44022 QY |
79 | static struct list *masters; |
80 | ||
6655966d | 81 | static void thread_free(struct thread_master *master, struct thread *thread); |
6b0655a2 | 82 | |
62f44022 | 83 | /* CLI start ---------------------------------------------------------------- */ |
d8b87afe | 84 | static unsigned int cpu_record_hash_key(const struct cpu_thread_history *a) |
e04ab74d | 85 | { |
883cc51d | 86 | int size = sizeof(a->func); |
bd74dc61 DS |
87 | |
88 | return jhash(&a->func, size, 0); | |
e04ab74d | 89 | } |
90 | ||
74df8d6d | 91 | static bool cpu_record_hash_cmp(const struct cpu_thread_history *a, |
d62a17ae | 92 | const struct cpu_thread_history *b) |
e04ab74d | 93 | { |
d62a17ae | 94 | return a->func == b->func; |
e04ab74d | 95 | } |
96 | ||
d62a17ae | 97 | static void *cpu_record_hash_alloc(struct cpu_thread_history *a) |
e04ab74d | 98 | { |
d62a17ae | 99 | struct cpu_thread_history *new; |
100 | new = XCALLOC(MTYPE_THREAD_STATS, sizeof(struct cpu_thread_history)); | |
101 | new->func = a->func; | |
102 | new->funcname = a->funcname; | |
103 | return new; | |
e04ab74d | 104 | } |
105 | ||
d62a17ae | 106 | static void cpu_record_hash_free(void *a) |
228da428 | 107 | { |
d62a17ae | 108 | struct cpu_thread_history *hist = a; |
109 | ||
110 | XFREE(MTYPE_THREAD_STATS, hist); | |
228da428 CC |
111 | } |
112 | ||
f75e802d | 113 | #ifndef EXCLUDE_CPU_TIME |
d62a17ae | 114 | static void vty_out_cpu_thread_history(struct vty *vty, |
115 | struct cpu_thread_history *a) | |
e04ab74d | 116 | { |
9a8a7b0e | 117 | vty_out(vty, "%5zu %10zu.%03zu %9zu %8zu %9zu %8zu %9zu", |
72327cf3 MS |
118 | a->total_active, a->cpu.total / 1000, a->cpu.total % 1000, |
119 | a->total_calls, (a->cpu.total / a->total_calls), a->cpu.max, | |
120 | (a->real.total / a->total_calls), a->real.max); | |
84d951d0 | 121 | vty_out(vty, " %c%c%c%c%c %s\n", |
d62a17ae | 122 | a->types & (1 << THREAD_READ) ? 'R' : ' ', |
123 | a->types & (1 << THREAD_WRITE) ? 'W' : ' ', | |
124 | a->types & (1 << THREAD_TIMER) ? 'T' : ' ', | |
125 | a->types & (1 << THREAD_EVENT) ? 'E' : ' ', | |
126 | a->types & (1 << THREAD_EXECUTE) ? 'X' : ' ', a->funcname); | |
e04ab74d | 127 | } |
128 | ||
e3b78da8 | 129 | static void cpu_record_hash_print(struct hash_bucket *bucket, void *args[]) |
e04ab74d | 130 | { |
d62a17ae | 131 | struct cpu_thread_history *totals = args[0]; |
fbcac826 | 132 | struct cpu_thread_history copy; |
d62a17ae | 133 | struct vty *vty = args[1]; |
fbcac826 | 134 | uint8_t *filter = args[2]; |
d62a17ae | 135 | |
136 | struct cpu_thread_history *a = bucket->data; | |
137 | ||
fbcac826 QY |
138 | copy.total_active = |
139 | atomic_load_explicit(&a->total_active, memory_order_seq_cst); | |
140 | copy.total_calls = | |
141 | atomic_load_explicit(&a->total_calls, memory_order_seq_cst); | |
142 | copy.cpu.total = | |
143 | atomic_load_explicit(&a->cpu.total, memory_order_seq_cst); | |
144 | copy.cpu.max = atomic_load_explicit(&a->cpu.max, memory_order_seq_cst); | |
145 | copy.real.total = | |
146 | atomic_load_explicit(&a->real.total, memory_order_seq_cst); | |
147 | copy.real.max = | |
148 | atomic_load_explicit(&a->real.max, memory_order_seq_cst); | |
149 | copy.types = atomic_load_explicit(&a->types, memory_order_seq_cst); | |
150 | copy.funcname = a->funcname; | |
151 | ||
152 | if (!(copy.types & *filter)) | |
d62a17ae | 153 | return; |
fbcac826 QY |
154 | |
155 | vty_out_cpu_thread_history(vty, ©); | |
156 | totals->total_active += copy.total_active; | |
157 | totals->total_calls += copy.total_calls; | |
158 | totals->real.total += copy.real.total; | |
159 | if (totals->real.max < copy.real.max) | |
160 | totals->real.max = copy.real.max; | |
161 | totals->cpu.total += copy.cpu.total; | |
162 | if (totals->cpu.max < copy.cpu.max) | |
163 | totals->cpu.max = copy.cpu.max; | |
e04ab74d | 164 | } |
165 | ||
fbcac826 | 166 | static void cpu_record_print(struct vty *vty, uint8_t filter) |
e04ab74d | 167 | { |
d62a17ae | 168 | struct cpu_thread_history tmp; |
169 | void *args[3] = {&tmp, vty, &filter}; | |
170 | struct thread_master *m; | |
171 | struct listnode *ln; | |
172 | ||
0d6f7fd6 | 173 | memset(&tmp, 0, sizeof(tmp)); |
d62a17ae | 174 | tmp.funcname = "TOTAL"; |
175 | tmp.types = filter; | |
176 | ||
00dffa8c | 177 | frr_with_mutex(&masters_mtx) { |
d62a17ae | 178 | for (ALL_LIST_ELEMENTS_RO(masters, ln, m)) { |
179 | const char *name = m->name ? m->name : "main"; | |
180 | ||
181 | char underline[strlen(name) + 1]; | |
182 | memset(underline, '-', sizeof(underline)); | |
4f113d60 | 183 | underline[sizeof(underline) - 1] = '\0'; |
d62a17ae | 184 | |
185 | vty_out(vty, "\n"); | |
186 | vty_out(vty, "Showing statistics for pthread %s\n", | |
187 | name); | |
188 | vty_out(vty, "-------------------------------%s\n", | |
189 | underline); | |
84d951d0 | 190 | vty_out(vty, "%30s %18s %18s\n", "", |
d62a17ae | 191 | "CPU (user+system):", "Real (wall-clock):"); |
192 | vty_out(vty, | |
193 | "Active Runtime(ms) Invoked Avg uSec Max uSecs"); | |
194 | vty_out(vty, " Avg uSec Max uSecs"); | |
195 | vty_out(vty, " Type Thread\n"); | |
196 | ||
197 | if (m->cpu_record->count) | |
198 | hash_iterate( | |
199 | m->cpu_record, | |
e3b78da8 | 200 | (void (*)(struct hash_bucket *, |
d62a17ae | 201 | void *))cpu_record_hash_print, |
202 | args); | |
203 | else | |
204 | vty_out(vty, "No data to display yet.\n"); | |
205 | ||
206 | vty_out(vty, "\n"); | |
207 | } | |
208 | } | |
d62a17ae | 209 | |
210 | vty_out(vty, "\n"); | |
211 | vty_out(vty, "Total thread statistics\n"); | |
212 | vty_out(vty, "-------------------------\n"); | |
84d951d0 | 213 | vty_out(vty, "%30s %18s %18s\n", "", |
d62a17ae | 214 | "CPU (user+system):", "Real (wall-clock):"); |
215 | vty_out(vty, "Active Runtime(ms) Invoked Avg uSec Max uSecs"); | |
216 | vty_out(vty, " Avg uSec Max uSecs"); | |
217 | vty_out(vty, " Type Thread\n"); | |
218 | ||
219 | if (tmp.total_calls > 0) | |
220 | vty_out_cpu_thread_history(vty, &tmp); | |
e04ab74d | 221 | } |
f75e802d | 222 | #endif |
e04ab74d | 223 | |
e3b78da8 | 224 | static void cpu_record_hash_clear(struct hash_bucket *bucket, void *args[]) |
e276eb82 | 225 | { |
fbcac826 | 226 | uint8_t *filter = args[0]; |
d62a17ae | 227 | struct hash *cpu_record = args[1]; |
228 | ||
229 | struct cpu_thread_history *a = bucket->data; | |
62f44022 | 230 | |
d62a17ae | 231 | if (!(a->types & *filter)) |
232 | return; | |
f48f65d2 | 233 | |
d62a17ae | 234 | hash_release(cpu_record, bucket->data); |
e276eb82 PJ |
235 | } |
236 | ||
fbcac826 | 237 | static void cpu_record_clear(uint8_t filter) |
e276eb82 | 238 | { |
fbcac826 | 239 | uint8_t *tmp = &filter; |
d62a17ae | 240 | struct thread_master *m; |
241 | struct listnode *ln; | |
242 | ||
00dffa8c | 243 | frr_with_mutex(&masters_mtx) { |
d62a17ae | 244 | for (ALL_LIST_ELEMENTS_RO(masters, ln, m)) { |
00dffa8c | 245 | frr_with_mutex(&m->mtx) { |
d62a17ae | 246 | void *args[2] = {tmp, m->cpu_record}; |
247 | hash_iterate( | |
248 | m->cpu_record, | |
e3b78da8 | 249 | (void (*)(struct hash_bucket *, |
d62a17ae | 250 | void *))cpu_record_hash_clear, |
251 | args); | |
252 | } | |
d62a17ae | 253 | } |
254 | } | |
62f44022 QY |
255 | } |
256 | ||
fbcac826 | 257 | static uint8_t parse_filter(const char *filterstr) |
62f44022 | 258 | { |
d62a17ae | 259 | int i = 0; |
260 | int filter = 0; | |
261 | ||
262 | while (filterstr[i] != '\0') { | |
263 | switch (filterstr[i]) { | |
264 | case 'r': | |
265 | case 'R': | |
266 | filter |= (1 << THREAD_READ); | |
267 | break; | |
268 | case 'w': | |
269 | case 'W': | |
270 | filter |= (1 << THREAD_WRITE); | |
271 | break; | |
272 | case 't': | |
273 | case 'T': | |
274 | filter |= (1 << THREAD_TIMER); | |
275 | break; | |
276 | case 'e': | |
277 | case 'E': | |
278 | filter |= (1 << THREAD_EVENT); | |
279 | break; | |
280 | case 'x': | |
281 | case 'X': | |
282 | filter |= (1 << THREAD_EXECUTE); | |
283 | break; | |
284 | default: | |
285 | break; | |
286 | } | |
287 | ++i; | |
288 | } | |
289 | return filter; | |
62f44022 QY |
290 | } |
291 | ||
f75e802d | 292 | #ifndef EXCLUDE_CPU_TIME |
62f44022 QY |
293 | DEFUN (show_thread_cpu, |
294 | show_thread_cpu_cmd, | |
295 | "show thread cpu [FILTER]", | |
296 | SHOW_STR | |
297 | "Thread information\n" | |
298 | "Thread CPU usage\n" | |
61fa0b97 | 299 | "Display filter (rwtex)\n") |
62f44022 | 300 | { |
fbcac826 | 301 | uint8_t filter = (uint8_t)-1U; |
d62a17ae | 302 | int idx = 0; |
303 | ||
304 | if (argv_find(argv, argc, "FILTER", &idx)) { | |
305 | filter = parse_filter(argv[idx]->arg); | |
306 | if (!filter) { | |
307 | vty_out(vty, | |
3efd0893 | 308 | "Invalid filter \"%s\" specified; must contain at leastone of 'RWTEXB'\n", |
d62a17ae | 309 | argv[idx]->arg); |
310 | return CMD_WARNING; | |
311 | } | |
312 | } | |
313 | ||
314 | cpu_record_print(vty, filter); | |
315 | return CMD_SUCCESS; | |
e276eb82 | 316 | } |
f75e802d | 317 | #endif |
e276eb82 | 318 | |
8872626b DS |
319 | static void show_thread_poll_helper(struct vty *vty, struct thread_master *m) |
320 | { | |
321 | const char *name = m->name ? m->name : "main"; | |
322 | char underline[strlen(name) + 1]; | |
a0b36ae6 | 323 | struct thread *thread; |
8872626b DS |
324 | uint32_t i; |
325 | ||
326 | memset(underline, '-', sizeof(underline)); | |
327 | underline[sizeof(underline) - 1] = '\0'; | |
328 | ||
329 | vty_out(vty, "\nShowing poll FD's for %s\n", name); | |
330 | vty_out(vty, "----------------------%s\n", underline); | |
6c19478a DS |
331 | vty_out(vty, "Count: %u/%d\n", (uint32_t)m->handler.pfdcount, |
332 | m->fd_limit); | |
a0b36ae6 DS |
333 | for (i = 0; i < m->handler.pfdcount; i++) { |
334 | vty_out(vty, "\t%6d fd:%6d events:%2d revents:%2d\t\t", i, | |
335 | m->handler.pfds[i].fd, m->handler.pfds[i].events, | |
8872626b | 336 | m->handler.pfds[i].revents); |
a0b36ae6 DS |
337 | |
338 | if (m->handler.pfds[i].events & POLLIN) { | |
339 | thread = m->read[m->handler.pfds[i].fd]; | |
340 | ||
341 | if (!thread) | |
342 | vty_out(vty, "ERROR "); | |
343 | else | |
60a3efec | 344 | vty_out(vty, "%s ", thread->xref->funcname); |
a0b36ae6 DS |
345 | } else |
346 | vty_out(vty, " "); | |
347 | ||
348 | if (m->handler.pfds[i].events & POLLOUT) { | |
349 | thread = m->write[m->handler.pfds[i].fd]; | |
350 | ||
351 | if (!thread) | |
352 | vty_out(vty, "ERROR\n"); | |
353 | else | |
60a3efec | 354 | vty_out(vty, "%s\n", thread->xref->funcname); |
a0b36ae6 DS |
355 | } else |
356 | vty_out(vty, "\n"); | |
357 | } | |
8872626b DS |
358 | } |
359 | ||
360 | DEFUN (show_thread_poll, | |
361 | show_thread_poll_cmd, | |
362 | "show thread poll", | |
363 | SHOW_STR | |
364 | "Thread information\n" | |
365 | "Show poll FD's and information\n") | |
366 | { | |
367 | struct listnode *node; | |
368 | struct thread_master *m; | |
369 | ||
00dffa8c | 370 | frr_with_mutex(&masters_mtx) { |
8872626b DS |
371 | for (ALL_LIST_ELEMENTS_RO(masters, node, m)) { |
372 | show_thread_poll_helper(vty, m); | |
373 | } | |
374 | } | |
8872626b DS |
375 | |
376 | return CMD_SUCCESS; | |
377 | } | |
378 | ||
379 | ||
49d41a26 DS |
380 | DEFUN (clear_thread_cpu, |
381 | clear_thread_cpu_cmd, | |
382 | "clear thread cpu [FILTER]", | |
62f44022 | 383 | "Clear stored data in all pthreads\n" |
49d41a26 DS |
384 | "Thread information\n" |
385 | "Thread CPU usage\n" | |
386 | "Display filter (rwtexb)\n") | |
e276eb82 | 387 | { |
fbcac826 | 388 | uint8_t filter = (uint8_t)-1U; |
d62a17ae | 389 | int idx = 0; |
390 | ||
391 | if (argv_find(argv, argc, "FILTER", &idx)) { | |
392 | filter = parse_filter(argv[idx]->arg); | |
393 | if (!filter) { | |
394 | vty_out(vty, | |
3efd0893 | 395 | "Invalid filter \"%s\" specified; must contain at leastone of 'RWTEXB'\n", |
d62a17ae | 396 | argv[idx]->arg); |
397 | return CMD_WARNING; | |
398 | } | |
399 | } | |
400 | ||
401 | cpu_record_clear(filter); | |
402 | return CMD_SUCCESS; | |
e276eb82 | 403 | } |
6b0655a2 | 404 | |
d62a17ae | 405 | void thread_cmd_init(void) |
0b84f294 | 406 | { |
f75e802d | 407 | #ifndef EXCLUDE_CPU_TIME |
d62a17ae | 408 | install_element(VIEW_NODE, &show_thread_cpu_cmd); |
f75e802d | 409 | #endif |
8872626b | 410 | install_element(VIEW_NODE, &show_thread_poll_cmd); |
d62a17ae | 411 | install_element(ENABLE_NODE, &clear_thread_cpu_cmd); |
0b84f294 | 412 | } |
62f44022 QY |
413 | /* CLI end ------------------------------------------------------------------ */ |
414 | ||
0b84f294 | 415 | |
d62a17ae | 416 | static void cancelreq_del(void *cr) |
63ccb9cb | 417 | { |
d62a17ae | 418 | XFREE(MTYPE_TMP, cr); |
63ccb9cb QY |
419 | } |
420 | ||
e0bebc7c | 421 | /* initializer, only ever called once */ |
4d762f26 | 422 | static void initializer(void) |
e0bebc7c | 423 | { |
d62a17ae | 424 | pthread_key_create(&thread_current, NULL); |
e0bebc7c QY |
425 | } |
426 | ||
d62a17ae | 427 | struct thread_master *thread_master_create(const char *name) |
718e3744 | 428 | { |
d62a17ae | 429 | struct thread_master *rv; |
430 | struct rlimit limit; | |
431 | ||
432 | pthread_once(&init_once, &initializer); | |
433 | ||
434 | rv = XCALLOC(MTYPE_THREAD_MASTER, sizeof(struct thread_master)); | |
d62a17ae | 435 | |
436 | /* Initialize master mutex */ | |
437 | pthread_mutex_init(&rv->mtx, NULL); | |
438 | pthread_cond_init(&rv->cancel_cond, NULL); | |
439 | ||
440 | /* Set name */ | |
7ffcd8bd QY |
441 | name = name ? name : "default"; |
442 | rv->name = XSTRDUP(MTYPE_THREAD_MASTER, name); | |
d62a17ae | 443 | |
444 | /* Initialize I/O task data structures */ | |
1a9f340b MS |
445 | |
446 | /* Use configured limit if present, ulimit otherwise. */ | |
447 | rv->fd_limit = frr_get_fd_limit(); | |
448 | if (rv->fd_limit == 0) { | |
449 | getrlimit(RLIMIT_NOFILE, &limit); | |
450 | rv->fd_limit = (int)limit.rlim_cur; | |
451 | } | |
452 | ||
a6f235f3 DS |
453 | rv->read = XCALLOC(MTYPE_THREAD_POLL, |
454 | sizeof(struct thread *) * rv->fd_limit); | |
455 | ||
456 | rv->write = XCALLOC(MTYPE_THREAD_POLL, | |
457 | sizeof(struct thread *) * rv->fd_limit); | |
d62a17ae | 458 | |
7ffcd8bd QY |
459 | char tmhashname[strlen(name) + 32]; |
460 | snprintf(tmhashname, sizeof(tmhashname), "%s - threadmaster event hash", | |
461 | name); | |
bd74dc61 | 462 | rv->cpu_record = hash_create_size( |
d8b87afe | 463 | 8, (unsigned int (*)(const void *))cpu_record_hash_key, |
74df8d6d | 464 | (bool (*)(const void *, const void *))cpu_record_hash_cmp, |
7ffcd8bd | 465 | tmhashname); |
d62a17ae | 466 | |
c284542b DL |
467 | thread_list_init(&rv->event); |
468 | thread_list_init(&rv->ready); | |
469 | thread_list_init(&rv->unuse); | |
27d29ced | 470 | thread_timer_list_init(&rv->timer); |
d62a17ae | 471 | |
472 | /* Initialize thread_fetch() settings */ | |
473 | rv->spin = true; | |
474 | rv->handle_signals = true; | |
475 | ||
476 | /* Set pthread owner, should be updated by actual owner */ | |
477 | rv->owner = pthread_self(); | |
478 | rv->cancel_req = list_new(); | |
479 | rv->cancel_req->del = cancelreq_del; | |
480 | rv->canceled = true; | |
481 | ||
482 | /* Initialize pipe poker */ | |
483 | pipe(rv->io_pipe); | |
484 | set_nonblocking(rv->io_pipe[0]); | |
485 | set_nonblocking(rv->io_pipe[1]); | |
486 | ||
487 | /* Initialize data structures for poll() */ | |
488 | rv->handler.pfdsize = rv->fd_limit; | |
489 | rv->handler.pfdcount = 0; | |
490 | rv->handler.pfds = XCALLOC(MTYPE_THREAD_MASTER, | |
491 | sizeof(struct pollfd) * rv->handler.pfdsize); | |
492 | rv->handler.copy = XCALLOC(MTYPE_THREAD_MASTER, | |
493 | sizeof(struct pollfd) * rv->handler.pfdsize); | |
494 | ||
eff09c66 | 495 | /* add to list of threadmasters */ |
00dffa8c | 496 | frr_with_mutex(&masters_mtx) { |
eff09c66 QY |
497 | if (!masters) |
498 | masters = list_new(); | |
499 | ||
d62a17ae | 500 | listnode_add(masters, rv); |
501 | } | |
d62a17ae | 502 | |
503 | return rv; | |
718e3744 | 504 | } |
505 | ||
d8a8a8de QY |
506 | void thread_master_set_name(struct thread_master *master, const char *name) |
507 | { | |
00dffa8c | 508 | frr_with_mutex(&master->mtx) { |
0a22ddfb | 509 | XFREE(MTYPE_THREAD_MASTER, master->name); |
d8a8a8de QY |
510 | master->name = XSTRDUP(MTYPE_THREAD_MASTER, name); |
511 | } | |
d8a8a8de QY |
512 | } |
513 | ||
6ed04aa2 DS |
514 | #define THREAD_UNUSED_DEPTH 10 |
515 | ||
718e3744 | 516 | /* Move thread to unuse list. */ |
d62a17ae | 517 | static void thread_add_unuse(struct thread_master *m, struct thread *thread) |
718e3744 | 518 | { |
6655966d RZ |
519 | pthread_mutex_t mtxc = thread->mtx; |
520 | ||
d62a17ae | 521 | assert(m != NULL && thread != NULL); |
d62a17ae | 522 | |
d62a17ae | 523 | thread->hist->total_active--; |
6ed04aa2 DS |
524 | memset(thread, 0, sizeof(struct thread)); |
525 | thread->type = THREAD_UNUSED; | |
526 | ||
6655966d RZ |
527 | /* Restore the thread mutex context. */ |
528 | thread->mtx = mtxc; | |
529 | ||
c284542b DL |
530 | if (thread_list_count(&m->unuse) < THREAD_UNUSED_DEPTH) { |
531 | thread_list_add_tail(&m->unuse, thread); | |
6655966d RZ |
532 | return; |
533 | } | |
534 | ||
535 | thread_free(m, thread); | |
718e3744 | 536 | } |
537 | ||
538 | /* Free all unused thread. */ | |
c284542b DL |
539 | static void thread_list_free(struct thread_master *m, |
540 | struct thread_list_head *list) | |
718e3744 | 541 | { |
d62a17ae | 542 | struct thread *t; |
d62a17ae | 543 | |
c284542b | 544 | while ((t = thread_list_pop(list))) |
6655966d | 545 | thread_free(m, t); |
718e3744 | 546 | } |
547 | ||
d62a17ae | 548 | static void thread_array_free(struct thread_master *m, |
549 | struct thread **thread_array) | |
308d14ae | 550 | { |
d62a17ae | 551 | struct thread *t; |
552 | int index; | |
553 | ||
554 | for (index = 0; index < m->fd_limit; ++index) { | |
555 | t = thread_array[index]; | |
556 | if (t) { | |
557 | thread_array[index] = NULL; | |
6655966d | 558 | thread_free(m, t); |
d62a17ae | 559 | } |
560 | } | |
a6f235f3 | 561 | XFREE(MTYPE_THREAD_POLL, thread_array); |
308d14ae DV |
562 | } |
563 | ||
495f0b13 DS |
564 | /* |
565 | * thread_master_free_unused | |
566 | * | |
567 | * As threads are finished with they are put on the | |
568 | * unuse list for later reuse. | |
569 | * If we are shutting down, Free up unused threads | |
570 | * So we can see if we forget to shut anything off | |
571 | */ | |
d62a17ae | 572 | void thread_master_free_unused(struct thread_master *m) |
495f0b13 | 573 | { |
00dffa8c | 574 | frr_with_mutex(&m->mtx) { |
d62a17ae | 575 | struct thread *t; |
c284542b | 576 | while ((t = thread_list_pop(&m->unuse))) |
6655966d | 577 | thread_free(m, t); |
d62a17ae | 578 | } |
495f0b13 DS |
579 | } |
580 | ||
718e3744 | 581 | /* Stop thread scheduler. */ |
d62a17ae | 582 | void thread_master_free(struct thread_master *m) |
718e3744 | 583 | { |
27d29ced DL |
584 | struct thread *t; |
585 | ||
00dffa8c | 586 | frr_with_mutex(&masters_mtx) { |
d62a17ae | 587 | listnode_delete(masters, m); |
eff09c66 | 588 | if (masters->count == 0) { |
6a154c88 | 589 | list_delete(&masters); |
eff09c66 | 590 | } |
d62a17ae | 591 | } |
d62a17ae | 592 | |
593 | thread_array_free(m, m->read); | |
594 | thread_array_free(m, m->write); | |
27d29ced DL |
595 | while ((t = thread_timer_list_pop(&m->timer))) |
596 | thread_free(m, t); | |
d62a17ae | 597 | thread_list_free(m, &m->event); |
598 | thread_list_free(m, &m->ready); | |
599 | thread_list_free(m, &m->unuse); | |
600 | pthread_mutex_destroy(&m->mtx); | |
33844bbe | 601 | pthread_cond_destroy(&m->cancel_cond); |
d62a17ae | 602 | close(m->io_pipe[0]); |
603 | close(m->io_pipe[1]); | |
6a154c88 | 604 | list_delete(&m->cancel_req); |
1a0a92ea | 605 | m->cancel_req = NULL; |
d62a17ae | 606 | |
607 | hash_clean(m->cpu_record, cpu_record_hash_free); | |
608 | hash_free(m->cpu_record); | |
609 | m->cpu_record = NULL; | |
610 | ||
0a22ddfb | 611 | XFREE(MTYPE_THREAD_MASTER, m->name); |
d62a17ae | 612 | XFREE(MTYPE_THREAD_MASTER, m->handler.pfds); |
613 | XFREE(MTYPE_THREAD_MASTER, m->handler.copy); | |
614 | XFREE(MTYPE_THREAD_MASTER, m); | |
718e3744 | 615 | } |
616 | ||
78ca0342 CF |
617 | /* Return remain time in miliseconds. */ |
618 | unsigned long thread_timer_remain_msec(struct thread *thread) | |
718e3744 | 619 | { |
d62a17ae | 620 | int64_t remain; |
1189d95f | 621 | |
00dffa8c | 622 | frr_with_mutex(&thread->mtx) { |
78ca0342 | 623 | remain = monotime_until(&thread->u.sands, NULL) / 1000LL; |
d62a17ae | 624 | } |
1189d95f | 625 | |
d62a17ae | 626 | return remain < 0 ? 0 : remain; |
718e3744 | 627 | } |
628 | ||
78ca0342 CF |
629 | /* Return remain time in seconds. */ |
630 | unsigned long thread_timer_remain_second(struct thread *thread) | |
631 | { | |
632 | return thread_timer_remain_msec(thread) / 1000LL; | |
633 | } | |
634 | ||
d62a17ae | 635 | struct timeval thread_timer_remain(struct thread *thread) |
6ac44687 | 636 | { |
d62a17ae | 637 | struct timeval remain; |
00dffa8c | 638 | frr_with_mutex(&thread->mtx) { |
d62a17ae | 639 | monotime_until(&thread->u.sands, &remain); |
640 | } | |
d62a17ae | 641 | return remain; |
6ac44687 CF |
642 | } |
643 | ||
0447957e AK |
644 | static int time_hhmmss(char *buf, int buf_size, long sec) |
645 | { | |
646 | long hh; | |
647 | long mm; | |
648 | int wr; | |
649 | ||
650 | zassert(buf_size >= 8); | |
651 | ||
652 | hh = sec / 3600; | |
653 | sec %= 3600; | |
654 | mm = sec / 60; | |
655 | sec %= 60; | |
656 | ||
657 | wr = snprintf(buf, buf_size, "%02ld:%02ld:%02ld", hh, mm, sec); | |
658 | ||
659 | return wr != 8; | |
660 | } | |
661 | ||
662 | char *thread_timer_to_hhmmss(char *buf, int buf_size, | |
663 | struct thread *t_timer) | |
664 | { | |
665 | if (t_timer) { | |
666 | time_hhmmss(buf, buf_size, | |
667 | thread_timer_remain_second(t_timer)); | |
668 | } else { | |
669 | snprintf(buf, buf_size, "--:--:--"); | |
670 | } | |
671 | return buf; | |
672 | } | |
673 | ||
718e3744 | 674 | /* Get new thread. */ |
d7c0a89a | 675 | static struct thread *thread_get(struct thread_master *m, uint8_t type, |
d62a17ae | 676 | int (*func)(struct thread *), void *arg, |
60a3efec | 677 | const struct xref_threadsched *xref) |
718e3744 | 678 | { |
c284542b | 679 | struct thread *thread = thread_list_pop(&m->unuse); |
d62a17ae | 680 | struct cpu_thread_history tmp; |
681 | ||
682 | if (!thread) { | |
683 | thread = XCALLOC(MTYPE_THREAD, sizeof(struct thread)); | |
684 | /* mutex only needs to be initialized at struct creation. */ | |
685 | pthread_mutex_init(&thread->mtx, NULL); | |
686 | m->alloc++; | |
687 | } | |
688 | ||
689 | thread->type = type; | |
690 | thread->add_type = type; | |
691 | thread->master = m; | |
692 | thread->arg = arg; | |
d62a17ae | 693 | thread->yield = THREAD_YIELD_TIME_SLOT; /* default */ |
694 | thread->ref = NULL; | |
695 | ||
696 | /* | |
697 | * So if the passed in funcname is not what we have | |
698 | * stored that means the thread->hist needs to be | |
699 | * updated. We keep the last one around in unused | |
700 | * under the assumption that we are probably | |
701 | * going to immediately allocate the same | |
702 | * type of thread. | |
703 | * This hopefully saves us some serious | |
704 | * hash_get lookups. | |
705 | */ | |
60a3efec DL |
706 | if ((thread->xref && thread->xref->funcname != xref->funcname) |
707 | || thread->func != func) { | |
d62a17ae | 708 | tmp.func = func; |
60a3efec | 709 | tmp.funcname = xref->funcname; |
d62a17ae | 710 | thread->hist = |
711 | hash_get(m->cpu_record, &tmp, | |
712 | (void *(*)(void *))cpu_record_hash_alloc); | |
713 | } | |
714 | thread->hist->total_active++; | |
715 | thread->func = func; | |
60a3efec | 716 | thread->xref = xref; |
d62a17ae | 717 | |
718 | return thread; | |
718e3744 | 719 | } |
720 | ||
6655966d RZ |
721 | static void thread_free(struct thread_master *master, struct thread *thread) |
722 | { | |
723 | /* Update statistics. */ | |
724 | assert(master->alloc > 0); | |
725 | master->alloc--; | |
726 | ||
727 | /* Free allocated resources. */ | |
728 | pthread_mutex_destroy(&thread->mtx); | |
729 | XFREE(MTYPE_THREAD, thread); | |
730 | } | |
731 | ||
d81ca9a3 MS |
732 | static int fd_poll(struct thread_master *m, const struct timeval *timer_wait, |
733 | bool *eintr_p) | |
209a72a6 | 734 | { |
d81ca9a3 MS |
735 | sigset_t origsigs; |
736 | unsigned char trash[64]; | |
737 | nfds_t count = m->handler.copycount; | |
738 | ||
d279ef57 DS |
739 | /* |
740 | * If timer_wait is null here, that means poll() should block | |
741 | * indefinitely, unless the thread_master has overridden it by setting | |
d62a17ae | 742 | * ->selectpoll_timeout. |
d279ef57 | 743 | * |
d62a17ae | 744 | * If the value is positive, it specifies the maximum number of |
d279ef57 DS |
745 | * milliseconds to wait. If the timeout is -1, it specifies that |
746 | * we should never wait and always return immediately even if no | |
747 | * event is detected. If the value is zero, the behavior is default. | |
748 | */ | |
d62a17ae | 749 | int timeout = -1; |
750 | ||
751 | /* number of file descriptors with events */ | |
752 | int num; | |
753 | ||
754 | if (timer_wait != NULL | |
755 | && m->selectpoll_timeout == 0) // use the default value | |
756 | timeout = (timer_wait->tv_sec * 1000) | |
757 | + (timer_wait->tv_usec / 1000); | |
758 | else if (m->selectpoll_timeout > 0) // use the user's timeout | |
759 | timeout = m->selectpoll_timeout; | |
760 | else if (m->selectpoll_timeout | |
761 | < 0) // effect a poll (return immediately) | |
762 | timeout = 0; | |
763 | ||
0bdeb5e5 | 764 | zlog_tls_buffer_flush(); |
3e41733f DL |
765 | rcu_read_unlock(); |
766 | rcu_assert_read_unlocked(); | |
767 | ||
d62a17ae | 768 | /* add poll pipe poker */ |
d81ca9a3 MS |
769 | assert(count + 1 < m->handler.pfdsize); |
770 | m->handler.copy[count].fd = m->io_pipe[0]; | |
771 | m->handler.copy[count].events = POLLIN; | |
772 | m->handler.copy[count].revents = 0x00; | |
773 | ||
774 | /* We need to deal with a signal-handling race here: we | |
775 | * don't want to miss a crucial signal, such as SIGTERM or SIGINT, | |
776 | * that may arrive just before we enter poll(). We will block the | |
777 | * key signals, then check whether any have arrived - if so, we return | |
778 | * before calling poll(). If not, we'll re-enable the signals | |
779 | * in the ppoll() call. | |
780 | */ | |
781 | ||
782 | sigemptyset(&origsigs); | |
783 | if (m->handle_signals) { | |
784 | /* Main pthread that handles the app signals */ | |
785 | if (frr_sigevent_check(&origsigs)) { | |
786 | /* Signal to process - restore signal mask and return */ | |
787 | pthread_sigmask(SIG_SETMASK, &origsigs, NULL); | |
788 | num = -1; | |
789 | *eintr_p = true; | |
790 | goto done; | |
791 | } | |
792 | } else { | |
793 | /* Don't make any changes for the non-main pthreads */ | |
794 | pthread_sigmask(SIG_SETMASK, NULL, &origsigs); | |
795 | } | |
d62a17ae | 796 | |
d81ca9a3 MS |
797 | #if defined(HAVE_PPOLL) |
798 | struct timespec ts, *tsp; | |
799 | ||
800 | if (timeout >= 0) { | |
801 | ts.tv_sec = timeout / 1000; | |
802 | ts.tv_nsec = (timeout % 1000) * 1000000; | |
803 | tsp = &ts; | |
804 | } else | |
805 | tsp = NULL; | |
806 | ||
807 | num = ppoll(m->handler.copy, count + 1, tsp, &origsigs); | |
808 | pthread_sigmask(SIG_SETMASK, &origsigs, NULL); | |
809 | #else | |
810 | /* Not ideal - there is a race after we restore the signal mask */ | |
811 | pthread_sigmask(SIG_SETMASK, &origsigs, NULL); | |
812 | num = poll(m->handler.copy, count + 1, timeout); | |
813 | #endif | |
d62a17ae | 814 | |
d81ca9a3 MS |
815 | done: |
816 | ||
817 | if (num < 0 && errno == EINTR) | |
818 | *eintr_p = true; | |
819 | ||
820 | if (num > 0 && m->handler.copy[count].revents != 0 && num--) | |
d62a17ae | 821 | while (read(m->io_pipe[0], &trash, sizeof(trash)) > 0) |
822 | ; | |
823 | ||
3e41733f DL |
824 | rcu_read_lock(); |
825 | ||
d62a17ae | 826 | return num; |
209a72a6 DS |
827 | } |
828 | ||
718e3744 | 829 | /* Add new read thread. */ |
60a3efec DL |
830 | struct thread *_thread_add_read_write(const struct xref_threadsched *xref, |
831 | struct thread_master *m, | |
832 | int (*func)(struct thread *), | |
833 | void *arg, int fd, struct thread **t_ptr) | |
718e3744 | 834 | { |
60a3efec | 835 | int dir = xref->thread_type; |
d62a17ae | 836 | struct thread *thread = NULL; |
1ef14bee | 837 | struct thread **thread_array; |
d62a17ae | 838 | |
abf96a87 | 839 | if (dir == THREAD_READ) |
6c3aa850 DL |
840 | frrtrace(9, frr_libfrr, schedule_read, m, |
841 | xref->funcname, xref->xref.file, xref->xref.line, | |
842 | t_ptr, fd, 0, arg, 0); | |
abf96a87 | 843 | else |
6c3aa850 DL |
844 | frrtrace(9, frr_libfrr, schedule_write, m, |
845 | xref->funcname, xref->xref.file, xref->xref.line, | |
846 | t_ptr, fd, 0, arg, 0); | |
abf96a87 | 847 | |
9b864cd3 | 848 | assert(fd >= 0 && fd < m->fd_limit); |
00dffa8c DL |
849 | frr_with_mutex(&m->mtx) { |
850 | if (t_ptr && *t_ptr) | |
851 | // thread is already scheduled; don't reschedule | |
852 | break; | |
d62a17ae | 853 | |
854 | /* default to a new pollfd */ | |
855 | nfds_t queuepos = m->handler.pfdcount; | |
856 | ||
1ef14bee DS |
857 | if (dir == THREAD_READ) |
858 | thread_array = m->read; | |
859 | else | |
860 | thread_array = m->write; | |
861 | ||
d62a17ae | 862 | /* if we already have a pollfd for our file descriptor, find and |
863 | * use it */ | |
864 | for (nfds_t i = 0; i < m->handler.pfdcount; i++) | |
865 | if (m->handler.pfds[i].fd == fd) { | |
866 | queuepos = i; | |
1ef14bee DS |
867 | |
868 | #ifdef DEV_BUILD | |
869 | /* | |
870 | * What happens if we have a thread already | |
871 | * created for this event? | |
872 | */ | |
873 | if (thread_array[fd]) | |
874 | assert(!"Thread already scheduled for file descriptor"); | |
875 | #endif | |
d62a17ae | 876 | break; |
877 | } | |
878 | ||
879 | /* make sure we have room for this fd + pipe poker fd */ | |
880 | assert(queuepos + 1 < m->handler.pfdsize); | |
881 | ||
60a3efec | 882 | thread = thread_get(m, dir, func, arg, xref); |
d62a17ae | 883 | |
884 | m->handler.pfds[queuepos].fd = fd; | |
885 | m->handler.pfds[queuepos].events |= | |
886 | (dir == THREAD_READ ? POLLIN : POLLOUT); | |
887 | ||
888 | if (queuepos == m->handler.pfdcount) | |
889 | m->handler.pfdcount++; | |
890 | ||
891 | if (thread) { | |
00dffa8c | 892 | frr_with_mutex(&thread->mtx) { |
d62a17ae | 893 | thread->u.fd = fd; |
1ef14bee | 894 | thread_array[thread->u.fd] = thread; |
d62a17ae | 895 | } |
d62a17ae | 896 | |
897 | if (t_ptr) { | |
898 | *t_ptr = thread; | |
899 | thread->ref = t_ptr; | |
900 | } | |
901 | } | |
902 | ||
903 | AWAKEN(m); | |
904 | } | |
d62a17ae | 905 | |
906 | return thread; | |
718e3744 | 907 | } |
908 | ||
56a94b36 | 909 | static struct thread * |
60a3efec DL |
910 | _thread_add_timer_timeval(const struct xref_threadsched *xref, |
911 | struct thread_master *m, int (*func)(struct thread *), | |
912 | int type, void *arg, struct timeval *time_relative, | |
913 | struct thread **t_ptr) | |
718e3744 | 914 | { |
d62a17ae | 915 | struct thread *thread; |
d62a17ae | 916 | |
917 | assert(m != NULL); | |
918 | ||
919 | assert(type == THREAD_TIMER); | |
920 | assert(time_relative); | |
921 | ||
6c3aa850 DL |
922 | frrtrace(9, frr_libfrr, schedule_timer, m, |
923 | xref->funcname, xref->xref.file, xref->xref.line, | |
c7bb4f00 | 924 | t_ptr, 0, 0, arg, (long)time_relative->tv_sec); |
abf96a87 | 925 | |
00dffa8c DL |
926 | frr_with_mutex(&m->mtx) { |
927 | if (t_ptr && *t_ptr) | |
d279ef57 | 928 | /* thread is already scheduled; don't reschedule */ |
d62a17ae | 929 | return NULL; |
d62a17ae | 930 | |
60a3efec | 931 | thread = thread_get(m, type, func, arg, xref); |
d62a17ae | 932 | |
00dffa8c | 933 | frr_with_mutex(&thread->mtx) { |
d62a17ae | 934 | monotime(&thread->u.sands); |
935 | timeradd(&thread->u.sands, time_relative, | |
936 | &thread->u.sands); | |
27d29ced | 937 | thread_timer_list_add(&m->timer, thread); |
d62a17ae | 938 | if (t_ptr) { |
939 | *t_ptr = thread; | |
940 | thread->ref = t_ptr; | |
941 | } | |
942 | } | |
d62a17ae | 943 | |
944 | AWAKEN(m); | |
945 | } | |
d62a17ae | 946 | |
947 | return thread; | |
9e867fe6 | 948 | } |
949 | ||
98c91ac6 | 950 | |
951 | /* Add timer event thread. */ | |
60a3efec DL |
952 | struct thread *_thread_add_timer(const struct xref_threadsched *xref, |
953 | struct thread_master *m, | |
954 | int (*func)(struct thread *), | |
955 | void *arg, long timer, struct thread **t_ptr) | |
9e867fe6 | 956 | { |
d62a17ae | 957 | struct timeval trel; |
9e867fe6 | 958 | |
d62a17ae | 959 | assert(m != NULL); |
9e867fe6 | 960 | |
d62a17ae | 961 | trel.tv_sec = timer; |
962 | trel.tv_usec = 0; | |
9e867fe6 | 963 | |
60a3efec DL |
964 | return _thread_add_timer_timeval(xref, m, func, THREAD_TIMER, arg, |
965 | &trel, t_ptr); | |
98c91ac6 | 966 | } |
9e867fe6 | 967 | |
98c91ac6 | 968 | /* Add timer event thread with "millisecond" resolution */ |
60a3efec DL |
969 | struct thread *_thread_add_timer_msec(const struct xref_threadsched *xref, |
970 | struct thread_master *m, | |
971 | int (*func)(struct thread *), | |
972 | void *arg, long timer, | |
973 | struct thread **t_ptr) | |
98c91ac6 | 974 | { |
d62a17ae | 975 | struct timeval trel; |
9e867fe6 | 976 | |
d62a17ae | 977 | assert(m != NULL); |
718e3744 | 978 | |
d62a17ae | 979 | trel.tv_sec = timer / 1000; |
980 | trel.tv_usec = 1000 * (timer % 1000); | |
98c91ac6 | 981 | |
60a3efec DL |
982 | return _thread_add_timer_timeval(xref, m, func, THREAD_TIMER, arg, |
983 | &trel, t_ptr); | |
a48b4e6d | 984 | } |
985 | ||
d03c4cbd | 986 | /* Add timer event thread with "millisecond" resolution */ |
60a3efec DL |
987 | struct thread *_thread_add_timer_tv(const struct xref_threadsched *xref, |
988 | struct thread_master *m, | |
989 | int (*func)(struct thread *), | |
990 | void *arg, struct timeval *tv, | |
991 | struct thread **t_ptr) | |
d03c4cbd | 992 | { |
60a3efec DL |
993 | return _thread_add_timer_timeval(xref, m, func, THREAD_TIMER, arg, tv, |
994 | t_ptr); | |
d03c4cbd DL |
995 | } |
996 | ||
718e3744 | 997 | /* Add simple event thread. */ |
60a3efec DL |
998 | struct thread *_thread_add_event(const struct xref_threadsched *xref, |
999 | struct thread_master *m, | |
1000 | int (*func)(struct thread *), | |
1001 | void *arg, int val, struct thread **t_ptr) | |
718e3744 | 1002 | { |
00dffa8c | 1003 | struct thread *thread = NULL; |
d62a17ae | 1004 | |
6c3aa850 DL |
1005 | frrtrace(9, frr_libfrr, schedule_event, m, |
1006 | xref->funcname, xref->xref.file, xref->xref.line, | |
c7bb4f00 | 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 | |
6c3aa850 DL |
1242 | frrtrace(9, frr_libfrr, thread_cancel, master, |
1243 | (*thread)->xref->funcname, (*thread)->xref->xref.file, | |
1244 | (*thread)->xref->xref.line, NULL, (*thread)->u.fd, | |
b4d6e855 | 1245 | (*thread)->u.val, (*thread)->arg, (*thread)->u.sands.tv_sec); |
abf96a87 | 1246 | |
6ed04aa2 DS |
1247 | assert(master->owner == pthread_self()); |
1248 | ||
00dffa8c | 1249 | frr_with_mutex(&master->mtx) { |
d62a17ae | 1250 | struct cancel_req *cr = |
1251 | XCALLOC(MTYPE_TMP, sizeof(struct cancel_req)); | |
b3d6bc6e | 1252 | cr->thread = *thread; |
6ed04aa2 DS |
1253 | listnode_add(master->cancel_req, cr); |
1254 | do_thread_cancel(master); | |
d62a17ae | 1255 | } |
b3d6bc6e MS |
1256 | |
1257 | *thread = NULL; | |
63ccb9cb | 1258 | } |
1189d95f | 1259 | |
63ccb9cb QY |
1260 | /** |
1261 | * Asynchronous cancellation. | |
1262 | * | |
8797240e QY |
1263 | * Called with either a struct thread ** or void * to an event argument, |
1264 | * this function posts the correct cancellation request and blocks until it is | |
1265 | * serviced. | |
63ccb9cb QY |
1266 | * |
1267 | * If the thread is currently running, execution blocks until it completes. | |
1268 | * | |
8797240e QY |
1269 | * The last two parameters are mutually exclusive, i.e. if you pass one the |
1270 | * other must be NULL. | |
1271 | * | |
1272 | * When the cancellation procedure executes on the target thread_master, the | |
1273 | * thread * provided is checked for nullity. If it is null, the thread is | |
1274 | * assumed to no longer exist and the cancellation request is a no-op. Thus | |
1275 | * users of this API must pass a back-reference when scheduling the original | |
1276 | * task. | |
1277 | * | |
63ccb9cb QY |
1278 | * MT-Safe |
1279 | * | |
8797240e QY |
1280 | * @param master the thread master with the relevant event / task |
1281 | * @param thread pointer to thread to cancel | |
1282 | * @param eventobj the event | |
63ccb9cb | 1283 | */ |
d62a17ae | 1284 | void thread_cancel_async(struct thread_master *master, struct thread **thread, |
1285 | void *eventobj) | |
63ccb9cb | 1286 | { |
d62a17ae | 1287 | assert(!(thread && eventobj) && (thread || eventobj)); |
abf96a87 QY |
1288 | |
1289 | if (thread && *thread) | |
c7bb4f00 | 1290 | frrtrace(9, frr_libfrr, thread_cancel_async, master, |
6c3aa850 DL |
1291 | (*thread)->xref->funcname, (*thread)->xref->xref.file, |
1292 | (*thread)->xref->xref.line, NULL, (*thread)->u.fd, | |
c7bb4f00 QY |
1293 | (*thread)->u.val, (*thread)->arg, |
1294 | (*thread)->u.sands.tv_sec); | |
abf96a87 | 1295 | else |
c7bb4f00 QY |
1296 | frrtrace(9, frr_libfrr, thread_cancel_async, master, NULL, NULL, |
1297 | 0, NULL, 0, 0, eventobj, 0); | |
abf96a87 | 1298 | |
d62a17ae | 1299 | assert(master->owner != pthread_self()); |
1300 | ||
00dffa8c | 1301 | frr_with_mutex(&master->mtx) { |
d62a17ae | 1302 | master->canceled = false; |
1303 | ||
1304 | if (thread) { | |
1305 | struct cancel_req *cr = | |
1306 | XCALLOC(MTYPE_TMP, sizeof(struct cancel_req)); | |
1307 | cr->threadref = thread; | |
1308 | listnode_add(master->cancel_req, cr); | |
1309 | } else if (eventobj) { | |
1310 | struct cancel_req *cr = | |
1311 | XCALLOC(MTYPE_TMP, sizeof(struct cancel_req)); | |
1312 | cr->eventobj = eventobj; | |
1313 | listnode_add(master->cancel_req, cr); | |
1314 | } | |
1315 | AWAKEN(master); | |
1316 | ||
1317 | while (!master->canceled) | |
1318 | pthread_cond_wait(&master->cancel_cond, &master->mtx); | |
1319 | } | |
50478845 MS |
1320 | |
1321 | if (thread) | |
1322 | *thread = NULL; | |
718e3744 | 1323 | } |
63ccb9cb | 1324 | /* ------------------------------------------------------------------------- */ |
718e3744 | 1325 | |
27d29ced | 1326 | static struct timeval *thread_timer_wait(struct thread_timer_list_head *timers, |
d62a17ae | 1327 | struct timeval *timer_val) |
718e3744 | 1328 | { |
27d29ced DL |
1329 | if (!thread_timer_list_count(timers)) |
1330 | return NULL; | |
1331 | ||
1332 | struct thread *next_timer = thread_timer_list_first(timers); | |
1333 | monotime_until(&next_timer->u.sands, timer_val); | |
1334 | return timer_val; | |
718e3744 | 1335 | } |
718e3744 | 1336 | |
d62a17ae | 1337 | static struct thread *thread_run(struct thread_master *m, struct thread *thread, |
1338 | struct thread *fetch) | |
718e3744 | 1339 | { |
d62a17ae | 1340 | *fetch = *thread; |
1341 | thread_add_unuse(m, thread); | |
1342 | return fetch; | |
718e3744 | 1343 | } |
1344 | ||
d62a17ae | 1345 | static int thread_process_io_helper(struct thread_master *m, |
45f3d590 DS |
1346 | struct thread *thread, short state, |
1347 | short actual_state, int pos) | |
5d4ccd4e | 1348 | { |
d62a17ae | 1349 | struct thread **thread_array; |
1350 | ||
45f3d590 DS |
1351 | /* |
1352 | * poll() clears the .events field, but the pollfd array we | |
1353 | * pass to poll() is a copy of the one used to schedule threads. | |
1354 | * We need to synchronize state between the two here by applying | |
1355 | * the same changes poll() made on the copy of the "real" pollfd | |
1356 | * array. | |
1357 | * | |
1358 | * This cleans up a possible infinite loop where we refuse | |
1359 | * to respond to a poll event but poll is insistent that | |
1360 | * we should. | |
1361 | */ | |
1362 | m->handler.pfds[pos].events &= ~(state); | |
1363 | ||
1364 | if (!thread) { | |
1365 | if ((actual_state & (POLLHUP|POLLIN)) != POLLHUP) | |
1366 | flog_err(EC_LIB_NO_THREAD, | |
1d5453d6 | 1367 | "Attempting to process an I/O event but for fd: %d(%d) no thread to handle this!", |
45f3d590 | 1368 | m->handler.pfds[pos].fd, actual_state); |
d62a17ae | 1369 | return 0; |
45f3d590 | 1370 | } |
d62a17ae | 1371 | |
1372 | if (thread->type == THREAD_READ) | |
1373 | thread_array = m->read; | |
1374 | else | |
1375 | thread_array = m->write; | |
1376 | ||
1377 | thread_array[thread->u.fd] = NULL; | |
c284542b | 1378 | thread_list_add_tail(&m->ready, thread); |
d62a17ae | 1379 | thread->type = THREAD_READY; |
45f3d590 | 1380 | |
d62a17ae | 1381 | return 1; |
5d4ccd4e DS |
1382 | } |
1383 | ||
8797240e QY |
1384 | /** |
1385 | * Process I/O events. | |
1386 | * | |
1387 | * Walks through file descriptor array looking for those pollfds whose .revents | |
1388 | * field has something interesting. Deletes any invalid file descriptors. | |
1389 | * | |
1390 | * @param m the thread master | |
1391 | * @param num the number of active file descriptors (return value of poll()) | |
1392 | */ | |
d62a17ae | 1393 | static void thread_process_io(struct thread_master *m, unsigned int num) |
0a95a0d0 | 1394 | { |
d62a17ae | 1395 | unsigned int ready = 0; |
1396 | struct pollfd *pfds = m->handler.copy; | |
1397 | ||
1398 | for (nfds_t i = 0; i < m->handler.copycount && ready < num; ++i) { | |
1399 | /* no event for current fd? immediately continue */ | |
1400 | if (pfds[i].revents == 0) | |
1401 | continue; | |
1402 | ||
1403 | ready++; | |
1404 | ||
d279ef57 DS |
1405 | /* |
1406 | * Unless someone has called thread_cancel from another | |
1407 | * pthread, the only thing that could have changed in | |
1408 | * m->handler.pfds while we were asleep is the .events | |
1409 | * field in a given pollfd. Barring thread_cancel() that | |
1410 | * value should be a superset of the values we have in our | |
1411 | * copy, so there's no need to update it. Similarily, | |
1412 | * barring deletion, the fd should still be a valid index | |
1413 | * into the master's pfds. | |
d142453d DS |
1414 | * |
1415 | * We are including POLLERR here to do a READ event | |
1416 | * this is because the read should fail and the | |
1417 | * read function should handle it appropriately | |
d279ef57 | 1418 | */ |
d142453d | 1419 | if (pfds[i].revents & (POLLIN | POLLHUP | POLLERR)) { |
d62a17ae | 1420 | thread_process_io_helper(m, m->read[pfds[i].fd], POLLIN, |
45f3d590 DS |
1421 | pfds[i].revents, i); |
1422 | } | |
d62a17ae | 1423 | if (pfds[i].revents & POLLOUT) |
1424 | thread_process_io_helper(m, m->write[pfds[i].fd], | |
45f3d590 | 1425 | POLLOUT, pfds[i].revents, i); |
d62a17ae | 1426 | |
1427 | /* if one of our file descriptors is garbage, remove the same | |
1428 | * from | |
1429 | * both pfds + update sizes and index */ | |
1430 | if (pfds[i].revents & POLLNVAL) { | |
1431 | memmove(m->handler.pfds + i, m->handler.pfds + i + 1, | |
1432 | (m->handler.pfdcount - i - 1) | |
1433 | * sizeof(struct pollfd)); | |
1434 | m->handler.pfdcount--; | |
e985cda0 S |
1435 | m->handler.pfds[m->handler.pfdcount].fd = 0; |
1436 | m->handler.pfds[m->handler.pfdcount].events = 0; | |
d62a17ae | 1437 | |
1438 | memmove(pfds + i, pfds + i + 1, | |
1439 | (m->handler.copycount - i - 1) | |
1440 | * sizeof(struct pollfd)); | |
1441 | m->handler.copycount--; | |
e985cda0 S |
1442 | m->handler.copy[m->handler.copycount].fd = 0; |
1443 | m->handler.copy[m->handler.copycount].events = 0; | |
d62a17ae | 1444 | |
1445 | i--; | |
1446 | } | |
1447 | } | |
718e3744 | 1448 | } |
1449 | ||
8b70d0b0 | 1450 | /* Add all timers that have popped to the ready list. */ |
27d29ced | 1451 | static unsigned int thread_process_timers(struct thread_timer_list_head *timers, |
d62a17ae | 1452 | struct timeval *timenow) |
a48b4e6d | 1453 | { |
d62a17ae | 1454 | struct thread *thread; |
1455 | unsigned int ready = 0; | |
1456 | ||
27d29ced | 1457 | while ((thread = thread_timer_list_first(timers))) { |
d62a17ae | 1458 | if (timercmp(timenow, &thread->u.sands, <)) |
1459 | return ready; | |
27d29ced | 1460 | thread_timer_list_pop(timers); |
d62a17ae | 1461 | thread->type = THREAD_READY; |
c284542b | 1462 | thread_list_add_tail(&thread->master->ready, thread); |
d62a17ae | 1463 | ready++; |
1464 | } | |
1465 | return ready; | |
a48b4e6d | 1466 | } |
1467 | ||
2613abe6 | 1468 | /* process a list en masse, e.g. for event thread lists */ |
c284542b | 1469 | static unsigned int thread_process(struct thread_list_head *list) |
2613abe6 | 1470 | { |
d62a17ae | 1471 | struct thread *thread; |
d62a17ae | 1472 | unsigned int ready = 0; |
1473 | ||
c284542b | 1474 | while ((thread = thread_list_pop(list))) { |
d62a17ae | 1475 | thread->type = THREAD_READY; |
c284542b | 1476 | thread_list_add_tail(&thread->master->ready, thread); |
d62a17ae | 1477 | ready++; |
1478 | } | |
1479 | return ready; | |
2613abe6 PJ |
1480 | } |
1481 | ||
1482 | ||
718e3744 | 1483 | /* Fetch next ready thread. */ |
d62a17ae | 1484 | struct thread *thread_fetch(struct thread_master *m, struct thread *fetch) |
718e3744 | 1485 | { |
d62a17ae | 1486 | struct thread *thread = NULL; |
1487 | struct timeval now; | |
1488 | struct timeval zerotime = {0, 0}; | |
1489 | struct timeval tv; | |
1490 | struct timeval *tw = NULL; | |
d81ca9a3 | 1491 | bool eintr_p = false; |
d62a17ae | 1492 | int num = 0; |
1493 | ||
1494 | do { | |
1495 | /* Handle signals if any */ | |
1496 | if (m->handle_signals) | |
1497 | quagga_sigevent_process(); | |
1498 | ||
1499 | pthread_mutex_lock(&m->mtx); | |
1500 | ||
1501 | /* Process any pending cancellation requests */ | |
1502 | do_thread_cancel(m); | |
1503 | ||
e3c9529e QY |
1504 | /* |
1505 | * Attempt to flush ready queue before going into poll(). | |
1506 | * This is performance-critical. Think twice before modifying. | |
1507 | */ | |
c284542b | 1508 | if ((thread = thread_list_pop(&m->ready))) { |
e3c9529e QY |
1509 | fetch = thread_run(m, thread, fetch); |
1510 | if (fetch->ref) | |
1511 | *fetch->ref = NULL; | |
1512 | pthread_mutex_unlock(&m->mtx); | |
1513 | break; | |
1514 | } | |
1515 | ||
1516 | /* otherwise, tick through scheduling sequence */ | |
1517 | ||
bca37d17 QY |
1518 | /* |
1519 | * Post events to ready queue. This must come before the | |
1520 | * following block since events should occur immediately | |
1521 | */ | |
d62a17ae | 1522 | thread_process(&m->event); |
1523 | ||
bca37d17 QY |
1524 | /* |
1525 | * If there are no tasks on the ready queue, we will poll() | |
1526 | * until a timer expires or we receive I/O, whichever comes | |
1527 | * first. The strategy for doing this is: | |
d62a17ae | 1528 | * |
1529 | * - If there are events pending, set the poll() timeout to zero | |
1530 | * - If there are no events pending, but there are timers | |
d279ef57 DS |
1531 | * pending, set the timeout to the smallest remaining time on |
1532 | * any timer. | |
d62a17ae | 1533 | * - If there are neither timers nor events pending, but there |
d279ef57 | 1534 | * are file descriptors pending, block indefinitely in poll() |
d62a17ae | 1535 | * - If nothing is pending, it's time for the application to die |
1536 | * | |
1537 | * In every case except the last, we need to hit poll() at least | |
bca37d17 QY |
1538 | * once per loop to avoid starvation by events |
1539 | */ | |
c284542b | 1540 | if (!thread_list_count(&m->ready)) |
27d29ced | 1541 | tw = thread_timer_wait(&m->timer, &tv); |
d62a17ae | 1542 | |
c284542b DL |
1543 | if (thread_list_count(&m->ready) || |
1544 | (tw && !timercmp(tw, &zerotime, >))) | |
d62a17ae | 1545 | tw = &zerotime; |
1546 | ||
1547 | if (!tw && m->handler.pfdcount == 0) { /* die */ | |
1548 | pthread_mutex_unlock(&m->mtx); | |
1549 | fetch = NULL; | |
1550 | break; | |
1551 | } | |
1552 | ||
bca37d17 QY |
1553 | /* |
1554 | * Copy pollfd array + # active pollfds in it. Not necessary to | |
1555 | * copy the array size as this is fixed. | |
1556 | */ | |
d62a17ae | 1557 | m->handler.copycount = m->handler.pfdcount; |
1558 | memcpy(m->handler.copy, m->handler.pfds, | |
1559 | m->handler.copycount * sizeof(struct pollfd)); | |
1560 | ||
e3c9529e QY |
1561 | pthread_mutex_unlock(&m->mtx); |
1562 | { | |
d81ca9a3 MS |
1563 | eintr_p = false; |
1564 | num = fd_poll(m, tw, &eintr_p); | |
e3c9529e QY |
1565 | } |
1566 | pthread_mutex_lock(&m->mtx); | |
d764d2cc | 1567 | |
e3c9529e QY |
1568 | /* Handle any errors received in poll() */ |
1569 | if (num < 0) { | |
d81ca9a3 | 1570 | if (eintr_p) { |
d62a17ae | 1571 | pthread_mutex_unlock(&m->mtx); |
e3c9529e QY |
1572 | /* loop around to signal handler */ |
1573 | continue; | |
d62a17ae | 1574 | } |
1575 | ||
e3c9529e | 1576 | /* else die */ |
450971aa | 1577 | flog_err(EC_LIB_SYSTEM_CALL, "poll() error: %s", |
9ef9495e | 1578 | safe_strerror(errno)); |
e3c9529e QY |
1579 | pthread_mutex_unlock(&m->mtx); |
1580 | fetch = NULL; | |
1581 | break; | |
bca37d17 | 1582 | } |
d62a17ae | 1583 | |
1584 | /* Post timers to ready queue. */ | |
1585 | monotime(&now); | |
27d29ced | 1586 | thread_process_timers(&m->timer, &now); |
d62a17ae | 1587 | |
1588 | /* Post I/O to ready queue. */ | |
1589 | if (num > 0) | |
1590 | thread_process_io(m, num); | |
1591 | ||
d62a17ae | 1592 | pthread_mutex_unlock(&m->mtx); |
1593 | ||
1594 | } while (!thread && m->spin); | |
1595 | ||
1596 | return fetch; | |
718e3744 | 1597 | } |
1598 | ||
d62a17ae | 1599 | static unsigned long timeval_elapsed(struct timeval a, struct timeval b) |
62f44022 | 1600 | { |
d62a17ae | 1601 | return (((a.tv_sec - b.tv_sec) * TIMER_SECOND_MICRO) |
1602 | + (a.tv_usec - b.tv_usec)); | |
62f44022 QY |
1603 | } |
1604 | ||
d62a17ae | 1605 | unsigned long thread_consumed_time(RUSAGE_T *now, RUSAGE_T *start, |
1606 | unsigned long *cputime) | |
718e3744 | 1607 | { |
d62a17ae | 1608 | /* This is 'user + sys' time. */ |
1609 | *cputime = timeval_elapsed(now->cpu.ru_utime, start->cpu.ru_utime) | |
1610 | + timeval_elapsed(now->cpu.ru_stime, start->cpu.ru_stime); | |
1611 | return timeval_elapsed(now->real, start->real); | |
8b70d0b0 | 1612 | } |
1613 | ||
50596be0 DS |
1614 | /* We should aim to yield after yield milliseconds, which defaults |
1615 | to THREAD_YIELD_TIME_SLOT . | |
8b70d0b0 | 1616 | Note: we are using real (wall clock) time for this calculation. |
1617 | It could be argued that CPU time may make more sense in certain | |
1618 | contexts. The things to consider are whether the thread may have | |
1619 | blocked (in which case wall time increases, but CPU time does not), | |
1620 | or whether the system is heavily loaded with other processes competing | |
d62a17ae | 1621 | for CPU time. On balance, wall clock time seems to make sense. |
8b70d0b0 | 1622 | Plus it has the added benefit that gettimeofday should be faster |
1623 | than calling getrusage. */ | |
d62a17ae | 1624 | int thread_should_yield(struct thread *thread) |
718e3744 | 1625 | { |
d62a17ae | 1626 | int result; |
00dffa8c | 1627 | frr_with_mutex(&thread->mtx) { |
d62a17ae | 1628 | result = monotime_since(&thread->real, NULL) |
1629 | > (int64_t)thread->yield; | |
1630 | } | |
d62a17ae | 1631 | return result; |
50596be0 DS |
1632 | } |
1633 | ||
d62a17ae | 1634 | void thread_set_yield_time(struct thread *thread, unsigned long yield_time) |
50596be0 | 1635 | { |
00dffa8c | 1636 | frr_with_mutex(&thread->mtx) { |
d62a17ae | 1637 | thread->yield = yield_time; |
1638 | } | |
718e3744 | 1639 | } |
1640 | ||
d62a17ae | 1641 | void thread_getrusage(RUSAGE_T *r) |
db9c0df9 | 1642 | { |
231db9a6 DS |
1643 | #if defined RUSAGE_THREAD |
1644 | #define FRR_RUSAGE RUSAGE_THREAD | |
1645 | #else | |
1646 | #define FRR_RUSAGE RUSAGE_SELF | |
1647 | #endif | |
d62a17ae | 1648 | monotime(&r->real); |
f75e802d | 1649 | #ifndef EXCLUDE_CPU_TIME |
231db9a6 | 1650 | getrusage(FRR_RUSAGE, &(r->cpu)); |
f75e802d | 1651 | #endif |
db9c0df9 PJ |
1652 | } |
1653 | ||
fbcac826 QY |
1654 | /* |
1655 | * Call a thread. | |
1656 | * | |
1657 | * This function will atomically update the thread's usage history. At present | |
1658 | * this is the only spot where usage history is written. Nevertheless the code | |
1659 | * has been written such that the introduction of writers in the future should | |
1660 | * not need to update it provided the writers atomically perform only the | |
1661 | * operations done here, i.e. updating the total and maximum times. In | |
1662 | * particular, the maximum real and cpu times must be monotonically increasing | |
1663 | * or this code is not correct. | |
1664 | */ | |
d62a17ae | 1665 | void thread_call(struct thread *thread) |
718e3744 | 1666 | { |
f75e802d | 1667 | #ifndef EXCLUDE_CPU_TIME |
fbcac826 QY |
1668 | _Atomic unsigned long realtime, cputime; |
1669 | unsigned long exp; | |
1670 | unsigned long helper; | |
f75e802d | 1671 | #endif |
d62a17ae | 1672 | RUSAGE_T before, after; |
cc8b13a0 | 1673 | |
d62a17ae | 1674 | GETRUSAGE(&before); |
1675 | thread->real = before.real; | |
718e3744 | 1676 | |
6c3aa850 DL |
1677 | frrtrace(9, frr_libfrr, thread_call, thread->master, |
1678 | thread->xref->funcname, thread->xref->xref.file, | |
1679 | thread->xref->xref.line, NULL, thread->u.fd, | |
c7bb4f00 | 1680 | thread->u.val, thread->arg, thread->u.sands.tv_sec); |
abf96a87 | 1681 | |
d62a17ae | 1682 | pthread_setspecific(thread_current, thread); |
1683 | (*thread->func)(thread); | |
1684 | pthread_setspecific(thread_current, NULL); | |
718e3744 | 1685 | |
d62a17ae | 1686 | GETRUSAGE(&after); |
718e3744 | 1687 | |
f75e802d | 1688 | #ifndef EXCLUDE_CPU_TIME |
fbcac826 QY |
1689 | realtime = thread_consumed_time(&after, &before, &helper); |
1690 | cputime = helper; | |
1691 | ||
1692 | /* update realtime */ | |
1693 | atomic_fetch_add_explicit(&thread->hist->real.total, realtime, | |
1694 | memory_order_seq_cst); | |
1695 | exp = atomic_load_explicit(&thread->hist->real.max, | |
1696 | memory_order_seq_cst); | |
1697 | while (exp < realtime | |
1698 | && !atomic_compare_exchange_weak_explicit( | |
1699 | &thread->hist->real.max, &exp, realtime, | |
1700 | memory_order_seq_cst, memory_order_seq_cst)) | |
1701 | ; | |
1702 | ||
1703 | /* update cputime */ | |
1704 | atomic_fetch_add_explicit(&thread->hist->cpu.total, cputime, | |
1705 | memory_order_seq_cst); | |
1706 | exp = atomic_load_explicit(&thread->hist->cpu.max, | |
1707 | memory_order_seq_cst); | |
1708 | while (exp < cputime | |
1709 | && !atomic_compare_exchange_weak_explicit( | |
1710 | &thread->hist->cpu.max, &exp, cputime, | |
1711 | memory_order_seq_cst, memory_order_seq_cst)) | |
1712 | ; | |
1713 | ||
1714 | atomic_fetch_add_explicit(&thread->hist->total_calls, 1, | |
1715 | memory_order_seq_cst); | |
1716 | atomic_fetch_or_explicit(&thread->hist->types, 1 << thread->add_type, | |
1717 | memory_order_seq_cst); | |
718e3744 | 1718 | |
924b9229 | 1719 | #ifdef CONSUMED_TIME_CHECK |
d62a17ae | 1720 | if (realtime > CONSUMED_TIME_CHECK) { |
1721 | /* | |
1722 | * We have a CPU Hog on our hands. | |
1723 | * Whinge about it now, so we're aware this is yet another task | |
1724 | * to fix. | |
1725 | */ | |
9ef9495e | 1726 | flog_warn( |
450971aa | 1727 | EC_LIB_SLOW_THREAD, |
d62a17ae | 1728 | "SLOW THREAD: task %s (%lx) ran for %lums (cpu time %lums)", |
60a3efec | 1729 | thread->xref->funcname, (unsigned long)thread->func, |
d62a17ae | 1730 | realtime / 1000, cputime / 1000); |
1731 | } | |
924b9229 | 1732 | #endif /* CONSUMED_TIME_CHECK */ |
f75e802d | 1733 | #endif /* Exclude CPU Time */ |
718e3744 | 1734 | } |
1735 | ||
1736 | /* Execute thread */ | |
60a3efec DL |
1737 | void _thread_execute(const struct xref_threadsched *xref, |
1738 | struct thread_master *m, int (*func)(struct thread *), | |
1739 | void *arg, int val) | |
718e3744 | 1740 | { |
c4345fbf | 1741 | struct thread *thread; |
718e3744 | 1742 | |
c4345fbf | 1743 | /* Get or allocate new thread to execute. */ |
00dffa8c | 1744 | frr_with_mutex(&m->mtx) { |
60a3efec | 1745 | thread = thread_get(m, THREAD_EVENT, func, arg, xref); |
9c7753e4 | 1746 | |
c4345fbf | 1747 | /* Set its event value. */ |
00dffa8c | 1748 | frr_with_mutex(&thread->mtx) { |
c4345fbf RZ |
1749 | thread->add_type = THREAD_EXECUTE; |
1750 | thread->u.val = val; | |
1751 | thread->ref = &thread; | |
1752 | } | |
c4345fbf | 1753 | } |
f7c62e11 | 1754 | |
c4345fbf RZ |
1755 | /* Execute thread doing all accounting. */ |
1756 | thread_call(thread); | |
9c7753e4 | 1757 | |
c4345fbf RZ |
1758 | /* Give back or free thread. */ |
1759 | thread_add_unuse(m, thread); | |
718e3744 | 1760 | } |
1543c387 MS |
1761 | |
1762 | /* Debug signal mask - if 'sigs' is NULL, use current effective mask. */ | |
1763 | void debug_signals(const sigset_t *sigs) | |
1764 | { | |
1765 | int i, found; | |
1766 | sigset_t tmpsigs; | |
1767 | char buf[300]; | |
1768 | ||
1769 | /* | |
1770 | * We're only looking at the non-realtime signals here, so we need | |
1771 | * some limit value. Platform differences mean at some point we just | |
1772 | * need to pick a reasonable value. | |
1773 | */ | |
1774 | #if defined SIGRTMIN | |
1775 | # define LAST_SIGNAL SIGRTMIN | |
1776 | #else | |
1777 | # define LAST_SIGNAL 32 | |
1778 | #endif | |
1779 | ||
1780 | ||
1781 | if (sigs == NULL) { | |
1782 | sigemptyset(&tmpsigs); | |
1783 | pthread_sigmask(SIG_BLOCK, NULL, &tmpsigs); | |
1784 | sigs = &tmpsigs; | |
1785 | } | |
1786 | ||
1787 | found = 0; | |
1788 | buf[0] = '\0'; | |
1789 | ||
1790 | for (i = 0; i < LAST_SIGNAL; i++) { | |
1791 | char tmp[20]; | |
1792 | ||
1793 | if (sigismember(sigs, i) > 0) { | |
1794 | if (found > 0) | |
1795 | strlcat(buf, ",", sizeof(buf)); | |
1796 | snprintf(tmp, sizeof(tmp), "%d", i); | |
1797 | strlcat(buf, tmp, sizeof(buf)); | |
1798 | found++; | |
1799 | } | |
1800 | } | |
1801 | ||
1802 | if (found == 0) | |
1803 | snprintf(buf, sizeof(buf), "<none>"); | |
1804 | ||
1805 | zlog_debug("%s: %s", __func__, buf); | |
1806 | } |