1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* Thread management routine
3 * Copyright (C) 1998, 2000 Kunihiro Ishiguro <kunihiro@zebra.org>
9 #include <sys/resource.h>
20 #include "frratomic.h"
21 #include "frr_pthread.h"
22 #include "lib_errors.h"
23 #include "libfrr_trace.h"
26 DEFINE_MTYPE_STATIC(LIB
, THREAD
, "Thread");
27 DEFINE_MTYPE_STATIC(LIB
, EVENT_MASTER
, "Thread master");
28 DEFINE_MTYPE_STATIC(LIB
, EVENT_POLL
, "Thread Poll Info");
29 DEFINE_MTYPE_STATIC(LIB
, EVENT_STATS
, "Thread stats");
31 DECLARE_LIST(event_list
, struct event
, eventitem
);
37 struct event
**threadref
;
40 /* Flags for task cancellation */
41 #define EVENT_CANCEL_FLAG_READY 0x01
43 static int event_timer_cmp(const struct event
*a
, const struct event
*b
)
45 if (a
->u
.sands
.tv_sec
< b
->u
.sands
.tv_sec
)
47 if (a
->u
.sands
.tv_sec
> b
->u
.sands
.tv_sec
)
49 if (a
->u
.sands
.tv_usec
< b
->u
.sands
.tv_usec
)
51 if (a
->u
.sands
.tv_usec
> b
->u
.sands
.tv_usec
)
56 DECLARE_HEAP(event_timer_list
, struct event
, timeritem
, event_timer_cmp
);
58 #if defined(__APPLE__)
59 #include <mach/mach.h>
60 #include <mach/mach_time.h>
65 const unsigned char wakebyte = 0x01; \
66 write(m->io_pipe[1], &wakebyte, 1); \
69 /* control variable for initializer */
70 static pthread_once_t init_once
= PTHREAD_ONCE_INIT
;
71 pthread_key_t thread_current
;
73 static pthread_mutex_t masters_mtx
= PTHREAD_MUTEX_INITIALIZER
;
74 static struct list
*masters
;
76 static void thread_free(struct event_loop
*master
, struct event
*thread
);
78 #ifndef EXCLUDE_CPU_TIME
79 #define EXCLUDE_CPU_TIME 0
81 #ifndef CONSUMED_TIME_CHECK
82 #define CONSUMED_TIME_CHECK 0
85 bool cputime_enabled
= !EXCLUDE_CPU_TIME
;
86 unsigned long cputime_threshold
= CONSUMED_TIME_CHECK
;
87 unsigned long walltime_threshold
= CONSUMED_TIME_CHECK
;
89 /* CLI start ---------------------------------------------------------------- */
90 #include "lib/event_clippy.c"
92 static unsigned int cpu_record_hash_key(const struct cpu_event_history
*a
)
94 int size
= sizeof(a
->func
);
96 return jhash(&a
->func
, size
, 0);
99 static bool cpu_record_hash_cmp(const struct cpu_event_history
*a
,
100 const struct cpu_event_history
*b
)
102 return a
->func
== b
->func
;
105 static void *cpu_record_hash_alloc(struct cpu_event_history
*a
)
107 struct cpu_event_history
*new;
108 new = XCALLOC(MTYPE_EVENT_STATS
, sizeof(struct cpu_event_history
));
110 new->funcname
= a
->funcname
;
114 static void cpu_record_hash_free(void *a
)
116 struct cpu_event_history
*hist
= a
;
118 XFREE(MTYPE_EVENT_STATS
, hist
);
121 static void vty_out_cpu_event_history(struct vty
*vty
,
122 struct cpu_event_history
*a
)
125 "%5zu %10zu.%03zu %9zu %8zu %9zu %8zu %9zu %9zu %9zu %10zu",
126 a
->total_active
, a
->cpu
.total
/ 1000, a
->cpu
.total
% 1000,
127 a
->total_calls
, (a
->cpu
.total
/ a
->total_calls
), a
->cpu
.max
,
128 (a
->real
.total
/ a
->total_calls
), a
->real
.max
,
129 a
->total_cpu_warn
, a
->total_wall_warn
, a
->total_starv_warn
);
130 vty_out(vty
, " %c%c%c%c%c %s\n",
131 a
->types
& (1 << EVENT_READ
) ? 'R' : ' ',
132 a
->types
& (1 << EVENT_WRITE
) ? 'W' : ' ',
133 a
->types
& (1 << EVENT_TIMER
) ? 'T' : ' ',
134 a
->types
& (1 << EVENT_EVENT
) ? 'E' : ' ',
135 a
->types
& (1 << EVENT_EXECUTE
) ? 'X' : ' ', a
->funcname
);
138 static void cpu_record_hash_print(struct hash_bucket
*bucket
, void *args
[])
140 struct cpu_event_history
*totals
= args
[0];
141 struct cpu_event_history copy
;
142 struct vty
*vty
= args
[1];
143 uint8_t *filter
= args
[2];
145 struct cpu_event_history
*a
= bucket
->data
;
148 atomic_load_explicit(&a
->total_active
, memory_order_seq_cst
);
150 atomic_load_explicit(&a
->total_calls
, memory_order_seq_cst
);
151 copy
.total_cpu_warn
=
152 atomic_load_explicit(&a
->total_cpu_warn
, memory_order_seq_cst
);
153 copy
.total_wall_warn
=
154 atomic_load_explicit(&a
->total_wall_warn
, memory_order_seq_cst
);
155 copy
.total_starv_warn
= atomic_load_explicit(&a
->total_starv_warn
,
156 memory_order_seq_cst
);
158 atomic_load_explicit(&a
->cpu
.total
, memory_order_seq_cst
);
159 copy
.cpu
.max
= atomic_load_explicit(&a
->cpu
.max
, memory_order_seq_cst
);
161 atomic_load_explicit(&a
->real
.total
, memory_order_seq_cst
);
163 atomic_load_explicit(&a
->real
.max
, memory_order_seq_cst
);
164 copy
.types
= atomic_load_explicit(&a
->types
, memory_order_seq_cst
);
165 copy
.funcname
= a
->funcname
;
167 if (!(copy
.types
& *filter
))
170 vty_out_cpu_event_history(vty
, ©
);
171 totals
->total_active
+= copy
.total_active
;
172 totals
->total_calls
+= copy
.total_calls
;
173 totals
->total_cpu_warn
+= copy
.total_cpu_warn
;
174 totals
->total_wall_warn
+= copy
.total_wall_warn
;
175 totals
->total_starv_warn
+= copy
.total_starv_warn
;
176 totals
->real
.total
+= copy
.real
.total
;
177 if (totals
->real
.max
< copy
.real
.max
)
178 totals
->real
.max
= copy
.real
.max
;
179 totals
->cpu
.total
+= copy
.cpu
.total
;
180 if (totals
->cpu
.max
< copy
.cpu
.max
)
181 totals
->cpu
.max
= copy
.cpu
.max
;
184 static void cpu_record_print(struct vty
*vty
, uint8_t filter
)
186 struct cpu_event_history tmp
;
187 void *args
[3] = {&tmp
, vty
, &filter
};
188 struct event_loop
*m
;
191 if (!cputime_enabled
)
194 "Collecting CPU time statistics is currently disabled. Following statistics\n"
195 "will be zero or may display data from when collection was enabled. Use the\n"
196 " \"service cputime-stats\" command to start collecting data.\n"
197 "\nCounters and wallclock times are always maintained and should be accurate.\n");
199 memset(&tmp
, 0, sizeof(tmp
));
200 tmp
.funcname
= "TOTAL";
203 frr_with_mutex (&masters_mtx
) {
204 for (ALL_LIST_ELEMENTS_RO(masters
, ln
, m
)) {
205 const char *name
= m
->name
? m
->name
: "main";
207 char underline
[strlen(name
) + 1];
208 memset(underline
, '-', sizeof(underline
));
209 underline
[sizeof(underline
) - 1] = '\0';
212 vty_out(vty
, "Showing statistics for pthread %s\n",
214 vty_out(vty
, "-------------------------------%s\n",
216 vty_out(vty
, "%30s %18s %18s\n", "",
217 "CPU (user+system):", "Real (wall-clock):");
219 "Active Runtime(ms) Invoked Avg uSec Max uSecs");
220 vty_out(vty
, " Avg uSec Max uSecs");
222 " CPU_Warn Wall_Warn Starv_Warn Type Thread\n");
224 if (m
->cpu_record
->count
)
227 (void (*)(struct hash_bucket
*,
228 void *))cpu_record_hash_print
,
231 vty_out(vty
, "No data to display yet.\n");
238 vty_out(vty
, "Total thread statistics\n");
239 vty_out(vty
, "-------------------------\n");
240 vty_out(vty
, "%30s %18s %18s\n", "",
241 "CPU (user+system):", "Real (wall-clock):");
242 vty_out(vty
, "Active Runtime(ms) Invoked Avg uSec Max uSecs");
243 vty_out(vty
, " Avg uSec Max uSecs CPU_Warn Wall_Warn");
244 vty_out(vty
, " Type Thread\n");
246 if (tmp
.total_calls
> 0)
247 vty_out_cpu_event_history(vty
, &tmp
);
250 static void cpu_record_hash_clear(struct hash_bucket
*bucket
, void *args
[])
252 uint8_t *filter
= args
[0];
253 struct hash
*cpu_record
= args
[1];
255 struct cpu_event_history
*a
= bucket
->data
;
257 if (!(a
->types
& *filter
))
260 hash_release(cpu_record
, bucket
->data
);
263 static void cpu_record_clear(uint8_t filter
)
265 uint8_t *tmp
= &filter
;
266 struct event_loop
*m
;
269 frr_with_mutex (&masters_mtx
) {
270 for (ALL_LIST_ELEMENTS_RO(masters
, ln
, m
)) {
271 frr_with_mutex (&m
->mtx
) {
272 void *args
[2] = {tmp
, m
->cpu_record
};
275 (void (*)(struct hash_bucket
*,
276 void *))cpu_record_hash_clear
,
283 static uint8_t parse_filter(const char *filterstr
)
288 while (filterstr
[i
] != '\0') {
289 switch (filterstr
[i
]) {
292 filter
|= (1 << EVENT_READ
);
296 filter
|= (1 << EVENT_WRITE
);
300 filter
|= (1 << EVENT_TIMER
);
304 filter
|= (1 << EVENT_EVENT
);
308 filter
|= (1 << EVENT_EXECUTE
);
318 DEFUN_NOSH (show_thread_cpu
,
320 "show thread cpu [FILTER]",
322 "Thread information\n"
324 "Display filter (rwtex)\n")
326 uint8_t filter
= (uint8_t)-1U;
329 if (argv_find(argv
, argc
, "FILTER", &idx
)) {
330 filter
= parse_filter(argv
[idx
]->arg
);
333 "Invalid filter \"%s\" specified; must contain at leastone of 'RWTEXB'\n",
339 cpu_record_print(vty
, filter
);
343 DEFPY (service_cputime_stats
,
344 service_cputime_stats_cmd
,
345 "[no] service cputime-stats",
347 "Set up miscellaneous service\n"
348 "Collect CPU usage statistics\n")
350 cputime_enabled
= !no
;
354 DEFPY (service_cputime_warning
,
355 service_cputime_warning_cmd
,
356 "[no] service cputime-warning (1-4294967295)",
358 "Set up miscellaneous service\n"
359 "Warn for tasks exceeding CPU usage threshold\n"
360 "Warning threshold in milliseconds\n")
363 cputime_threshold
= 0;
365 cputime_threshold
= cputime_warning
* 1000;
369 ALIAS (service_cputime_warning
,
370 no_service_cputime_warning_cmd
,
371 "no service cputime-warning",
373 "Set up miscellaneous service\n"
374 "Warn for tasks exceeding CPU usage threshold\n")
376 DEFPY (service_walltime_warning
,
377 service_walltime_warning_cmd
,
378 "[no] service walltime-warning (1-4294967295)",
380 "Set up miscellaneous service\n"
381 "Warn for tasks exceeding total wallclock threshold\n"
382 "Warning threshold in milliseconds\n")
385 walltime_threshold
= 0;
387 walltime_threshold
= walltime_warning
* 1000;
391 ALIAS (service_walltime_warning
,
392 no_service_walltime_warning_cmd
,
393 "no service walltime-warning",
395 "Set up miscellaneous service\n"
396 "Warn for tasks exceeding total wallclock threshold\n")
398 static void show_thread_poll_helper(struct vty
*vty
, struct event_loop
*m
)
400 const char *name
= m
->name
? m
->name
: "main";
401 char underline
[strlen(name
) + 1];
402 struct event
*thread
;
405 memset(underline
, '-', sizeof(underline
));
406 underline
[sizeof(underline
) - 1] = '\0';
408 vty_out(vty
, "\nShowing poll FD's for %s\n", name
);
409 vty_out(vty
, "----------------------%s\n", underline
);
410 vty_out(vty
, "Count: %u/%d\n", (uint32_t)m
->handler
.pfdcount
,
412 for (i
= 0; i
< m
->handler
.pfdcount
; i
++) {
413 vty_out(vty
, "\t%6d fd:%6d events:%2d revents:%2d\t\t", i
,
414 m
->handler
.pfds
[i
].fd
, m
->handler
.pfds
[i
].events
,
415 m
->handler
.pfds
[i
].revents
);
417 if (m
->handler
.pfds
[i
].events
& POLLIN
) {
418 thread
= m
->read
[m
->handler
.pfds
[i
].fd
];
421 vty_out(vty
, "ERROR ");
423 vty_out(vty
, "%s ", thread
->xref
->funcname
);
427 if (m
->handler
.pfds
[i
].events
& POLLOUT
) {
428 thread
= m
->write
[m
->handler
.pfds
[i
].fd
];
431 vty_out(vty
, "ERROR\n");
433 vty_out(vty
, "%s\n", thread
->xref
->funcname
);
439 DEFUN_NOSH (show_thread_poll
,
440 show_thread_poll_cmd
,
443 "Thread information\n"
444 "Show poll FD's and information\n")
446 struct listnode
*node
;
447 struct event_loop
*m
;
449 frr_with_mutex (&masters_mtx
) {
450 for (ALL_LIST_ELEMENTS_RO(masters
, node
, m
)) {
451 show_thread_poll_helper(vty
, m
);
459 DEFUN (clear_thread_cpu
,
460 clear_thread_cpu_cmd
,
461 "clear thread cpu [FILTER]",
462 "Clear stored data in all pthreads\n"
463 "Thread information\n"
465 "Display filter (rwtexb)\n")
467 uint8_t filter
= (uint8_t)-1U;
470 if (argv_find(argv
, argc
, "FILTER", &idx
)) {
471 filter
= parse_filter(argv
[idx
]->arg
);
474 "Invalid filter \"%s\" specified; must contain at leastone of 'RWTEXB'\n",
480 cpu_record_clear(filter
);
484 static void show_thread_timers_helper(struct vty
*vty
, struct event_loop
*m
)
486 const char *name
= m
->name
? m
->name
: "main";
487 char underline
[strlen(name
) + 1];
488 struct event
*thread
;
490 memset(underline
, '-', sizeof(underline
));
491 underline
[sizeof(underline
) - 1] = '\0';
493 vty_out(vty
, "\nShowing timers for %s\n", name
);
494 vty_out(vty
, "-------------------%s\n", underline
);
496 frr_each (event_timer_list
, &m
->timer
, thread
) {
497 vty_out(vty
, " %-50s%pTH\n", thread
->hist
->funcname
, thread
);
501 DEFPY_NOSH (show_thread_timers
,
502 show_thread_timers_cmd
,
503 "show thread timers",
505 "Thread information\n"
506 "Show all timers and how long they have in the system\n")
508 struct listnode
*node
;
509 struct event_loop
*m
;
511 frr_with_mutex (&masters_mtx
) {
512 for (ALL_LIST_ELEMENTS_RO(masters
, node
, m
))
513 show_thread_timers_helper(vty
, m
);
519 void event_cmd_init(void)
521 install_element(VIEW_NODE
, &show_thread_cpu_cmd
);
522 install_element(VIEW_NODE
, &show_thread_poll_cmd
);
523 install_element(ENABLE_NODE
, &clear_thread_cpu_cmd
);
525 install_element(CONFIG_NODE
, &service_cputime_stats_cmd
);
526 install_element(CONFIG_NODE
, &service_cputime_warning_cmd
);
527 install_element(CONFIG_NODE
, &no_service_cputime_warning_cmd
);
528 install_element(CONFIG_NODE
, &service_walltime_warning_cmd
);
529 install_element(CONFIG_NODE
, &no_service_walltime_warning_cmd
);
531 install_element(VIEW_NODE
, &show_thread_timers_cmd
);
533 /* CLI end ------------------------------------------------------------------ */
536 static void cancelreq_del(void *cr
)
538 XFREE(MTYPE_TMP
, cr
);
541 /* initializer, only ever called once */
542 static void initializer(void)
544 pthread_key_create(&thread_current
, NULL
);
547 struct event_loop
*event_master_create(const char *name
)
549 struct event_loop
*rv
;
552 pthread_once(&init_once
, &initializer
);
554 rv
= XCALLOC(MTYPE_EVENT_MASTER
, sizeof(struct event_loop
));
556 /* Initialize master mutex */
557 pthread_mutex_init(&rv
->mtx
, NULL
);
558 pthread_cond_init(&rv
->cancel_cond
, NULL
);
561 name
= name
? name
: "default";
562 rv
->name
= XSTRDUP(MTYPE_EVENT_MASTER
, name
);
564 /* Initialize I/O task data structures */
566 /* Use configured limit if present, ulimit otherwise. */
567 rv
->fd_limit
= frr_get_fd_limit();
568 if (rv
->fd_limit
== 0) {
569 getrlimit(RLIMIT_NOFILE
, &limit
);
570 rv
->fd_limit
= (int)limit
.rlim_cur
;
573 rv
->read
= XCALLOC(MTYPE_EVENT_POLL
,
574 sizeof(struct event
*) * rv
->fd_limit
);
576 rv
->write
= XCALLOC(MTYPE_EVENT_POLL
,
577 sizeof(struct event
*) * rv
->fd_limit
);
579 char tmhashname
[strlen(name
) + 32];
580 snprintf(tmhashname
, sizeof(tmhashname
), "%s - threadmaster event hash",
582 rv
->cpu_record
= hash_create_size(
583 8, (unsigned int (*)(const void *))cpu_record_hash_key
,
584 (bool (*)(const void *, const void *))cpu_record_hash_cmp
,
587 event_list_init(&rv
->event
);
588 event_list_init(&rv
->ready
);
589 event_list_init(&rv
->unuse
);
590 event_timer_list_init(&rv
->timer
);
592 /* Initialize event_fetch() settings */
594 rv
->handle_signals
= true;
596 /* Set pthread owner, should be updated by actual owner */
597 rv
->owner
= pthread_self();
598 rv
->cancel_req
= list_new();
599 rv
->cancel_req
->del
= cancelreq_del
;
602 /* Initialize pipe poker */
604 set_nonblocking(rv
->io_pipe
[0]);
605 set_nonblocking(rv
->io_pipe
[1]);
607 /* Initialize data structures for poll() */
608 rv
->handler
.pfdsize
= rv
->fd_limit
;
609 rv
->handler
.pfdcount
= 0;
610 rv
->handler
.pfds
= XCALLOC(MTYPE_EVENT_MASTER
,
611 sizeof(struct pollfd
) * rv
->handler
.pfdsize
);
612 rv
->handler
.copy
= XCALLOC(MTYPE_EVENT_MASTER
,
613 sizeof(struct pollfd
) * rv
->handler
.pfdsize
);
615 /* add to list of threadmasters */
616 frr_with_mutex (&masters_mtx
) {
618 masters
= list_new();
620 listnode_add(masters
, rv
);
626 void event_master_set_name(struct event_loop
*master
, const char *name
)
628 frr_with_mutex (&master
->mtx
) {
629 XFREE(MTYPE_EVENT_MASTER
, master
->name
);
630 master
->name
= XSTRDUP(MTYPE_EVENT_MASTER
, name
);
634 #define EVENT_UNUSED_DEPTH 10
636 /* Move thread to unuse list. */
637 static void thread_add_unuse(struct event_loop
*m
, struct event
*thread
)
639 pthread_mutex_t mtxc
= thread
->mtx
;
641 assert(m
!= NULL
&& thread
!= NULL
);
643 thread
->hist
->total_active
--;
644 memset(thread
, 0, sizeof(struct event
));
645 thread
->type
= EVENT_UNUSED
;
647 /* Restore the thread mutex context. */
650 if (event_list_count(&m
->unuse
) < EVENT_UNUSED_DEPTH
) {
651 event_list_add_tail(&m
->unuse
, thread
);
655 thread_free(m
, thread
);
658 /* Free all unused thread. */
659 static void thread_list_free(struct event_loop
*m
, struct event_list_head
*list
)
663 while ((t
= event_list_pop(list
)))
667 static void thread_array_free(struct event_loop
*m
, struct event
**thread_array
)
672 for (index
= 0; index
< m
->fd_limit
; ++index
) {
673 t
= thread_array
[index
];
675 thread_array
[index
] = NULL
;
679 XFREE(MTYPE_EVENT_POLL
, thread_array
);
683 * event_master_free_unused
685 * As threads are finished with they are put on the
686 * unuse list for later reuse.
687 * If we are shutting down, Free up unused threads
688 * So we can see if we forget to shut anything off
690 void event_master_free_unused(struct event_loop
*m
)
692 frr_with_mutex (&m
->mtx
) {
694 while ((t
= event_list_pop(&m
->unuse
)))
699 /* Stop thread scheduler. */
700 void event_master_free(struct event_loop
*m
)
704 frr_with_mutex (&masters_mtx
) {
705 listnode_delete(masters
, m
);
706 if (masters
->count
== 0) {
707 list_delete(&masters
);
711 thread_array_free(m
, m
->read
);
712 thread_array_free(m
, m
->write
);
713 while ((t
= event_timer_list_pop(&m
->timer
)))
715 thread_list_free(m
, &m
->event
);
716 thread_list_free(m
, &m
->ready
);
717 thread_list_free(m
, &m
->unuse
);
718 pthread_mutex_destroy(&m
->mtx
);
719 pthread_cond_destroy(&m
->cancel_cond
);
720 close(m
->io_pipe
[0]);
721 close(m
->io_pipe
[1]);
722 list_delete(&m
->cancel_req
);
723 m
->cancel_req
= NULL
;
725 hash_clean_and_free(&m
->cpu_record
, cpu_record_hash_free
);
727 XFREE(MTYPE_EVENT_MASTER
, m
->name
);
728 XFREE(MTYPE_EVENT_MASTER
, m
->handler
.pfds
);
729 XFREE(MTYPE_EVENT_MASTER
, m
->handler
.copy
);
730 XFREE(MTYPE_EVENT_MASTER
, m
);
733 /* Return remain time in milliseconds. */
734 unsigned long event_timer_remain_msec(struct event
*thread
)
738 if (!event_is_scheduled(thread
))
741 frr_with_mutex (&thread
->mtx
) {
742 remain
= monotime_until(&thread
->u
.sands
, NULL
) / 1000LL;
745 return remain
< 0 ? 0 : remain
;
748 /* Return remain time in seconds. */
749 unsigned long event_timer_remain_second(struct event
*thread
)
751 return event_timer_remain_msec(thread
) / 1000LL;
754 struct timeval
event_timer_remain(struct event
*thread
)
756 struct timeval remain
;
757 frr_with_mutex (&thread
->mtx
) {
758 monotime_until(&thread
->u
.sands
, &remain
);
763 static int time_hhmmss(char *buf
, int buf_size
, long sec
)
769 assert(buf_size
>= 8);
776 wr
= snprintf(buf
, buf_size
, "%02ld:%02ld:%02ld", hh
, mm
, sec
);
781 char *event_timer_to_hhmmss(char *buf
, int buf_size
, struct event
*t_timer
)
784 time_hhmmss(buf
, buf_size
, event_timer_remain_second(t_timer
));
786 snprintf(buf
, buf_size
, "--:--:--");
791 /* Get new thread. */
792 static struct event
*thread_get(struct event_loop
*m
, uint8_t type
,
793 void (*func
)(struct event
*), void *arg
,
794 const struct xref_eventsched
*xref
)
796 struct event
*thread
= event_list_pop(&m
->unuse
);
797 struct cpu_event_history tmp
;
800 thread
= XCALLOC(MTYPE_THREAD
, sizeof(struct event
));
801 /* mutex only needs to be initialized at struct creation. */
802 pthread_mutex_init(&thread
->mtx
, NULL
);
807 thread
->add_type
= type
;
810 thread
->yield
= EVENT_YIELD_TIME_SLOT
; /* default */
812 thread
->ignore_timer_late
= false;
815 * So if the passed in funcname is not what we have
816 * stored that means the thread->hist needs to be
817 * updated. We keep the last one around in unused
818 * under the assumption that we are probably
819 * going to immediately allocate the same
821 * This hopefully saves us some serious
824 if ((thread
->xref
&& thread
->xref
->funcname
!= xref
->funcname
)
825 || thread
->func
!= func
) {
827 tmp
.funcname
= xref
->funcname
;
829 hash_get(m
->cpu_record
, &tmp
,
830 (void *(*)(void *))cpu_record_hash_alloc
);
832 thread
->hist
->total_active
++;
839 static void thread_free(struct event_loop
*master
, struct event
*thread
)
841 /* Update statistics. */
842 assert(master
->alloc
> 0);
845 /* Free allocated resources. */
846 pthread_mutex_destroy(&thread
->mtx
);
847 XFREE(MTYPE_THREAD
, thread
);
850 static int fd_poll(struct event_loop
*m
, const struct timeval
*timer_wait
,
854 unsigned char trash
[64];
855 nfds_t count
= m
->handler
.copycount
;
858 * If timer_wait is null here, that means poll() should block
859 * indefinitely, unless the event_master has overridden it by setting
860 * ->selectpoll_timeout.
862 * If the value is positive, it specifies the maximum number of
863 * milliseconds to wait. If the timeout is -1, it specifies that
864 * we should never wait and always return immediately even if no
865 * event is detected. If the value is zero, the behavior is default.
869 /* number of file descriptors with events */
872 if (timer_wait
!= NULL
873 && m
->selectpoll_timeout
== 0) // use the default value
874 timeout
= (timer_wait
->tv_sec
* 1000)
875 + (timer_wait
->tv_usec
/ 1000);
876 else if (m
->selectpoll_timeout
> 0) // use the user's timeout
877 timeout
= m
->selectpoll_timeout
;
878 else if (m
->selectpoll_timeout
879 < 0) // effect a poll (return immediately)
882 zlog_tls_buffer_flush();
884 rcu_assert_read_unlocked();
886 /* add poll pipe poker */
887 assert(count
+ 1 < m
->handler
.pfdsize
);
888 m
->handler
.copy
[count
].fd
= m
->io_pipe
[0];
889 m
->handler
.copy
[count
].events
= POLLIN
;
890 m
->handler
.copy
[count
].revents
= 0x00;
892 /* We need to deal with a signal-handling race here: we
893 * don't want to miss a crucial signal, such as SIGTERM or SIGINT,
894 * that may arrive just before we enter poll(). We will block the
895 * key signals, then check whether any have arrived - if so, we return
896 * before calling poll(). If not, we'll re-enable the signals
897 * in the ppoll() call.
900 sigemptyset(&origsigs
);
901 if (m
->handle_signals
) {
902 /* Main pthread that handles the app signals */
903 if (frr_sigevent_check(&origsigs
)) {
904 /* Signal to process - restore signal mask and return */
905 pthread_sigmask(SIG_SETMASK
, &origsigs
, NULL
);
911 /* Don't make any changes for the non-main pthreads */
912 pthread_sigmask(SIG_SETMASK
, NULL
, &origsigs
);
915 #if defined(HAVE_PPOLL)
916 struct timespec ts
, *tsp
;
919 ts
.tv_sec
= timeout
/ 1000;
920 ts
.tv_nsec
= (timeout
% 1000) * 1000000;
925 num
= ppoll(m
->handler
.copy
, count
+ 1, tsp
, &origsigs
);
926 pthread_sigmask(SIG_SETMASK
, &origsigs
, NULL
);
928 /* Not ideal - there is a race after we restore the signal mask */
929 pthread_sigmask(SIG_SETMASK
, &origsigs
, NULL
);
930 num
= poll(m
->handler
.copy
, count
+ 1, timeout
);
935 if (num
< 0 && errno
== EINTR
)
938 if (num
> 0 && m
->handler
.copy
[count
].revents
!= 0 && num
--)
939 while (read(m
->io_pipe
[0], &trash
, sizeof(trash
)) > 0)
947 /* Add new read thread. */
948 void _event_add_read_write(const struct xref_eventsched
*xref
,
949 struct event_loop
*m
, void (*func
)(struct event
*),
950 void *arg
, int fd
, struct event
**t_ptr
)
952 int dir
= xref
->event_type
;
953 struct event
*thread
= NULL
;
954 struct event
**thread_array
;
956 if (dir
== EVENT_READ
)
957 frrtrace(9, frr_libfrr
, schedule_read
, m
,
958 xref
->funcname
, xref
->xref
.file
, xref
->xref
.line
,
959 t_ptr
, fd
, 0, arg
, 0);
961 frrtrace(9, frr_libfrr
, schedule_write
, m
,
962 xref
->funcname
, xref
->xref
.file
, xref
->xref
.line
,
963 t_ptr
, fd
, 0, arg
, 0);
966 if (fd
>= m
->fd_limit
)
967 assert(!"Number of FD's open is greater than FRR currently configured to handle, aborting");
969 frr_with_mutex (&m
->mtx
) {
971 // thread is already scheduled; don't reschedule
974 /* default to a new pollfd */
975 nfds_t queuepos
= m
->handler
.pfdcount
;
977 if (dir
== EVENT_READ
)
978 thread_array
= m
->read
;
980 thread_array
= m
->write
;
982 /* if we already have a pollfd for our file descriptor, find and
984 for (nfds_t i
= 0; i
< m
->handler
.pfdcount
; i
++)
985 if (m
->handler
.pfds
[i
].fd
== fd
) {
990 * What happens if we have a thread already
991 * created for this event?
993 if (thread_array
[fd
])
994 assert(!"Thread already scheduled for file descriptor");
999 /* make sure we have room for this fd + pipe poker fd */
1000 assert(queuepos
+ 1 < m
->handler
.pfdsize
);
1002 thread
= thread_get(m
, dir
, func
, arg
, xref
);
1004 m
->handler
.pfds
[queuepos
].fd
= fd
;
1005 m
->handler
.pfds
[queuepos
].events
|=
1006 (dir
== EVENT_READ
? POLLIN
: POLLOUT
);
1008 if (queuepos
== m
->handler
.pfdcount
)
1009 m
->handler
.pfdcount
++;
1012 frr_with_mutex (&thread
->mtx
) {
1014 thread_array
[thread
->u
.fd
] = thread
;
1019 thread
->ref
= t_ptr
;
1027 static void _event_add_timer_timeval(const struct xref_eventsched
*xref
,
1028 struct event_loop
*m
,
1029 void (*func
)(struct event
*), void *arg
,
1030 struct timeval
*time_relative
,
1031 struct event
**t_ptr
)
1033 struct event
*thread
;
1038 assert(time_relative
);
1040 frrtrace(9, frr_libfrr
, schedule_timer
, m
,
1041 xref
->funcname
, xref
->xref
.file
, xref
->xref
.line
,
1042 t_ptr
, 0, 0, arg
, (long)time_relative
->tv_sec
);
1044 /* Compute expiration/deadline time. */
1046 timeradd(&t
, time_relative
, &t
);
1048 frr_with_mutex (&m
->mtx
) {
1049 if (t_ptr
&& *t_ptr
)
1050 /* thread is already scheduled; don't reschedule */
1053 thread
= thread_get(m
, EVENT_TIMER
, func
, arg
, xref
);
1055 frr_with_mutex (&thread
->mtx
) {
1056 thread
->u
.sands
= t
;
1057 event_timer_list_add(&m
->timer
, thread
);
1060 thread
->ref
= t_ptr
;
1064 /* The timer list is sorted - if this new timer
1065 * might change the time we'll wait for, give the pthread
1066 * a chance to re-compute.
1068 if (event_timer_list_first(&m
->timer
) == thread
)
1071 #define ONEYEAR2SEC (60 * 60 * 24 * 365)
1072 if (time_relative
->tv_sec
> ONEYEAR2SEC
)
1074 EC_LIB_TIMER_TOO_LONG
,
1075 "Timer: %pTHD is created with an expiration that is greater than 1 year",
1080 /* Add timer event thread. */
1081 void _event_add_timer(const struct xref_eventsched
*xref
, struct event_loop
*m
,
1082 void (*func
)(struct event
*), void *arg
, long timer
,
1083 struct event
**t_ptr
)
1085 struct timeval trel
;
1089 trel
.tv_sec
= timer
;
1092 _event_add_timer_timeval(xref
, m
, func
, arg
, &trel
, t_ptr
);
1095 /* Add timer event thread with "millisecond" resolution */
1096 void _event_add_timer_msec(const struct xref_eventsched
*xref
,
1097 struct event_loop
*m
, void (*func
)(struct event
*),
1098 void *arg
, long timer
, struct event
**t_ptr
)
1100 struct timeval trel
;
1104 trel
.tv_sec
= timer
/ 1000;
1105 trel
.tv_usec
= 1000 * (timer
% 1000);
1107 _event_add_timer_timeval(xref
, m
, func
, arg
, &trel
, t_ptr
);
1110 /* Add timer event thread with "timeval" resolution */
1111 void _event_add_timer_tv(const struct xref_eventsched
*xref
,
1112 struct event_loop
*m
, void (*func
)(struct event
*),
1113 void *arg
, struct timeval
*tv
, struct event
**t_ptr
)
1115 _event_add_timer_timeval(xref
, m
, func
, arg
, tv
, t_ptr
);
1118 /* Add simple event thread. */
1119 void _event_add_event(const struct xref_eventsched
*xref
, struct event_loop
*m
,
1120 void (*func
)(struct event
*), void *arg
, int val
,
1121 struct event
**t_ptr
)
1123 struct event
*thread
= NULL
;
1125 frrtrace(9, frr_libfrr
, schedule_event
, m
,
1126 xref
->funcname
, xref
->xref
.file
, xref
->xref
.line
,
1127 t_ptr
, 0, val
, arg
, 0);
1131 frr_with_mutex (&m
->mtx
) {
1132 if (t_ptr
&& *t_ptr
)
1133 /* thread is already scheduled; don't reschedule */
1136 thread
= thread_get(m
, EVENT_EVENT
, func
, arg
, xref
);
1137 frr_with_mutex (&thread
->mtx
) {
1138 thread
->u
.val
= val
;
1139 event_list_add_tail(&m
->event
, thread
);
1144 thread
->ref
= t_ptr
;
1151 /* Thread cancellation ------------------------------------------------------ */
1154 * NOT's out the .events field of pollfd corresponding to the given file
1155 * descriptor. The event to be NOT'd is passed in the 'state' parameter.
1157 * This needs to happen for both copies of pollfd's. See 'event_fetch'
1158 * implementation for details.
1162 * @param state the event to cancel. One or more (OR'd together) of the
1167 static void event_cancel_rw(struct event_loop
*master
, int fd
, short state
,
1172 /* find the index of corresponding pollfd */
1175 /* Cancel POLLHUP too just in case some bozo set it */
1178 /* Some callers know the index of the pfd already */
1179 if (idx_hint
>= 0) {
1183 /* Have to look for the fd in the pfd array */
1184 for (i
= 0; i
< master
->handler
.pfdcount
; i
++)
1185 if (master
->handler
.pfds
[i
].fd
== fd
) {
1193 "[!] Received cancellation request for nonexistent rw job");
1194 zlog_debug("[!] threadmaster: %s | fd: %d",
1195 master
->name
? master
->name
: "", fd
);
1199 /* NOT out event. */
1200 master
->handler
.pfds
[i
].events
&= ~(state
);
1202 /* If all events are canceled, delete / resize the pollfd array. */
1203 if (master
->handler
.pfds
[i
].events
== 0) {
1204 memmove(master
->handler
.pfds
+ i
, master
->handler
.pfds
+ i
+ 1,
1205 (master
->handler
.pfdcount
- i
- 1)
1206 * sizeof(struct pollfd
));
1207 master
->handler
.pfdcount
--;
1208 master
->handler
.pfds
[master
->handler
.pfdcount
].fd
= 0;
1209 master
->handler
.pfds
[master
->handler
.pfdcount
].events
= 0;
1212 /* If we have the same pollfd in the copy, perform the same operations,
1213 * otherwise return. */
1214 if (i
>= master
->handler
.copycount
)
1217 master
->handler
.copy
[i
].events
&= ~(state
);
1219 if (master
->handler
.copy
[i
].events
== 0) {
1220 memmove(master
->handler
.copy
+ i
, master
->handler
.copy
+ i
+ 1,
1221 (master
->handler
.copycount
- i
- 1)
1222 * sizeof(struct pollfd
));
1223 master
->handler
.copycount
--;
1224 master
->handler
.copy
[master
->handler
.copycount
].fd
= 0;
1225 master
->handler
.copy
[master
->handler
.copycount
].events
= 0;
1230 * Process task cancellation given a task argument: iterate through the
1231 * various lists of tasks, looking for any that match the argument.
1233 static void cancel_arg_helper(struct event_loop
*master
,
1234 const struct cancel_req
*cr
)
1241 /* We're only processing arg-based cancellations here. */
1242 if (cr
->eventobj
== NULL
)
1245 /* First process the ready lists. */
1246 frr_each_safe (event_list
, &master
->event
, t
) {
1247 if (t
->arg
!= cr
->eventobj
)
1249 event_list_del(&master
->event
, t
);
1252 thread_add_unuse(master
, t
);
1255 frr_each_safe (event_list
, &master
->ready
, t
) {
1256 if (t
->arg
!= cr
->eventobj
)
1258 event_list_del(&master
->ready
, t
);
1261 thread_add_unuse(master
, t
);
1264 /* If requested, stop here and ignore io and timers */
1265 if (CHECK_FLAG(cr
->flags
, EVENT_CANCEL_FLAG_READY
))
1268 /* Check the io tasks */
1269 for (i
= 0; i
< master
->handler
.pfdcount
;) {
1270 pfd
= master
->handler
.pfds
+ i
;
1272 if (pfd
->events
& POLLIN
)
1273 t
= master
->read
[pfd
->fd
];
1275 t
= master
->write
[pfd
->fd
];
1277 if (t
&& t
->arg
== cr
->eventobj
) {
1280 /* Found a match to cancel: clean up fd arrays */
1281 event_cancel_rw(master
, pfd
->fd
, pfd
->events
, i
);
1283 /* Clean up thread arrays */
1284 master
->read
[fd
] = NULL
;
1285 master
->write
[fd
] = NULL
;
1287 /* Clear caller's ref */
1291 thread_add_unuse(master
, t
);
1293 /* Don't increment 'i' since the cancellation will have
1294 * removed the entry from the pfd array
1300 /* Check the timer tasks */
1301 t
= event_timer_list_first(&master
->timer
);
1303 struct event
*t_next
;
1305 t_next
= event_timer_list_next(&master
->timer
, t
);
1307 if (t
->arg
== cr
->eventobj
) {
1308 event_timer_list_del(&master
->timer
, t
);
1311 thread_add_unuse(master
, t
);
1319 * Process cancellation requests.
1321 * This may only be run from the pthread which owns the event_master.
1323 * @param master the thread master to process
1324 * @REQUIRE master->mtx
1326 static void do_event_cancel(struct event_loop
*master
)
1328 struct event_list_head
*list
= NULL
;
1329 struct event
**thread_array
= NULL
;
1330 struct event
*thread
;
1331 struct cancel_req
*cr
;
1332 struct listnode
*ln
;
1334 for (ALL_LIST_ELEMENTS_RO(master
->cancel_req
, ln
, cr
)) {
1336 * If this is an event object cancellation, search
1337 * through task lists deleting any tasks which have the
1338 * specified argument - use this handy helper function.
1341 cancel_arg_helper(master
, cr
);
1346 * The pointer varies depending on whether the cancellation
1347 * request was made asynchronously or not. If it was, we
1348 * need to check whether the thread even exists anymore
1349 * before cancelling it.
1351 thread
= (cr
->thread
) ? cr
->thread
: *cr
->threadref
;
1357 thread_array
= NULL
;
1359 /* Determine the appropriate queue to cancel the thread from */
1360 switch (thread
->type
) {
1362 event_cancel_rw(master
, thread
->u
.fd
, POLLIN
, -1);
1363 thread_array
= master
->read
;
1366 event_cancel_rw(master
, thread
->u
.fd
, POLLOUT
, -1);
1367 thread_array
= master
->write
;
1370 event_timer_list_del(&master
->timer
, thread
);
1373 list
= &master
->event
;
1376 list
= &master
->ready
;
1385 event_list_del(list
, thread
);
1386 } else if (thread_array
) {
1387 thread_array
[thread
->u
.fd
] = NULL
;
1391 *thread
->ref
= NULL
;
1393 thread_add_unuse(thread
->master
, thread
);
1396 /* Delete and free all cancellation requests */
1397 if (master
->cancel_req
)
1398 list_delete_all_node(master
->cancel_req
);
1400 /* Wake up any threads which may be blocked in event_cancel_async() */
1401 master
->canceled
= true;
1402 pthread_cond_broadcast(&master
->cancel_cond
);
1406 * Helper function used for multiple flavors of arg-based cancellation.
1408 static void cancel_event_helper(struct event_loop
*m
, void *arg
, int flags
)
1410 struct cancel_req
*cr
;
1412 assert(m
->owner
== pthread_self());
1414 /* Only worth anything if caller supplies an arg. */
1418 cr
= XCALLOC(MTYPE_TMP
, sizeof(struct cancel_req
));
1422 frr_with_mutex (&m
->mtx
) {
1424 listnode_add(m
->cancel_req
, cr
);
1430 * Cancel any events which have the specified argument.
1434 * @param m the event_master to cancel from
1435 * @param arg the argument passed when creating the event
1437 void event_cancel_event(struct event_loop
*master
, void *arg
)
1439 cancel_event_helper(master
, arg
, 0);
1443 * Cancel ready tasks with an arg matching 'arg'
1447 * @param m the event_master to cancel from
1448 * @param arg the argument passed when creating the event
1450 void event_cancel_event_ready(struct event_loop
*m
, void *arg
)
1453 /* Only cancel ready/event tasks */
1454 cancel_event_helper(m
, arg
, EVENT_CANCEL_FLAG_READY
);
1458 * Cancel a specific task.
1462 * @param thread task to cancel
1464 void event_cancel(struct event
**thread
)
1466 struct event_loop
*master
;
1468 if (thread
== NULL
|| *thread
== NULL
)
1471 master
= (*thread
)->master
;
1473 frrtrace(9, frr_libfrr
, event_cancel
, master
, (*thread
)->xref
->funcname
,
1474 (*thread
)->xref
->xref
.file
, (*thread
)->xref
->xref
.line
, NULL
,
1475 (*thread
)->u
.fd
, (*thread
)->u
.val
, (*thread
)->arg
,
1476 (*thread
)->u
.sands
.tv_sec
);
1478 assert(master
->owner
== pthread_self());
1480 frr_with_mutex (&master
->mtx
) {
1481 struct cancel_req
*cr
=
1482 XCALLOC(MTYPE_TMP
, sizeof(struct cancel_req
));
1483 cr
->thread
= *thread
;
1484 listnode_add(master
->cancel_req
, cr
);
1485 do_event_cancel(master
);
1492 * Asynchronous cancellation.
1494 * Called with either a struct event ** or void * to an event argument,
1495 * this function posts the correct cancellation request and blocks until it is
1498 * If the thread is currently running, execution blocks until it completes.
1500 * The last two parameters are mutually exclusive, i.e. if you pass one the
1501 * other must be NULL.
1503 * When the cancellation procedure executes on the target event_master, the
1504 * thread * provided is checked for nullity. If it is null, the thread is
1505 * assumed to no longer exist and the cancellation request is a no-op. Thus
1506 * users of this API must pass a back-reference when scheduling the original
1511 * @param master the thread master with the relevant event / task
1512 * @param thread pointer to thread to cancel
1513 * @param eventobj the event
1515 void event_cancel_async(struct event_loop
*master
, struct event
**thread
,
1518 assert(!(thread
&& eventobj
) && (thread
|| eventobj
));
1520 if (thread
&& *thread
)
1521 frrtrace(9, frr_libfrr
, event_cancel_async
, master
,
1522 (*thread
)->xref
->funcname
, (*thread
)->xref
->xref
.file
,
1523 (*thread
)->xref
->xref
.line
, NULL
, (*thread
)->u
.fd
,
1524 (*thread
)->u
.val
, (*thread
)->arg
,
1525 (*thread
)->u
.sands
.tv_sec
);
1527 frrtrace(9, frr_libfrr
, event_cancel_async
, master
, NULL
, NULL
,
1528 0, NULL
, 0, 0, eventobj
, 0);
1530 assert(master
->owner
!= pthread_self());
1532 frr_with_mutex (&master
->mtx
) {
1533 master
->canceled
= false;
1536 struct cancel_req
*cr
=
1537 XCALLOC(MTYPE_TMP
, sizeof(struct cancel_req
));
1538 cr
->threadref
= thread
;
1539 listnode_add(master
->cancel_req
, cr
);
1540 } else if (eventobj
) {
1541 struct cancel_req
*cr
=
1542 XCALLOC(MTYPE_TMP
, sizeof(struct cancel_req
));
1543 cr
->eventobj
= eventobj
;
1544 listnode_add(master
->cancel_req
, cr
);
1548 while (!master
->canceled
)
1549 pthread_cond_wait(&master
->cancel_cond
, &master
->mtx
);
1555 /* ------------------------------------------------------------------------- */
1557 static struct timeval
*thread_timer_wait(struct event_timer_list_head
*timers
,
1558 struct timeval
*timer_val
)
1560 if (!event_timer_list_count(timers
))
1563 struct event
*next_timer
= event_timer_list_first(timers
);
1564 monotime_until(&next_timer
->u
.sands
, timer_val
);
1568 static struct event
*thread_run(struct event_loop
*m
, struct event
*thread
,
1569 struct event
*fetch
)
1572 thread_add_unuse(m
, thread
);
1576 static int thread_process_io_helper(struct event_loop
*m
, struct event
*thread
,
1577 short state
, short actual_state
, int pos
)
1579 struct event
**thread_array
;
1582 * poll() clears the .events field, but the pollfd array we
1583 * pass to poll() is a copy of the one used to schedule threads.
1584 * We need to synchronize state between the two here by applying
1585 * the same changes poll() made on the copy of the "real" pollfd
1588 * This cleans up a possible infinite loop where we refuse
1589 * to respond to a poll event but poll is insistent that
1592 m
->handler
.pfds
[pos
].events
&= ~(state
);
1595 if ((actual_state
& (POLLHUP
|POLLIN
)) != POLLHUP
)
1596 flog_err(EC_LIB_NO_THREAD
,
1597 "Attempting to process an I/O event but for fd: %d(%d) no thread to handle this!",
1598 m
->handler
.pfds
[pos
].fd
, actual_state
);
1602 if (thread
->type
== EVENT_READ
)
1603 thread_array
= m
->read
;
1605 thread_array
= m
->write
;
1607 thread_array
[thread
->u
.fd
] = NULL
;
1608 event_list_add_tail(&m
->ready
, thread
);
1609 thread
->type
= EVENT_READY
;
1615 * Process I/O events.
1617 * Walks through file descriptor array looking for those pollfds whose .revents
1618 * field has something interesting. Deletes any invalid file descriptors.
1620 * @param m the thread master
1621 * @param num the number of active file descriptors (return value of poll())
1623 static void thread_process_io(struct event_loop
*m
, unsigned int num
)
1625 unsigned int ready
= 0;
1626 struct pollfd
*pfds
= m
->handler
.copy
;
1628 for (nfds_t i
= 0; i
< m
->handler
.copycount
&& ready
< num
; ++i
) {
1629 /* no event for current fd? immediately continue */
1630 if (pfds
[i
].revents
== 0)
1636 * Unless someone has called event_cancel from another
1637 * pthread, the only thing that could have changed in
1638 * m->handler.pfds while we were asleep is the .events
1639 * field in a given pollfd. Barring event_cancel() that
1640 * value should be a superset of the values we have in our
1641 * copy, so there's no need to update it. Similarily,
1642 * barring deletion, the fd should still be a valid index
1643 * into the master's pfds.
1645 * We are including POLLERR here to do a READ event
1646 * this is because the read should fail and the
1647 * read function should handle it appropriately
1649 if (pfds
[i
].revents
& (POLLIN
| POLLHUP
| POLLERR
)) {
1650 thread_process_io_helper(m
, m
->read
[pfds
[i
].fd
], POLLIN
,
1651 pfds
[i
].revents
, i
);
1653 if (pfds
[i
].revents
& POLLOUT
)
1654 thread_process_io_helper(m
, m
->write
[pfds
[i
].fd
],
1655 POLLOUT
, pfds
[i
].revents
, i
);
1657 /* if one of our file descriptors is garbage, remove the same
1659 * both pfds + update sizes and index */
1660 if (pfds
[i
].revents
& POLLNVAL
) {
1661 memmove(m
->handler
.pfds
+ i
, m
->handler
.pfds
+ i
+ 1,
1662 (m
->handler
.pfdcount
- i
- 1)
1663 * sizeof(struct pollfd
));
1664 m
->handler
.pfdcount
--;
1665 m
->handler
.pfds
[m
->handler
.pfdcount
].fd
= 0;
1666 m
->handler
.pfds
[m
->handler
.pfdcount
].events
= 0;
1668 memmove(pfds
+ i
, pfds
+ i
+ 1,
1669 (m
->handler
.copycount
- i
- 1)
1670 * sizeof(struct pollfd
));
1671 m
->handler
.copycount
--;
1672 m
->handler
.copy
[m
->handler
.copycount
].fd
= 0;
1673 m
->handler
.copy
[m
->handler
.copycount
].events
= 0;
1680 /* Add all timers that have popped to the ready list. */
1681 static unsigned int thread_process_timers(struct event_loop
*m
,
1682 struct timeval
*timenow
)
1684 struct timeval prev
= *timenow
;
1685 bool displayed
= false;
1686 struct event
*thread
;
1687 unsigned int ready
= 0;
1689 while ((thread
= event_timer_list_first(&m
->timer
))) {
1690 if (timercmp(timenow
, &thread
->u
.sands
, <))
1692 prev
= thread
->u
.sands
;
1695 * If the timer would have popped 4 seconds in the
1696 * past then we are in a situation where we are
1697 * really getting behind on handling of events.
1698 * Let's log it and do the right thing with it.
1700 if (timercmp(timenow
, &prev
, >)) {
1701 atomic_fetch_add_explicit(
1702 &thread
->hist
->total_starv_warn
, 1,
1703 memory_order_seq_cst
);
1704 if (!displayed
&& !thread
->ignore_timer_late
) {
1706 EC_LIB_STARVE_THREAD
,
1707 "Thread Starvation: %pTHD was scheduled to pop greater than 4s ago",
1713 event_timer_list_pop(&m
->timer
);
1714 thread
->type
= EVENT_READY
;
1715 event_list_add_tail(&m
->ready
, thread
);
1722 /* process a list en masse, e.g. for event thread lists */
1723 static unsigned int thread_process(struct event_list_head
*list
)
1725 struct event
*thread
;
1726 unsigned int ready
= 0;
1728 while ((thread
= event_list_pop(list
))) {
1729 thread
->type
= EVENT_READY
;
1730 event_list_add_tail(&thread
->master
->ready
, thread
);
1737 /* Fetch next ready thread. */
1738 struct event
*event_fetch(struct event_loop
*m
, struct event
*fetch
)
1740 struct event
*thread
= NULL
;
1742 struct timeval zerotime
= {0, 0};
1744 struct timeval
*tw
= NULL
;
1745 bool eintr_p
= false;
1749 /* Handle signals if any */
1750 if (m
->handle_signals
)
1751 frr_sigevent_process();
1753 pthread_mutex_lock(&m
->mtx
);
1755 /* Process any pending cancellation requests */
1759 * Attempt to flush ready queue before going into poll().
1760 * This is performance-critical. Think twice before modifying.
1762 if ((thread
= event_list_pop(&m
->ready
))) {
1763 fetch
= thread_run(m
, thread
, fetch
);
1766 pthread_mutex_unlock(&m
->mtx
);
1767 if (!m
->ready_run_loop
)
1768 GETRUSAGE(&m
->last_getrusage
);
1769 m
->ready_run_loop
= true;
1773 m
->ready_run_loop
= false;
1774 /* otherwise, tick through scheduling sequence */
1777 * Post events to ready queue. This must come before the
1778 * following block since events should occur immediately
1780 thread_process(&m
->event
);
1783 * If there are no tasks on the ready queue, we will poll()
1784 * until a timer expires or we receive I/O, whichever comes
1785 * first. The strategy for doing this is:
1787 * - If there are events pending, set the poll() timeout to zero
1788 * - If there are no events pending, but there are timers
1789 * pending, set the timeout to the smallest remaining time on
1791 * - If there are neither timers nor events pending, but there
1792 * are file descriptors pending, block indefinitely in poll()
1793 * - If nothing is pending, it's time for the application to die
1795 * In every case except the last, we need to hit poll() at least
1796 * once per loop to avoid starvation by events
1798 if (!event_list_count(&m
->ready
))
1799 tw
= thread_timer_wait(&m
->timer
, &tv
);
1801 if (event_list_count(&m
->ready
) ||
1802 (tw
&& !timercmp(tw
, &zerotime
, >)))
1805 if (!tw
&& m
->handler
.pfdcount
== 0) { /* die */
1806 pthread_mutex_unlock(&m
->mtx
);
1812 * Copy pollfd array + # active pollfds in it. Not necessary to
1813 * copy the array size as this is fixed.
1815 m
->handler
.copycount
= m
->handler
.pfdcount
;
1816 memcpy(m
->handler
.copy
, m
->handler
.pfds
,
1817 m
->handler
.copycount
* sizeof(struct pollfd
));
1819 pthread_mutex_unlock(&m
->mtx
);
1822 num
= fd_poll(m
, tw
, &eintr_p
);
1824 pthread_mutex_lock(&m
->mtx
);
1826 /* Handle any errors received in poll() */
1829 pthread_mutex_unlock(&m
->mtx
);
1830 /* loop around to signal handler */
1835 flog_err(EC_LIB_SYSTEM_CALL
, "poll() error: %s",
1836 safe_strerror(errno
));
1837 pthread_mutex_unlock(&m
->mtx
);
1842 /* Post timers to ready queue. */
1844 thread_process_timers(m
, &now
);
1846 /* Post I/O to ready queue. */
1848 thread_process_io(m
, num
);
1850 pthread_mutex_unlock(&m
->mtx
);
1852 } while (!thread
&& m
->spin
);
1857 static unsigned long timeval_elapsed(struct timeval a
, struct timeval b
)
1859 return (((a
.tv_sec
- b
.tv_sec
) * TIMER_SECOND_MICRO
)
1860 + (a
.tv_usec
- b
.tv_usec
));
1863 unsigned long event_consumed_time(RUSAGE_T
*now
, RUSAGE_T
*start
,
1864 unsigned long *cputime
)
1866 #ifdef HAVE_CLOCK_THREAD_CPUTIME_ID
1870 * FreeBSD appears to have an issue when calling clock_gettime
1871 * with CLOCK_THREAD_CPUTIME_ID really close to each other
1872 * occassionally the now time will be before the start time.
1873 * This is not good and FRR is ending up with CPU HOG's
1874 * when the subtraction wraps to very large numbers
1876 * What we are going to do here is cheat a little bit
1877 * and notice that this is a problem and just correct
1878 * it so that it is impossible to happen
1880 if (start
->cpu
.tv_sec
== now
->cpu
.tv_sec
&&
1881 start
->cpu
.tv_nsec
> now
->cpu
.tv_nsec
)
1882 now
->cpu
.tv_nsec
= start
->cpu
.tv_nsec
+ 1;
1883 else if (start
->cpu
.tv_sec
> now
->cpu
.tv_sec
) {
1884 now
->cpu
.tv_sec
= start
->cpu
.tv_sec
;
1885 now
->cpu
.tv_nsec
= start
->cpu
.tv_nsec
+ 1;
1888 *cputime
= (now
->cpu
.tv_sec
- start
->cpu
.tv_sec
) * TIMER_SECOND_MICRO
1889 + (now
->cpu
.tv_nsec
- start
->cpu
.tv_nsec
) / 1000;
1891 /* This is 'user + sys' time. */
1892 *cputime
= timeval_elapsed(now
->cpu
.ru_utime
, start
->cpu
.ru_utime
)
1893 + timeval_elapsed(now
->cpu
.ru_stime
, start
->cpu
.ru_stime
);
1895 return timeval_elapsed(now
->real
, start
->real
);
1898 /* We should aim to yield after yield milliseconds, which defaults
1899 to EVENT_YIELD_TIME_SLOT .
1900 Note: we are using real (wall clock) time for this calculation.
1901 It could be argued that CPU time may make more sense in certain
1902 contexts. The things to consider are whether the thread may have
1903 blocked (in which case wall time increases, but CPU time does not),
1904 or whether the system is heavily loaded with other processes competing
1905 for CPU time. On balance, wall clock time seems to make sense.
1906 Plus it has the added benefit that gettimeofday should be faster
1907 than calling getrusage. */
1908 int event_should_yield(struct event
*thread
)
1911 frr_with_mutex (&thread
->mtx
) {
1912 result
= monotime_since(&thread
->real
, NULL
)
1913 > (int64_t)thread
->yield
;
1918 void event_set_yield_time(struct event
*thread
, unsigned long yield_time
)
1920 frr_with_mutex (&thread
->mtx
) {
1921 thread
->yield
= yield_time
;
1925 void event_getrusage(RUSAGE_T
*r
)
1928 if (!cputime_enabled
) {
1929 memset(&r
->cpu
, 0, sizeof(r
->cpu
));
1933 #ifdef HAVE_CLOCK_THREAD_CPUTIME_ID
1934 /* not currently implemented in Linux's vDSO, but maybe at some point
1937 clock_gettime(CLOCK_THREAD_CPUTIME_ID
, &r
->cpu
);
1938 #else /* !HAVE_CLOCK_THREAD_CPUTIME_ID */
1939 #if defined RUSAGE_THREAD
1940 #define FRR_RUSAGE RUSAGE_THREAD
1942 #define FRR_RUSAGE RUSAGE_SELF
1944 getrusage(FRR_RUSAGE
, &(r
->cpu
));
1951 * This function will atomically update the thread's usage history. At present
1952 * this is the only spot where usage history is written. Nevertheless the code
1953 * has been written such that the introduction of writers in the future should
1954 * not need to update it provided the writers atomically perform only the
1955 * operations done here, i.e. updating the total and maximum times. In
1956 * particular, the maximum real and cpu times must be monotonically increasing
1957 * or this code is not correct.
1959 void event_call(struct event
*thread
)
1961 RUSAGE_T before
, after
;
1963 /* if the thread being called is the CLI, it may change cputime_enabled
1964 * ("service cputime-stats" command), which can result in nonsensical
1965 * and very confusing warnings
1967 bool cputime_enabled_here
= cputime_enabled
;
1969 if (thread
->master
->ready_run_loop
)
1970 before
= thread
->master
->last_getrusage
;
1974 thread
->real
= before
.real
;
1976 frrtrace(9, frr_libfrr
, event_call
, thread
->master
,
1977 thread
->xref
->funcname
, thread
->xref
->xref
.file
,
1978 thread
->xref
->xref
.line
, NULL
, thread
->u
.fd
, thread
->u
.val
,
1979 thread
->arg
, thread
->u
.sands
.tv_sec
);
1981 pthread_setspecific(thread_current
, thread
);
1982 (*thread
->func
)(thread
);
1983 pthread_setspecific(thread_current
, NULL
);
1986 thread
->master
->last_getrusage
= after
;
1988 unsigned long walltime
, cputime
;
1991 walltime
= event_consumed_time(&after
, &before
, &cputime
);
1993 /* update walltime */
1994 atomic_fetch_add_explicit(&thread
->hist
->real
.total
, walltime
,
1995 memory_order_seq_cst
);
1996 exp
= atomic_load_explicit(&thread
->hist
->real
.max
,
1997 memory_order_seq_cst
);
1998 while (exp
< walltime
1999 && !atomic_compare_exchange_weak_explicit(
2000 &thread
->hist
->real
.max
, &exp
, walltime
,
2001 memory_order_seq_cst
, memory_order_seq_cst
))
2004 if (cputime_enabled_here
&& cputime_enabled
) {
2005 /* update cputime */
2006 atomic_fetch_add_explicit(&thread
->hist
->cpu
.total
, cputime
,
2007 memory_order_seq_cst
);
2008 exp
= atomic_load_explicit(&thread
->hist
->cpu
.max
,
2009 memory_order_seq_cst
);
2010 while (exp
< cputime
2011 && !atomic_compare_exchange_weak_explicit(
2012 &thread
->hist
->cpu
.max
, &exp
, cputime
,
2013 memory_order_seq_cst
, memory_order_seq_cst
))
2017 atomic_fetch_add_explicit(&thread
->hist
->total_calls
, 1,
2018 memory_order_seq_cst
);
2019 atomic_fetch_or_explicit(&thread
->hist
->types
, 1 << thread
->add_type
,
2020 memory_order_seq_cst
);
2022 if (cputime_enabled_here
&& cputime_enabled
&& cputime_threshold
2023 && cputime
> cputime_threshold
) {
2025 * We have a CPU Hog on our hands. The time FRR has spent
2026 * doing actual work (not sleeping) is greater than 5 seconds.
2027 * Whinge about it now, so we're aware this is yet another task
2030 atomic_fetch_add_explicit(&thread
->hist
->total_cpu_warn
,
2031 1, memory_order_seq_cst
);
2033 EC_LIB_SLOW_THREAD_CPU
,
2034 "CPU HOG: task %s (%lx) ran for %lums (cpu time %lums)",
2035 thread
->xref
->funcname
, (unsigned long)thread
->func
,
2036 walltime
/ 1000, cputime
/ 1000);
2038 } else if (walltime_threshold
&& walltime
> walltime_threshold
) {
2040 * The runtime for a task is greater than 5 seconds, but the
2041 * cpu time is under 5 seconds. Let's whine about this because
2042 * this could imply some sort of scheduling issue.
2044 atomic_fetch_add_explicit(&thread
->hist
->total_wall_warn
,
2045 1, memory_order_seq_cst
);
2047 EC_LIB_SLOW_THREAD_WALL
,
2048 "STARVATION: task %s (%lx) ran for %lums (cpu time %lums)",
2049 thread
->xref
->funcname
, (unsigned long)thread
->func
,
2050 walltime
/ 1000, cputime
/ 1000);
2054 /* Execute thread */
2055 void _event_execute(const struct xref_eventsched
*xref
, struct event_loop
*m
,
2056 void (*func
)(struct event
*), void *arg
, int val
)
2058 struct event
*thread
;
2060 /* Get or allocate new thread to execute. */
2061 frr_with_mutex (&m
->mtx
) {
2062 thread
= thread_get(m
, EVENT_EVENT
, func
, arg
, xref
);
2064 /* Set its event value. */
2065 frr_with_mutex (&thread
->mtx
) {
2066 thread
->add_type
= EVENT_EXECUTE
;
2067 thread
->u
.val
= val
;
2068 thread
->ref
= &thread
;
2072 /* Execute thread doing all accounting. */
2075 /* Give back or free thread. */
2076 thread_add_unuse(m
, thread
);
2079 /* Debug signal mask - if 'sigs' is NULL, use current effective mask. */
2080 void debug_signals(const sigset_t
*sigs
)
2087 * We're only looking at the non-realtime signals here, so we need
2088 * some limit value. Platform differences mean at some point we just
2089 * need to pick a reasonable value.
2091 #if defined SIGRTMIN
2092 # define LAST_SIGNAL SIGRTMIN
2094 # define LAST_SIGNAL 32
2099 sigemptyset(&tmpsigs
);
2100 pthread_sigmask(SIG_BLOCK
, NULL
, &tmpsigs
);
2107 for (i
= 0; i
< LAST_SIGNAL
; i
++) {
2110 if (sigismember(sigs
, i
) > 0) {
2112 strlcat(buf
, ",", sizeof(buf
));
2113 snprintf(tmp
, sizeof(tmp
), "%d", i
);
2114 strlcat(buf
, tmp
, sizeof(buf
));
2120 snprintf(buf
, sizeof(buf
), "<none>");
2122 zlog_debug("%s: %s", __func__
, buf
);
2125 static ssize_t
printfrr_thread_dbg(struct fbuf
*buf
, struct printfrr_eargs
*ea
,
2126 const struct event
*thread
)
2128 static const char *const types
[] = {
2129 [EVENT_READ
] = "read", [EVENT_WRITE
] = "write",
2130 [EVENT_TIMER
] = "timer", [EVENT_EVENT
] = "event",
2131 [EVENT_READY
] = "ready", [EVENT_UNUSED
] = "unused",
2132 [EVENT_EXECUTE
] = "exec",
2138 return bputs(buf
, "{(thread *)NULL}");
2140 rv
+= bprintfrr(buf
, "{(thread *)%p arg=%p", thread
, thread
->arg
);
2142 if (thread
->type
< array_size(types
) && types
[thread
->type
])
2143 rv
+= bprintfrr(buf
, " %-6s", types
[thread
->type
]);
2145 rv
+= bprintfrr(buf
, " INVALID(%u)", thread
->type
);
2147 switch (thread
->type
) {
2150 snprintfrr(info
, sizeof(info
), "fd=%d", thread
->u
.fd
);
2154 snprintfrr(info
, sizeof(info
), "r=%pTVMud", &thread
->u
.sands
);
2163 rv
+= bprintfrr(buf
, " %-12s %s() %s from %s:%d}", info
,
2164 thread
->xref
->funcname
, thread
->xref
->dest
,
2165 thread
->xref
->xref
.file
, thread
->xref
->xref
.line
);
2169 printfrr_ext_autoreg_p("TH", printfrr_thread
);
2170 static ssize_t
printfrr_thread(struct fbuf
*buf
, struct printfrr_eargs
*ea
,
2173 const struct event
*thread
= ptr
;
2174 struct timespec remain
= {};
2176 if (ea
->fmt
[0] == 'D') {
2178 return printfrr_thread_dbg(buf
, ea
, thread
);
2182 /* need to jump over time formatting flag characters in the
2183 * input format string, i.e. adjust ea->fmt!
2185 printfrr_time(buf
, ea
, &remain
,
2186 TIMEFMT_TIMER_DEADLINE
| TIMEFMT_SKIP
);
2187 return bputch(buf
, '-');
2190 TIMEVAL_TO_TIMESPEC(&thread
->u
.sands
, &remain
);
2191 return printfrr_time(buf
, ea
, &remain
, TIMEFMT_TIMER_DEADLINE
);