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_master
*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_thread_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_thread_history
*a
,
100 const struct cpu_thread_history
*b
)
102 return a
->func
== b
->func
;
105 static void *cpu_record_hash_alloc(struct cpu_thread_history
*a
)
107 struct cpu_thread_history
*new;
108 new = XCALLOC(MTYPE_EVENT_STATS
, sizeof(struct cpu_thread_history
));
110 new->funcname
= a
->funcname
;
114 static void cpu_record_hash_free(void *a
)
116 struct cpu_thread_history
*hist
= a
;
118 XFREE(MTYPE_EVENT_STATS
, hist
);
121 static void vty_out_cpu_thread_history(struct vty
*vty
,
122 struct cpu_thread_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_thread_history
*totals
= args
[0];
141 struct cpu_thread_history copy
;
142 struct vty
*vty
= args
[1];
143 uint8_t *filter
= args
[2];
145 struct cpu_thread_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_thread_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_thread_history tmp
;
187 void *args
[3] = {&tmp
, vty
, &filter
};
188 struct event_master
*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_thread_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_thread_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_master
*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_master
*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_master
*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_master
*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_master
*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_master
*thread_master_create(const char *name
)
549 struct event_master
*rv
;
552 pthread_once(&init_once
, &initializer
);
554 rv
= XCALLOC(MTYPE_EVENT_MASTER
, sizeof(struct event_master
));
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 thread_master_set_name(struct event_master
*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_master
*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_master
*m
,
660 struct event_list_head
*list
)
664 while ((t
= event_list_pop(list
)))
668 static void thread_array_free(struct event_master
*m
,
669 struct event
**thread_array
)
674 for (index
= 0; index
< m
->fd_limit
; ++index
) {
675 t
= thread_array
[index
];
677 thread_array
[index
] = NULL
;
681 XFREE(MTYPE_EVENT_POLL
, thread_array
);
685 * thread_master_free_unused
687 * As threads are finished with they are put on the
688 * unuse list for later reuse.
689 * If we are shutting down, Free up unused threads
690 * So we can see if we forget to shut anything off
692 void thread_master_free_unused(struct event_master
*m
)
694 frr_with_mutex (&m
->mtx
) {
696 while ((t
= event_list_pop(&m
->unuse
)))
701 /* Stop thread scheduler. */
702 void thread_master_free(struct event_master
*m
)
706 frr_with_mutex (&masters_mtx
) {
707 listnode_delete(masters
, m
);
708 if (masters
->count
== 0) {
709 list_delete(&masters
);
713 thread_array_free(m
, m
->read
);
714 thread_array_free(m
, m
->write
);
715 while ((t
= event_timer_list_pop(&m
->timer
)))
717 thread_list_free(m
, &m
->event
);
718 thread_list_free(m
, &m
->ready
);
719 thread_list_free(m
, &m
->unuse
);
720 pthread_mutex_destroy(&m
->mtx
);
721 pthread_cond_destroy(&m
->cancel_cond
);
722 close(m
->io_pipe
[0]);
723 close(m
->io_pipe
[1]);
724 list_delete(&m
->cancel_req
);
725 m
->cancel_req
= NULL
;
727 hash_clean_and_free(&m
->cpu_record
, cpu_record_hash_free
);
729 XFREE(MTYPE_EVENT_MASTER
, m
->name
);
730 XFREE(MTYPE_EVENT_MASTER
, m
->handler
.pfds
);
731 XFREE(MTYPE_EVENT_MASTER
, m
->handler
.copy
);
732 XFREE(MTYPE_EVENT_MASTER
, m
);
735 /* Return remain time in milliseconds. */
736 unsigned long event_timer_remain_msec(struct event
*thread
)
740 if (!event_is_scheduled(thread
))
743 frr_with_mutex (&thread
->mtx
) {
744 remain
= monotime_until(&thread
->u
.sands
, NULL
) / 1000LL;
747 return remain
< 0 ? 0 : remain
;
750 /* Return remain time in seconds. */
751 unsigned long event_timer_remain_second(struct event
*thread
)
753 return event_timer_remain_msec(thread
) / 1000LL;
756 struct timeval
event_timer_remain(struct event
*thread
)
758 struct timeval remain
;
759 frr_with_mutex (&thread
->mtx
) {
760 monotime_until(&thread
->u
.sands
, &remain
);
765 static int time_hhmmss(char *buf
, int buf_size
, long sec
)
771 assert(buf_size
>= 8);
778 wr
= snprintf(buf
, buf_size
, "%02ld:%02ld:%02ld", hh
, mm
, sec
);
783 char *event_timer_to_hhmmss(char *buf
, int buf_size
, struct event
*t_timer
)
786 time_hhmmss(buf
, buf_size
, event_timer_remain_second(t_timer
));
788 snprintf(buf
, buf_size
, "--:--:--");
793 /* Get new thread. */
794 static struct event
*thread_get(struct event_master
*m
, uint8_t type
,
795 void (*func
)(struct event
*), void *arg
,
796 const struct xref_eventsched
*xref
)
798 struct event
*thread
= event_list_pop(&m
->unuse
);
799 struct cpu_thread_history tmp
;
802 thread
= XCALLOC(MTYPE_THREAD
, sizeof(struct event
));
803 /* mutex only needs to be initialized at struct creation. */
804 pthread_mutex_init(&thread
->mtx
, NULL
);
809 thread
->add_type
= type
;
812 thread
->yield
= EVENT_YIELD_TIME_SLOT
; /* default */
814 thread
->ignore_timer_late
= false;
817 * So if the passed in funcname is not what we have
818 * stored that means the thread->hist needs to be
819 * updated. We keep the last one around in unused
820 * under the assumption that we are probably
821 * going to immediately allocate the same
823 * This hopefully saves us some serious
826 if ((thread
->xref
&& thread
->xref
->funcname
!= xref
->funcname
)
827 || thread
->func
!= func
) {
829 tmp
.funcname
= xref
->funcname
;
831 hash_get(m
->cpu_record
, &tmp
,
832 (void *(*)(void *))cpu_record_hash_alloc
);
834 thread
->hist
->total_active
++;
841 static void thread_free(struct event_master
*master
, struct event
*thread
)
843 /* Update statistics. */
844 assert(master
->alloc
> 0);
847 /* Free allocated resources. */
848 pthread_mutex_destroy(&thread
->mtx
);
849 XFREE(MTYPE_THREAD
, thread
);
852 static int fd_poll(struct event_master
*m
, const struct timeval
*timer_wait
,
856 unsigned char trash
[64];
857 nfds_t count
= m
->handler
.copycount
;
860 * If timer_wait is null here, that means poll() should block
861 * indefinitely, unless the thread_master has overridden it by setting
862 * ->selectpoll_timeout.
864 * If the value is positive, it specifies the maximum number of
865 * milliseconds to wait. If the timeout is -1, it specifies that
866 * we should never wait and always return immediately even if no
867 * event is detected. If the value is zero, the behavior is default.
871 /* number of file descriptors with events */
874 if (timer_wait
!= NULL
875 && m
->selectpoll_timeout
== 0) // use the default value
876 timeout
= (timer_wait
->tv_sec
* 1000)
877 + (timer_wait
->tv_usec
/ 1000);
878 else if (m
->selectpoll_timeout
> 0) // use the user's timeout
879 timeout
= m
->selectpoll_timeout
;
880 else if (m
->selectpoll_timeout
881 < 0) // effect a poll (return immediately)
884 zlog_tls_buffer_flush();
886 rcu_assert_read_unlocked();
888 /* add poll pipe poker */
889 assert(count
+ 1 < m
->handler
.pfdsize
);
890 m
->handler
.copy
[count
].fd
= m
->io_pipe
[0];
891 m
->handler
.copy
[count
].events
= POLLIN
;
892 m
->handler
.copy
[count
].revents
= 0x00;
894 /* We need to deal with a signal-handling race here: we
895 * don't want to miss a crucial signal, such as SIGTERM or SIGINT,
896 * that may arrive just before we enter poll(). We will block the
897 * key signals, then check whether any have arrived - if so, we return
898 * before calling poll(). If not, we'll re-enable the signals
899 * in the ppoll() call.
902 sigemptyset(&origsigs
);
903 if (m
->handle_signals
) {
904 /* Main pthread that handles the app signals */
905 if (frr_sigevent_check(&origsigs
)) {
906 /* Signal to process - restore signal mask and return */
907 pthread_sigmask(SIG_SETMASK
, &origsigs
, NULL
);
913 /* Don't make any changes for the non-main pthreads */
914 pthread_sigmask(SIG_SETMASK
, NULL
, &origsigs
);
917 #if defined(HAVE_PPOLL)
918 struct timespec ts
, *tsp
;
921 ts
.tv_sec
= timeout
/ 1000;
922 ts
.tv_nsec
= (timeout
% 1000) * 1000000;
927 num
= ppoll(m
->handler
.copy
, count
+ 1, tsp
, &origsigs
);
928 pthread_sigmask(SIG_SETMASK
, &origsigs
, NULL
);
930 /* Not ideal - there is a race after we restore the signal mask */
931 pthread_sigmask(SIG_SETMASK
, &origsigs
, NULL
);
932 num
= poll(m
->handler
.copy
, count
+ 1, timeout
);
937 if (num
< 0 && errno
== EINTR
)
940 if (num
> 0 && m
->handler
.copy
[count
].revents
!= 0 && num
--)
941 while (read(m
->io_pipe
[0], &trash
, sizeof(trash
)) > 0)
949 /* Add new read thread. */
950 void _event_add_read_write(const struct xref_eventsched
*xref
,
951 struct event_master
*m
, void (*func
)(struct event
*),
952 void *arg
, int fd
, struct event
**t_ptr
)
954 int dir
= xref
->event_type
;
955 struct event
*thread
= NULL
;
956 struct event
**thread_array
;
958 if (dir
== EVENT_READ
)
959 frrtrace(9, frr_libfrr
, schedule_read
, m
,
960 xref
->funcname
, xref
->xref
.file
, xref
->xref
.line
,
961 t_ptr
, fd
, 0, arg
, 0);
963 frrtrace(9, frr_libfrr
, schedule_write
, m
,
964 xref
->funcname
, xref
->xref
.file
, xref
->xref
.line
,
965 t_ptr
, fd
, 0, arg
, 0);
968 if (fd
>= m
->fd_limit
)
969 assert(!"Number of FD's open is greater than FRR currently configured to handle, aborting");
971 frr_with_mutex (&m
->mtx
) {
973 // thread is already scheduled; don't reschedule
976 /* default to a new pollfd */
977 nfds_t queuepos
= m
->handler
.pfdcount
;
979 if (dir
== EVENT_READ
)
980 thread_array
= m
->read
;
982 thread_array
= m
->write
;
984 /* if we already have a pollfd for our file descriptor, find and
986 for (nfds_t i
= 0; i
< m
->handler
.pfdcount
; i
++)
987 if (m
->handler
.pfds
[i
].fd
== fd
) {
992 * What happens if we have a thread already
993 * created for this event?
995 if (thread_array
[fd
])
996 assert(!"Thread already scheduled for file descriptor");
1001 /* make sure we have room for this fd + pipe poker fd */
1002 assert(queuepos
+ 1 < m
->handler
.pfdsize
);
1004 thread
= thread_get(m
, dir
, func
, arg
, xref
);
1006 m
->handler
.pfds
[queuepos
].fd
= fd
;
1007 m
->handler
.pfds
[queuepos
].events
|=
1008 (dir
== EVENT_READ
? POLLIN
: POLLOUT
);
1010 if (queuepos
== m
->handler
.pfdcount
)
1011 m
->handler
.pfdcount
++;
1014 frr_with_mutex (&thread
->mtx
) {
1016 thread_array
[thread
->u
.fd
] = thread
;
1021 thread
->ref
= t_ptr
;
1029 static void _event_add_timer_timeval(const struct xref_eventsched
*xref
,
1030 struct event_master
*m
,
1031 void (*func
)(struct event
*), void *arg
,
1032 struct timeval
*time_relative
,
1033 struct event
**t_ptr
)
1035 struct event
*thread
;
1040 assert(time_relative
);
1042 frrtrace(9, frr_libfrr
, schedule_timer
, m
,
1043 xref
->funcname
, xref
->xref
.file
, xref
->xref
.line
,
1044 t_ptr
, 0, 0, arg
, (long)time_relative
->tv_sec
);
1046 /* Compute expiration/deadline time. */
1048 timeradd(&t
, time_relative
, &t
);
1050 frr_with_mutex (&m
->mtx
) {
1051 if (t_ptr
&& *t_ptr
)
1052 /* thread is already scheduled; don't reschedule */
1055 thread
= thread_get(m
, EVENT_TIMER
, func
, arg
, xref
);
1057 frr_with_mutex (&thread
->mtx
) {
1058 thread
->u
.sands
= t
;
1059 event_timer_list_add(&m
->timer
, thread
);
1062 thread
->ref
= t_ptr
;
1066 /* The timer list is sorted - if this new timer
1067 * might change the time we'll wait for, give the pthread
1068 * a chance to re-compute.
1070 if (event_timer_list_first(&m
->timer
) == thread
)
1073 #define ONEYEAR2SEC (60 * 60 * 24 * 365)
1074 if (time_relative
->tv_sec
> ONEYEAR2SEC
)
1076 EC_LIB_TIMER_TOO_LONG
,
1077 "Timer: %pTHD is created with an expiration that is greater than 1 year",
1082 /* Add timer event thread. */
1083 void _event_add_timer(const struct xref_eventsched
*xref
,
1084 struct event_master
*m
, void (*func
)(struct event
*),
1085 void *arg
, long timer
, struct event
**t_ptr
)
1087 struct timeval trel
;
1091 trel
.tv_sec
= timer
;
1094 _event_add_timer_timeval(xref
, m
, func
, arg
, &trel
, t_ptr
);
1097 /* Add timer event thread with "millisecond" resolution */
1098 void _event_add_timer_msec(const struct xref_eventsched
*xref
,
1099 struct event_master
*m
, void (*func
)(struct event
*),
1100 void *arg
, long timer
, struct event
**t_ptr
)
1102 struct timeval trel
;
1106 trel
.tv_sec
= timer
/ 1000;
1107 trel
.tv_usec
= 1000 * (timer
% 1000);
1109 _event_add_timer_timeval(xref
, m
, func
, arg
, &trel
, t_ptr
);
1112 /* Add timer event thread with "timeval" resolution */
1113 void _event_add_timer_tv(const struct xref_eventsched
*xref
,
1114 struct event_master
*m
, void (*func
)(struct event
*),
1115 void *arg
, struct timeval
*tv
, struct event
**t_ptr
)
1117 _event_add_timer_timeval(xref
, m
, func
, arg
, tv
, t_ptr
);
1120 /* Add simple event thread. */
1121 void _event_add_event(const struct xref_eventsched
*xref
,
1122 struct event_master
*m
, void (*func
)(struct event
*),
1123 void *arg
, int val
, struct event
**t_ptr
)
1125 struct event
*thread
= NULL
;
1127 frrtrace(9, frr_libfrr
, schedule_event
, m
,
1128 xref
->funcname
, xref
->xref
.file
, xref
->xref
.line
,
1129 t_ptr
, 0, val
, arg
, 0);
1133 frr_with_mutex (&m
->mtx
) {
1134 if (t_ptr
&& *t_ptr
)
1135 /* thread is already scheduled; don't reschedule */
1138 thread
= thread_get(m
, EVENT_EVENT
, func
, arg
, xref
);
1139 frr_with_mutex (&thread
->mtx
) {
1140 thread
->u
.val
= val
;
1141 event_list_add_tail(&m
->event
, thread
);
1146 thread
->ref
= t_ptr
;
1153 /* Thread cancellation ------------------------------------------------------ */
1156 * NOT's out the .events field of pollfd corresponding to the given file
1157 * descriptor. The event to be NOT'd is passed in the 'state' parameter.
1159 * This needs to happen for both copies of pollfd's. See 'event_fetch'
1160 * implementation for details.
1164 * @param state the event to cancel. One or more (OR'd together) of the
1169 static void event_cancel_rw(struct event_master
*master
, int fd
, short state
,
1174 /* find the index of corresponding pollfd */
1177 /* Cancel POLLHUP too just in case some bozo set it */
1180 /* Some callers know the index of the pfd already */
1181 if (idx_hint
>= 0) {
1185 /* Have to look for the fd in the pfd array */
1186 for (i
= 0; i
< master
->handler
.pfdcount
; i
++)
1187 if (master
->handler
.pfds
[i
].fd
== fd
) {
1195 "[!] Received cancellation request for nonexistent rw job");
1196 zlog_debug("[!] threadmaster: %s | fd: %d",
1197 master
->name
? master
->name
: "", fd
);
1201 /* NOT out event. */
1202 master
->handler
.pfds
[i
].events
&= ~(state
);
1204 /* If all events are canceled, delete / resize the pollfd array. */
1205 if (master
->handler
.pfds
[i
].events
== 0) {
1206 memmove(master
->handler
.pfds
+ i
, master
->handler
.pfds
+ i
+ 1,
1207 (master
->handler
.pfdcount
- i
- 1)
1208 * sizeof(struct pollfd
));
1209 master
->handler
.pfdcount
--;
1210 master
->handler
.pfds
[master
->handler
.pfdcount
].fd
= 0;
1211 master
->handler
.pfds
[master
->handler
.pfdcount
].events
= 0;
1214 /* If we have the same pollfd in the copy, perform the same operations,
1215 * otherwise return. */
1216 if (i
>= master
->handler
.copycount
)
1219 master
->handler
.copy
[i
].events
&= ~(state
);
1221 if (master
->handler
.copy
[i
].events
== 0) {
1222 memmove(master
->handler
.copy
+ i
, master
->handler
.copy
+ i
+ 1,
1223 (master
->handler
.copycount
- i
- 1)
1224 * sizeof(struct pollfd
));
1225 master
->handler
.copycount
--;
1226 master
->handler
.copy
[master
->handler
.copycount
].fd
= 0;
1227 master
->handler
.copy
[master
->handler
.copycount
].events
= 0;
1232 * Process task cancellation given a task argument: iterate through the
1233 * various lists of tasks, looking for any that match the argument.
1235 static void cancel_arg_helper(struct event_master
*master
,
1236 const struct cancel_req
*cr
)
1243 /* We're only processing arg-based cancellations here. */
1244 if (cr
->eventobj
== NULL
)
1247 /* First process the ready lists. */
1248 frr_each_safe (event_list
, &master
->event
, t
) {
1249 if (t
->arg
!= cr
->eventobj
)
1251 event_list_del(&master
->event
, t
);
1254 thread_add_unuse(master
, t
);
1257 frr_each_safe (event_list
, &master
->ready
, t
) {
1258 if (t
->arg
!= cr
->eventobj
)
1260 event_list_del(&master
->ready
, t
);
1263 thread_add_unuse(master
, t
);
1266 /* If requested, stop here and ignore io and timers */
1267 if (CHECK_FLAG(cr
->flags
, EVENT_CANCEL_FLAG_READY
))
1270 /* Check the io tasks */
1271 for (i
= 0; i
< master
->handler
.pfdcount
;) {
1272 pfd
= master
->handler
.pfds
+ i
;
1274 if (pfd
->events
& POLLIN
)
1275 t
= master
->read
[pfd
->fd
];
1277 t
= master
->write
[pfd
->fd
];
1279 if (t
&& t
->arg
== cr
->eventobj
) {
1282 /* Found a match to cancel: clean up fd arrays */
1283 event_cancel_rw(master
, pfd
->fd
, pfd
->events
, i
);
1285 /* Clean up thread arrays */
1286 master
->read
[fd
] = NULL
;
1287 master
->write
[fd
] = NULL
;
1289 /* Clear caller's ref */
1293 thread_add_unuse(master
, t
);
1295 /* Don't increment 'i' since the cancellation will have
1296 * removed the entry from the pfd array
1302 /* Check the timer tasks */
1303 t
= event_timer_list_first(&master
->timer
);
1305 struct event
*t_next
;
1307 t_next
= event_timer_list_next(&master
->timer
, t
);
1309 if (t
->arg
== cr
->eventobj
) {
1310 event_timer_list_del(&master
->timer
, t
);
1313 thread_add_unuse(master
, t
);
1321 * Process cancellation requests.
1323 * This may only be run from the pthread which owns the thread_master.
1325 * @param master the thread master to process
1326 * @REQUIRE master->mtx
1328 static void do_event_cancel(struct event_master
*master
)
1330 struct event_list_head
*list
= NULL
;
1331 struct event
**thread_array
= NULL
;
1332 struct event
*thread
;
1333 struct cancel_req
*cr
;
1334 struct listnode
*ln
;
1336 for (ALL_LIST_ELEMENTS_RO(master
->cancel_req
, ln
, cr
)) {
1338 * If this is an event object cancellation, search
1339 * through task lists deleting any tasks which have the
1340 * specified argument - use this handy helper function.
1343 cancel_arg_helper(master
, cr
);
1348 * The pointer varies depending on whether the cancellation
1349 * request was made asynchronously or not. If it was, we
1350 * need to check whether the thread even exists anymore
1351 * before cancelling it.
1353 thread
= (cr
->thread
) ? cr
->thread
: *cr
->threadref
;
1359 thread_array
= NULL
;
1361 /* Determine the appropriate queue to cancel the thread from */
1362 switch (thread
->type
) {
1364 event_cancel_rw(master
, thread
->u
.fd
, POLLIN
, -1);
1365 thread_array
= master
->read
;
1368 event_cancel_rw(master
, thread
->u
.fd
, POLLOUT
, -1);
1369 thread_array
= master
->write
;
1372 event_timer_list_del(&master
->timer
, thread
);
1375 list
= &master
->event
;
1378 list
= &master
->ready
;
1387 event_list_del(list
, thread
);
1388 } else if (thread_array
) {
1389 thread_array
[thread
->u
.fd
] = NULL
;
1393 *thread
->ref
= NULL
;
1395 thread_add_unuse(thread
->master
, thread
);
1398 /* Delete and free all cancellation requests */
1399 if (master
->cancel_req
)
1400 list_delete_all_node(master
->cancel_req
);
1402 /* Wake up any threads which may be blocked in event_cancel_async() */
1403 master
->canceled
= true;
1404 pthread_cond_broadcast(&master
->cancel_cond
);
1408 * Helper function used for multiple flavors of arg-based cancellation.
1410 static void cancel_event_helper(struct event_master
*m
, void *arg
, int flags
)
1412 struct cancel_req
*cr
;
1414 assert(m
->owner
== pthread_self());
1416 /* Only worth anything if caller supplies an arg. */
1420 cr
= XCALLOC(MTYPE_TMP
, sizeof(struct cancel_req
));
1424 frr_with_mutex (&m
->mtx
) {
1426 listnode_add(m
->cancel_req
, cr
);
1432 * Cancel any events which have the specified argument.
1436 * @param m the thread_master to cancel from
1437 * @param arg the argument passed when creating the event
1439 void event_cancel_event(struct event_master
*master
, void *arg
)
1441 cancel_event_helper(master
, arg
, 0);
1445 * Cancel ready tasks with an arg matching 'arg'
1449 * @param m the thread_master to cancel from
1450 * @param arg the argument passed when creating the event
1452 void event_cancel_event_ready(struct event_master
*m
, void *arg
)
1455 /* Only cancel ready/event tasks */
1456 cancel_event_helper(m
, arg
, EVENT_CANCEL_FLAG_READY
);
1460 * Cancel a specific task.
1464 * @param thread task to cancel
1466 void event_cancel(struct event
**thread
)
1468 struct event_master
*master
;
1470 if (thread
== NULL
|| *thread
== NULL
)
1473 master
= (*thread
)->master
;
1475 frrtrace(9, frr_libfrr
, event_cancel
, master
, (*thread
)->xref
->funcname
,
1476 (*thread
)->xref
->xref
.file
, (*thread
)->xref
->xref
.line
, NULL
,
1477 (*thread
)->u
.fd
, (*thread
)->u
.val
, (*thread
)->arg
,
1478 (*thread
)->u
.sands
.tv_sec
);
1480 assert(master
->owner
== pthread_self());
1482 frr_with_mutex (&master
->mtx
) {
1483 struct cancel_req
*cr
=
1484 XCALLOC(MTYPE_TMP
, sizeof(struct cancel_req
));
1485 cr
->thread
= *thread
;
1486 listnode_add(master
->cancel_req
, cr
);
1487 do_event_cancel(master
);
1494 * Asynchronous cancellation.
1496 * Called with either a struct event ** or void * to an event argument,
1497 * this function posts the correct cancellation request and blocks until it is
1500 * If the thread is currently running, execution blocks until it completes.
1502 * The last two parameters are mutually exclusive, i.e. if you pass one the
1503 * other must be NULL.
1505 * When the cancellation procedure executes on the target thread_master, the
1506 * thread * provided is checked for nullity. If it is null, the thread is
1507 * assumed to no longer exist and the cancellation request is a no-op. Thus
1508 * users of this API must pass a back-reference when scheduling the original
1513 * @param master the thread master with the relevant event / task
1514 * @param thread pointer to thread to cancel
1515 * @param eventobj the event
1517 void event_cancel_async(struct event_master
*master
, struct event
**thread
,
1520 assert(!(thread
&& eventobj
) && (thread
|| eventobj
));
1522 if (thread
&& *thread
)
1523 frrtrace(9, frr_libfrr
, event_cancel_async
, master
,
1524 (*thread
)->xref
->funcname
, (*thread
)->xref
->xref
.file
,
1525 (*thread
)->xref
->xref
.line
, NULL
, (*thread
)->u
.fd
,
1526 (*thread
)->u
.val
, (*thread
)->arg
,
1527 (*thread
)->u
.sands
.tv_sec
);
1529 frrtrace(9, frr_libfrr
, event_cancel_async
, master
, NULL
, NULL
,
1530 0, NULL
, 0, 0, eventobj
, 0);
1532 assert(master
->owner
!= pthread_self());
1534 frr_with_mutex (&master
->mtx
) {
1535 master
->canceled
= false;
1538 struct cancel_req
*cr
=
1539 XCALLOC(MTYPE_TMP
, sizeof(struct cancel_req
));
1540 cr
->threadref
= thread
;
1541 listnode_add(master
->cancel_req
, cr
);
1542 } else if (eventobj
) {
1543 struct cancel_req
*cr
=
1544 XCALLOC(MTYPE_TMP
, sizeof(struct cancel_req
));
1545 cr
->eventobj
= eventobj
;
1546 listnode_add(master
->cancel_req
, cr
);
1550 while (!master
->canceled
)
1551 pthread_cond_wait(&master
->cancel_cond
, &master
->mtx
);
1557 /* ------------------------------------------------------------------------- */
1559 static struct timeval
*thread_timer_wait(struct event_timer_list_head
*timers
,
1560 struct timeval
*timer_val
)
1562 if (!event_timer_list_count(timers
))
1565 struct event
*next_timer
= event_timer_list_first(timers
);
1566 monotime_until(&next_timer
->u
.sands
, timer_val
);
1570 static struct event
*thread_run(struct event_master
*m
, struct event
*thread
,
1571 struct event
*fetch
)
1574 thread_add_unuse(m
, thread
);
1578 static int thread_process_io_helper(struct event_master
*m
,
1579 struct event
*thread
, short state
,
1580 short actual_state
, int pos
)
1582 struct event
**thread_array
;
1585 * poll() clears the .events field, but the pollfd array we
1586 * pass to poll() is a copy of the one used to schedule threads.
1587 * We need to synchronize state between the two here by applying
1588 * the same changes poll() made on the copy of the "real" pollfd
1591 * This cleans up a possible infinite loop where we refuse
1592 * to respond to a poll event but poll is insistent that
1595 m
->handler
.pfds
[pos
].events
&= ~(state
);
1598 if ((actual_state
& (POLLHUP
|POLLIN
)) != POLLHUP
)
1599 flog_err(EC_LIB_NO_THREAD
,
1600 "Attempting to process an I/O event but for fd: %d(%d) no thread to handle this!",
1601 m
->handler
.pfds
[pos
].fd
, actual_state
);
1605 if (thread
->type
== EVENT_READ
)
1606 thread_array
= m
->read
;
1608 thread_array
= m
->write
;
1610 thread_array
[thread
->u
.fd
] = NULL
;
1611 event_list_add_tail(&m
->ready
, thread
);
1612 thread
->type
= EVENT_READY
;
1618 * Process I/O events.
1620 * Walks through file descriptor array looking for those pollfds whose .revents
1621 * field has something interesting. Deletes any invalid file descriptors.
1623 * @param m the thread master
1624 * @param num the number of active file descriptors (return value of poll())
1626 static void thread_process_io(struct event_master
*m
, unsigned int num
)
1628 unsigned int ready
= 0;
1629 struct pollfd
*pfds
= m
->handler
.copy
;
1631 for (nfds_t i
= 0; i
< m
->handler
.copycount
&& ready
< num
; ++i
) {
1632 /* no event for current fd? immediately continue */
1633 if (pfds
[i
].revents
== 0)
1639 * Unless someone has called event_cancel from another
1640 * pthread, the only thing that could have changed in
1641 * m->handler.pfds while we were asleep is the .events
1642 * field in a given pollfd. Barring event_cancel() that
1643 * value should be a superset of the values we have in our
1644 * copy, so there's no need to update it. Similarily,
1645 * barring deletion, the fd should still be a valid index
1646 * into the master's pfds.
1648 * We are including POLLERR here to do a READ event
1649 * this is because the read should fail and the
1650 * read function should handle it appropriately
1652 if (pfds
[i
].revents
& (POLLIN
| POLLHUP
| POLLERR
)) {
1653 thread_process_io_helper(m
, m
->read
[pfds
[i
].fd
], POLLIN
,
1654 pfds
[i
].revents
, i
);
1656 if (pfds
[i
].revents
& POLLOUT
)
1657 thread_process_io_helper(m
, m
->write
[pfds
[i
].fd
],
1658 POLLOUT
, pfds
[i
].revents
, i
);
1660 /* if one of our file descriptors is garbage, remove the same
1662 * both pfds + update sizes and index */
1663 if (pfds
[i
].revents
& POLLNVAL
) {
1664 memmove(m
->handler
.pfds
+ i
, m
->handler
.pfds
+ i
+ 1,
1665 (m
->handler
.pfdcount
- i
- 1)
1666 * sizeof(struct pollfd
));
1667 m
->handler
.pfdcount
--;
1668 m
->handler
.pfds
[m
->handler
.pfdcount
].fd
= 0;
1669 m
->handler
.pfds
[m
->handler
.pfdcount
].events
= 0;
1671 memmove(pfds
+ i
, pfds
+ i
+ 1,
1672 (m
->handler
.copycount
- i
- 1)
1673 * sizeof(struct pollfd
));
1674 m
->handler
.copycount
--;
1675 m
->handler
.copy
[m
->handler
.copycount
].fd
= 0;
1676 m
->handler
.copy
[m
->handler
.copycount
].events
= 0;
1683 /* Add all timers that have popped to the ready list. */
1684 static unsigned int thread_process_timers(struct event_master
*m
,
1685 struct timeval
*timenow
)
1687 struct timeval prev
= *timenow
;
1688 bool displayed
= false;
1689 struct event
*thread
;
1690 unsigned int ready
= 0;
1692 while ((thread
= event_timer_list_first(&m
->timer
))) {
1693 if (timercmp(timenow
, &thread
->u
.sands
, <))
1695 prev
= thread
->u
.sands
;
1698 * If the timer would have popped 4 seconds in the
1699 * past then we are in a situation where we are
1700 * really getting behind on handling of events.
1701 * Let's log it and do the right thing with it.
1703 if (timercmp(timenow
, &prev
, >)) {
1704 atomic_fetch_add_explicit(
1705 &thread
->hist
->total_starv_warn
, 1,
1706 memory_order_seq_cst
);
1707 if (!displayed
&& !thread
->ignore_timer_late
) {
1709 EC_LIB_STARVE_THREAD
,
1710 "Thread Starvation: %pTHD was scheduled to pop greater than 4s ago",
1716 event_timer_list_pop(&m
->timer
);
1717 thread
->type
= EVENT_READY
;
1718 event_list_add_tail(&m
->ready
, thread
);
1725 /* process a list en masse, e.g. for event thread lists */
1726 static unsigned int thread_process(struct event_list_head
*list
)
1728 struct event
*thread
;
1729 unsigned int ready
= 0;
1731 while ((thread
= event_list_pop(list
))) {
1732 thread
->type
= EVENT_READY
;
1733 event_list_add_tail(&thread
->master
->ready
, thread
);
1740 /* Fetch next ready thread. */
1741 struct event
*event_fetch(struct event_master
*m
, struct event
*fetch
)
1743 struct event
*thread
= NULL
;
1745 struct timeval zerotime
= {0, 0};
1747 struct timeval
*tw
= NULL
;
1748 bool eintr_p
= false;
1752 /* Handle signals if any */
1753 if (m
->handle_signals
)
1754 frr_sigevent_process();
1756 pthread_mutex_lock(&m
->mtx
);
1758 /* Process any pending cancellation requests */
1762 * Attempt to flush ready queue before going into poll().
1763 * This is performance-critical. Think twice before modifying.
1765 if ((thread
= event_list_pop(&m
->ready
))) {
1766 fetch
= thread_run(m
, thread
, fetch
);
1769 pthread_mutex_unlock(&m
->mtx
);
1770 if (!m
->ready_run_loop
)
1771 GETRUSAGE(&m
->last_getrusage
);
1772 m
->ready_run_loop
= true;
1776 m
->ready_run_loop
= false;
1777 /* otherwise, tick through scheduling sequence */
1780 * Post events to ready queue. This must come before the
1781 * following block since events should occur immediately
1783 thread_process(&m
->event
);
1786 * If there are no tasks on the ready queue, we will poll()
1787 * until a timer expires or we receive I/O, whichever comes
1788 * first. The strategy for doing this is:
1790 * - If there are events pending, set the poll() timeout to zero
1791 * - If there are no events pending, but there are timers
1792 * pending, set the timeout to the smallest remaining time on
1794 * - If there are neither timers nor events pending, but there
1795 * are file descriptors pending, block indefinitely in poll()
1796 * - If nothing is pending, it's time for the application to die
1798 * In every case except the last, we need to hit poll() at least
1799 * once per loop to avoid starvation by events
1801 if (!event_list_count(&m
->ready
))
1802 tw
= thread_timer_wait(&m
->timer
, &tv
);
1804 if (event_list_count(&m
->ready
) ||
1805 (tw
&& !timercmp(tw
, &zerotime
, >)))
1808 if (!tw
&& m
->handler
.pfdcount
== 0) { /* die */
1809 pthread_mutex_unlock(&m
->mtx
);
1815 * Copy pollfd array + # active pollfds in it. Not necessary to
1816 * copy the array size as this is fixed.
1818 m
->handler
.copycount
= m
->handler
.pfdcount
;
1819 memcpy(m
->handler
.copy
, m
->handler
.pfds
,
1820 m
->handler
.copycount
* sizeof(struct pollfd
));
1822 pthread_mutex_unlock(&m
->mtx
);
1825 num
= fd_poll(m
, tw
, &eintr_p
);
1827 pthread_mutex_lock(&m
->mtx
);
1829 /* Handle any errors received in poll() */
1832 pthread_mutex_unlock(&m
->mtx
);
1833 /* loop around to signal handler */
1838 flog_err(EC_LIB_SYSTEM_CALL
, "poll() error: %s",
1839 safe_strerror(errno
));
1840 pthread_mutex_unlock(&m
->mtx
);
1845 /* Post timers to ready queue. */
1847 thread_process_timers(m
, &now
);
1849 /* Post I/O to ready queue. */
1851 thread_process_io(m
, num
);
1853 pthread_mutex_unlock(&m
->mtx
);
1855 } while (!thread
&& m
->spin
);
1860 static unsigned long timeval_elapsed(struct timeval a
, struct timeval b
)
1862 return (((a
.tv_sec
- b
.tv_sec
) * TIMER_SECOND_MICRO
)
1863 + (a
.tv_usec
- b
.tv_usec
));
1866 unsigned long event_consumed_time(RUSAGE_T
*now
, RUSAGE_T
*start
,
1867 unsigned long *cputime
)
1869 #ifdef HAVE_CLOCK_THREAD_CPUTIME_ID
1873 * FreeBSD appears to have an issue when calling clock_gettime
1874 * with CLOCK_THREAD_CPUTIME_ID really close to each other
1875 * occassionally the now time will be before the start time.
1876 * This is not good and FRR is ending up with CPU HOG's
1877 * when the subtraction wraps to very large numbers
1879 * What we are going to do here is cheat a little bit
1880 * and notice that this is a problem and just correct
1881 * it so that it is impossible to happen
1883 if (start
->cpu
.tv_sec
== now
->cpu
.tv_sec
&&
1884 start
->cpu
.tv_nsec
> now
->cpu
.tv_nsec
)
1885 now
->cpu
.tv_nsec
= start
->cpu
.tv_nsec
+ 1;
1886 else if (start
->cpu
.tv_sec
> now
->cpu
.tv_sec
) {
1887 now
->cpu
.tv_sec
= start
->cpu
.tv_sec
;
1888 now
->cpu
.tv_nsec
= start
->cpu
.tv_nsec
+ 1;
1891 *cputime
= (now
->cpu
.tv_sec
- start
->cpu
.tv_sec
) * TIMER_SECOND_MICRO
1892 + (now
->cpu
.tv_nsec
- start
->cpu
.tv_nsec
) / 1000;
1894 /* This is 'user + sys' time. */
1895 *cputime
= timeval_elapsed(now
->cpu
.ru_utime
, start
->cpu
.ru_utime
)
1896 + timeval_elapsed(now
->cpu
.ru_stime
, start
->cpu
.ru_stime
);
1898 return timeval_elapsed(now
->real
, start
->real
);
1901 /* We should aim to yield after yield milliseconds, which defaults
1902 to EVENT_YIELD_TIME_SLOT .
1903 Note: we are using real (wall clock) time for this calculation.
1904 It could be argued that CPU time may make more sense in certain
1905 contexts. The things to consider are whether the thread may have
1906 blocked (in which case wall time increases, but CPU time does not),
1907 or whether the system is heavily loaded with other processes competing
1908 for CPU time. On balance, wall clock time seems to make sense.
1909 Plus it has the added benefit that gettimeofday should be faster
1910 than calling getrusage. */
1911 int event_should_yield(struct event
*thread
)
1914 frr_with_mutex (&thread
->mtx
) {
1915 result
= monotime_since(&thread
->real
, NULL
)
1916 > (int64_t)thread
->yield
;
1921 void event_set_yield_time(struct event
*thread
, unsigned long yield_time
)
1923 frr_with_mutex (&thread
->mtx
) {
1924 thread
->yield
= yield_time
;
1928 void event_getrusage(RUSAGE_T
*r
)
1931 if (!cputime_enabled
) {
1932 memset(&r
->cpu
, 0, sizeof(r
->cpu
));
1936 #ifdef HAVE_CLOCK_THREAD_CPUTIME_ID
1937 /* not currently implemented in Linux's vDSO, but maybe at some point
1940 clock_gettime(CLOCK_THREAD_CPUTIME_ID
, &r
->cpu
);
1941 #else /* !HAVE_CLOCK_THREAD_CPUTIME_ID */
1942 #if defined RUSAGE_THREAD
1943 #define FRR_RUSAGE RUSAGE_THREAD
1945 #define FRR_RUSAGE RUSAGE_SELF
1947 getrusage(FRR_RUSAGE
, &(r
->cpu
));
1954 * This function will atomically update the thread's usage history. At present
1955 * this is the only spot where usage history is written. Nevertheless the code
1956 * has been written such that the introduction of writers in the future should
1957 * not need to update it provided the writers atomically perform only the
1958 * operations done here, i.e. updating the total and maximum times. In
1959 * particular, the maximum real and cpu times must be monotonically increasing
1960 * or this code is not correct.
1962 void event_call(struct event
*thread
)
1964 RUSAGE_T before
, after
;
1966 /* if the thread being called is the CLI, it may change cputime_enabled
1967 * ("service cputime-stats" command), which can result in nonsensical
1968 * and very confusing warnings
1970 bool cputime_enabled_here
= cputime_enabled
;
1972 if (thread
->master
->ready_run_loop
)
1973 before
= thread
->master
->last_getrusage
;
1977 thread
->real
= before
.real
;
1979 frrtrace(9, frr_libfrr
, event_call
, thread
->master
,
1980 thread
->xref
->funcname
, thread
->xref
->xref
.file
,
1981 thread
->xref
->xref
.line
, NULL
, thread
->u
.fd
, thread
->u
.val
,
1982 thread
->arg
, thread
->u
.sands
.tv_sec
);
1984 pthread_setspecific(thread_current
, thread
);
1985 (*thread
->func
)(thread
);
1986 pthread_setspecific(thread_current
, NULL
);
1989 thread
->master
->last_getrusage
= after
;
1991 unsigned long walltime
, cputime
;
1994 walltime
= event_consumed_time(&after
, &before
, &cputime
);
1996 /* update walltime */
1997 atomic_fetch_add_explicit(&thread
->hist
->real
.total
, walltime
,
1998 memory_order_seq_cst
);
1999 exp
= atomic_load_explicit(&thread
->hist
->real
.max
,
2000 memory_order_seq_cst
);
2001 while (exp
< walltime
2002 && !atomic_compare_exchange_weak_explicit(
2003 &thread
->hist
->real
.max
, &exp
, walltime
,
2004 memory_order_seq_cst
, memory_order_seq_cst
))
2007 if (cputime_enabled_here
&& cputime_enabled
) {
2008 /* update cputime */
2009 atomic_fetch_add_explicit(&thread
->hist
->cpu
.total
, cputime
,
2010 memory_order_seq_cst
);
2011 exp
= atomic_load_explicit(&thread
->hist
->cpu
.max
,
2012 memory_order_seq_cst
);
2013 while (exp
< cputime
2014 && !atomic_compare_exchange_weak_explicit(
2015 &thread
->hist
->cpu
.max
, &exp
, cputime
,
2016 memory_order_seq_cst
, memory_order_seq_cst
))
2020 atomic_fetch_add_explicit(&thread
->hist
->total_calls
, 1,
2021 memory_order_seq_cst
);
2022 atomic_fetch_or_explicit(&thread
->hist
->types
, 1 << thread
->add_type
,
2023 memory_order_seq_cst
);
2025 if (cputime_enabled_here
&& cputime_enabled
&& cputime_threshold
2026 && cputime
> cputime_threshold
) {
2028 * We have a CPU Hog on our hands. The time FRR has spent
2029 * doing actual work (not sleeping) is greater than 5 seconds.
2030 * Whinge about it now, so we're aware this is yet another task
2033 atomic_fetch_add_explicit(&thread
->hist
->total_cpu_warn
,
2034 1, memory_order_seq_cst
);
2036 EC_LIB_SLOW_THREAD_CPU
,
2037 "CPU HOG: task %s (%lx) ran for %lums (cpu time %lums)",
2038 thread
->xref
->funcname
, (unsigned long)thread
->func
,
2039 walltime
/ 1000, cputime
/ 1000);
2041 } else if (walltime_threshold
&& walltime
> walltime_threshold
) {
2043 * The runtime for a task is greater than 5 seconds, but the
2044 * cpu time is under 5 seconds. Let's whine about this because
2045 * this could imply some sort of scheduling issue.
2047 atomic_fetch_add_explicit(&thread
->hist
->total_wall_warn
,
2048 1, memory_order_seq_cst
);
2050 EC_LIB_SLOW_THREAD_WALL
,
2051 "STARVATION: task %s (%lx) ran for %lums (cpu time %lums)",
2052 thread
->xref
->funcname
, (unsigned long)thread
->func
,
2053 walltime
/ 1000, cputime
/ 1000);
2057 /* Execute thread */
2058 void _event_execute(const struct xref_eventsched
*xref
, struct event_master
*m
,
2059 void (*func
)(struct event
*), void *arg
, int val
)
2061 struct event
*thread
;
2063 /* Get or allocate new thread to execute. */
2064 frr_with_mutex (&m
->mtx
) {
2065 thread
= thread_get(m
, EVENT_EVENT
, func
, arg
, xref
);
2067 /* Set its event value. */
2068 frr_with_mutex (&thread
->mtx
) {
2069 thread
->add_type
= EVENT_EXECUTE
;
2070 thread
->u
.val
= val
;
2071 thread
->ref
= &thread
;
2075 /* Execute thread doing all accounting. */
2078 /* Give back or free thread. */
2079 thread_add_unuse(m
, thread
);
2082 /* Debug signal mask - if 'sigs' is NULL, use current effective mask. */
2083 void debug_signals(const sigset_t
*sigs
)
2090 * We're only looking at the non-realtime signals here, so we need
2091 * some limit value. Platform differences mean at some point we just
2092 * need to pick a reasonable value.
2094 #if defined SIGRTMIN
2095 # define LAST_SIGNAL SIGRTMIN
2097 # define LAST_SIGNAL 32
2102 sigemptyset(&tmpsigs
);
2103 pthread_sigmask(SIG_BLOCK
, NULL
, &tmpsigs
);
2110 for (i
= 0; i
< LAST_SIGNAL
; i
++) {
2113 if (sigismember(sigs
, i
) > 0) {
2115 strlcat(buf
, ",", sizeof(buf
));
2116 snprintf(tmp
, sizeof(tmp
), "%d", i
);
2117 strlcat(buf
, tmp
, sizeof(buf
));
2123 snprintf(buf
, sizeof(buf
), "<none>");
2125 zlog_debug("%s: %s", __func__
, buf
);
2128 static ssize_t
printfrr_thread_dbg(struct fbuf
*buf
, struct printfrr_eargs
*ea
,
2129 const struct event
*thread
)
2131 static const char *const types
[] = {
2132 [EVENT_READ
] = "read", [EVENT_WRITE
] = "write",
2133 [EVENT_TIMER
] = "timer", [EVENT_EVENT
] = "event",
2134 [EVENT_READY
] = "ready", [EVENT_UNUSED
] = "unused",
2135 [EVENT_EXECUTE
] = "exec",
2141 return bputs(buf
, "{(thread *)NULL}");
2143 rv
+= bprintfrr(buf
, "{(thread *)%p arg=%p", thread
, thread
->arg
);
2145 if (thread
->type
< array_size(types
) && types
[thread
->type
])
2146 rv
+= bprintfrr(buf
, " %-6s", types
[thread
->type
]);
2148 rv
+= bprintfrr(buf
, " INVALID(%u)", thread
->type
);
2150 switch (thread
->type
) {
2153 snprintfrr(info
, sizeof(info
), "fd=%d", thread
->u
.fd
);
2157 snprintfrr(info
, sizeof(info
), "r=%pTVMud", &thread
->u
.sands
);
2166 rv
+= bprintfrr(buf
, " %-12s %s() %s from %s:%d}", info
,
2167 thread
->xref
->funcname
, thread
->xref
->dest
,
2168 thread
->xref
->xref
.file
, thread
->xref
->xref
.line
);
2172 printfrr_ext_autoreg_p("TH", printfrr_thread
);
2173 static ssize_t
printfrr_thread(struct fbuf
*buf
, struct printfrr_eargs
*ea
,
2176 const struct event
*thread
= ptr
;
2177 struct timespec remain
= {};
2179 if (ea
->fmt
[0] == 'D') {
2181 return printfrr_thread_dbg(buf
, ea
, thread
);
2185 /* need to jump over time formatting flag characters in the
2186 * input format string, i.e. adjust ea->fmt!
2188 printfrr_time(buf
, ea
, &remain
,
2189 TIMEFMT_TIMER_DEADLINE
| TIMEFMT_SKIP
);
2190 return bputch(buf
, '-');
2193 TIMEVAL_TO_TIMESPEC(&thread
->u
.sands
, &remain
);
2194 return printfrr_time(buf
, ea
, &remain
, TIMEFMT_TIMER_DEADLINE
);