1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c) 2018, Matthew Macy <[email protected]>
5 *
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
8 * are met:
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
14 *
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25 * SUCH DAMAGE.
26 *
27 */
28
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/counter.h>
35 #include <sys/epoch.h>
36 #include <sys/gtaskqueue.h>
37 #include <sys/kernel.h>
38 #include <sys/limits.h>
39 #include <sys/lock.h>
40 #include <sys/malloc.h>
41 #include <sys/mutex.h>
42 #include <sys/pcpu.h>
43 #include <sys/proc.h>
44 #include <sys/sched.h>
45 #include <sys/sx.h>
46 #include <sys/smp.h>
47 #include <sys/sysctl.h>
48 #include <sys/turnstile.h>
49 #ifdef EPOCH_TRACE
50 #include <machine/stdarg.h>
51 #include <sys/stack.h>
52 #include <sys/tree.h>
53 #endif
54 #include <vm/vm.h>
55 #include <vm/vm_extern.h>
56 #include <vm/vm_kern.h>
57 #include <vm/uma.h>
58
59 #include <ck_epoch.h>
60
61 #ifdef __amd64__
62 #define EPOCH_ALIGN CACHE_LINE_SIZE*2
63 #else
64 #define EPOCH_ALIGN CACHE_LINE_SIZE
65 #endif
66
67 TAILQ_HEAD (epoch_tdlist, epoch_tracker);
68 typedef struct epoch_record {
69 ck_epoch_record_t er_record;
70 struct epoch_context er_drain_ctx;
71 struct epoch *er_parent;
72 volatile struct epoch_tdlist er_tdlist;
73 volatile uint32_t er_gen;
74 uint32_t er_cpuid;
75 #ifdef INVARIANTS
76 /* Used to verify record ownership for non-preemptible epochs. */
77 struct thread *er_td;
78 #endif
79 } __aligned(EPOCH_ALIGN) *epoch_record_t;
80
81 struct epoch {
82 struct ck_epoch e_epoch __aligned(EPOCH_ALIGN);
83 epoch_record_t e_pcpu_record;
84 int e_in_use;
85 int e_flags;
86 struct sx e_drain_sx;
87 struct mtx e_drain_mtx;
88 volatile int e_drain_count;
89 const char *e_name;
90 };
91
92 /* arbitrary --- needs benchmarking */
93 #define MAX_ADAPTIVE_SPIN 100
94 #define MAX_EPOCHS 64
95
96 CTASSERT(sizeof(ck_epoch_entry_t) == sizeof(struct epoch_context));
97 SYSCTL_NODE(_kern, OID_AUTO, epoch, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
98 "epoch information");
99 SYSCTL_NODE(_kern_epoch, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
100 "epoch stats");
101
102 /* Stats. */
103 static counter_u64_t block_count;
104
105 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, nblocked, CTLFLAG_RW,
106 &block_count, "# of times a thread was in an epoch when epoch_wait was called");
107 static counter_u64_t migrate_count;
108
109 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, migrations, CTLFLAG_RW,
110 &migrate_count, "# of times thread was migrated to another CPU in epoch_wait");
111 static counter_u64_t turnstile_count;
112
113 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, ncontended, CTLFLAG_RW,
114 &turnstile_count, "# of times a thread was blocked on a lock in an epoch during an epoch_wait");
115 static counter_u64_t switch_count;
116
117 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, switches, CTLFLAG_RW,
118 &switch_count, "# of times a thread voluntarily context switched in epoch_wait");
119 static counter_u64_t epoch_call_count;
120
121 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_calls, CTLFLAG_RW,
122 &epoch_call_count, "# of times a callback was deferred");
123 static counter_u64_t epoch_call_task_count;
124
125 SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, epoch_call_tasks, CTLFLAG_RW,
126 &epoch_call_task_count, "# of times a callback task was run");
127
128 TAILQ_HEAD (threadlist, thread);
129
130 CK_STACK_CONTAINER(struct ck_epoch_entry, stack_entry,
131 ck_epoch_entry_container)
132
133 static struct epoch epoch_array[MAX_EPOCHS];
134
135 DPCPU_DEFINE(struct grouptask, epoch_cb_task);
136 DPCPU_DEFINE(int, epoch_cb_count);
137
138 static __read_mostly int inited;
139 __read_mostly epoch_t global_epoch;
140 __read_mostly epoch_t global_epoch_preempt;
141
142 static void epoch_call_task(void *context __unused);
143 static uma_zone_t pcpu_zone_record;
144
145 static struct sx epoch_sx;
146
147 #define EPOCH_LOCK() sx_xlock(&epoch_sx)
148 #define EPOCH_UNLOCK() sx_xunlock(&epoch_sx)
149
150 #ifdef EPOCH_TRACE
151 struct stackentry {
152 RB_ENTRY(stackentry) se_node;
153 struct stack se_stack;
154 };
155
156 static int
stackentry_compare(struct stackentry * a,struct stackentry * b)157 stackentry_compare(struct stackentry *a, struct stackentry *b)
158 {
159
160 if (a->se_stack.depth > b->se_stack.depth)
161 return (1);
162 if (a->se_stack.depth < b->se_stack.depth)
163 return (-1);
164 for (int i = 0; i < a->se_stack.depth; i++) {
165 if (a->se_stack.pcs[i] > b->se_stack.pcs[i])
166 return (1);
167 if (a->se_stack.pcs[i] < b->se_stack.pcs[i])
168 return (-1);
169 }
170
171 return (0);
172 }
173
174 RB_HEAD(stacktree, stackentry) epoch_stacks = RB_INITIALIZER(&epoch_stacks);
175 RB_GENERATE_STATIC(stacktree, stackentry, se_node, stackentry_compare);
176
177 static struct mtx epoch_stacks_lock;
178 MTX_SYSINIT(epochstacks, &epoch_stacks_lock, "epoch_stacks", MTX_DEF);
179
180 static bool epoch_trace_stack_print = true;
181 SYSCTL_BOOL(_kern_epoch, OID_AUTO, trace_stack_print, CTLFLAG_RWTUN,
182 &epoch_trace_stack_print, 0, "Print stack traces on epoch reports");
183
184 static void epoch_trace_report(const char *fmt, ...) __printflike(1, 2);
185 static inline void
epoch_trace_report(const char * fmt,...)186 epoch_trace_report(const char *fmt, ...)
187 {
188 va_list ap;
189 struct stackentry se, *new;
190
191 stack_zero(&se.se_stack); /* XXX: is it really needed? */
192 stack_save(&se.se_stack);
193
194 /* Tree is never reduced - go lockless. */
195 if (RB_FIND(stacktree, &epoch_stacks, &se) != NULL)
196 return;
197
198 new = malloc(sizeof(*new), M_STACK, M_NOWAIT);
199 if (new != NULL) {
200 bcopy(&se.se_stack, &new->se_stack, sizeof(struct stack));
201
202 mtx_lock(&epoch_stacks_lock);
203 new = RB_INSERT(stacktree, &epoch_stacks, new);
204 mtx_unlock(&epoch_stacks_lock);
205 if (new != NULL)
206 free(new, M_STACK);
207 }
208
209 va_start(ap, fmt);
210 (void)vprintf(fmt, ap);
211 va_end(ap);
212 if (epoch_trace_stack_print)
213 stack_print_ddb(&se.se_stack);
214 }
215
216 static inline void
epoch_trace_enter(struct thread * td,epoch_t epoch,epoch_tracker_t et,const char * file,int line)217 epoch_trace_enter(struct thread *td, epoch_t epoch, epoch_tracker_t et,
218 const char *file, int line)
219 {
220 epoch_tracker_t iet;
221
222 SLIST_FOREACH(iet, &td->td_epochs, et_tlink)
223 if (iet->et_epoch == epoch)
224 epoch_trace_report("Recursively entering epoch %s "
225 "at %s:%d, previously entered at %s:%d\n",
226 epoch->e_name, file, line,
227 iet->et_file, iet->et_line);
228 et->et_epoch = epoch;
229 et->et_file = file;
230 et->et_line = line;
231 SLIST_INSERT_HEAD(&td->td_epochs, et, et_tlink);
232 }
233
234 static inline void
epoch_trace_exit(struct thread * td,epoch_t epoch,epoch_tracker_t et,const char * file,int line)235 epoch_trace_exit(struct thread *td, epoch_t epoch, epoch_tracker_t et,
236 const char *file, int line)
237 {
238
239 if (SLIST_FIRST(&td->td_epochs) != et) {
240 epoch_trace_report("Exiting epoch %s in a not nested order "
241 "at %s:%d. Most recently entered %s at %s:%d\n",
242 epoch->e_name,
243 file, line,
244 SLIST_FIRST(&td->td_epochs)->et_epoch->e_name,
245 SLIST_FIRST(&td->td_epochs)->et_file,
246 SLIST_FIRST(&td->td_epochs)->et_line);
247 /* This will panic if et is not anywhere on td_epochs. */
248 SLIST_REMOVE(&td->td_epochs, et, epoch_tracker, et_tlink);
249 } else
250 SLIST_REMOVE_HEAD(&td->td_epochs, et_tlink);
251 }
252
253 /* Used by assertions that check thread state before going to sleep. */
254 void
epoch_trace_list(struct thread * td)255 epoch_trace_list(struct thread *td)
256 {
257 epoch_tracker_t iet;
258
259 SLIST_FOREACH(iet, &td->td_epochs, et_tlink)
260 printf("Epoch %s entered at %s:%d\n", iet->et_epoch->e_name,
261 iet->et_file, iet->et_line);
262 }
263 #endif /* EPOCH_TRACE */
264
265 static void
epoch_init(void * arg __unused)266 epoch_init(void *arg __unused)
267 {
268 int cpu;
269
270 block_count = counter_u64_alloc(M_WAITOK);
271 migrate_count = counter_u64_alloc(M_WAITOK);
272 turnstile_count = counter_u64_alloc(M_WAITOK);
273 switch_count = counter_u64_alloc(M_WAITOK);
274 epoch_call_count = counter_u64_alloc(M_WAITOK);
275 epoch_call_task_count = counter_u64_alloc(M_WAITOK);
276
277 pcpu_zone_record = uma_zcreate("epoch_record pcpu",
278 sizeof(struct epoch_record), NULL, NULL, NULL, NULL,
279 UMA_ALIGN_PTR, UMA_ZONE_PCPU);
280 CPU_FOREACH(cpu) {
281 GROUPTASK_INIT(DPCPU_ID_PTR(cpu, epoch_cb_task), 0,
282 epoch_call_task, NULL);
283 #ifndef FSTACK
284 taskqgroup_attach_cpu(qgroup_softirq,
285 DPCPU_ID_PTR(cpu, epoch_cb_task), NULL, cpu, NULL, NULL,
286 "epoch call task");
287 #endif
288 }
289 #ifdef EPOCH_TRACE
290 SLIST_INIT(&thread0.td_epochs);
291 #endif
292 sx_init(&epoch_sx, "epoch-sx");
293 inited = 1;
294 global_epoch = epoch_alloc("Global", 0);
295 global_epoch_preempt = epoch_alloc("Global preemptible", EPOCH_PREEMPT);
296 }
297 SYSINIT(epoch, SI_SUB_EPOCH, SI_ORDER_FIRST, epoch_init, NULL);
298
299 #if !defined(EARLY_AP_STARTUP)
300 static void
epoch_init_smp(void * dummy __unused)301 epoch_init_smp(void *dummy __unused)
302 {
303 inited = 2;
304 }
305 SYSINIT(epoch_smp, SI_SUB_SMP + 1, SI_ORDER_FIRST, epoch_init_smp, NULL);
306 #endif
307
308 static void
epoch_ctor(epoch_t epoch)309 epoch_ctor(epoch_t epoch)
310 {
311 epoch_record_t er;
312 int cpu;
313
314 epoch->e_pcpu_record = uma_zalloc_pcpu(pcpu_zone_record, M_WAITOK);
315 CPU_FOREACH(cpu) {
316 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
317 bzero(er, sizeof(*er));
318 ck_epoch_register(&epoch->e_epoch, &er->er_record, NULL);
319 TAILQ_INIT((struct threadlist *)(uintptr_t)&er->er_tdlist);
320 er->er_cpuid = cpu;
321 er->er_parent = epoch;
322 }
323 }
324
325 static void
epoch_adjust_prio(struct thread * td,u_char prio)326 epoch_adjust_prio(struct thread *td, u_char prio)
327 {
328
329 thread_lock(td);
330 sched_prio(td, prio);
331 thread_unlock(td);
332 }
333
334 epoch_t
epoch_alloc(const char * name,int flags)335 epoch_alloc(const char *name, int flags)
336 {
337 epoch_t epoch;
338 int i;
339
340 MPASS(name != NULL);
341
342 if (__predict_false(!inited))
343 panic("%s called too early in boot", __func__);
344
345 EPOCH_LOCK();
346
347 /*
348 * Find a free index in the epoch array. If no free index is
349 * found, try to use the index after the last one.
350 */
351 for (i = 0;; i++) {
352 /*
353 * If too many epochs are currently allocated,
354 * return NULL.
355 */
356 if (i == MAX_EPOCHS) {
357 epoch = NULL;
358 goto done;
359 }
360 if (epoch_array[i].e_in_use == 0)
361 break;
362 }
363
364 epoch = epoch_array + i;
365 ck_epoch_init(&epoch->e_epoch);
366 epoch_ctor(epoch);
367 epoch->e_flags = flags;
368 epoch->e_name = name;
369 sx_init(&epoch->e_drain_sx, "epoch-drain-sx");
370 mtx_init(&epoch->e_drain_mtx, "epoch-drain-mtx", NULL, MTX_DEF);
371
372 /*
373 * Set e_in_use last, because when this field is set the
374 * epoch_call_task() function will start scanning this epoch
375 * structure.
376 */
377 atomic_store_rel_int(&epoch->e_in_use, 1);
378 done:
379 EPOCH_UNLOCK();
380 return (epoch);
381 }
382
383 void
epoch_free(epoch_t epoch)384 epoch_free(epoch_t epoch)
385 {
386 #ifdef INVARIANTS
387 int cpu;
388 #endif
389
390 EPOCH_LOCK();
391
392 MPASS(epoch->e_in_use != 0);
393
394 epoch_drain_callbacks(epoch);
395
396 atomic_store_rel_int(&epoch->e_in_use, 0);
397 /*
398 * Make sure the epoch_call_task() function see e_in_use equal
399 * to zero, by calling epoch_wait() on the global_epoch:
400 */
401 epoch_wait(global_epoch);
402 #ifdef INVARIANTS
403 CPU_FOREACH(cpu) {
404 epoch_record_t er;
405
406 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
407
408 /*
409 * Sanity check: none of the records should be in use anymore.
410 * We drained callbacks above and freeing the pcpu records is
411 * imminent.
412 */
413 MPASS(er->er_td == NULL);
414 MPASS(TAILQ_EMPTY(&er->er_tdlist));
415 }
416 #endif
417 uma_zfree_pcpu(pcpu_zone_record, epoch->e_pcpu_record);
418 mtx_destroy(&epoch->e_drain_mtx);
419 sx_destroy(&epoch->e_drain_sx);
420 memset(epoch, 0, sizeof(*epoch));
421
422 EPOCH_UNLOCK();
423 }
424
425 static epoch_record_t
epoch_currecord(epoch_t epoch)426 epoch_currecord(epoch_t epoch)
427 {
428
429 return (zpcpu_get(epoch->e_pcpu_record));
430 }
431
432 #define INIT_CHECK(epoch) \
433 do { \
434 if (__predict_false((epoch) == NULL)) \
435 return; \
436 } while (0)
437
438 void
_epoch_enter_preempt(epoch_t epoch,epoch_tracker_t et EPOCH_FILE_LINE)439 _epoch_enter_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
440 {
441 struct epoch_record *er;
442 struct thread *td;
443
444 MPASS(cold || epoch != NULL);
445 MPASS(epoch->e_flags & EPOCH_PREEMPT);
446 td = curthread;
447 MPASS((vm_offset_t)et >= td->td_kstack &&
448 (vm_offset_t)et + sizeof(struct epoch_tracker) <=
449 td->td_kstack + td->td_kstack_pages * PAGE_SIZE);
450
451 INIT_CHECK(epoch);
452 #ifdef EPOCH_TRACE
453 epoch_trace_enter(td, epoch, et, file, line);
454 #endif
455 et->et_td = td;
456 THREAD_NO_SLEEPING();
457 critical_enter();
458 sched_pin();
459 td->td_pre_epoch_prio = td->td_priority;
460 er = epoch_currecord(epoch);
461 /* Record-level tracking is reserved for non-preemptible epochs. */
462 MPASS(er->er_td == NULL);
463 TAILQ_INSERT_TAIL(&er->er_tdlist, et, et_link);
464 ck_epoch_begin(&er->er_record, &et->et_section);
465 critical_exit();
466 }
467
468 void
epoch_enter(epoch_t epoch)469 epoch_enter(epoch_t epoch)
470 {
471 epoch_record_t er;
472
473 MPASS(cold || epoch != NULL);
474 INIT_CHECK(epoch);
475 critical_enter();
476 er = epoch_currecord(epoch);
477 #ifdef INVARIANTS
478 if (er->er_record.active == 0) {
479 MPASS(er->er_td == NULL);
480 er->er_td = curthread;
481 } else {
482 /* We've recursed, just make sure our accounting isn't wrong. */
483 MPASS(er->er_td == curthread);
484 }
485 #endif
486 ck_epoch_begin(&er->er_record, NULL);
487 }
488
489 void
_epoch_exit_preempt(epoch_t epoch,epoch_tracker_t et EPOCH_FILE_LINE)490 _epoch_exit_preempt(epoch_t epoch, epoch_tracker_t et EPOCH_FILE_LINE)
491 {
492 struct epoch_record *er;
493 struct thread *td;
494
495 INIT_CHECK(epoch);
496 td = curthread;
497 critical_enter();
498 sched_unpin();
499 THREAD_SLEEPING_OK();
500 er = epoch_currecord(epoch);
501 MPASS(epoch->e_flags & EPOCH_PREEMPT);
502 MPASS(et != NULL);
503 MPASS(et->et_td == td);
504 #ifdef INVARIANTS
505 et->et_td = (void*)0xDEADBEEF;
506 /* Record-level tracking is reserved for non-preemptible epochs. */
507 MPASS(er->er_td == NULL);
508 #endif
509 ck_epoch_end(&er->er_record, &et->et_section);
510 TAILQ_REMOVE(&er->er_tdlist, et, et_link);
511 er->er_gen++;
512 if (__predict_false(td->td_pre_epoch_prio != td->td_priority))
513 epoch_adjust_prio(td, td->td_pre_epoch_prio);
514 critical_exit();
515 #ifdef EPOCH_TRACE
516 epoch_trace_exit(td, epoch, et, file, line);
517 #endif
518 }
519
520 void
epoch_exit(epoch_t epoch)521 epoch_exit(epoch_t epoch)
522 {
523 epoch_record_t er;
524
525 INIT_CHECK(epoch);
526 er = epoch_currecord(epoch);
527 ck_epoch_end(&er->er_record, NULL);
528 #ifdef INVARIANTS
529 MPASS(er->er_td == curthread);
530 if (er->er_record.active == 0)
531 er->er_td = NULL;
532 #endif
533 critical_exit();
534 }
535
536 /*
537 * epoch_block_handler_preempt() is a callback from the CK code when another
538 * thread is currently in an epoch section.
539 */
540 static void
epoch_block_handler_preempt(struct ck_epoch * global __unused,ck_epoch_record_t * cr,void * arg __unused)541 epoch_block_handler_preempt(struct ck_epoch *global __unused,
542 ck_epoch_record_t *cr, void *arg __unused)
543 {
544 epoch_record_t record;
545 struct thread *td, *owner, *curwaittd;
546 struct epoch_tracker *tdwait;
547 struct turnstile *ts;
548 struct lock_object *lock;
549 int spincount, gen;
550 int locksheld __unused;
551
552 record = __containerof(cr, struct epoch_record, er_record);
553 td = curthread;
554 locksheld = td->td_locks;
555 spincount = 0;
556 counter_u64_add(block_count, 1);
557 /*
558 * We lost a race and there's no longer any threads
559 * on the CPU in an epoch section.
560 */
561 if (TAILQ_EMPTY(&record->er_tdlist))
562 return;
563
564 if (record->er_cpuid != curcpu) {
565 /*
566 * If the head of the list is running, we can wait for it
567 * to remove itself from the list and thus save us the
568 * overhead of a migration
569 */
570 gen = record->er_gen;
571 thread_unlock(td);
572 /*
573 * We can't actually check if the waiting thread is running
574 * so we simply poll for it to exit before giving up and
575 * migrating.
576 */
577 do {
578 cpu_spinwait();
579 } while (!TAILQ_EMPTY(&record->er_tdlist) &&
580 gen == record->er_gen &&
581 spincount++ < MAX_ADAPTIVE_SPIN);
582 thread_lock(td);
583 /*
584 * If the generation has changed we can poll again
585 * otherwise we need to migrate.
586 */
587 if (gen != record->er_gen)
588 return;
589 /*
590 * Being on the same CPU as that of the record on which
591 * we need to wait allows us access to the thread
592 * list associated with that CPU. We can then examine the
593 * oldest thread in the queue and wait on its turnstile
594 * until it resumes and so on until a grace period
595 * elapses.
596 *
597 */
598 counter_u64_add(migrate_count, 1);
599 sched_bind(td, record->er_cpuid);
600 /*
601 * At this point we need to return to the ck code
602 * to scan to see if a grace period has elapsed.
603 * We can't move on to check the thread list, because
604 * in the meantime new threads may have arrived that
605 * in fact belong to a different epoch.
606 */
607 return;
608 }
609 /*
610 * Try to find a thread in an epoch section on this CPU
611 * waiting on a turnstile. Otherwise find the lowest
612 * priority thread (highest prio value) and drop our priority
613 * to match to allow it to run.
614 */
615 TAILQ_FOREACH(tdwait, &record->er_tdlist, et_link) {
616 /*
617 * Propagate our priority to any other waiters to prevent us
618 * from starving them. They will have their original priority
619 * restore on exit from epoch_wait().
620 */
621 curwaittd = tdwait->et_td;
622 if (!TD_IS_INHIBITED(curwaittd) && curwaittd->td_priority > td->td_priority) {
623 critical_enter();
624 thread_unlock(td);
625 thread_lock(curwaittd);
626 sched_prio(curwaittd, td->td_priority);
627 thread_unlock(curwaittd);
628 thread_lock(td);
629 critical_exit();
630 }
631 if (TD_IS_INHIBITED(curwaittd) && TD_ON_LOCK(curwaittd) &&
632 ((ts = curwaittd->td_blocked) != NULL)) {
633 /*
634 * We unlock td to allow turnstile_wait to reacquire
635 * the thread lock. Before unlocking it we enter a
636 * critical section to prevent preemption after we
637 * reenable interrupts by dropping the thread lock in
638 * order to prevent curwaittd from getting to run.
639 */
640 critical_enter();
641 thread_unlock(td);
642
643 if (turnstile_lock(ts, &lock, &owner)) {
644 if (ts == curwaittd->td_blocked) {
645 MPASS(TD_IS_INHIBITED(curwaittd) &&
646 TD_ON_LOCK(curwaittd));
647 critical_exit();
648 turnstile_wait(ts, owner,
649 curwaittd->td_tsqueue);
650 counter_u64_add(turnstile_count, 1);
651 thread_lock(td);
652 return;
653 }
654 turnstile_unlock(ts, lock);
655 }
656 thread_lock(td);
657 critical_exit();
658 KASSERT(td->td_locks == locksheld,
659 ("%d extra locks held", td->td_locks - locksheld));
660 }
661 }
662 /*
663 * We didn't find any threads actually blocked on a lock
664 * so we have nothing to do except context switch away.
665 */
666 counter_u64_add(switch_count, 1);
667 mi_switch(SW_VOL | SWT_RELINQUISH);
668 /*
669 * It is important the thread lock is dropped while yielding
670 * to allow other threads to acquire the lock pointed to by
671 * TDQ_LOCKPTR(td). Currently mi_switch() will unlock the
672 * thread lock before returning. Else a deadlock like
673 * situation might happen.
674 */
675 thread_lock(td);
676 }
677
678 void
epoch_wait_preempt(epoch_t epoch)679 epoch_wait_preempt(epoch_t epoch)
680 {
681 struct thread *td;
682 int was_bound;
683 int old_cpu;
684 int old_pinned;
685 u_char old_prio;
686 int locks __unused;
687
688 MPASS(cold || epoch != NULL);
689 INIT_CHECK(epoch);
690 td = curthread;
691 #ifdef INVARIANTS
692 locks = curthread->td_locks;
693 MPASS(epoch->e_flags & EPOCH_PREEMPT);
694 if ((epoch->e_flags & EPOCH_LOCKED) == 0)
695 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
696 "epoch_wait() can be long running");
697 KASSERT(!in_epoch(epoch), ("epoch_wait_preempt() called in the middle "
698 "of an epoch section of the same epoch"));
699 #endif
700 DROP_GIANT();
701 thread_lock(td);
702
703 old_cpu = PCPU_GET(cpuid);
704 old_pinned = td->td_pinned;
705 old_prio = td->td_priority;
706 was_bound = sched_is_bound(td);
707 sched_unbind(td);
708 td->td_pinned = 0;
709 sched_bind(td, old_cpu);
710
711 ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler_preempt,
712 NULL);
713
714 /* restore CPU binding, if any */
715 if (was_bound != 0) {
716 sched_bind(td, old_cpu);
717 } else {
718 /* get thread back to initial CPU, if any */
719 if (old_pinned != 0)
720 sched_bind(td, old_cpu);
721 sched_unbind(td);
722 }
723 /* restore pinned after bind */
724 td->td_pinned = old_pinned;
725
726 /* restore thread priority */
727 sched_prio(td, old_prio);
728 thread_unlock(td);
729 PICKUP_GIANT();
730 KASSERT(td->td_locks == locks,
731 ("%d residual locks held", td->td_locks - locks));
732 }
733
734 static void
epoch_block_handler(struct ck_epoch * g __unused,ck_epoch_record_t * c __unused,void * arg __unused)735 epoch_block_handler(struct ck_epoch *g __unused, ck_epoch_record_t *c __unused,
736 void *arg __unused)
737 {
738 cpu_spinwait();
739 }
740
741 void
epoch_wait(epoch_t epoch)742 epoch_wait(epoch_t epoch)
743 {
744
745 MPASS(cold || epoch != NULL);
746 INIT_CHECK(epoch);
747 MPASS(epoch->e_flags == 0);
748 critical_enter();
749 ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler, NULL);
750 critical_exit();
751 }
752
753 void
epoch_call(epoch_t epoch,epoch_callback_t callback,epoch_context_t ctx)754 epoch_call(epoch_t epoch, epoch_callback_t callback, epoch_context_t ctx)
755 {
756 epoch_record_t er;
757 ck_epoch_entry_t *cb;
758
759 cb = (void *)ctx;
760
761 MPASS(callback);
762 /* too early in boot to have epoch set up */
763 if (__predict_false(epoch == NULL))
764 goto boottime;
765 #if !defined(EARLY_AP_STARTUP)
766 if (__predict_false(inited < 2))
767 goto boottime;
768 #endif
769
770 critical_enter();
771 *DPCPU_PTR(epoch_cb_count) += 1;
772 er = epoch_currecord(epoch);
773 ck_epoch_call(&er->er_record, cb, (ck_epoch_cb_t *)callback);
774 critical_exit();
775 return;
776 boottime:
777 callback(ctx);
778 }
779
780 static void
epoch_call_task(void * arg __unused)781 epoch_call_task(void *arg __unused)
782 {
783 ck_stack_entry_t *cursor, *head, *next;
784 ck_epoch_record_t *record;
785 epoch_record_t er;
786 epoch_t epoch;
787 ck_stack_t cb_stack;
788 int i, npending, total;
789
790 ck_stack_init(&cb_stack);
791 critical_enter();
792 epoch_enter(global_epoch);
793 for (total = i = 0; i != MAX_EPOCHS; i++) {
794 epoch = epoch_array + i;
795 if (__predict_false(
796 atomic_load_acq_int(&epoch->e_in_use) == 0))
797 continue;
798 er = epoch_currecord(epoch);
799 record = &er->er_record;
800 if ((npending = record->n_pending) == 0)
801 continue;
802 ck_epoch_poll_deferred(record, &cb_stack);
803 total += npending - record->n_pending;
804 }
805 epoch_exit(global_epoch);
806 *DPCPU_PTR(epoch_cb_count) -= total;
807 critical_exit();
808
809 counter_u64_add(epoch_call_count, total);
810 counter_u64_add(epoch_call_task_count, 1);
811
812 head = ck_stack_batch_pop_npsc(&cb_stack);
813 for (cursor = head; cursor != NULL; cursor = next) {
814 struct ck_epoch_entry *entry =
815 ck_epoch_entry_container(cursor);
816
817 next = CK_STACK_NEXT(cursor);
818 entry->function(entry);
819 }
820 }
821
822 static int
in_epoch_verbose_preempt(epoch_t epoch,int dump_onfail)823 in_epoch_verbose_preempt(epoch_t epoch, int dump_onfail)
824 {
825 epoch_record_t er;
826 struct epoch_tracker *tdwait;
827 struct thread *td;
828
829 MPASS(epoch != NULL);
830 MPASS((epoch->e_flags & EPOCH_PREEMPT) != 0);
831 td = curthread;
832 if (THREAD_CAN_SLEEP())
833 return (0);
834 critical_enter();
835 er = epoch_currecord(epoch);
836 TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
837 if (tdwait->et_td == td) {
838 critical_exit();
839 return (1);
840 }
841 #ifdef INVARIANTS
842 if (dump_onfail) {
843 MPASS(td->td_pinned);
844 printf("cpu: %d id: %d\n", curcpu, td->td_tid);
845 TAILQ_FOREACH(tdwait, &er->er_tdlist, et_link)
846 printf("td_tid: %d ", tdwait->et_td->td_tid);
847 printf("\n");
848 }
849 #endif
850 critical_exit();
851 return (0);
852 }
853
854 #ifdef INVARIANTS
855 static void
epoch_assert_nocpu(epoch_t epoch,struct thread * td)856 epoch_assert_nocpu(epoch_t epoch, struct thread *td)
857 {
858 epoch_record_t er;
859 int cpu;
860 bool crit;
861
862 crit = td->td_critnest > 0;
863
864 /* Check for a critical section mishap. */
865 CPU_FOREACH(cpu) {
866 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
867 KASSERT(er->er_td != td,
868 ("%s critical section in epoch '%s', from cpu %d",
869 (crit ? "exited" : "re-entered"), epoch->e_name, cpu));
870 }
871 }
872 #else
873 #define epoch_assert_nocpu(e, td)
874 #endif
875
876 int
in_epoch_verbose(epoch_t epoch,int dump_onfail)877 in_epoch_verbose(epoch_t epoch, int dump_onfail)
878 {
879 epoch_record_t er;
880 struct thread *td;
881
882 if (__predict_false((epoch) == NULL))
883 return (0);
884 if ((epoch->e_flags & EPOCH_PREEMPT) != 0)
885 return (in_epoch_verbose_preempt(epoch, dump_onfail));
886
887 /*
888 * The thread being in a critical section is a necessary
889 * condition to be correctly inside a non-preemptible epoch,
890 * so it's definitely not in this epoch.
891 */
892 td = curthread;
893 if (td->td_critnest == 0) {
894 epoch_assert_nocpu(epoch, td);
895 return (0);
896 }
897
898 /*
899 * The current cpu is in a critical section, so the epoch record will be
900 * stable for the rest of this function. Knowing that the record is not
901 * active is sufficient for knowing whether we're in this epoch or not,
902 * since it's a pcpu record.
903 */
904 er = epoch_currecord(epoch);
905 if (er->er_record.active == 0) {
906 epoch_assert_nocpu(epoch, td);
907 return (0);
908 }
909
910 MPASS(er->er_td == td);
911 return (1);
912 }
913
914 int
in_epoch(epoch_t epoch)915 in_epoch(epoch_t epoch)
916 {
917 return (in_epoch_verbose(epoch, 0));
918 }
919
920 static void
epoch_drain_cb(struct epoch_context * ctx)921 epoch_drain_cb(struct epoch_context *ctx)
922 {
923 struct epoch *epoch =
924 __containerof(ctx, struct epoch_record, er_drain_ctx)->er_parent;
925
926 if (atomic_fetchadd_int(&epoch->e_drain_count, -1) == 1) {
927 mtx_lock(&epoch->e_drain_mtx);
928 wakeup(epoch);
929 mtx_unlock(&epoch->e_drain_mtx);
930 }
931 }
932
933 void
epoch_drain_callbacks(epoch_t epoch)934 epoch_drain_callbacks(epoch_t epoch)
935 {
936 epoch_record_t er;
937 struct thread *td;
938 int was_bound;
939 int old_pinned;
940 int old_cpu;
941 int cpu;
942
943 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
944 "epoch_drain_callbacks() may sleep!");
945
946 /* too early in boot to have epoch set up */
947 if (__predict_false(epoch == NULL))
948 return;
949 #if !defined(EARLY_AP_STARTUP)
950 if (__predict_false(inited < 2))
951 return;
952 #endif
953 DROP_GIANT();
954
955 sx_xlock(&epoch->e_drain_sx);
956 mtx_lock(&epoch->e_drain_mtx);
957
958 td = curthread;
959 thread_lock(td);
960 old_cpu = PCPU_GET(cpuid);
961 old_pinned = td->td_pinned;
962 was_bound = sched_is_bound(td);
963 sched_unbind(td);
964 td->td_pinned = 0;
965
966 CPU_FOREACH(cpu)
967 epoch->e_drain_count++;
968 CPU_FOREACH(cpu) {
969 er = zpcpu_get_cpu(epoch->e_pcpu_record, cpu);
970 sched_bind(td, cpu);
971 epoch_call(epoch, &epoch_drain_cb, &er->er_drain_ctx);
972 }
973
974 /* restore CPU binding, if any */
975 if (was_bound != 0) {
976 sched_bind(td, old_cpu);
977 } else {
978 /* get thread back to initial CPU, if any */
979 if (old_pinned != 0)
980 sched_bind(td, old_cpu);
981 sched_unbind(td);
982 }
983 /* restore pinned after bind */
984 td->td_pinned = old_pinned;
985
986 thread_unlock(td);
987
988 while (epoch->e_drain_count != 0)
989 msleep(epoch, &epoch->e_drain_mtx, PZERO, "EDRAIN", 0);
990
991 mtx_unlock(&epoch->e_drain_mtx);
992 sx_xunlock(&epoch->e_drain_sx);
993
994 PICKUP_GIANT();
995 }
996