92 "Number of CPUs online");
93
94int smp_topology = 0; /* Which topology we're using. */
95SYSCTL_INT(_kern_smp, OID_AUTO, topology, CTLFLAG_RD, &smp_topology, 0,
96 "Topology override setting; 0 is default provided by hardware.");
97TUNABLE_INT("kern.smp.topology", &smp_topology);
98
99#ifdef SMP
100/* Enable forwarding of a signal to a process running on a different CPU */
101static int forward_signal_enabled = 1;
102SYSCTL_INT(_kern_smp, OID_AUTO, forward_signal_enabled, CTLFLAG_RW,
103 &forward_signal_enabled, 0,
104 "Forwarding of a signal to a process on a different CPU");
105
106/* Variables needed for SMP rendezvous. */
107static volatile int smp_rv_ncpus;
108static void (*volatile smp_rv_setup_func)(void *arg);
109static void (*volatile smp_rv_action_func)(void *arg);
110static void (*volatile smp_rv_teardown_func)(void *arg);
111static void *volatile smp_rv_func_arg;
112static volatile int smp_rv_waiters[3];
113static volatile int smp_rv_generation;
114
115/*
116 * Shared mutex to restrict busywaits between smp_rendezvous() and
117 * smp(_targeted)_tlb_shootdown(). A deadlock occurs if both of these
118 * functions trigger at once and cause multiple CPUs to busywait with
119 * interrupts disabled.
120 */
121struct mtx smp_ipi_mtx;
122
123/*
124 * Let the MD SMP code initialize mp_maxid very early if it can.
125 */
126static void
127mp_setmaxid(void *dummy)
128{
129 cpu_mp_setmaxid();
130}
131SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL);
132
133/*
134 * Call the MD SMP initialization code.
135 */
136static void
137mp_start(void *dummy)
138{
139
140 mtx_init(&smp_ipi_mtx, "smp rendezvous", NULL, MTX_SPIN);
141
142 /* Probe for MP hardware. */
143 if (smp_disabled != 0 || cpu_mp_probe() == 0) {
144 mp_ncpus = 1;
145 CPU_SETOF(PCPU_GET(cpuid), &all_cpus);
146 return;
147 }
148
149 cpu_mp_start();
150 printf("FreeBSD/SMP: Multiprocessor System Detected: %d CPUs\n",
151 mp_ncpus);
152 cpu_mp_announce();
153}
154SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL);
155
156void
157forward_signal(struct thread *td)
158{
159 int id;
160
161 /*
162 * signotify() has already set TDF_ASTPENDING and TDF_NEEDSIGCHECK on
163 * this thread, so all we need to do is poke it if it is currently
164 * executing so that it executes ast().
165 */
166 THREAD_LOCK_ASSERT(td, MA_OWNED);
167 KASSERT(TD_IS_RUNNING(td),
168 ("forward_signal: thread is not TDS_RUNNING"));
169
170 CTR1(KTR_SMP, "forward_signal(%p)", td->td_proc);
171
172 if (!smp_started || cold || panicstr)
173 return;
174 if (!forward_signal_enabled)
175 return;
176
177 /* No need to IPI ourself. */
178 if (td == curthread)
179 return;
180
181 id = td->td_oncpu;
182 if (id == NOCPU)
183 return;
184 ipi_cpu(id, IPI_AST);
185}
186
187/*
188 * When called the executing CPU will send an IPI to all other CPUs
189 * requesting that they halt execution.
190 *
191 * Usually (but not necessarily) called with 'other_cpus' as its arg.
192 *
193 * - Signals all CPUs in map to stop.
194 * - Waits for each to stop.
195 *
196 * Returns:
197 * -1: error
198 * 0: NA
199 * 1: ok
200 *
201 */
202static int
203generic_stop_cpus(cpuset_t map, u_int type)
204{
205#ifdef KTR
206 char cpusetbuf[CPUSETBUFSIZ];
207#endif
208 static volatile u_int stopping_cpu = NOCPU;
209 int i;
210
211 KASSERT(
212#if defined(__amd64__)
213 type == IPI_STOP || type == IPI_STOP_HARD || type == IPI_SUSPEND,
214#else
215 type == IPI_STOP || type == IPI_STOP_HARD,
216#endif
217 ("%s: invalid stop type", __func__));
218
219 if (!smp_started)
220 return (0);
221
222 CTR2(KTR_SMP, "stop_cpus(%s) with %u type",
223 cpusetobj_strprint(cpusetbuf, &map), type);
224
225 if (stopping_cpu != PCPU_GET(cpuid))
226 while (atomic_cmpset_int(&stopping_cpu, NOCPU,
227 PCPU_GET(cpuid)) == 0)
228 while (stopping_cpu != NOCPU)
229 cpu_spinwait(); /* spin */
230
231 /* send the stop IPI to all CPUs in map */
232 ipi_selected(map, type);
233
234 i = 0;
235 while (!CPU_SUBSET(&stopped_cpus, &map)) {
236 /* spin */
237 cpu_spinwait();
238 i++;
239 if (i == 100000000) {
240 printf("timeout stopping cpus\n");
241 break;
242 }
243 }
244
245 stopping_cpu = NOCPU;
246 return (1);
247}
248
249int
250stop_cpus(cpuset_t map)
251{
252
253 return (generic_stop_cpus(map, IPI_STOP));
254}
255
256int
257stop_cpus_hard(cpuset_t map)
258{
259
260 return (generic_stop_cpus(map, IPI_STOP_HARD));
261}
262
263#if defined(__amd64__)
264int
265suspend_cpus(cpuset_t map)
266{
267
268 return (generic_stop_cpus(map, IPI_SUSPEND));
269}
270#endif
271
272/*
273 * Called by a CPU to restart stopped CPUs.
274 *
275 * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
276 *
277 * - Signals all CPUs in map to restart.
278 * - Waits for each to restart.
279 *
280 * Returns:
281 * -1: error
282 * 0: NA
283 * 1: ok
284 */
285int
286restart_cpus(cpuset_t map)
287{
288#ifdef KTR
289 char cpusetbuf[CPUSETBUFSIZ];
290#endif
291
292 if (!smp_started)
293 return 0;
294
295 CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map));
296
297 /* signal other cpus to restart */
298 CPU_COPY_STORE_REL(&map, &started_cpus);
299
300 /* wait for each to clear its bit */
301 while (CPU_OVERLAP(&stopped_cpus, &map))
302 cpu_spinwait();
303
304 return 1;
305}
306
307/*
308 * All-CPU rendezvous. CPUs are signalled, all execute the setup function
309 * (if specified), rendezvous, execute the action function (if specified),
310 * rendezvous again, execute the teardown function (if specified), and then
311 * resume.
312 *
313 * Note that the supplied external functions _must_ be reentrant and aware
314 * that they are running in parallel and in an unknown lock context.
315 */
316void
317smp_rendezvous_action(void)
318{
319 struct thread *td;
320 void *local_func_arg;
321 void (*local_setup_func)(void*);
322 void (*local_action_func)(void*);
323 void (*local_teardown_func)(void*);
324 int generation;
325#ifdef INVARIANTS
326 int owepreempt;
327#endif
328
329 /* Ensure we have up-to-date values. */
330 atomic_add_acq_int(&smp_rv_waiters[0], 1);
331 while (smp_rv_waiters[0] < smp_rv_ncpus)
332 cpu_spinwait();
333
334 /* Fetch rendezvous parameters after acquire barrier. */
335 local_func_arg = smp_rv_func_arg;
336 local_setup_func = smp_rv_setup_func;
337 local_action_func = smp_rv_action_func;
338 local_teardown_func = smp_rv_teardown_func;
339 generation = smp_rv_generation;
340
341 /*
342 * Use a nested critical section to prevent any preemptions
343 * from occurring during a rendezvous action routine.
344 * Specifically, if a rendezvous handler is invoked via an IPI
345 * and the interrupted thread was in the critical_exit()
346 * function after setting td_critnest to 0 but before
347 * performing a deferred preemption, this routine can be
348 * invoked with td_critnest set to 0 and td_owepreempt true.
349 * In that case, a critical_exit() during the rendezvous
350 * action would trigger a preemption which is not permitted in
351 * a rendezvous action. To fix this, wrap all of the
352 * rendezvous action handlers in a critical section. We
353 * cannot use a regular critical section however as having
354 * critical_exit() preempt from this routine would also be
355 * problematic (the preemption must not occur before the IPI
356 * has been acknowledged via an EOI). Instead, we
357 * intentionally ignore td_owepreempt when leaving the
358 * critical section. This should be harmless because we do
359 * not permit rendezvous action routines to schedule threads,
360 * and thus td_owepreempt should never transition from 0 to 1
361 * during this routine.
362 */
363 td = curthread;
364 td->td_critnest++;
365#ifdef INVARIANTS
366 owepreempt = td->td_owepreempt;
367#endif
368
369 /*
370 * If requested, run a setup function before the main action
371 * function. Ensure all CPUs have completed the setup
372 * function before moving on to the action function.
373 */
374 if (local_setup_func != smp_no_rendevous_barrier) {
375 if (smp_rv_setup_func != NULL)
376 smp_rv_setup_func(smp_rv_func_arg);
377 atomic_add_int(&smp_rv_waiters[1], 1);
378 while (smp_rv_waiters[1] < smp_rv_ncpus)
379 cpu_spinwait();
380 }
381
382 if (local_action_func != NULL)
383 local_action_func(local_func_arg);
384
385 /*
386 * Signal that the main action has been completed. If a
387 * full exit rendezvous is requested, then all CPUs will
388 * wait here until all CPUs have finished the main action.
389 *
390 * Note that the write by the last CPU to finish the action
391 * may become visible to different CPUs at different times.
392 * As a result, the CPU that initiated the rendezvous may
393 * exit the rendezvous and drop the lock allowing another
394 * rendezvous to be initiated on the same CPU or a different
395 * CPU. In that case the exit sentinel may be cleared before
396 * all CPUs have noticed causing those CPUs to hang forever.
397 * Workaround this by using a generation count to notice when
398 * this race occurs and to exit the rendezvous in that case.
399 */
400 MPASS(generation == smp_rv_generation);
401 atomic_add_int(&smp_rv_waiters[2], 1);
402 if (local_teardown_func != smp_no_rendevous_barrier) {
403 while (smp_rv_waiters[2] < smp_rv_ncpus &&
404 generation == smp_rv_generation)
405 cpu_spinwait();
406
407 if (local_teardown_func != NULL)
408 local_teardown_func(local_func_arg);
409 }
410
411 td->td_critnest--;
412 KASSERT(owepreempt == td->td_owepreempt,
413 ("rendezvous action changed td_owepreempt"));
414}
415
416void
417smp_rendezvous_cpus(cpuset_t map,
418 void (* setup_func)(void *),
419 void (* action_func)(void *),
420 void (* teardown_func)(void *),
421 void *arg)
422{
423 int curcpumap, i, ncpus = 0;
424
425 if (!smp_started) {
426 if (setup_func != NULL)
427 setup_func(arg);
428 if (action_func != NULL)
429 action_func(arg);
430 if (teardown_func != NULL)
431 teardown_func(arg);
432 return;
433 }
434
435 CPU_FOREACH(i) {
436 if (CPU_ISSET(i, &map))
437 ncpus++;
438 }
439 if (ncpus == 0)
440 panic("ncpus is 0 with non-zero map");
441
442 mtx_lock_spin(&smp_ipi_mtx);
443
444 atomic_add_acq_int(&smp_rv_generation, 1);
445
446 /* Pass rendezvous parameters via global variables. */
447 smp_rv_ncpus = ncpus;
448 smp_rv_setup_func = setup_func;
449 smp_rv_action_func = action_func;
450 smp_rv_teardown_func = teardown_func;
451 smp_rv_func_arg = arg;
452 smp_rv_waiters[1] = 0;
453 smp_rv_waiters[2] = 0;
454 atomic_store_rel_int(&smp_rv_waiters[0], 0);
455
456 /*
457 * Signal other processors, which will enter the IPI with
458 * interrupts off.
459 */
460 curcpumap = CPU_ISSET(curcpu, &map);
461 CPU_CLR(curcpu, &map);
462 ipi_selected(map, IPI_RENDEZVOUS);
463
464 /* Check if the current CPU is in the map */
465 if (curcpumap != 0)
466 smp_rendezvous_action();
467
468 /*
469 * If the caller did not request an exit barrier to be enforced
470 * on each CPU, ensure that this CPU waits for all the other
471 * CPUs to finish the rendezvous.
472 */
473 if (teardown_func == smp_no_rendevous_barrier)
474 while (atomic_load_acq_int(&smp_rv_waiters[2]) < ncpus)
475 cpu_spinwait();
476
477 mtx_unlock_spin(&smp_ipi_mtx);
478}
479
480void
481smp_rendezvous(void (* setup_func)(void *),
482 void (* action_func)(void *),
483 void (* teardown_func)(void *),
484 void *arg)
485{
486 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg);
487}
488
489static struct cpu_group group[MAXCPU];
490
491struct cpu_group *
492smp_topo(void)
493{
494 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
495 struct cpu_group *top;
496
497 /*
498 * Check for a fake topology request for debugging purposes.
499 */
500 switch (smp_topology) {
501 case 1:
502 /* Dual core with no sharing. */
503 top = smp_topo_1level(CG_SHARE_NONE, 2, 0);
504 break;
505 case 2:
506 /* No topology, all cpus are equal. */
507 top = smp_topo_none();
508 break;
509 case 3:
510 /* Dual core with shared L2. */
511 top = smp_topo_1level(CG_SHARE_L2, 2, 0);
512 break;
513 case 4:
514 /* quad core, shared l3 among each package, private l2. */
515 top = smp_topo_1level(CG_SHARE_L3, 4, 0);
516 break;
517 case 5:
518 /* quad core, 2 dualcore parts on each package share l2. */
519 top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0);
520 break;
521 case 6:
522 /* Single-core 2xHTT */
523 top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT);
524 break;
525 case 7:
526 /* quad core with a shared l3, 8 threads sharing L2. */
527 top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8,
528 CG_FLAG_SMT);
529 break;
530 default:
531 /* Default, ask the system what it wants. */
532 top = cpu_topo();
533 break;
534 }
535 /*
536 * Verify the returned topology.
537 */
538 if (top->cg_count != mp_ncpus)
539 panic("Built bad topology at %p. CPU count %d != %d",
540 top, top->cg_count, mp_ncpus);
541 if (CPU_CMP(&top->cg_mask, &all_cpus))
542 panic("Built bad topology at %p. CPU mask (%s) != (%s)",
543 top, cpusetobj_strprint(cpusetbuf, &top->cg_mask),
544 cpusetobj_strprint(cpusetbuf2, &all_cpus));
545 return (top);
546}
547
548struct cpu_group *
549smp_topo_none(void)
550{
551 struct cpu_group *top;
552
553 top = &group[0];
554 top->cg_parent = NULL;
555 top->cg_child = NULL;
556 top->cg_mask = all_cpus;
557 top->cg_count = mp_ncpus;
558 top->cg_children = 0;
559 top->cg_level = CG_SHARE_NONE;
560 top->cg_flags = 0;
561
562 return (top);
563}
564
565static int
566smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share,
567 int count, int flags, int start)
568{
569 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
570 cpuset_t mask;
571 int i;
572
573 CPU_ZERO(&mask);
574 for (i = 0; i < count; i++, start++)
575 CPU_SET(start, &mask);
576 child->cg_parent = parent;
577 child->cg_child = NULL;
578 child->cg_children = 0;
579 child->cg_level = share;
580 child->cg_count = count;
581 child->cg_flags = flags;
582 child->cg_mask = mask;
583 parent->cg_children++;
584 for (; parent != NULL; parent = parent->cg_parent) {
585 if (CPU_OVERLAP(&parent->cg_mask, &child->cg_mask))
586 panic("Duplicate children in %p. mask (%s) child (%s)",
587 parent,
588 cpusetobj_strprint(cpusetbuf, &parent->cg_mask),
589 cpusetobj_strprint(cpusetbuf2, &child->cg_mask));
590 CPU_OR(&parent->cg_mask, &child->cg_mask);
591 parent->cg_count += child->cg_count;
592 }
593
594 return (start);
595}
596
597struct cpu_group *
598smp_topo_1level(int share, int count, int flags)
599{
600 struct cpu_group *child;
601 struct cpu_group *top;
602 int packages;
603 int cpu;
604 int i;
605
606 cpu = 0;
607 top = &group[0];
608 packages = mp_ncpus / count;
609 top->cg_child = child = &group[1];
610 top->cg_level = CG_SHARE_NONE;
611 for (i = 0; i < packages; i++, child++)
612 cpu = smp_topo_addleaf(top, child, share, count, flags, cpu);
613 return (top);
614}
615
616struct cpu_group *
617smp_topo_2level(int l2share, int l2count, int l1share, int l1count,
618 int l1flags)
619{
620 struct cpu_group *top;
621 struct cpu_group *l1g;
622 struct cpu_group *l2g;
623 int cpu;
624 int i;
625 int j;
626
627 cpu = 0;
628 top = &group[0];
629 l2g = &group[1];
630 top->cg_child = l2g;
631 top->cg_level = CG_SHARE_NONE;
632 top->cg_children = mp_ncpus / (l2count * l1count);
633 l1g = l2g + top->cg_children;
634 for (i = 0; i < top->cg_children; i++, l2g++) {
635 l2g->cg_parent = top;
636 l2g->cg_child = l1g;
637 l2g->cg_level = l2share;
638 for (j = 0; j < l2count; j++, l1g++)
639 cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count,
640 l1flags, cpu);
641 }
642 return (top);
643}
644
645
646struct cpu_group *
647smp_topo_find(struct cpu_group *top, int cpu)
648{
649 struct cpu_group *cg;
650 cpuset_t mask;
651 int children;
652 int i;
653
654 CPU_SETOF(cpu, &mask);
655 cg = top;
656 for (;;) {
657 if (!CPU_OVERLAP(&cg->cg_mask, &mask))
658 return (NULL);
659 if (cg->cg_children == 0)
660 return (cg);
661 children = cg->cg_children;
662 for (i = 0, cg = cg->cg_child; i < children; cg++, i++)
663 if (CPU_OVERLAP(&cg->cg_mask, &mask))
664 break;
665 }
666 return (NULL);
667}
668#else /* !SMP */
669
670void
671smp_rendezvous_cpus(cpuset_t map,
672 void (*setup_func)(void *),
673 void (*action_func)(void *),
674 void (*teardown_func)(void *),
675 void *arg)
676{
677 if (setup_func != NULL)
678 setup_func(arg);
679 if (action_func != NULL)
680 action_func(arg);
681 if (teardown_func != NULL)
682 teardown_func(arg);
683}
684
685void
686smp_rendezvous(void (*setup_func)(void *),
687 void (*action_func)(void *),
688 void (*teardown_func)(void *),
689 void *arg)
690{
691
692 if (setup_func != NULL)
693 setup_func(arg);
694 if (action_func != NULL)
695 action_func(arg);
696 if (teardown_func != NULL)
697 teardown_func(arg);
698}
699
700/*
701 * Provide dummy SMP support for UP kernels. Modules that need to use SMP
702 * APIs will still work using this dummy support.
703 */
704static void
705mp_setvariables_for_up(void *dummy)
706{
707 mp_ncpus = 1;
708 mp_maxid = PCPU_GET(cpuid);
709 CPU_SETOF(mp_maxid, &all_cpus);
710 KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero"));
711}
712SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST,
713 mp_setvariables_for_up, NULL);
714#endif /* SMP */
715
716void
717smp_no_rendevous_barrier(void *dummy)
718{
719#ifdef SMP
720 KASSERT((!smp_started),("smp_no_rendevous called and smp is started"));
721#endif
722}
|