xref: /freebsd/sys/kern/subr_smp.c (revision 48c5129f93c5eb5419c87b08e4677d51513f1dc0)
1 /*-
2  * Copyright (c) 2001, John Baldwin <jhb@FreeBSD.org>.
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 3. Neither the name of the author nor the names of any co-contributors
14  *    may be used to endorse or promote products derived from this software
15  *    without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 /*
31  * This module holds the global variables and machine independent functions
32  * used for the kernel SMP support.
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
41 #include <sys/ktr.h>
42 #include <sys/proc.h>
43 #include <sys/bus.h>
44 #include <sys/lock.h>
45 #include <sys/mutex.h>
46 #include <sys/pcpu.h>
47 #include <sys/sched.h>
48 #include <sys/smp.h>
49 #include <sys/sysctl.h>
50 
51 #include <machine/cpu.h>
52 #include <machine/smp.h>
53 
54 #include "opt_sched.h"
55 
56 #ifdef SMP
57 volatile cpuset_t stopped_cpus;
58 volatile cpuset_t started_cpus;
59 volatile cpuset_t suspended_cpus;
60 cpuset_t hlt_cpus_mask;
61 cpuset_t logical_cpus_mask;
62 
63 void (*cpustop_restartfunc)(void);
64 #endif
65 /* This is used in modules that need to work in both SMP and UP. */
66 cpuset_t all_cpus;
67 
68 int mp_ncpus;
69 /* export this for libkvm consumers. */
70 int mp_maxcpus = MAXCPU;
71 
72 volatile int smp_started;
73 u_int mp_maxid;
74 
75 static SYSCTL_NODE(_kern, OID_AUTO, smp, CTLFLAG_RD|CTLFLAG_CAPRD, NULL,
76     "Kernel SMP");
77 
78 SYSCTL_INT(_kern_smp, OID_AUTO, maxid, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxid, 0,
79     "Max CPU ID.");
80 
81 SYSCTL_INT(_kern_smp, OID_AUTO, maxcpus, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxcpus,
82     0, "Max number of CPUs that the system was compiled for.");
83 
84 int smp_active = 0;	/* are the APs allowed to run? */
85 SYSCTL_INT(_kern_smp, OID_AUTO, active, CTLFLAG_RW, &smp_active, 0,
86     "Number of Auxillary Processors (APs) that were successfully started");
87 
88 int smp_disabled = 0;	/* has smp been disabled? */
89 SYSCTL_INT(_kern_smp, OID_AUTO, disabled, CTLFLAG_RDTUN|CTLFLAG_CAPRD,
90     &smp_disabled, 0, "SMP has been disabled from the loader");
91 TUNABLE_INT("kern.smp.disabled", &smp_disabled);
92 
93 int smp_cpus = 1;	/* how many cpu's running */
94 SYSCTL_INT(_kern_smp, OID_AUTO, cpus, CTLFLAG_RD|CTLFLAG_CAPRD, &smp_cpus, 0,
95     "Number of CPUs online");
96 
97 int smp_topology = 0;	/* Which topology we're using. */
98 SYSCTL_INT(_kern_smp, OID_AUTO, topology, CTLFLAG_RD, &smp_topology, 0,
99     "Topology override setting; 0 is default provided by hardware.");
100 TUNABLE_INT("kern.smp.topology", &smp_topology);
101 
102 #ifdef SMP
103 /* Enable forwarding of a signal to a process running on a different CPU */
104 static int forward_signal_enabled = 1;
105 SYSCTL_INT(_kern_smp, OID_AUTO, forward_signal_enabled, CTLFLAG_RW,
106 	   &forward_signal_enabled, 0,
107 	   "Forwarding of a signal to a process on a different CPU");
108 
109 /* Variables needed for SMP rendezvous. */
110 static volatile int smp_rv_ncpus;
111 static void (*volatile smp_rv_setup_func)(void *arg);
112 static void (*volatile smp_rv_action_func)(void *arg);
113 static void (*volatile smp_rv_teardown_func)(void *arg);
114 static void *volatile smp_rv_func_arg;
115 static volatile int smp_rv_waiters[4];
116 
117 /*
118  * Shared mutex to restrict busywaits between smp_rendezvous() and
119  * smp(_targeted)_tlb_shootdown().  A deadlock occurs if both of these
120  * functions trigger at once and cause multiple CPUs to busywait with
121  * interrupts disabled.
122  */
123 struct mtx smp_ipi_mtx;
124 
125 /*
126  * Let the MD SMP code initialize mp_maxid very early if it can.
127  */
128 static void
129 mp_setmaxid(void *dummy)
130 {
131 	cpu_mp_setmaxid();
132 }
133 SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL);
134 
135 /*
136  * Call the MD SMP initialization code.
137  */
138 static void
139 mp_start(void *dummy)
140 {
141 
142 	mtx_init(&smp_ipi_mtx, "smp rendezvous", NULL, MTX_SPIN);
143 
144 	/* Probe for MP hardware. */
145 	if (smp_disabled != 0 || cpu_mp_probe() == 0) {
146 		mp_ncpus = 1;
147 		CPU_SETOF(PCPU_GET(cpuid), &all_cpus);
148 		return;
149 	}
150 
151 	cpu_mp_start();
152 	printf("FreeBSD/SMP: Multiprocessor System Detected: %d CPUs\n",
153 	    mp_ncpus);
154 	cpu_mp_announce();
155 }
156 SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL);
157 
158 void
159 forward_signal(struct thread *td)
160 {
161 	int id;
162 
163 	/*
164 	 * signotify() has already set TDF_ASTPENDING and TDF_NEEDSIGCHECK on
165 	 * this thread, so all we need to do is poke it if it is currently
166 	 * executing so that it executes ast().
167 	 */
168 	THREAD_LOCK_ASSERT(td, MA_OWNED);
169 	KASSERT(TD_IS_RUNNING(td),
170 	    ("forward_signal: thread is not TDS_RUNNING"));
171 
172 	CTR1(KTR_SMP, "forward_signal(%p)", td->td_proc);
173 
174 	if (!smp_started || cold || panicstr)
175 		return;
176 	if (!forward_signal_enabled)
177 		return;
178 
179 	/* No need to IPI ourself. */
180 	if (td == curthread)
181 		return;
182 
183 	id = td->td_oncpu;
184 	if (id == NOCPU)
185 		return;
186 	ipi_cpu(id, IPI_AST);
187 }
188 
189 /*
190  * When called the executing CPU will send an IPI to all other CPUs
191  *  requesting that they halt execution.
192  *
193  * Usually (but not necessarily) called with 'other_cpus' as its arg.
194  *
195  *  - Signals all CPUs in map to stop.
196  *  - Waits for each to stop.
197  *
198  * Returns:
199  *  -1: error
200  *   0: NA
201  *   1: ok
202  *
203  */
204 static int
205 generic_stop_cpus(cpuset_t map, u_int type)
206 {
207 #ifdef KTR
208 	char cpusetbuf[CPUSETBUFSIZ];
209 #endif
210 	static volatile u_int stopping_cpu = NOCPU;
211 	int i;
212 	volatile cpuset_t *cpus;
213 
214 	KASSERT(
215 #if defined(__amd64__) || defined(__i386__)
216 	    type == IPI_STOP || type == IPI_STOP_HARD || type == IPI_SUSPEND,
217 #else
218 	    type == IPI_STOP || type == IPI_STOP_HARD,
219 #endif
220 	    ("%s: invalid stop type", __func__));
221 
222 	if (!smp_started)
223 		return (0);
224 
225 	CTR2(KTR_SMP, "stop_cpus(%s) with %u type",
226 	    cpusetobj_strprint(cpusetbuf, &map), type);
227 
228 	if (stopping_cpu != PCPU_GET(cpuid))
229 		while (atomic_cmpset_int(&stopping_cpu, NOCPU,
230 		    PCPU_GET(cpuid)) == 0)
231 			while (stopping_cpu != NOCPU)
232 				cpu_spinwait(); /* spin */
233 
234 	/* send the stop IPI to all CPUs in map */
235 	ipi_selected(map, type);
236 
237 #if defined(__amd64__) || defined(__i386__)
238 	if (type == IPI_SUSPEND)
239 		cpus = &suspended_cpus;
240 	else
241 #endif
242 		cpus = &stopped_cpus;
243 
244 	i = 0;
245 	while (!CPU_SUBSET(cpus, &map)) {
246 		/* spin */
247 		cpu_spinwait();
248 		i++;
249 		if (i == 100000000) {
250 			printf("timeout stopping cpus\n");
251 			break;
252 		}
253 	}
254 
255 	stopping_cpu = NOCPU;
256 	return (1);
257 }
258 
259 int
260 stop_cpus(cpuset_t map)
261 {
262 
263 	return (generic_stop_cpus(map, IPI_STOP));
264 }
265 
266 int
267 stop_cpus_hard(cpuset_t map)
268 {
269 
270 	return (generic_stop_cpus(map, IPI_STOP_HARD));
271 }
272 
273 #if defined(__amd64__) || defined(__i386__)
274 int
275 suspend_cpus(cpuset_t map)
276 {
277 
278 	return (generic_stop_cpus(map, IPI_SUSPEND));
279 }
280 #endif
281 
282 /*
283  * Called by a CPU to restart stopped CPUs.
284  *
285  * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
286  *
287  *  - Signals all CPUs in map to restart.
288  *  - Waits for each to restart.
289  *
290  * Returns:
291  *  -1: error
292  *   0: NA
293  *   1: ok
294  */
295 int
296 restart_cpus(cpuset_t map)
297 {
298 #ifdef KTR
299 	char cpusetbuf[CPUSETBUFSIZ];
300 #endif
301 
302 	if (!smp_started)
303 		return 0;
304 
305 	CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map));
306 
307 	/* signal other cpus to restart */
308 	CPU_COPY_STORE_REL(&map, &started_cpus);
309 
310 	/* wait for each to clear its bit */
311 	while (CPU_OVERLAP(&stopped_cpus, &map))
312 		cpu_spinwait();
313 
314 	return 1;
315 }
316 
317 /*
318  * All-CPU rendezvous.  CPUs are signalled, all execute the setup function
319  * (if specified), rendezvous, execute the action function (if specified),
320  * rendezvous again, execute the teardown function (if specified), and then
321  * resume.
322  *
323  * Note that the supplied external functions _must_ be reentrant and aware
324  * that they are running in parallel and in an unknown lock context.
325  */
326 void
327 smp_rendezvous_action(void)
328 {
329 	struct thread *td;
330 	void *local_func_arg;
331 	void (*local_setup_func)(void*);
332 	void (*local_action_func)(void*);
333 	void (*local_teardown_func)(void*);
334 #ifdef INVARIANTS
335 	int owepreempt;
336 #endif
337 
338 	/* Ensure we have up-to-date values. */
339 	atomic_add_acq_int(&smp_rv_waiters[0], 1);
340 	while (smp_rv_waiters[0] < smp_rv_ncpus)
341 		cpu_spinwait();
342 
343 	/* Fetch rendezvous parameters after acquire barrier. */
344 	local_func_arg = smp_rv_func_arg;
345 	local_setup_func = smp_rv_setup_func;
346 	local_action_func = smp_rv_action_func;
347 	local_teardown_func = smp_rv_teardown_func;
348 
349 	/*
350 	 * Use a nested critical section to prevent any preemptions
351 	 * from occurring during a rendezvous action routine.
352 	 * Specifically, if a rendezvous handler is invoked via an IPI
353 	 * and the interrupted thread was in the critical_exit()
354 	 * function after setting td_critnest to 0 but before
355 	 * performing a deferred preemption, this routine can be
356 	 * invoked with td_critnest set to 0 and td_owepreempt true.
357 	 * In that case, a critical_exit() during the rendezvous
358 	 * action would trigger a preemption which is not permitted in
359 	 * a rendezvous action.  To fix this, wrap all of the
360 	 * rendezvous action handlers in a critical section.  We
361 	 * cannot use a regular critical section however as having
362 	 * critical_exit() preempt from this routine would also be
363 	 * problematic (the preemption must not occur before the IPI
364 	 * has been acknowledged via an EOI).  Instead, we
365 	 * intentionally ignore td_owepreempt when leaving the
366 	 * critical section.  This should be harmless because we do
367 	 * not permit rendezvous action routines to schedule threads,
368 	 * and thus td_owepreempt should never transition from 0 to 1
369 	 * during this routine.
370 	 */
371 	td = curthread;
372 	td->td_critnest++;
373 #ifdef INVARIANTS
374 	owepreempt = td->td_owepreempt;
375 #endif
376 
377 	/*
378 	 * If requested, run a setup function before the main action
379 	 * function.  Ensure all CPUs have completed the setup
380 	 * function before moving on to the action function.
381 	 */
382 	if (local_setup_func != smp_no_rendevous_barrier) {
383 		if (smp_rv_setup_func != NULL)
384 			smp_rv_setup_func(smp_rv_func_arg);
385 		atomic_add_int(&smp_rv_waiters[1], 1);
386 		while (smp_rv_waiters[1] < smp_rv_ncpus)
387                 	cpu_spinwait();
388 	}
389 
390 	if (local_action_func != NULL)
391 		local_action_func(local_func_arg);
392 
393 	if (local_teardown_func != smp_no_rendevous_barrier) {
394 		/*
395 		 * Signal that the main action has been completed.  If a
396 		 * full exit rendezvous is requested, then all CPUs will
397 		 * wait here until all CPUs have finished the main action.
398 		 */
399 		atomic_add_int(&smp_rv_waiters[2], 1);
400 		while (smp_rv_waiters[2] < smp_rv_ncpus)
401 			cpu_spinwait();
402 
403 		if (local_teardown_func != NULL)
404 			local_teardown_func(local_func_arg);
405 	}
406 
407 	/*
408 	 * Signal that the rendezvous is fully completed by this CPU.
409 	 * This means that no member of smp_rv_* pseudo-structure will be
410 	 * accessed by this target CPU after this point; in particular,
411 	 * memory pointed by smp_rv_func_arg.
412 	 */
413 	atomic_add_int(&smp_rv_waiters[3], 1);
414 
415 	td->td_critnest--;
416 	KASSERT(owepreempt == td->td_owepreempt,
417 	    ("rendezvous action changed td_owepreempt"));
418 }
419 
420 void
421 smp_rendezvous_cpus(cpuset_t map,
422 	void (* setup_func)(void *),
423 	void (* action_func)(void *),
424 	void (* teardown_func)(void *),
425 	void *arg)
426 {
427 	int curcpumap, i, ncpus = 0;
428 
429 	/* Look comments in the !SMP case. */
430 	if (!smp_started) {
431 		spinlock_enter();
432 		if (setup_func != NULL)
433 			setup_func(arg);
434 		if (action_func != NULL)
435 			action_func(arg);
436 		if (teardown_func != NULL)
437 			teardown_func(arg);
438 		spinlock_exit();
439 		return;
440 	}
441 
442 	CPU_FOREACH(i) {
443 		if (CPU_ISSET(i, &map))
444 			ncpus++;
445 	}
446 	if (ncpus == 0)
447 		panic("ncpus is 0 with non-zero map");
448 
449 	mtx_lock_spin(&smp_ipi_mtx);
450 
451 	/* Pass rendezvous parameters via global variables. */
452 	smp_rv_ncpus = ncpus;
453 	smp_rv_setup_func = setup_func;
454 	smp_rv_action_func = action_func;
455 	smp_rv_teardown_func = teardown_func;
456 	smp_rv_func_arg = arg;
457 	smp_rv_waiters[1] = 0;
458 	smp_rv_waiters[2] = 0;
459 	smp_rv_waiters[3] = 0;
460 	atomic_store_rel_int(&smp_rv_waiters[0], 0);
461 
462 	/*
463 	 * Signal other processors, which will enter the IPI with
464 	 * interrupts off.
465 	 */
466 	curcpumap = CPU_ISSET(curcpu, &map);
467 	CPU_CLR(curcpu, &map);
468 	ipi_selected(map, IPI_RENDEZVOUS);
469 
470 	/* Check if the current CPU is in the map */
471 	if (curcpumap != 0)
472 		smp_rendezvous_action();
473 
474 	/*
475 	 * Ensure that the master CPU waits for all the other
476 	 * CPUs to finish the rendezvous, so that smp_rv_*
477 	 * pseudo-structure and the arg are guaranteed to not
478 	 * be in use.
479 	 */
480 	while (atomic_load_acq_int(&smp_rv_waiters[3]) < ncpus)
481 		cpu_spinwait();
482 
483 	mtx_unlock_spin(&smp_ipi_mtx);
484 }
485 
486 void
487 smp_rendezvous(void (* setup_func)(void *),
488 	       void (* action_func)(void *),
489 	       void (* teardown_func)(void *),
490 	       void *arg)
491 {
492 	smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg);
493 }
494 
495 static struct cpu_group group[MAXCPU];
496 
497 struct cpu_group *
498 smp_topo(void)
499 {
500 	char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
501 	struct cpu_group *top;
502 
503 	/*
504 	 * Check for a fake topology request for debugging purposes.
505 	 */
506 	switch (smp_topology) {
507 	case 1:
508 		/* Dual core with no sharing.  */
509 		top = smp_topo_1level(CG_SHARE_NONE, 2, 0);
510 		break;
511 	case 2:
512 		/* No topology, all cpus are equal. */
513 		top = smp_topo_none();
514 		break;
515 	case 3:
516 		/* Dual core with shared L2.  */
517 		top = smp_topo_1level(CG_SHARE_L2, 2, 0);
518 		break;
519 	case 4:
520 		/* quad core, shared l3 among each package, private l2.  */
521 		top = smp_topo_1level(CG_SHARE_L3, 4, 0);
522 		break;
523 	case 5:
524 		/* quad core,  2 dualcore parts on each package share l2.  */
525 		top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0);
526 		break;
527 	case 6:
528 		/* Single-core 2xHTT */
529 		top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT);
530 		break;
531 	case 7:
532 		/* quad core with a shared l3, 8 threads sharing L2.  */
533 		top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8,
534 		    CG_FLAG_SMT);
535 		break;
536 	default:
537 		/* Default, ask the system what it wants. */
538 		top = cpu_topo();
539 		break;
540 	}
541 	/*
542 	 * Verify the returned topology.
543 	 */
544 	if (top->cg_count != mp_ncpus)
545 		panic("Built bad topology at %p.  CPU count %d != %d",
546 		    top, top->cg_count, mp_ncpus);
547 	if (CPU_CMP(&top->cg_mask, &all_cpus))
548 		panic("Built bad topology at %p.  CPU mask (%s) != (%s)",
549 		    top, cpusetobj_strprint(cpusetbuf, &top->cg_mask),
550 		    cpusetobj_strprint(cpusetbuf2, &all_cpus));
551 	return (top);
552 }
553 
554 struct cpu_group *
555 smp_topo_none(void)
556 {
557 	struct cpu_group *top;
558 
559 	top = &group[0];
560 	top->cg_parent = NULL;
561 	top->cg_child = NULL;
562 	top->cg_mask = all_cpus;
563 	top->cg_count = mp_ncpus;
564 	top->cg_children = 0;
565 	top->cg_level = CG_SHARE_NONE;
566 	top->cg_flags = 0;
567 
568 	return (top);
569 }
570 
571 static int
572 smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share,
573     int count, int flags, int start)
574 {
575 	char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
576 	cpuset_t mask;
577 	int i;
578 
579 	CPU_ZERO(&mask);
580 	for (i = 0; i < count; i++, start++)
581 		CPU_SET(start, &mask);
582 	child->cg_parent = parent;
583 	child->cg_child = NULL;
584 	child->cg_children = 0;
585 	child->cg_level = share;
586 	child->cg_count = count;
587 	child->cg_flags = flags;
588 	child->cg_mask = mask;
589 	parent->cg_children++;
590 	for (; parent != NULL; parent = parent->cg_parent) {
591 		if (CPU_OVERLAP(&parent->cg_mask, &child->cg_mask))
592 			panic("Duplicate children in %p.  mask (%s) child (%s)",
593 			    parent,
594 			    cpusetobj_strprint(cpusetbuf, &parent->cg_mask),
595 			    cpusetobj_strprint(cpusetbuf2, &child->cg_mask));
596 		CPU_OR(&parent->cg_mask, &child->cg_mask);
597 		parent->cg_count += child->cg_count;
598 	}
599 
600 	return (start);
601 }
602 
603 struct cpu_group *
604 smp_topo_1level(int share, int count, int flags)
605 {
606 	struct cpu_group *child;
607 	struct cpu_group *top;
608 	int packages;
609 	int cpu;
610 	int i;
611 
612 	cpu = 0;
613 	top = &group[0];
614 	packages = mp_ncpus / count;
615 	top->cg_child = child = &group[1];
616 	top->cg_level = CG_SHARE_NONE;
617 	for (i = 0; i < packages; i++, child++)
618 		cpu = smp_topo_addleaf(top, child, share, count, flags, cpu);
619 	return (top);
620 }
621 
622 struct cpu_group *
623 smp_topo_2level(int l2share, int l2count, int l1share, int l1count,
624     int l1flags)
625 {
626 	struct cpu_group *top;
627 	struct cpu_group *l1g;
628 	struct cpu_group *l2g;
629 	int cpu;
630 	int i;
631 	int j;
632 
633 	cpu = 0;
634 	top = &group[0];
635 	l2g = &group[1];
636 	top->cg_child = l2g;
637 	top->cg_level = CG_SHARE_NONE;
638 	top->cg_children = mp_ncpus / (l2count * l1count);
639 	l1g = l2g + top->cg_children;
640 	for (i = 0; i < top->cg_children; i++, l2g++) {
641 		l2g->cg_parent = top;
642 		l2g->cg_child = l1g;
643 		l2g->cg_level = l2share;
644 		for (j = 0; j < l2count; j++, l1g++)
645 			cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count,
646 			    l1flags, cpu);
647 	}
648 	return (top);
649 }
650 
651 
652 struct cpu_group *
653 smp_topo_find(struct cpu_group *top, int cpu)
654 {
655 	struct cpu_group *cg;
656 	cpuset_t mask;
657 	int children;
658 	int i;
659 
660 	CPU_SETOF(cpu, &mask);
661 	cg = top;
662 	for (;;) {
663 		if (!CPU_OVERLAP(&cg->cg_mask, &mask))
664 			return (NULL);
665 		if (cg->cg_children == 0)
666 			return (cg);
667 		children = cg->cg_children;
668 		for (i = 0, cg = cg->cg_child; i < children; cg++, i++)
669 			if (CPU_OVERLAP(&cg->cg_mask, &mask))
670 				break;
671 	}
672 	return (NULL);
673 }
674 #else /* !SMP */
675 
676 void
677 smp_rendezvous_cpus(cpuset_t map,
678 	void (*setup_func)(void *),
679 	void (*action_func)(void *),
680 	void (*teardown_func)(void *),
681 	void *arg)
682 {
683 	/*
684 	 * In the !SMP case we just need to ensure the same initial conditions
685 	 * as the SMP case.
686 	 */
687 	spinlock_enter();
688 	if (setup_func != NULL)
689 		setup_func(arg);
690 	if (action_func != NULL)
691 		action_func(arg);
692 	if (teardown_func != NULL)
693 		teardown_func(arg);
694 	spinlock_exit();
695 }
696 
697 void
698 smp_rendezvous(void (*setup_func)(void *),
699 	       void (*action_func)(void *),
700 	       void (*teardown_func)(void *),
701 	       void *arg)
702 {
703 
704 	/* Look comments in the smp_rendezvous_cpus() case. */
705 	spinlock_enter();
706 	if (setup_func != NULL)
707 		setup_func(arg);
708 	if (action_func != NULL)
709 		action_func(arg);
710 	if (teardown_func != NULL)
711 		teardown_func(arg);
712 	spinlock_exit();
713 }
714 
715 /*
716  * Provide dummy SMP support for UP kernels.  Modules that need to use SMP
717  * APIs will still work using this dummy support.
718  */
719 static void
720 mp_setvariables_for_up(void *dummy)
721 {
722 	mp_ncpus = 1;
723 	mp_maxid = PCPU_GET(cpuid);
724 	CPU_SETOF(mp_maxid, &all_cpus);
725 	KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero"));
726 }
727 SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST,
728     mp_setvariables_for_up, NULL);
729 #endif /* SMP */
730 
731 void
732 smp_no_rendevous_barrier(void *dummy)
733 {
734 #ifdef SMP
735 	KASSERT((!smp_started),("smp_no_rendevous called and smp is started"));
736 #endif
737 }
738 
739 /*
740  * Wait specified idle threads to switch once.  This ensures that even
741  * preempted threads have cycled through the switch function once,
742  * exiting their codepaths.  This allows us to change global pointers
743  * with no other synchronization.
744  */
745 int
746 quiesce_cpus(cpuset_t map, const char *wmesg, int prio)
747 {
748 	struct pcpu *pcpu;
749 	u_int gen[MAXCPU];
750 	int error;
751 	int cpu;
752 
753 	error = 0;
754 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
755 		if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu))
756 			continue;
757 		pcpu = pcpu_find(cpu);
758 		gen[cpu] = pcpu->pc_idlethread->td_generation;
759 	}
760 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
761 		if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu))
762 			continue;
763 		pcpu = pcpu_find(cpu);
764 		thread_lock(curthread);
765 		sched_bind(curthread, cpu);
766 		thread_unlock(curthread);
767 		while (gen[cpu] == pcpu->pc_idlethread->td_generation) {
768 			error = tsleep(quiesce_cpus, prio, wmesg, 1);
769 			if (error)
770 				goto out;
771 		}
772 	}
773 out:
774 	thread_lock(curthread);
775 	sched_unbind(curthread);
776 	thread_unlock(curthread);
777 
778 	return (error);
779 }
780 
781 int
782 quiesce_all_cpus(const char *wmesg, int prio)
783 {
784 
785 	return quiesce_cpus(all_cpus, wmesg, prio);
786 }
787