xref: /freebsd/sys/kern/subr_smp.c (revision d6eb98610fa65663bf0df4574b7cb2c5c4ffda71)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2001, John Baldwin <jhb@FreeBSD.org>.
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  * This module holds the global variables and machine independent functions
30  * used for the kernel SMP support.
31  */
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/kernel.h>
39 #include <sys/ktr.h>
40 #include <sys/proc.h>
41 #include <sys/bus.h>
42 #include <sys/lock.h>
43 #include <sys/malloc.h>
44 #include <sys/mutex.h>
45 #include <sys/pcpu.h>
46 #include <sys/sched.h>
47 #include <sys/smp.h>
48 #include <sys/sysctl.h>
49 
50 #include <machine/cpu.h>
51 #include <machine/smp.h>
52 
53 #include "opt_sched.h"
54 
55 #ifdef SMP
56 MALLOC_DEFINE(M_TOPO, "toponodes", "SMP topology data");
57 
58 volatile cpuset_t stopped_cpus;
59 volatile cpuset_t started_cpus;
60 volatile cpuset_t suspended_cpus;
61 cpuset_t hlt_cpus_mask;
62 cpuset_t logical_cpus_mask;
63 
64 void (*cpustop_restartfunc)(void);
65 #endif
66 
67 static int sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS);
68 
69 /* This is used in modules that need to work in both SMP and UP. */
70 cpuset_t all_cpus;
71 
72 int mp_ncpus;
73 /* export this for libkvm consumers. */
74 int mp_maxcpus = MAXCPU;
75 
76 volatile int smp_started;
77 u_int mp_maxid;
78 
79 static SYSCTL_NODE(_kern, OID_AUTO, smp, CTLFLAG_RD|CTLFLAG_CAPRD, NULL,
80     "Kernel SMP");
81 
82 SYSCTL_INT(_kern_smp, OID_AUTO, maxid, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxid, 0,
83     "Max CPU ID.");
84 
85 SYSCTL_INT(_kern_smp, OID_AUTO, maxcpus, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxcpus,
86     0, "Max number of CPUs that the system was compiled for.");
87 
88 SYSCTL_PROC(_kern_smp, OID_AUTO, active, CTLFLAG_RD|CTLTYPE_INT|CTLFLAG_MPSAFE,
89     NULL, 0, sysctl_kern_smp_active, "I",
90     "Indicates system is running in SMP mode");
91 
92 int smp_disabled = 0;	/* has smp been disabled? */
93 SYSCTL_INT(_kern_smp, OID_AUTO, disabled, CTLFLAG_RDTUN|CTLFLAG_CAPRD,
94     &smp_disabled, 0, "SMP has been disabled from the loader");
95 
96 int smp_cpus = 1;	/* how many cpu's running */
97 SYSCTL_INT(_kern_smp, OID_AUTO, cpus, CTLFLAG_RD|CTLFLAG_CAPRD, &smp_cpus, 0,
98     "Number of CPUs online");
99 
100 int smp_threads_per_core = 1;	/* how many SMT threads are running per core */
101 SYSCTL_INT(_kern_smp, OID_AUTO, threads_per_core, CTLFLAG_RD|CTLFLAG_CAPRD,
102     &smp_threads_per_core, 0, "Number of SMT threads online per core");
103 
104 int mp_ncores = -1;	/* how many physical cores running */
105 SYSCTL_INT(_kern_smp, OID_AUTO, cores, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_ncores, 0,
106     "Number of CPUs online");
107 
108 int smp_topology = 0;	/* Which topology we're using. */
109 SYSCTL_INT(_kern_smp, OID_AUTO, topology, CTLFLAG_RDTUN, &smp_topology, 0,
110     "Topology override setting; 0 is default provided by hardware.");
111 
112 #ifdef SMP
113 /* Enable forwarding of a signal to a process running on a different CPU */
114 static int forward_signal_enabled = 1;
115 SYSCTL_INT(_kern_smp, OID_AUTO, forward_signal_enabled, CTLFLAG_RW,
116 	   &forward_signal_enabled, 0,
117 	   "Forwarding of a signal to a process on a different CPU");
118 
119 /* Variables needed for SMP rendezvous. */
120 static volatile int smp_rv_ncpus;
121 static void (*volatile smp_rv_setup_func)(void *arg);
122 static void (*volatile smp_rv_action_func)(void *arg);
123 static void (*volatile smp_rv_teardown_func)(void *arg);
124 static void *volatile smp_rv_func_arg;
125 static volatile int smp_rv_waiters[4];
126 
127 /*
128  * Shared mutex to restrict busywaits between smp_rendezvous() and
129  * smp(_targeted)_tlb_shootdown().  A deadlock occurs if both of these
130  * functions trigger at once and cause multiple CPUs to busywait with
131  * interrupts disabled.
132  */
133 struct mtx smp_ipi_mtx;
134 
135 /*
136  * Let the MD SMP code initialize mp_maxid very early if it can.
137  */
138 static void
139 mp_setmaxid(void *dummy)
140 {
141 
142 	cpu_mp_setmaxid();
143 
144 	KASSERT(mp_ncpus >= 1, ("%s: CPU count < 1", __func__));
145 	KASSERT(mp_ncpus > 1 || mp_maxid == 0,
146 	    ("%s: one CPU but mp_maxid is not zero", __func__));
147 	KASSERT(mp_maxid >= mp_ncpus - 1,
148 	    ("%s: counters out of sync: max %d, count %d", __func__,
149 		mp_maxid, mp_ncpus));
150 }
151 SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL);
152 
153 /*
154  * Call the MD SMP initialization code.
155  */
156 static void
157 mp_start(void *dummy)
158 {
159 
160 	mtx_init(&smp_ipi_mtx, "smp rendezvous", NULL, MTX_SPIN);
161 
162 	/* Probe for MP hardware. */
163 	if (smp_disabled != 0 || cpu_mp_probe() == 0) {
164 		mp_ncores = 1;
165 		mp_ncpus = 1;
166 		CPU_SETOF(PCPU_GET(cpuid), &all_cpus);
167 		return;
168 	}
169 
170 	cpu_mp_start();
171 	printf("FreeBSD/SMP: Multiprocessor System Detected: %d CPUs\n",
172 	    mp_ncpus);
173 
174 	/* Provide a default for most architectures that don't have SMT/HTT. */
175 	if (mp_ncores < 0)
176 		mp_ncores = mp_ncpus;
177 
178 	cpu_mp_announce();
179 }
180 SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL);
181 
182 void
183 forward_signal(struct thread *td)
184 {
185 	int id;
186 
187 	/*
188 	 * signotify() has already set TDF_ASTPENDING and TDF_NEEDSIGCHECK on
189 	 * this thread, so all we need to do is poke it if it is currently
190 	 * executing so that it executes ast().
191 	 */
192 	THREAD_LOCK_ASSERT(td, MA_OWNED);
193 	KASSERT(TD_IS_RUNNING(td),
194 	    ("forward_signal: thread is not TDS_RUNNING"));
195 
196 	CTR1(KTR_SMP, "forward_signal(%p)", td->td_proc);
197 
198 	if (!smp_started || cold || panicstr)
199 		return;
200 	if (!forward_signal_enabled)
201 		return;
202 
203 	/* No need to IPI ourself. */
204 	if (td == curthread)
205 		return;
206 
207 	id = td->td_oncpu;
208 	if (id == NOCPU)
209 		return;
210 	ipi_cpu(id, IPI_AST);
211 }
212 
213 /*
214  * When called the executing CPU will send an IPI to all other CPUs
215  *  requesting that they halt execution.
216  *
217  * Usually (but not necessarily) called with 'other_cpus' as its arg.
218  *
219  *  - Signals all CPUs in map to stop.
220  *  - Waits for each to stop.
221  *
222  * Returns:
223  *  -1: error
224  *   0: NA
225  *   1: ok
226  *
227  */
228 #if defined(__amd64__) || defined(__i386__)
229 #define	X86	1
230 #else
231 #define	X86	0
232 #endif
233 static int
234 generic_stop_cpus(cpuset_t map, u_int type)
235 {
236 #ifdef KTR
237 	char cpusetbuf[CPUSETBUFSIZ];
238 #endif
239 	static volatile u_int stopping_cpu = NOCPU;
240 	int i;
241 	volatile cpuset_t *cpus;
242 
243 	KASSERT(
244 	    type == IPI_STOP || type == IPI_STOP_HARD
245 #if X86
246 	    || type == IPI_SUSPEND
247 #endif
248 	    , ("%s: invalid stop type", __func__));
249 
250 	if (!smp_started)
251 		return (0);
252 
253 	CTR2(KTR_SMP, "stop_cpus(%s) with %u type",
254 	    cpusetobj_strprint(cpusetbuf, &map), type);
255 
256 #if X86
257 	/*
258 	 * When suspending, ensure there are are no IPIs in progress.
259 	 * IPIs that have been issued, but not yet delivered (e.g.
260 	 * not pending on a vCPU when running under virtualization)
261 	 * will be lost, violating FreeBSD's assumption of reliable
262 	 * IPI delivery.
263 	 */
264 	if (type == IPI_SUSPEND)
265 		mtx_lock_spin(&smp_ipi_mtx);
266 #endif
267 
268 #if X86
269 	if (!nmi_is_broadcast || nmi_kdb_lock == 0) {
270 #endif
271 	if (stopping_cpu != PCPU_GET(cpuid))
272 		while (atomic_cmpset_int(&stopping_cpu, NOCPU,
273 		    PCPU_GET(cpuid)) == 0)
274 			while (stopping_cpu != NOCPU)
275 				cpu_spinwait(); /* spin */
276 
277 	/* send the stop IPI to all CPUs in map */
278 	ipi_selected(map, type);
279 #if X86
280 	}
281 #endif
282 
283 #if X86
284 	if (type == IPI_SUSPEND)
285 		cpus = &suspended_cpus;
286 	else
287 #endif
288 		cpus = &stopped_cpus;
289 
290 	i = 0;
291 	while (!CPU_SUBSET(cpus, &map)) {
292 		/* spin */
293 		cpu_spinwait();
294 		i++;
295 		if (i == 100000000) {
296 			printf("timeout stopping cpus\n");
297 			break;
298 		}
299 	}
300 
301 #if X86
302 	if (type == IPI_SUSPEND)
303 		mtx_unlock_spin(&smp_ipi_mtx);
304 #endif
305 
306 	stopping_cpu = NOCPU;
307 	return (1);
308 }
309 
310 int
311 stop_cpus(cpuset_t map)
312 {
313 
314 	return (generic_stop_cpus(map, IPI_STOP));
315 }
316 
317 int
318 stop_cpus_hard(cpuset_t map)
319 {
320 
321 	return (generic_stop_cpus(map, IPI_STOP_HARD));
322 }
323 
324 #if X86
325 int
326 suspend_cpus(cpuset_t map)
327 {
328 
329 	return (generic_stop_cpus(map, IPI_SUSPEND));
330 }
331 #endif
332 
333 /*
334  * Called by a CPU to restart stopped CPUs.
335  *
336  * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
337  *
338  *  - Signals all CPUs in map to restart.
339  *  - Waits for each to restart.
340  *
341  * Returns:
342  *  -1: error
343  *   0: NA
344  *   1: ok
345  */
346 static int
347 generic_restart_cpus(cpuset_t map, u_int type)
348 {
349 #ifdef KTR
350 	char cpusetbuf[CPUSETBUFSIZ];
351 #endif
352 	volatile cpuset_t *cpus;
353 
354 #if X86
355 	KASSERT(type == IPI_STOP || type == IPI_STOP_HARD
356 	    || type == IPI_SUSPEND, ("%s: invalid stop type", __func__));
357 
358 	if (!smp_started)
359 		return (0);
360 
361 	CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map));
362 
363 	if (type == IPI_SUSPEND)
364 		cpus = &resuming_cpus;
365 	else
366 		cpus = &stopped_cpus;
367 
368 	/* signal other cpus to restart */
369 	if (type == IPI_SUSPEND)
370 		CPU_COPY_STORE_REL(&map, &toresume_cpus);
371 	else
372 		CPU_COPY_STORE_REL(&map, &started_cpus);
373 
374 	/*
375 	 * Wake up any CPUs stopped with MWAIT.  From MI code we can't tell if
376 	 * MONITOR/MWAIT is enabled, but the potentially redundant writes are
377 	 * relatively inexpensive.
378 	 */
379 	if (type == IPI_STOP) {
380 		struct monitorbuf *mb;
381 		u_int id;
382 
383 		CPU_FOREACH(id) {
384 			if (!CPU_ISSET(id, &map))
385 				continue;
386 
387 			mb = &pcpu_find(id)->pc_monitorbuf;
388 			atomic_store_int(&mb->stop_state,
389 			    MONITOR_STOPSTATE_RUNNING);
390 		}
391 	}
392 
393 	if (!nmi_is_broadcast || nmi_kdb_lock == 0) {
394 		/* wait for each to clear its bit */
395 		while (CPU_OVERLAP(cpus, &map))
396 			cpu_spinwait();
397 	}
398 #else /* !X86 */
399 	KASSERT(type == IPI_STOP || type == IPI_STOP_HARD,
400 	    ("%s: invalid stop type", __func__));
401 
402 	if (!smp_started)
403 		return (0);
404 
405 	CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map));
406 
407 	cpus = &stopped_cpus;
408 
409 	/* signal other cpus to restart */
410 	CPU_COPY_STORE_REL(&map, &started_cpus);
411 
412 	/* wait for each to clear its bit */
413 	while (CPU_OVERLAP(cpus, &map))
414 		cpu_spinwait();
415 #endif
416 	return (1);
417 }
418 
419 int
420 restart_cpus(cpuset_t map)
421 {
422 
423 	return (generic_restart_cpus(map, IPI_STOP));
424 }
425 
426 #if X86
427 int
428 resume_cpus(cpuset_t map)
429 {
430 
431 	return (generic_restart_cpus(map, IPI_SUSPEND));
432 }
433 #endif
434 #undef X86
435 
436 /*
437  * All-CPU rendezvous.  CPUs are signalled, all execute the setup function
438  * (if specified), rendezvous, execute the action function (if specified),
439  * rendezvous again, execute the teardown function (if specified), and then
440  * resume.
441  *
442  * Note that the supplied external functions _must_ be reentrant and aware
443  * that they are running in parallel and in an unknown lock context.
444  */
445 void
446 smp_rendezvous_action(void)
447 {
448 	struct thread *td;
449 	void *local_func_arg;
450 	void (*local_setup_func)(void*);
451 	void (*local_action_func)(void*);
452 	void (*local_teardown_func)(void*);
453 #ifdef INVARIANTS
454 	int owepreempt;
455 #endif
456 
457 	/* Ensure we have up-to-date values. */
458 	atomic_add_acq_int(&smp_rv_waiters[0], 1);
459 	while (smp_rv_waiters[0] < smp_rv_ncpus)
460 		cpu_spinwait();
461 
462 	/* Fetch rendezvous parameters after acquire barrier. */
463 	local_func_arg = smp_rv_func_arg;
464 	local_setup_func = smp_rv_setup_func;
465 	local_action_func = smp_rv_action_func;
466 	local_teardown_func = smp_rv_teardown_func;
467 
468 	/*
469 	 * Use a nested critical section to prevent any preemptions
470 	 * from occurring during a rendezvous action routine.
471 	 * Specifically, if a rendezvous handler is invoked via an IPI
472 	 * and the interrupted thread was in the critical_exit()
473 	 * function after setting td_critnest to 0 but before
474 	 * performing a deferred preemption, this routine can be
475 	 * invoked with td_critnest set to 0 and td_owepreempt true.
476 	 * In that case, a critical_exit() during the rendezvous
477 	 * action would trigger a preemption which is not permitted in
478 	 * a rendezvous action.  To fix this, wrap all of the
479 	 * rendezvous action handlers in a critical section.  We
480 	 * cannot use a regular critical section however as having
481 	 * critical_exit() preempt from this routine would also be
482 	 * problematic (the preemption must not occur before the IPI
483 	 * has been acknowledged via an EOI).  Instead, we
484 	 * intentionally ignore td_owepreempt when leaving the
485 	 * critical section.  This should be harmless because we do
486 	 * not permit rendezvous action routines to schedule threads,
487 	 * and thus td_owepreempt should never transition from 0 to 1
488 	 * during this routine.
489 	 */
490 	td = curthread;
491 	td->td_critnest++;
492 #ifdef INVARIANTS
493 	owepreempt = td->td_owepreempt;
494 #endif
495 
496 	/*
497 	 * If requested, run a setup function before the main action
498 	 * function.  Ensure all CPUs have completed the setup
499 	 * function before moving on to the action function.
500 	 */
501 	if (local_setup_func != smp_no_rendezvous_barrier) {
502 		if (smp_rv_setup_func != NULL)
503 			smp_rv_setup_func(smp_rv_func_arg);
504 		atomic_add_int(&smp_rv_waiters[1], 1);
505 		while (smp_rv_waiters[1] < smp_rv_ncpus)
506                 	cpu_spinwait();
507 	}
508 
509 	if (local_action_func != NULL)
510 		local_action_func(local_func_arg);
511 
512 	if (local_teardown_func != smp_no_rendezvous_barrier) {
513 		/*
514 		 * Signal that the main action has been completed.  If a
515 		 * full exit rendezvous is requested, then all CPUs will
516 		 * wait here until all CPUs have finished the main action.
517 		 */
518 		atomic_add_int(&smp_rv_waiters[2], 1);
519 		while (smp_rv_waiters[2] < smp_rv_ncpus)
520 			cpu_spinwait();
521 
522 		if (local_teardown_func != NULL)
523 			local_teardown_func(local_func_arg);
524 	}
525 
526 	/*
527 	 * Signal that the rendezvous is fully completed by this CPU.
528 	 * This means that no member of smp_rv_* pseudo-structure will be
529 	 * accessed by this target CPU after this point; in particular,
530 	 * memory pointed by smp_rv_func_arg.
531 	 *
532 	 * The release semantic ensures that all accesses performed by
533 	 * the current CPU are visible when smp_rendezvous_cpus()
534 	 * returns, by synchronizing with the
535 	 * atomic_load_acq_int(&smp_rv_waiters[3]).
536 	 */
537 	atomic_add_rel_int(&smp_rv_waiters[3], 1);
538 
539 	td->td_critnest--;
540 	KASSERT(owepreempt == td->td_owepreempt,
541 	    ("rendezvous action changed td_owepreempt"));
542 }
543 
544 void
545 smp_rendezvous_cpus(cpuset_t map,
546 	void (* setup_func)(void *),
547 	void (* action_func)(void *),
548 	void (* teardown_func)(void *),
549 	void *arg)
550 {
551 	int curcpumap, i, ncpus = 0;
552 
553 	/* Look comments in the !SMP case. */
554 	if (!smp_started) {
555 		spinlock_enter();
556 		if (setup_func != NULL)
557 			setup_func(arg);
558 		if (action_func != NULL)
559 			action_func(arg);
560 		if (teardown_func != NULL)
561 			teardown_func(arg);
562 		spinlock_exit();
563 		return;
564 	}
565 
566 	CPU_FOREACH(i) {
567 		if (CPU_ISSET(i, &map))
568 			ncpus++;
569 	}
570 	if (ncpus == 0)
571 		panic("ncpus is 0 with non-zero map");
572 
573 	mtx_lock_spin(&smp_ipi_mtx);
574 
575 	/* Pass rendezvous parameters via global variables. */
576 	smp_rv_ncpus = ncpus;
577 	smp_rv_setup_func = setup_func;
578 	smp_rv_action_func = action_func;
579 	smp_rv_teardown_func = teardown_func;
580 	smp_rv_func_arg = arg;
581 	smp_rv_waiters[1] = 0;
582 	smp_rv_waiters[2] = 0;
583 	smp_rv_waiters[3] = 0;
584 	atomic_store_rel_int(&smp_rv_waiters[0], 0);
585 
586 	/*
587 	 * Signal other processors, which will enter the IPI with
588 	 * interrupts off.
589 	 */
590 	curcpumap = CPU_ISSET(curcpu, &map);
591 	CPU_CLR(curcpu, &map);
592 	ipi_selected(map, IPI_RENDEZVOUS);
593 
594 	/* Check if the current CPU is in the map */
595 	if (curcpumap != 0)
596 		smp_rendezvous_action();
597 
598 	/*
599 	 * Ensure that the master CPU waits for all the other
600 	 * CPUs to finish the rendezvous, so that smp_rv_*
601 	 * pseudo-structure and the arg are guaranteed to not
602 	 * be in use.
603 	 *
604 	 * Load acquire synchronizes with the release add in
605 	 * smp_rendezvous_action(), which ensures that our caller sees
606 	 * all memory actions done by the called functions on other
607 	 * CPUs.
608 	 */
609 	while (atomic_load_acq_int(&smp_rv_waiters[3]) < ncpus)
610 		cpu_spinwait();
611 
612 	mtx_unlock_spin(&smp_ipi_mtx);
613 }
614 
615 void
616 smp_rendezvous(void (* setup_func)(void *),
617 	       void (* action_func)(void *),
618 	       void (* teardown_func)(void *),
619 	       void *arg)
620 {
621 	smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg);
622 }
623 
624 static struct cpu_group group[MAXCPU * MAX_CACHE_LEVELS + 1];
625 
626 struct cpu_group *
627 smp_topo(void)
628 {
629 	char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
630 	struct cpu_group *top;
631 
632 	/*
633 	 * Check for a fake topology request for debugging purposes.
634 	 */
635 	switch (smp_topology) {
636 	case 1:
637 		/* Dual core with no sharing.  */
638 		top = smp_topo_1level(CG_SHARE_NONE, 2, 0);
639 		break;
640 	case 2:
641 		/* No topology, all cpus are equal. */
642 		top = smp_topo_none();
643 		break;
644 	case 3:
645 		/* Dual core with shared L2.  */
646 		top = smp_topo_1level(CG_SHARE_L2, 2, 0);
647 		break;
648 	case 4:
649 		/* quad core, shared l3 among each package, private l2.  */
650 		top = smp_topo_1level(CG_SHARE_L3, 4, 0);
651 		break;
652 	case 5:
653 		/* quad core,  2 dualcore parts on each package share l2.  */
654 		top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0);
655 		break;
656 	case 6:
657 		/* Single-core 2xHTT */
658 		top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT);
659 		break;
660 	case 7:
661 		/* quad core with a shared l3, 8 threads sharing L2.  */
662 		top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8,
663 		    CG_FLAG_SMT);
664 		break;
665 	default:
666 		/* Default, ask the system what it wants. */
667 		top = cpu_topo();
668 		break;
669 	}
670 	/*
671 	 * Verify the returned topology.
672 	 */
673 	if (top->cg_count != mp_ncpus)
674 		panic("Built bad topology at %p.  CPU count %d != %d",
675 		    top, top->cg_count, mp_ncpus);
676 	if (CPU_CMP(&top->cg_mask, &all_cpus))
677 		panic("Built bad topology at %p.  CPU mask (%s) != (%s)",
678 		    top, cpusetobj_strprint(cpusetbuf, &top->cg_mask),
679 		    cpusetobj_strprint(cpusetbuf2, &all_cpus));
680 
681 	/*
682 	 * Collapse nonsense levels that may be created out of convenience by
683 	 * the MD layers.  They cause extra work in the search functions.
684 	 */
685 	while (top->cg_children == 1) {
686 		top = &top->cg_child[0];
687 		top->cg_parent = NULL;
688 	}
689 	return (top);
690 }
691 
692 struct cpu_group *
693 smp_topo_alloc(u_int count)
694 {
695 	static u_int index;
696 	u_int curr;
697 
698 	curr = index;
699 	index += count;
700 	return (&group[curr]);
701 }
702 
703 struct cpu_group *
704 smp_topo_none(void)
705 {
706 	struct cpu_group *top;
707 
708 	top = &group[0];
709 	top->cg_parent = NULL;
710 	top->cg_child = NULL;
711 	top->cg_mask = all_cpus;
712 	top->cg_count = mp_ncpus;
713 	top->cg_children = 0;
714 	top->cg_level = CG_SHARE_NONE;
715 	top->cg_flags = 0;
716 
717 	return (top);
718 }
719 
720 static int
721 smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share,
722     int count, int flags, int start)
723 {
724 	char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
725 	cpuset_t mask;
726 	int i;
727 
728 	CPU_ZERO(&mask);
729 	for (i = 0; i < count; i++, start++)
730 		CPU_SET(start, &mask);
731 	child->cg_parent = parent;
732 	child->cg_child = NULL;
733 	child->cg_children = 0;
734 	child->cg_level = share;
735 	child->cg_count = count;
736 	child->cg_flags = flags;
737 	child->cg_mask = mask;
738 	parent->cg_children++;
739 	for (; parent != NULL; parent = parent->cg_parent) {
740 		if (CPU_OVERLAP(&parent->cg_mask, &child->cg_mask))
741 			panic("Duplicate children in %p.  mask (%s) child (%s)",
742 			    parent,
743 			    cpusetobj_strprint(cpusetbuf, &parent->cg_mask),
744 			    cpusetobj_strprint(cpusetbuf2, &child->cg_mask));
745 		CPU_OR(&parent->cg_mask, &child->cg_mask);
746 		parent->cg_count += child->cg_count;
747 	}
748 
749 	return (start);
750 }
751 
752 struct cpu_group *
753 smp_topo_1level(int share, int count, int flags)
754 {
755 	struct cpu_group *child;
756 	struct cpu_group *top;
757 	int packages;
758 	int cpu;
759 	int i;
760 
761 	cpu = 0;
762 	top = &group[0];
763 	packages = mp_ncpus / count;
764 	top->cg_child = child = &group[1];
765 	top->cg_level = CG_SHARE_NONE;
766 	for (i = 0; i < packages; i++, child++)
767 		cpu = smp_topo_addleaf(top, child, share, count, flags, cpu);
768 	return (top);
769 }
770 
771 struct cpu_group *
772 smp_topo_2level(int l2share, int l2count, int l1share, int l1count,
773     int l1flags)
774 {
775 	struct cpu_group *top;
776 	struct cpu_group *l1g;
777 	struct cpu_group *l2g;
778 	int cpu;
779 	int i;
780 	int j;
781 
782 	cpu = 0;
783 	top = &group[0];
784 	l2g = &group[1];
785 	top->cg_child = l2g;
786 	top->cg_level = CG_SHARE_NONE;
787 	top->cg_children = mp_ncpus / (l2count * l1count);
788 	l1g = l2g + top->cg_children;
789 	for (i = 0; i < top->cg_children; i++, l2g++) {
790 		l2g->cg_parent = top;
791 		l2g->cg_child = l1g;
792 		l2g->cg_level = l2share;
793 		for (j = 0; j < l2count; j++, l1g++)
794 			cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count,
795 			    l1flags, cpu);
796 	}
797 	return (top);
798 }
799 
800 
801 struct cpu_group *
802 smp_topo_find(struct cpu_group *top, int cpu)
803 {
804 	struct cpu_group *cg;
805 	cpuset_t mask;
806 	int children;
807 	int i;
808 
809 	CPU_SETOF(cpu, &mask);
810 	cg = top;
811 	for (;;) {
812 		if (!CPU_OVERLAP(&cg->cg_mask, &mask))
813 			return (NULL);
814 		if (cg->cg_children == 0)
815 			return (cg);
816 		children = cg->cg_children;
817 		for (i = 0, cg = cg->cg_child; i < children; cg++, i++)
818 			if (CPU_OVERLAP(&cg->cg_mask, &mask))
819 				break;
820 	}
821 	return (NULL);
822 }
823 #else /* !SMP */
824 
825 void
826 smp_rendezvous_cpus(cpuset_t map,
827 	void (*setup_func)(void *),
828 	void (*action_func)(void *),
829 	void (*teardown_func)(void *),
830 	void *arg)
831 {
832 	/*
833 	 * In the !SMP case we just need to ensure the same initial conditions
834 	 * as the SMP case.
835 	 */
836 	spinlock_enter();
837 	if (setup_func != NULL)
838 		setup_func(arg);
839 	if (action_func != NULL)
840 		action_func(arg);
841 	if (teardown_func != NULL)
842 		teardown_func(arg);
843 	spinlock_exit();
844 }
845 
846 void
847 smp_rendezvous(void (*setup_func)(void *),
848 	       void (*action_func)(void *),
849 	       void (*teardown_func)(void *),
850 	       void *arg)
851 {
852 
853 	smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func,
854 	    arg);
855 }
856 
857 /*
858  * Provide dummy SMP support for UP kernels.  Modules that need to use SMP
859  * APIs will still work using this dummy support.
860  */
861 static void
862 mp_setvariables_for_up(void *dummy)
863 {
864 	mp_ncpus = 1;
865 	mp_ncores = 1;
866 	mp_maxid = PCPU_GET(cpuid);
867 	CPU_SETOF(mp_maxid, &all_cpus);
868 	KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero"));
869 }
870 SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST,
871     mp_setvariables_for_up, NULL);
872 #endif /* SMP */
873 
874 void
875 smp_no_rendezvous_barrier(void *dummy)
876 {
877 #ifdef SMP
878 	KASSERT((!smp_started),("smp_no_rendezvous called and smp is started"));
879 #endif
880 }
881 
882 /*
883  * Wait for specified idle threads to switch once.  This ensures that even
884  * preempted threads have cycled through the switch function once,
885  * exiting their codepaths.  This allows us to change global pointers
886  * with no other synchronization.
887  */
888 int
889 quiesce_cpus(cpuset_t map, const char *wmesg, int prio)
890 {
891 	struct pcpu *pcpu;
892 	u_int gen[MAXCPU];
893 	int error;
894 	int cpu;
895 
896 	error = 0;
897 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
898 		if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu))
899 			continue;
900 		pcpu = pcpu_find(cpu);
901 		gen[cpu] = pcpu->pc_idlethread->td_generation;
902 	}
903 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
904 		if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu))
905 			continue;
906 		pcpu = pcpu_find(cpu);
907 		thread_lock(curthread);
908 		sched_bind(curthread, cpu);
909 		thread_unlock(curthread);
910 		while (gen[cpu] == pcpu->pc_idlethread->td_generation) {
911 			error = tsleep(quiesce_cpus, prio, wmesg, 1);
912 			if (error != EWOULDBLOCK)
913 				goto out;
914 			error = 0;
915 		}
916 	}
917 out:
918 	thread_lock(curthread);
919 	sched_unbind(curthread);
920 	thread_unlock(curthread);
921 
922 	return (error);
923 }
924 
925 int
926 quiesce_all_cpus(const char *wmesg, int prio)
927 {
928 
929 	return quiesce_cpus(all_cpus, wmesg, prio);
930 }
931 
932 /* Extra care is taken with this sysctl because the data type is volatile */
933 static int
934 sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS)
935 {
936 	int error, active;
937 
938 	active = smp_started;
939 	error = SYSCTL_OUT(req, &active, sizeof(active));
940 	return (error);
941 }
942 
943 
944 #ifdef SMP
945 void
946 topo_init_node(struct topo_node *node)
947 {
948 
949 	bzero(node, sizeof(*node));
950 	TAILQ_INIT(&node->children);
951 }
952 
953 void
954 topo_init_root(struct topo_node *root)
955 {
956 
957 	topo_init_node(root);
958 	root->type = TOPO_TYPE_SYSTEM;
959 }
960 
961 /*
962  * Add a child node with the given ID under the given parent.
963  * Do nothing if there is already a child with that ID.
964  */
965 struct topo_node *
966 topo_add_node_by_hwid(struct topo_node *parent, int hwid,
967     topo_node_type type, uintptr_t subtype)
968 {
969 	struct topo_node *node;
970 
971 	TAILQ_FOREACH_REVERSE(node, &parent->children,
972 	    topo_children, siblings) {
973 		if (node->hwid == hwid
974 		    && node->type == type && node->subtype == subtype) {
975 			return (node);
976 		}
977 	}
978 
979 	node = malloc(sizeof(*node), M_TOPO, M_WAITOK);
980 	topo_init_node(node);
981 	node->parent = parent;
982 	node->hwid = hwid;
983 	node->type = type;
984 	node->subtype = subtype;
985 	TAILQ_INSERT_TAIL(&parent->children, node, siblings);
986 	parent->nchildren++;
987 
988 	return (node);
989 }
990 
991 /*
992  * Find a child node with the given ID under the given parent.
993  */
994 struct topo_node *
995 topo_find_node_by_hwid(struct topo_node *parent, int hwid,
996     topo_node_type type, uintptr_t subtype)
997 {
998 
999 	struct topo_node *node;
1000 
1001 	TAILQ_FOREACH(node, &parent->children, siblings) {
1002 		if (node->hwid == hwid
1003 		    && node->type == type && node->subtype == subtype) {
1004 			return (node);
1005 		}
1006 	}
1007 
1008 	return (NULL);
1009 }
1010 
1011 /*
1012  * Given a node change the order of its parent's child nodes such
1013  * that the node becomes the firt child while preserving the cyclic
1014  * order of the children.  In other words, the given node is promoted
1015  * by rotation.
1016  */
1017 void
1018 topo_promote_child(struct topo_node *child)
1019 {
1020 	struct topo_node *next;
1021 	struct topo_node *node;
1022 	struct topo_node *parent;
1023 
1024 	parent = child->parent;
1025 	next = TAILQ_NEXT(child, siblings);
1026 	TAILQ_REMOVE(&parent->children, child, siblings);
1027 	TAILQ_INSERT_HEAD(&parent->children, child, siblings);
1028 
1029 	while (next != NULL) {
1030 		node = next;
1031 		next = TAILQ_NEXT(node, siblings);
1032 		TAILQ_REMOVE(&parent->children, node, siblings);
1033 		TAILQ_INSERT_AFTER(&parent->children, child, node, siblings);
1034 		child = node;
1035 	}
1036 }
1037 
1038 /*
1039  * Iterate to the next node in the depth-first search (traversal) of
1040  * the topology tree.
1041  */
1042 struct topo_node *
1043 topo_next_node(struct topo_node *top, struct topo_node *node)
1044 {
1045 	struct topo_node *next;
1046 
1047 	if ((next = TAILQ_FIRST(&node->children)) != NULL)
1048 		return (next);
1049 
1050 	if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1051 		return (next);
1052 
1053 	while (node != top && (node = node->parent) != top)
1054 		if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1055 			return (next);
1056 
1057 	return (NULL);
1058 }
1059 
1060 /*
1061  * Iterate to the next node in the depth-first search of the topology tree,
1062  * but without descending below the current node.
1063  */
1064 struct topo_node *
1065 topo_next_nonchild_node(struct topo_node *top, struct topo_node *node)
1066 {
1067 	struct topo_node *next;
1068 
1069 	if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1070 		return (next);
1071 
1072 	while (node != top && (node = node->parent) != top)
1073 		if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1074 			return (next);
1075 
1076 	return (NULL);
1077 }
1078 
1079 /*
1080  * Assign the given ID to the given topology node that represents a logical
1081  * processor.
1082  */
1083 void
1084 topo_set_pu_id(struct topo_node *node, cpuid_t id)
1085 {
1086 
1087 	KASSERT(node->type == TOPO_TYPE_PU,
1088 	    ("topo_set_pu_id: wrong node type: %u", node->type));
1089 	KASSERT(CPU_EMPTY(&node->cpuset) && node->cpu_count == 0,
1090 	    ("topo_set_pu_id: cpuset already not empty"));
1091 	node->id = id;
1092 	CPU_SET(id, &node->cpuset);
1093 	node->cpu_count = 1;
1094 	node->subtype = 1;
1095 
1096 	while ((node = node->parent) != NULL) {
1097 		KASSERT(!CPU_ISSET(id, &node->cpuset),
1098 		    ("logical ID %u is already set in node %p", id, node));
1099 		CPU_SET(id, &node->cpuset);
1100 		node->cpu_count++;
1101 	}
1102 }
1103 
1104 static struct topology_spec {
1105 	topo_node_type	type;
1106 	bool		match_subtype;
1107 	uintptr_t	subtype;
1108 } topology_level_table[TOPO_LEVEL_COUNT] = {
1109 	[TOPO_LEVEL_PKG] = { .type = TOPO_TYPE_PKG, },
1110 	[TOPO_LEVEL_GROUP] = { .type = TOPO_TYPE_GROUP, },
1111 	[TOPO_LEVEL_CACHEGROUP] = {
1112 		.type = TOPO_TYPE_CACHE,
1113 		.match_subtype = true,
1114 		.subtype = CG_SHARE_L3,
1115 	},
1116 	[TOPO_LEVEL_CORE] = { .type = TOPO_TYPE_CORE, },
1117 	[TOPO_LEVEL_THREAD] = { .type = TOPO_TYPE_PU, },
1118 };
1119 
1120 static bool
1121 topo_analyze_table(struct topo_node *root, int all, enum topo_level level,
1122     struct topo_analysis *results)
1123 {
1124 	struct topology_spec *spec;
1125 	struct topo_node *node;
1126 	int count;
1127 
1128 	if (level >= TOPO_LEVEL_COUNT)
1129 		return (true);
1130 
1131 	spec = &topology_level_table[level];
1132 	count = 0;
1133 	node = topo_next_node(root, root);
1134 
1135 	while (node != NULL) {
1136 		if (node->type != spec->type ||
1137 		    (spec->match_subtype && node->subtype != spec->subtype)) {
1138 			node = topo_next_node(root, node);
1139 			continue;
1140 		}
1141 		if (!all && CPU_EMPTY(&node->cpuset)) {
1142 			node = topo_next_nonchild_node(root, node);
1143 			continue;
1144 		}
1145 
1146 		count++;
1147 
1148 		if (!topo_analyze_table(node, all, level + 1, results))
1149 			return (false);
1150 
1151 		node = topo_next_nonchild_node(root, node);
1152 	}
1153 
1154 	/* No explicit subgroups is essentially one subgroup. */
1155 	if (count == 0) {
1156 		count = 1;
1157 
1158 		if (!topo_analyze_table(root, all, level + 1, results))
1159 			return (false);
1160 	}
1161 
1162 	if (results->entities[level] == -1)
1163 		results->entities[level] = count;
1164 	else if (results->entities[level] != count)
1165 		return (false);
1166 
1167 	return (true);
1168 }
1169 
1170 /*
1171  * Check if the topology is uniform, that is, each package has the same number
1172  * of cores in it and each core has the same number of threads (logical
1173  * processors) in it.  If so, calculate the number of packages, the number of
1174  * groups per package, the number of cachegroups per group, and the number of
1175  * logical processors per cachegroup.  'all' parameter tells whether to include
1176  * administratively disabled logical processors into the analysis.
1177  */
1178 int
1179 topo_analyze(struct topo_node *topo_root, int all,
1180     struct topo_analysis *results)
1181 {
1182 
1183 	results->entities[TOPO_LEVEL_PKG] = -1;
1184 	results->entities[TOPO_LEVEL_CORE] = -1;
1185 	results->entities[TOPO_LEVEL_THREAD] = -1;
1186 	results->entities[TOPO_LEVEL_GROUP] = -1;
1187 	results->entities[TOPO_LEVEL_CACHEGROUP] = -1;
1188 
1189 	if (!topo_analyze_table(topo_root, all, TOPO_LEVEL_PKG, results))
1190 		return (0);
1191 
1192 	KASSERT(results->entities[TOPO_LEVEL_PKG] > 0,
1193 		("bug in topology or analysis"));
1194 
1195 	return (1);
1196 }
1197 
1198 #endif /* SMP */
1199 
1200