xref: /freebsd/sys/kern/subr_smp.c (revision 721351876cd4d3a8a700f62d2061331fa951a488)
1 /*-
2  * Copyright (c) 2001
3  *	John Baldwin <jhb@FreeBSD.org>.  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  * 4. 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 JOHN BALDWIN 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 JOHN BALDWIN OR THE VOICES IN HIS HEAD
21  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
22  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
23  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
24  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
25  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
26  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
27  * THE POSSIBILITY OF 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/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 volatile cpumask_t stopped_cpus;
57 volatile cpumask_t started_cpus;
58 cpumask_t idle_cpus_mask;
59 cpumask_t hlt_cpus_mask;
60 cpumask_t logical_cpus_mask;
61 
62 void (*cpustop_restartfunc)(void);
63 #endif
64 /* This is used in modules that need to work in both SMP and UP. */
65 cpumask_t all_cpus;
66 
67 int mp_ncpus;
68 /* export this for libkvm consumers. */
69 int mp_maxcpus = MAXCPU;
70 
71 volatile int smp_started;
72 u_int mp_maxid;
73 
74 SYSCTL_NODE(_kern, OID_AUTO, smp, CTLFLAG_RD, NULL, "Kernel SMP");
75 
76 SYSCTL_INT(_kern_smp, OID_AUTO, maxid, CTLFLAG_RD, &mp_maxid, 0,
77     "Max CPU ID.");
78 
79 SYSCTL_INT(_kern_smp, OID_AUTO, maxcpus, CTLFLAG_RD, &mp_maxcpus, 0,
80     "Max number of CPUs that the system was compiled for.");
81 
82 int smp_active = 0;	/* are the APs allowed to run? */
83 SYSCTL_INT(_kern_smp, OID_AUTO, active, CTLFLAG_RW, &smp_active, 0,
84     "Number of Auxillary Processors (APs) that were successfully started");
85 
86 int smp_disabled = 0;	/* has smp been disabled? */
87 SYSCTL_INT(_kern_smp, OID_AUTO, disabled, CTLFLAG_RDTUN, &smp_disabled, 0,
88     "SMP has been disabled from the loader");
89 TUNABLE_INT("kern.smp.disabled", &smp_disabled);
90 
91 int smp_cpus = 1;	/* how many cpu's running */
92 SYSCTL_INT(_kern_smp, OID_AUTO, cpus, CTLFLAG_RD, &smp_cpus, 0,
93     "Number of CPUs online");
94 
95 int smp_topology = 0;	/* Which topology we're using. */
96 SYSCTL_INT(_kern_smp, OID_AUTO, topology, CTLFLAG_RD, &smp_topology, 0,
97     "Topology override setting; 0 is default provided by hardware.");
98 TUNABLE_INT("kern.smp.topology", &smp_topology);
99 
100 #ifdef SMP
101 /* Enable forwarding of a signal to a process running on a different CPU */
102 static int forward_signal_enabled = 1;
103 SYSCTL_INT(_kern_smp, OID_AUTO, forward_signal_enabled, CTLFLAG_RW,
104 	   &forward_signal_enabled, 0,
105 	   "Forwarding of a signal to a process on a different CPU");
106 
107 /* Enable forwarding of roundrobin to all other cpus */
108 static int forward_roundrobin_enabled = 1;
109 SYSCTL_INT(_kern_smp, OID_AUTO, forward_roundrobin_enabled, CTLFLAG_RW,
110 	   &forward_roundrobin_enabled, 0,
111 	   "Forwarding of roundrobin to all other CPUs");
112 
113 /* Variables needed for SMP rendezvous. */
114 static volatile cpumask_t smp_rv_cpumask;
115 static void (*volatile smp_rv_setup_func)(void *arg);
116 static void (*volatile smp_rv_action_func)(void *arg);
117 static void (*volatile smp_rv_teardown_func)(void *arg);
118 static void * volatile smp_rv_func_arg;
119 static volatile int smp_rv_waiters[3];
120 
121 /*
122  * Shared mutex to restrict busywaits between smp_rendezvous() and
123  * smp(_targeted)_tlb_shootdown().  A deadlock occurs if both of these
124  * functions trigger at once and cause multiple CPUs to busywait with
125  * interrupts disabled.
126  */
127 struct mtx smp_ipi_mtx;
128 
129 /*
130  * Let the MD SMP code initialize mp_maxid very early if it can.
131  */
132 static void
133 mp_setmaxid(void *dummy)
134 {
135 	cpu_mp_setmaxid();
136 }
137 SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL);
138 
139 /*
140  * Call the MD SMP initialization code.
141  */
142 static void
143 mp_start(void *dummy)
144 {
145 
146 	/* Probe for MP hardware. */
147 	if (smp_disabled != 0 || cpu_mp_probe() == 0) {
148 		mp_ncpus = 1;
149 		all_cpus = PCPU_GET(cpumask);
150 		return;
151 	}
152 
153 	mtx_init(&smp_ipi_mtx, "smp rendezvous", NULL, MTX_SPIN);
154 	cpu_mp_start();
155 	printf("FreeBSD/SMP: Multiprocessor System Detected: %d CPUs\n",
156 	    mp_ncpus);
157 	cpu_mp_announce();
158 }
159 SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL);
160 
161 void
162 forward_signal(struct thread *td)
163 {
164 	int id;
165 
166 	/*
167 	 * signotify() has already set TDF_ASTPENDING and TDF_NEEDSIGCHECK on
168 	 * this thread, so all we need to do is poke it if it is currently
169 	 * executing so that it executes ast().
170 	 */
171 	THREAD_LOCK_ASSERT(td, MA_OWNED);
172 	KASSERT(TD_IS_RUNNING(td),
173 	    ("forward_signal: thread is not TDS_RUNNING"));
174 
175 	CTR1(KTR_SMP, "forward_signal(%p)", td->td_proc);
176 
177 	if (!smp_started || cold || panicstr)
178 		return;
179 	if (!forward_signal_enabled)
180 		return;
181 
182 	/* No need to IPI ourself. */
183 	if (td == curthread)
184 		return;
185 
186 	id = td->td_oncpu;
187 	if (id == NOCPU)
188 		return;
189 	ipi_selected(1 << id, IPI_AST);
190 }
191 
192 void
193 forward_roundrobin(void)
194 {
195 	struct pcpu *pc;
196 	struct thread *td;
197 	cpumask_t id, map, me;
198 
199 	CTR0(KTR_SMP, "forward_roundrobin()");
200 
201 	if (!smp_started || cold || panicstr)
202 		return;
203 	if (!forward_roundrobin_enabled)
204 		return;
205 	map = 0;
206 	me = PCPU_GET(cpumask);
207 	SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
208 		td = pc->pc_curthread;
209 		id = pc->pc_cpumask;
210 		if (id != me && (id & stopped_cpus) == 0 &&
211 		    !TD_IS_IDLETHREAD(td)) {
212 			td->td_flags |= TDF_NEEDRESCHED;
213 			map |= id;
214 		}
215 	}
216 	ipi_selected(map, IPI_AST);
217 }
218 
219 /*
220  * When called the executing CPU will send an IPI to all other CPUs
221  *  requesting that they halt execution.
222  *
223  * Usually (but not necessarily) called with 'other_cpus' as its arg.
224  *
225  *  - Signals all CPUs in map to stop.
226  *  - Waits for each to stop.
227  *
228  * Returns:
229  *  -1: error
230  *   0: NA
231  *   1: ok
232  *
233  * XXX FIXME: this is not MP-safe, needs a lock to prevent multiple CPUs
234  *            from executing at same time.
235  */
236 int
237 stop_cpus(cpumask_t map)
238 {
239 	int i;
240 
241 	if (!smp_started)
242 		return 0;
243 
244 	CTR1(KTR_SMP, "stop_cpus(%x)", map);
245 
246 	/* send the stop IPI to all CPUs in map */
247 	ipi_selected(map, IPI_STOP);
248 
249 	i = 0;
250 	while ((stopped_cpus & map) != map) {
251 		/* spin */
252 		cpu_spinwait();
253 		i++;
254 #ifdef DIAGNOSTIC
255 		if (i == 100000) {
256 			printf("timeout stopping cpus\n");
257 			break;
258 		}
259 #endif
260 	}
261 
262 	return 1;
263 }
264 
265 /*
266  * Called by a CPU to restart stopped CPUs.
267  *
268  * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
269  *
270  *  - Signals all CPUs in map to restart.
271  *  - Waits for each to restart.
272  *
273  * Returns:
274  *  -1: error
275  *   0: NA
276  *   1: ok
277  */
278 int
279 restart_cpus(cpumask_t map)
280 {
281 
282 	if (!smp_started)
283 		return 0;
284 
285 	CTR1(KTR_SMP, "restart_cpus(%x)", map);
286 
287 	/* signal other cpus to restart */
288 	atomic_store_rel_int(&started_cpus, map);
289 
290 	/* wait for each to clear its bit */
291 	while ((stopped_cpus & map) != 0)
292 		cpu_spinwait();
293 
294 	return 1;
295 }
296 
297 /*
298  * All-CPU rendezvous.  CPUs are signalled, all execute the setup function
299  * (if specified), rendezvous, execute the action function (if specified),
300  * rendezvous again, execute the teardown function (if specified), and then
301  * resume.
302  *
303  * Note that the supplied external functions _must_ be reentrant and aware
304  * that they are running in parallel and in an unknown lock context.
305  */
306 void
307 smp_rendezvous_action(void)
308 {
309 	cpumask_t map = smp_rv_cpumask;
310 	int i, ncpus = 0;
311 	void* local_func_arg = smp_rv_func_arg;
312 	void (*local_setup_func)(void*)   = smp_rv_setup_func;
313 	void (*local_action_func)(void*)   = smp_rv_action_func;
314 	void (*local_teardown_func)(void*) = smp_rv_teardown_func;
315 
316 	for (i = 0; i < MAXCPU; i++)
317 		if (((1 << i) & map) != 0 && pcpu_find(i) != NULL)
318 			ncpus++;
319 
320 	/* Ensure we have up-to-date values. */
321 	atomic_add_acq_int(&smp_rv_waiters[0], 1);
322 	while (smp_rv_waiters[0] < ncpus)
323 		cpu_spinwait();
324 
325 	/* setup function */
326 	if (local_setup_func != smp_no_rendevous_barrier) {
327 		if (smp_rv_setup_func != NULL)
328 			smp_rv_setup_func(smp_rv_func_arg);
329 
330 		/* spin on entry rendezvous */
331 		atomic_add_int(&smp_rv_waiters[1], 1);
332 		while (smp_rv_waiters[1] < ncpus)
333                 	cpu_spinwait();
334 	}
335 
336 	/* action function */
337 	if (local_action_func != NULL)
338 		local_action_func(local_func_arg);
339 
340 	/* spin on exit rendezvous */
341 	atomic_add_int(&smp_rv_waiters[2], 1);
342 	if (local_teardown_func == smp_no_rendevous_barrier)
343                 return;
344 	while (smp_rv_waiters[2] < ncpus)
345 		cpu_spinwait();
346 
347 	/* teardown function */
348 	if (local_teardown_func != NULL)
349 		local_teardown_func(local_func_arg);
350 }
351 
352 void
353 smp_rendezvous_cpus(cpumask_t map,
354 	void (* setup_func)(void *),
355 	void (* action_func)(void *),
356 	void (* teardown_func)(void *),
357 	void *arg)
358 {
359 	int i, ncpus = 0;
360 
361 	if (!smp_started) {
362 		if (setup_func != NULL)
363 			setup_func(arg);
364 		if (action_func != NULL)
365 			action_func(arg);
366 		if (teardown_func != NULL)
367 			teardown_func(arg);
368 		return;
369 	}
370 
371 	for (i = 0; i < MAXCPU; i++)
372 		if (((1 << i) & map) != 0 && pcpu_find(i) != NULL)
373 			ncpus++;
374 
375 	/* obtain rendezvous lock */
376 	mtx_lock_spin(&smp_ipi_mtx);
377 
378 	/* set static function pointers */
379 	smp_rv_cpumask = map;
380 	smp_rv_setup_func = setup_func;
381 	smp_rv_action_func = action_func;
382 	smp_rv_teardown_func = teardown_func;
383 	smp_rv_func_arg = arg;
384 	smp_rv_waiters[1] = 0;
385 	smp_rv_waiters[2] = 0;
386 	atomic_store_rel_int(&smp_rv_waiters[0], 0);
387 
388 	/* signal other processors, which will enter the IPI with interrupts off */
389 	ipi_selected(map & ~(1 << curcpu), IPI_RENDEZVOUS);
390 
391 	/* Check if the current CPU is in the map */
392 	if ((map & (1 << curcpu)) != 0)
393 		smp_rendezvous_action();
394 
395 	if (teardown_func == smp_no_rendevous_barrier)
396 		while (atomic_load_acq_int(&smp_rv_waiters[2]) < ncpus)
397 			cpu_spinwait();
398 
399 	/* release lock */
400 	mtx_unlock_spin(&smp_ipi_mtx);
401 }
402 
403 void
404 smp_rendezvous(void (* setup_func)(void *),
405 	       void (* action_func)(void *),
406 	       void (* teardown_func)(void *),
407 	       void *arg)
408 {
409 	smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg);
410 }
411 
412 static struct cpu_group group[MAXCPU];
413 
414 struct cpu_group *
415 smp_topo(void)
416 {
417 	struct cpu_group *top;
418 
419 	/*
420 	 * Check for a fake topology request for debugging purposes.
421 	 */
422 	switch (smp_topology) {
423 	case 1:
424 		/* Dual core with no sharing.  */
425 		top = smp_topo_1level(CG_SHARE_NONE, 2, 0);
426 		break;
427 	case 2:
428 		/* No topology, all cpus are equal. */
429 		top = smp_topo_none();
430 		break;
431 	case 3:
432 		/* Dual core with shared L2.  */
433 		top = smp_topo_1level(CG_SHARE_L2, 2, 0);
434 		break;
435 	case 4:
436 		/* quad core, shared l3 among each package, private l2.  */
437 		top = smp_topo_1level(CG_SHARE_L3, 4, 0);
438 		break;
439 	case 5:
440 		/* quad core,  2 dualcore parts on each package share l2.  */
441 		top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0);
442 		break;
443 	case 6:
444 		/* Single-core 2xHTT */
445 		top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT);
446 		break;
447 	case 7:
448 		/* quad core with a shared l3, 8 threads sharing L2.  */
449 		top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8,
450 		    CG_FLAG_THREAD);
451 		break;
452 	default:
453 		/* Default, ask the system what it wants. */
454 		top = cpu_topo();
455 		break;
456 	}
457 	/*
458 	 * Verify the returned topology.
459 	 */
460 	if (top->cg_count != mp_ncpus)
461 		panic("Built bad topology at %p.  CPU count %d != %d",
462 		    top, top->cg_count, mp_ncpus);
463 	if (top->cg_mask != all_cpus)
464 		panic("Built bad topology at %p.  CPU mask 0x%X != 0x%X",
465 		    top, top->cg_mask, all_cpus);
466 	return (top);
467 }
468 
469 struct cpu_group *
470 smp_topo_none(void)
471 {
472 	struct cpu_group *top;
473 
474 	top = &group[0];
475 	top->cg_parent = NULL;
476 	top->cg_child = NULL;
477 	top->cg_mask = (1 << mp_ncpus) - 1;
478 	top->cg_count = mp_ncpus;
479 	top->cg_children = 0;
480 	top->cg_level = CG_SHARE_NONE;
481 	top->cg_flags = 0;
482 
483 	return (top);
484 }
485 
486 static int
487 smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share,
488     int count, int flags, int start)
489 {
490 	cpumask_t mask;
491 	int i;
492 
493 	for (mask = 0, i = 0; i < count; i++, start++)
494 		mask |= (1 << start);
495 	child->cg_parent = parent;
496 	child->cg_child = NULL;
497 	child->cg_children = 0;
498 	child->cg_level = share;
499 	child->cg_count = count;
500 	child->cg_flags = flags;
501 	child->cg_mask = mask;
502 	parent->cg_children++;
503 	for (; parent != NULL; parent = parent->cg_parent) {
504 		if ((parent->cg_mask & child->cg_mask) != 0)
505 			panic("Duplicate children in %p.  mask 0x%X child 0x%X",
506 			    parent, parent->cg_mask, child->cg_mask);
507 		parent->cg_mask |= child->cg_mask;
508 		parent->cg_count += child->cg_count;
509 	}
510 
511 	return (start);
512 }
513 
514 struct cpu_group *
515 smp_topo_1level(int share, int count, int flags)
516 {
517 	struct cpu_group *child;
518 	struct cpu_group *top;
519 	int packages;
520 	int cpu;
521 	int i;
522 
523 	cpu = 0;
524 	top = &group[0];
525 	packages = mp_ncpus / count;
526 	top->cg_child = child = &group[1];
527 	top->cg_level = CG_SHARE_NONE;
528 	for (i = 0; i < packages; i++, child++)
529 		cpu = smp_topo_addleaf(top, child, share, count, flags, cpu);
530 	return (top);
531 }
532 
533 struct cpu_group *
534 smp_topo_2level(int l2share, int l2count, int l1share, int l1count,
535     int l1flags)
536 {
537 	struct cpu_group *top;
538 	struct cpu_group *l1g;
539 	struct cpu_group *l2g;
540 	int cpu;
541 	int i;
542 	int j;
543 
544 	cpu = 0;
545 	top = &group[0];
546 	l2g = &group[1];
547 	top->cg_child = l2g;
548 	top->cg_level = CG_SHARE_NONE;
549 	top->cg_children = mp_ncpus / (l2count * l1count);
550 	l1g = l2g + top->cg_children;
551 	for (i = 0; i < top->cg_children; i++, l2g++) {
552 		l2g->cg_parent = top;
553 		l2g->cg_child = l1g;
554 		l2g->cg_level = l2share;
555 		for (j = 0; j < l2count; j++, l1g++)
556 			cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count,
557 			    l1flags, cpu);
558 	}
559 	return (top);
560 }
561 
562 
563 struct cpu_group *
564 smp_topo_find(struct cpu_group *top, int cpu)
565 {
566 	struct cpu_group *cg;
567 	cpumask_t mask;
568 	int children;
569 	int i;
570 
571 	mask = (1 << cpu);
572 	cg = top;
573 	for (;;) {
574 		if ((cg->cg_mask & mask) == 0)
575 			return (NULL);
576 		if (cg->cg_children == 0)
577 			return (cg);
578 		children = cg->cg_children;
579 		for (i = 0, cg = cg->cg_child; i < children; cg++, i++)
580 			if ((cg->cg_mask & mask) != 0)
581 				break;
582 	}
583 	return (NULL);
584 }
585 #else /* !SMP */
586 
587 void
588 smp_rendezvous_cpus(cpumask_t map,
589 	void (*setup_func)(void *),
590 	void (*action_func)(void *),
591 	void (*teardown_func)(void *),
592 	void *arg)
593 {
594 	if (setup_func != NULL)
595 		setup_func(arg);
596 	if (action_func != NULL)
597 		action_func(arg);
598 	if (teardown_func != NULL)
599 		teardown_func(arg);
600 }
601 
602 void
603 smp_rendezvous(void (*setup_func)(void *),
604 	       void (*action_func)(void *),
605 	       void (*teardown_func)(void *),
606 	       void *arg)
607 {
608 
609 	if (setup_func != NULL)
610 		setup_func(arg);
611 	if (action_func != NULL)
612 		action_func(arg);
613 	if (teardown_func != NULL)
614 		teardown_func(arg);
615 }
616 
617 /*
618  * Provide dummy SMP support for UP kernels.  Modules that need to use SMP
619  * APIs will still work using this dummy support.
620  */
621 static void
622 mp_setvariables_for_up(void *dummy)
623 {
624 	mp_ncpus = 1;
625 	mp_maxid = PCPU_GET(cpuid);
626 	all_cpus = PCPU_GET(cpumask);
627 	KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero"));
628 }
629 SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST,
630     mp_setvariables_for_up, NULL);
631 #endif /* SMP */
632 
633 void
634 smp_no_rendevous_barrier(void *dummy)
635 {
636 #ifdef SMP
637 	KASSERT((!smp_started),("smp_no_rendevous called and smp is started"));
638 #endif
639 }
640