xref: /freebsd/sys/kern/subr_smp.c (revision b2db760808f74bb53c232900091c9da801ebbfcc)
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/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 /* Variables needed for SMP rendezvous. */
108 static volatile int smp_rv_ncpus;
109 static void (*volatile smp_rv_setup_func)(void *arg);
110 static void (*volatile smp_rv_action_func)(void *arg);
111 static void (*volatile smp_rv_teardown_func)(void *arg);
112 static void *volatile smp_rv_func_arg;
113 static volatile int smp_rv_waiters[3];
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  */
121 struct mtx smp_ipi_mtx;
122 
123 /*
124  * Let the MD SMP code initialize mp_maxid very early if it can.
125  */
126 static void
127 mp_setmaxid(void *dummy)
128 {
129 	cpu_mp_setmaxid();
130 }
131 SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL);
132 
133 /*
134  * Call the MD SMP initialization code.
135  */
136 static void
137 mp_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 		all_cpus = PCPU_GET(cpumask);
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 }
154 SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL);
155 
156 void
157 forward_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  * XXX FIXME: this is not MP-safe, needs a lock to prevent multiple CPUs
202  *            from executing at same time.
203  */
204 static int
205 generic_stop_cpus(cpumask_t map, u_int type)
206 {
207 	int i;
208 
209 	KASSERT(type == IPI_STOP || type == IPI_STOP_HARD,
210 	    ("%s: invalid stop type", __func__));
211 
212 	if (!smp_started)
213 		return 0;
214 
215 	CTR2(KTR_SMP, "stop_cpus(%x) with %u type", map, type);
216 
217 	/* send the stop IPI to all CPUs in map */
218 	ipi_selected(map, type);
219 
220 	i = 0;
221 	while ((stopped_cpus & map) != map) {
222 		/* spin */
223 		cpu_spinwait();
224 		i++;
225 #ifdef DIAGNOSTIC
226 		if (i == 100000) {
227 			printf("timeout stopping cpus\n");
228 			break;
229 		}
230 #endif
231 	}
232 
233 	return 1;
234 }
235 
236 int
237 stop_cpus(cpumask_t map)
238 {
239 
240 	return (generic_stop_cpus(map, IPI_STOP));
241 }
242 
243 int
244 stop_cpus_hard(cpumask_t map)
245 {
246 
247 	return (generic_stop_cpus(map, IPI_STOP_HARD));
248 }
249 
250 #if defined(__amd64__)
251 /*
252  * When called the executing CPU will send an IPI to all other CPUs
253  *  requesting that they halt execution.
254  *
255  * Usually (but not necessarily) called with 'other_cpus' as its arg.
256  *
257  *  - Signals all CPUs in map to suspend.
258  *  - Waits for each to suspend.
259  *
260  * Returns:
261  *  -1: error
262  *   0: NA
263  *   1: ok
264  *
265  * XXX FIXME: this is not MP-safe, needs a lock to prevent multiple CPUs
266  *            from executing at same time.
267  */
268 int
269 suspend_cpus(cpumask_t map)
270 {
271 	int i;
272 
273 	if (!smp_started)
274 		return (0);
275 
276 	CTR1(KTR_SMP, "suspend_cpus(%x)", map);
277 
278 	/* send the suspend IPI to all CPUs in map */
279 	ipi_selected(map, IPI_SUSPEND);
280 
281 	i = 0;
282 	while ((stopped_cpus & map) != map) {
283 		/* spin */
284 		cpu_spinwait();
285 		i++;
286 #ifdef DIAGNOSTIC
287 		if (i == 100000) {
288 			printf("timeout suspending cpus\n");
289 			break;
290 		}
291 #endif
292 	}
293 
294 	return (1);
295 }
296 #endif
297 
298 /*
299  * Called by a CPU to restart stopped CPUs.
300  *
301  * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
302  *
303  *  - Signals all CPUs in map to restart.
304  *  - Waits for each to restart.
305  *
306  * Returns:
307  *  -1: error
308  *   0: NA
309  *   1: ok
310  */
311 int
312 restart_cpus(cpumask_t map)
313 {
314 
315 	if (!smp_started)
316 		return 0;
317 
318 	CTR1(KTR_SMP, "restart_cpus(%x)", map);
319 
320 	/* signal other cpus to restart */
321 	atomic_store_rel_int(&started_cpus, map);
322 
323 	/* wait for each to clear its bit */
324 	while ((stopped_cpus & map) != 0)
325 		cpu_spinwait();
326 
327 	return 1;
328 }
329 
330 /*
331  * All-CPU rendezvous.  CPUs are signalled, all execute the setup function
332  * (if specified), rendezvous, execute the action function (if specified),
333  * rendezvous again, execute the teardown function (if specified), and then
334  * resume.
335  *
336  * Note that the supplied external functions _must_ be reentrant and aware
337  * that they are running in parallel and in an unknown lock context.
338  */
339 void
340 smp_rendezvous_action(void)
341 {
342 	void* local_func_arg = smp_rv_func_arg;
343 	void (*local_setup_func)(void*)   = smp_rv_setup_func;
344 	void (*local_action_func)(void*)   = smp_rv_action_func;
345 	void (*local_teardown_func)(void*) = smp_rv_teardown_func;
346 
347 	/* Ensure we have up-to-date values. */
348 	atomic_add_acq_int(&smp_rv_waiters[0], 1);
349 	while (smp_rv_waiters[0] < smp_rv_ncpus)
350 		cpu_spinwait();
351 
352 	/* setup function */
353 	if (local_setup_func != smp_no_rendevous_barrier) {
354 		if (smp_rv_setup_func != NULL)
355 			smp_rv_setup_func(smp_rv_func_arg);
356 
357 		/* spin on entry rendezvous */
358 		atomic_add_int(&smp_rv_waiters[1], 1);
359 		while (smp_rv_waiters[1] < smp_rv_ncpus)
360                 	cpu_spinwait();
361 	}
362 
363 	/* action function */
364 	if (local_action_func != NULL)
365 		local_action_func(local_func_arg);
366 
367 	/* spin on exit rendezvous */
368 	atomic_add_int(&smp_rv_waiters[2], 1);
369 	if (local_teardown_func == smp_no_rendevous_barrier)
370                 return;
371 	while (smp_rv_waiters[2] < smp_rv_ncpus)
372 		cpu_spinwait();
373 
374 	/* teardown function */
375 	if (local_teardown_func != NULL)
376 		local_teardown_func(local_func_arg);
377 }
378 
379 void
380 smp_rendezvous_cpus(cpumask_t map,
381 	void (* setup_func)(void *),
382 	void (* action_func)(void *),
383 	void (* teardown_func)(void *),
384 	void *arg)
385 {
386 	int i, ncpus = 0;
387 
388 	if (!smp_started) {
389 		if (setup_func != NULL)
390 			setup_func(arg);
391 		if (action_func != NULL)
392 			action_func(arg);
393 		if (teardown_func != NULL)
394 			teardown_func(arg);
395 		return;
396 	}
397 
398 	CPU_FOREACH(i) {
399 		if (((1 << i) & map) != 0)
400 			ncpus++;
401 	}
402 	if (ncpus == 0)
403 		panic("ncpus is 0 with map=0x%x", map);
404 
405 	/* obtain rendezvous lock */
406 	mtx_lock_spin(&smp_ipi_mtx);
407 
408 	/* set static function pointers */
409 	smp_rv_ncpus = ncpus;
410 	smp_rv_setup_func = setup_func;
411 	smp_rv_action_func = action_func;
412 	smp_rv_teardown_func = teardown_func;
413 	smp_rv_func_arg = arg;
414 	smp_rv_waiters[1] = 0;
415 	smp_rv_waiters[2] = 0;
416 	atomic_store_rel_int(&smp_rv_waiters[0], 0);
417 
418 	/* signal other processors, which will enter the IPI with interrupts off */
419 	ipi_selected(map & ~(1 << curcpu), IPI_RENDEZVOUS);
420 
421 	/* Check if the current CPU is in the map */
422 	if ((map & (1 << curcpu)) != 0)
423 		smp_rendezvous_action();
424 
425 	if (teardown_func == smp_no_rendevous_barrier)
426 		while (atomic_load_acq_int(&smp_rv_waiters[2]) < ncpus)
427 			cpu_spinwait();
428 
429 	/* release lock */
430 	mtx_unlock_spin(&smp_ipi_mtx);
431 }
432 
433 void
434 smp_rendezvous(void (* setup_func)(void *),
435 	       void (* action_func)(void *),
436 	       void (* teardown_func)(void *),
437 	       void *arg)
438 {
439 	smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg);
440 }
441 
442 static struct cpu_group group[MAXCPU];
443 
444 struct cpu_group *
445 smp_topo(void)
446 {
447 	struct cpu_group *top;
448 
449 	/*
450 	 * Check for a fake topology request for debugging purposes.
451 	 */
452 	switch (smp_topology) {
453 	case 1:
454 		/* Dual core with no sharing.  */
455 		top = smp_topo_1level(CG_SHARE_NONE, 2, 0);
456 		break;
457 	case 2:
458 		/* No topology, all cpus are equal. */
459 		top = smp_topo_none();
460 		break;
461 	case 3:
462 		/* Dual core with shared L2.  */
463 		top = smp_topo_1level(CG_SHARE_L2, 2, 0);
464 		break;
465 	case 4:
466 		/* quad core, shared l3 among each package, private l2.  */
467 		top = smp_topo_1level(CG_SHARE_L3, 4, 0);
468 		break;
469 	case 5:
470 		/* quad core,  2 dualcore parts on each package share l2.  */
471 		top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0);
472 		break;
473 	case 6:
474 		/* Single-core 2xHTT */
475 		top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT);
476 		break;
477 	case 7:
478 		/* quad core with a shared l3, 8 threads sharing L2.  */
479 		top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8,
480 		    CG_FLAG_SMT);
481 		break;
482 	default:
483 		/* Default, ask the system what it wants. */
484 		top = cpu_topo();
485 		break;
486 	}
487 	/*
488 	 * Verify the returned topology.
489 	 */
490 	if (top->cg_count != mp_ncpus)
491 		panic("Built bad topology at %p.  CPU count %d != %d",
492 		    top, top->cg_count, mp_ncpus);
493 	if (top->cg_mask != all_cpus)
494 		panic("Built bad topology at %p.  CPU mask 0x%X != 0x%X",
495 		    top, top->cg_mask, all_cpus);
496 	return (top);
497 }
498 
499 struct cpu_group *
500 smp_topo_none(void)
501 {
502 	struct cpu_group *top;
503 
504 	top = &group[0];
505 	top->cg_parent = NULL;
506 	top->cg_child = NULL;
507 	top->cg_mask = ~0U >> (32 - mp_ncpus);
508 	top->cg_count = mp_ncpus;
509 	top->cg_children = 0;
510 	top->cg_level = CG_SHARE_NONE;
511 	top->cg_flags = 0;
512 
513 	return (top);
514 }
515 
516 static int
517 smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share,
518     int count, int flags, int start)
519 {
520 	cpumask_t mask;
521 	int i;
522 
523 	for (mask = 0, i = 0; i < count; i++, start++)
524 		mask |= (1 << start);
525 	child->cg_parent = parent;
526 	child->cg_child = NULL;
527 	child->cg_children = 0;
528 	child->cg_level = share;
529 	child->cg_count = count;
530 	child->cg_flags = flags;
531 	child->cg_mask = mask;
532 	parent->cg_children++;
533 	for (; parent != NULL; parent = parent->cg_parent) {
534 		if ((parent->cg_mask & child->cg_mask) != 0)
535 			panic("Duplicate children in %p.  mask 0x%X child 0x%X",
536 			    parent, parent->cg_mask, child->cg_mask);
537 		parent->cg_mask |= child->cg_mask;
538 		parent->cg_count += child->cg_count;
539 	}
540 
541 	return (start);
542 }
543 
544 struct cpu_group *
545 smp_topo_1level(int share, int count, int flags)
546 {
547 	struct cpu_group *child;
548 	struct cpu_group *top;
549 	int packages;
550 	int cpu;
551 	int i;
552 
553 	cpu = 0;
554 	top = &group[0];
555 	packages = mp_ncpus / count;
556 	top->cg_child = child = &group[1];
557 	top->cg_level = CG_SHARE_NONE;
558 	for (i = 0; i < packages; i++, child++)
559 		cpu = smp_topo_addleaf(top, child, share, count, flags, cpu);
560 	return (top);
561 }
562 
563 struct cpu_group *
564 smp_topo_2level(int l2share, int l2count, int l1share, int l1count,
565     int l1flags)
566 {
567 	struct cpu_group *top;
568 	struct cpu_group *l1g;
569 	struct cpu_group *l2g;
570 	int cpu;
571 	int i;
572 	int j;
573 
574 	cpu = 0;
575 	top = &group[0];
576 	l2g = &group[1];
577 	top->cg_child = l2g;
578 	top->cg_level = CG_SHARE_NONE;
579 	top->cg_children = mp_ncpus / (l2count * l1count);
580 	l1g = l2g + top->cg_children;
581 	for (i = 0; i < top->cg_children; i++, l2g++) {
582 		l2g->cg_parent = top;
583 		l2g->cg_child = l1g;
584 		l2g->cg_level = l2share;
585 		for (j = 0; j < l2count; j++, l1g++)
586 			cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count,
587 			    l1flags, cpu);
588 	}
589 	return (top);
590 }
591 
592 
593 struct cpu_group *
594 smp_topo_find(struct cpu_group *top, int cpu)
595 {
596 	struct cpu_group *cg;
597 	cpumask_t mask;
598 	int children;
599 	int i;
600 
601 	mask = (1 << cpu);
602 	cg = top;
603 	for (;;) {
604 		if ((cg->cg_mask & mask) == 0)
605 			return (NULL);
606 		if (cg->cg_children == 0)
607 			return (cg);
608 		children = cg->cg_children;
609 		for (i = 0, cg = cg->cg_child; i < children; cg++, i++)
610 			if ((cg->cg_mask & mask) != 0)
611 				break;
612 	}
613 	return (NULL);
614 }
615 #else /* !SMP */
616 
617 void
618 smp_rendezvous_cpus(cpumask_t map,
619 	void (*setup_func)(void *),
620 	void (*action_func)(void *),
621 	void (*teardown_func)(void *),
622 	void *arg)
623 {
624 	if (setup_func != NULL)
625 		setup_func(arg);
626 	if (action_func != NULL)
627 		action_func(arg);
628 	if (teardown_func != NULL)
629 		teardown_func(arg);
630 }
631 
632 void
633 smp_rendezvous(void (*setup_func)(void *),
634 	       void (*action_func)(void *),
635 	       void (*teardown_func)(void *),
636 	       void *arg)
637 {
638 
639 	if (setup_func != NULL)
640 		setup_func(arg);
641 	if (action_func != NULL)
642 		action_func(arg);
643 	if (teardown_func != NULL)
644 		teardown_func(arg);
645 }
646 
647 /*
648  * Provide dummy SMP support for UP kernels.  Modules that need to use SMP
649  * APIs will still work using this dummy support.
650  */
651 static void
652 mp_setvariables_for_up(void *dummy)
653 {
654 	mp_ncpus = 1;
655 	mp_maxid = PCPU_GET(cpuid);
656 	all_cpus = PCPU_GET(cpumask);
657 	KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero"));
658 }
659 SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST,
660     mp_setvariables_for_up, NULL);
661 #endif /* SMP */
662 
663 void
664 smp_no_rendevous_barrier(void *dummy)
665 {
666 #ifdef SMP
667 	KASSERT((!smp_started),("smp_no_rendevous called and smp is started"));
668 #endif
669 }
670