FreeBSD/Linux Kernel Cross Reference
sys/kern/subr_smp.c
1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c) 2001, John Baldwin <jhb@FreeBSD.org>.
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26 * SUCH DAMAGE.
27 */
28
29 /*
30 * This module holds the global variables and machine independent functions
31 * used for the kernel SMP support.
32 */
33
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD: releng/12.0/sys/kern/subr_smp.c 327056 2017-12-21 09:17:48Z bde $");
36
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
40 #include <sys/ktr.h>
41 #include <sys/proc.h>
42 #include <sys/bus.h>
43 #include <sys/lock.h>
44 #include <sys/malloc.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 MALLOC_DEFINE(M_TOPO, "toponodes", "SMP topology data");
58
59 volatile cpuset_t stopped_cpus;
60 volatile cpuset_t started_cpus;
61 volatile cpuset_t suspended_cpus;
62 cpuset_t hlt_cpus_mask;
63 cpuset_t logical_cpus_mask;
64
65 void (*cpustop_restartfunc)(void);
66 #endif
67
68 static int sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS);
69
70 /* This is used in modules that need to work in both SMP and UP. */
71 cpuset_t all_cpus;
72
73 int mp_ncpus;
74 /* export this for libkvm consumers. */
75 int mp_maxcpus = MAXCPU;
76
77 volatile int smp_started;
78 u_int mp_maxid;
79
80 static SYSCTL_NODE(_kern, OID_AUTO, smp, CTLFLAG_RD|CTLFLAG_CAPRD, NULL,
81 "Kernel SMP");
82
83 SYSCTL_INT(_kern_smp, OID_AUTO, maxid, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxid, 0,
84 "Max CPU ID.");
85
86 SYSCTL_INT(_kern_smp, OID_AUTO, maxcpus, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxcpus,
87 0, "Max number of CPUs that the system was compiled for.");
88
89 SYSCTL_PROC(_kern_smp, OID_AUTO, active, CTLFLAG_RD|CTLTYPE_INT|CTLFLAG_MPSAFE,
90 NULL, 0, sysctl_kern_smp_active, "I",
91 "Indicates system is running in SMP mode");
92
93 int smp_disabled = 0; /* has smp been disabled? */
94 SYSCTL_INT(_kern_smp, OID_AUTO, disabled, CTLFLAG_RDTUN|CTLFLAG_CAPRD,
95 &smp_disabled, 0, "SMP has been disabled from the loader");
96
97 int smp_cpus = 1; /* how many cpu's running */
98 SYSCTL_INT(_kern_smp, OID_AUTO, cpus, CTLFLAG_RD|CTLFLAG_CAPRD, &smp_cpus, 0,
99 "Number of CPUs online");
100
101 int smp_topology = 0; /* Which topology we're using. */
102 SYSCTL_INT(_kern_smp, OID_AUTO, topology, CTLFLAG_RDTUN, &smp_topology, 0,
103 "Topology override setting; 0 is default provided by hardware.");
104
105 #ifdef SMP
106 /* Enable forwarding of a signal to a process running on a different CPU */
107 static int forward_signal_enabled = 1;
108 SYSCTL_INT(_kern_smp, OID_AUTO, forward_signal_enabled, CTLFLAG_RW,
109 &forward_signal_enabled, 0,
110 "Forwarding of a signal to a process on a different CPU");
111
112 /* Variables needed for SMP rendezvous. */
113 static volatile int smp_rv_ncpus;
114 static void (*volatile smp_rv_setup_func)(void *arg);
115 static void (*volatile smp_rv_action_func)(void *arg);
116 static void (*volatile smp_rv_teardown_func)(void *arg);
117 static void *volatile smp_rv_func_arg;
118 static volatile int smp_rv_waiters[4];
119
120 /*
121 * Shared mutex to restrict busywaits between smp_rendezvous() and
122 * smp(_targeted)_tlb_shootdown(). A deadlock occurs if both of these
123 * functions trigger at once and cause multiple CPUs to busywait with
124 * interrupts disabled.
125 */
126 struct mtx smp_ipi_mtx;
127
128 /*
129 * Let the MD SMP code initialize mp_maxid very early if it can.
130 */
131 static void
132 mp_setmaxid(void *dummy)
133 {
134
135 cpu_mp_setmaxid();
136
137 KASSERT(mp_ncpus >= 1, ("%s: CPU count < 1", __func__));
138 KASSERT(mp_ncpus > 1 || mp_maxid == 0,
139 ("%s: one CPU but mp_maxid is not zero", __func__));
140 KASSERT(mp_maxid >= mp_ncpus - 1,
141 ("%s: counters out of sync: max %d, count %d", __func__,
142 mp_maxid, mp_ncpus));
143 }
144 SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL);
145
146 /*
147 * Call the MD SMP initialization code.
148 */
149 static void
150 mp_start(void *dummy)
151 {
152
153 mtx_init(&smp_ipi_mtx, "smp rendezvous", NULL, MTX_SPIN);
154
155 /* Probe for MP hardware. */
156 if (smp_disabled != 0 || cpu_mp_probe() == 0) {
157 mp_ncpus = 1;
158 CPU_SETOF(PCPU_GET(cpuid), &all_cpus);
159 return;
160 }
161
162 cpu_mp_start();
163 printf("FreeBSD/SMP: Multiprocessor System Detected: %d CPUs\n",
164 mp_ncpus);
165 cpu_mp_announce();
166 }
167 SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL);
168
169 void
170 forward_signal(struct thread *td)
171 {
172 int id;
173
174 /*
175 * signotify() has already set TDF_ASTPENDING and TDF_NEEDSIGCHECK on
176 * this thread, so all we need to do is poke it if it is currently
177 * executing so that it executes ast().
178 */
179 THREAD_LOCK_ASSERT(td, MA_OWNED);
180 KASSERT(TD_IS_RUNNING(td),
181 ("forward_signal: thread is not TDS_RUNNING"));
182
183 CTR1(KTR_SMP, "forward_signal(%p)", td->td_proc);
184
185 if (!smp_started || cold || panicstr)
186 return;
187 if (!forward_signal_enabled)
188 return;
189
190 /* No need to IPI ourself. */
191 if (td == curthread)
192 return;
193
194 id = td->td_oncpu;
195 if (id == NOCPU)
196 return;
197 ipi_cpu(id, IPI_AST);
198 }
199
200 /*
201 * When called the executing CPU will send an IPI to all other CPUs
202 * requesting that they halt execution.
203 *
204 * Usually (but not necessarily) called with 'other_cpus' as its arg.
205 *
206 * - Signals all CPUs in map to stop.
207 * - Waits for each to stop.
208 *
209 * Returns:
210 * -1: error
211 * 0: NA
212 * 1: ok
213 *
214 */
215 #if defined(__amd64__) || defined(__i386__)
216 #define X86 1
217 #else
218 #define X86 0
219 #endif
220 static int
221 generic_stop_cpus(cpuset_t map, u_int type)
222 {
223 #ifdef KTR
224 char cpusetbuf[CPUSETBUFSIZ];
225 #endif
226 static volatile u_int stopping_cpu = NOCPU;
227 int i;
228 volatile cpuset_t *cpus;
229
230 KASSERT(
231 type == IPI_STOP || type == IPI_STOP_HARD
232 #if X86
233 || type == IPI_SUSPEND
234 #endif
235 , ("%s: invalid stop type", __func__));
236
237 if (!smp_started)
238 return (0);
239
240 CTR2(KTR_SMP, "stop_cpus(%s) with %u type",
241 cpusetobj_strprint(cpusetbuf, &map), type);
242
243 #if X86
244 /*
245 * When suspending, ensure there are are no IPIs in progress.
246 * IPIs that have been issued, but not yet delivered (e.g.
247 * not pending on a vCPU when running under virtualization)
248 * will be lost, violating FreeBSD's assumption of reliable
249 * IPI delivery.
250 */
251 if (type == IPI_SUSPEND)
252 mtx_lock_spin(&smp_ipi_mtx);
253 #endif
254
255 #if X86
256 if (!nmi_is_broadcast || nmi_kdb_lock == 0) {
257 #endif
258 if (stopping_cpu != PCPU_GET(cpuid))
259 while (atomic_cmpset_int(&stopping_cpu, NOCPU,
260 PCPU_GET(cpuid)) == 0)
261 while (stopping_cpu != NOCPU)
262 cpu_spinwait(); /* spin */
263
264 /* send the stop IPI to all CPUs in map */
265 ipi_selected(map, type);
266 #if X86
267 }
268 #endif
269
270 #if X86
271 if (type == IPI_SUSPEND)
272 cpus = &suspended_cpus;
273 else
274 #endif
275 cpus = &stopped_cpus;
276
277 i = 0;
278 while (!CPU_SUBSET(cpus, &map)) {
279 /* spin */
280 cpu_spinwait();
281 i++;
282 if (i == 100000000) {
283 printf("timeout stopping cpus\n");
284 break;
285 }
286 }
287
288 #if X86
289 if (type == IPI_SUSPEND)
290 mtx_unlock_spin(&smp_ipi_mtx);
291 #endif
292
293 stopping_cpu = NOCPU;
294 return (1);
295 }
296
297 int
298 stop_cpus(cpuset_t map)
299 {
300
301 return (generic_stop_cpus(map, IPI_STOP));
302 }
303
304 int
305 stop_cpus_hard(cpuset_t map)
306 {
307
308 return (generic_stop_cpus(map, IPI_STOP_HARD));
309 }
310
311 #if X86
312 int
313 suspend_cpus(cpuset_t map)
314 {
315
316 return (generic_stop_cpus(map, IPI_SUSPEND));
317 }
318 #endif
319
320 /*
321 * Called by a CPU to restart stopped CPUs.
322 *
323 * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
324 *
325 * - Signals all CPUs in map to restart.
326 * - Waits for each to restart.
327 *
328 * Returns:
329 * -1: error
330 * 0: NA
331 * 1: ok
332 */
333 static int
334 generic_restart_cpus(cpuset_t map, u_int type)
335 {
336 #ifdef KTR
337 char cpusetbuf[CPUSETBUFSIZ];
338 #endif
339 volatile cpuset_t *cpus;
340
341 KASSERT(type == IPI_STOP || type == IPI_STOP_HARD
342 #if X86
343 || type == IPI_SUSPEND
344 #endif
345 , ("%s: invalid stop type", __func__));
346
347 if (!smp_started)
348 return (0);
349
350 CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map));
351
352 #if X86
353 if (type == IPI_SUSPEND)
354 cpus = &resuming_cpus;
355 else
356 #endif
357 cpus = &stopped_cpus;
358
359 /* signal other cpus to restart */
360 #if X86
361 if (type == IPI_SUSPEND)
362 CPU_COPY_STORE_REL(&map, &toresume_cpus);
363 else
364 #endif
365 CPU_COPY_STORE_REL(&map, &started_cpus);
366
367 #if X86
368 if (!nmi_is_broadcast || nmi_kdb_lock == 0) {
369 #endif
370 /* wait for each to clear its bit */
371 while (CPU_OVERLAP(cpus, &map))
372 cpu_spinwait();
373 #if X86
374 }
375 #endif
376
377 return (1);
378 }
379
380 int
381 restart_cpus(cpuset_t map)
382 {
383
384 return (generic_restart_cpus(map, IPI_STOP));
385 }
386
387 #if X86
388 int
389 resume_cpus(cpuset_t map)
390 {
391
392 return (generic_restart_cpus(map, IPI_SUSPEND));
393 }
394 #endif
395 #undef X86
396
397 /*
398 * All-CPU rendezvous. CPUs are signalled, all execute the setup function
399 * (if specified), rendezvous, execute the action function (if specified),
400 * rendezvous again, execute the teardown function (if specified), and then
401 * resume.
402 *
403 * Note that the supplied external functions _must_ be reentrant and aware
404 * that they are running in parallel and in an unknown lock context.
405 */
406 void
407 smp_rendezvous_action(void)
408 {
409 struct thread *td;
410 void *local_func_arg;
411 void (*local_setup_func)(void*);
412 void (*local_action_func)(void*);
413 void (*local_teardown_func)(void*);
414 #ifdef INVARIANTS
415 int owepreempt;
416 #endif
417
418 /* Ensure we have up-to-date values. */
419 atomic_add_acq_int(&smp_rv_waiters[0], 1);
420 while (smp_rv_waiters[0] < smp_rv_ncpus)
421 cpu_spinwait();
422
423 /* Fetch rendezvous parameters after acquire barrier. */
424 local_func_arg = smp_rv_func_arg;
425 local_setup_func = smp_rv_setup_func;
426 local_action_func = smp_rv_action_func;
427 local_teardown_func = smp_rv_teardown_func;
428
429 /*
430 * Use a nested critical section to prevent any preemptions
431 * from occurring during a rendezvous action routine.
432 * Specifically, if a rendezvous handler is invoked via an IPI
433 * and the interrupted thread was in the critical_exit()
434 * function after setting td_critnest to 0 but before
435 * performing a deferred preemption, this routine can be
436 * invoked with td_critnest set to 0 and td_owepreempt true.
437 * In that case, a critical_exit() during the rendezvous
438 * action would trigger a preemption which is not permitted in
439 * a rendezvous action. To fix this, wrap all of the
440 * rendezvous action handlers in a critical section. We
441 * cannot use a regular critical section however as having
442 * critical_exit() preempt from this routine would also be
443 * problematic (the preemption must not occur before the IPI
444 * has been acknowledged via an EOI). Instead, we
445 * intentionally ignore td_owepreempt when leaving the
446 * critical section. This should be harmless because we do
447 * not permit rendezvous action routines to schedule threads,
448 * and thus td_owepreempt should never transition from 0 to 1
449 * during this routine.
450 */
451 td = curthread;
452 td->td_critnest++;
453 #ifdef INVARIANTS
454 owepreempt = td->td_owepreempt;
455 #endif
456
457 /*
458 * If requested, run a setup function before the main action
459 * function. Ensure all CPUs have completed the setup
460 * function before moving on to the action function.
461 */
462 if (local_setup_func != smp_no_rendezvous_barrier) {
463 if (smp_rv_setup_func != NULL)
464 smp_rv_setup_func(smp_rv_func_arg);
465 atomic_add_int(&smp_rv_waiters[1], 1);
466 while (smp_rv_waiters[1] < smp_rv_ncpus)
467 cpu_spinwait();
468 }
469
470 if (local_action_func != NULL)
471 local_action_func(local_func_arg);
472
473 if (local_teardown_func != smp_no_rendezvous_barrier) {
474 /*
475 * Signal that the main action has been completed. If a
476 * full exit rendezvous is requested, then all CPUs will
477 * wait here until all CPUs have finished the main action.
478 */
479 atomic_add_int(&smp_rv_waiters[2], 1);
480 while (smp_rv_waiters[2] < smp_rv_ncpus)
481 cpu_spinwait();
482
483 if (local_teardown_func != NULL)
484 local_teardown_func(local_func_arg);
485 }
486
487 /*
488 * Signal that the rendezvous is fully completed by this CPU.
489 * This means that no member of smp_rv_* pseudo-structure will be
490 * accessed by this target CPU after this point; in particular,
491 * memory pointed by smp_rv_func_arg.
492 *
493 * The release semantic ensures that all accesses performed by
494 * the current CPU are visible when smp_rendezvous_cpus()
495 * returns, by synchronizing with the
496 * atomic_load_acq_int(&smp_rv_waiters[3]).
497 */
498 atomic_add_rel_int(&smp_rv_waiters[3], 1);
499
500 td->td_critnest--;
501 KASSERT(owepreempt == td->td_owepreempt,
502 ("rendezvous action changed td_owepreempt"));
503 }
504
505 void
506 smp_rendezvous_cpus(cpuset_t map,
507 void (* setup_func)(void *),
508 void (* action_func)(void *),
509 void (* teardown_func)(void *),
510 void *arg)
511 {
512 int curcpumap, i, ncpus = 0;
513
514 /* Look comments in the !SMP case. */
515 if (!smp_started) {
516 spinlock_enter();
517 if (setup_func != NULL)
518 setup_func(arg);
519 if (action_func != NULL)
520 action_func(arg);
521 if (teardown_func != NULL)
522 teardown_func(arg);
523 spinlock_exit();
524 return;
525 }
526
527 CPU_FOREACH(i) {
528 if (CPU_ISSET(i, &map))
529 ncpus++;
530 }
531 if (ncpus == 0)
532 panic("ncpus is 0 with non-zero map");
533
534 mtx_lock_spin(&smp_ipi_mtx);
535
536 /* Pass rendezvous parameters via global variables. */
537 smp_rv_ncpus = ncpus;
538 smp_rv_setup_func = setup_func;
539 smp_rv_action_func = action_func;
540 smp_rv_teardown_func = teardown_func;
541 smp_rv_func_arg = arg;
542 smp_rv_waiters[1] = 0;
543 smp_rv_waiters[2] = 0;
544 smp_rv_waiters[3] = 0;
545 atomic_store_rel_int(&smp_rv_waiters[0], 0);
546
547 /*
548 * Signal other processors, which will enter the IPI with
549 * interrupts off.
550 */
551 curcpumap = CPU_ISSET(curcpu, &map);
552 CPU_CLR(curcpu, &map);
553 ipi_selected(map, IPI_RENDEZVOUS);
554
555 /* Check if the current CPU is in the map */
556 if (curcpumap != 0)
557 smp_rendezvous_action();
558
559 /*
560 * Ensure that the master CPU waits for all the other
561 * CPUs to finish the rendezvous, so that smp_rv_*
562 * pseudo-structure and the arg are guaranteed to not
563 * be in use.
564 *
565 * Load acquire synchronizes with the release add in
566 * smp_rendezvous_action(), which ensures that our caller sees
567 * all memory actions done by the called functions on other
568 * CPUs.
569 */
570 while (atomic_load_acq_int(&smp_rv_waiters[3]) < ncpus)
571 cpu_spinwait();
572
573 mtx_unlock_spin(&smp_ipi_mtx);
574 }
575
576 void
577 smp_rendezvous(void (* setup_func)(void *),
578 void (* action_func)(void *),
579 void (* teardown_func)(void *),
580 void *arg)
581 {
582 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg);
583 }
584
585 static struct cpu_group group[MAXCPU * MAX_CACHE_LEVELS + 1];
586
587 struct cpu_group *
588 smp_topo(void)
589 {
590 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
591 struct cpu_group *top;
592
593 /*
594 * Check for a fake topology request for debugging purposes.
595 */
596 switch (smp_topology) {
597 case 1:
598 /* Dual core with no sharing. */
599 top = smp_topo_1level(CG_SHARE_NONE, 2, 0);
600 break;
601 case 2:
602 /* No topology, all cpus are equal. */
603 top = smp_topo_none();
604 break;
605 case 3:
606 /* Dual core with shared L2. */
607 top = smp_topo_1level(CG_SHARE_L2, 2, 0);
608 break;
609 case 4:
610 /* quad core, shared l3 among each package, private l2. */
611 top = smp_topo_1level(CG_SHARE_L3, 4, 0);
612 break;
613 case 5:
614 /* quad core, 2 dualcore parts on each package share l2. */
615 top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0);
616 break;
617 case 6:
618 /* Single-core 2xHTT */
619 top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT);
620 break;
621 case 7:
622 /* quad core with a shared l3, 8 threads sharing L2. */
623 top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8,
624 CG_FLAG_SMT);
625 break;
626 default:
627 /* Default, ask the system what it wants. */
628 top = cpu_topo();
629 break;
630 }
631 /*
632 * Verify the returned topology.
633 */
634 if (top->cg_count != mp_ncpus)
635 panic("Built bad topology at %p. CPU count %d != %d",
636 top, top->cg_count, mp_ncpus);
637 if (CPU_CMP(&top->cg_mask, &all_cpus))
638 panic("Built bad topology at %p. CPU mask (%s) != (%s)",
639 top, cpusetobj_strprint(cpusetbuf, &top->cg_mask),
640 cpusetobj_strprint(cpusetbuf2, &all_cpus));
641
642 /*
643 * Collapse nonsense levels that may be created out of convenience by
644 * the MD layers. They cause extra work in the search functions.
645 */
646 while (top->cg_children == 1) {
647 top = &top->cg_child[0];
648 top->cg_parent = NULL;
649 }
650 return (top);
651 }
652
653 struct cpu_group *
654 smp_topo_alloc(u_int count)
655 {
656 static u_int index;
657 u_int curr;
658
659 curr = index;
660 index += count;
661 return (&group[curr]);
662 }
663
664 struct cpu_group *
665 smp_topo_none(void)
666 {
667 struct cpu_group *top;
668
669 top = &group[0];
670 top->cg_parent = NULL;
671 top->cg_child = NULL;
672 top->cg_mask = all_cpus;
673 top->cg_count = mp_ncpus;
674 top->cg_children = 0;
675 top->cg_level = CG_SHARE_NONE;
676 top->cg_flags = 0;
677
678 return (top);
679 }
680
681 static int
682 smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share,
683 int count, int flags, int start)
684 {
685 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
686 cpuset_t mask;
687 int i;
688
689 CPU_ZERO(&mask);
690 for (i = 0; i < count; i++, start++)
691 CPU_SET(start, &mask);
692 child->cg_parent = parent;
693 child->cg_child = NULL;
694 child->cg_children = 0;
695 child->cg_level = share;
696 child->cg_count = count;
697 child->cg_flags = flags;
698 child->cg_mask = mask;
699 parent->cg_children++;
700 for (; parent != NULL; parent = parent->cg_parent) {
701 if (CPU_OVERLAP(&parent->cg_mask, &child->cg_mask))
702 panic("Duplicate children in %p. mask (%s) child (%s)",
703 parent,
704 cpusetobj_strprint(cpusetbuf, &parent->cg_mask),
705 cpusetobj_strprint(cpusetbuf2, &child->cg_mask));
706 CPU_OR(&parent->cg_mask, &child->cg_mask);
707 parent->cg_count += child->cg_count;
708 }
709
710 return (start);
711 }
712
713 struct cpu_group *
714 smp_topo_1level(int share, int count, int flags)
715 {
716 struct cpu_group *child;
717 struct cpu_group *top;
718 int packages;
719 int cpu;
720 int i;
721
722 cpu = 0;
723 top = &group[0];
724 packages = mp_ncpus / count;
725 top->cg_child = child = &group[1];
726 top->cg_level = CG_SHARE_NONE;
727 for (i = 0; i < packages; i++, child++)
728 cpu = smp_topo_addleaf(top, child, share, count, flags, cpu);
729 return (top);
730 }
731
732 struct cpu_group *
733 smp_topo_2level(int l2share, int l2count, int l1share, int l1count,
734 int l1flags)
735 {
736 struct cpu_group *top;
737 struct cpu_group *l1g;
738 struct cpu_group *l2g;
739 int cpu;
740 int i;
741 int j;
742
743 cpu = 0;
744 top = &group[0];
745 l2g = &group[1];
746 top->cg_child = l2g;
747 top->cg_level = CG_SHARE_NONE;
748 top->cg_children = mp_ncpus / (l2count * l1count);
749 l1g = l2g + top->cg_children;
750 for (i = 0; i < top->cg_children; i++, l2g++) {
751 l2g->cg_parent = top;
752 l2g->cg_child = l1g;
753 l2g->cg_level = l2share;
754 for (j = 0; j < l2count; j++, l1g++)
755 cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count,
756 l1flags, cpu);
757 }
758 return (top);
759 }
760
761
762 struct cpu_group *
763 smp_topo_find(struct cpu_group *top, int cpu)
764 {
765 struct cpu_group *cg;
766 cpuset_t mask;
767 int children;
768 int i;
769
770 CPU_SETOF(cpu, &mask);
771 cg = top;
772 for (;;) {
773 if (!CPU_OVERLAP(&cg->cg_mask, &mask))
774 return (NULL);
775 if (cg->cg_children == 0)
776 return (cg);
777 children = cg->cg_children;
778 for (i = 0, cg = cg->cg_child; i < children; cg++, i++)
779 if (CPU_OVERLAP(&cg->cg_mask, &mask))
780 break;
781 }
782 return (NULL);
783 }
784 #else /* !SMP */
785
786 void
787 smp_rendezvous_cpus(cpuset_t map,
788 void (*setup_func)(void *),
789 void (*action_func)(void *),
790 void (*teardown_func)(void *),
791 void *arg)
792 {
793 /*
794 * In the !SMP case we just need to ensure the same initial conditions
795 * as the SMP case.
796 */
797 spinlock_enter();
798 if (setup_func != NULL)
799 setup_func(arg);
800 if (action_func != NULL)
801 action_func(arg);
802 if (teardown_func != NULL)
803 teardown_func(arg);
804 spinlock_exit();
805 }
806
807 void
808 smp_rendezvous(void (*setup_func)(void *),
809 void (*action_func)(void *),
810 void (*teardown_func)(void *),
811 void *arg)
812 {
813
814 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func,
815 arg);
816 }
817
818 /*
819 * Provide dummy SMP support for UP kernels. Modules that need to use SMP
820 * APIs will still work using this dummy support.
821 */
822 static void
823 mp_setvariables_for_up(void *dummy)
824 {
825 mp_ncpus = 1;
826 mp_maxid = PCPU_GET(cpuid);
827 CPU_SETOF(mp_maxid, &all_cpus);
828 KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero"));
829 }
830 SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST,
831 mp_setvariables_for_up, NULL);
832 #endif /* SMP */
833
834 void
835 smp_no_rendezvous_barrier(void *dummy)
836 {
837 #ifdef SMP
838 KASSERT((!smp_started),("smp_no_rendezvous called and smp is started"));
839 #endif
840 }
841
842 /*
843 * Wait for specified idle threads to switch once. This ensures that even
844 * preempted threads have cycled through the switch function once,
845 * exiting their codepaths. This allows us to change global pointers
846 * with no other synchronization.
847 */
848 int
849 quiesce_cpus(cpuset_t map, const char *wmesg, int prio)
850 {
851 struct pcpu *pcpu;
852 u_int gen[MAXCPU];
853 int error;
854 int cpu;
855
856 error = 0;
857 for (cpu = 0; cpu <= mp_maxid; cpu++) {
858 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu))
859 continue;
860 pcpu = pcpu_find(cpu);
861 gen[cpu] = pcpu->pc_idlethread->td_generation;
862 }
863 for (cpu = 0; cpu <= mp_maxid; cpu++) {
864 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu))
865 continue;
866 pcpu = pcpu_find(cpu);
867 thread_lock(curthread);
868 sched_bind(curthread, cpu);
869 thread_unlock(curthread);
870 while (gen[cpu] == pcpu->pc_idlethread->td_generation) {
871 error = tsleep(quiesce_cpus, prio, wmesg, 1);
872 if (error != EWOULDBLOCK)
873 goto out;
874 error = 0;
875 }
876 }
877 out:
878 thread_lock(curthread);
879 sched_unbind(curthread);
880 thread_unlock(curthread);
881
882 return (error);
883 }
884
885 int
886 quiesce_all_cpus(const char *wmesg, int prio)
887 {
888
889 return quiesce_cpus(all_cpus, wmesg, prio);
890 }
891
892 /* Extra care is taken with this sysctl because the data type is volatile */
893 static int
894 sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS)
895 {
896 int error, active;
897
898 active = smp_started;
899 error = SYSCTL_OUT(req, &active, sizeof(active));
900 return (error);
901 }
902
903
904 #ifdef SMP
905 void
906 topo_init_node(struct topo_node *node)
907 {
908
909 bzero(node, sizeof(*node));
910 TAILQ_INIT(&node->children);
911 }
912
913 void
914 topo_init_root(struct topo_node *root)
915 {
916
917 topo_init_node(root);
918 root->type = TOPO_TYPE_SYSTEM;
919 }
920
921 /*
922 * Add a child node with the given ID under the given parent.
923 * Do nothing if there is already a child with that ID.
924 */
925 struct topo_node *
926 topo_add_node_by_hwid(struct topo_node *parent, int hwid,
927 topo_node_type type, uintptr_t subtype)
928 {
929 struct topo_node *node;
930
931 TAILQ_FOREACH_REVERSE(node, &parent->children,
932 topo_children, siblings) {
933 if (node->hwid == hwid
934 && node->type == type && node->subtype == subtype) {
935 return (node);
936 }
937 }
938
939 node = malloc(sizeof(*node), M_TOPO, M_WAITOK);
940 topo_init_node(node);
941 node->parent = parent;
942 node->hwid = hwid;
943 node->type = type;
944 node->subtype = subtype;
945 TAILQ_INSERT_TAIL(&parent->children, node, siblings);
946 parent->nchildren++;
947
948 return (node);
949 }
950
951 /*
952 * Find a child node with the given ID under the given parent.
953 */
954 struct topo_node *
955 topo_find_node_by_hwid(struct topo_node *parent, int hwid,
956 topo_node_type type, uintptr_t subtype)
957 {
958
959 struct topo_node *node;
960
961 TAILQ_FOREACH(node, &parent->children, siblings) {
962 if (node->hwid == hwid
963 && node->type == type && node->subtype == subtype) {
964 return (node);
965 }
966 }
967
968 return (NULL);
969 }
970
971 /*
972 * Given a node change the order of its parent's child nodes such
973 * that the node becomes the firt child while preserving the cyclic
974 * order of the children. In other words, the given node is promoted
975 * by rotation.
976 */
977 void
978 topo_promote_child(struct topo_node *child)
979 {
980 struct topo_node *next;
981 struct topo_node *node;
982 struct topo_node *parent;
983
984 parent = child->parent;
985 next = TAILQ_NEXT(child, siblings);
986 TAILQ_REMOVE(&parent->children, child, siblings);
987 TAILQ_INSERT_HEAD(&parent->children, child, siblings);
988
989 while (next != NULL) {
990 node = next;
991 next = TAILQ_NEXT(node, siblings);
992 TAILQ_REMOVE(&parent->children, node, siblings);
993 TAILQ_INSERT_AFTER(&parent->children, child, node, siblings);
994 child = node;
995 }
996 }
997
998 /*
999 * Iterate to the next node in the depth-first search (traversal) of
1000 * the topology tree.
1001 */
1002 struct topo_node *
1003 topo_next_node(struct topo_node *top, struct topo_node *node)
1004 {
1005 struct topo_node *next;
1006
1007 if ((next = TAILQ_FIRST(&node->children)) != NULL)
1008 return (next);
1009
1010 if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1011 return (next);
1012
1013 while (node != top && (node = node->parent) != top)
1014 if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1015 return (next);
1016
1017 return (NULL);
1018 }
1019
1020 /*
1021 * Iterate to the next node in the depth-first search of the topology tree,
1022 * but without descending below the current node.
1023 */
1024 struct topo_node *
1025 topo_next_nonchild_node(struct topo_node *top, struct topo_node *node)
1026 {
1027 struct topo_node *next;
1028
1029 if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1030 return (next);
1031
1032 while (node != top && (node = node->parent) != top)
1033 if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1034 return (next);
1035
1036 return (NULL);
1037 }
1038
1039 /*
1040 * Assign the given ID to the given topology node that represents a logical
1041 * processor.
1042 */
1043 void
1044 topo_set_pu_id(struct topo_node *node, cpuid_t id)
1045 {
1046
1047 KASSERT(node->type == TOPO_TYPE_PU,
1048 ("topo_set_pu_id: wrong node type: %u", node->type));
1049 KASSERT(CPU_EMPTY(&node->cpuset) && node->cpu_count == 0,
1050 ("topo_set_pu_id: cpuset already not empty"));
1051 node->id = id;
1052 CPU_SET(id, &node->cpuset);
1053 node->cpu_count = 1;
1054 node->subtype = 1;
1055
1056 while ((node = node->parent) != NULL) {
1057 KASSERT(!CPU_ISSET(id, &node->cpuset),
1058 ("logical ID %u is already set in node %p", id, node));
1059 CPU_SET(id, &node->cpuset);
1060 node->cpu_count++;
1061 }
1062 }
1063
1064 static struct topology_spec {
1065 topo_node_type type;
1066 bool match_subtype;
1067 uintptr_t subtype;
1068 } topology_level_table[TOPO_LEVEL_COUNT] = {
1069 [TOPO_LEVEL_PKG] = { .type = TOPO_TYPE_PKG, },
1070 [TOPO_LEVEL_GROUP] = { .type = TOPO_TYPE_GROUP, },
1071 [TOPO_LEVEL_CACHEGROUP] = {
1072 .type = TOPO_TYPE_CACHE,
1073 .match_subtype = true,
1074 .subtype = CG_SHARE_L3,
1075 },
1076 [TOPO_LEVEL_CORE] = { .type = TOPO_TYPE_CORE, },
1077 [TOPO_LEVEL_THREAD] = { .type = TOPO_TYPE_PU, },
1078 };
1079
1080 static bool
1081 topo_analyze_table(struct topo_node *root, int all, enum topo_level level,
1082 struct topo_analysis *results)
1083 {
1084 struct topology_spec *spec;
1085 struct topo_node *node;
1086 int count;
1087
1088 if (level >= TOPO_LEVEL_COUNT)
1089 return (true);
1090
1091 spec = &topology_level_table[level];
1092 count = 0;
1093 node = topo_next_node(root, root);
1094
1095 while (node != NULL) {
1096 if (node->type != spec->type ||
1097 (spec->match_subtype && node->subtype != spec->subtype)) {
1098 node = topo_next_node(root, node);
1099 continue;
1100 }
1101 if (!all && CPU_EMPTY(&node->cpuset)) {
1102 node = topo_next_nonchild_node(root, node);
1103 continue;
1104 }
1105
1106 count++;
1107
1108 if (!topo_analyze_table(node, all, level + 1, results))
1109 return (false);
1110
1111 node = topo_next_nonchild_node(root, node);
1112 }
1113
1114 /* No explicit subgroups is essentially one subgroup. */
1115 if (count == 0) {
1116 count = 1;
1117
1118 if (!topo_analyze_table(root, all, level + 1, results))
1119 return (false);
1120 }
1121
1122 if (results->entities[level] == -1)
1123 results->entities[level] = count;
1124 else if (results->entities[level] != count)
1125 return (false);
1126
1127 return (true);
1128 }
1129
1130 /*
1131 * Check if the topology is uniform, that is, each package has the same number
1132 * of cores in it and each core has the same number of threads (logical
1133 * processors) in it. If so, calculate the number of packages, the number of
1134 * groups per package, the number of cachegroups per group, and the number of
1135 * logical processors per cachegroup. 'all' parameter tells whether to include
1136 * administratively disabled logical processors into the analysis.
1137 */
1138 int
1139 topo_analyze(struct topo_node *topo_root, int all,
1140 struct topo_analysis *results)
1141 {
1142
1143 results->entities[TOPO_LEVEL_PKG] = -1;
1144 results->entities[TOPO_LEVEL_CORE] = -1;
1145 results->entities[TOPO_LEVEL_THREAD] = -1;
1146 results->entities[TOPO_LEVEL_GROUP] = -1;
1147 results->entities[TOPO_LEVEL_CACHEGROUP] = -1;
1148
1149 if (!topo_analyze_table(topo_root, all, TOPO_LEVEL_PKG, results))
1150 return (0);
1151
1152 KASSERT(results->entities[TOPO_LEVEL_PKG] > 0,
1153 ("bug in topology or analysis"));
1154
1155 return (1);
1156 }
1157
1158 #endif /* SMP */
1159
Cache object: 0535f8c897513491d7cf039f0a6a74ad
|