2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2001, John Baldwin <jhb@FreeBSD.org>.
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
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.
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
29 * This module holds the global variables and machine independent functions
30 * used for the kernel SMP support.
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/kernel.h>
43 #include <sys/malloc.h>
44 #include <sys/mutex.h>
46 #include <sys/sched.h>
48 #include <sys/sysctl.h>
50 #include <machine/cpu.h>
51 #include <machine/smp.h>
53 #include "opt_sched.h"
56 MALLOC_DEFINE(M_TOPO, "toponodes", "SMP topology data");
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;
64 void (*cpustop_restartfunc)(void);
67 static int sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS);
69 /* This is used in modules that need to work in both SMP and UP. */
73 /* export this for libkvm consumers. */
74 int mp_maxcpus = MAXCPU;
76 volatile int smp_started;
79 static SYSCTL_NODE(_kern, OID_AUTO, smp,
80 CTLFLAG_RD | CTLFLAG_CAPRD | CTLFLAG_MPSAFE, NULL,
83 SYSCTL_INT(_kern_smp, OID_AUTO, maxid, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxid, 0,
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.");
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");
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");
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");
101 int smp_threads_per_core = 1; /* how many SMT threads are running per core */
102 SYSCTL_INT(_kern_smp, OID_AUTO, threads_per_core, CTLFLAG_RD|CTLFLAG_CAPRD,
103 &smp_threads_per_core, 0, "Number of SMT threads online per core");
105 int mp_ncores = -1; /* how many physical cores running */
106 SYSCTL_INT(_kern_smp, OID_AUTO, cores, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_ncores, 0,
107 "Number of physical cores online");
109 int smp_topology = 0; /* Which topology we're using. */
110 SYSCTL_INT(_kern_smp, OID_AUTO, topology, CTLFLAG_RDTUN, &smp_topology, 0,
111 "Topology override setting; 0 is default provided by hardware.");
114 /* Enable forwarding of a signal to a process running on a different CPU */
115 static int forward_signal_enabled = 1;
116 SYSCTL_INT(_kern_smp, OID_AUTO, forward_signal_enabled, CTLFLAG_RW,
117 &forward_signal_enabled, 0,
118 "Forwarding of a signal to a process on a different CPU");
120 /* Variables needed for SMP rendezvous. */
121 static volatile int smp_rv_ncpus;
122 static void (*volatile smp_rv_setup_func)(void *arg);
123 static void (*volatile smp_rv_action_func)(void *arg);
124 static void (*volatile smp_rv_teardown_func)(void *arg);
125 static void *volatile smp_rv_func_arg;
126 static volatile int smp_rv_waiters[4];
129 * Shared mutex to restrict busywaits between smp_rendezvous() and
130 * smp(_targeted)_tlb_shootdown(). A deadlock occurs if both of these
131 * functions trigger at once and cause multiple CPUs to busywait with
132 * interrupts disabled.
134 struct mtx smp_ipi_mtx;
137 * Let the MD SMP code initialize mp_maxid very early if it can.
140 mp_setmaxid(void *dummy)
145 KASSERT(mp_ncpus >= 1, ("%s: CPU count < 1", __func__));
146 KASSERT(mp_ncpus > 1 || mp_maxid == 0,
147 ("%s: one CPU but mp_maxid is not zero", __func__));
148 KASSERT(mp_maxid >= mp_ncpus - 1,
149 ("%s: counters out of sync: max %d, count %d", __func__,
150 mp_maxid, mp_ncpus));
152 SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL);
155 * Call the MD SMP initialization code.
158 mp_start(void *dummy)
161 mtx_init(&smp_ipi_mtx, "smp rendezvous", NULL, MTX_SPIN);
163 /* Probe for MP hardware. */
164 if (smp_disabled != 0 || cpu_mp_probe() == 0) {
167 CPU_SETOF(PCPU_GET(cpuid), &all_cpus);
172 printf("FreeBSD/SMP: Multiprocessor System Detected: %d CPUs\n",
175 /* Provide a default for most architectures that don't have SMT/HTT. */
177 mp_ncores = mp_ncpus;
181 SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL);
184 forward_signal(struct thread *td)
189 * signotify() has already set TDF_ASTPENDING and TDF_NEEDSIGCHECK on
190 * this thread, so all we need to do is poke it if it is currently
191 * executing so that it executes ast().
193 THREAD_LOCK_ASSERT(td, MA_OWNED);
194 KASSERT(TD_IS_RUNNING(td),
195 ("forward_signal: thread is not TDS_RUNNING"));
197 CTR1(KTR_SMP, "forward_signal(%p)", td->td_proc);
199 if (!smp_started || cold || KERNEL_PANICKED())
201 if (!forward_signal_enabled)
204 /* No need to IPI ourself. */
211 ipi_cpu(id, IPI_AST);
215 * When called the executing CPU will send an IPI to all other CPUs
216 * requesting that they halt execution.
218 * Usually (but not necessarily) called with 'other_cpus' as its arg.
220 * - Signals all CPUs in map to stop.
221 * - Waits for each to stop.
229 #if defined(__amd64__) || defined(__i386__)
235 generic_stop_cpus(cpuset_t map, u_int type)
238 char cpusetbuf[CPUSETBUFSIZ];
240 static volatile u_int stopping_cpu = NOCPU;
242 volatile cpuset_t *cpus;
245 type == IPI_STOP || type == IPI_STOP_HARD
247 || type == IPI_SUSPEND
249 , ("%s: invalid stop type", __func__));
254 CTR2(KTR_SMP, "stop_cpus(%s) with %u type",
255 cpusetobj_strprint(cpusetbuf, &map), type);
259 * When suspending, ensure there are are no IPIs in progress.
260 * IPIs that have been issued, but not yet delivered (e.g.
261 * not pending on a vCPU when running under virtualization)
262 * will be lost, violating FreeBSD's assumption of reliable
265 if (type == IPI_SUSPEND)
266 mtx_lock_spin(&smp_ipi_mtx);
270 if (!nmi_is_broadcast || nmi_kdb_lock == 0) {
272 if (stopping_cpu != PCPU_GET(cpuid))
273 while (atomic_cmpset_int(&stopping_cpu, NOCPU,
274 PCPU_GET(cpuid)) == 0)
275 while (stopping_cpu != NOCPU)
276 cpu_spinwait(); /* spin */
278 /* send the stop IPI to all CPUs in map */
279 ipi_selected(map, type);
285 if (type == IPI_SUSPEND)
286 cpus = &suspended_cpus;
289 cpus = &stopped_cpus;
292 while (!CPU_SUBSET(cpus, &map)) {
296 if (i == 100000000) {
297 printf("timeout stopping cpus\n");
303 if (type == IPI_SUSPEND)
304 mtx_unlock_spin(&smp_ipi_mtx);
307 stopping_cpu = NOCPU;
312 stop_cpus(cpuset_t map)
315 return (generic_stop_cpus(map, IPI_STOP));
319 stop_cpus_hard(cpuset_t map)
322 return (generic_stop_cpus(map, IPI_STOP_HARD));
327 suspend_cpus(cpuset_t map)
330 return (generic_stop_cpus(map, IPI_SUSPEND));
335 * Called by a CPU to restart stopped CPUs.
337 * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
339 * - Signals all CPUs in map to restart.
340 * - Waits for each to restart.
348 generic_restart_cpus(cpuset_t map, u_int type)
351 char cpusetbuf[CPUSETBUFSIZ];
353 volatile cpuset_t *cpus;
356 KASSERT(type == IPI_STOP || type == IPI_STOP_HARD
357 || type == IPI_SUSPEND, ("%s: invalid stop type", __func__));
362 CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map));
364 if (type == IPI_SUSPEND)
365 cpus = &resuming_cpus;
367 cpus = &stopped_cpus;
369 /* signal other cpus to restart */
370 if (type == IPI_SUSPEND)
371 CPU_COPY_STORE_REL(&map, &toresume_cpus);
373 CPU_COPY_STORE_REL(&map, &started_cpus);
376 * Wake up any CPUs stopped with MWAIT. From MI code we can't tell if
377 * MONITOR/MWAIT is enabled, but the potentially redundant writes are
378 * relatively inexpensive.
380 if (type == IPI_STOP) {
381 struct monitorbuf *mb;
385 if (!CPU_ISSET(id, &map))
388 mb = &pcpu_find(id)->pc_monitorbuf;
389 atomic_store_int(&mb->stop_state,
390 MONITOR_STOPSTATE_RUNNING);
394 if (!nmi_is_broadcast || nmi_kdb_lock == 0) {
395 /* wait for each to clear its bit */
396 while (CPU_OVERLAP(cpus, &map))
400 KASSERT(type == IPI_STOP || type == IPI_STOP_HARD,
401 ("%s: invalid stop type", __func__));
406 CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map));
408 cpus = &stopped_cpus;
410 /* signal other cpus to restart */
411 CPU_COPY_STORE_REL(&map, &started_cpus);
413 /* wait for each to clear its bit */
414 while (CPU_OVERLAP(cpus, &map))
421 restart_cpus(cpuset_t map)
424 return (generic_restart_cpus(map, IPI_STOP));
429 resume_cpus(cpuset_t map)
432 return (generic_restart_cpus(map, IPI_SUSPEND));
438 * All-CPU rendezvous. CPUs are signalled, all execute the setup function
439 * (if specified), rendezvous, execute the action function (if specified),
440 * rendezvous again, execute the teardown function (if specified), and then
443 * Note that the supplied external functions _must_ be reentrant and aware
444 * that they are running in parallel and in an unknown lock context.
447 smp_rendezvous_action(void)
450 void *local_func_arg;
451 void (*local_setup_func)(void*);
452 void (*local_action_func)(void*);
453 void (*local_teardown_func)(void*);
458 /* Ensure we have up-to-date values. */
459 atomic_add_acq_int(&smp_rv_waiters[0], 1);
460 while (smp_rv_waiters[0] < smp_rv_ncpus)
463 /* Fetch rendezvous parameters after acquire barrier. */
464 local_func_arg = smp_rv_func_arg;
465 local_setup_func = smp_rv_setup_func;
466 local_action_func = smp_rv_action_func;
467 local_teardown_func = smp_rv_teardown_func;
470 * Use a nested critical section to prevent any preemptions
471 * from occurring during a rendezvous action routine.
472 * Specifically, if a rendezvous handler is invoked via an IPI
473 * and the interrupted thread was in the critical_exit()
474 * function after setting td_critnest to 0 but before
475 * performing a deferred preemption, this routine can be
476 * invoked with td_critnest set to 0 and td_owepreempt true.
477 * In that case, a critical_exit() during the rendezvous
478 * action would trigger a preemption which is not permitted in
479 * a rendezvous action. To fix this, wrap all of the
480 * rendezvous action handlers in a critical section. We
481 * cannot use a regular critical section however as having
482 * critical_exit() preempt from this routine would also be
483 * problematic (the preemption must not occur before the IPI
484 * has been acknowledged via an EOI). Instead, we
485 * intentionally ignore td_owepreempt when leaving the
486 * critical section. This should be harmless because we do
487 * not permit rendezvous action routines to schedule threads,
488 * and thus td_owepreempt should never transition from 0 to 1
489 * during this routine.
494 owepreempt = td->td_owepreempt;
498 * If requested, run a setup function before the main action
499 * function. Ensure all CPUs have completed the setup
500 * function before moving on to the action function.
502 if (local_setup_func != smp_no_rendezvous_barrier) {
503 if (local_setup_func != NULL)
504 local_setup_func(local_func_arg);
505 atomic_add_int(&smp_rv_waiters[1], 1);
506 while (smp_rv_waiters[1] < smp_rv_ncpus)
510 if (local_action_func != NULL)
511 local_action_func(local_func_arg);
513 if (local_teardown_func != smp_no_rendezvous_barrier) {
515 * Signal that the main action has been completed. If a
516 * full exit rendezvous is requested, then all CPUs will
517 * wait here until all CPUs have finished the main action.
519 atomic_add_int(&smp_rv_waiters[2], 1);
520 while (smp_rv_waiters[2] < smp_rv_ncpus)
523 if (local_teardown_func != NULL)
524 local_teardown_func(local_func_arg);
528 * Signal that the rendezvous is fully completed by this CPU.
529 * This means that no member of smp_rv_* pseudo-structure will be
530 * accessed by this target CPU after this point; in particular,
531 * memory pointed by smp_rv_func_arg.
533 * The release semantic ensures that all accesses performed by
534 * the current CPU are visible when smp_rendezvous_cpus()
535 * returns, by synchronizing with the
536 * atomic_load_acq_int(&smp_rv_waiters[3]).
538 atomic_add_rel_int(&smp_rv_waiters[3], 1);
541 KASSERT(owepreempt == td->td_owepreempt,
542 ("rendezvous action changed td_owepreempt"));
546 smp_rendezvous_cpus(cpuset_t map,
547 void (* setup_func)(void *),
548 void (* action_func)(void *),
549 void (* teardown_func)(void *),
552 int curcpumap, i, ncpus = 0;
554 /* See comments in the !SMP case. */
557 if (setup_func != NULL)
559 if (action_func != NULL)
561 if (teardown_func != NULL)
568 * Make sure we come here with interrupts enabled. Otherwise we
569 * livelock if smp_ipi_mtx is owned by a thread which sent us an IPI.
571 MPASS(curthread->td_md.md_spinlock_count == 0);
574 if (CPU_ISSET(i, &map))
578 panic("ncpus is 0 with non-zero map");
580 mtx_lock_spin(&smp_ipi_mtx);
582 /* Pass rendezvous parameters via global variables. */
583 smp_rv_ncpus = ncpus;
584 smp_rv_setup_func = setup_func;
585 smp_rv_action_func = action_func;
586 smp_rv_teardown_func = teardown_func;
587 smp_rv_func_arg = arg;
588 smp_rv_waiters[1] = 0;
589 smp_rv_waiters[2] = 0;
590 smp_rv_waiters[3] = 0;
591 atomic_store_rel_int(&smp_rv_waiters[0], 0);
594 * Signal other processors, which will enter the IPI with
597 curcpumap = CPU_ISSET(curcpu, &map);
598 CPU_CLR(curcpu, &map);
599 ipi_selected(map, IPI_RENDEZVOUS);
601 /* Check if the current CPU is in the map */
603 smp_rendezvous_action();
606 * Ensure that the master CPU waits for all the other
607 * CPUs to finish the rendezvous, so that smp_rv_*
608 * pseudo-structure and the arg are guaranteed to not
611 * Load acquire synchronizes with the release add in
612 * smp_rendezvous_action(), which ensures that our caller sees
613 * all memory actions done by the called functions on other
616 while (atomic_load_acq_int(&smp_rv_waiters[3]) < ncpus)
619 mtx_unlock_spin(&smp_ipi_mtx);
623 smp_rendezvous(void (* setup_func)(void *),
624 void (* action_func)(void *),
625 void (* teardown_func)(void *),
628 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg);
631 static struct cpu_group group[MAXCPU * MAX_CACHE_LEVELS + 1];
634 smp_topo_fill(struct cpu_group *cg)
638 for (c = 0; c < cg->cg_children; c++)
639 smp_topo_fill(&cg->cg_child[c]);
640 cg->cg_first = CPU_FFS(&cg->cg_mask) - 1;
641 cg->cg_last = CPU_FLS(&cg->cg_mask) - 1;
647 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
648 struct cpu_group *top;
651 * Check for a fake topology request for debugging purposes.
653 switch (smp_topology) {
655 /* Dual core with no sharing. */
656 top = smp_topo_1level(CG_SHARE_NONE, 2, 0);
659 /* No topology, all cpus are equal. */
660 top = smp_topo_none();
663 /* Dual core with shared L2. */
664 top = smp_topo_1level(CG_SHARE_L2, 2, 0);
667 /* quad core, shared l3 among each package, private l2. */
668 top = smp_topo_1level(CG_SHARE_L3, 4, 0);
671 /* quad core, 2 dualcore parts on each package share l2. */
672 top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0);
675 /* Single-core 2xHTT */
676 top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT);
679 /* quad core with a shared l3, 8 threads sharing L2. */
680 top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8,
684 /* Default, ask the system what it wants. */
689 * Verify the returned topology.
691 if (top->cg_count != mp_ncpus)
692 panic("Built bad topology at %p. CPU count %d != %d",
693 top, top->cg_count, mp_ncpus);
694 if (CPU_CMP(&top->cg_mask, &all_cpus))
695 panic("Built bad topology at %p. CPU mask (%s) != (%s)",
696 top, cpusetobj_strprint(cpusetbuf, &top->cg_mask),
697 cpusetobj_strprint(cpusetbuf2, &all_cpus));
700 * Collapse nonsense levels that may be created out of convenience by
701 * the MD layers. They cause extra work in the search functions.
703 while (top->cg_children == 1) {
704 top = &top->cg_child[0];
705 top->cg_parent = NULL;
712 smp_topo_alloc(u_int count)
719 return (&group[curr]);
725 struct cpu_group *top;
728 top->cg_parent = NULL;
729 top->cg_child = NULL;
730 top->cg_mask = all_cpus;
731 top->cg_count = mp_ncpus;
732 top->cg_children = 0;
733 top->cg_level = CG_SHARE_NONE;
740 smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share,
741 int count, int flags, int start)
743 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
748 for (i = 0; i < count; i++, start++)
749 CPU_SET(start, &mask);
750 child->cg_parent = parent;
751 child->cg_child = NULL;
752 child->cg_children = 0;
753 child->cg_level = share;
754 child->cg_count = count;
755 child->cg_flags = flags;
756 child->cg_mask = mask;
757 parent->cg_children++;
758 for (; parent != NULL; parent = parent->cg_parent) {
759 if (CPU_OVERLAP(&parent->cg_mask, &child->cg_mask))
760 panic("Duplicate children in %p. mask (%s) child (%s)",
762 cpusetobj_strprint(cpusetbuf, &parent->cg_mask),
763 cpusetobj_strprint(cpusetbuf2, &child->cg_mask));
764 CPU_OR(&parent->cg_mask, &parent->cg_mask, &child->cg_mask);
765 parent->cg_count += child->cg_count;
772 smp_topo_1level(int share, int count, int flags)
774 struct cpu_group *child;
775 struct cpu_group *top;
782 packages = mp_ncpus / count;
783 top->cg_child = child = &group[1];
784 top->cg_level = CG_SHARE_NONE;
785 for (i = 0; i < packages; i++, child++)
786 cpu = smp_topo_addleaf(top, child, share, count, flags, cpu);
791 smp_topo_2level(int l2share, int l2count, int l1share, int l1count,
794 struct cpu_group *top;
795 struct cpu_group *l1g;
796 struct cpu_group *l2g;
805 top->cg_level = CG_SHARE_NONE;
806 top->cg_children = mp_ncpus / (l2count * l1count);
807 l1g = l2g + top->cg_children;
808 for (i = 0; i < top->cg_children; i++, l2g++) {
809 l2g->cg_parent = top;
811 l2g->cg_level = l2share;
812 for (j = 0; j < l2count; j++, l1g++)
813 cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count,
820 smp_topo_find(struct cpu_group *top, int cpu)
822 struct cpu_group *cg;
827 CPU_SETOF(cpu, &mask);
830 if (!CPU_OVERLAP(&cg->cg_mask, &mask))
832 if (cg->cg_children == 0)
834 children = cg->cg_children;
835 for (i = 0, cg = cg->cg_child; i < children; cg++, i++)
836 if (CPU_OVERLAP(&cg->cg_mask, &mask))
844 smp_rendezvous_cpus(cpuset_t map,
845 void (*setup_func)(void *),
846 void (*action_func)(void *),
847 void (*teardown_func)(void *),
851 * In the !SMP case we just need to ensure the same initial conditions
855 if (setup_func != NULL)
857 if (action_func != NULL)
859 if (teardown_func != NULL)
865 smp_rendezvous(void (*setup_func)(void *),
866 void (*action_func)(void *),
867 void (*teardown_func)(void *),
871 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func,
876 * Provide dummy SMP support for UP kernels. Modules that need to use SMP
877 * APIs will still work using this dummy support.
880 mp_setvariables_for_up(void *dummy)
884 mp_maxid = PCPU_GET(cpuid);
885 CPU_SETOF(mp_maxid, &all_cpus);
886 KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero"));
888 SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST,
889 mp_setvariables_for_up, NULL);
893 smp_no_rendezvous_barrier(void *dummy)
896 KASSERT((!smp_started),("smp_no_rendezvous called and smp is started"));
901 smp_rendezvous_cpus_retry(cpuset_t map,
902 void (* setup_func)(void *),
903 void (* action_func)(void *),
904 void (* teardown_func)(void *),
905 void (* wait_func)(void *, int),
906 struct smp_rendezvous_cpus_retry_arg *arg)
910 CPU_COPY(&map, &arg->cpus);
913 * Only one CPU to execute on.
917 if (setup_func != NULL)
919 if (action_func != NULL)
921 if (teardown_func != NULL)
928 * Execute an action on all specified CPUs while retrying until they
929 * all acknowledge completion.
939 if (CPU_EMPTY(&arg->cpus))
943 if (!CPU_ISSET(cpu, &arg->cpus))
951 smp_rendezvous_cpus_done(struct smp_rendezvous_cpus_retry_arg *arg)
954 CPU_CLR_ATOMIC(curcpu, &arg->cpus);
958 * If (prio & PDROP) == 0:
959 * Wait for specified idle threads to switch once. This ensures that even
960 * preempted threads have cycled through the switch function once,
961 * exiting their codepaths. This allows us to change global pointers
962 * with no other synchronization.
963 * If (prio & PDROP) != 0:
964 * Force the specified CPUs to switch context at least once.
967 quiesce_cpus(cpuset_t map, const char *wmesg, int prio)
975 if ((prio & PDROP) == 0) {
976 gen = malloc(sizeof(u_int) * MAXCPU, M_TEMP, M_WAITOK);
977 for (cpu = 0; cpu <= mp_maxid; cpu++) {
978 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu))
980 pcpu = pcpu_find(cpu);
981 gen[cpu] = pcpu->pc_idlethread->td_generation;
984 for (cpu = 0; cpu <= mp_maxid; cpu++) {
985 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu))
987 pcpu = pcpu_find(cpu);
988 thread_lock(curthread);
989 sched_bind(curthread, cpu);
990 thread_unlock(curthread);
991 if ((prio & PDROP) != 0)
993 while (gen[cpu] == pcpu->pc_idlethread->td_generation) {
994 error = tsleep(quiesce_cpus, prio & ~PDROP, wmesg, 1);
995 if (error != EWOULDBLOCK)
1001 thread_lock(curthread);
1002 sched_unbind(curthread);
1003 thread_unlock(curthread);
1004 if ((prio & PDROP) == 0)
1011 quiesce_all_cpus(const char *wmesg, int prio)
1014 return quiesce_cpus(all_cpus, wmesg, prio);
1018 * Observe all CPUs not executing in critical section.
1019 * We are not in one so the check for us is safe. If the found
1020 * thread changes to something else we know the section was
1024 quiesce_all_critical(void)
1026 struct thread *td, *newtd;
1030 MPASS(curthread->td_critnest == 0);
1033 pcpu = cpuid_to_pcpu[cpu];
1034 td = pcpu->pc_curthread;
1036 if (td->td_critnest == 0)
1039 newtd = (struct thread *)
1040 atomic_load_acq_ptr((void *)pcpu->pc_curthread);
1048 cpus_fence_seq_cst_issue(void *arg __unused)
1051 atomic_thread_fence_seq_cst();
1055 * Send an IPI forcing a sequentially consistent fence.
1057 * Allows replacement of an explicitly fence with a compiler barrier.
1058 * Trades speed up during normal execution for a significant slowdown when
1059 * the barrier is needed.
1062 cpus_fence_seq_cst(void)
1067 smp_no_rendezvous_barrier,
1068 cpus_fence_seq_cst_issue,
1069 smp_no_rendezvous_barrier,
1073 cpus_fence_seq_cst_issue(NULL);
1077 /* Extra care is taken with this sysctl because the data type is volatile */
1079 sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS)
1083 active = smp_started;
1084 error = SYSCTL_OUT(req, &active, sizeof(active));
1090 topo_init_node(struct topo_node *node)
1093 bzero(node, sizeof(*node));
1094 TAILQ_INIT(&node->children);
1098 topo_init_root(struct topo_node *root)
1101 topo_init_node(root);
1102 root->type = TOPO_TYPE_SYSTEM;
1106 * Add a child node with the given ID under the given parent.
1107 * Do nothing if there is already a child with that ID.
1110 topo_add_node_by_hwid(struct topo_node *parent, int hwid,
1111 topo_node_type type, uintptr_t subtype)
1113 struct topo_node *node;
1115 TAILQ_FOREACH_REVERSE(node, &parent->children,
1116 topo_children, siblings) {
1117 if (node->hwid == hwid
1118 && node->type == type && node->subtype == subtype) {
1123 node = malloc(sizeof(*node), M_TOPO, M_WAITOK);
1124 topo_init_node(node);
1125 node->parent = parent;
1128 node->subtype = subtype;
1129 TAILQ_INSERT_TAIL(&parent->children, node, siblings);
1130 parent->nchildren++;
1136 * Find a child node with the given ID under the given parent.
1139 topo_find_node_by_hwid(struct topo_node *parent, int hwid,
1140 topo_node_type type, uintptr_t subtype)
1143 struct topo_node *node;
1145 TAILQ_FOREACH(node, &parent->children, siblings) {
1146 if (node->hwid == hwid
1147 && node->type == type && node->subtype == subtype) {
1156 * Given a node change the order of its parent's child nodes such
1157 * that the node becomes the firt child while preserving the cyclic
1158 * order of the children. In other words, the given node is promoted
1162 topo_promote_child(struct topo_node *child)
1164 struct topo_node *next;
1165 struct topo_node *node;
1166 struct topo_node *parent;
1168 parent = child->parent;
1169 next = TAILQ_NEXT(child, siblings);
1170 TAILQ_REMOVE(&parent->children, child, siblings);
1171 TAILQ_INSERT_HEAD(&parent->children, child, siblings);
1173 while (next != NULL) {
1175 next = TAILQ_NEXT(node, siblings);
1176 TAILQ_REMOVE(&parent->children, node, siblings);
1177 TAILQ_INSERT_AFTER(&parent->children, child, node, siblings);
1183 * Iterate to the next node in the depth-first search (traversal) of
1184 * the topology tree.
1187 topo_next_node(struct topo_node *top, struct topo_node *node)
1189 struct topo_node *next;
1191 if ((next = TAILQ_FIRST(&node->children)) != NULL)
1194 if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1197 while (node != top && (node = node->parent) != top)
1198 if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1205 * Iterate to the next node in the depth-first search of the topology tree,
1206 * but without descending below the current node.
1209 topo_next_nonchild_node(struct topo_node *top, struct topo_node *node)
1211 struct topo_node *next;
1213 if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1216 while (node != top && (node = node->parent) != top)
1217 if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1224 * Assign the given ID to the given topology node that represents a logical
1228 topo_set_pu_id(struct topo_node *node, cpuid_t id)
1231 KASSERT(node->type == TOPO_TYPE_PU,
1232 ("topo_set_pu_id: wrong node type: %u", node->type));
1233 KASSERT(CPU_EMPTY(&node->cpuset) && node->cpu_count == 0,
1234 ("topo_set_pu_id: cpuset already not empty"));
1236 CPU_SET(id, &node->cpuset);
1237 node->cpu_count = 1;
1240 while ((node = node->parent) != NULL) {
1241 KASSERT(!CPU_ISSET(id, &node->cpuset),
1242 ("logical ID %u is already set in node %p", id, node));
1243 CPU_SET(id, &node->cpuset);
1248 static struct topology_spec {
1249 topo_node_type type;
1252 } topology_level_table[TOPO_LEVEL_COUNT] = {
1253 [TOPO_LEVEL_PKG] = { .type = TOPO_TYPE_PKG, },
1254 [TOPO_LEVEL_GROUP] = { .type = TOPO_TYPE_GROUP, },
1255 [TOPO_LEVEL_CACHEGROUP] = {
1256 .type = TOPO_TYPE_CACHE,
1257 .match_subtype = true,
1258 .subtype = CG_SHARE_L3,
1260 [TOPO_LEVEL_CORE] = { .type = TOPO_TYPE_CORE, },
1261 [TOPO_LEVEL_THREAD] = { .type = TOPO_TYPE_PU, },
1265 topo_analyze_table(struct topo_node *root, int all, enum topo_level level,
1266 struct topo_analysis *results)
1268 struct topology_spec *spec;
1269 struct topo_node *node;
1272 if (level >= TOPO_LEVEL_COUNT)
1275 spec = &topology_level_table[level];
1277 node = topo_next_node(root, root);
1279 while (node != NULL) {
1280 if (node->type != spec->type ||
1281 (spec->match_subtype && node->subtype != spec->subtype)) {
1282 node = topo_next_node(root, node);
1285 if (!all && CPU_EMPTY(&node->cpuset)) {
1286 node = topo_next_nonchild_node(root, node);
1292 if (!topo_analyze_table(node, all, level + 1, results))
1295 node = topo_next_nonchild_node(root, node);
1298 /* No explicit subgroups is essentially one subgroup. */
1302 if (!topo_analyze_table(root, all, level + 1, results))
1306 if (results->entities[level] == -1)
1307 results->entities[level] = count;
1308 else if (results->entities[level] != count)
1315 * Check if the topology is uniform, that is, each package has the same number
1316 * of cores in it and each core has the same number of threads (logical
1317 * processors) in it. If so, calculate the number of packages, the number of
1318 * groups per package, the number of cachegroups per group, and the number of
1319 * logical processors per cachegroup. 'all' parameter tells whether to include
1320 * administratively disabled logical processors into the analysis.
1323 topo_analyze(struct topo_node *topo_root, int all,
1324 struct topo_analysis *results)
1327 results->entities[TOPO_LEVEL_PKG] = -1;
1328 results->entities[TOPO_LEVEL_CORE] = -1;
1329 results->entities[TOPO_LEVEL_THREAD] = -1;
1330 results->entities[TOPO_LEVEL_GROUP] = -1;
1331 results->entities[TOPO_LEVEL_CACHEGROUP] = -1;
1333 if (!topo_analyze_table(topo_root, all, TOPO_LEVEL_PKG, results))
1336 KASSERT(results->entities[TOPO_LEVEL_PKG] > 0,
1337 ("bug in topology or analysis"));