2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2001, John Baldwin <jhb@FreeBSD.org>.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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.
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
30 * This module holds the global variables and machine independent functions
31 * used for the kernel SMP support.
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
44 #include <sys/malloc.h>
45 #include <sys/mutex.h>
47 #include <sys/sched.h>
49 #include <sys/sysctl.h>
51 #include <machine/cpu.h>
52 #include <machine/smp.h>
54 #include "opt_sched.h"
57 MALLOC_DEFINE(M_TOPO, "toponodes", "SMP topology data");
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;
65 void (*cpustop_restartfunc)(void);
68 static int sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS);
70 /* This is used in modules that need to work in both SMP and UP. */
74 /* export this for libkvm consumers. */
75 int mp_maxcpus = MAXCPU;
77 volatile int smp_started;
80 static SYSCTL_NODE(_kern, OID_AUTO, smp, CTLFLAG_RD|CTLFLAG_CAPRD, 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_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.");
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");
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];
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.
126 struct mtx smp_ipi_mtx;
129 * Let the MD SMP code initialize mp_maxid very early if it can.
132 mp_setmaxid(void *dummy)
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));
144 SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL);
147 * Call the MD SMP initialization code.
150 mp_start(void *dummy)
153 mtx_init(&smp_ipi_mtx, "smp rendezvous", NULL, MTX_SPIN);
155 /* Probe for MP hardware. */
156 if (smp_disabled != 0 || cpu_mp_probe() == 0) {
158 CPU_SETOF(PCPU_GET(cpuid), &all_cpus);
163 printf("FreeBSD/SMP: Multiprocessor System Detected: %d CPUs\n",
167 SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL);
170 forward_signal(struct thread *td)
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().
179 THREAD_LOCK_ASSERT(td, MA_OWNED);
180 KASSERT(TD_IS_RUNNING(td),
181 ("forward_signal: thread is not TDS_RUNNING"));
183 CTR1(KTR_SMP, "forward_signal(%p)", td->td_proc);
185 if (!smp_started || cold || panicstr)
187 if (!forward_signal_enabled)
190 /* No need to IPI ourself. */
197 ipi_cpu(id, IPI_AST);
201 * When called the executing CPU will send an IPI to all other CPUs
202 * requesting that they halt execution.
204 * Usually (but not necessarily) called with 'other_cpus' as its arg.
206 * - Signals all CPUs in map to stop.
207 * - Waits for each to stop.
215 #if defined(__amd64__) || defined(__i386__)
221 generic_stop_cpus(cpuset_t map, u_int type)
224 char cpusetbuf[CPUSETBUFSIZ];
226 static volatile u_int stopping_cpu = NOCPU;
228 volatile cpuset_t *cpus;
231 type == IPI_STOP || type == IPI_STOP_HARD
233 || type == IPI_SUSPEND
235 , ("%s: invalid stop type", __func__));
240 CTR2(KTR_SMP, "stop_cpus(%s) with %u type",
241 cpusetobj_strprint(cpusetbuf, &map), type);
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
251 if (type == IPI_SUSPEND)
252 mtx_lock_spin(&smp_ipi_mtx);
256 if (!nmi_is_broadcast || nmi_kdb_lock == 0) {
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 */
264 /* send the stop IPI to all CPUs in map */
265 ipi_selected(map, type);
271 if (type == IPI_SUSPEND)
272 cpus = &suspended_cpus;
275 cpus = &stopped_cpus;
278 while (!CPU_SUBSET(cpus, &map)) {
282 if (i == 100000000) {
283 printf("timeout stopping cpus\n");
289 if (type == IPI_SUSPEND)
290 mtx_unlock_spin(&smp_ipi_mtx);
293 stopping_cpu = NOCPU;
298 stop_cpus(cpuset_t map)
301 return (generic_stop_cpus(map, IPI_STOP));
305 stop_cpus_hard(cpuset_t map)
308 return (generic_stop_cpus(map, IPI_STOP_HARD));
313 suspend_cpus(cpuset_t map)
316 return (generic_stop_cpus(map, IPI_SUSPEND));
321 * Called by a CPU to restart stopped CPUs.
323 * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
325 * - Signals all CPUs in map to restart.
326 * - Waits for each to restart.
334 generic_restart_cpus(cpuset_t map, u_int type)
337 char cpusetbuf[CPUSETBUFSIZ];
339 volatile cpuset_t *cpus;
341 KASSERT(type == IPI_STOP || type == IPI_STOP_HARD
343 || type == IPI_SUSPEND
345 , ("%s: invalid stop type", __func__));
350 CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map));
353 if (type == IPI_SUSPEND)
354 cpus = &suspended_cpus;
357 cpus = &stopped_cpus;
359 /* signal other cpus to restart */
360 CPU_COPY_STORE_REL(&map, &started_cpus);
363 if (!nmi_is_broadcast || nmi_kdb_lock == 0) {
365 /* wait for each to clear its bit */
366 while (CPU_OVERLAP(cpus, &map))
376 restart_cpus(cpuset_t map)
379 return (generic_restart_cpus(map, IPI_STOP));
384 resume_cpus(cpuset_t map)
387 return (generic_restart_cpus(map, IPI_SUSPEND));
393 * All-CPU rendezvous. CPUs are signalled, all execute the setup function
394 * (if specified), rendezvous, execute the action function (if specified),
395 * rendezvous again, execute the teardown function (if specified), and then
398 * Note that the supplied external functions _must_ be reentrant and aware
399 * that they are running in parallel and in an unknown lock context.
402 smp_rendezvous_action(void)
405 void *local_func_arg;
406 void (*local_setup_func)(void*);
407 void (*local_action_func)(void*);
408 void (*local_teardown_func)(void*);
413 /* Ensure we have up-to-date values. */
414 atomic_add_acq_int(&smp_rv_waiters[0], 1);
415 while (smp_rv_waiters[0] < smp_rv_ncpus)
418 /* Fetch rendezvous parameters after acquire barrier. */
419 local_func_arg = smp_rv_func_arg;
420 local_setup_func = smp_rv_setup_func;
421 local_action_func = smp_rv_action_func;
422 local_teardown_func = smp_rv_teardown_func;
425 * Use a nested critical section to prevent any preemptions
426 * from occurring during a rendezvous action routine.
427 * Specifically, if a rendezvous handler is invoked via an IPI
428 * and the interrupted thread was in the critical_exit()
429 * function after setting td_critnest to 0 but before
430 * performing a deferred preemption, this routine can be
431 * invoked with td_critnest set to 0 and td_owepreempt true.
432 * In that case, a critical_exit() during the rendezvous
433 * action would trigger a preemption which is not permitted in
434 * a rendezvous action. To fix this, wrap all of the
435 * rendezvous action handlers in a critical section. We
436 * cannot use a regular critical section however as having
437 * critical_exit() preempt from this routine would also be
438 * problematic (the preemption must not occur before the IPI
439 * has been acknowledged via an EOI). Instead, we
440 * intentionally ignore td_owepreempt when leaving the
441 * critical section. This should be harmless because we do
442 * not permit rendezvous action routines to schedule threads,
443 * and thus td_owepreempt should never transition from 0 to 1
444 * during this routine.
449 owepreempt = td->td_owepreempt;
453 * If requested, run a setup function before the main action
454 * function. Ensure all CPUs have completed the setup
455 * function before moving on to the action function.
457 if (local_setup_func != smp_no_rendezvous_barrier) {
458 if (smp_rv_setup_func != NULL)
459 smp_rv_setup_func(smp_rv_func_arg);
460 atomic_add_int(&smp_rv_waiters[1], 1);
461 while (smp_rv_waiters[1] < smp_rv_ncpus)
465 if (local_action_func != NULL)
466 local_action_func(local_func_arg);
468 if (local_teardown_func != smp_no_rendezvous_barrier) {
470 * Signal that the main action has been completed. If a
471 * full exit rendezvous is requested, then all CPUs will
472 * wait here until all CPUs have finished the main action.
474 atomic_add_int(&smp_rv_waiters[2], 1);
475 while (smp_rv_waiters[2] < smp_rv_ncpus)
478 if (local_teardown_func != NULL)
479 local_teardown_func(local_func_arg);
483 * Signal that the rendezvous is fully completed by this CPU.
484 * This means that no member of smp_rv_* pseudo-structure will be
485 * accessed by this target CPU after this point; in particular,
486 * memory pointed by smp_rv_func_arg.
488 * The release semantic ensures that all accesses performed by
489 * the current CPU are visible when smp_rendezvous_cpus()
490 * returns, by synchronizing with the
491 * atomic_load_acq_int(&smp_rv_waiters[3]).
493 atomic_add_rel_int(&smp_rv_waiters[3], 1);
496 KASSERT(owepreempt == td->td_owepreempt,
497 ("rendezvous action changed td_owepreempt"));
501 smp_rendezvous_cpus(cpuset_t map,
502 void (* setup_func)(void *),
503 void (* action_func)(void *),
504 void (* teardown_func)(void *),
507 int curcpumap, i, ncpus = 0;
509 /* Look comments in the !SMP case. */
512 if (setup_func != NULL)
514 if (action_func != NULL)
516 if (teardown_func != NULL)
523 if (CPU_ISSET(i, &map))
527 panic("ncpus is 0 with non-zero map");
529 mtx_lock_spin(&smp_ipi_mtx);
531 /* Pass rendezvous parameters via global variables. */
532 smp_rv_ncpus = ncpus;
533 smp_rv_setup_func = setup_func;
534 smp_rv_action_func = action_func;
535 smp_rv_teardown_func = teardown_func;
536 smp_rv_func_arg = arg;
537 smp_rv_waiters[1] = 0;
538 smp_rv_waiters[2] = 0;
539 smp_rv_waiters[3] = 0;
540 atomic_store_rel_int(&smp_rv_waiters[0], 0);
543 * Signal other processors, which will enter the IPI with
546 curcpumap = CPU_ISSET(curcpu, &map);
547 CPU_CLR(curcpu, &map);
548 ipi_selected(map, IPI_RENDEZVOUS);
550 /* Check if the current CPU is in the map */
552 smp_rendezvous_action();
555 * Ensure that the master CPU waits for all the other
556 * CPUs to finish the rendezvous, so that smp_rv_*
557 * pseudo-structure and the arg are guaranteed to not
560 * Load acquire synchronizes with the release add in
561 * smp_rendezvous_action(), which ensures that our caller sees
562 * all memory actions done by the called functions on other
565 while (atomic_load_acq_int(&smp_rv_waiters[3]) < ncpus)
568 mtx_unlock_spin(&smp_ipi_mtx);
572 smp_rendezvous(void (* setup_func)(void *),
573 void (* action_func)(void *),
574 void (* teardown_func)(void *),
577 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg);
580 static struct cpu_group group[MAXCPU * MAX_CACHE_LEVELS + 1];
585 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
586 struct cpu_group *top;
589 * Check for a fake topology request for debugging purposes.
591 switch (smp_topology) {
593 /* Dual core with no sharing. */
594 top = smp_topo_1level(CG_SHARE_NONE, 2, 0);
597 /* No topology, all cpus are equal. */
598 top = smp_topo_none();
601 /* Dual core with shared L2. */
602 top = smp_topo_1level(CG_SHARE_L2, 2, 0);
605 /* quad core, shared l3 among each package, private l2. */
606 top = smp_topo_1level(CG_SHARE_L3, 4, 0);
609 /* quad core, 2 dualcore parts on each package share l2. */
610 top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0);
613 /* Single-core 2xHTT */
614 top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT);
617 /* quad core with a shared l3, 8 threads sharing L2. */
618 top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8,
622 /* Default, ask the system what it wants. */
627 * Verify the returned topology.
629 if (top->cg_count != mp_ncpus)
630 panic("Built bad topology at %p. CPU count %d != %d",
631 top, top->cg_count, mp_ncpus);
632 if (CPU_CMP(&top->cg_mask, &all_cpus))
633 panic("Built bad topology at %p. CPU mask (%s) != (%s)",
634 top, cpusetobj_strprint(cpusetbuf, &top->cg_mask),
635 cpusetobj_strprint(cpusetbuf2, &all_cpus));
638 * Collapse nonsense levels that may be created out of convenience by
639 * the MD layers. They cause extra work in the search functions.
641 while (top->cg_children == 1) {
642 top = &top->cg_child[0];
643 top->cg_parent = NULL;
649 smp_topo_alloc(u_int count)
656 return (&group[curr]);
662 struct cpu_group *top;
665 top->cg_parent = NULL;
666 top->cg_child = NULL;
667 top->cg_mask = all_cpus;
668 top->cg_count = mp_ncpus;
669 top->cg_children = 0;
670 top->cg_level = CG_SHARE_NONE;
677 smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share,
678 int count, int flags, int start)
680 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
685 for (i = 0; i < count; i++, start++)
686 CPU_SET(start, &mask);
687 child->cg_parent = parent;
688 child->cg_child = NULL;
689 child->cg_children = 0;
690 child->cg_level = share;
691 child->cg_count = count;
692 child->cg_flags = flags;
693 child->cg_mask = mask;
694 parent->cg_children++;
695 for (; parent != NULL; parent = parent->cg_parent) {
696 if (CPU_OVERLAP(&parent->cg_mask, &child->cg_mask))
697 panic("Duplicate children in %p. mask (%s) child (%s)",
699 cpusetobj_strprint(cpusetbuf, &parent->cg_mask),
700 cpusetobj_strprint(cpusetbuf2, &child->cg_mask));
701 CPU_OR(&parent->cg_mask, &child->cg_mask);
702 parent->cg_count += child->cg_count;
709 smp_topo_1level(int share, int count, int flags)
711 struct cpu_group *child;
712 struct cpu_group *top;
719 packages = mp_ncpus / count;
720 top->cg_child = child = &group[1];
721 top->cg_level = CG_SHARE_NONE;
722 for (i = 0; i < packages; i++, child++)
723 cpu = smp_topo_addleaf(top, child, share, count, flags, cpu);
728 smp_topo_2level(int l2share, int l2count, int l1share, int l1count,
731 struct cpu_group *top;
732 struct cpu_group *l1g;
733 struct cpu_group *l2g;
742 top->cg_level = CG_SHARE_NONE;
743 top->cg_children = mp_ncpus / (l2count * l1count);
744 l1g = l2g + top->cg_children;
745 for (i = 0; i < top->cg_children; i++, l2g++) {
746 l2g->cg_parent = top;
748 l2g->cg_level = l2share;
749 for (j = 0; j < l2count; j++, l1g++)
750 cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count,
758 smp_topo_find(struct cpu_group *top, int cpu)
760 struct cpu_group *cg;
765 CPU_SETOF(cpu, &mask);
768 if (!CPU_OVERLAP(&cg->cg_mask, &mask))
770 if (cg->cg_children == 0)
772 children = cg->cg_children;
773 for (i = 0, cg = cg->cg_child; i < children; cg++, i++)
774 if (CPU_OVERLAP(&cg->cg_mask, &mask))
782 smp_rendezvous_cpus(cpuset_t map,
783 void (*setup_func)(void *),
784 void (*action_func)(void *),
785 void (*teardown_func)(void *),
789 * In the !SMP case we just need to ensure the same initial conditions
793 if (setup_func != NULL)
795 if (action_func != NULL)
797 if (teardown_func != NULL)
803 smp_rendezvous(void (*setup_func)(void *),
804 void (*action_func)(void *),
805 void (*teardown_func)(void *),
809 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func,
814 * Provide dummy SMP support for UP kernels. Modules that need to use SMP
815 * APIs will still work using this dummy support.
818 mp_setvariables_for_up(void *dummy)
821 mp_maxid = PCPU_GET(cpuid);
822 CPU_SETOF(mp_maxid, &all_cpus);
823 KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero"));
825 SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST,
826 mp_setvariables_for_up, NULL);
830 smp_no_rendezvous_barrier(void *dummy)
833 KASSERT((!smp_started),("smp_no_rendezvous called and smp is started"));
838 * Wait for specified idle threads to switch once. This ensures that even
839 * preempted threads have cycled through the switch function once,
840 * exiting their codepaths. This allows us to change global pointers
841 * with no other synchronization.
844 quiesce_cpus(cpuset_t map, const char *wmesg, int prio)
852 for (cpu = 0; cpu <= mp_maxid; cpu++) {
853 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu))
855 pcpu = pcpu_find(cpu);
856 gen[cpu] = pcpu->pc_idlethread->td_generation;
858 for (cpu = 0; cpu <= mp_maxid; cpu++) {
859 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu))
861 pcpu = pcpu_find(cpu);
862 thread_lock(curthread);
863 sched_bind(curthread, cpu);
864 thread_unlock(curthread);
865 while (gen[cpu] == pcpu->pc_idlethread->td_generation) {
866 error = tsleep(quiesce_cpus, prio, wmesg, 1);
867 if (error != EWOULDBLOCK)
873 thread_lock(curthread);
874 sched_unbind(curthread);
875 thread_unlock(curthread);
881 quiesce_all_cpus(const char *wmesg, int prio)
884 return quiesce_cpus(all_cpus, wmesg, prio);
887 /* Extra care is taken with this sysctl because the data type is volatile */
889 sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS)
893 active = smp_started;
894 error = SYSCTL_OUT(req, &active, sizeof(active));
901 topo_init_node(struct topo_node *node)
904 bzero(node, sizeof(*node));
905 TAILQ_INIT(&node->children);
909 topo_init_root(struct topo_node *root)
912 topo_init_node(root);
913 root->type = TOPO_TYPE_SYSTEM;
917 * Add a child node with the given ID under the given parent.
918 * Do nothing if there is already a child with that ID.
921 topo_add_node_by_hwid(struct topo_node *parent, int hwid,
922 topo_node_type type, uintptr_t subtype)
924 struct topo_node *node;
926 TAILQ_FOREACH_REVERSE(node, &parent->children,
927 topo_children, siblings) {
928 if (node->hwid == hwid
929 && node->type == type && node->subtype == subtype) {
934 node = malloc(sizeof(*node), M_TOPO, M_WAITOK);
935 topo_init_node(node);
936 node->parent = parent;
939 node->subtype = subtype;
940 TAILQ_INSERT_TAIL(&parent->children, node, siblings);
947 * Find a child node with the given ID under the given parent.
950 topo_find_node_by_hwid(struct topo_node *parent, int hwid,
951 topo_node_type type, uintptr_t subtype)
954 struct topo_node *node;
956 TAILQ_FOREACH(node, &parent->children, siblings) {
957 if (node->hwid == hwid
958 && node->type == type && node->subtype == subtype) {
967 * Given a node change the order of its parent's child nodes such
968 * that the node becomes the firt child while preserving the cyclic
969 * order of the children. In other words, the given node is promoted
973 topo_promote_child(struct topo_node *child)
975 struct topo_node *next;
976 struct topo_node *node;
977 struct topo_node *parent;
979 parent = child->parent;
980 next = TAILQ_NEXT(child, siblings);
981 TAILQ_REMOVE(&parent->children, child, siblings);
982 TAILQ_INSERT_HEAD(&parent->children, child, siblings);
984 while (next != NULL) {
986 next = TAILQ_NEXT(node, siblings);
987 TAILQ_REMOVE(&parent->children, node, siblings);
988 TAILQ_INSERT_AFTER(&parent->children, child, node, siblings);
994 * Iterate to the next node in the depth-first search (traversal) of
998 topo_next_node(struct topo_node *top, struct topo_node *node)
1000 struct topo_node *next;
1002 if ((next = TAILQ_FIRST(&node->children)) != NULL)
1005 if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1008 while (node != top && (node = node->parent) != top)
1009 if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1016 * Iterate to the next node in the depth-first search of the topology tree,
1017 * but without descending below the current node.
1020 topo_next_nonchild_node(struct topo_node *top, struct topo_node *node)
1022 struct topo_node *next;
1024 if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1027 while (node != top && (node = node->parent) != top)
1028 if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1035 * Assign the given ID to the given topology node that represents a logical
1039 topo_set_pu_id(struct topo_node *node, cpuid_t id)
1042 KASSERT(node->type == TOPO_TYPE_PU,
1043 ("topo_set_pu_id: wrong node type: %u", node->type));
1044 KASSERT(CPU_EMPTY(&node->cpuset) && node->cpu_count == 0,
1045 ("topo_set_pu_id: cpuset already not empty"));
1047 CPU_SET(id, &node->cpuset);
1048 node->cpu_count = 1;
1051 while ((node = node->parent) != NULL) {
1052 KASSERT(!CPU_ISSET(id, &node->cpuset),
1053 ("logical ID %u is already set in node %p", id, node));
1054 CPU_SET(id, &node->cpuset);
1059 static struct topology_spec {
1060 topo_node_type type;
1063 } topology_level_table[TOPO_LEVEL_COUNT] = {
1064 [TOPO_LEVEL_PKG] = { .type = TOPO_TYPE_PKG, },
1065 [TOPO_LEVEL_GROUP] = { .type = TOPO_TYPE_GROUP, },
1066 [TOPO_LEVEL_CACHEGROUP] = {
1067 .type = TOPO_TYPE_CACHE,
1068 .match_subtype = true,
1069 .subtype = CG_SHARE_L3,
1071 [TOPO_LEVEL_CORE] = { .type = TOPO_TYPE_CORE, },
1072 [TOPO_LEVEL_THREAD] = { .type = TOPO_TYPE_PU, },
1076 topo_analyze_table(struct topo_node *root, int all, enum topo_level level,
1077 struct topo_analysis *results)
1079 struct topology_spec *spec;
1080 struct topo_node *node;
1083 if (level >= TOPO_LEVEL_COUNT)
1086 spec = &topology_level_table[level];
1088 node = topo_next_node(root, root);
1090 while (node != NULL) {
1091 if (node->type != spec->type ||
1092 (spec->match_subtype && node->subtype != spec->subtype)) {
1093 node = topo_next_node(root, node);
1096 if (!all && CPU_EMPTY(&node->cpuset)) {
1097 node = topo_next_nonchild_node(root, node);
1103 if (!topo_analyze_table(node, all, level + 1, results))
1106 node = topo_next_nonchild_node(root, node);
1109 /* No explicit subgroups is essentially one subgroup. */
1113 if (!topo_analyze_table(root, all, level + 1, results))
1117 if (results->entities[level] == -1)
1118 results->entities[level] = count;
1119 else if (results->entities[level] != count)
1126 * Check if the topology is uniform, that is, each package has the same number
1127 * of cores in it and each core has the same number of threads (logical
1128 * processors) in it. If so, calculate the number of packages, the number of
1129 * groups per package, the number of cachegroups per group, and the number of
1130 * logical processors per cachegroup. 'all' parameter tells whether to include
1131 * administratively disabled logical processors into the analysis.
1134 topo_analyze(struct topo_node *topo_root, int all,
1135 struct topo_analysis *results)
1138 results->entities[TOPO_LEVEL_PKG] = -1;
1139 results->entities[TOPO_LEVEL_CORE] = -1;
1140 results->entities[TOPO_LEVEL_THREAD] = -1;
1141 results->entities[TOPO_LEVEL_GROUP] = -1;
1142 results->entities[TOPO_LEVEL_CACHEGROUP] = -1;
1144 if (!topo_analyze_table(topo_root, all, TOPO_LEVEL_PKG, results))
1147 KASSERT(results->entities[TOPO_LEVEL_PKG] > 0,
1148 ("bug in topology or analysis"));