2 * Copyright (c) 2001, John Baldwin <jhb@FreeBSD.org>.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
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.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 * This module holds the global variables and machine independent functions
29 * used for the kernel SMP support.
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
42 #include <sys/malloc.h>
43 #include <sys/mutex.h>
45 #include <sys/sched.h>
47 #include <sys/sysctl.h>
49 #include <machine/cpu.h>
50 #include <machine/smp.h>
52 #include "opt_sched.h"
55 MALLOC_DEFINE(M_TOPO, "toponodes", "SMP topology data");
57 volatile cpuset_t stopped_cpus;
58 volatile cpuset_t started_cpus;
59 volatile cpuset_t suspended_cpus;
60 cpuset_t hlt_cpus_mask;
61 cpuset_t logical_cpus_mask;
63 void (*cpustop_restartfunc)(void);
66 static int sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS);
68 /* This is used in modules that need to work in both SMP and UP. */
72 /* export this for libkvm consumers. */
73 int mp_maxcpus = MAXCPU;
75 volatile int smp_started;
78 static SYSCTL_NODE(_kern, OID_AUTO, smp, CTLFLAG_RD|CTLFLAG_CAPRD, NULL,
81 SYSCTL_INT(_kern_smp, OID_AUTO, maxid, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxid, 0,
84 SYSCTL_INT(_kern_smp, OID_AUTO, maxcpus, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxcpus,
85 0, "Max number of CPUs that the system was compiled for.");
87 SYSCTL_PROC(_kern_smp, OID_AUTO, active, CTLFLAG_RD | CTLTYPE_INT, NULL, 0,
88 sysctl_kern_smp_active, "I", "Indicates system is running in SMP mode");
90 int smp_disabled = 0; /* has smp been disabled? */
91 SYSCTL_INT(_kern_smp, OID_AUTO, disabled, CTLFLAG_RDTUN|CTLFLAG_CAPRD,
92 &smp_disabled, 0, "SMP has been disabled from the loader");
94 int smp_cpus = 1; /* how many cpu's running */
95 SYSCTL_INT(_kern_smp, OID_AUTO, cpus, CTLFLAG_RD|CTLFLAG_CAPRD, &smp_cpus, 0,
96 "Number of CPUs online");
98 int smp_topology = 0; /* Which topology we're using. */
99 SYSCTL_INT(_kern_smp, OID_AUTO, topology, CTLFLAG_RDTUN, &smp_topology, 0,
100 "Topology override setting; 0 is default provided by hardware.");
103 /* Enable forwarding of a signal to a process running on a different CPU */
104 static int forward_signal_enabled = 1;
105 SYSCTL_INT(_kern_smp, OID_AUTO, forward_signal_enabled, CTLFLAG_RW,
106 &forward_signal_enabled, 0,
107 "Forwarding of a signal to a process on a different CPU");
109 /* Variables needed for SMP rendezvous. */
110 static volatile int smp_rv_ncpus;
111 static void (*volatile smp_rv_setup_func)(void *arg);
112 static void (*volatile smp_rv_action_func)(void *arg);
113 static void (*volatile smp_rv_teardown_func)(void *arg);
114 static void *volatile smp_rv_func_arg;
115 static volatile int smp_rv_waiters[4];
118 * Shared mutex to restrict busywaits between smp_rendezvous() and
119 * smp(_targeted)_tlb_shootdown(). A deadlock occurs if both of these
120 * functions trigger at once and cause multiple CPUs to busywait with
121 * interrupts disabled.
123 struct mtx smp_ipi_mtx;
126 * Let the MD SMP code initialize mp_maxid very early if it can.
129 mp_setmaxid(void *dummy)
134 KASSERT(mp_ncpus >= 1, ("%s: CPU count < 1", __func__));
135 KASSERT(mp_ncpus > 1 || mp_maxid == 0,
136 ("%s: one CPU but mp_maxid is not zero", __func__));
137 KASSERT(mp_maxid >= mp_ncpus - 1,
138 ("%s: counters out of sync: max %d, count %d", __func__,
139 mp_maxid, mp_ncpus));
141 SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL);
144 * Call the MD SMP initialization code.
147 mp_start(void *dummy)
150 mtx_init(&smp_ipi_mtx, "smp rendezvous", NULL, MTX_SPIN);
152 /* Probe for MP hardware. */
153 if (smp_disabled != 0 || cpu_mp_probe() == 0) {
155 CPU_SETOF(PCPU_GET(cpuid), &all_cpus);
160 printf("FreeBSD/SMP: Multiprocessor System Detected: %d CPUs\n",
164 SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL);
167 forward_signal(struct thread *td)
172 * signotify() has already set TDF_ASTPENDING and TDF_NEEDSIGCHECK on
173 * this thread, so all we need to do is poke it if it is currently
174 * executing so that it executes ast().
176 THREAD_LOCK_ASSERT(td, MA_OWNED);
177 KASSERT(TD_IS_RUNNING(td),
178 ("forward_signal: thread is not TDS_RUNNING"));
180 CTR1(KTR_SMP, "forward_signal(%p)", td->td_proc);
182 if (!smp_started || cold || panicstr)
184 if (!forward_signal_enabled)
187 /* No need to IPI ourself. */
194 ipi_cpu(id, IPI_AST);
198 * When called the executing CPU will send an IPI to all other CPUs
199 * requesting that they halt execution.
201 * Usually (but not necessarily) called with 'other_cpus' as its arg.
203 * - Signals all CPUs in map to stop.
204 * - Waits for each to stop.
212 #if defined(__amd64__) || defined(__i386__)
218 generic_stop_cpus(cpuset_t map, u_int type)
221 char cpusetbuf[CPUSETBUFSIZ];
223 static volatile u_int stopping_cpu = NOCPU;
225 volatile cpuset_t *cpus;
228 type == IPI_STOP || type == IPI_STOP_HARD
230 || type == IPI_SUSPEND
232 , ("%s: invalid stop type", __func__));
237 CTR2(KTR_SMP, "stop_cpus(%s) with %u type",
238 cpusetobj_strprint(cpusetbuf, &map), type);
242 * When suspending, ensure there are are no IPIs in progress.
243 * IPIs that have been issued, but not yet delivered (e.g.
244 * not pending on a vCPU when running under virtualization)
245 * will be lost, violating FreeBSD's assumption of reliable
248 if (type == IPI_SUSPEND)
249 mtx_lock_spin(&smp_ipi_mtx);
253 if (!nmi_is_broadcast || nmi_kdb_lock == 0) {
255 if (stopping_cpu != PCPU_GET(cpuid))
256 while (atomic_cmpset_int(&stopping_cpu, NOCPU,
257 PCPU_GET(cpuid)) == 0)
258 while (stopping_cpu != NOCPU)
259 cpu_spinwait(); /* spin */
261 /* send the stop IPI to all CPUs in map */
262 ipi_selected(map, type);
268 if (type == IPI_SUSPEND)
269 cpus = &suspended_cpus;
272 cpus = &stopped_cpus;
275 while (!CPU_SUBSET(cpus, &map)) {
279 if (i == 100000000) {
280 printf("timeout stopping cpus\n");
286 if (type == IPI_SUSPEND)
287 mtx_unlock_spin(&smp_ipi_mtx);
290 stopping_cpu = NOCPU;
295 stop_cpus(cpuset_t map)
298 return (generic_stop_cpus(map, IPI_STOP));
302 stop_cpus_hard(cpuset_t map)
305 return (generic_stop_cpus(map, IPI_STOP_HARD));
310 suspend_cpus(cpuset_t map)
313 return (generic_stop_cpus(map, IPI_SUSPEND));
318 * Called by a CPU to restart stopped CPUs.
320 * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
322 * - Signals all CPUs in map to restart.
323 * - Waits for each to restart.
331 generic_restart_cpus(cpuset_t map, u_int type)
334 char cpusetbuf[CPUSETBUFSIZ];
336 volatile cpuset_t *cpus;
338 KASSERT(type == IPI_STOP || type == IPI_STOP_HARD
340 || type == IPI_SUSPEND
342 , ("%s: invalid stop type", __func__));
347 CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map));
350 if (type == IPI_SUSPEND)
351 cpus = &resuming_cpus;
354 cpus = &stopped_cpus;
356 /* signal other cpus to restart */
358 if (type == IPI_SUSPEND)
359 CPU_COPY_STORE_REL(&map, &toresume_cpus);
362 CPU_COPY_STORE_REL(&map, &started_cpus);
365 if (!nmi_is_broadcast || nmi_kdb_lock == 0) {
367 /* wait for each to clear its bit */
368 while (CPU_OVERLAP(cpus, &map))
378 restart_cpus(cpuset_t map)
381 return (generic_restart_cpus(map, IPI_STOP));
386 resume_cpus(cpuset_t map)
389 return (generic_restart_cpus(map, IPI_SUSPEND));
395 * All-CPU rendezvous. CPUs are signalled, all execute the setup function
396 * (if specified), rendezvous, execute the action function (if specified),
397 * rendezvous again, execute the teardown function (if specified), and then
400 * Note that the supplied external functions _must_ be reentrant and aware
401 * that they are running in parallel and in an unknown lock context.
404 smp_rendezvous_action(void)
407 void *local_func_arg;
408 void (*local_setup_func)(void*);
409 void (*local_action_func)(void*);
410 void (*local_teardown_func)(void*);
415 /* Ensure we have up-to-date values. */
416 atomic_add_acq_int(&smp_rv_waiters[0], 1);
417 while (smp_rv_waiters[0] < smp_rv_ncpus)
420 /* Fetch rendezvous parameters after acquire barrier. */
421 local_func_arg = smp_rv_func_arg;
422 local_setup_func = smp_rv_setup_func;
423 local_action_func = smp_rv_action_func;
424 local_teardown_func = smp_rv_teardown_func;
427 * Use a nested critical section to prevent any preemptions
428 * from occurring during a rendezvous action routine.
429 * Specifically, if a rendezvous handler is invoked via an IPI
430 * and the interrupted thread was in the critical_exit()
431 * function after setting td_critnest to 0 but before
432 * performing a deferred preemption, this routine can be
433 * invoked with td_critnest set to 0 and td_owepreempt true.
434 * In that case, a critical_exit() during the rendezvous
435 * action would trigger a preemption which is not permitted in
436 * a rendezvous action. To fix this, wrap all of the
437 * rendezvous action handlers in a critical section. We
438 * cannot use a regular critical section however as having
439 * critical_exit() preempt from this routine would also be
440 * problematic (the preemption must not occur before the IPI
441 * has been acknowledged via an EOI). Instead, we
442 * intentionally ignore td_owepreempt when leaving the
443 * critical section. This should be harmless because we do
444 * not permit rendezvous action routines to schedule threads,
445 * and thus td_owepreempt should never transition from 0 to 1
446 * during this routine.
451 owepreempt = td->td_owepreempt;
455 * If requested, run a setup function before the main action
456 * function. Ensure all CPUs have completed the setup
457 * function before moving on to the action function.
459 if (local_setup_func != smp_no_rendezvous_barrier) {
460 if (smp_rv_setup_func != NULL)
461 smp_rv_setup_func(smp_rv_func_arg);
462 atomic_add_int(&smp_rv_waiters[1], 1);
463 while (smp_rv_waiters[1] < smp_rv_ncpus)
467 if (local_action_func != NULL)
468 local_action_func(local_func_arg);
470 if (local_teardown_func != smp_no_rendezvous_barrier) {
472 * Signal that the main action has been completed. If a
473 * full exit rendezvous is requested, then all CPUs will
474 * wait here until all CPUs have finished the main action.
476 atomic_add_int(&smp_rv_waiters[2], 1);
477 while (smp_rv_waiters[2] < smp_rv_ncpus)
480 if (local_teardown_func != NULL)
481 local_teardown_func(local_func_arg);
485 * Signal that the rendezvous is fully completed by this CPU.
486 * This means that no member of smp_rv_* pseudo-structure will be
487 * accessed by this target CPU after this point; in particular,
488 * memory pointed by smp_rv_func_arg.
490 * The release semantic ensures that all accesses performed by
491 * the current CPU are visible when smp_rendezvous_cpus()
492 * returns, by synchronizing with the
493 * atomic_load_acq_int(&smp_rv_waiters[3]).
495 atomic_add_rel_int(&smp_rv_waiters[3], 1);
498 KASSERT(owepreempt == td->td_owepreempt,
499 ("rendezvous action changed td_owepreempt"));
503 smp_rendezvous_cpus(cpuset_t map,
504 void (* setup_func)(void *),
505 void (* action_func)(void *),
506 void (* teardown_func)(void *),
509 int curcpumap, i, ncpus = 0;
511 /* Look comments in the !SMP case. */
514 if (setup_func != NULL)
516 if (action_func != NULL)
518 if (teardown_func != NULL)
525 if (CPU_ISSET(i, &map))
529 panic("ncpus is 0 with non-zero map");
531 mtx_lock_spin(&smp_ipi_mtx);
533 /* Pass rendezvous parameters via global variables. */
534 smp_rv_ncpus = ncpus;
535 smp_rv_setup_func = setup_func;
536 smp_rv_action_func = action_func;
537 smp_rv_teardown_func = teardown_func;
538 smp_rv_func_arg = arg;
539 smp_rv_waiters[1] = 0;
540 smp_rv_waiters[2] = 0;
541 smp_rv_waiters[3] = 0;
542 atomic_store_rel_int(&smp_rv_waiters[0], 0);
545 * Signal other processors, which will enter the IPI with
548 curcpumap = CPU_ISSET(curcpu, &map);
549 CPU_CLR(curcpu, &map);
550 ipi_selected(map, IPI_RENDEZVOUS);
552 /* Check if the current CPU is in the map */
554 smp_rendezvous_action();
557 * Ensure that the master CPU waits for all the other
558 * CPUs to finish the rendezvous, so that smp_rv_*
559 * pseudo-structure and the arg are guaranteed to not
562 * Load acquire synchronizes with the release add in
563 * smp_rendezvous_action(), which ensures that our caller sees
564 * all memory actions done by the called functions on other
567 while (atomic_load_acq_int(&smp_rv_waiters[3]) < ncpus)
570 mtx_unlock_spin(&smp_ipi_mtx);
574 smp_rendezvous(void (* setup_func)(void *),
575 void (* action_func)(void *),
576 void (* teardown_func)(void *),
579 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg);
582 static struct cpu_group group[MAXCPU * MAX_CACHE_LEVELS + 1];
587 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
588 struct cpu_group *top;
591 * Check for a fake topology request for debugging purposes.
593 switch (smp_topology) {
595 /* Dual core with no sharing. */
596 top = smp_topo_1level(CG_SHARE_NONE, 2, 0);
599 /* No topology, all cpus are equal. */
600 top = smp_topo_none();
603 /* Dual core with shared L2. */
604 top = smp_topo_1level(CG_SHARE_L2, 2, 0);
607 /* quad core, shared l3 among each package, private l2. */
608 top = smp_topo_1level(CG_SHARE_L3, 4, 0);
611 /* quad core, 2 dualcore parts on each package share l2. */
612 top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0);
615 /* Single-core 2xHTT */
616 top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT);
619 /* quad core with a shared l3, 8 threads sharing L2. */
620 top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8,
624 /* Default, ask the system what it wants. */
629 * Verify the returned topology.
631 if (top->cg_count != mp_ncpus)
632 panic("Built bad topology at %p. CPU count %d != %d",
633 top, top->cg_count, mp_ncpus);
634 if (CPU_CMP(&top->cg_mask, &all_cpus))
635 panic("Built bad topology at %p. CPU mask (%s) != (%s)",
636 top, cpusetobj_strprint(cpusetbuf, &top->cg_mask),
637 cpusetobj_strprint(cpusetbuf2, &all_cpus));
642 smp_topo_alloc(u_int count)
649 return (&group[curr]);
655 struct cpu_group *top;
658 top->cg_parent = NULL;
659 top->cg_child = NULL;
660 top->cg_mask = all_cpus;
661 top->cg_count = mp_ncpus;
662 top->cg_children = 0;
663 top->cg_level = CG_SHARE_NONE;
670 smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share,
671 int count, int flags, int start)
673 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
678 for (i = 0; i < count; i++, start++)
679 CPU_SET(start, &mask);
680 child->cg_parent = parent;
681 child->cg_child = NULL;
682 child->cg_children = 0;
683 child->cg_level = share;
684 child->cg_count = count;
685 child->cg_flags = flags;
686 child->cg_mask = mask;
687 parent->cg_children++;
688 for (; parent != NULL; parent = parent->cg_parent) {
689 if (CPU_OVERLAP(&parent->cg_mask, &child->cg_mask))
690 panic("Duplicate children in %p. mask (%s) child (%s)",
692 cpusetobj_strprint(cpusetbuf, &parent->cg_mask),
693 cpusetobj_strprint(cpusetbuf2, &child->cg_mask));
694 CPU_OR(&parent->cg_mask, &child->cg_mask);
695 parent->cg_count += child->cg_count;
702 smp_topo_1level(int share, int count, int flags)
704 struct cpu_group *child;
705 struct cpu_group *top;
712 packages = mp_ncpus / count;
713 top->cg_child = child = &group[1];
714 top->cg_level = CG_SHARE_NONE;
715 for (i = 0; i < packages; i++, child++)
716 cpu = smp_topo_addleaf(top, child, share, count, flags, cpu);
721 smp_topo_2level(int l2share, int l2count, int l1share, int l1count,
724 struct cpu_group *top;
725 struct cpu_group *l1g;
726 struct cpu_group *l2g;
735 top->cg_level = CG_SHARE_NONE;
736 top->cg_children = mp_ncpus / (l2count * l1count);
737 l1g = l2g + top->cg_children;
738 for (i = 0; i < top->cg_children; i++, l2g++) {
739 l2g->cg_parent = top;
741 l2g->cg_level = l2share;
742 for (j = 0; j < l2count; j++, l1g++)
743 cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count,
751 smp_topo_find(struct cpu_group *top, int cpu)
753 struct cpu_group *cg;
758 CPU_SETOF(cpu, &mask);
761 if (!CPU_OVERLAP(&cg->cg_mask, &mask))
763 if (cg->cg_children == 0)
765 children = cg->cg_children;
766 for (i = 0, cg = cg->cg_child; i < children; cg++, i++)
767 if (CPU_OVERLAP(&cg->cg_mask, &mask))
775 smp_rendezvous_cpus(cpuset_t map,
776 void (*setup_func)(void *),
777 void (*action_func)(void *),
778 void (*teardown_func)(void *),
782 * In the !SMP case we just need to ensure the same initial conditions
786 if (setup_func != NULL)
788 if (action_func != NULL)
790 if (teardown_func != NULL)
796 smp_rendezvous(void (*setup_func)(void *),
797 void (*action_func)(void *),
798 void (*teardown_func)(void *),
802 /* Look comments in the smp_rendezvous_cpus() case. */
804 if (setup_func != NULL)
806 if (action_func != NULL)
808 if (teardown_func != NULL)
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_rendevous_barrier was renamed to smp_no_rendezvous_barrier
831 * in __FreeBSD_version 1101508, with the old name remaining in 11.x
832 * as an alias for compatibility. The old name will be gone in 12.0
833 * (__FreeBSD_version >= 1200028).
835 __strong_reference(smp_no_rendezvous_barrier, smp_no_rendevous_barrier);
837 smp_no_rendezvous_barrier(void *dummy)
840 KASSERT((!smp_started),("smp_no_rendezvous called and smp is started"));
845 * Wait specified idle threads to switch once. This ensures that even
846 * preempted threads have cycled through the switch function once,
847 * exiting their codepaths. This allows us to change global pointers
848 * with no other synchronization.
851 quiesce_cpus(cpuset_t map, const char *wmesg, int prio)
859 for (cpu = 0; cpu <= mp_maxid; cpu++) {
860 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu))
862 pcpu = pcpu_find(cpu);
863 gen[cpu] = pcpu->pc_idlethread->td_generation;
865 for (cpu = 0; cpu <= mp_maxid; cpu++) {
866 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu))
868 pcpu = pcpu_find(cpu);
869 thread_lock(curthread);
870 sched_bind(curthread, cpu);
871 thread_unlock(curthread);
872 while (gen[cpu] == pcpu->pc_idlethread->td_generation) {
873 error = tsleep(quiesce_cpus, prio, wmesg, 1);
874 if (error != EWOULDBLOCK)
880 thread_lock(curthread);
881 sched_unbind(curthread);
882 thread_unlock(curthread);
888 quiesce_all_cpus(const char *wmesg, int prio)
891 return quiesce_cpus(all_cpus, wmesg, prio);
894 /* Extra care is taken with this sysctl because the data type is volatile */
896 sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS)
900 active = smp_started;
901 error = SYSCTL_OUT(req, &active, sizeof(active));
908 topo_init_node(struct topo_node *node)
911 bzero(node, sizeof(*node));
912 TAILQ_INIT(&node->children);
916 topo_init_root(struct topo_node *root)
919 topo_init_node(root);
920 root->type = TOPO_TYPE_SYSTEM;
924 * Add a child node with the given ID under the given parent.
925 * Do nothing if there is already a child with that ID.
928 topo_add_node_by_hwid(struct topo_node *parent, int hwid,
929 topo_node_type type, uintptr_t subtype)
931 struct topo_node *node;
933 TAILQ_FOREACH_REVERSE(node, &parent->children,
934 topo_children, siblings) {
935 if (node->hwid == hwid
936 && node->type == type && node->subtype == subtype) {
941 node = malloc(sizeof(*node), M_TOPO, M_WAITOK);
942 topo_init_node(node);
943 node->parent = parent;
946 node->subtype = subtype;
947 TAILQ_INSERT_TAIL(&parent->children, node, siblings);
954 * Find a child node with the given ID under the given parent.
957 topo_find_node_by_hwid(struct topo_node *parent, int hwid,
958 topo_node_type type, uintptr_t subtype)
961 struct topo_node *node;
963 TAILQ_FOREACH(node, &parent->children, siblings) {
964 if (node->hwid == hwid
965 && node->type == type && node->subtype == subtype) {
974 * Given a node change the order of its parent's child nodes such
975 * that the node becomes the firt child while preserving the cyclic
976 * order of the children. In other words, the given node is promoted
980 topo_promote_child(struct topo_node *child)
982 struct topo_node *next;
983 struct topo_node *node;
984 struct topo_node *parent;
986 parent = child->parent;
987 next = TAILQ_NEXT(child, siblings);
988 TAILQ_REMOVE(&parent->children, child, siblings);
989 TAILQ_INSERT_HEAD(&parent->children, child, siblings);
991 while (next != NULL) {
993 next = TAILQ_NEXT(node, siblings);
994 TAILQ_REMOVE(&parent->children, node, siblings);
995 TAILQ_INSERT_AFTER(&parent->children, child, node, siblings);
1001 * Iterate to the next node in the depth-first search (traversal) of
1002 * the topology tree.
1005 topo_next_node(struct topo_node *top, struct topo_node *node)
1007 struct topo_node *next;
1009 if ((next = TAILQ_FIRST(&node->children)) != NULL)
1012 if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1015 while ((node = node->parent) != top)
1016 if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1023 * Iterate to the next node in the depth-first search of the topology tree,
1024 * but without descending below the current node.
1027 topo_next_nonchild_node(struct topo_node *top, struct topo_node *node)
1029 struct topo_node *next;
1031 if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1034 while ((node = node->parent) != top)
1035 if ((next = TAILQ_NEXT(node, siblings)) != NULL)
1042 * Assign the given ID to the given topology node that represents a logical
1046 topo_set_pu_id(struct topo_node *node, cpuid_t id)
1049 KASSERT(node->type == TOPO_TYPE_PU,
1050 ("topo_set_pu_id: wrong node type: %u", node->type));
1051 KASSERT(CPU_EMPTY(&node->cpuset) && node->cpu_count == 0,
1052 ("topo_set_pu_id: cpuset already not empty"));
1054 CPU_SET(id, &node->cpuset);
1055 node->cpu_count = 1;
1058 while ((node = node->parent) != NULL) {
1059 KASSERT(!CPU_ISSET(id, &node->cpuset),
1060 ("logical ID %u is already set in node %p", id, node));
1061 CPU_SET(id, &node->cpuset);
1067 * Check if the topology is uniform, that is, each package has the same number
1068 * of cores in it and each core has the same number of threads (logical
1069 * processors) in it. If so, calculate the number of package, the number of
1070 * cores per package and the number of logical processors per core.
1071 * 'all' parameter tells whether to include administratively disabled logical
1072 * processors into the analysis.
1075 topo_analyze(struct topo_node *topo_root, int all,
1076 int *pkg_count, int *cores_per_pkg, int *thrs_per_core)
1078 struct topo_node *pkg_node;
1079 struct topo_node *core_node;
1080 struct topo_node *pu_node;
1087 *cores_per_pkg = -1;
1088 *thrs_per_core = -1;
1090 pkg_node = topo_root;
1091 while (pkg_node != NULL) {
1092 if (pkg_node->type != TOPO_TYPE_PKG) {
1093 pkg_node = topo_next_node(topo_root, pkg_node);
1096 if (!all && CPU_EMPTY(&pkg_node->cpuset)) {
1097 pkg_node = topo_next_nonchild_node(topo_root, pkg_node);
1105 core_node = pkg_node;
1106 while (core_node != NULL) {
1107 if (core_node->type == TOPO_TYPE_CORE) {
1108 if (!all && CPU_EMPTY(&core_node->cpuset)) {
1110 topo_next_nonchild_node(pkg_node,
1118 pu_node = core_node;
1119 while (pu_node != NULL) {
1120 if (pu_node->type == TOPO_TYPE_PU &&
1121 (all || !CPU_EMPTY(&pu_node->cpuset)))
1123 pu_node = topo_next_node(core_node,
1127 if (*thrs_per_core == -1)
1128 *thrs_per_core = tpc_counter;
1129 else if (*thrs_per_core != tpc_counter)
1132 core_node = topo_next_nonchild_node(pkg_node,
1135 /* PU node directly under PKG. */
1136 if (core_node->type == TOPO_TYPE_PU &&
1137 (all || !CPU_EMPTY(&core_node->cpuset)))
1139 core_node = topo_next_node(pkg_node,
1144 if (*cores_per_pkg == -1)
1145 *cores_per_pkg = cpp_counter;
1146 else if (*cores_per_pkg != cpp_counter)
1148 if (thrs_per_pkg == -1)
1149 thrs_per_pkg = tpp_counter;
1150 else if (thrs_per_pkg != tpp_counter)
1153 pkg_node = topo_next_nonchild_node(topo_root, pkg_node);
1156 KASSERT(*pkg_count > 0,
1157 ("bug in topology or analysis"));
1158 if (*cores_per_pkg == 0) {
1159 KASSERT(*thrs_per_core == -1 && thrs_per_pkg > 0,
1160 ("bug in topology or analysis"));
1161 *thrs_per_core = thrs_per_pkg;