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/mutex.h>
44 #include <sys/sched.h>
46 #include <sys/sysctl.h>
48 #include <machine/cpu.h>
49 #include <machine/smp.h>
51 #include "opt_sched.h"
54 volatile cpuset_t stopped_cpus;
55 volatile cpuset_t started_cpus;
56 volatile cpuset_t suspended_cpus;
57 cpuset_t hlt_cpus_mask;
58 cpuset_t logical_cpus_mask;
60 void (*cpustop_restartfunc)(void);
63 static int sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS);
65 /* This is used in modules that need to work in both SMP and UP. */
69 /* export this for libkvm consumers. */
70 int mp_maxcpus = MAXCPU;
72 volatile int smp_started;
75 static SYSCTL_NODE(_kern, OID_AUTO, smp, CTLFLAG_RD|CTLFLAG_CAPRD, NULL,
78 SYSCTL_INT(_kern_smp, OID_AUTO, maxid, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxid, 0,
81 SYSCTL_INT(_kern_smp, OID_AUTO, maxcpus, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxcpus,
82 0, "Max number of CPUs that the system was compiled for.");
84 SYSCTL_PROC(_kern_smp, OID_AUTO, active, CTLFLAG_RD | CTLTYPE_INT, NULL, 0,
85 sysctl_kern_smp_active, "I", "Indicates system is running in SMP mode");
87 int smp_disabled = 0; /* has smp been disabled? */
88 SYSCTL_INT(_kern_smp, OID_AUTO, disabled, CTLFLAG_RDTUN|CTLFLAG_CAPRD,
89 &smp_disabled, 0, "SMP has been disabled from the loader");
91 int smp_cpus = 1; /* how many cpu's running */
92 SYSCTL_INT(_kern_smp, OID_AUTO, cpus, CTLFLAG_RD|CTLFLAG_CAPRD, &smp_cpus, 0,
93 "Number of CPUs online");
95 int smp_topology = 0; /* Which topology we're using. */
96 SYSCTL_INT(_kern_smp, OID_AUTO, topology, CTLFLAG_RDTUN, &smp_topology, 0,
97 "Topology override setting; 0 is default provided by hardware.");
100 /* Enable forwarding of a signal to a process running on a different CPU */
101 static int forward_signal_enabled = 1;
102 SYSCTL_INT(_kern_smp, OID_AUTO, forward_signal_enabled, CTLFLAG_RW,
103 &forward_signal_enabled, 0,
104 "Forwarding of a signal to a process on a different CPU");
106 /* Variables needed for SMP rendezvous. */
107 static volatile int smp_rv_ncpus;
108 static void (*volatile smp_rv_setup_func)(void *arg);
109 static void (*volatile smp_rv_action_func)(void *arg);
110 static void (*volatile smp_rv_teardown_func)(void *arg);
111 static void *volatile smp_rv_func_arg;
112 static volatile int smp_rv_waiters[4];
115 * Shared mutex to restrict busywaits between smp_rendezvous() and
116 * smp(_targeted)_tlb_shootdown(). A deadlock occurs if both of these
117 * functions trigger at once and cause multiple CPUs to busywait with
118 * interrupts disabled.
120 struct mtx smp_ipi_mtx;
123 * Let the MD SMP code initialize mp_maxid very early if it can.
126 mp_setmaxid(void *dummy)
130 SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL);
133 * Call the MD SMP initialization code.
136 mp_start(void *dummy)
139 mtx_init(&smp_ipi_mtx, "smp rendezvous", NULL, MTX_SPIN);
141 /* Probe for MP hardware. */
142 if (smp_disabled != 0 || cpu_mp_probe() == 0) {
144 CPU_SETOF(PCPU_GET(cpuid), &all_cpus);
149 printf("FreeBSD/SMP: Multiprocessor System Detected: %d CPUs\n",
153 SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL);
156 forward_signal(struct thread *td)
161 * signotify() has already set TDF_ASTPENDING and TDF_NEEDSIGCHECK on
162 * this thread, so all we need to do is poke it if it is currently
163 * executing so that it executes ast().
165 THREAD_LOCK_ASSERT(td, MA_OWNED);
166 KASSERT(TD_IS_RUNNING(td),
167 ("forward_signal: thread is not TDS_RUNNING"));
169 CTR1(KTR_SMP, "forward_signal(%p)", td->td_proc);
171 if (!smp_started || cold || panicstr)
173 if (!forward_signal_enabled)
176 /* No need to IPI ourself. */
183 ipi_cpu(id, IPI_AST);
187 * When called the executing CPU will send an IPI to all other CPUs
188 * requesting that they halt execution.
190 * Usually (but not necessarily) called with 'other_cpus' as its arg.
192 * - Signals all CPUs in map to stop.
193 * - Waits for each to stop.
202 generic_stop_cpus(cpuset_t map, u_int type)
205 char cpusetbuf[CPUSETBUFSIZ];
207 static volatile u_int stopping_cpu = NOCPU;
209 volatile cpuset_t *cpus;
212 #if defined(__amd64__) || defined(__i386__)
213 type == IPI_STOP || type == IPI_STOP_HARD || type == IPI_SUSPEND,
215 type == IPI_STOP || type == IPI_STOP_HARD,
217 ("%s: invalid stop type", __func__));
222 CTR2(KTR_SMP, "stop_cpus(%s) with %u type",
223 cpusetobj_strprint(cpusetbuf, &map), type);
225 #if defined(__amd64__) || defined(__i386__)
227 * When suspending, ensure there are are no IPIs in progress.
228 * IPIs that have been issued, but not yet delivered (e.g.
229 * not pending on a vCPU when running under virtualization)
230 * will be lost, violating FreeBSD's assumption of reliable
233 if (type == IPI_SUSPEND)
234 mtx_lock_spin(&smp_ipi_mtx);
237 if (stopping_cpu != PCPU_GET(cpuid))
238 while (atomic_cmpset_int(&stopping_cpu, NOCPU,
239 PCPU_GET(cpuid)) == 0)
240 while (stopping_cpu != NOCPU)
241 cpu_spinwait(); /* spin */
243 /* send the stop IPI to all CPUs in map */
244 ipi_selected(map, type);
246 #if defined(__amd64__) || defined(__i386__)
247 if (type == IPI_SUSPEND)
248 cpus = &suspended_cpus;
251 cpus = &stopped_cpus;
254 while (!CPU_SUBSET(cpus, &map)) {
258 if (i == 100000000) {
259 printf("timeout stopping cpus\n");
264 #if defined(__amd64__) || defined(__i386__)
265 if (type == IPI_SUSPEND)
266 mtx_unlock_spin(&smp_ipi_mtx);
269 stopping_cpu = NOCPU;
274 stop_cpus(cpuset_t map)
277 return (generic_stop_cpus(map, IPI_STOP));
281 stop_cpus_hard(cpuset_t map)
284 return (generic_stop_cpus(map, IPI_STOP_HARD));
287 #if defined(__amd64__) || defined(__i386__)
289 suspend_cpus(cpuset_t map)
292 return (generic_stop_cpus(map, IPI_SUSPEND));
297 * Called by a CPU to restart stopped CPUs.
299 * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
301 * - Signals all CPUs in map to restart.
302 * - Waits for each to restart.
310 generic_restart_cpus(cpuset_t map, u_int type)
313 char cpusetbuf[CPUSETBUFSIZ];
315 volatile cpuset_t *cpus;
318 #if defined(__amd64__) || defined(__i386__)
319 type == IPI_STOP || type == IPI_STOP_HARD || type == IPI_SUSPEND,
321 type == IPI_STOP || type == IPI_STOP_HARD,
323 ("%s: invalid stop type", __func__));
328 CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map));
330 #if defined(__amd64__) || defined(__i386__)
331 if (type == IPI_SUSPEND)
332 cpus = &suspended_cpus;
335 cpus = &stopped_cpus;
337 /* signal other cpus to restart */
338 CPU_COPY_STORE_REL(&map, &started_cpus);
340 /* wait for each to clear its bit */
341 while (CPU_OVERLAP(cpus, &map))
348 restart_cpus(cpuset_t map)
351 return (generic_restart_cpus(map, IPI_STOP));
354 #if defined(__amd64__) || defined(__i386__)
356 resume_cpus(cpuset_t map)
359 return (generic_restart_cpus(map, IPI_SUSPEND));
364 * All-CPU rendezvous. CPUs are signalled, all execute the setup function
365 * (if specified), rendezvous, execute the action function (if specified),
366 * rendezvous again, execute the teardown function (if specified), and then
369 * Note that the supplied external functions _must_ be reentrant and aware
370 * that they are running in parallel and in an unknown lock context.
373 smp_rendezvous_action(void)
376 void *local_func_arg;
377 void (*local_setup_func)(void*);
378 void (*local_action_func)(void*);
379 void (*local_teardown_func)(void*);
384 /* Ensure we have up-to-date values. */
385 atomic_add_acq_int(&smp_rv_waiters[0], 1);
386 while (smp_rv_waiters[0] < smp_rv_ncpus)
389 /* Fetch rendezvous parameters after acquire barrier. */
390 local_func_arg = smp_rv_func_arg;
391 local_setup_func = smp_rv_setup_func;
392 local_action_func = smp_rv_action_func;
393 local_teardown_func = smp_rv_teardown_func;
396 * Use a nested critical section to prevent any preemptions
397 * from occurring during a rendezvous action routine.
398 * Specifically, if a rendezvous handler is invoked via an IPI
399 * and the interrupted thread was in the critical_exit()
400 * function after setting td_critnest to 0 but before
401 * performing a deferred preemption, this routine can be
402 * invoked with td_critnest set to 0 and td_owepreempt true.
403 * In that case, a critical_exit() during the rendezvous
404 * action would trigger a preemption which is not permitted in
405 * a rendezvous action. To fix this, wrap all of the
406 * rendezvous action handlers in a critical section. We
407 * cannot use a regular critical section however as having
408 * critical_exit() preempt from this routine would also be
409 * problematic (the preemption must not occur before the IPI
410 * has been acknowledged via an EOI). Instead, we
411 * intentionally ignore td_owepreempt when leaving the
412 * critical section. This should be harmless because we do
413 * not permit rendezvous action routines to schedule threads,
414 * and thus td_owepreempt should never transition from 0 to 1
415 * during this routine.
420 owepreempt = td->td_owepreempt;
424 * If requested, run a setup function before the main action
425 * function. Ensure all CPUs have completed the setup
426 * function before moving on to the action function.
428 if (local_setup_func != smp_no_rendevous_barrier) {
429 if (smp_rv_setup_func != NULL)
430 smp_rv_setup_func(smp_rv_func_arg);
431 atomic_add_int(&smp_rv_waiters[1], 1);
432 while (smp_rv_waiters[1] < smp_rv_ncpus)
436 if (local_action_func != NULL)
437 local_action_func(local_func_arg);
439 if (local_teardown_func != smp_no_rendevous_barrier) {
441 * Signal that the main action has been completed. If a
442 * full exit rendezvous is requested, then all CPUs will
443 * wait here until all CPUs have finished the main action.
445 atomic_add_int(&smp_rv_waiters[2], 1);
446 while (smp_rv_waiters[2] < smp_rv_ncpus)
449 if (local_teardown_func != NULL)
450 local_teardown_func(local_func_arg);
454 * Signal that the rendezvous is fully completed by this CPU.
455 * This means that no member of smp_rv_* pseudo-structure will be
456 * accessed by this target CPU after this point; in particular,
457 * memory pointed by smp_rv_func_arg.
459 atomic_add_int(&smp_rv_waiters[3], 1);
462 KASSERT(owepreempt == td->td_owepreempt,
463 ("rendezvous action changed td_owepreempt"));
467 smp_rendezvous_cpus(cpuset_t map,
468 void (* setup_func)(void *),
469 void (* action_func)(void *),
470 void (* teardown_func)(void *),
473 int curcpumap, i, ncpus = 0;
475 /* Look comments in the !SMP case. */
478 if (setup_func != NULL)
480 if (action_func != NULL)
482 if (teardown_func != NULL)
489 if (CPU_ISSET(i, &map))
493 panic("ncpus is 0 with non-zero map");
495 mtx_lock_spin(&smp_ipi_mtx);
497 /* Pass rendezvous parameters via global variables. */
498 smp_rv_ncpus = ncpus;
499 smp_rv_setup_func = setup_func;
500 smp_rv_action_func = action_func;
501 smp_rv_teardown_func = teardown_func;
502 smp_rv_func_arg = arg;
503 smp_rv_waiters[1] = 0;
504 smp_rv_waiters[2] = 0;
505 smp_rv_waiters[3] = 0;
506 atomic_store_rel_int(&smp_rv_waiters[0], 0);
509 * Signal other processors, which will enter the IPI with
512 curcpumap = CPU_ISSET(curcpu, &map);
513 CPU_CLR(curcpu, &map);
514 ipi_selected(map, IPI_RENDEZVOUS);
516 /* Check if the current CPU is in the map */
518 smp_rendezvous_action();
521 * Ensure that the master CPU waits for all the other
522 * CPUs to finish the rendezvous, so that smp_rv_*
523 * pseudo-structure and the arg are guaranteed to not
526 while (atomic_load_acq_int(&smp_rv_waiters[3]) < ncpus)
529 mtx_unlock_spin(&smp_ipi_mtx);
533 smp_rendezvous(void (* setup_func)(void *),
534 void (* action_func)(void *),
535 void (* teardown_func)(void *),
538 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg);
541 static struct cpu_group group[MAXCPU];
546 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
547 struct cpu_group *top;
550 * Check for a fake topology request for debugging purposes.
552 switch (smp_topology) {
554 /* Dual core with no sharing. */
555 top = smp_topo_1level(CG_SHARE_NONE, 2, 0);
558 /* No topology, all cpus are equal. */
559 top = smp_topo_none();
562 /* Dual core with shared L2. */
563 top = smp_topo_1level(CG_SHARE_L2, 2, 0);
566 /* quad core, shared l3 among each package, private l2. */
567 top = smp_topo_1level(CG_SHARE_L3, 4, 0);
570 /* quad core, 2 dualcore parts on each package share l2. */
571 top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0);
574 /* Single-core 2xHTT */
575 top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT);
578 /* quad core with a shared l3, 8 threads sharing L2. */
579 top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8,
583 /* Default, ask the system what it wants. */
588 * Verify the returned topology.
590 if (top->cg_count != mp_ncpus)
591 panic("Built bad topology at %p. CPU count %d != %d",
592 top, top->cg_count, mp_ncpus);
593 if (CPU_CMP(&top->cg_mask, &all_cpus))
594 panic("Built bad topology at %p. CPU mask (%s) != (%s)",
595 top, cpusetobj_strprint(cpusetbuf, &top->cg_mask),
596 cpusetobj_strprint(cpusetbuf2, &all_cpus));
603 struct cpu_group *top;
606 top->cg_parent = NULL;
607 top->cg_child = NULL;
608 top->cg_mask = all_cpus;
609 top->cg_count = mp_ncpus;
610 top->cg_children = 0;
611 top->cg_level = CG_SHARE_NONE;
618 smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share,
619 int count, int flags, int start)
621 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
626 for (i = 0; i < count; i++, start++)
627 CPU_SET(start, &mask);
628 child->cg_parent = parent;
629 child->cg_child = NULL;
630 child->cg_children = 0;
631 child->cg_level = share;
632 child->cg_count = count;
633 child->cg_flags = flags;
634 child->cg_mask = mask;
635 parent->cg_children++;
636 for (; parent != NULL; parent = parent->cg_parent) {
637 if (CPU_OVERLAP(&parent->cg_mask, &child->cg_mask))
638 panic("Duplicate children in %p. mask (%s) child (%s)",
640 cpusetobj_strprint(cpusetbuf, &parent->cg_mask),
641 cpusetobj_strprint(cpusetbuf2, &child->cg_mask));
642 CPU_OR(&parent->cg_mask, &child->cg_mask);
643 parent->cg_count += child->cg_count;
650 smp_topo_1level(int share, int count, int flags)
652 struct cpu_group *child;
653 struct cpu_group *top;
660 packages = mp_ncpus / count;
661 top->cg_child = child = &group[1];
662 top->cg_level = CG_SHARE_NONE;
663 for (i = 0; i < packages; i++, child++)
664 cpu = smp_topo_addleaf(top, child, share, count, flags, cpu);
669 smp_topo_2level(int l2share, int l2count, int l1share, int l1count,
672 struct cpu_group *top;
673 struct cpu_group *l1g;
674 struct cpu_group *l2g;
683 top->cg_level = CG_SHARE_NONE;
684 top->cg_children = mp_ncpus / (l2count * l1count);
685 l1g = l2g + top->cg_children;
686 for (i = 0; i < top->cg_children; i++, l2g++) {
687 l2g->cg_parent = top;
689 l2g->cg_level = l2share;
690 for (j = 0; j < l2count; j++, l1g++)
691 cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count,
699 smp_topo_find(struct cpu_group *top, int cpu)
701 struct cpu_group *cg;
706 CPU_SETOF(cpu, &mask);
709 if (!CPU_OVERLAP(&cg->cg_mask, &mask))
711 if (cg->cg_children == 0)
713 children = cg->cg_children;
714 for (i = 0, cg = cg->cg_child; i < children; cg++, i++)
715 if (CPU_OVERLAP(&cg->cg_mask, &mask))
723 smp_rendezvous_cpus(cpuset_t map,
724 void (*setup_func)(void *),
725 void (*action_func)(void *),
726 void (*teardown_func)(void *),
730 * In the !SMP case we just need to ensure the same initial conditions
734 if (setup_func != NULL)
736 if (action_func != NULL)
738 if (teardown_func != NULL)
744 smp_rendezvous(void (*setup_func)(void *),
745 void (*action_func)(void *),
746 void (*teardown_func)(void *),
750 /* Look comments in the smp_rendezvous_cpus() case. */
752 if (setup_func != NULL)
754 if (action_func != NULL)
756 if (teardown_func != NULL)
762 * Provide dummy SMP support for UP kernels. Modules that need to use SMP
763 * APIs will still work using this dummy support.
766 mp_setvariables_for_up(void *dummy)
769 mp_maxid = PCPU_GET(cpuid);
770 CPU_SETOF(mp_maxid, &all_cpus);
771 KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero"));
773 SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST,
774 mp_setvariables_for_up, NULL);
778 smp_no_rendevous_barrier(void *dummy)
781 KASSERT((!smp_started),("smp_no_rendevous called and smp is started"));
786 * Wait specified idle threads to switch once. This ensures that even
787 * preempted threads have cycled through the switch function once,
788 * exiting their codepaths. This allows us to change global pointers
789 * with no other synchronization.
792 quiesce_cpus(cpuset_t map, const char *wmesg, int prio)
800 for (cpu = 0; cpu <= mp_maxid; cpu++) {
801 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu))
803 pcpu = pcpu_find(cpu);
804 gen[cpu] = pcpu->pc_idlethread->td_generation;
806 for (cpu = 0; cpu <= mp_maxid; cpu++) {
807 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu))
809 pcpu = pcpu_find(cpu);
810 thread_lock(curthread);
811 sched_bind(curthread, cpu);
812 thread_unlock(curthread);
813 while (gen[cpu] == pcpu->pc_idlethread->td_generation) {
814 error = tsleep(quiesce_cpus, prio, wmesg, 1);
815 if (error != EWOULDBLOCK)
821 thread_lock(curthread);
822 sched_unbind(curthread);
823 thread_unlock(curthread);
829 quiesce_all_cpus(const char *wmesg, int prio)
832 return quiesce_cpus(all_cpus, wmesg, prio);
835 /* Extra care is taken with this sysctl because the data type is volatile */
837 sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS)
841 active = smp_started;
842 error = SYSCTL_OUT(req, &active, sizeof(active));