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.
13 * 3. Neither the name of the author nor the names of any co-contributors
14 * may be used to endorse or promote products derived from this software
15 * without specific prior written permission.
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * This module holds the global variables and machine independent functions
32 * used for the kernel SMP support.
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
45 #include <sys/mutex.h>
48 #include <sys/sysctl.h>
50 #include <machine/cpu.h>
51 #include <machine/smp.h>
53 #include "opt_sched.h"
56 volatile cpuset_t stopped_cpus;
57 volatile cpuset_t started_cpus;
58 cpuset_t hlt_cpus_mask;
59 cpuset_t logical_cpus_mask;
61 void (*cpustop_restartfunc)(void);
63 /* This is used in modules that need to work in both SMP and UP. */
67 /* export this for libkvm consumers. */
68 int mp_maxcpus = MAXCPU;
70 volatile int smp_started;
73 SYSCTL_NODE(_kern, OID_AUTO, smp, CTLFLAG_RD|CTLFLAG_CAPRD, NULL, "Kernel SMP");
75 SYSCTL_INT(_kern_smp, OID_AUTO, maxid, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxid, 0,
78 SYSCTL_INT(_kern_smp, OID_AUTO, maxcpus, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxcpus,
79 0, "Max number of CPUs that the system was compiled for.");
81 int smp_active = 0; /* are the APs allowed to run? */
82 SYSCTL_INT(_kern_smp, OID_AUTO, active, CTLFLAG_RW, &smp_active, 0,
83 "Number of Auxillary Processors (APs) that were successfully started");
85 int smp_disabled = 0; /* has smp been disabled? */
86 SYSCTL_INT(_kern_smp, OID_AUTO, disabled, CTLFLAG_RDTUN|CTLFLAG_CAPRD,
87 &smp_disabled, 0, "SMP has been disabled from the loader");
88 TUNABLE_INT("kern.smp.disabled", &smp_disabled);
90 int smp_cpus = 1; /* how many cpu's running */
91 SYSCTL_INT(_kern_smp, OID_AUTO, cpus, CTLFLAG_RD|CTLFLAG_CAPRD, &smp_cpus, 0,
92 "Number of CPUs online");
94 int smp_topology = 0; /* Which topology we're using. */
95 SYSCTL_INT(_kern_smp, OID_AUTO, topology, CTLFLAG_RD, &smp_topology, 0,
96 "Topology override setting; 0 is default provided by hardware.");
97 TUNABLE_INT("kern.smp.topology", &smp_topology);
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;
211 #if defined(__amd64__) || defined(__i386__)
212 type == IPI_STOP || type == IPI_STOP_HARD || type == IPI_SUSPEND,
214 type == IPI_STOP || type == IPI_STOP_HARD,
216 ("%s: invalid stop type", __func__));
221 CTR2(KTR_SMP, "stop_cpus(%s) with %u type",
222 cpusetobj_strprint(cpusetbuf, &map), type);
224 if (stopping_cpu != PCPU_GET(cpuid))
225 while (atomic_cmpset_int(&stopping_cpu, NOCPU,
226 PCPU_GET(cpuid)) == 0)
227 while (stopping_cpu != NOCPU)
228 cpu_spinwait(); /* spin */
230 /* send the stop IPI to all CPUs in map */
231 ipi_selected(map, type);
234 while (!CPU_SUBSET(&stopped_cpus, &map)) {
238 if (i == 100000000) {
239 printf("timeout stopping cpus\n");
244 stopping_cpu = NOCPU;
249 stop_cpus(cpuset_t map)
252 return (generic_stop_cpus(map, IPI_STOP));
256 stop_cpus_hard(cpuset_t map)
259 return (generic_stop_cpus(map, IPI_STOP_HARD));
262 #if defined(__amd64__) || defined(__i386__)
264 suspend_cpus(cpuset_t map)
267 return (generic_stop_cpus(map, IPI_SUSPEND));
272 * Called by a CPU to restart stopped CPUs.
274 * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
276 * - Signals all CPUs in map to restart.
277 * - Waits for each to restart.
285 restart_cpus(cpuset_t map)
288 char cpusetbuf[CPUSETBUFSIZ];
294 CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map));
296 /* signal other cpus to restart */
297 CPU_COPY_STORE_REL(&map, &started_cpus);
299 /* wait for each to clear its bit */
300 while (CPU_OVERLAP(&stopped_cpus, &map))
307 * All-CPU rendezvous. CPUs are signalled, all execute the setup function
308 * (if specified), rendezvous, execute the action function (if specified),
309 * rendezvous again, execute the teardown function (if specified), and then
312 * Note that the supplied external functions _must_ be reentrant and aware
313 * that they are running in parallel and in an unknown lock context.
316 smp_rendezvous_action(void)
319 void *local_func_arg;
320 void (*local_setup_func)(void*);
321 void (*local_action_func)(void*);
322 void (*local_teardown_func)(void*);
327 /* Ensure we have up-to-date values. */
328 atomic_add_acq_int(&smp_rv_waiters[0], 1);
329 while (smp_rv_waiters[0] < smp_rv_ncpus)
332 /* Fetch rendezvous parameters after acquire barrier. */
333 local_func_arg = smp_rv_func_arg;
334 local_setup_func = smp_rv_setup_func;
335 local_action_func = smp_rv_action_func;
336 local_teardown_func = smp_rv_teardown_func;
339 * Use a nested critical section to prevent any preemptions
340 * from occurring during a rendezvous action routine.
341 * Specifically, if a rendezvous handler is invoked via an IPI
342 * and the interrupted thread was in the critical_exit()
343 * function after setting td_critnest to 0 but before
344 * performing a deferred preemption, this routine can be
345 * invoked with td_critnest set to 0 and td_owepreempt true.
346 * In that case, a critical_exit() during the rendezvous
347 * action would trigger a preemption which is not permitted in
348 * a rendezvous action. To fix this, wrap all of the
349 * rendezvous action handlers in a critical section. We
350 * cannot use a regular critical section however as having
351 * critical_exit() preempt from this routine would also be
352 * problematic (the preemption must not occur before the IPI
353 * has been acknowledged via an EOI). Instead, we
354 * intentionally ignore td_owepreempt when leaving the
355 * critical section. This should be harmless because we do
356 * not permit rendezvous action routines to schedule threads,
357 * and thus td_owepreempt should never transition from 0 to 1
358 * during this routine.
363 owepreempt = td->td_owepreempt;
367 * If requested, run a setup function before the main action
368 * function. Ensure all CPUs have completed the setup
369 * function before moving on to the action function.
371 if (local_setup_func != smp_no_rendevous_barrier) {
372 if (smp_rv_setup_func != NULL)
373 smp_rv_setup_func(smp_rv_func_arg);
374 atomic_add_int(&smp_rv_waiters[1], 1);
375 while (smp_rv_waiters[1] < smp_rv_ncpus)
379 if (local_action_func != NULL)
380 local_action_func(local_func_arg);
382 if (local_teardown_func != smp_no_rendevous_barrier) {
384 * Signal that the main action has been completed. If a
385 * full exit rendezvous is requested, then all CPUs will
386 * wait here until all CPUs have finished the main action.
388 atomic_add_int(&smp_rv_waiters[2], 1);
389 while (smp_rv_waiters[2] < smp_rv_ncpus)
392 if (local_teardown_func != NULL)
393 local_teardown_func(local_func_arg);
397 * Signal that the rendezvous is fully completed by this CPU.
398 * This means that no member of smp_rv_* pseudo-structure will be
399 * accessed by this target CPU after this point; in particular,
400 * memory pointed by smp_rv_func_arg.
402 atomic_add_int(&smp_rv_waiters[3], 1);
405 KASSERT(owepreempt == td->td_owepreempt,
406 ("rendezvous action changed td_owepreempt"));
410 smp_rendezvous_cpus(cpuset_t map,
411 void (* setup_func)(void *),
412 void (* action_func)(void *),
413 void (* teardown_func)(void *),
416 int curcpumap, i, ncpus = 0;
418 /* Look comments in the !SMP case. */
421 if (setup_func != NULL)
423 if (action_func != NULL)
425 if (teardown_func != NULL)
432 if (CPU_ISSET(i, &map))
436 panic("ncpus is 0 with non-zero map");
438 mtx_lock_spin(&smp_ipi_mtx);
440 /* Pass rendezvous parameters via global variables. */
441 smp_rv_ncpus = ncpus;
442 smp_rv_setup_func = setup_func;
443 smp_rv_action_func = action_func;
444 smp_rv_teardown_func = teardown_func;
445 smp_rv_func_arg = arg;
446 smp_rv_waiters[1] = 0;
447 smp_rv_waiters[2] = 0;
448 smp_rv_waiters[3] = 0;
449 atomic_store_rel_int(&smp_rv_waiters[0], 0);
452 * Signal other processors, which will enter the IPI with
455 curcpumap = CPU_ISSET(curcpu, &map);
456 CPU_CLR(curcpu, &map);
457 ipi_selected(map, IPI_RENDEZVOUS);
459 /* Check if the current CPU is in the map */
461 smp_rendezvous_action();
464 * Ensure that the master CPU waits for all the other
465 * CPUs to finish the rendezvous, so that smp_rv_*
466 * pseudo-structure and the arg are guaranteed to not
469 while (atomic_load_acq_int(&smp_rv_waiters[3]) < ncpus)
472 mtx_unlock_spin(&smp_ipi_mtx);
476 smp_rendezvous(void (* setup_func)(void *),
477 void (* action_func)(void *),
478 void (* teardown_func)(void *),
481 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg);
484 static struct cpu_group group[MAXCPU];
489 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
490 struct cpu_group *top;
493 * Check for a fake topology request for debugging purposes.
495 switch (smp_topology) {
497 /* Dual core with no sharing. */
498 top = smp_topo_1level(CG_SHARE_NONE, 2, 0);
501 /* No topology, all cpus are equal. */
502 top = smp_topo_none();
505 /* Dual core with shared L2. */
506 top = smp_topo_1level(CG_SHARE_L2, 2, 0);
509 /* quad core, shared l3 among each package, private l2. */
510 top = smp_topo_1level(CG_SHARE_L3, 4, 0);
513 /* quad core, 2 dualcore parts on each package share l2. */
514 top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0);
517 /* Single-core 2xHTT */
518 top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT);
521 /* quad core with a shared l3, 8 threads sharing L2. */
522 top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8,
526 /* Default, ask the system what it wants. */
531 * Verify the returned topology.
533 if (top->cg_count != mp_ncpus)
534 panic("Built bad topology at %p. CPU count %d != %d",
535 top, top->cg_count, mp_ncpus);
536 if (CPU_CMP(&top->cg_mask, &all_cpus))
537 panic("Built bad topology at %p. CPU mask (%s) != (%s)",
538 top, cpusetobj_strprint(cpusetbuf, &top->cg_mask),
539 cpusetobj_strprint(cpusetbuf2, &all_cpus));
546 struct cpu_group *top;
549 top->cg_parent = NULL;
550 top->cg_child = NULL;
551 top->cg_mask = all_cpus;
552 top->cg_count = mp_ncpus;
553 top->cg_children = 0;
554 top->cg_level = CG_SHARE_NONE;
561 smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share,
562 int count, int flags, int start)
564 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
569 for (i = 0; i < count; i++, start++)
570 CPU_SET(start, &mask);
571 child->cg_parent = parent;
572 child->cg_child = NULL;
573 child->cg_children = 0;
574 child->cg_level = share;
575 child->cg_count = count;
576 child->cg_flags = flags;
577 child->cg_mask = mask;
578 parent->cg_children++;
579 for (; parent != NULL; parent = parent->cg_parent) {
580 if (CPU_OVERLAP(&parent->cg_mask, &child->cg_mask))
581 panic("Duplicate children in %p. mask (%s) child (%s)",
583 cpusetobj_strprint(cpusetbuf, &parent->cg_mask),
584 cpusetobj_strprint(cpusetbuf2, &child->cg_mask));
585 CPU_OR(&parent->cg_mask, &child->cg_mask);
586 parent->cg_count += child->cg_count;
593 smp_topo_1level(int share, int count, int flags)
595 struct cpu_group *child;
596 struct cpu_group *top;
603 packages = mp_ncpus / count;
604 top->cg_child = child = &group[1];
605 top->cg_level = CG_SHARE_NONE;
606 for (i = 0; i < packages; i++, child++)
607 cpu = smp_topo_addleaf(top, child, share, count, flags, cpu);
612 smp_topo_2level(int l2share, int l2count, int l1share, int l1count,
615 struct cpu_group *top;
616 struct cpu_group *l1g;
617 struct cpu_group *l2g;
626 top->cg_level = CG_SHARE_NONE;
627 top->cg_children = mp_ncpus / (l2count * l1count);
628 l1g = l2g + top->cg_children;
629 for (i = 0; i < top->cg_children; i++, l2g++) {
630 l2g->cg_parent = top;
632 l2g->cg_level = l2share;
633 for (j = 0; j < l2count; j++, l1g++)
634 cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count,
642 smp_topo_find(struct cpu_group *top, int cpu)
644 struct cpu_group *cg;
649 CPU_SETOF(cpu, &mask);
652 if (!CPU_OVERLAP(&cg->cg_mask, &mask))
654 if (cg->cg_children == 0)
656 children = cg->cg_children;
657 for (i = 0, cg = cg->cg_child; i < children; cg++, i++)
658 if (CPU_OVERLAP(&cg->cg_mask, &mask))
666 smp_rendezvous_cpus(cpuset_t map,
667 void (*setup_func)(void *),
668 void (*action_func)(void *),
669 void (*teardown_func)(void *),
673 * In the !SMP case we just need to ensure the same initial conditions
677 if (setup_func != NULL)
679 if (action_func != NULL)
681 if (teardown_func != NULL)
687 smp_rendezvous(void (*setup_func)(void *),
688 void (*action_func)(void *),
689 void (*teardown_func)(void *),
693 /* Look comments in the smp_rendezvous_cpus() case. */
695 if (setup_func != NULL)
697 if (action_func != NULL)
699 if (teardown_func != NULL)
705 * Provide dummy SMP support for UP kernels. Modules that need to use SMP
706 * APIs will still work using this dummy support.
709 mp_setvariables_for_up(void *dummy)
712 mp_maxid = PCPU_GET(cpuid);
713 CPU_SETOF(mp_maxid, &all_cpus);
714 KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero"));
716 SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST,
717 mp_setvariables_for_up, NULL);
721 smp_no_rendevous_barrier(void *dummy)
724 KASSERT((!smp_started),("smp_no_rendevous called and smp is started"));