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>
45 #include <sys/sysctl.h>
47 #include <machine/cpu.h>
48 #include <machine/smp.h>
50 #include "opt_sched.h"
53 volatile cpuset_t stopped_cpus;
54 volatile cpuset_t started_cpus;
55 cpuset_t hlt_cpus_mask;
56 cpuset_t logical_cpus_mask;
58 void (*cpustop_restartfunc)(void);
60 /* This is used in modules that need to work in both SMP and UP. */
64 /* export this for libkvm consumers. */
65 int mp_maxcpus = MAXCPU;
67 volatile int smp_started;
70 static SYSCTL_NODE(_kern, OID_AUTO, smp, CTLFLAG_RD|CTLFLAG_CAPRD, NULL,
73 SYSCTL_INT(_kern_smp, OID_AUTO, maxid, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxid, 0,
76 SYSCTL_INT(_kern_smp, OID_AUTO, maxcpus, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxcpus,
77 0, "Max number of CPUs that the system was compiled for.");
79 int smp_active = 0; /* are the APs allowed to run? */
80 SYSCTL_INT(_kern_smp, OID_AUTO, active, CTLFLAG_RW, &smp_active, 0,
81 "Number of Auxillary Processors (APs) that were successfully started");
83 int smp_disabled = 0; /* has smp been disabled? */
84 SYSCTL_INT(_kern_smp, OID_AUTO, disabled, CTLFLAG_RDTUN|CTLFLAG_CAPRD,
85 &smp_disabled, 0, "SMP has been disabled from the loader");
86 TUNABLE_INT("kern.smp.disabled", &smp_disabled);
88 int smp_cpus = 1; /* how many cpu's running */
89 SYSCTL_INT(_kern_smp, OID_AUTO, cpus, CTLFLAG_RD|CTLFLAG_CAPRD, &smp_cpus, 0,
90 "Number of CPUs online");
92 int smp_topology = 0; /* Which topology we're using. */
93 SYSCTL_INT(_kern_smp, OID_AUTO, topology, CTLFLAG_RD, &smp_topology, 0,
94 "Topology override setting; 0 is default provided by hardware.");
95 TUNABLE_INT("kern.smp.topology", &smp_topology);
98 /* Enable forwarding of a signal to a process running on a different CPU */
99 static int forward_signal_enabled = 1;
100 SYSCTL_INT(_kern_smp, OID_AUTO, forward_signal_enabled, CTLFLAG_RW,
101 &forward_signal_enabled, 0,
102 "Forwarding of a signal to a process on a different CPU");
104 /* Variables needed for SMP rendezvous. */
105 static volatile int smp_rv_ncpus;
106 static void (*volatile smp_rv_setup_func)(void *arg);
107 static void (*volatile smp_rv_action_func)(void *arg);
108 static void (*volatile smp_rv_teardown_func)(void *arg);
109 static void *volatile smp_rv_func_arg;
110 static volatile int smp_rv_waiters[4];
113 * Shared mutex to restrict busywaits between smp_rendezvous() and
114 * smp(_targeted)_tlb_shootdown(). A deadlock occurs if both of these
115 * functions trigger at once and cause multiple CPUs to busywait with
116 * interrupts disabled.
118 struct mtx smp_ipi_mtx;
121 * Let the MD SMP code initialize mp_maxid very early if it can.
124 mp_setmaxid(void *dummy)
128 SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL);
131 * Call the MD SMP initialization code.
134 mp_start(void *dummy)
137 mtx_init(&smp_ipi_mtx, "smp rendezvous", NULL, MTX_SPIN);
139 /* Probe for MP hardware. */
140 if (smp_disabled != 0 || cpu_mp_probe() == 0) {
142 CPU_SETOF(PCPU_GET(cpuid), &all_cpus);
147 printf("FreeBSD/SMP: Multiprocessor System Detected: %d CPUs\n",
151 SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL);
154 forward_signal(struct thread *td)
159 * signotify() has already set TDF_ASTPENDING and TDF_NEEDSIGCHECK on
160 * this thread, so all we need to do is poke it if it is currently
161 * executing so that it executes ast().
163 THREAD_LOCK_ASSERT(td, MA_OWNED);
164 KASSERT(TD_IS_RUNNING(td),
165 ("forward_signal: thread is not TDS_RUNNING"));
167 CTR1(KTR_SMP, "forward_signal(%p)", td->td_proc);
169 if (!smp_started || cold || panicstr)
171 if (!forward_signal_enabled)
174 /* No need to IPI ourself. */
181 ipi_cpu(id, IPI_AST);
185 * When called the executing CPU will send an IPI to all other CPUs
186 * requesting that they halt execution.
188 * Usually (but not necessarily) called with 'other_cpus' as its arg.
190 * - Signals all CPUs in map to stop.
191 * - Waits for each to stop.
200 generic_stop_cpus(cpuset_t map, u_int type)
203 char cpusetbuf[CPUSETBUFSIZ];
205 static volatile u_int stopping_cpu = NOCPU;
209 #if defined(__amd64__) || defined(__i386__)
210 type == IPI_STOP || type == IPI_STOP_HARD || type == IPI_SUSPEND,
212 type == IPI_STOP || type == IPI_STOP_HARD,
214 ("%s: invalid stop type", __func__));
219 CTR2(KTR_SMP, "stop_cpus(%s) with %u type",
220 cpusetobj_strprint(cpusetbuf, &map), type);
222 if (stopping_cpu != PCPU_GET(cpuid))
223 while (atomic_cmpset_int(&stopping_cpu, NOCPU,
224 PCPU_GET(cpuid)) == 0)
225 while (stopping_cpu != NOCPU)
226 cpu_spinwait(); /* spin */
228 /* send the stop IPI to all CPUs in map */
229 ipi_selected(map, type);
232 while (!CPU_SUBSET(&stopped_cpus, &map)) {
236 if (i == 100000000) {
237 printf("timeout stopping cpus\n");
242 stopping_cpu = NOCPU;
247 stop_cpus(cpuset_t map)
250 return (generic_stop_cpus(map, IPI_STOP));
254 stop_cpus_hard(cpuset_t map)
257 return (generic_stop_cpus(map, IPI_STOP_HARD));
260 #if defined(__amd64__) || defined(__i386__)
262 suspend_cpus(cpuset_t map)
265 return (generic_stop_cpus(map, IPI_SUSPEND));
270 * Called by a CPU to restart stopped CPUs.
272 * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
274 * - Signals all CPUs in map to restart.
275 * - Waits for each to restart.
283 restart_cpus(cpuset_t map)
286 char cpusetbuf[CPUSETBUFSIZ];
292 CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map));
294 /* signal other cpus to restart */
295 CPU_COPY_STORE_REL(&map, &started_cpus);
297 /* wait for each to clear its bit */
298 while (CPU_OVERLAP(&stopped_cpus, &map))
305 * All-CPU rendezvous. CPUs are signalled, all execute the setup function
306 * (if specified), rendezvous, execute the action function (if specified),
307 * rendezvous again, execute the teardown function (if specified), and then
310 * Note that the supplied external functions _must_ be reentrant and aware
311 * that they are running in parallel and in an unknown lock context.
314 smp_rendezvous_action(void)
317 void *local_func_arg;
318 void (*local_setup_func)(void*);
319 void (*local_action_func)(void*);
320 void (*local_teardown_func)(void*);
325 /* Ensure we have up-to-date values. */
326 atomic_add_acq_int(&smp_rv_waiters[0], 1);
327 while (smp_rv_waiters[0] < smp_rv_ncpus)
330 /* Fetch rendezvous parameters after acquire barrier. */
331 local_func_arg = smp_rv_func_arg;
332 local_setup_func = smp_rv_setup_func;
333 local_action_func = smp_rv_action_func;
334 local_teardown_func = smp_rv_teardown_func;
337 * Use a nested critical section to prevent any preemptions
338 * from occurring during a rendezvous action routine.
339 * Specifically, if a rendezvous handler is invoked via an IPI
340 * and the interrupted thread was in the critical_exit()
341 * function after setting td_critnest to 0 but before
342 * performing a deferred preemption, this routine can be
343 * invoked with td_critnest set to 0 and td_owepreempt true.
344 * In that case, a critical_exit() during the rendezvous
345 * action would trigger a preemption which is not permitted in
346 * a rendezvous action. To fix this, wrap all of the
347 * rendezvous action handlers in a critical section. We
348 * cannot use a regular critical section however as having
349 * critical_exit() preempt from this routine would also be
350 * problematic (the preemption must not occur before the IPI
351 * has been acknowledged via an EOI). Instead, we
352 * intentionally ignore td_owepreempt when leaving the
353 * critical section. This should be harmless because we do
354 * not permit rendezvous action routines to schedule threads,
355 * and thus td_owepreempt should never transition from 0 to 1
356 * during this routine.
361 owepreempt = td->td_owepreempt;
365 * If requested, run a setup function before the main action
366 * function. Ensure all CPUs have completed the setup
367 * function before moving on to the action function.
369 if (local_setup_func != smp_no_rendevous_barrier) {
370 if (smp_rv_setup_func != NULL)
371 smp_rv_setup_func(smp_rv_func_arg);
372 atomic_add_int(&smp_rv_waiters[1], 1);
373 while (smp_rv_waiters[1] < smp_rv_ncpus)
377 if (local_action_func != NULL)
378 local_action_func(local_func_arg);
380 if (local_teardown_func != smp_no_rendevous_barrier) {
382 * Signal that the main action has been completed. If a
383 * full exit rendezvous is requested, then all CPUs will
384 * wait here until all CPUs have finished the main action.
386 atomic_add_int(&smp_rv_waiters[2], 1);
387 while (smp_rv_waiters[2] < smp_rv_ncpus)
390 if (local_teardown_func != NULL)
391 local_teardown_func(local_func_arg);
395 * Signal that the rendezvous is fully completed by this CPU.
396 * This means that no member of smp_rv_* pseudo-structure will be
397 * accessed by this target CPU after this point; in particular,
398 * memory pointed by smp_rv_func_arg.
400 atomic_add_int(&smp_rv_waiters[3], 1);
403 KASSERT(owepreempt == td->td_owepreempt,
404 ("rendezvous action changed td_owepreempt"));
408 smp_rendezvous_cpus(cpuset_t map,
409 void (* setup_func)(void *),
410 void (* action_func)(void *),
411 void (* teardown_func)(void *),
414 int curcpumap, i, ncpus = 0;
416 /* Look comments in the !SMP case. */
419 if (setup_func != NULL)
421 if (action_func != NULL)
423 if (teardown_func != NULL)
430 if (CPU_ISSET(i, &map))
434 panic("ncpus is 0 with non-zero map");
436 mtx_lock_spin(&smp_ipi_mtx);
438 /* Pass rendezvous parameters via global variables. */
439 smp_rv_ncpus = ncpus;
440 smp_rv_setup_func = setup_func;
441 smp_rv_action_func = action_func;
442 smp_rv_teardown_func = teardown_func;
443 smp_rv_func_arg = arg;
444 smp_rv_waiters[1] = 0;
445 smp_rv_waiters[2] = 0;
446 smp_rv_waiters[3] = 0;
447 atomic_store_rel_int(&smp_rv_waiters[0], 0);
450 * Signal other processors, which will enter the IPI with
453 curcpumap = CPU_ISSET(curcpu, &map);
454 CPU_CLR(curcpu, &map);
455 ipi_selected(map, IPI_RENDEZVOUS);
457 /* Check if the current CPU is in the map */
459 smp_rendezvous_action();
462 * Ensure that the master CPU waits for all the other
463 * CPUs to finish the rendezvous, so that smp_rv_*
464 * pseudo-structure and the arg are guaranteed to not
467 while (atomic_load_acq_int(&smp_rv_waiters[3]) < ncpus)
470 mtx_unlock_spin(&smp_ipi_mtx);
474 smp_rendezvous(void (* setup_func)(void *),
475 void (* action_func)(void *),
476 void (* teardown_func)(void *),
479 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg);
482 static struct cpu_group group[MAXCPU];
487 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
488 struct cpu_group *top;
491 * Check for a fake topology request for debugging purposes.
493 switch (smp_topology) {
495 /* Dual core with no sharing. */
496 top = smp_topo_1level(CG_SHARE_NONE, 2, 0);
499 /* No topology, all cpus are equal. */
500 top = smp_topo_none();
503 /* Dual core with shared L2. */
504 top = smp_topo_1level(CG_SHARE_L2, 2, 0);
507 /* quad core, shared l3 among each package, private l2. */
508 top = smp_topo_1level(CG_SHARE_L3, 4, 0);
511 /* quad core, 2 dualcore parts on each package share l2. */
512 top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0);
515 /* Single-core 2xHTT */
516 top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT);
519 /* quad core with a shared l3, 8 threads sharing L2. */
520 top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8,
524 /* Default, ask the system what it wants. */
529 * Verify the returned topology.
531 if (top->cg_count != mp_ncpus)
532 panic("Built bad topology at %p. CPU count %d != %d",
533 top, top->cg_count, mp_ncpus);
534 if (CPU_CMP(&top->cg_mask, &all_cpus))
535 panic("Built bad topology at %p. CPU mask (%s) != (%s)",
536 top, cpusetobj_strprint(cpusetbuf, &top->cg_mask),
537 cpusetobj_strprint(cpusetbuf2, &all_cpus));
544 struct cpu_group *top;
547 top->cg_parent = NULL;
548 top->cg_child = NULL;
549 top->cg_mask = all_cpus;
550 top->cg_count = mp_ncpus;
551 top->cg_children = 0;
552 top->cg_level = CG_SHARE_NONE;
559 smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share,
560 int count, int flags, int start)
562 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
567 for (i = 0; i < count; i++, start++)
568 CPU_SET(start, &mask);
569 child->cg_parent = parent;
570 child->cg_child = NULL;
571 child->cg_children = 0;
572 child->cg_level = share;
573 child->cg_count = count;
574 child->cg_flags = flags;
575 child->cg_mask = mask;
576 parent->cg_children++;
577 for (; parent != NULL; parent = parent->cg_parent) {
578 if (CPU_OVERLAP(&parent->cg_mask, &child->cg_mask))
579 panic("Duplicate children in %p. mask (%s) child (%s)",
581 cpusetobj_strprint(cpusetbuf, &parent->cg_mask),
582 cpusetobj_strprint(cpusetbuf2, &child->cg_mask));
583 CPU_OR(&parent->cg_mask, &child->cg_mask);
584 parent->cg_count += child->cg_count;
591 smp_topo_1level(int share, int count, int flags)
593 struct cpu_group *child;
594 struct cpu_group *top;
601 packages = mp_ncpus / count;
602 top->cg_child = child = &group[1];
603 top->cg_level = CG_SHARE_NONE;
604 for (i = 0; i < packages; i++, child++)
605 cpu = smp_topo_addleaf(top, child, share, count, flags, cpu);
610 smp_topo_2level(int l2share, int l2count, int l1share, int l1count,
613 struct cpu_group *top;
614 struct cpu_group *l1g;
615 struct cpu_group *l2g;
624 top->cg_level = CG_SHARE_NONE;
625 top->cg_children = mp_ncpus / (l2count * l1count);
626 l1g = l2g + top->cg_children;
627 for (i = 0; i < top->cg_children; i++, l2g++) {
628 l2g->cg_parent = top;
630 l2g->cg_level = l2share;
631 for (j = 0; j < l2count; j++, l1g++)
632 cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count,
640 smp_topo_find(struct cpu_group *top, int cpu)
642 struct cpu_group *cg;
647 CPU_SETOF(cpu, &mask);
650 if (!CPU_OVERLAP(&cg->cg_mask, &mask))
652 if (cg->cg_children == 0)
654 children = cg->cg_children;
655 for (i = 0, cg = cg->cg_child; i < children; cg++, i++)
656 if (CPU_OVERLAP(&cg->cg_mask, &mask))
664 smp_rendezvous_cpus(cpuset_t map,
665 void (*setup_func)(void *),
666 void (*action_func)(void *),
667 void (*teardown_func)(void *),
671 * In the !SMP case we just need to ensure the same initial conditions
675 if (setup_func != NULL)
677 if (action_func != NULL)
679 if (teardown_func != NULL)
685 smp_rendezvous(void (*setup_func)(void *),
686 void (*action_func)(void *),
687 void (*teardown_func)(void *),
691 /* Look comments in the smp_rendezvous_cpus() case. */
693 if (setup_func != NULL)
695 if (action_func != NULL)
697 if (teardown_func != NULL)
703 * Provide dummy SMP support for UP kernels. Modules that need to use SMP
704 * APIs will still work using this dummy support.
707 mp_setvariables_for_up(void *dummy)
710 mp_maxid = PCPU_GET(cpuid);
711 CPU_SETOF(mp_maxid, &all_cpus);
712 KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero"));
714 SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST,
715 mp_setvariables_for_up, NULL);
719 smp_no_rendevous_barrier(void *dummy)
722 KASSERT((!smp_started),("smp_no_rendevous called and smp is started"));