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 static SYSCTL_NODE(_kern, OID_AUTO, smp, CTLFLAG_RD|CTLFLAG_CAPRD, NULL,
76 SYSCTL_INT(_kern_smp, OID_AUTO, maxid, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxid, 0,
79 SYSCTL_INT(_kern_smp, OID_AUTO, maxcpus, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxcpus,
80 0, "Max number of CPUs that the system was compiled for.");
82 int smp_active = 0; /* are the APs allowed to run? */
83 SYSCTL_INT(_kern_smp, OID_AUTO, active, CTLFLAG_RW, &smp_active, 0,
84 "Number of Auxillary Processors (APs) that were successfully started");
86 int smp_disabled = 0; /* has smp been disabled? */
87 SYSCTL_INT(_kern_smp, OID_AUTO, disabled, CTLFLAG_RDTUN|CTLFLAG_CAPRD,
88 &smp_disabled, 0, "SMP has been disabled from the loader");
89 TUNABLE_INT("kern.smp.disabled", &smp_disabled);
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_RD, &smp_topology, 0,
97 "Topology override setting; 0 is default provided by hardware.");
98 TUNABLE_INT("kern.smp.topology", &smp_topology);
101 /* Enable forwarding of a signal to a process running on a different CPU */
102 static int forward_signal_enabled = 1;
103 SYSCTL_INT(_kern_smp, OID_AUTO, forward_signal_enabled, CTLFLAG_RW,
104 &forward_signal_enabled, 0,
105 "Forwarding of a signal to a process on a different CPU");
107 /* Variables needed for SMP rendezvous. */
108 static volatile int smp_rv_ncpus;
109 static void (*volatile smp_rv_setup_func)(void *arg);
110 static void (*volatile smp_rv_action_func)(void *arg);
111 static void (*volatile smp_rv_teardown_func)(void *arg);
112 static void *volatile smp_rv_func_arg;
113 static volatile int smp_rv_waiters[4];
116 * Shared mutex to restrict busywaits between smp_rendezvous() and
117 * smp(_targeted)_tlb_shootdown(). A deadlock occurs if both of these
118 * functions trigger at once and cause multiple CPUs to busywait with
119 * interrupts disabled.
121 struct mtx smp_ipi_mtx;
124 * Let the MD SMP code initialize mp_maxid very early if it can.
127 mp_setmaxid(void *dummy)
131 SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL);
134 * Call the MD SMP initialization code.
137 mp_start(void *dummy)
140 mtx_init(&smp_ipi_mtx, "smp rendezvous", NULL, MTX_SPIN);
142 /* Probe for MP hardware. */
143 if (smp_disabled != 0 || cpu_mp_probe() == 0) {
145 CPU_SETOF(PCPU_GET(cpuid), &all_cpus);
150 printf("FreeBSD/SMP: Multiprocessor System Detected: %d CPUs\n",
154 SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL);
157 forward_signal(struct thread *td)
162 * signotify() has already set TDF_ASTPENDING and TDF_NEEDSIGCHECK on
163 * this thread, so all we need to do is poke it if it is currently
164 * executing so that it executes ast().
166 THREAD_LOCK_ASSERT(td, MA_OWNED);
167 KASSERT(TD_IS_RUNNING(td),
168 ("forward_signal: thread is not TDS_RUNNING"));
170 CTR1(KTR_SMP, "forward_signal(%p)", td->td_proc);
172 if (!smp_started || cold || panicstr)
174 if (!forward_signal_enabled)
177 /* No need to IPI ourself. */
184 ipi_cpu(id, IPI_AST);
188 * When called the executing CPU will send an IPI to all other CPUs
189 * requesting that they halt execution.
191 * Usually (but not necessarily) called with 'other_cpus' as its arg.
193 * - Signals all CPUs in map to stop.
194 * - Waits for each to stop.
203 generic_stop_cpus(cpuset_t map, u_int type)
206 char cpusetbuf[CPUSETBUFSIZ];
208 static volatile u_int stopping_cpu = NOCPU;
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 (stopping_cpu != PCPU_GET(cpuid))
226 while (atomic_cmpset_int(&stopping_cpu, NOCPU,
227 PCPU_GET(cpuid)) == 0)
228 while (stopping_cpu != NOCPU)
229 cpu_spinwait(); /* spin */
231 /* send the stop IPI to all CPUs in map */
232 ipi_selected(map, type);
235 while (!CPU_SUBSET(&stopped_cpus, &map)) {
239 if (i == 100000000) {
240 printf("timeout stopping cpus\n");
245 stopping_cpu = NOCPU;
250 stop_cpus(cpuset_t map)
253 return (generic_stop_cpus(map, IPI_STOP));
257 stop_cpus_hard(cpuset_t map)
260 return (generic_stop_cpus(map, IPI_STOP_HARD));
263 #if defined(__amd64__) || defined(__i386__)
265 suspend_cpus(cpuset_t map)
268 return (generic_stop_cpus(map, IPI_SUSPEND));
273 * Called by a CPU to restart stopped CPUs.
275 * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
277 * - Signals all CPUs in map to restart.
278 * - Waits for each to restart.
286 restart_cpus(cpuset_t map)
289 char cpusetbuf[CPUSETBUFSIZ];
295 CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map));
297 /* signal other cpus to restart */
298 CPU_COPY_STORE_REL(&map, &started_cpus);
300 /* wait for each to clear its bit */
301 while (CPU_OVERLAP(&stopped_cpus, &map))
308 * All-CPU rendezvous. CPUs are signalled, all execute the setup function
309 * (if specified), rendezvous, execute the action function (if specified),
310 * rendezvous again, execute the teardown function (if specified), and then
313 * Note that the supplied external functions _must_ be reentrant and aware
314 * that they are running in parallel and in an unknown lock context.
317 smp_rendezvous_action(void)
320 void *local_func_arg;
321 void (*local_setup_func)(void*);
322 void (*local_action_func)(void*);
323 void (*local_teardown_func)(void*);
328 /* Ensure we have up-to-date values. */
329 atomic_add_acq_int(&smp_rv_waiters[0], 1);
330 while (smp_rv_waiters[0] < smp_rv_ncpus)
333 /* Fetch rendezvous parameters after acquire barrier. */
334 local_func_arg = smp_rv_func_arg;
335 local_setup_func = smp_rv_setup_func;
336 local_action_func = smp_rv_action_func;
337 local_teardown_func = smp_rv_teardown_func;
340 * Use a nested critical section to prevent any preemptions
341 * from occurring during a rendezvous action routine.
342 * Specifically, if a rendezvous handler is invoked via an IPI
343 * and the interrupted thread was in the critical_exit()
344 * function after setting td_critnest to 0 but before
345 * performing a deferred preemption, this routine can be
346 * invoked with td_critnest set to 0 and td_owepreempt true.
347 * In that case, a critical_exit() during the rendezvous
348 * action would trigger a preemption which is not permitted in
349 * a rendezvous action. To fix this, wrap all of the
350 * rendezvous action handlers in a critical section. We
351 * cannot use a regular critical section however as having
352 * critical_exit() preempt from this routine would also be
353 * problematic (the preemption must not occur before the IPI
354 * has been acknowledged via an EOI). Instead, we
355 * intentionally ignore td_owepreempt when leaving the
356 * critical section. This should be harmless because we do
357 * not permit rendezvous action routines to schedule threads,
358 * and thus td_owepreempt should never transition from 0 to 1
359 * during this routine.
364 owepreempt = td->td_owepreempt;
368 * If requested, run a setup function before the main action
369 * function. Ensure all CPUs have completed the setup
370 * function before moving on to the action function.
372 if (local_setup_func != smp_no_rendevous_barrier) {
373 if (smp_rv_setup_func != NULL)
374 smp_rv_setup_func(smp_rv_func_arg);
375 atomic_add_int(&smp_rv_waiters[1], 1);
376 while (smp_rv_waiters[1] < smp_rv_ncpus)
380 if (local_action_func != NULL)
381 local_action_func(local_func_arg);
383 if (local_teardown_func != smp_no_rendevous_barrier) {
385 * Signal that the main action has been completed. If a
386 * full exit rendezvous is requested, then all CPUs will
387 * wait here until all CPUs have finished the main action.
389 atomic_add_int(&smp_rv_waiters[2], 1);
390 while (smp_rv_waiters[2] < smp_rv_ncpus)
393 if (local_teardown_func != NULL)
394 local_teardown_func(local_func_arg);
398 * Signal that the rendezvous is fully completed by this CPU.
399 * This means that no member of smp_rv_* pseudo-structure will be
400 * accessed by this target CPU after this point; in particular,
401 * memory pointed by smp_rv_func_arg.
403 atomic_add_int(&smp_rv_waiters[3], 1);
406 KASSERT(owepreempt == td->td_owepreempt,
407 ("rendezvous action changed td_owepreempt"));
411 smp_rendezvous_cpus(cpuset_t map,
412 void (* setup_func)(void *),
413 void (* action_func)(void *),
414 void (* teardown_func)(void *),
417 int curcpumap, i, ncpus = 0;
419 /* Look comments in the !SMP case. */
422 if (setup_func != NULL)
424 if (action_func != NULL)
426 if (teardown_func != NULL)
433 if (CPU_ISSET(i, &map))
437 panic("ncpus is 0 with non-zero map");
439 mtx_lock_spin(&smp_ipi_mtx);
441 /* Pass rendezvous parameters via global variables. */
442 smp_rv_ncpus = ncpus;
443 smp_rv_setup_func = setup_func;
444 smp_rv_action_func = action_func;
445 smp_rv_teardown_func = teardown_func;
446 smp_rv_func_arg = arg;
447 smp_rv_waiters[1] = 0;
448 smp_rv_waiters[2] = 0;
449 smp_rv_waiters[3] = 0;
450 atomic_store_rel_int(&smp_rv_waiters[0], 0);
453 * Signal other processors, which will enter the IPI with
456 curcpumap = CPU_ISSET(curcpu, &map);
457 CPU_CLR(curcpu, &map);
458 ipi_selected(map, IPI_RENDEZVOUS);
460 /* Check if the current CPU is in the map */
462 smp_rendezvous_action();
465 * Ensure that the master CPU waits for all the other
466 * CPUs to finish the rendezvous, so that smp_rv_*
467 * pseudo-structure and the arg are guaranteed to not
470 while (atomic_load_acq_int(&smp_rv_waiters[3]) < ncpus)
473 mtx_unlock_spin(&smp_ipi_mtx);
477 smp_rendezvous(void (* setup_func)(void *),
478 void (* action_func)(void *),
479 void (* teardown_func)(void *),
482 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg);
485 static struct cpu_group group[MAXCPU];
490 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
491 struct cpu_group *top;
494 * Check for a fake topology request for debugging purposes.
496 switch (smp_topology) {
498 /* Dual core with no sharing. */
499 top = smp_topo_1level(CG_SHARE_NONE, 2, 0);
502 /* No topology, all cpus are equal. */
503 top = smp_topo_none();
506 /* Dual core with shared L2. */
507 top = smp_topo_1level(CG_SHARE_L2, 2, 0);
510 /* quad core, shared l3 among each package, private l2. */
511 top = smp_topo_1level(CG_SHARE_L3, 4, 0);
514 /* quad core, 2 dualcore parts on each package share l2. */
515 top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0);
518 /* Single-core 2xHTT */
519 top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT);
522 /* quad core with a shared l3, 8 threads sharing L2. */
523 top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8,
527 /* Default, ask the system what it wants. */
532 * Verify the returned topology.
534 if (top->cg_count != mp_ncpus)
535 panic("Built bad topology at %p. CPU count %d != %d",
536 top, top->cg_count, mp_ncpus);
537 if (CPU_CMP(&top->cg_mask, &all_cpus))
538 panic("Built bad topology at %p. CPU mask (%s) != (%s)",
539 top, cpusetobj_strprint(cpusetbuf, &top->cg_mask),
540 cpusetobj_strprint(cpusetbuf2, &all_cpus));
547 struct cpu_group *top;
550 top->cg_parent = NULL;
551 top->cg_child = NULL;
552 top->cg_mask = all_cpus;
553 top->cg_count = mp_ncpus;
554 top->cg_children = 0;
555 top->cg_level = CG_SHARE_NONE;
562 smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share,
563 int count, int flags, int start)
565 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ];
570 for (i = 0; i < count; i++, start++)
571 CPU_SET(start, &mask);
572 child->cg_parent = parent;
573 child->cg_child = NULL;
574 child->cg_children = 0;
575 child->cg_level = share;
576 child->cg_count = count;
577 child->cg_flags = flags;
578 child->cg_mask = mask;
579 parent->cg_children++;
580 for (; parent != NULL; parent = parent->cg_parent) {
581 if (CPU_OVERLAP(&parent->cg_mask, &child->cg_mask))
582 panic("Duplicate children in %p. mask (%s) child (%s)",
584 cpusetobj_strprint(cpusetbuf, &parent->cg_mask),
585 cpusetobj_strprint(cpusetbuf2, &child->cg_mask));
586 CPU_OR(&parent->cg_mask, &child->cg_mask);
587 parent->cg_count += child->cg_count;
594 smp_topo_1level(int share, int count, int flags)
596 struct cpu_group *child;
597 struct cpu_group *top;
604 packages = mp_ncpus / count;
605 top->cg_child = child = &group[1];
606 top->cg_level = CG_SHARE_NONE;
607 for (i = 0; i < packages; i++, child++)
608 cpu = smp_topo_addleaf(top, child, share, count, flags, cpu);
613 smp_topo_2level(int l2share, int l2count, int l1share, int l1count,
616 struct cpu_group *top;
617 struct cpu_group *l1g;
618 struct cpu_group *l2g;
627 top->cg_level = CG_SHARE_NONE;
628 top->cg_children = mp_ncpus / (l2count * l1count);
629 l1g = l2g + top->cg_children;
630 for (i = 0; i < top->cg_children; i++, l2g++) {
631 l2g->cg_parent = top;
633 l2g->cg_level = l2share;
634 for (j = 0; j < l2count; j++, l1g++)
635 cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count,
643 smp_topo_find(struct cpu_group *top, int cpu)
645 struct cpu_group *cg;
650 CPU_SETOF(cpu, &mask);
653 if (!CPU_OVERLAP(&cg->cg_mask, &mask))
655 if (cg->cg_children == 0)
657 children = cg->cg_children;
658 for (i = 0, cg = cg->cg_child; i < children; cg++, i++)
659 if (CPU_OVERLAP(&cg->cg_mask, &mask))
667 smp_rendezvous_cpus(cpuset_t map,
668 void (*setup_func)(void *),
669 void (*action_func)(void *),
670 void (*teardown_func)(void *),
674 * In the !SMP case we just need to ensure the same initial conditions
678 if (setup_func != NULL)
680 if (action_func != NULL)
682 if (teardown_func != NULL)
688 smp_rendezvous(void (*setup_func)(void *),
689 void (*action_func)(void *),
690 void (*teardown_func)(void *),
694 /* Look comments in the smp_rendezvous_cpus() case. */
696 if (setup_func != NULL)
698 if (action_func != NULL)
700 if (teardown_func != NULL)
706 * Provide dummy SMP support for UP kernels. Modules that need to use SMP
707 * APIs will still work using this dummy support.
710 mp_setvariables_for_up(void *dummy)
713 mp_maxid = PCPU_GET(cpuid);
714 CPU_SETOF(mp_maxid, &all_cpus);
715 KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero"));
717 SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST,
718 mp_setvariables_for_up, NULL);
722 smp_no_rendevous_barrier(void *dummy)
725 KASSERT((!smp_started),("smp_no_rendevous called and smp is started"));