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 cpumask_t stopped_cpus;
57 volatile cpumask_t started_cpus;
58 cpumask_t idle_cpus_mask;
59 cpumask_t hlt_cpus_mask;
60 cpumask_t logical_cpus_mask;
62 void (*cpustop_restartfunc)(void);
64 /* This is used in modules that need to work in both SMP and UP. */
68 /* export this for libkvm consumers. */
69 int mp_maxcpus = MAXCPU;
71 volatile int smp_started;
74 SYSCTL_NODE(_kern, OID_AUTO, smp, CTLFLAG_RD, NULL, "Kernel SMP");
76 SYSCTL_INT(_kern_smp, OID_AUTO, maxid, CTLFLAG_RD, &mp_maxid, 0,
79 SYSCTL_INT(_kern_smp, OID_AUTO, maxcpus, CTLFLAG_RD, &mp_maxcpus, 0,
80 "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, &smp_disabled, 0,
88 "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, &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[3];
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 all_cpus = PCPU_GET(cpumask);
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(cpumask_t map, u_int type)
205 static volatile u_int stopping_cpu = NOCPU;
209 #if defined(__amd64__)
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(%x) with %u type", map, type);
221 if (stopping_cpu != PCPU_GET(cpuid))
222 while (atomic_cmpset_int(&stopping_cpu, NOCPU,
223 PCPU_GET(cpuid)) == 0)
224 while (stopping_cpu != NOCPU)
225 cpu_spinwait(); /* spin */
227 /* send the stop IPI to all CPUs in map */
228 ipi_selected(map, type);
231 while ((stopped_cpus & map) != map) {
237 printf("timeout stopping cpus\n");
243 stopping_cpu = NOCPU;
248 stop_cpus(cpumask_t map)
251 return (generic_stop_cpus(map, IPI_STOP));
255 stop_cpus_hard(cpumask_t map)
258 return (generic_stop_cpus(map, IPI_STOP_HARD));
261 #if defined(__amd64__)
263 suspend_cpus(cpumask_t map)
266 return (generic_stop_cpus(map, IPI_SUSPEND));
271 * Called by a CPU to restart stopped CPUs.
273 * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
275 * - Signals all CPUs in map to restart.
276 * - Waits for each to restart.
284 restart_cpus(cpumask_t map)
290 CTR1(KTR_SMP, "restart_cpus(%x)", map);
292 /* signal other cpus to restart */
293 atomic_store_rel_int(&started_cpus, map);
295 /* wait for each to clear its bit */
296 while ((stopped_cpus & map) != 0)
303 * All-CPU rendezvous. CPUs are signalled, all execute the setup function
304 * (if specified), rendezvous, execute the action function (if specified),
305 * rendezvous again, execute the teardown function (if specified), and then
308 * Note that the supplied external functions _must_ be reentrant and aware
309 * that they are running in parallel and in an unknown lock context.
312 smp_rendezvous_action(void)
314 void* local_func_arg = smp_rv_func_arg;
315 void (*local_setup_func)(void*) = smp_rv_setup_func;
316 void (*local_action_func)(void*) = smp_rv_action_func;
317 void (*local_teardown_func)(void*) = smp_rv_teardown_func;
319 /* Ensure we have up-to-date values. */
320 atomic_add_acq_int(&smp_rv_waiters[0], 1);
321 while (smp_rv_waiters[0] < smp_rv_ncpus)
325 if (local_setup_func != smp_no_rendevous_barrier) {
326 if (smp_rv_setup_func != NULL)
327 smp_rv_setup_func(smp_rv_func_arg);
329 /* spin on entry rendezvous */
330 atomic_add_int(&smp_rv_waiters[1], 1);
331 while (smp_rv_waiters[1] < smp_rv_ncpus)
335 /* action function */
336 if (local_action_func != NULL)
337 local_action_func(local_func_arg);
339 /* spin on exit rendezvous */
340 atomic_add_int(&smp_rv_waiters[2], 1);
341 if (local_teardown_func == smp_no_rendevous_barrier)
343 while (smp_rv_waiters[2] < smp_rv_ncpus)
346 /* teardown function */
347 if (local_teardown_func != NULL)
348 local_teardown_func(local_func_arg);
352 smp_rendezvous_cpus(cpumask_t map,
353 void (* setup_func)(void *),
354 void (* action_func)(void *),
355 void (* teardown_func)(void *),
361 if (setup_func != NULL)
363 if (action_func != NULL)
365 if (teardown_func != NULL)
371 if (((1 << i) & map) != 0)
375 panic("ncpus is 0 with map=0x%x", map);
377 /* obtain rendezvous lock */
378 mtx_lock_spin(&smp_ipi_mtx);
380 /* set static function pointers */
381 smp_rv_ncpus = ncpus;
382 smp_rv_setup_func = setup_func;
383 smp_rv_action_func = action_func;
384 smp_rv_teardown_func = teardown_func;
385 smp_rv_func_arg = arg;
386 smp_rv_waiters[1] = 0;
387 smp_rv_waiters[2] = 0;
388 atomic_store_rel_int(&smp_rv_waiters[0], 0);
390 /* signal other processors, which will enter the IPI with interrupts off */
391 ipi_selected(map & ~(1 << curcpu), IPI_RENDEZVOUS);
393 /* Check if the current CPU is in the map */
394 if ((map & (1 << curcpu)) != 0)
395 smp_rendezvous_action();
397 if (teardown_func == smp_no_rendevous_barrier)
398 while (atomic_load_acq_int(&smp_rv_waiters[2]) < ncpus)
402 mtx_unlock_spin(&smp_ipi_mtx);
406 smp_rendezvous(void (* setup_func)(void *),
407 void (* action_func)(void *),
408 void (* teardown_func)(void *),
411 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg);
414 static struct cpu_group group[MAXCPU];
419 struct cpu_group *top;
422 * Check for a fake topology request for debugging purposes.
424 switch (smp_topology) {
426 /* Dual core with no sharing. */
427 top = smp_topo_1level(CG_SHARE_NONE, 2, 0);
430 /* No topology, all cpus are equal. */
431 top = smp_topo_none();
434 /* Dual core with shared L2. */
435 top = smp_topo_1level(CG_SHARE_L2, 2, 0);
438 /* quad core, shared l3 among each package, private l2. */
439 top = smp_topo_1level(CG_SHARE_L3, 4, 0);
442 /* quad core, 2 dualcore parts on each package share l2. */
443 top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0);
446 /* Single-core 2xHTT */
447 top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT);
450 /* quad core with a shared l3, 8 threads sharing L2. */
451 top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8,
455 /* Default, ask the system what it wants. */
460 * Verify the returned topology.
462 if (top->cg_count != mp_ncpus)
463 panic("Built bad topology at %p. CPU count %d != %d",
464 top, top->cg_count, mp_ncpus);
465 if (top->cg_mask != all_cpus)
466 panic("Built bad topology at %p. CPU mask 0x%X != 0x%X",
467 top, top->cg_mask, all_cpus);
474 struct cpu_group *top;
477 top->cg_parent = NULL;
478 top->cg_child = NULL;
479 top->cg_mask = ~0U >> (32 - mp_ncpus);
480 top->cg_count = mp_ncpus;
481 top->cg_children = 0;
482 top->cg_level = CG_SHARE_NONE;
489 smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share,
490 int count, int flags, int start)
495 for (mask = 0, i = 0; i < count; i++, start++)
496 mask |= (1 << start);
497 child->cg_parent = parent;
498 child->cg_child = NULL;
499 child->cg_children = 0;
500 child->cg_level = share;
501 child->cg_count = count;
502 child->cg_flags = flags;
503 child->cg_mask = mask;
504 parent->cg_children++;
505 for (; parent != NULL; parent = parent->cg_parent) {
506 if ((parent->cg_mask & child->cg_mask) != 0)
507 panic("Duplicate children in %p. mask 0x%X child 0x%X",
508 parent, parent->cg_mask, child->cg_mask);
509 parent->cg_mask |= child->cg_mask;
510 parent->cg_count += child->cg_count;
517 smp_topo_1level(int share, int count, int flags)
519 struct cpu_group *child;
520 struct cpu_group *top;
527 packages = mp_ncpus / count;
528 top->cg_child = child = &group[1];
529 top->cg_level = CG_SHARE_NONE;
530 for (i = 0; i < packages; i++, child++)
531 cpu = smp_topo_addleaf(top, child, share, count, flags, cpu);
536 smp_topo_2level(int l2share, int l2count, int l1share, int l1count,
539 struct cpu_group *top;
540 struct cpu_group *l1g;
541 struct cpu_group *l2g;
550 top->cg_level = CG_SHARE_NONE;
551 top->cg_children = mp_ncpus / (l2count * l1count);
552 l1g = l2g + top->cg_children;
553 for (i = 0; i < top->cg_children; i++, l2g++) {
554 l2g->cg_parent = top;
556 l2g->cg_level = l2share;
557 for (j = 0; j < l2count; j++, l1g++)
558 cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count,
566 smp_topo_find(struct cpu_group *top, int cpu)
568 struct cpu_group *cg;
576 if ((cg->cg_mask & mask) == 0)
578 if (cg->cg_children == 0)
580 children = cg->cg_children;
581 for (i = 0, cg = cg->cg_child; i < children; cg++, i++)
582 if ((cg->cg_mask & mask) != 0)
590 smp_rendezvous_cpus(cpumask_t map,
591 void (*setup_func)(void *),
592 void (*action_func)(void *),
593 void (*teardown_func)(void *),
596 if (setup_func != NULL)
598 if (action_func != NULL)
600 if (teardown_func != NULL)
605 smp_rendezvous(void (*setup_func)(void *),
606 void (*action_func)(void *),
607 void (*teardown_func)(void *),
611 if (setup_func != NULL)
613 if (action_func != NULL)
615 if (teardown_func != NULL)
620 * Provide dummy SMP support for UP kernels. Modules that need to use SMP
621 * APIs will still work using this dummy support.
624 mp_setvariables_for_up(void *dummy)
627 mp_maxid = PCPU_GET(cpuid);
628 all_cpus = PCPU_GET(cpumask);
629 KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero"));
631 SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST,
632 mp_setvariables_for_up, NULL);
636 smp_no_rendevous_barrier(void *dummy)
639 KASSERT((!smp_started),("smp_no_rendevous called and smp is started"));