2 * linux/arch/arm/kernel/smp.c
4 * Copyright (C) 2002 ARM Limited, All Rights Reserved.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 as
8 * published by the Free Software Foundation.
10 #include <linux/module.h>
11 #include <linux/delay.h>
12 #include <linux/init.h>
13 #include <linux/spinlock.h>
14 #include <linux/sched.h>
15 #include <linux/interrupt.h>
16 #include <linux/cache.h>
17 #include <linux/profile.h>
18 #include <linux/errno.h>
20 #include <linux/err.h>
21 #include <linux/cpu.h>
22 #include <linux/seq_file.h>
23 #include <linux/irq.h>
24 #include <linux/percpu.h>
25 #include <linux/clockchips.h>
26 #include <linux/completion.h>
27 #include <linux/cpufreq.h>
29 #include <linux/atomic.h>
31 #include <asm/cacheflush.h>
33 #include <asm/cputype.h>
34 #include <asm/exception.h>
35 #include <asm/idmap.h>
36 #include <asm/topology.h>
37 #include <asm/mmu_context.h>
38 #include <asm/pgtable.h>
39 #include <asm/pgalloc.h>
40 #include <asm/processor.h>
41 #include <asm/sections.h>
42 #include <asm/tlbflush.h>
43 #include <asm/ptrace.h>
44 #include <asm/localtimer.h>
45 #include <asm/smp_plat.h>
47 #include <asm/mach/arch.h>
50 * as from 2.5, kernels no longer have an init_tasks structure
51 * so we need some other way of telling a new secondary core
52 * where to place its SVC stack
54 struct secondary_data secondary_data;
57 * control for which core is the next to come out of the secondary
60 volatile int __cpuinitdata pen_release = -1;
71 static DECLARE_COMPLETION(cpu_running);
73 static struct smp_operations smp_ops;
75 void __init smp_set_ops(struct smp_operations *ops)
81 int __cpuinit __cpu_up(unsigned int cpu, struct task_struct *idle)
86 * We need to tell the secondary core where to find
87 * its stack and the page tables.
89 secondary_data.stack = task_stack_page(idle) + THREAD_START_SP;
90 secondary_data.pgdir = virt_to_phys(idmap_pgd);
91 secondary_data.swapper_pg_dir = virt_to_phys(swapper_pg_dir);
92 __cpuc_flush_dcache_area(&secondary_data, sizeof(secondary_data));
93 outer_clean_range(__pa(&secondary_data), __pa(&secondary_data + 1));
96 * Now bring the CPU into our world.
98 ret = boot_secondary(cpu, idle);
101 * CPU was successfully started, wait for it
102 * to come online or time out.
104 wait_for_completion_timeout(&cpu_running,
105 msecs_to_jiffies(1000));
107 if (!cpu_online(cpu)) {
108 pr_crit("CPU%u: failed to come online\n", cpu);
112 pr_err("CPU%u: failed to boot: %d\n", cpu, ret);
115 secondary_data.stack = NULL;
116 secondary_data.pgdir = 0;
121 /* platform specific SMP operations */
122 void __init smp_init_cpus(void)
124 if (smp_ops.smp_init_cpus)
125 smp_ops.smp_init_cpus();
128 int __cpuinit boot_secondary(unsigned int cpu, struct task_struct *idle)
130 if (smp_ops.smp_boot_secondary)
131 return smp_ops.smp_boot_secondary(cpu, idle);
135 #ifdef CONFIG_HOTPLUG_CPU
136 static void percpu_timer_stop(void);
138 static int platform_cpu_kill(unsigned int cpu)
140 if (smp_ops.cpu_kill)
141 return smp_ops.cpu_kill(cpu);
145 static int platform_cpu_disable(unsigned int cpu)
147 if (smp_ops.cpu_disable)
148 return smp_ops.cpu_disable(cpu);
151 * By default, allow disabling all CPUs except the first one,
152 * since this is special on a lot of platforms, e.g. because
153 * of clock tick interrupts.
155 return cpu == 0 ? -EPERM : 0;
158 * __cpu_disable runs on the processor to be shutdown.
160 int __cpuinit __cpu_disable(void)
162 unsigned int cpu = smp_processor_id();
165 ret = platform_cpu_disable(cpu);
170 * Take this CPU offline. Once we clear this, we can't return,
171 * and we must not schedule until we're ready to give up the cpu.
173 set_cpu_online(cpu, false);
176 * OK - migrate IRQs away from this CPU
181 * Stop the local timer for this CPU.
186 * Flush user cache and TLB mappings, and then remove this CPU
187 * from the vm mask set of all processes.
189 * Caches are flushed to the Level of Unification Inner Shareable
190 * to write-back dirty lines to unified caches shared by all CPUs.
193 local_flush_tlb_all();
195 clear_tasks_mm_cpumask(cpu);
200 static DECLARE_COMPLETION(cpu_died);
203 * called on the thread which is asking for a CPU to be shutdown -
204 * waits until shutdown has completed, or it is timed out.
206 void __cpuinit __cpu_die(unsigned int cpu)
208 if (!wait_for_completion_timeout(&cpu_died, msecs_to_jiffies(5000))) {
209 pr_err("CPU%u: cpu didn't die\n", cpu);
212 printk(KERN_NOTICE "CPU%u: shutdown\n", cpu);
215 * platform_cpu_kill() is generally expected to do the powering off
216 * and/or cutting of clocks to the dying CPU. Optionally, this may
217 * be done by the CPU which is dying in preference to supporting
218 * this call, but that means there is _no_ synchronisation between
219 * the requesting CPU and the dying CPU actually losing power.
221 if (!platform_cpu_kill(cpu))
222 printk("CPU%u: unable to kill\n", cpu);
226 * Called from the idle thread for the CPU which has been shutdown.
228 * Note that we disable IRQs here, but do not re-enable them
229 * before returning to the caller. This is also the behaviour
230 * of the other hotplug-cpu capable cores, so presumably coming
231 * out of idle fixes this.
233 void __ref cpu_die(void)
235 unsigned int cpu = smp_processor_id();
242 * Flush the data out of the L1 cache for this CPU. This must be
243 * before the completion to ensure that data is safely written out
244 * before platform_cpu_kill() gets called - which may disable
245 * *this* CPU and power down its cache.
250 * Tell __cpu_die() that this CPU is now safe to dispose of. Once
251 * this returns, power and/or clocks can be removed at any point
252 * from this CPU and its cache by platform_cpu_kill().
257 * Ensure that the cache lines associated with that completion are
258 * written out. This covers the case where _this_ CPU is doing the
259 * powering down, to ensure that the completion is visible to the
260 * CPU waiting for this one.
265 * The actual CPU shutdown procedure is at least platform (if not
266 * CPU) specific. This may remove power, or it may simply spin.
268 * Platforms are generally expected *NOT* to return from this call,
269 * although there are some which do because they have no way to
270 * power down the CPU. These platforms are the _only_ reason we
271 * have a return path which uses the fragment of assembly below.
273 * The return path should not be used for platforms which can
277 smp_ops.cpu_die(cpu);
280 * Do not return to the idle loop - jump back to the secondary
281 * cpu initialisation. There's some initialisation which needs
282 * to be repeated to undo the effects of taking the CPU offline.
284 __asm__("mov sp, %0\n"
286 " b secondary_start_kernel"
288 : "r" (task_stack_page(current) + THREAD_SIZE - 8));
290 #endif /* CONFIG_HOTPLUG_CPU */
293 * Called by both boot and secondaries to move global data into
294 * per-processor storage.
296 static void __cpuinit smp_store_cpu_info(unsigned int cpuid)
298 struct cpuinfo_arm *cpu_info = &per_cpu(cpu_data, cpuid);
300 cpu_info->loops_per_jiffy = loops_per_jiffy;
301 cpu_info->cpuid = read_cpuid_id();
303 store_cpu_topology(cpuid);
306 static void percpu_timer_setup(void);
309 * This is the secondary CPU boot entry. We're using this CPUs
310 * idle thread stack, but a set of temporary page tables.
312 asmlinkage void __cpuinit secondary_start_kernel(void)
314 struct mm_struct *mm = &init_mm;
318 * The identity mapping is uncached (strongly ordered), so
319 * switch away from it before attempting any exclusive accesses.
321 cpu_switch_mm(mm->pgd, mm);
322 local_flush_bp_all();
323 enter_lazy_tlb(mm, current);
324 local_flush_tlb_all();
327 * All kernel threads share the same mm context; grab a
328 * reference and switch to it.
330 cpu = smp_processor_id();
331 atomic_inc(&mm->mm_count);
332 current->active_mm = mm;
333 cpumask_set_cpu(cpu, mm_cpumask(mm));
337 printk("CPU%u: Booted secondary processor\n", cpu);
340 trace_hardirqs_off();
343 * Give the platform a chance to do its own initialisation.
345 if (smp_ops.smp_secondary_init)
346 smp_ops.smp_secondary_init(cpu);
348 notify_cpu_starting(cpu);
352 smp_store_cpu_info(cpu);
355 * OK, now it's safe to let the boot CPU continue. Wait for
356 * the CPU migration code to notice that the CPU is online
357 * before we continue - which happens after __cpu_up returns.
359 set_cpu_online(cpu, true);
360 complete(&cpu_running);
363 * Setup the percpu timer for this CPU.
365 percpu_timer_setup();
371 * OK, it's off to the idle thread for us
373 cpu_startup_entry(CPUHP_ONLINE);
376 void __init smp_cpus_done(unsigned int max_cpus)
379 unsigned long bogosum = 0;
381 for_each_online_cpu(cpu)
382 bogosum += per_cpu(cpu_data, cpu).loops_per_jiffy;
384 printk(KERN_INFO "SMP: Total of %d processors activated "
385 "(%lu.%02lu BogoMIPS).\n",
387 bogosum / (500000/HZ),
388 (bogosum / (5000/HZ)) % 100);
393 void __init smp_prepare_boot_cpu(void)
395 set_my_cpu_offset(per_cpu_offset(smp_processor_id()));
398 void __init smp_prepare_cpus(unsigned int max_cpus)
400 unsigned int ncores = num_possible_cpus();
404 smp_store_cpu_info(smp_processor_id());
407 * are we trying to boot more cores than exist?
409 if (max_cpus > ncores)
411 if (ncores > 1 && max_cpus) {
413 * Enable the local timer or broadcast device for the
414 * boot CPU, but only if we have more than one CPU.
416 percpu_timer_setup();
419 * Initialise the present map, which describes the set of CPUs
420 * actually populated at the present time. A platform should
421 * re-initialize the map in the platforms smp_prepare_cpus()
422 * if present != possible (e.g. physical hotplug).
424 init_cpu_present(cpu_possible_mask);
427 * Initialise the SCU if there are more than one CPU
428 * and let them know where to start.
430 if (smp_ops.smp_prepare_cpus)
431 smp_ops.smp_prepare_cpus(max_cpus);
435 static void (*smp_cross_call)(const struct cpumask *, unsigned int);
437 void __init set_smp_cross_call(void (*fn)(const struct cpumask *, unsigned int))
443 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
445 smp_cross_call(mask, IPI_CALL_FUNC);
448 void arch_send_wakeup_ipi_mask(const struct cpumask *mask)
450 smp_cross_call(mask, IPI_WAKEUP);
453 void arch_send_call_function_single_ipi(int cpu)
455 smp_cross_call(cpumask_of(cpu), IPI_CALL_FUNC_SINGLE);
458 static const char *ipi_types[NR_IPI] = {
459 #define S(x,s) [x] = s
460 S(IPI_WAKEUP, "CPU wakeup interrupts"),
461 S(IPI_TIMER, "Timer broadcast interrupts"),
462 S(IPI_RESCHEDULE, "Rescheduling interrupts"),
463 S(IPI_CALL_FUNC, "Function call interrupts"),
464 S(IPI_CALL_FUNC_SINGLE, "Single function call interrupts"),
465 S(IPI_CPU_STOP, "CPU stop interrupts"),
468 void show_ipi_list(struct seq_file *p, int prec)
472 for (i = 0; i < NR_IPI; i++) {
473 seq_printf(p, "%*s%u: ", prec - 1, "IPI", i);
475 for_each_online_cpu(cpu)
476 seq_printf(p, "%10u ",
477 __get_irq_stat(cpu, ipi_irqs[i]));
479 seq_printf(p, " %s\n", ipi_types[i]);
483 u64 smp_irq_stat_cpu(unsigned int cpu)
488 for (i = 0; i < NR_IPI; i++)
489 sum += __get_irq_stat(cpu, ipi_irqs[i]);
495 * Timer (local or broadcast) support
497 static DEFINE_PER_CPU(struct clock_event_device, percpu_clockevent);
499 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
500 void tick_broadcast(const struct cpumask *mask)
502 smp_cross_call(mask, IPI_TIMER);
506 static void broadcast_timer_set_mode(enum clock_event_mode mode,
507 struct clock_event_device *evt)
511 static void __cpuinit broadcast_timer_setup(struct clock_event_device *evt)
513 evt->name = "dummy_timer";
514 evt->features = CLOCK_EVT_FEAT_ONESHOT |
515 CLOCK_EVT_FEAT_PERIODIC |
516 CLOCK_EVT_FEAT_DUMMY;
519 evt->set_mode = broadcast_timer_set_mode;
521 clockevents_register_device(evt);
524 static struct local_timer_ops *lt_ops;
526 #ifdef CONFIG_LOCAL_TIMERS
527 int local_timer_register(struct local_timer_ops *ops)
529 if (!is_smp() || !setup_max_cpus)
540 static void __cpuinit percpu_timer_setup(void)
542 unsigned int cpu = smp_processor_id();
543 struct clock_event_device *evt = &per_cpu(percpu_clockevent, cpu);
545 evt->cpumask = cpumask_of(cpu);
547 if (!lt_ops || lt_ops->setup(evt))
548 broadcast_timer_setup(evt);
551 #ifdef CONFIG_HOTPLUG_CPU
553 * The generic clock events code purposely does not stop the local timer
554 * on CPU_DEAD/CPU_DEAD_FROZEN hotplug events, so we have to do it
557 static void percpu_timer_stop(void)
559 unsigned int cpu = smp_processor_id();
560 struct clock_event_device *evt = &per_cpu(percpu_clockevent, cpu);
567 static DEFINE_RAW_SPINLOCK(stop_lock);
570 * ipi_cpu_stop - handle IPI from smp_send_stop()
572 static void ipi_cpu_stop(unsigned int cpu)
574 if (system_state == SYSTEM_BOOTING ||
575 system_state == SYSTEM_RUNNING) {
576 raw_spin_lock(&stop_lock);
577 printk(KERN_CRIT "CPU%u: stopping\n", cpu);
579 raw_spin_unlock(&stop_lock);
582 set_cpu_online(cpu, false);
592 * Main handler for inter-processor interrupts
594 asmlinkage void __exception_irq_entry do_IPI(int ipinr, struct pt_regs *regs)
596 handle_IPI(ipinr, regs);
599 void handle_IPI(int ipinr, struct pt_regs *regs)
601 unsigned int cpu = smp_processor_id();
602 struct pt_regs *old_regs = set_irq_regs(regs);
605 __inc_irq_stat(cpu, ipi_irqs[ipinr]);
611 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
614 tick_receive_broadcast();
625 generic_smp_call_function_interrupt();
629 case IPI_CALL_FUNC_SINGLE:
631 generic_smp_call_function_single_interrupt();
642 printk(KERN_CRIT "CPU%u: Unknown IPI message 0x%x\n",
646 set_irq_regs(old_regs);
649 void smp_send_reschedule(int cpu)
651 smp_cross_call(cpumask_of(cpu), IPI_RESCHEDULE);
654 void smp_send_stop(void)
656 unsigned long timeout;
659 cpumask_copy(&mask, cpu_online_mask);
660 cpumask_clear_cpu(smp_processor_id(), &mask);
661 if (!cpumask_empty(&mask))
662 smp_cross_call(&mask, IPI_CPU_STOP);
664 /* Wait up to one second for other CPUs to stop */
665 timeout = USEC_PER_SEC;
666 while (num_online_cpus() > 1 && timeout--)
669 if (num_online_cpus() > 1)
670 pr_warning("SMP: failed to stop secondary CPUs\n");
676 int setup_profiling_timer(unsigned int multiplier)
681 #ifdef CONFIG_CPU_FREQ
683 static DEFINE_PER_CPU(unsigned long, l_p_j_ref);
684 static DEFINE_PER_CPU(unsigned long, l_p_j_ref_freq);
685 static unsigned long global_l_p_j_ref;
686 static unsigned long global_l_p_j_ref_freq;
688 static int cpufreq_callback(struct notifier_block *nb,
689 unsigned long val, void *data)
691 struct cpufreq_freqs *freq = data;
694 if (freq->flags & CPUFREQ_CONST_LOOPS)
697 if (!per_cpu(l_p_j_ref, cpu)) {
698 per_cpu(l_p_j_ref, cpu) =
699 per_cpu(cpu_data, cpu).loops_per_jiffy;
700 per_cpu(l_p_j_ref_freq, cpu) = freq->old;
701 if (!global_l_p_j_ref) {
702 global_l_p_j_ref = loops_per_jiffy;
703 global_l_p_j_ref_freq = freq->old;
707 if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) ||
708 (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
709 (val == CPUFREQ_RESUMECHANGE || val == CPUFREQ_SUSPENDCHANGE)) {
710 loops_per_jiffy = cpufreq_scale(global_l_p_j_ref,
711 global_l_p_j_ref_freq,
713 per_cpu(cpu_data, cpu).loops_per_jiffy =
714 cpufreq_scale(per_cpu(l_p_j_ref, cpu),
715 per_cpu(l_p_j_ref_freq, cpu),
721 static struct notifier_block cpufreq_notifier = {
722 .notifier_call = cpufreq_callback,
725 static int __init register_cpufreq_notifier(void)
727 return cpufreq_register_notifier(&cpufreq_notifier,
728 CPUFREQ_TRANSITION_NOTIFIER);
730 core_initcall(register_cpufreq_notifier);