1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/arch/arm/kernel/smp.c
5 * Copyright (C) 2002 ARM Limited, All Rights Reserved.
7 #include <linux/module.h>
8 #include <linux/delay.h>
9 #include <linux/init.h>
10 #include <linux/spinlock.h>
11 #include <linux/sched/mm.h>
12 #include <linux/sched/hotplug.h>
13 #include <linux/sched/task_stack.h>
14 #include <linux/interrupt.h>
15 #include <linux/cache.h>
16 #include <linux/profile.h>
17 #include <linux/errno.h>
19 #include <linux/err.h>
20 #include <linux/cpu.h>
21 #include <linux/seq_file.h>
22 #include <linux/irq.h>
23 #include <linux/nmi.h>
24 #include <linux/percpu.h>
25 #include <linux/clockchips.h>
26 #include <linux/completion.h>
27 #include <linux/cpufreq.h>
28 #include <linux/irq_work.h>
29 #include <linux/kernel_stat.h>
31 #include <linux/atomic.h>
34 #include <asm/cacheflush.h>
36 #include <asm/cputype.h>
37 #include <asm/exception.h>
38 #include <asm/idmap.h>
39 #include <asm/topology.h>
40 #include <asm/mmu_context.h>
41 #include <asm/procinfo.h>
42 #include <asm/processor.h>
43 #include <asm/sections.h>
44 #include <asm/tlbflush.h>
45 #include <asm/ptrace.h>
46 #include <asm/smp_plat.h>
48 #include <asm/mach/arch.h>
51 #define CREATE_TRACE_POINTS
52 #include <trace/events/ipi.h>
55 * as from 2.5, kernels no longer have an init_tasks structure
56 * so we need some other way of telling a new secondary core
57 * where to place its SVC stack
59 struct secondary_data secondary_data;
71 * CPU_BACKTRACE is special and not included in NR_IPI
72 * or tracable with trace_ipi_*
74 IPI_CPU_BACKTRACE = NR_IPI,
76 * SGI8-15 can be reserved by secure firmware, and thus may
77 * not be usable by the kernel. Please keep the above limited
78 * to at most 8 entries.
83 static int ipi_irq_base __read_mostly;
84 static int nr_ipi __read_mostly = NR_IPI;
85 static struct irq_desc *ipi_desc[MAX_IPI] __read_mostly;
87 static void ipi_setup(int cpu);
89 static DECLARE_COMPLETION(cpu_running);
91 static struct smp_operations smp_ops __ro_after_init;
93 void __init smp_set_ops(const struct smp_operations *ops)
99 static unsigned long get_arch_pgd(pgd_t *pgd)
101 #ifdef CONFIG_ARM_LPAE
102 return __phys_to_pfn(virt_to_phys(pgd));
104 return virt_to_phys(pgd);
108 #if defined(CONFIG_BIG_LITTLE) && defined(CONFIG_HARDEN_BRANCH_PREDICTOR)
109 static int secondary_biglittle_prepare(unsigned int cpu)
111 if (!cpu_vtable[cpu])
112 cpu_vtable[cpu] = kzalloc(sizeof(*cpu_vtable[cpu]), GFP_KERNEL);
114 return cpu_vtable[cpu] ? 0 : -ENOMEM;
117 static void secondary_biglittle_init(void)
119 init_proc_vtable(lookup_processor(read_cpuid_id())->proc);
122 static int secondary_biglittle_prepare(unsigned int cpu)
127 static void secondary_biglittle_init(void)
132 int __cpu_up(unsigned int cpu, struct task_struct *idle)
136 if (!smp_ops.smp_boot_secondary)
139 ret = secondary_biglittle_prepare(cpu);
144 * We need to tell the secondary core where to find
145 * its stack and the page tables.
147 secondary_data.stack = task_stack_page(idle) + THREAD_START_SP;
148 #ifdef CONFIG_ARM_MPU
149 secondary_data.mpu_rgn_info = &mpu_rgn_info;
153 secondary_data.pgdir = virt_to_phys(idmap_pgd);
154 secondary_data.swapper_pg_dir = get_arch_pgd(swapper_pg_dir);
156 secondary_data.task = idle;
157 if (IS_ENABLED(CONFIG_THREAD_INFO_IN_TASK))
158 task_thread_info(idle)->cpu = cpu;
160 sync_cache_w(&secondary_data);
163 * Now bring the CPU into our world.
165 ret = smp_ops.smp_boot_secondary(cpu, idle);
168 * CPU was successfully started, wait for it
169 * to come online or time out.
171 wait_for_completion_timeout(&cpu_running,
172 msecs_to_jiffies(1000));
174 if (!cpu_online(cpu)) {
175 pr_crit("CPU%u: failed to come online\n", cpu);
179 pr_err("CPU%u: failed to boot: %d\n", cpu, ret);
183 memset(&secondary_data, 0, sizeof(secondary_data));
187 /* platform specific SMP operations */
188 void __init smp_init_cpus(void)
190 if (smp_ops.smp_init_cpus)
191 smp_ops.smp_init_cpus();
194 int platform_can_secondary_boot(void)
196 return !!smp_ops.smp_boot_secondary;
199 int platform_can_cpu_hotplug(void)
201 #ifdef CONFIG_HOTPLUG_CPU
202 if (smp_ops.cpu_kill)
209 #ifdef CONFIG_HOTPLUG_CPU
210 static int platform_cpu_kill(unsigned int cpu)
212 if (smp_ops.cpu_kill)
213 return smp_ops.cpu_kill(cpu);
217 static int platform_cpu_disable(unsigned int cpu)
219 if (smp_ops.cpu_disable)
220 return smp_ops.cpu_disable(cpu);
225 int platform_can_hotplug_cpu(unsigned int cpu)
227 /* cpu_die must be specified to support hotplug */
228 if (!smp_ops.cpu_die)
231 if (smp_ops.cpu_can_disable)
232 return smp_ops.cpu_can_disable(cpu);
235 * By default, allow disabling all CPUs except the first one,
236 * since this is special on a lot of platforms, e.g. because
237 * of clock tick interrupts.
242 static void ipi_teardown(int cpu)
246 if (WARN_ON_ONCE(!ipi_irq_base))
249 for (i = 0; i < nr_ipi; i++)
250 disable_percpu_irq(ipi_irq_base + i);
254 * __cpu_disable runs on the processor to be shutdown.
256 int __cpu_disable(void)
258 unsigned int cpu = smp_processor_id();
261 ret = platform_cpu_disable(cpu);
265 #ifdef CONFIG_GENERIC_ARCH_TOPOLOGY
266 remove_cpu_topology(cpu);
270 * Take this CPU offline. Once we clear this, we can't return,
271 * and we must not schedule until we're ready to give up the cpu.
273 set_cpu_online(cpu, false);
277 * OK - migrate IRQs away from this CPU
279 irq_migrate_all_off_this_cpu();
282 * Flush user cache and TLB mappings, and then remove this CPU
283 * from the vm mask set of all processes.
285 * Caches are flushed to the Level of Unification Inner Shareable
286 * to write-back dirty lines to unified caches shared by all CPUs.
289 local_flush_tlb_all();
295 * called on the thread which is asking for a CPU to be shutdown -
296 * waits until shutdown has completed, or it is timed out.
298 void __cpu_die(unsigned int cpu)
300 if (!cpu_wait_death(cpu, 5)) {
301 pr_err("CPU%u: cpu didn't die\n", cpu);
304 pr_debug("CPU%u: shutdown\n", cpu);
306 clear_tasks_mm_cpumask(cpu);
308 * platform_cpu_kill() is generally expected to do the powering off
309 * and/or cutting of clocks to the dying CPU. Optionally, this may
310 * be done by the CPU which is dying in preference to supporting
311 * this call, but that means there is _no_ synchronisation between
312 * the requesting CPU and the dying CPU actually losing power.
314 if (!platform_cpu_kill(cpu))
315 pr_err("CPU%u: unable to kill\n", cpu);
319 * Called from the idle thread for the CPU which has been shutdown.
321 * Note that we disable IRQs here, but do not re-enable them
322 * before returning to the caller. This is also the behaviour
323 * of the other hotplug-cpu capable cores, so presumably coming
324 * out of idle fixes this.
326 void arch_cpu_idle_dead(void)
328 unsigned int cpu = smp_processor_id();
335 * Flush the data out of the L1 cache for this CPU. This must be
336 * before the completion to ensure that data is safely written out
337 * before platform_cpu_kill() gets called - which may disable
338 * *this* CPU and power down its cache.
343 * Tell __cpu_die() that this CPU is now safe to dispose of. Once
344 * this returns, power and/or clocks can be removed at any point
345 * from this CPU and its cache by platform_cpu_kill().
347 (void)cpu_report_death();
350 * Ensure that the cache lines associated with that completion are
351 * written out. This covers the case where _this_ CPU is doing the
352 * powering down, to ensure that the completion is visible to the
353 * CPU waiting for this one.
358 * The actual CPU shutdown procedure is at least platform (if not
359 * CPU) specific. This may remove power, or it may simply spin.
361 * Platforms are generally expected *NOT* to return from this call,
362 * although there are some which do because they have no way to
363 * power down the CPU. These platforms are the _only_ reason we
364 * have a return path which uses the fragment of assembly below.
366 * The return path should not be used for platforms which can
370 smp_ops.cpu_die(cpu);
372 pr_warn("CPU%u: smp_ops.cpu_die() returned, trying to resuscitate\n",
376 * Do not return to the idle loop - jump back to the secondary
377 * cpu initialisation. There's some initialisation which needs
378 * to be repeated to undo the effects of taking the CPU offline.
380 __asm__("mov sp, %0\n"
383 " b secondary_start_kernel"
385 : "r" (task_stack_page(current) + THREAD_SIZE - 8),
389 #endif /* CONFIG_HOTPLUG_CPU */
392 * Called by both boot and secondaries to move global data into
393 * per-processor storage.
395 static void smp_store_cpu_info(unsigned int cpuid)
397 struct cpuinfo_arm *cpu_info = &per_cpu(cpu_data, cpuid);
399 cpu_info->loops_per_jiffy = loops_per_jiffy;
400 cpu_info->cpuid = read_cpuid_id();
402 store_cpu_topology(cpuid);
403 check_cpu_icache_size(cpuid);
407 * This is the secondary CPU boot entry. We're using this CPUs
408 * idle thread stack, but a set of temporary page tables.
410 asmlinkage void secondary_start_kernel(struct task_struct *task)
412 struct mm_struct *mm = &init_mm;
417 secondary_biglittle_init();
420 * The identity mapping is uncached (strongly ordered), so
421 * switch away from it before attempting any exclusive accesses.
423 cpu_switch_mm(mm->pgd, mm);
424 local_flush_bp_all();
425 enter_lazy_tlb(mm, current);
426 local_flush_tlb_all();
429 * All kernel threads share the same mm context; grab a
430 * reference and switch to it.
432 cpu = smp_processor_id();
434 current->active_mm = mm;
435 cpumask_set_cpu(cpu, mm_cpumask(mm));
440 setup_vectors_base();
442 pr_debug("CPU%u: Booted secondary processor\n", cpu);
444 trace_hardirqs_off();
447 * Give the platform a chance to do its own initialisation.
449 if (smp_ops.smp_secondary_init)
450 smp_ops.smp_secondary_init(cpu);
452 notify_cpu_starting(cpu);
458 smp_store_cpu_info(cpu);
461 * OK, now it's safe to let the boot CPU continue. Wait for
462 * the CPU migration code to notice that the CPU is online
463 * before we continue - which happens after __cpu_up returns.
465 set_cpu_online(cpu, true);
469 complete(&cpu_running);
476 * OK, it's off to the idle thread for us
478 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
481 void __init smp_cpus_done(unsigned int max_cpus)
484 unsigned long bogosum = 0;
486 for_each_online_cpu(cpu)
487 bogosum += per_cpu(cpu_data, cpu).loops_per_jiffy;
489 printk(KERN_INFO "SMP: Total of %d processors activated "
490 "(%lu.%02lu BogoMIPS).\n",
492 bogosum / (500000/HZ),
493 (bogosum / (5000/HZ)) % 100);
498 void __init smp_prepare_boot_cpu(void)
500 set_my_cpu_offset(per_cpu_offset(smp_processor_id()));
503 void __init smp_prepare_cpus(unsigned int max_cpus)
505 unsigned int ncores = num_possible_cpus();
509 smp_store_cpu_info(smp_processor_id());
512 * are we trying to boot more cores than exist?
514 if (max_cpus > ncores)
516 if (ncores > 1 && max_cpus) {
518 * Initialise the present map, which describes the set of CPUs
519 * actually populated at the present time. A platform should
520 * re-initialize the map in the platforms smp_prepare_cpus()
521 * if present != possible (e.g. physical hotplug).
523 init_cpu_present(cpu_possible_mask);
526 * Initialise the SCU if there are more than one CPU
527 * and let them know where to start.
529 if (smp_ops.smp_prepare_cpus)
530 smp_ops.smp_prepare_cpus(max_cpus);
534 static const char *ipi_types[NR_IPI] __tracepoint_string = {
535 [IPI_WAKEUP] = "CPU wakeup interrupts",
536 [IPI_TIMER] = "Timer broadcast interrupts",
537 [IPI_RESCHEDULE] = "Rescheduling interrupts",
538 [IPI_CALL_FUNC] = "Function call interrupts",
539 [IPI_CPU_STOP] = "CPU stop interrupts",
540 [IPI_IRQ_WORK] = "IRQ work interrupts",
541 [IPI_COMPLETION] = "completion interrupts",
544 static void smp_cross_call(const struct cpumask *target, unsigned int ipinr);
546 void show_ipi_list(struct seq_file *p, int prec)
550 for (i = 0; i < NR_IPI; i++) {
554 seq_printf(p, "%*s%u: ", prec - 1, "IPI", i);
556 for_each_online_cpu(cpu)
557 seq_printf(p, "%10u ", irq_desc_kstat_cpu(ipi_desc[i], cpu));
559 seq_printf(p, " %s\n", ipi_types[i]);
563 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
565 smp_cross_call(mask, IPI_CALL_FUNC);
568 void arch_send_wakeup_ipi_mask(const struct cpumask *mask)
570 smp_cross_call(mask, IPI_WAKEUP);
573 void arch_send_call_function_single_ipi(int cpu)
575 smp_cross_call(cpumask_of(cpu), IPI_CALL_FUNC);
578 #ifdef CONFIG_IRQ_WORK
579 void arch_irq_work_raise(void)
581 if (arch_irq_work_has_interrupt())
582 smp_cross_call(cpumask_of(smp_processor_id()), IPI_IRQ_WORK);
586 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
587 void tick_broadcast(const struct cpumask *mask)
589 smp_cross_call(mask, IPI_TIMER);
593 static DEFINE_RAW_SPINLOCK(stop_lock);
596 * ipi_cpu_stop - handle IPI from smp_send_stop()
598 static void ipi_cpu_stop(unsigned int cpu)
600 if (system_state <= SYSTEM_RUNNING) {
601 raw_spin_lock(&stop_lock);
602 pr_crit("CPU%u: stopping\n", cpu);
604 raw_spin_unlock(&stop_lock);
607 set_cpu_online(cpu, false);
618 static DEFINE_PER_CPU(struct completion *, cpu_completion);
620 int register_ipi_completion(struct completion *completion, int cpu)
622 per_cpu(cpu_completion, cpu) = completion;
623 return IPI_COMPLETION;
626 static void ipi_complete(unsigned int cpu)
628 complete(per_cpu(cpu_completion, cpu));
632 * Main handler for inter-processor interrupts
634 asmlinkage void __exception_irq_entry do_IPI(int ipinr, struct pt_regs *regs)
636 handle_IPI(ipinr, regs);
639 static void do_handle_IPI(int ipinr)
641 unsigned int cpu = smp_processor_id();
643 if ((unsigned)ipinr < NR_IPI)
644 trace_ipi_entry_rcuidle(ipi_types[ipinr]);
650 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
652 tick_receive_broadcast();
661 generic_smp_call_function_interrupt();
668 #ifdef CONFIG_IRQ_WORK
678 case IPI_CPU_BACKTRACE:
679 printk_deferred_enter();
680 nmi_cpu_backtrace(get_irq_regs());
681 printk_deferred_exit();
685 pr_crit("CPU%u: Unknown IPI message 0x%x\n",
690 if ((unsigned)ipinr < NR_IPI)
691 trace_ipi_exit_rcuidle(ipi_types[ipinr]);
694 /* Legacy version, should go away once all irqchips have been converted */
695 void handle_IPI(int ipinr, struct pt_regs *regs)
697 struct pt_regs *old_regs = set_irq_regs(regs);
700 do_handle_IPI(ipinr);
703 set_irq_regs(old_regs);
706 static irqreturn_t ipi_handler(int irq, void *data)
708 do_handle_IPI(irq - ipi_irq_base);
712 static void smp_cross_call(const struct cpumask *target, unsigned int ipinr)
714 trace_ipi_raise_rcuidle(target, ipi_types[ipinr]);
715 __ipi_send_mask(ipi_desc[ipinr], target);
718 static void ipi_setup(int cpu)
722 if (WARN_ON_ONCE(!ipi_irq_base))
725 for (i = 0; i < nr_ipi; i++)
726 enable_percpu_irq(ipi_irq_base + i, 0);
729 void __init set_smp_ipi_range(int ipi_base, int n)
733 WARN_ON(n < MAX_IPI);
734 nr_ipi = min(n, MAX_IPI);
736 for (i = 0; i < nr_ipi; i++) {
739 err = request_percpu_irq(ipi_base + i, ipi_handler,
743 ipi_desc[i] = irq_to_desc(ipi_base + i);
744 irq_set_status_flags(ipi_base + i, IRQ_HIDDEN);
747 ipi_irq_base = ipi_base;
749 /* Setup the boot CPU immediately */
750 ipi_setup(smp_processor_id());
753 void smp_send_reschedule(int cpu)
755 smp_cross_call(cpumask_of(cpu), IPI_RESCHEDULE);
758 void smp_send_stop(void)
760 unsigned long timeout;
763 cpumask_copy(&mask, cpu_online_mask);
764 cpumask_clear_cpu(smp_processor_id(), &mask);
765 if (!cpumask_empty(&mask))
766 smp_cross_call(&mask, IPI_CPU_STOP);
768 /* Wait up to one second for other CPUs to stop */
769 timeout = USEC_PER_SEC;
770 while (num_online_cpus() > 1 && timeout--)
773 if (num_online_cpus() > 1)
774 pr_warn("SMP: failed to stop secondary CPUs\n");
777 /* In case panic() and panic() called at the same time on CPU1 and CPU2,
778 * and CPU 1 calls panic_smp_self_stop() before crash_smp_send_stop()
779 * CPU1 can't receive the ipi irqs from CPU2, CPU1 will be always online,
780 * kdump fails. So split out the panic_smp_self_stop() and add
781 * set_cpu_online(smp_processor_id(), false).
783 void panic_smp_self_stop(void)
785 pr_debug("CPU %u will stop doing anything useful since another CPU has paniced\n",
787 set_cpu_online(smp_processor_id(), false);
795 int setup_profiling_timer(unsigned int multiplier)
800 #ifdef CONFIG_CPU_FREQ
802 static DEFINE_PER_CPU(unsigned long, l_p_j_ref);
803 static DEFINE_PER_CPU(unsigned long, l_p_j_ref_freq);
804 static unsigned long global_l_p_j_ref;
805 static unsigned long global_l_p_j_ref_freq;
807 static int cpufreq_callback(struct notifier_block *nb,
808 unsigned long val, void *data)
810 struct cpufreq_freqs *freq = data;
811 struct cpumask *cpus = freq->policy->cpus;
812 int cpu, first = cpumask_first(cpus);
815 if (freq->flags & CPUFREQ_CONST_LOOPS)
818 if (!per_cpu(l_p_j_ref, first)) {
819 for_each_cpu(cpu, cpus) {
820 per_cpu(l_p_j_ref, cpu) =
821 per_cpu(cpu_data, cpu).loops_per_jiffy;
822 per_cpu(l_p_j_ref_freq, cpu) = freq->old;
825 if (!global_l_p_j_ref) {
826 global_l_p_j_ref = loops_per_jiffy;
827 global_l_p_j_ref_freq = freq->old;
831 if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) ||
832 (val == CPUFREQ_POSTCHANGE && freq->old > freq->new)) {
833 loops_per_jiffy = cpufreq_scale(global_l_p_j_ref,
834 global_l_p_j_ref_freq,
837 lpj = cpufreq_scale(per_cpu(l_p_j_ref, first),
838 per_cpu(l_p_j_ref_freq, first), freq->new);
839 for_each_cpu(cpu, cpus)
840 per_cpu(cpu_data, cpu).loops_per_jiffy = lpj;
845 static struct notifier_block cpufreq_notifier = {
846 .notifier_call = cpufreq_callback,
849 static int __init register_cpufreq_notifier(void)
851 return cpufreq_register_notifier(&cpufreq_notifier,
852 CPUFREQ_TRANSITION_NOTIFIER);
854 core_initcall(register_cpufreq_notifier);
858 static void raise_nmi(cpumask_t *mask)
860 __ipi_send_mask(ipi_desc[IPI_CPU_BACKTRACE], mask);
863 void arch_trigger_cpumask_backtrace(const cpumask_t *mask, bool exclude_self)
865 nmi_trigger_cpumask_backtrace(mask, exclude_self, raise_nmi);