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 sync_cache_w(&secondary_data);
159 * Now bring the CPU into our world.
161 ret = smp_ops.smp_boot_secondary(cpu, idle);
164 * CPU was successfully started, wait for it
165 * to come online or time out.
167 wait_for_completion_timeout(&cpu_running,
168 msecs_to_jiffies(1000));
170 if (!cpu_online(cpu)) {
171 pr_crit("CPU%u: failed to come online\n", cpu);
175 pr_err("CPU%u: failed to boot: %d\n", cpu, ret);
179 memset(&secondary_data, 0, sizeof(secondary_data));
183 /* platform specific SMP operations */
184 void __init smp_init_cpus(void)
186 if (smp_ops.smp_init_cpus)
187 smp_ops.smp_init_cpus();
190 int platform_can_secondary_boot(void)
192 return !!smp_ops.smp_boot_secondary;
195 int platform_can_cpu_hotplug(void)
197 #ifdef CONFIG_HOTPLUG_CPU
198 if (smp_ops.cpu_kill)
205 #ifdef CONFIG_HOTPLUG_CPU
206 static int platform_cpu_kill(unsigned int cpu)
208 if (smp_ops.cpu_kill)
209 return smp_ops.cpu_kill(cpu);
213 static int platform_cpu_disable(unsigned int cpu)
215 if (smp_ops.cpu_disable)
216 return smp_ops.cpu_disable(cpu);
221 int platform_can_hotplug_cpu(unsigned int cpu)
223 /* cpu_die must be specified to support hotplug */
224 if (!smp_ops.cpu_die)
227 if (smp_ops.cpu_can_disable)
228 return smp_ops.cpu_can_disable(cpu);
231 * By default, allow disabling all CPUs except the first one,
232 * since this is special on a lot of platforms, e.g. because
233 * of clock tick interrupts.
238 static void ipi_teardown(int cpu)
242 if (WARN_ON_ONCE(!ipi_irq_base))
245 for (i = 0; i < nr_ipi; i++)
246 disable_percpu_irq(ipi_irq_base + i);
250 * __cpu_disable runs on the processor to be shutdown.
252 int __cpu_disable(void)
254 unsigned int cpu = smp_processor_id();
257 ret = platform_cpu_disable(cpu);
261 #ifdef CONFIG_GENERIC_ARCH_TOPOLOGY
262 remove_cpu_topology(cpu);
266 * Take this CPU offline. Once we clear this, we can't return,
267 * and we must not schedule until we're ready to give up the cpu.
269 set_cpu_online(cpu, false);
273 * OK - migrate IRQs away from this CPU
275 irq_migrate_all_off_this_cpu();
278 * Flush user cache and TLB mappings, and then remove this CPU
279 * from the vm mask set of all processes.
281 * Caches are flushed to the Level of Unification Inner Shareable
282 * to write-back dirty lines to unified caches shared by all CPUs.
285 local_flush_tlb_all();
291 * called on the thread which is asking for a CPU to be shutdown -
292 * waits until shutdown has completed, or it is timed out.
294 void __cpu_die(unsigned int cpu)
296 if (!cpu_wait_death(cpu, 5)) {
297 pr_err("CPU%u: cpu didn't die\n", cpu);
300 pr_debug("CPU%u: shutdown\n", cpu);
302 clear_tasks_mm_cpumask(cpu);
304 * platform_cpu_kill() is generally expected to do the powering off
305 * and/or cutting of clocks to the dying CPU. Optionally, this may
306 * be done by the CPU which is dying in preference to supporting
307 * this call, but that means there is _no_ synchronisation between
308 * the requesting CPU and the dying CPU actually losing power.
310 if (!platform_cpu_kill(cpu))
311 pr_err("CPU%u: unable to kill\n", cpu);
315 * Called from the idle thread for the CPU which has been shutdown.
317 * Note that we disable IRQs here, but do not re-enable them
318 * before returning to the caller. This is also the behaviour
319 * of the other hotplug-cpu capable cores, so presumably coming
320 * out of idle fixes this.
322 void arch_cpu_idle_dead(void)
324 unsigned int cpu = smp_processor_id();
331 * Flush the data out of the L1 cache for this CPU. This must be
332 * before the completion to ensure that data is safely written out
333 * before platform_cpu_kill() gets called - which may disable
334 * *this* CPU and power down its cache.
339 * Tell __cpu_die() that this CPU is now safe to dispose of. Once
340 * this returns, power and/or clocks can be removed at any point
341 * from this CPU and its cache by platform_cpu_kill().
343 (void)cpu_report_death();
346 * Ensure that the cache lines associated with that completion are
347 * written out. This covers the case where _this_ CPU is doing the
348 * powering down, to ensure that the completion is visible to the
349 * CPU waiting for this one.
354 * The actual CPU shutdown procedure is at least platform (if not
355 * CPU) specific. This may remove power, or it may simply spin.
357 * Platforms are generally expected *NOT* to return from this call,
358 * although there are some which do because they have no way to
359 * power down the CPU. These platforms are the _only_ reason we
360 * have a return path which uses the fragment of assembly below.
362 * The return path should not be used for platforms which can
366 smp_ops.cpu_die(cpu);
368 pr_warn("CPU%u: smp_ops.cpu_die() returned, trying to resuscitate\n",
372 * Do not return to the idle loop - jump back to the secondary
373 * cpu initialisation. There's some initialisation which needs
374 * to be repeated to undo the effects of taking the CPU offline.
376 __asm__("mov sp, %0\n"
378 " b secondary_start_kernel"
380 : "r" (task_stack_page(current) + THREAD_SIZE - 8));
382 #endif /* CONFIG_HOTPLUG_CPU */
385 * Called by both boot and secondaries to move global data into
386 * per-processor storage.
388 static void smp_store_cpu_info(unsigned int cpuid)
390 struct cpuinfo_arm *cpu_info = &per_cpu(cpu_data, cpuid);
392 cpu_info->loops_per_jiffy = loops_per_jiffy;
393 cpu_info->cpuid = read_cpuid_id();
395 store_cpu_topology(cpuid);
396 check_cpu_icache_size(cpuid);
400 * This is the secondary CPU boot entry. We're using this CPUs
401 * idle thread stack, but a set of temporary page tables.
403 asmlinkage void secondary_start_kernel(void)
405 struct mm_struct *mm = &init_mm;
408 secondary_biglittle_init();
411 * The identity mapping is uncached (strongly ordered), so
412 * switch away from it before attempting any exclusive accesses.
414 cpu_switch_mm(mm->pgd, mm);
415 local_flush_bp_all();
416 enter_lazy_tlb(mm, current);
417 local_flush_tlb_all();
420 * All kernel threads share the same mm context; grab a
421 * reference and switch to it.
423 cpu = smp_processor_id();
425 current->active_mm = mm;
426 cpumask_set_cpu(cpu, mm_cpumask(mm));
431 setup_vectors_base();
433 pr_debug("CPU%u: Booted secondary processor\n", cpu);
435 trace_hardirqs_off();
438 * Give the platform a chance to do its own initialisation.
440 if (smp_ops.smp_secondary_init)
441 smp_ops.smp_secondary_init(cpu);
443 notify_cpu_starting(cpu);
449 smp_store_cpu_info(cpu);
452 * OK, now it's safe to let the boot CPU continue. Wait for
453 * the CPU migration code to notice that the CPU is online
454 * before we continue - which happens after __cpu_up returns.
456 set_cpu_online(cpu, true);
460 complete(&cpu_running);
467 * OK, it's off to the idle thread for us
469 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
472 void __init smp_cpus_done(unsigned int max_cpus)
475 unsigned long bogosum = 0;
477 for_each_online_cpu(cpu)
478 bogosum += per_cpu(cpu_data, cpu).loops_per_jiffy;
480 printk(KERN_INFO "SMP: Total of %d processors activated "
481 "(%lu.%02lu BogoMIPS).\n",
483 bogosum / (500000/HZ),
484 (bogosum / (5000/HZ)) % 100);
489 void __init smp_prepare_boot_cpu(void)
491 set_my_cpu_offset(per_cpu_offset(smp_processor_id()));
494 void __init smp_prepare_cpus(unsigned int max_cpus)
496 unsigned int ncores = num_possible_cpus();
500 smp_store_cpu_info(smp_processor_id());
503 * are we trying to boot more cores than exist?
505 if (max_cpus > ncores)
507 if (ncores > 1 && max_cpus) {
509 * Initialise the present map, which describes the set of CPUs
510 * actually populated at the present time. A platform should
511 * re-initialize the map in the platforms smp_prepare_cpus()
512 * if present != possible (e.g. physical hotplug).
514 init_cpu_present(cpu_possible_mask);
517 * Initialise the SCU if there are more than one CPU
518 * and let them know where to start.
520 if (smp_ops.smp_prepare_cpus)
521 smp_ops.smp_prepare_cpus(max_cpus);
525 static const char *ipi_types[NR_IPI] __tracepoint_string = {
526 #define S(x,s) [x] = s
527 S(IPI_WAKEUP, "CPU wakeup interrupts"),
528 S(IPI_TIMER, "Timer broadcast interrupts"),
529 S(IPI_RESCHEDULE, "Rescheduling interrupts"),
530 S(IPI_CALL_FUNC, "Function call interrupts"),
531 S(IPI_CPU_STOP, "CPU stop interrupts"),
532 S(IPI_IRQ_WORK, "IRQ work interrupts"),
533 S(IPI_COMPLETION, "completion interrupts"),
536 static void smp_cross_call(const struct cpumask *target, unsigned int ipinr);
538 void show_ipi_list(struct seq_file *p, int prec)
542 for (i = 0; i < NR_IPI; i++) {
548 irq = irq_desc_get_irq(ipi_desc[i]);
549 seq_printf(p, "%*s%u: ", prec - 1, "IPI", i);
551 for_each_online_cpu(cpu)
552 seq_printf(p, "%10u ", kstat_irqs_cpu(irq, cpu));
554 seq_printf(p, " %s\n", ipi_types[i]);
558 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
560 smp_cross_call(mask, IPI_CALL_FUNC);
563 void arch_send_wakeup_ipi_mask(const struct cpumask *mask)
565 smp_cross_call(mask, IPI_WAKEUP);
568 void arch_send_call_function_single_ipi(int cpu)
570 smp_cross_call(cpumask_of(cpu), IPI_CALL_FUNC);
573 #ifdef CONFIG_IRQ_WORK
574 void arch_irq_work_raise(void)
576 if (arch_irq_work_has_interrupt())
577 smp_cross_call(cpumask_of(smp_processor_id()), IPI_IRQ_WORK);
581 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
582 void tick_broadcast(const struct cpumask *mask)
584 smp_cross_call(mask, IPI_TIMER);
588 static DEFINE_RAW_SPINLOCK(stop_lock);
591 * ipi_cpu_stop - handle IPI from smp_send_stop()
593 static void ipi_cpu_stop(unsigned int cpu)
595 if (system_state <= SYSTEM_RUNNING) {
596 raw_spin_lock(&stop_lock);
597 pr_crit("CPU%u: stopping\n", cpu);
599 raw_spin_unlock(&stop_lock);
602 set_cpu_online(cpu, false);
613 static DEFINE_PER_CPU(struct completion *, cpu_completion);
615 int register_ipi_completion(struct completion *completion, int cpu)
617 per_cpu(cpu_completion, cpu) = completion;
618 return IPI_COMPLETION;
621 static void ipi_complete(unsigned int cpu)
623 complete(per_cpu(cpu_completion, cpu));
627 * Main handler for inter-processor interrupts
629 asmlinkage void __exception_irq_entry do_IPI(int ipinr, struct pt_regs *regs)
631 handle_IPI(ipinr, regs);
634 static void do_handle_IPI(int ipinr)
636 unsigned int cpu = smp_processor_id();
638 if ((unsigned)ipinr < NR_IPI)
639 trace_ipi_entry_rcuidle(ipi_types[ipinr]);
645 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
647 tick_receive_broadcast();
656 generic_smp_call_function_interrupt();
663 #ifdef CONFIG_IRQ_WORK
673 case IPI_CPU_BACKTRACE:
675 nmi_cpu_backtrace(get_irq_regs());
680 pr_crit("CPU%u: Unknown IPI message 0x%x\n",
685 if ((unsigned)ipinr < NR_IPI)
686 trace_ipi_exit_rcuidle(ipi_types[ipinr]);
689 /* Legacy version, should go away once all irqchips have been converted */
690 void handle_IPI(int ipinr, struct pt_regs *regs)
692 struct pt_regs *old_regs = set_irq_regs(regs);
695 do_handle_IPI(ipinr);
698 set_irq_regs(old_regs);
701 static irqreturn_t ipi_handler(int irq, void *data)
703 do_handle_IPI(irq - ipi_irq_base);
707 static void smp_cross_call(const struct cpumask *target, unsigned int ipinr)
709 trace_ipi_raise_rcuidle(target, ipi_types[ipinr]);
710 __ipi_send_mask(ipi_desc[ipinr], target);
713 static void ipi_setup(int cpu)
717 if (WARN_ON_ONCE(!ipi_irq_base))
720 for (i = 0; i < nr_ipi; i++)
721 enable_percpu_irq(ipi_irq_base + i, 0);
724 void __init set_smp_ipi_range(int ipi_base, int n)
728 WARN_ON(n < MAX_IPI);
729 nr_ipi = min(n, MAX_IPI);
731 for (i = 0; i < nr_ipi; i++) {
734 err = request_percpu_irq(ipi_base + i, ipi_handler,
738 ipi_desc[i] = irq_to_desc(ipi_base + i);
739 irq_set_status_flags(ipi_base + i, IRQ_HIDDEN);
742 ipi_irq_base = ipi_base;
744 /* Setup the boot CPU immediately */
745 ipi_setup(smp_processor_id());
748 void smp_send_reschedule(int cpu)
750 smp_cross_call(cpumask_of(cpu), IPI_RESCHEDULE);
753 void smp_send_stop(void)
755 unsigned long timeout;
758 cpumask_copy(&mask, cpu_online_mask);
759 cpumask_clear_cpu(smp_processor_id(), &mask);
760 if (!cpumask_empty(&mask))
761 smp_cross_call(&mask, IPI_CPU_STOP);
763 /* Wait up to one second for other CPUs to stop */
764 timeout = USEC_PER_SEC;
765 while (num_online_cpus() > 1 && timeout--)
768 if (num_online_cpus() > 1)
769 pr_warn("SMP: failed to stop secondary CPUs\n");
772 /* In case panic() and panic() called at the same time on CPU1 and CPU2,
773 * and CPU 1 calls panic_smp_self_stop() before crash_smp_send_stop()
774 * CPU1 can't receive the ipi irqs from CPU2, CPU1 will be always online,
775 * kdump fails. So split out the panic_smp_self_stop() and add
776 * set_cpu_online(smp_processor_id(), false).
778 void panic_smp_self_stop(void)
780 pr_debug("CPU %u will stop doing anything useful since another CPU has paniced\n",
782 set_cpu_online(smp_processor_id(), false);
790 int setup_profiling_timer(unsigned int multiplier)
795 #ifdef CONFIG_CPU_FREQ
797 static DEFINE_PER_CPU(unsigned long, l_p_j_ref);
798 static DEFINE_PER_CPU(unsigned long, l_p_j_ref_freq);
799 static unsigned long global_l_p_j_ref;
800 static unsigned long global_l_p_j_ref_freq;
802 static int cpufreq_callback(struct notifier_block *nb,
803 unsigned long val, void *data)
805 struct cpufreq_freqs *freq = data;
806 struct cpumask *cpus = freq->policy->cpus;
807 int cpu, first = cpumask_first(cpus);
810 if (freq->flags & CPUFREQ_CONST_LOOPS)
813 if (!per_cpu(l_p_j_ref, first)) {
814 for_each_cpu(cpu, cpus) {
815 per_cpu(l_p_j_ref, cpu) =
816 per_cpu(cpu_data, cpu).loops_per_jiffy;
817 per_cpu(l_p_j_ref_freq, cpu) = freq->old;
820 if (!global_l_p_j_ref) {
821 global_l_p_j_ref = loops_per_jiffy;
822 global_l_p_j_ref_freq = freq->old;
826 if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) ||
827 (val == CPUFREQ_POSTCHANGE && freq->old > freq->new)) {
828 loops_per_jiffy = cpufreq_scale(global_l_p_j_ref,
829 global_l_p_j_ref_freq,
832 lpj = cpufreq_scale(per_cpu(l_p_j_ref, first),
833 per_cpu(l_p_j_ref_freq, first), freq->new);
834 for_each_cpu(cpu, cpus)
835 per_cpu(cpu_data, cpu).loops_per_jiffy = lpj;
840 static struct notifier_block cpufreq_notifier = {
841 .notifier_call = cpufreq_callback,
844 static int __init register_cpufreq_notifier(void)
846 return cpufreq_register_notifier(&cpufreq_notifier,
847 CPUFREQ_TRANSITION_NOTIFIER);
849 core_initcall(register_cpufreq_notifier);
853 static void raise_nmi(cpumask_t *mask)
855 __ipi_send_mask(ipi_desc[IPI_CPU_BACKTRACE], mask);
858 void arch_trigger_cpumask_backtrace(const cpumask_t *mask, bool exclude_self)
860 nmi_trigger_cpumask_backtrace(mask, exclude_self, raise_nmi);