4 * Written by Cort Dougan (cort@cs.nmt.edu) borrowing a great
5 * deal of code from the sparc and intel versions.
7 * Copyright (C) 1999 Cort Dougan <cort@cs.nmt.edu>
9 * PowerPC-64 Support added by Dave Engebretsen, Peter Bergner, and
10 * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
12 * This program is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU General Public License
14 * as published by the Free Software Foundation; either version
15 * 2 of the License, or (at your option) any later version.
20 #include <linux/kernel.h>
21 #include <linux/export.h>
22 #include <linux/sched.h>
23 #include <linux/smp.h>
24 #include <linux/interrupt.h>
25 #include <linux/delay.h>
26 #include <linux/init.h>
27 #include <linux/spinlock.h>
28 #include <linux/cache.h>
29 #include <linux/err.h>
30 #include <linux/device.h>
31 #include <linux/cpu.h>
32 #include <linux/notifier.h>
33 #include <linux/topology.h>
35 #include <asm/ptrace.h>
36 #include <linux/atomic.h>
39 #include <asm/pgtable.h>
43 #include <asm/machdep.h>
44 #include <asm/cputhreads.h>
45 #include <asm/cputable.h>
47 #include <asm/vdso_datapage.h>
52 #include <asm/debug.h>
56 #define DBG(fmt...) udbg_printf(fmt)
61 #ifdef CONFIG_HOTPLUG_CPU
62 /* State of each CPU during hotplug phases */
63 static DEFINE_PER_CPU(int, cpu_state) = { 0 };
66 struct thread_info *secondary_ti;
68 DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
69 DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
71 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
72 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
74 /* SMP operations for this machine */
75 struct smp_ops_t *smp_ops;
77 /* Can't be static due to PowerMac hackery */
78 volatile unsigned int cpu_callin_map[NR_CPUS];
80 int smt_enabled_at_boot = 1;
82 static void (*crash_ipi_function_ptr)(struct pt_regs *) = NULL;
85 int smp_generic_kick_cpu(int nr)
87 BUG_ON(nr < 0 || nr >= NR_CPUS);
90 * The processor is currently spinning, waiting for the
91 * cpu_start field to become non-zero After we set cpu_start,
92 * the processor will continue on to secondary_start
94 if (!paca[nr].cpu_start) {
95 paca[nr].cpu_start = 1;
100 #ifdef CONFIG_HOTPLUG_CPU
102 * Ok it's not there, so it might be soft-unplugged, let's
103 * try to bring it back
105 generic_set_cpu_up(nr);
107 smp_send_reschedule(nr);
108 #endif /* CONFIG_HOTPLUG_CPU */
112 #endif /* CONFIG_PPC64 */
114 static irqreturn_t call_function_action(int irq, void *data)
116 generic_smp_call_function_interrupt();
120 static irqreturn_t reschedule_action(int irq, void *data)
126 static irqreturn_t call_function_single_action(int irq, void *data)
128 generic_smp_call_function_single_interrupt();
132 static irqreturn_t debug_ipi_action(int irq, void *data)
134 if (crash_ipi_function_ptr) {
135 crash_ipi_function_ptr(get_irq_regs());
139 #ifdef CONFIG_DEBUGGER
140 debugger_ipi(get_irq_regs());
141 #endif /* CONFIG_DEBUGGER */
146 static irq_handler_t smp_ipi_action[] = {
147 [PPC_MSG_CALL_FUNCTION] = call_function_action,
148 [PPC_MSG_RESCHEDULE] = reschedule_action,
149 [PPC_MSG_CALL_FUNC_SINGLE] = call_function_single_action,
150 [PPC_MSG_DEBUGGER_BREAK] = debug_ipi_action,
153 const char *smp_ipi_name[] = {
154 [PPC_MSG_CALL_FUNCTION] = "ipi call function",
155 [PPC_MSG_RESCHEDULE] = "ipi reschedule",
156 [PPC_MSG_CALL_FUNC_SINGLE] = "ipi call function single",
157 [PPC_MSG_DEBUGGER_BREAK] = "ipi debugger",
160 /* optional function to request ipi, for controllers with >= 4 ipis */
161 int smp_request_message_ipi(int virq, int msg)
165 if (msg < 0 || msg > PPC_MSG_DEBUGGER_BREAK) {
168 #if !defined(CONFIG_DEBUGGER) && !defined(CONFIG_KEXEC)
169 if (msg == PPC_MSG_DEBUGGER_BREAK) {
173 err = request_irq(virq, smp_ipi_action[msg],
174 IRQF_PERCPU | IRQF_NO_THREAD | IRQF_NO_SUSPEND,
175 smp_ipi_name[msg], 0);
176 WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n",
177 virq, smp_ipi_name[msg], err);
182 #ifdef CONFIG_PPC_SMP_MUXED_IPI
183 struct cpu_messages {
184 int messages; /* current messages */
185 unsigned long data; /* data for cause ipi */
187 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages, ipi_message);
189 void smp_muxed_ipi_set_data(int cpu, unsigned long data)
191 struct cpu_messages *info = &per_cpu(ipi_message, cpu);
196 void smp_muxed_ipi_message_pass(int cpu, int msg)
198 struct cpu_messages *info = &per_cpu(ipi_message, cpu);
199 char *message = (char *)&info->messages;
202 * Order previous accesses before accesses in the IPI handler.
207 * cause_ipi functions are required to include a full barrier
208 * before doing whatever causes the IPI.
210 smp_ops->cause_ipi(cpu, info->data);
213 irqreturn_t smp_ipi_demux(void)
215 struct cpu_messages *info = &__get_cpu_var(ipi_message);
218 mb(); /* order any irq clear */
221 all = xchg(&info->messages, 0);
224 if (all & (1 << (24 - 8 * PPC_MSG_CALL_FUNCTION)))
225 generic_smp_call_function_interrupt();
226 if (all & (1 << (24 - 8 * PPC_MSG_RESCHEDULE)))
228 if (all & (1 << (24 - 8 * PPC_MSG_CALL_FUNC_SINGLE)))
229 generic_smp_call_function_single_interrupt();
230 if (all & (1 << (24 - 8 * PPC_MSG_DEBUGGER_BREAK)))
231 debug_ipi_action(0, NULL);
233 #error Unsupported ENDIAN
235 } while (info->messages);
239 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
241 static inline void do_message_pass(int cpu, int msg)
243 if (smp_ops->message_pass)
244 smp_ops->message_pass(cpu, msg);
245 #ifdef CONFIG_PPC_SMP_MUXED_IPI
247 smp_muxed_ipi_message_pass(cpu, msg);
251 void smp_send_reschedule(int cpu)
254 do_message_pass(cpu, PPC_MSG_RESCHEDULE);
256 EXPORT_SYMBOL_GPL(smp_send_reschedule);
258 void arch_send_call_function_single_ipi(int cpu)
260 do_message_pass(cpu, PPC_MSG_CALL_FUNC_SINGLE);
263 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
267 for_each_cpu(cpu, mask)
268 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
271 #if defined(CONFIG_DEBUGGER) || defined(CONFIG_KEXEC)
272 void smp_send_debugger_break(void)
275 int me = raw_smp_processor_id();
277 if (unlikely(!smp_ops))
280 for_each_online_cpu(cpu)
282 do_message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
287 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
289 crash_ipi_function_ptr = crash_ipi_callback;
290 if (crash_ipi_callback) {
292 smp_send_debugger_break();
297 static void stop_this_cpu(void *dummy)
299 /* Remove this CPU */
300 set_cpu_online(smp_processor_id(), false);
307 void smp_send_stop(void)
309 smp_call_function(stop_this_cpu, NULL, 0);
312 struct thread_info *current_set[NR_CPUS];
314 static void smp_store_cpu_info(int id)
316 per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
317 #ifdef CONFIG_PPC_FSL_BOOK3E
318 per_cpu(next_tlbcam_idx, id)
319 = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) - 1;
323 void __init smp_prepare_cpus(unsigned int max_cpus)
327 DBG("smp_prepare_cpus\n");
330 * setup_cpu may need to be called on the boot cpu. We havent
331 * spun any cpus up but lets be paranoid.
333 BUG_ON(boot_cpuid != smp_processor_id());
336 smp_store_cpu_info(boot_cpuid);
337 cpu_callin_map[boot_cpuid] = 1;
339 for_each_possible_cpu(cpu) {
340 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
341 GFP_KERNEL, cpu_to_node(cpu));
342 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
343 GFP_KERNEL, cpu_to_node(cpu));
346 cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
347 cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
351 max_cpus = smp_ops->probe();
358 void smp_prepare_boot_cpu(void)
360 BUG_ON(smp_processor_id() != boot_cpuid);
362 paca[boot_cpuid].__current = current;
364 current_set[boot_cpuid] = task_thread_info(current);
367 #ifdef CONFIG_HOTPLUG_CPU
369 int generic_cpu_disable(void)
371 unsigned int cpu = smp_processor_id();
373 if (cpu == boot_cpuid)
376 set_cpu_online(cpu, false);
378 vdso_data->processorCount--;
384 void generic_cpu_die(unsigned int cpu)
388 for (i = 0; i < 100; i++) {
390 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
394 printk(KERN_ERR "CPU%d didn't die...\n", cpu);
397 void generic_mach_cpu_die(void)
403 cpu = smp_processor_id();
404 printk(KERN_DEBUG "CPU%d offline\n", cpu);
405 __get_cpu_var(cpu_state) = CPU_DEAD;
407 while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
411 void generic_set_cpu_dead(unsigned int cpu)
413 per_cpu(cpu_state, cpu) = CPU_DEAD;
417 * The cpu_state should be set to CPU_UP_PREPARE in kick_cpu(), otherwise
418 * the cpu_state is always CPU_DEAD after calling generic_set_cpu_dead(),
419 * which makes the delay in generic_cpu_die() not happen.
421 void generic_set_cpu_up(unsigned int cpu)
423 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
426 int generic_check_cpu_restart(unsigned int cpu)
428 return per_cpu(cpu_state, cpu) == CPU_UP_PREPARE;
431 static atomic_t secondary_inhibit_count;
434 * Don't allow secondary CPU threads to come online
436 void inhibit_secondary_onlining(void)
439 * This makes secondary_inhibit_count stable during cpu
440 * online/offline operations.
444 atomic_inc(&secondary_inhibit_count);
447 EXPORT_SYMBOL_GPL(inhibit_secondary_onlining);
450 * Allow secondary CPU threads to come online again
452 void uninhibit_secondary_onlining(void)
455 atomic_dec(&secondary_inhibit_count);
458 EXPORT_SYMBOL_GPL(uninhibit_secondary_onlining);
460 static int secondaries_inhibited(void)
462 return atomic_read(&secondary_inhibit_count);
465 #else /* HOTPLUG_CPU */
467 #define secondaries_inhibited() 0
471 static void cpu_idle_thread_init(unsigned int cpu, struct task_struct *idle)
473 struct thread_info *ti = task_thread_info(idle);
476 paca[cpu].__current = idle;
477 paca[cpu].kstack = (unsigned long)ti + THREAD_SIZE - STACK_FRAME_OVERHEAD;
480 secondary_ti = current_set[cpu] = ti;
483 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
488 * Don't allow secondary threads to come online if inhibited
490 if (threads_per_core > 1 && secondaries_inhibited() &&
491 cpu % threads_per_core != 0)
494 if (smp_ops == NULL ||
495 (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
498 cpu_idle_thread_init(cpu, tidle);
500 /* Make sure callin-map entry is 0 (can be leftover a CPU
503 cpu_callin_map[cpu] = 0;
505 /* The information for processor bringup must
506 * be written out to main store before we release
512 DBG("smp: kicking cpu %d\n", cpu);
513 rc = smp_ops->kick_cpu(cpu);
515 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
520 * wait to see if the cpu made a callin (is actually up).
521 * use this value that I found through experimentation.
524 if (system_state < SYSTEM_RUNNING)
525 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
527 #ifdef CONFIG_HOTPLUG_CPU
530 * CPUs can take much longer to come up in the
531 * hotplug case. Wait five seconds.
533 for (c = 5000; c && !cpu_callin_map[cpu]; c--)
537 if (!cpu_callin_map[cpu]) {
538 printk(KERN_ERR "Processor %u is stuck.\n", cpu);
542 DBG("Processor %u found.\n", cpu);
544 if (smp_ops->give_timebase)
545 smp_ops->give_timebase();
547 /* Wait until cpu puts itself in the online map */
548 while (!cpu_online(cpu))
554 /* Return the value of the reg property corresponding to the given
557 int cpu_to_core_id(int cpu)
559 struct device_node *np;
563 np = of_get_cpu_node(cpu, NULL);
567 reg = of_get_property(np, "reg", NULL);
577 /* Helper routines for cpu to core mapping */
578 int cpu_core_index_of_thread(int cpu)
580 return cpu >> threads_shift;
582 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
584 int cpu_first_thread_of_core(int core)
586 return core << threads_shift;
588 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
590 /* Must be called when no change can occur to cpu_present_mask,
591 * i.e. during cpu online or offline.
593 static struct device_node *cpu_to_l2cache(int cpu)
595 struct device_node *np;
596 struct device_node *cache;
598 if (!cpu_present(cpu))
601 np = of_get_cpu_node(cpu, NULL);
605 cache = of_find_next_cache_node(np);
612 /* Activate a secondary processor. */
613 void start_secondary(void *unused)
615 unsigned int cpu = smp_processor_id();
616 struct device_node *l2_cache;
619 atomic_inc(&init_mm.mm_count);
620 current->active_mm = &init_mm;
622 smp_store_cpu_info(cpu);
623 set_dec(tb_ticks_per_jiffy);
625 cpu_callin_map[cpu] = 1;
627 if (smp_ops->setup_cpu)
628 smp_ops->setup_cpu(cpu);
629 if (smp_ops->take_timebase)
630 smp_ops->take_timebase();
632 secondary_cpu_time_init();
635 if (system_state == SYSTEM_RUNNING)
636 vdso_data->processorCount++;
640 /* Update sibling maps */
641 base = cpu_first_thread_sibling(cpu);
642 for (i = 0; i < threads_per_core; i++) {
643 if (cpu_is_offline(base + i) && (cpu != base + i))
645 cpumask_set_cpu(cpu, cpu_sibling_mask(base + i));
646 cpumask_set_cpu(base + i, cpu_sibling_mask(cpu));
648 /* cpu_core_map should be a superset of
649 * cpu_sibling_map even if we don't have cache
650 * information, so update the former here, too.
652 cpumask_set_cpu(cpu, cpu_core_mask(base + i));
653 cpumask_set_cpu(base + i, cpu_core_mask(cpu));
655 l2_cache = cpu_to_l2cache(cpu);
656 for_each_online_cpu(i) {
657 struct device_node *np = cpu_to_l2cache(i);
660 if (np == l2_cache) {
661 cpumask_set_cpu(cpu, cpu_core_mask(i));
662 cpumask_set_cpu(i, cpu_core_mask(cpu));
666 of_node_put(l2_cache);
669 notify_cpu_starting(cpu);
670 set_cpu_online(cpu, true);
674 cpu_startup_entry(CPUHP_ONLINE);
679 int setup_profiling_timer(unsigned int multiplier)
684 void __init smp_cpus_done(unsigned int max_cpus)
686 cpumask_var_t old_mask;
688 /* We want the setup_cpu() here to be called from CPU 0, but our
689 * init thread may have been "borrowed" by another CPU in the meantime
690 * se we pin us down to CPU 0 for a short while
692 alloc_cpumask_var(&old_mask, GFP_NOWAIT);
693 cpumask_copy(old_mask, tsk_cpus_allowed(current));
694 set_cpus_allowed_ptr(current, cpumask_of(boot_cpuid));
696 if (smp_ops && smp_ops->setup_cpu)
697 smp_ops->setup_cpu(boot_cpuid);
699 set_cpus_allowed_ptr(current, old_mask);
701 free_cpumask_var(old_mask);
703 if (smp_ops && smp_ops->bringup_done)
704 smp_ops->bringup_done();
706 dump_numa_cpu_topology();
710 int arch_sd_sibling_asym_packing(void)
712 if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
713 printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
714 return SD_ASYM_PACKING;
719 #ifdef CONFIG_HOTPLUG_CPU
720 int __cpu_disable(void)
722 struct device_node *l2_cache;
723 int cpu = smp_processor_id();
727 if (!smp_ops->cpu_disable)
730 err = smp_ops->cpu_disable();
734 /* Update sibling maps */
735 base = cpu_first_thread_sibling(cpu);
736 for (i = 0; i < threads_per_core; i++) {
737 cpumask_clear_cpu(cpu, cpu_sibling_mask(base + i));
738 cpumask_clear_cpu(base + i, cpu_sibling_mask(cpu));
739 cpumask_clear_cpu(cpu, cpu_core_mask(base + i));
740 cpumask_clear_cpu(base + i, cpu_core_mask(cpu));
743 l2_cache = cpu_to_l2cache(cpu);
744 for_each_present_cpu(i) {
745 struct device_node *np = cpu_to_l2cache(i);
748 if (np == l2_cache) {
749 cpumask_clear_cpu(cpu, cpu_core_mask(i));
750 cpumask_clear_cpu(i, cpu_core_mask(cpu));
754 of_node_put(l2_cache);
760 void __cpu_die(unsigned int cpu)
762 if (smp_ops->cpu_die)
763 smp_ops->cpu_die(cpu);
766 static DEFINE_MUTEX(powerpc_cpu_hotplug_driver_mutex);
768 void cpu_hotplug_driver_lock()
770 mutex_lock(&powerpc_cpu_hotplug_driver_mutex);
773 void cpu_hotplug_driver_unlock()
775 mutex_unlock(&powerpc_cpu_hotplug_driver_mutex);
783 /* If we return, we re-enter start_secondary */
784 start_secondary_resume();