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 * Returns 1 if the specified cpu should be brought up during boot.
86 * Used to inhibit booting threads if they've been disabled or
87 * limited on the command line
89 int smp_generic_cpu_bootable(unsigned int nr)
91 /* Special case - we inhibit secondary thread startup
92 * during boot if the user requests it.
94 if (system_state == SYSTEM_BOOTING && cpu_has_feature(CPU_FTR_SMT)) {
95 if (!smt_enabled_at_boot && cpu_thread_in_core(nr) != 0)
97 if (smt_enabled_at_boot
98 && cpu_thread_in_core(nr) >= smt_enabled_at_boot)
107 int smp_generic_kick_cpu(int nr)
109 BUG_ON(nr < 0 || nr >= NR_CPUS);
112 * The processor is currently spinning, waiting for the
113 * cpu_start field to become non-zero After we set cpu_start,
114 * the processor will continue on to secondary_start
116 if (!paca[nr].cpu_start) {
117 paca[nr].cpu_start = 1;
122 #ifdef CONFIG_HOTPLUG_CPU
124 * Ok it's not there, so it might be soft-unplugged, let's
125 * try to bring it back
127 generic_set_cpu_up(nr);
129 smp_send_reschedule(nr);
130 #endif /* CONFIG_HOTPLUG_CPU */
134 #endif /* CONFIG_PPC64 */
136 static irqreturn_t call_function_action(int irq, void *data)
138 generic_smp_call_function_interrupt();
142 static irqreturn_t reschedule_action(int irq, void *data)
148 static irqreturn_t call_function_single_action(int irq, void *data)
150 generic_smp_call_function_single_interrupt();
154 static irqreturn_t debug_ipi_action(int irq, void *data)
156 if (crash_ipi_function_ptr) {
157 crash_ipi_function_ptr(get_irq_regs());
161 #ifdef CONFIG_DEBUGGER
162 debugger_ipi(get_irq_regs());
163 #endif /* CONFIG_DEBUGGER */
168 static irq_handler_t smp_ipi_action[] = {
169 [PPC_MSG_CALL_FUNCTION] = call_function_action,
170 [PPC_MSG_RESCHEDULE] = reschedule_action,
171 [PPC_MSG_CALL_FUNC_SINGLE] = call_function_single_action,
172 [PPC_MSG_DEBUGGER_BREAK] = debug_ipi_action,
175 const char *smp_ipi_name[] = {
176 [PPC_MSG_CALL_FUNCTION] = "ipi call function",
177 [PPC_MSG_RESCHEDULE] = "ipi reschedule",
178 [PPC_MSG_CALL_FUNC_SINGLE] = "ipi call function single",
179 [PPC_MSG_DEBUGGER_BREAK] = "ipi debugger",
182 /* optional function to request ipi, for controllers with >= 4 ipis */
183 int smp_request_message_ipi(int virq, int msg)
187 if (msg < 0 || msg > PPC_MSG_DEBUGGER_BREAK) {
190 #if !defined(CONFIG_DEBUGGER) && !defined(CONFIG_KEXEC)
191 if (msg == PPC_MSG_DEBUGGER_BREAK) {
195 err = request_irq(virq, smp_ipi_action[msg],
196 IRQF_PERCPU | IRQF_NO_THREAD | IRQF_NO_SUSPEND,
197 smp_ipi_name[msg], NULL);
198 WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n",
199 virq, smp_ipi_name[msg], err);
204 #ifdef CONFIG_PPC_SMP_MUXED_IPI
205 struct cpu_messages {
206 int messages; /* current messages */
207 unsigned long data; /* data for cause ipi */
209 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages, ipi_message);
211 void smp_muxed_ipi_set_data(int cpu, unsigned long data)
213 struct cpu_messages *info = &per_cpu(ipi_message, cpu);
218 void smp_muxed_ipi_message_pass(int cpu, int msg)
220 struct cpu_messages *info = &per_cpu(ipi_message, cpu);
221 char *message = (char *)&info->messages;
224 * Order previous accesses before accesses in the IPI handler.
229 * cause_ipi functions are required to include a full barrier
230 * before doing whatever causes the IPI.
232 smp_ops->cause_ipi(cpu, info->data);
235 #ifdef __BIG_ENDIAN__
236 #define IPI_MESSAGE(A) (1 << (24 - 8 * (A)))
238 #define IPI_MESSAGE(A) (1 << (8 * (A)))
241 irqreturn_t smp_ipi_demux(void)
243 struct cpu_messages *info = &__get_cpu_var(ipi_message);
246 mb(); /* order any irq clear */
249 all = xchg(&info->messages, 0);
250 if (all & IPI_MESSAGE(PPC_MSG_CALL_FUNCTION))
251 generic_smp_call_function_interrupt();
252 if (all & IPI_MESSAGE(PPC_MSG_RESCHEDULE))
254 if (all & IPI_MESSAGE(PPC_MSG_CALL_FUNC_SINGLE))
255 generic_smp_call_function_single_interrupt();
256 if (all & IPI_MESSAGE(PPC_MSG_DEBUGGER_BREAK))
257 debug_ipi_action(0, NULL);
258 } while (info->messages);
262 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
264 static inline void do_message_pass(int cpu, int msg)
266 if (smp_ops->message_pass)
267 smp_ops->message_pass(cpu, msg);
268 #ifdef CONFIG_PPC_SMP_MUXED_IPI
270 smp_muxed_ipi_message_pass(cpu, msg);
274 void smp_send_reschedule(int cpu)
277 do_message_pass(cpu, PPC_MSG_RESCHEDULE);
279 EXPORT_SYMBOL_GPL(smp_send_reschedule);
281 void arch_send_call_function_single_ipi(int cpu)
283 do_message_pass(cpu, PPC_MSG_CALL_FUNC_SINGLE);
286 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
290 for_each_cpu(cpu, mask)
291 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
294 #if defined(CONFIG_DEBUGGER) || defined(CONFIG_KEXEC)
295 void smp_send_debugger_break(void)
298 int me = raw_smp_processor_id();
300 if (unlikely(!smp_ops))
303 for_each_online_cpu(cpu)
305 do_message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
310 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
312 crash_ipi_function_ptr = crash_ipi_callback;
313 if (crash_ipi_callback) {
315 smp_send_debugger_break();
320 static void stop_this_cpu(void *dummy)
322 /* Remove this CPU */
323 set_cpu_online(smp_processor_id(), false);
330 void smp_send_stop(void)
332 smp_call_function(stop_this_cpu, NULL, 0);
335 struct thread_info *current_set[NR_CPUS];
337 static void smp_store_cpu_info(int id)
339 per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
340 #ifdef CONFIG_PPC_FSL_BOOK3E
341 per_cpu(next_tlbcam_idx, id)
342 = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) - 1;
346 void __init smp_prepare_cpus(unsigned int max_cpus)
350 DBG("smp_prepare_cpus\n");
353 * setup_cpu may need to be called on the boot cpu. We havent
354 * spun any cpus up but lets be paranoid.
356 BUG_ON(boot_cpuid != smp_processor_id());
359 smp_store_cpu_info(boot_cpuid);
360 cpu_callin_map[boot_cpuid] = 1;
362 for_each_possible_cpu(cpu) {
363 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
364 GFP_KERNEL, cpu_to_node(cpu));
365 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
366 GFP_KERNEL, cpu_to_node(cpu));
369 cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
370 cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
372 if (smp_ops && smp_ops->probe)
376 void smp_prepare_boot_cpu(void)
378 BUG_ON(smp_processor_id() != boot_cpuid);
380 paca[boot_cpuid].__current = current;
382 current_set[boot_cpuid] = task_thread_info(current);
385 #ifdef CONFIG_HOTPLUG_CPU
387 int generic_cpu_disable(void)
389 unsigned int cpu = smp_processor_id();
391 if (cpu == boot_cpuid)
394 set_cpu_online(cpu, false);
396 vdso_data->processorCount--;
402 void generic_cpu_die(unsigned int cpu)
406 for (i = 0; i < 100; i++) {
408 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
412 printk(KERN_ERR "CPU%d didn't die...\n", cpu);
415 void generic_mach_cpu_die(void)
421 cpu = smp_processor_id();
422 printk(KERN_DEBUG "CPU%d offline\n", cpu);
423 __get_cpu_var(cpu_state) = CPU_DEAD;
425 while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
429 void generic_set_cpu_dead(unsigned int cpu)
431 per_cpu(cpu_state, cpu) = CPU_DEAD;
435 * The cpu_state should be set to CPU_UP_PREPARE in kick_cpu(), otherwise
436 * the cpu_state is always CPU_DEAD after calling generic_set_cpu_dead(),
437 * which makes the delay in generic_cpu_die() not happen.
439 void generic_set_cpu_up(unsigned int cpu)
441 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
444 int generic_check_cpu_restart(unsigned int cpu)
446 return per_cpu(cpu_state, cpu) == CPU_UP_PREPARE;
449 static atomic_t secondary_inhibit_count;
452 * Don't allow secondary CPU threads to come online
454 void inhibit_secondary_onlining(void)
457 * This makes secondary_inhibit_count stable during cpu
458 * online/offline operations.
462 atomic_inc(&secondary_inhibit_count);
465 EXPORT_SYMBOL_GPL(inhibit_secondary_onlining);
468 * Allow secondary CPU threads to come online again
470 void uninhibit_secondary_onlining(void)
473 atomic_dec(&secondary_inhibit_count);
476 EXPORT_SYMBOL_GPL(uninhibit_secondary_onlining);
478 static int secondaries_inhibited(void)
480 return atomic_read(&secondary_inhibit_count);
483 #else /* HOTPLUG_CPU */
485 #define secondaries_inhibited() 0
489 static void cpu_idle_thread_init(unsigned int cpu, struct task_struct *idle)
491 struct thread_info *ti = task_thread_info(idle);
494 paca[cpu].__current = idle;
495 paca[cpu].kstack = (unsigned long)ti + THREAD_SIZE - STACK_FRAME_OVERHEAD;
498 secondary_ti = current_set[cpu] = ti;
501 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
506 * Don't allow secondary threads to come online if inhibited
508 if (threads_per_core > 1 && secondaries_inhibited() &&
509 cpu % threads_per_core != 0)
512 if (smp_ops == NULL ||
513 (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
516 cpu_idle_thread_init(cpu, tidle);
518 /* Make sure callin-map entry is 0 (can be leftover a CPU
521 cpu_callin_map[cpu] = 0;
523 /* The information for processor bringup must
524 * be written out to main store before we release
530 DBG("smp: kicking cpu %d\n", cpu);
531 rc = smp_ops->kick_cpu(cpu);
533 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
538 * wait to see if the cpu made a callin (is actually up).
539 * use this value that I found through experimentation.
542 if (system_state < SYSTEM_RUNNING)
543 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
545 #ifdef CONFIG_HOTPLUG_CPU
548 * CPUs can take much longer to come up in the
549 * hotplug case. Wait five seconds.
551 for (c = 5000; c && !cpu_callin_map[cpu]; c--)
555 if (!cpu_callin_map[cpu]) {
556 printk(KERN_ERR "Processor %u is stuck.\n", cpu);
560 DBG("Processor %u found.\n", cpu);
562 if (smp_ops->give_timebase)
563 smp_ops->give_timebase();
565 /* Wait until cpu puts itself in the online map */
566 while (!cpu_online(cpu))
572 /* Return the value of the reg property corresponding to the given
575 int cpu_to_core_id(int cpu)
577 struct device_node *np;
581 np = of_get_cpu_node(cpu, NULL);
585 reg = of_get_property(np, "reg", NULL);
595 /* Helper routines for cpu to core mapping */
596 int cpu_core_index_of_thread(int cpu)
598 return cpu >> threads_shift;
600 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
602 int cpu_first_thread_of_core(int core)
604 return core << threads_shift;
606 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
608 static void traverse_siblings_chip_id(int cpu, bool add, int chipid)
610 const struct cpumask *mask;
611 struct device_node *np;
615 mask = add ? cpu_online_mask : cpu_present_mask;
616 for_each_cpu(i, mask) {
617 np = of_get_cpu_node(i, NULL);
620 prop = of_get_property(np, "ibm,chip-id", &plen);
621 if (prop && plen == sizeof(int) &&
622 of_read_number(prop, 1) == chipid) {
624 cpumask_set_cpu(cpu, cpu_core_mask(i));
625 cpumask_set_cpu(i, cpu_core_mask(cpu));
627 cpumask_clear_cpu(cpu, cpu_core_mask(i));
628 cpumask_clear_cpu(i, cpu_core_mask(cpu));
635 /* Must be called when no change can occur to cpu_present_mask,
636 * i.e. during cpu online or offline.
638 static struct device_node *cpu_to_l2cache(int cpu)
640 struct device_node *np;
641 struct device_node *cache;
643 if (!cpu_present(cpu))
646 np = of_get_cpu_node(cpu, NULL);
650 cache = of_find_next_cache_node(np);
657 static void traverse_core_siblings(int cpu, bool add)
659 struct device_node *l2_cache, *np;
660 const struct cpumask *mask;
664 /* First see if we have ibm,chip-id properties in cpu nodes */
665 np = of_get_cpu_node(cpu, NULL);
668 prop = of_get_property(np, "ibm,chip-id", &plen);
669 if (prop && plen == sizeof(int))
670 chip = of_read_number(prop, 1);
673 traverse_siblings_chip_id(cpu, add, chip);
678 l2_cache = cpu_to_l2cache(cpu);
679 mask = add ? cpu_online_mask : cpu_present_mask;
680 for_each_cpu(i, mask) {
681 np = cpu_to_l2cache(i);
684 if (np == l2_cache) {
686 cpumask_set_cpu(cpu, cpu_core_mask(i));
687 cpumask_set_cpu(i, cpu_core_mask(cpu));
689 cpumask_clear_cpu(cpu, cpu_core_mask(i));
690 cpumask_clear_cpu(i, cpu_core_mask(cpu));
695 of_node_put(l2_cache);
698 /* Activate a secondary processor. */
699 void start_secondary(void *unused)
701 unsigned int cpu = smp_processor_id();
704 atomic_inc(&init_mm.mm_count);
705 current->active_mm = &init_mm;
707 smp_store_cpu_info(cpu);
708 set_dec(tb_ticks_per_jiffy);
710 cpu_callin_map[cpu] = 1;
712 if (smp_ops->setup_cpu)
713 smp_ops->setup_cpu(cpu);
714 if (smp_ops->take_timebase)
715 smp_ops->take_timebase();
717 secondary_cpu_time_init();
720 if (system_state == SYSTEM_RUNNING)
721 vdso_data->processorCount++;
725 /* Update sibling maps */
726 base = cpu_first_thread_sibling(cpu);
727 for (i = 0; i < threads_per_core; i++) {
728 if (cpu_is_offline(base + i) && (cpu != base + i))
730 cpumask_set_cpu(cpu, cpu_sibling_mask(base + i));
731 cpumask_set_cpu(base + i, cpu_sibling_mask(cpu));
733 /* cpu_core_map should be a superset of
734 * cpu_sibling_map even if we don't have cache
735 * information, so update the former here, too.
737 cpumask_set_cpu(cpu, cpu_core_mask(base + i));
738 cpumask_set_cpu(base + i, cpu_core_mask(cpu));
740 traverse_core_siblings(cpu, true);
743 notify_cpu_starting(cpu);
744 set_cpu_online(cpu, true);
748 cpu_startup_entry(CPUHP_ONLINE);
753 int setup_profiling_timer(unsigned int multiplier)
758 void __init smp_cpus_done(unsigned int max_cpus)
760 cpumask_var_t old_mask;
762 /* We want the setup_cpu() here to be called from CPU 0, but our
763 * init thread may have been "borrowed" by another CPU in the meantime
764 * se we pin us down to CPU 0 for a short while
766 alloc_cpumask_var(&old_mask, GFP_NOWAIT);
767 cpumask_copy(old_mask, tsk_cpus_allowed(current));
768 set_cpus_allowed_ptr(current, cpumask_of(boot_cpuid));
770 if (smp_ops && smp_ops->setup_cpu)
771 smp_ops->setup_cpu(boot_cpuid);
773 set_cpus_allowed_ptr(current, old_mask);
775 free_cpumask_var(old_mask);
777 if (smp_ops && smp_ops->bringup_done)
778 smp_ops->bringup_done();
780 dump_numa_cpu_topology();
784 int arch_sd_sibling_asym_packing(void)
786 if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
787 printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
788 return SD_ASYM_PACKING;
793 #ifdef CONFIG_HOTPLUG_CPU
794 int __cpu_disable(void)
796 int cpu = smp_processor_id();
800 if (!smp_ops->cpu_disable)
803 err = smp_ops->cpu_disable();
807 /* Update sibling maps */
808 base = cpu_first_thread_sibling(cpu);
809 for (i = 0; i < threads_per_core; i++) {
810 cpumask_clear_cpu(cpu, cpu_sibling_mask(base + i));
811 cpumask_clear_cpu(base + i, cpu_sibling_mask(cpu));
812 cpumask_clear_cpu(cpu, cpu_core_mask(base + i));
813 cpumask_clear_cpu(base + i, cpu_core_mask(cpu));
815 traverse_core_siblings(cpu, false);
820 void __cpu_die(unsigned int cpu)
822 if (smp_ops->cpu_die)
823 smp_ops->cpu_die(cpu);
831 /* If we return, we re-enter start_secondary */
832 start_secondary_resume();