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>
38 #include <asm/hw_irq.h>
39 #include <asm/kvm_ppc.h>
41 #include <asm/pgtable.h>
45 #include <asm/machdep.h>
46 #include <asm/cputhreads.h>
47 #include <asm/cputable.h>
49 #include <asm/vdso_datapage.h>
54 #include <asm/debug.h>
58 #define DBG(fmt...) udbg_printf(fmt)
63 #ifdef CONFIG_HOTPLUG_CPU
64 /* State of each CPU during hotplug phases */
65 static DEFINE_PER_CPU(int, cpu_state) = { 0 };
68 struct thread_info *secondary_ti;
70 DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
71 DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
73 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
74 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
76 /* SMP operations for this machine */
77 struct smp_ops_t *smp_ops;
79 /* Can't be static due to PowerMac hackery */
80 volatile unsigned int cpu_callin_map[NR_CPUS];
82 int smt_enabled_at_boot = 1;
84 static void (*crash_ipi_function_ptr)(struct pt_regs *) = NULL;
87 * Returns 1 if the specified cpu should be brought up during boot.
88 * Used to inhibit booting threads if they've been disabled or
89 * limited on the command line
91 int smp_generic_cpu_bootable(unsigned int nr)
93 /* Special case - we inhibit secondary thread startup
94 * during boot if the user requests it.
96 if (system_state == SYSTEM_BOOTING && cpu_has_feature(CPU_FTR_SMT)) {
97 if (!smt_enabled_at_boot && cpu_thread_in_core(nr) != 0)
99 if (smt_enabled_at_boot
100 && cpu_thread_in_core(nr) >= smt_enabled_at_boot)
109 int smp_generic_kick_cpu(int nr)
111 BUG_ON(nr < 0 || nr >= NR_CPUS);
114 * The processor is currently spinning, waiting for the
115 * cpu_start field to become non-zero After we set cpu_start,
116 * the processor will continue on to secondary_start
118 if (!paca[nr].cpu_start) {
119 paca[nr].cpu_start = 1;
124 #ifdef CONFIG_HOTPLUG_CPU
126 * Ok it's not there, so it might be soft-unplugged, let's
127 * try to bring it back
129 generic_set_cpu_up(nr);
131 smp_send_reschedule(nr);
132 #endif /* CONFIG_HOTPLUG_CPU */
136 #endif /* CONFIG_PPC64 */
138 static irqreturn_t call_function_action(int irq, void *data)
140 generic_smp_call_function_interrupt();
144 static irqreturn_t reschedule_action(int irq, void *data)
150 static irqreturn_t tick_broadcast_ipi_action(int irq, void *data)
152 tick_broadcast_ipi_handler();
156 static irqreturn_t debug_ipi_action(int irq, void *data)
158 if (crash_ipi_function_ptr) {
159 crash_ipi_function_ptr(get_irq_regs());
163 #ifdef CONFIG_DEBUGGER
164 debugger_ipi(get_irq_regs());
165 #endif /* CONFIG_DEBUGGER */
170 static irq_handler_t smp_ipi_action[] = {
171 [PPC_MSG_CALL_FUNCTION] = call_function_action,
172 [PPC_MSG_RESCHEDULE] = reschedule_action,
173 [PPC_MSG_TICK_BROADCAST] = tick_broadcast_ipi_action,
174 [PPC_MSG_DEBUGGER_BREAK] = debug_ipi_action,
177 const char *smp_ipi_name[] = {
178 [PPC_MSG_CALL_FUNCTION] = "ipi call function",
179 [PPC_MSG_RESCHEDULE] = "ipi reschedule",
180 [PPC_MSG_TICK_BROADCAST] = "ipi tick-broadcast",
181 [PPC_MSG_DEBUGGER_BREAK] = "ipi debugger",
184 /* optional function to request ipi, for controllers with >= 4 ipis */
185 int smp_request_message_ipi(int virq, int msg)
189 if (msg < 0 || msg > PPC_MSG_DEBUGGER_BREAK) {
192 #if !defined(CONFIG_DEBUGGER) && !defined(CONFIG_KEXEC)
193 if (msg == PPC_MSG_DEBUGGER_BREAK) {
197 err = request_irq(virq, smp_ipi_action[msg],
198 IRQF_PERCPU | IRQF_NO_THREAD | IRQF_NO_SUSPEND,
199 smp_ipi_name[msg], NULL);
200 WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n",
201 virq, smp_ipi_name[msg], err);
206 #ifdef CONFIG_PPC_SMP_MUXED_IPI
207 struct cpu_messages {
208 int messages; /* current messages */
209 unsigned long data; /* data for cause ipi */
211 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages, ipi_message);
213 void smp_muxed_ipi_set_data(int cpu, unsigned long data)
215 struct cpu_messages *info = &per_cpu(ipi_message, cpu);
220 void smp_muxed_ipi_message_pass(int cpu, int msg)
222 struct cpu_messages *info = &per_cpu(ipi_message, cpu);
223 char *message = (char *)&info->messages;
226 * Order previous accesses before accesses in the IPI handler.
231 * cause_ipi functions are required to include a full barrier
232 * before doing whatever causes the IPI.
234 smp_ops->cause_ipi(cpu, info->data);
237 #ifdef __BIG_ENDIAN__
238 #define IPI_MESSAGE(A) (1 << (24 - 8 * (A)))
240 #define IPI_MESSAGE(A) (1 << (8 * (A)))
243 irqreturn_t smp_ipi_demux(void)
245 struct cpu_messages *info = &__get_cpu_var(ipi_message);
248 mb(); /* order any irq clear */
251 all = xchg(&info->messages, 0);
252 if (all & IPI_MESSAGE(PPC_MSG_CALL_FUNCTION))
253 generic_smp_call_function_interrupt();
254 if (all & IPI_MESSAGE(PPC_MSG_RESCHEDULE))
256 if (all & IPI_MESSAGE(PPC_MSG_TICK_BROADCAST))
257 tick_broadcast_ipi_handler();
258 if (all & IPI_MESSAGE(PPC_MSG_DEBUGGER_BREAK))
259 debug_ipi_action(0, NULL);
260 } while (info->messages);
264 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
266 static inline void do_message_pass(int cpu, int msg)
268 if (smp_ops->message_pass)
269 smp_ops->message_pass(cpu, msg);
270 #ifdef CONFIG_PPC_SMP_MUXED_IPI
272 smp_muxed_ipi_message_pass(cpu, msg);
276 void smp_send_reschedule(int cpu)
279 do_message_pass(cpu, PPC_MSG_RESCHEDULE);
281 EXPORT_SYMBOL_GPL(smp_send_reschedule);
283 void arch_send_call_function_single_ipi(int cpu)
285 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
288 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
292 for_each_cpu(cpu, mask)
293 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
296 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
297 void tick_broadcast(const struct cpumask *mask)
301 for_each_cpu(cpu, mask)
302 do_message_pass(cpu, PPC_MSG_TICK_BROADCAST);
306 #if defined(CONFIG_DEBUGGER) || defined(CONFIG_KEXEC)
307 void smp_send_debugger_break(void)
310 int me = raw_smp_processor_id();
312 if (unlikely(!smp_ops))
315 for_each_online_cpu(cpu)
317 do_message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
322 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
324 crash_ipi_function_ptr = crash_ipi_callback;
325 if (crash_ipi_callback) {
327 smp_send_debugger_break();
332 static void stop_this_cpu(void *dummy)
334 /* Remove this CPU */
335 set_cpu_online(smp_processor_id(), false);
342 void smp_send_stop(void)
344 smp_call_function(stop_this_cpu, NULL, 0);
347 struct thread_info *current_set[NR_CPUS];
349 static void smp_store_cpu_info(int id)
351 per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
352 #ifdef CONFIG_PPC_FSL_BOOK3E
353 per_cpu(next_tlbcam_idx, id)
354 = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) - 1;
358 void __init smp_prepare_cpus(unsigned int max_cpus)
362 DBG("smp_prepare_cpus\n");
365 * setup_cpu may need to be called on the boot cpu. We havent
366 * spun any cpus up but lets be paranoid.
368 BUG_ON(boot_cpuid != smp_processor_id());
371 smp_store_cpu_info(boot_cpuid);
372 cpu_callin_map[boot_cpuid] = 1;
374 for_each_possible_cpu(cpu) {
375 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
376 GFP_KERNEL, cpu_to_node(cpu));
377 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
378 GFP_KERNEL, cpu_to_node(cpu));
381 cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
382 cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
384 if (smp_ops && smp_ops->probe)
388 void smp_prepare_boot_cpu(void)
390 BUG_ON(smp_processor_id() != boot_cpuid);
392 paca[boot_cpuid].__current = current;
394 set_numa_node(numa_cpu_lookup_table[boot_cpuid]);
395 current_set[boot_cpuid] = task_thread_info(current);
398 #ifdef CONFIG_HOTPLUG_CPU
400 int generic_cpu_disable(void)
402 unsigned int cpu = smp_processor_id();
404 if (cpu == boot_cpuid)
407 set_cpu_online(cpu, false);
409 vdso_data->processorCount--;
415 void generic_cpu_die(unsigned int cpu)
419 for (i = 0; i < 100; i++) {
421 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
425 printk(KERN_ERR "CPU%d didn't die...\n", cpu);
428 void generic_mach_cpu_die(void)
434 cpu = smp_processor_id();
435 printk(KERN_DEBUG "CPU%d offline\n", cpu);
436 __get_cpu_var(cpu_state) = CPU_DEAD;
438 while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
442 void generic_set_cpu_dead(unsigned int cpu)
444 per_cpu(cpu_state, cpu) = CPU_DEAD;
448 * The cpu_state should be set to CPU_UP_PREPARE in kick_cpu(), otherwise
449 * the cpu_state is always CPU_DEAD after calling generic_set_cpu_dead(),
450 * which makes the delay in generic_cpu_die() not happen.
452 void generic_set_cpu_up(unsigned int cpu)
454 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
457 int generic_check_cpu_restart(unsigned int cpu)
459 return per_cpu(cpu_state, cpu) == CPU_UP_PREPARE;
462 static bool secondaries_inhibited(void)
464 return kvm_hv_mode_active();
467 #else /* HOTPLUG_CPU */
469 #define secondaries_inhibited() 0
473 static void cpu_idle_thread_init(unsigned int cpu, struct task_struct *idle)
475 struct thread_info *ti = task_thread_info(idle);
478 paca[cpu].__current = idle;
479 paca[cpu].kstack = (unsigned long)ti + THREAD_SIZE - STACK_FRAME_OVERHEAD;
482 secondary_ti = current_set[cpu] = ti;
485 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
490 * Don't allow secondary threads to come online if inhibited
492 if (threads_per_core > 1 && secondaries_inhibited() &&
493 cpu_thread_in_subcore(cpu))
496 if (smp_ops == NULL ||
497 (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
500 cpu_idle_thread_init(cpu, tidle);
502 /* Make sure callin-map entry is 0 (can be leftover a CPU
505 cpu_callin_map[cpu] = 0;
507 /* The information for processor bringup must
508 * be written out to main store before we release
514 DBG("smp: kicking cpu %d\n", cpu);
515 rc = smp_ops->kick_cpu(cpu);
517 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
522 * wait to see if the cpu made a callin (is actually up).
523 * use this value that I found through experimentation.
526 if (system_state < SYSTEM_RUNNING)
527 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
529 #ifdef CONFIG_HOTPLUG_CPU
532 * CPUs can take much longer to come up in the
533 * hotplug case. Wait five seconds.
535 for (c = 5000; c && !cpu_callin_map[cpu]; c--)
539 if (!cpu_callin_map[cpu]) {
540 printk(KERN_ERR "Processor %u is stuck.\n", cpu);
544 DBG("Processor %u found.\n", cpu);
546 if (smp_ops->give_timebase)
547 smp_ops->give_timebase();
549 /* Wait until cpu puts itself in the online map */
550 while (!cpu_online(cpu))
556 /* Return the value of the reg property corresponding to the given
559 int cpu_to_core_id(int cpu)
561 struct device_node *np;
565 np = of_get_cpu_node(cpu, NULL);
569 reg = of_get_property(np, "reg", NULL);
573 id = be32_to_cpup(reg);
579 /* Helper routines for cpu to core mapping */
580 int cpu_core_index_of_thread(int cpu)
582 return cpu >> threads_shift;
584 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
586 int cpu_first_thread_of_core(int core)
588 return core << threads_shift;
590 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
592 static void traverse_siblings_chip_id(int cpu, bool add, int chipid)
594 const struct cpumask *mask;
595 struct device_node *np;
599 mask = add ? cpu_online_mask : cpu_present_mask;
600 for_each_cpu(i, mask) {
601 np = of_get_cpu_node(i, NULL);
604 prop = of_get_property(np, "ibm,chip-id", &plen);
605 if (prop && plen == sizeof(int) &&
606 of_read_number(prop, 1) == chipid) {
608 cpumask_set_cpu(cpu, cpu_core_mask(i));
609 cpumask_set_cpu(i, cpu_core_mask(cpu));
611 cpumask_clear_cpu(cpu, cpu_core_mask(i));
612 cpumask_clear_cpu(i, cpu_core_mask(cpu));
619 /* Must be called when no change can occur to cpu_present_mask,
620 * i.e. during cpu online or offline.
622 static struct device_node *cpu_to_l2cache(int cpu)
624 struct device_node *np;
625 struct device_node *cache;
627 if (!cpu_present(cpu))
630 np = of_get_cpu_node(cpu, NULL);
634 cache = of_find_next_cache_node(np);
641 static void traverse_core_siblings(int cpu, bool add)
643 struct device_node *l2_cache, *np;
644 const struct cpumask *mask;
648 /* First see if we have ibm,chip-id properties in cpu nodes */
649 np = of_get_cpu_node(cpu, NULL);
652 prop = of_get_property(np, "ibm,chip-id", &plen);
653 if (prop && plen == sizeof(int))
654 chip = of_read_number(prop, 1);
657 traverse_siblings_chip_id(cpu, add, chip);
662 l2_cache = cpu_to_l2cache(cpu);
663 mask = add ? cpu_online_mask : cpu_present_mask;
664 for_each_cpu(i, mask) {
665 np = cpu_to_l2cache(i);
668 if (np == l2_cache) {
670 cpumask_set_cpu(cpu, cpu_core_mask(i));
671 cpumask_set_cpu(i, cpu_core_mask(cpu));
673 cpumask_clear_cpu(cpu, cpu_core_mask(i));
674 cpumask_clear_cpu(i, cpu_core_mask(cpu));
679 of_node_put(l2_cache);
682 /* Activate a secondary processor. */
683 void start_secondary(void *unused)
685 unsigned int cpu = smp_processor_id();
688 atomic_inc(&init_mm.mm_count);
689 current->active_mm = &init_mm;
691 smp_store_cpu_info(cpu);
692 set_dec(tb_ticks_per_jiffy);
694 cpu_callin_map[cpu] = 1;
696 if (smp_ops->setup_cpu)
697 smp_ops->setup_cpu(cpu);
698 if (smp_ops->take_timebase)
699 smp_ops->take_timebase();
701 secondary_cpu_time_init();
704 if (system_state == SYSTEM_RUNNING)
705 vdso_data->processorCount++;
709 /* Update sibling maps */
710 base = cpu_first_thread_sibling(cpu);
711 for (i = 0; i < threads_per_core; i++) {
712 if (cpu_is_offline(base + i) && (cpu != base + i))
714 cpumask_set_cpu(cpu, cpu_sibling_mask(base + i));
715 cpumask_set_cpu(base + i, cpu_sibling_mask(cpu));
717 /* cpu_core_map should be a superset of
718 * cpu_sibling_map even if we don't have cache
719 * information, so update the former here, too.
721 cpumask_set_cpu(cpu, cpu_core_mask(base + i));
722 cpumask_set_cpu(base + i, cpu_core_mask(cpu));
724 traverse_core_siblings(cpu, true);
727 * numa_node_id() works after this.
729 set_numa_node(numa_cpu_lookup_table[cpu]);
730 set_numa_mem(local_memory_node(numa_cpu_lookup_table[cpu]));
733 notify_cpu_starting(cpu);
734 set_cpu_online(cpu, true);
738 cpu_startup_entry(CPUHP_ONLINE);
743 int setup_profiling_timer(unsigned int multiplier)
748 #ifdef CONFIG_SCHED_SMT
749 /* cpumask of CPUs with asymetric SMT dependancy */
750 static int powerpc_smt_flags(void)
752 int flags = SD_SHARE_CPUCAPACITY | SD_SHARE_PKG_RESOURCES;
754 if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
755 printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
756 flags |= SD_ASYM_PACKING;
762 static struct sched_domain_topology_level powerpc_topology[] = {
763 #ifdef CONFIG_SCHED_SMT
764 { cpu_smt_mask, powerpc_smt_flags, SD_INIT_NAME(SMT) },
766 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
770 void __init smp_cpus_done(unsigned int max_cpus)
772 cpumask_var_t old_mask;
774 /* We want the setup_cpu() here to be called from CPU 0, but our
775 * init thread may have been "borrowed" by another CPU in the meantime
776 * se we pin us down to CPU 0 for a short while
778 alloc_cpumask_var(&old_mask, GFP_NOWAIT);
779 cpumask_copy(old_mask, tsk_cpus_allowed(current));
780 set_cpus_allowed_ptr(current, cpumask_of(boot_cpuid));
782 if (smp_ops && smp_ops->setup_cpu)
783 smp_ops->setup_cpu(boot_cpuid);
785 set_cpus_allowed_ptr(current, old_mask);
787 free_cpumask_var(old_mask);
789 if (smp_ops && smp_ops->bringup_done)
790 smp_ops->bringup_done();
792 dump_numa_cpu_topology();
794 set_sched_topology(powerpc_topology);
798 #ifdef CONFIG_HOTPLUG_CPU
799 int __cpu_disable(void)
801 int cpu = smp_processor_id();
805 if (!smp_ops->cpu_disable)
808 err = smp_ops->cpu_disable();
812 /* Update sibling maps */
813 base = cpu_first_thread_sibling(cpu);
814 for (i = 0; i < threads_per_core; i++) {
815 cpumask_clear_cpu(cpu, cpu_sibling_mask(base + i));
816 cpumask_clear_cpu(base + i, cpu_sibling_mask(cpu));
817 cpumask_clear_cpu(cpu, cpu_core_mask(base + i));
818 cpumask_clear_cpu(base + i, cpu_core_mask(cpu));
820 traverse_core_siblings(cpu, false);
825 void __cpu_die(unsigned int cpu)
827 if (smp_ops->cpu_die)
828 smp_ops->cpu_die(cpu);
836 /* If we return, we re-enter start_secondary */
837 start_secondary_resume();