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>
55 #include <asm/kexec.h>
59 #define DBG(fmt...) udbg_printf(fmt)
64 #ifdef CONFIG_HOTPLUG_CPU
65 /* State of each CPU during hotplug phases */
66 static DEFINE_PER_CPU(int, cpu_state) = { 0 };
69 struct thread_info *secondary_ti;
71 DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
72 DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
74 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
75 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
77 /* SMP operations for this machine */
78 struct smp_ops_t *smp_ops;
80 /* Can't be static due to PowerMac hackery */
81 volatile unsigned int cpu_callin_map[NR_CPUS];
83 int smt_enabled_at_boot = 1;
85 static void (*crash_ipi_function_ptr)(struct pt_regs *) = NULL;
88 * Returns 1 if the specified cpu should be brought up during boot.
89 * Used to inhibit booting threads if they've been disabled or
90 * limited on the command line
92 int smp_generic_cpu_bootable(unsigned int nr)
94 /* Special case - we inhibit secondary thread startup
95 * during boot if the user requests it.
97 if (system_state == SYSTEM_BOOTING && cpu_has_feature(CPU_FTR_SMT)) {
98 if (!smt_enabled_at_boot && cpu_thread_in_core(nr) != 0)
100 if (smt_enabled_at_boot
101 && cpu_thread_in_core(nr) >= smt_enabled_at_boot)
110 int smp_generic_kick_cpu(int nr)
112 BUG_ON(nr < 0 || nr >= NR_CPUS);
115 * The processor is currently spinning, waiting for the
116 * cpu_start field to become non-zero After we set cpu_start,
117 * the processor will continue on to secondary_start
119 if (!paca[nr].cpu_start) {
120 paca[nr].cpu_start = 1;
125 #ifdef CONFIG_HOTPLUG_CPU
127 * Ok it's not there, so it might be soft-unplugged, let's
128 * try to bring it back
130 generic_set_cpu_up(nr);
132 smp_send_reschedule(nr);
133 #endif /* CONFIG_HOTPLUG_CPU */
137 #endif /* CONFIG_PPC64 */
139 static irqreturn_t call_function_action(int irq, void *data)
141 generic_smp_call_function_interrupt();
145 static irqreturn_t reschedule_action(int irq, void *data)
151 static irqreturn_t tick_broadcast_ipi_action(int irq, void *data)
153 tick_broadcast_ipi_handler();
157 static irqreturn_t debug_ipi_action(int irq, void *data)
159 if (crash_ipi_function_ptr) {
160 crash_ipi_function_ptr(get_irq_regs());
164 #ifdef CONFIG_DEBUGGER
165 debugger_ipi(get_irq_regs());
166 #endif /* CONFIG_DEBUGGER */
171 static irq_handler_t smp_ipi_action[] = {
172 [PPC_MSG_CALL_FUNCTION] = call_function_action,
173 [PPC_MSG_RESCHEDULE] = reschedule_action,
174 [PPC_MSG_TICK_BROADCAST] = tick_broadcast_ipi_action,
175 [PPC_MSG_DEBUGGER_BREAK] = debug_ipi_action,
178 const char *smp_ipi_name[] = {
179 [PPC_MSG_CALL_FUNCTION] = "ipi call function",
180 [PPC_MSG_RESCHEDULE] = "ipi reschedule",
181 [PPC_MSG_TICK_BROADCAST] = "ipi tick-broadcast",
182 [PPC_MSG_DEBUGGER_BREAK] = "ipi debugger",
185 /* optional function to request ipi, for controllers with >= 4 ipis */
186 int smp_request_message_ipi(int virq, int msg)
190 if (msg < 0 || msg > PPC_MSG_DEBUGGER_BREAK) {
193 #if !defined(CONFIG_DEBUGGER) && !defined(CONFIG_KEXEC)
194 if (msg == PPC_MSG_DEBUGGER_BREAK) {
198 err = request_irq(virq, smp_ipi_action[msg],
199 IRQF_PERCPU | IRQF_NO_THREAD | IRQF_NO_SUSPEND,
200 smp_ipi_name[msg], NULL);
201 WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n",
202 virq, smp_ipi_name[msg], err);
207 #ifdef CONFIG_PPC_SMP_MUXED_IPI
208 struct cpu_messages {
209 int messages; /* current messages */
210 unsigned long data; /* data for cause ipi */
212 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages, ipi_message);
214 void smp_muxed_ipi_set_data(int cpu, unsigned long data)
216 struct cpu_messages *info = &per_cpu(ipi_message, cpu);
221 void smp_muxed_ipi_message_pass(int cpu, int msg)
223 struct cpu_messages *info = &per_cpu(ipi_message, cpu);
224 char *message = (char *)&info->messages;
227 * Order previous accesses before accesses in the IPI handler.
232 * cause_ipi functions are required to include a full barrier
233 * before doing whatever causes the IPI.
235 smp_ops->cause_ipi(cpu, info->data);
238 #ifdef __BIG_ENDIAN__
239 #define IPI_MESSAGE(A) (1 << (24 - 8 * (A)))
241 #define IPI_MESSAGE(A) (1 << (8 * (A)))
244 irqreturn_t smp_ipi_demux(void)
246 struct cpu_messages *info = this_cpu_ptr(&ipi_message);
249 mb(); /* order any irq clear */
252 all = xchg(&info->messages, 0);
253 if (all & IPI_MESSAGE(PPC_MSG_CALL_FUNCTION))
254 generic_smp_call_function_interrupt();
255 if (all & IPI_MESSAGE(PPC_MSG_RESCHEDULE))
257 if (all & IPI_MESSAGE(PPC_MSG_TICK_BROADCAST))
258 tick_broadcast_ipi_handler();
259 if (all & IPI_MESSAGE(PPC_MSG_DEBUGGER_BREAK))
260 debug_ipi_action(0, NULL);
261 } while (info->messages);
265 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
267 static inline void do_message_pass(int cpu, int msg)
269 if (smp_ops->message_pass)
270 smp_ops->message_pass(cpu, msg);
271 #ifdef CONFIG_PPC_SMP_MUXED_IPI
273 smp_muxed_ipi_message_pass(cpu, msg);
277 void smp_send_reschedule(int cpu)
280 do_message_pass(cpu, PPC_MSG_RESCHEDULE);
282 EXPORT_SYMBOL_GPL(smp_send_reschedule);
284 void arch_send_call_function_single_ipi(int cpu)
286 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
289 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
293 for_each_cpu(cpu, mask)
294 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
297 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
298 void tick_broadcast(const struct cpumask *mask)
302 for_each_cpu(cpu, mask)
303 do_message_pass(cpu, PPC_MSG_TICK_BROADCAST);
307 #if defined(CONFIG_DEBUGGER) || defined(CONFIG_KEXEC)
308 void smp_send_debugger_break(void)
311 int me = raw_smp_processor_id();
313 if (unlikely(!smp_ops))
316 for_each_online_cpu(cpu)
318 do_message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
323 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
325 crash_ipi_function_ptr = crash_ipi_callback;
326 if (crash_ipi_callback) {
328 smp_send_debugger_break();
333 static void stop_this_cpu(void *dummy)
335 /* Remove this CPU */
336 set_cpu_online(smp_processor_id(), false);
343 void smp_send_stop(void)
345 smp_call_function(stop_this_cpu, NULL, 0);
348 struct thread_info *current_set[NR_CPUS];
350 static void smp_store_cpu_info(int id)
352 per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
353 #ifdef CONFIG_PPC_FSL_BOOK3E
354 per_cpu(next_tlbcam_idx, id)
355 = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) - 1;
359 void __init smp_prepare_cpus(unsigned int max_cpus)
363 DBG("smp_prepare_cpus\n");
366 * setup_cpu may need to be called on the boot cpu. We havent
367 * spun any cpus up but lets be paranoid.
369 BUG_ON(boot_cpuid != smp_processor_id());
372 smp_store_cpu_info(boot_cpuid);
373 cpu_callin_map[boot_cpuid] = 1;
375 for_each_possible_cpu(cpu) {
376 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
377 GFP_KERNEL, cpu_to_node(cpu));
378 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
379 GFP_KERNEL, cpu_to_node(cpu));
381 * numa_node_id() works after this.
383 if (cpu_present(cpu)) {
384 set_cpu_numa_node(cpu, numa_cpu_lookup_table[cpu]);
385 set_cpu_numa_mem(cpu,
386 local_memory_node(numa_cpu_lookup_table[cpu]));
390 cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
391 cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
393 if (smp_ops && smp_ops->probe)
397 void smp_prepare_boot_cpu(void)
399 BUG_ON(smp_processor_id() != boot_cpuid);
401 paca[boot_cpuid].__current = current;
403 set_numa_node(numa_cpu_lookup_table[boot_cpuid]);
404 current_set[boot_cpuid] = task_thread_info(current);
407 #ifdef CONFIG_HOTPLUG_CPU
409 int generic_cpu_disable(void)
411 unsigned int cpu = smp_processor_id();
413 if (cpu == boot_cpuid)
416 set_cpu_online(cpu, false);
418 vdso_data->processorCount--;
424 void generic_cpu_die(unsigned int cpu)
428 for (i = 0; i < 100; i++) {
430 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
434 printk(KERN_ERR "CPU%d didn't die...\n", cpu);
437 void generic_mach_cpu_die(void)
443 cpu = smp_processor_id();
444 printk(KERN_DEBUG "CPU%d offline\n", cpu);
445 __this_cpu_write(cpu_state, CPU_DEAD);
447 while (__this_cpu_read(cpu_state) != CPU_UP_PREPARE)
451 void generic_set_cpu_dead(unsigned int cpu)
453 per_cpu(cpu_state, cpu) = CPU_DEAD;
457 * The cpu_state should be set to CPU_UP_PREPARE in kick_cpu(), otherwise
458 * the cpu_state is always CPU_DEAD after calling generic_set_cpu_dead(),
459 * which makes the delay in generic_cpu_die() not happen.
461 void generic_set_cpu_up(unsigned int cpu)
463 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
466 int generic_check_cpu_restart(unsigned int cpu)
468 return per_cpu(cpu_state, cpu) == CPU_UP_PREPARE;
471 static bool secondaries_inhibited(void)
473 return kvm_hv_mode_active();
476 #else /* HOTPLUG_CPU */
478 #define secondaries_inhibited() 0
482 static void cpu_idle_thread_init(unsigned int cpu, struct task_struct *idle)
484 struct thread_info *ti = task_thread_info(idle);
487 paca[cpu].__current = idle;
488 paca[cpu].kstack = (unsigned long)ti + THREAD_SIZE - STACK_FRAME_OVERHEAD;
491 secondary_ti = current_set[cpu] = ti;
494 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
499 * Don't allow secondary threads to come online if inhibited
501 if (threads_per_core > 1 && secondaries_inhibited() &&
502 cpu_thread_in_subcore(cpu))
505 if (smp_ops == NULL ||
506 (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
509 cpu_idle_thread_init(cpu, tidle);
511 /* Make sure callin-map entry is 0 (can be leftover a CPU
514 cpu_callin_map[cpu] = 0;
516 /* The information for processor bringup must
517 * be written out to main store before we release
523 DBG("smp: kicking cpu %d\n", cpu);
524 rc = smp_ops->kick_cpu(cpu);
526 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
531 * wait to see if the cpu made a callin (is actually up).
532 * use this value that I found through experimentation.
535 if (system_state < SYSTEM_RUNNING)
536 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
538 #ifdef CONFIG_HOTPLUG_CPU
541 * CPUs can take much longer to come up in the
542 * hotplug case. Wait five seconds.
544 for (c = 5000; c && !cpu_callin_map[cpu]; c--)
548 if (!cpu_callin_map[cpu]) {
549 printk(KERN_ERR "Processor %u is stuck.\n", cpu);
553 DBG("Processor %u found.\n", cpu);
555 if (smp_ops->give_timebase)
556 smp_ops->give_timebase();
558 /* Wait until cpu puts itself in the online map */
559 while (!cpu_online(cpu))
565 /* Return the value of the reg property corresponding to the given
568 int cpu_to_core_id(int cpu)
570 struct device_node *np;
574 np = of_get_cpu_node(cpu, NULL);
578 reg = of_get_property(np, "reg", NULL);
582 id = be32_to_cpup(reg);
588 /* Helper routines for cpu to core mapping */
589 int cpu_core_index_of_thread(int cpu)
591 return cpu >> threads_shift;
593 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
595 int cpu_first_thread_of_core(int core)
597 return core << threads_shift;
599 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
601 static void traverse_siblings_chip_id(int cpu, bool add, int chipid)
603 const struct cpumask *mask;
604 struct device_node *np;
608 mask = add ? cpu_online_mask : cpu_present_mask;
609 for_each_cpu(i, mask) {
610 np = of_get_cpu_node(i, NULL);
613 prop = of_get_property(np, "ibm,chip-id", &plen);
614 if (prop && plen == sizeof(int) &&
615 of_read_number(prop, 1) == chipid) {
617 cpumask_set_cpu(cpu, cpu_core_mask(i));
618 cpumask_set_cpu(i, cpu_core_mask(cpu));
620 cpumask_clear_cpu(cpu, cpu_core_mask(i));
621 cpumask_clear_cpu(i, cpu_core_mask(cpu));
628 /* Must be called when no change can occur to cpu_present_mask,
629 * i.e. during cpu online or offline.
631 static struct device_node *cpu_to_l2cache(int cpu)
633 struct device_node *np;
634 struct device_node *cache;
636 if (!cpu_present(cpu))
639 np = of_get_cpu_node(cpu, NULL);
643 cache = of_find_next_cache_node(np);
650 static void traverse_core_siblings(int cpu, bool add)
652 struct device_node *l2_cache, *np;
653 const struct cpumask *mask;
657 /* First see if we have ibm,chip-id properties in cpu nodes */
658 np = of_get_cpu_node(cpu, NULL);
661 prop = of_get_property(np, "ibm,chip-id", &plen);
662 if (prop && plen == sizeof(int))
663 chip = of_read_number(prop, 1);
666 traverse_siblings_chip_id(cpu, add, chip);
671 l2_cache = cpu_to_l2cache(cpu);
672 mask = add ? cpu_online_mask : cpu_present_mask;
673 for_each_cpu(i, mask) {
674 np = cpu_to_l2cache(i);
677 if (np == l2_cache) {
679 cpumask_set_cpu(cpu, cpu_core_mask(i));
680 cpumask_set_cpu(i, cpu_core_mask(cpu));
682 cpumask_clear_cpu(cpu, cpu_core_mask(i));
683 cpumask_clear_cpu(i, cpu_core_mask(cpu));
688 of_node_put(l2_cache);
691 /* Activate a secondary processor. */
692 void start_secondary(void *unused)
694 unsigned int cpu = smp_processor_id();
697 atomic_inc(&init_mm.mm_count);
698 current->active_mm = &init_mm;
700 smp_store_cpu_info(cpu);
701 set_dec(tb_ticks_per_jiffy);
703 cpu_callin_map[cpu] = 1;
705 if (smp_ops->setup_cpu)
706 smp_ops->setup_cpu(cpu);
707 if (smp_ops->take_timebase)
708 smp_ops->take_timebase();
710 secondary_cpu_time_init();
713 if (system_state == SYSTEM_RUNNING)
714 vdso_data->processorCount++;
718 /* Update sibling maps */
719 base = cpu_first_thread_sibling(cpu);
720 for (i = 0; i < threads_per_core; i++) {
721 if (cpu_is_offline(base + i) && (cpu != base + i))
723 cpumask_set_cpu(cpu, cpu_sibling_mask(base + i));
724 cpumask_set_cpu(base + i, cpu_sibling_mask(cpu));
726 /* cpu_core_map should be a superset of
727 * cpu_sibling_map even if we don't have cache
728 * information, so update the former here, too.
730 cpumask_set_cpu(cpu, cpu_core_mask(base + i));
731 cpumask_set_cpu(base + i, cpu_core_mask(cpu));
733 traverse_core_siblings(cpu, true);
735 set_numa_node(numa_cpu_lookup_table[cpu]);
736 set_numa_mem(local_memory_node(numa_cpu_lookup_table[cpu]));
739 notify_cpu_starting(cpu);
740 set_cpu_online(cpu, true);
744 cpu_startup_entry(CPUHP_ONLINE);
749 int setup_profiling_timer(unsigned int multiplier)
754 #ifdef CONFIG_SCHED_SMT
755 /* cpumask of CPUs with asymetric SMT dependancy */
756 static int powerpc_smt_flags(void)
758 int flags = SD_SHARE_CPUCAPACITY | SD_SHARE_PKG_RESOURCES;
760 if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
761 printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
762 flags |= SD_ASYM_PACKING;
768 static struct sched_domain_topology_level powerpc_topology[] = {
769 #ifdef CONFIG_SCHED_SMT
770 { cpu_smt_mask, powerpc_smt_flags, SD_INIT_NAME(SMT) },
772 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
776 void __init smp_cpus_done(unsigned int max_cpus)
778 cpumask_var_t old_mask;
780 /* We want the setup_cpu() here to be called from CPU 0, but our
781 * init thread may have been "borrowed" by another CPU in the meantime
782 * se we pin us down to CPU 0 for a short while
784 alloc_cpumask_var(&old_mask, GFP_NOWAIT);
785 cpumask_copy(old_mask, tsk_cpus_allowed(current));
786 set_cpus_allowed_ptr(current, cpumask_of(boot_cpuid));
788 if (smp_ops && smp_ops->setup_cpu)
789 smp_ops->setup_cpu(boot_cpuid);
791 set_cpus_allowed_ptr(current, old_mask);
793 free_cpumask_var(old_mask);
795 if (smp_ops && smp_ops->bringup_done)
796 smp_ops->bringup_done();
798 dump_numa_cpu_topology();
800 set_sched_topology(powerpc_topology);
804 #ifdef CONFIG_HOTPLUG_CPU
805 int __cpu_disable(void)
807 int cpu = smp_processor_id();
811 if (!smp_ops->cpu_disable)
814 err = smp_ops->cpu_disable();
818 /* Update sibling maps */
819 base = cpu_first_thread_sibling(cpu);
820 for (i = 0; i < threads_per_core; i++) {
821 cpumask_clear_cpu(cpu, cpu_sibling_mask(base + i));
822 cpumask_clear_cpu(base + i, cpu_sibling_mask(cpu));
823 cpumask_clear_cpu(cpu, cpu_core_mask(base + i));
824 cpumask_clear_cpu(base + i, cpu_core_mask(cpu));
826 traverse_core_siblings(cpu, false);
831 void __cpu_die(unsigned int cpu)
833 if (smp_ops->cpu_die)
834 smp_ops->cpu_die(cpu);
842 /* If we return, we re-enter start_secondary */
843 start_secondary_resume();