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/module.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/sysdev.h>
31 #include <linux/cpu.h>
32 #include <linux/notifier.h>
33 #include <linux/topology.h>
35 #include <asm/ptrace.h>
36 #include <asm/atomic.h>
39 #include <asm/pgtable.h>
43 #include <asm/machdep.h>
44 #include <asm/cputhreads.h>
45 #include <asm/cputable.h>
46 #include <asm/system.h>
48 #include <asm/vdso_datapage.h>
55 #define DBG(fmt...) udbg_printf(fmt)
61 /* Store all idle threads, this can be reused instead of creating
62 * a new thread. Also avoids complicated thread destroy functionality
65 #ifdef CONFIG_HOTPLUG_CPU
67 * Needed only for CONFIG_HOTPLUG_CPU because __cpuinitdata is
68 * removed after init for !CONFIG_HOTPLUG_CPU.
70 static DEFINE_PER_CPU(struct task_struct *, idle_thread_array);
71 #define get_idle_for_cpu(x) (per_cpu(idle_thread_array, x))
72 #define set_idle_for_cpu(x, p) (per_cpu(idle_thread_array, x) = (p))
74 static struct task_struct *idle_thread_array[NR_CPUS] __cpuinitdata ;
75 #define get_idle_for_cpu(x) (idle_thread_array[(x)])
76 #define set_idle_for_cpu(x, p) (idle_thread_array[(x)] = (p))
79 struct thread_info *secondary_ti;
81 DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
82 DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
84 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
85 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
87 /* SMP operations for this machine */
88 struct smp_ops_t *smp_ops;
90 /* Can't be static due to PowerMac hackery */
91 volatile unsigned int cpu_callin_map[NR_CPUS];
93 int smt_enabled_at_boot = 1;
95 static void (*crash_ipi_function_ptr)(struct pt_regs *) = NULL;
98 int __devinit smp_generic_kick_cpu(int nr)
100 BUG_ON(nr < 0 || nr >= NR_CPUS);
103 * The processor is currently spinning, waiting for the
104 * cpu_start field to become non-zero After we set cpu_start,
105 * the processor will continue on to secondary_start
107 paca[nr].cpu_start = 1;
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], IRQF_DISABLED|IRQF_PERCPU,
174 smp_ipi_name[msg], 0);
175 WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n",
176 virq, smp_ipi_name[msg], err);
181 #ifdef CONFIG_PPC_SMP_MUXED_IPI
182 struct cpu_messages {
183 int messages; /* current messages */
184 unsigned long data; /* data for cause ipi */
186 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages, ipi_message);
188 void smp_muxed_ipi_set_data(int cpu, unsigned long data)
190 struct cpu_messages *info = &per_cpu(ipi_message, cpu);
195 void smp_muxed_ipi_message_pass(int cpu, int msg)
197 struct cpu_messages *info = &per_cpu(ipi_message, cpu);
198 char *message = (char *)&info->messages;
202 smp_ops->cause_ipi(cpu, info->data);
205 irqreturn_t smp_ipi_demux(void)
207 struct cpu_messages *info = &__get_cpu_var(ipi_message);
210 mb(); /* order any irq clear */
213 all = xchg_local(&info->messages, 0);
216 if (all & (1 << (24 - 8 * PPC_MSG_CALL_FUNCTION)))
217 generic_smp_call_function_interrupt();
218 if (all & (1 << (24 - 8 * PPC_MSG_RESCHEDULE)))
220 if (all & (1 << (24 - 8 * PPC_MSG_CALL_FUNC_SINGLE)))
221 generic_smp_call_function_single_interrupt();
222 if (all & (1 << (24 - 8 * PPC_MSG_DEBUGGER_BREAK)))
223 debug_ipi_action(0, NULL);
225 #error Unsupported ENDIAN
227 } while (info->messages);
231 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
233 static inline void do_message_pass(int cpu, int msg)
235 if (smp_ops->message_pass)
236 smp_ops->message_pass(cpu, msg);
237 #ifdef CONFIG_PPC_SMP_MUXED_IPI
239 smp_muxed_ipi_message_pass(cpu, msg);
243 void smp_send_reschedule(int cpu)
246 do_message_pass(cpu, PPC_MSG_RESCHEDULE);
249 void arch_send_call_function_single_ipi(int cpu)
251 do_message_pass(cpu, PPC_MSG_CALL_FUNC_SINGLE);
254 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
258 for_each_cpu(cpu, mask)
259 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
262 #if defined(CONFIG_DEBUGGER) || defined(CONFIG_KEXEC)
263 void smp_send_debugger_break(void)
266 int me = raw_smp_processor_id();
268 if (unlikely(!smp_ops))
271 for_each_online_cpu(cpu)
273 do_message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
278 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
280 crash_ipi_function_ptr = crash_ipi_callback;
281 if (crash_ipi_callback) {
283 smp_send_debugger_break();
288 static void stop_this_cpu(void *dummy)
290 /* Remove this CPU */
291 set_cpu_online(smp_processor_id(), false);
298 void smp_send_stop(void)
300 smp_call_function(stop_this_cpu, NULL, 0);
303 struct thread_info *current_set[NR_CPUS];
305 static void __devinit smp_store_cpu_info(int id)
307 per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
308 #ifdef CONFIG_PPC_FSL_BOOK3E
309 per_cpu(next_tlbcam_idx, id)
310 = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) - 1;
314 void __init smp_prepare_cpus(unsigned int max_cpus)
318 DBG("smp_prepare_cpus\n");
321 * setup_cpu may need to be called on the boot cpu. We havent
322 * spun any cpus up but lets be paranoid.
324 BUG_ON(boot_cpuid != smp_processor_id());
327 smp_store_cpu_info(boot_cpuid);
328 cpu_callin_map[boot_cpuid] = 1;
330 for_each_possible_cpu(cpu) {
331 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
332 GFP_KERNEL, cpu_to_node(cpu));
333 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
334 GFP_KERNEL, cpu_to_node(cpu));
337 cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
338 cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
342 max_cpus = smp_ops->probe();
349 void __devinit smp_prepare_boot_cpu(void)
351 BUG_ON(smp_processor_id() != boot_cpuid);
353 paca[boot_cpuid].__current = current;
355 current_set[boot_cpuid] = task_thread_info(current);
358 #ifdef CONFIG_HOTPLUG_CPU
359 /* State of each CPU during hotplug phases */
360 static DEFINE_PER_CPU(int, cpu_state) = { 0 };
362 int generic_cpu_disable(void)
364 unsigned int cpu = smp_processor_id();
366 if (cpu == boot_cpuid)
369 set_cpu_online(cpu, false);
371 vdso_data->processorCount--;
377 void generic_cpu_die(unsigned int cpu)
381 for (i = 0; i < 100; i++) {
383 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
387 printk(KERN_ERR "CPU%d didn't die...\n", cpu);
390 void generic_mach_cpu_die(void)
396 cpu = smp_processor_id();
397 printk(KERN_DEBUG "CPU%d offline\n", cpu);
398 __get_cpu_var(cpu_state) = CPU_DEAD;
400 while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
404 void generic_set_cpu_dead(unsigned int cpu)
406 per_cpu(cpu_state, cpu) = CPU_DEAD;
411 struct work_struct work;
412 struct task_struct *idle;
413 struct completion done;
417 static void __cpuinit do_fork_idle(struct work_struct *work)
419 struct create_idle *c_idle =
420 container_of(work, struct create_idle, work);
422 c_idle->idle = fork_idle(c_idle->cpu);
423 complete(&c_idle->done);
426 static int __cpuinit create_idle(unsigned int cpu)
428 struct thread_info *ti;
429 struct create_idle c_idle = {
431 .done = COMPLETION_INITIALIZER_ONSTACK(c_idle.done),
433 INIT_WORK_ONSTACK(&c_idle.work, do_fork_idle);
435 c_idle.idle = get_idle_for_cpu(cpu);
437 /* We can't use kernel_thread since we must avoid to
438 * reschedule the child. We use a workqueue because
439 * we want to fork from a kernel thread, not whatever
440 * userspace process happens to be trying to online us.
443 schedule_work(&c_idle.work);
444 wait_for_completion(&c_idle.done);
446 init_idle(c_idle.idle, cpu);
447 if (IS_ERR(c_idle.idle)) {
448 pr_err("Failed fork for CPU %u: %li", cpu, PTR_ERR(c_idle.idle));
449 return PTR_ERR(c_idle.idle);
451 ti = task_thread_info(c_idle.idle);
454 paca[cpu].__current = c_idle.idle;
455 paca[cpu].kstack = (unsigned long)ti + THREAD_SIZE - STACK_FRAME_OVERHEAD;
458 current_set[cpu] = ti;
463 int __cpuinit __cpu_up(unsigned int cpu)
467 if (smp_ops == NULL ||
468 (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
471 /* Make sure we have an idle thread */
472 rc = create_idle(cpu);
476 secondary_ti = current_set[cpu];
478 /* Make sure callin-map entry is 0 (can be leftover a CPU
481 cpu_callin_map[cpu] = 0;
483 /* The information for processor bringup must
484 * be written out to main store before we release
490 DBG("smp: kicking cpu %d\n", cpu);
491 rc = smp_ops->kick_cpu(cpu);
493 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
498 * wait to see if the cpu made a callin (is actually up).
499 * use this value that I found through experimentation.
502 if (system_state < SYSTEM_RUNNING)
503 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
505 #ifdef CONFIG_HOTPLUG_CPU
508 * CPUs can take much longer to come up in the
509 * hotplug case. Wait five seconds.
511 for (c = 5000; c && !cpu_callin_map[cpu]; c--)
515 if (!cpu_callin_map[cpu]) {
516 printk(KERN_ERR "Processor %u is stuck.\n", cpu);
520 DBG("Processor %u found.\n", cpu);
522 if (smp_ops->give_timebase)
523 smp_ops->give_timebase();
525 /* Wait until cpu puts itself in the online map */
526 while (!cpu_online(cpu))
532 /* Return the value of the reg property corresponding to the given
535 int cpu_to_core_id(int cpu)
537 struct device_node *np;
541 np = of_get_cpu_node(cpu, NULL);
545 reg = of_get_property(np, "reg", NULL);
555 /* Helper routines for cpu to core mapping */
556 int cpu_core_index_of_thread(int cpu)
558 return cpu >> threads_shift;
560 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
562 int cpu_first_thread_of_core(int core)
564 return core << threads_shift;
566 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
568 /* Must be called when no change can occur to cpu_present_mask,
569 * i.e. during cpu online or offline.
571 static struct device_node *cpu_to_l2cache(int cpu)
573 struct device_node *np;
574 struct device_node *cache;
576 if (!cpu_present(cpu))
579 np = of_get_cpu_node(cpu, NULL);
583 cache = of_find_next_cache_node(np);
590 /* Activate a secondary processor. */
591 void __devinit start_secondary(void *unused)
593 unsigned int cpu = smp_processor_id();
594 struct device_node *l2_cache;
597 atomic_inc(&init_mm.mm_count);
598 current->active_mm = &init_mm;
600 smp_store_cpu_info(cpu);
601 set_dec(tb_ticks_per_jiffy);
603 cpu_callin_map[cpu] = 1;
605 if (smp_ops->setup_cpu)
606 smp_ops->setup_cpu(cpu);
607 if (smp_ops->take_timebase)
608 smp_ops->take_timebase();
610 secondary_cpu_time_init();
613 if (system_state == SYSTEM_RUNNING)
614 vdso_data->processorCount++;
617 notify_cpu_starting(cpu);
618 set_cpu_online(cpu, true);
619 /* Update sibling maps */
620 base = cpu_first_thread_sibling(cpu);
621 for (i = 0; i < threads_per_core; i++) {
622 if (cpu_is_offline(base + i))
624 cpumask_set_cpu(cpu, cpu_sibling_mask(base + i));
625 cpumask_set_cpu(base + i, cpu_sibling_mask(cpu));
627 /* cpu_core_map should be a superset of
628 * cpu_sibling_map even if we don't have cache
629 * information, so update the former here, too.
631 cpumask_set_cpu(cpu, cpu_core_mask(base + i));
632 cpumask_set_cpu(base + i, cpu_core_mask(cpu));
634 l2_cache = cpu_to_l2cache(cpu);
635 for_each_online_cpu(i) {
636 struct device_node *np = cpu_to_l2cache(i);
639 if (np == l2_cache) {
640 cpumask_set_cpu(cpu, cpu_core_mask(i));
641 cpumask_set_cpu(i, cpu_core_mask(cpu));
645 of_node_put(l2_cache);
655 int setup_profiling_timer(unsigned int multiplier)
660 void __init smp_cpus_done(unsigned int max_cpus)
662 cpumask_var_t old_mask;
664 /* We want the setup_cpu() here to be called from CPU 0, but our
665 * init thread may have been "borrowed" by another CPU in the meantime
666 * se we pin us down to CPU 0 for a short while
668 alloc_cpumask_var(&old_mask, GFP_NOWAIT);
669 cpumask_copy(old_mask, tsk_cpus_allowed(current));
670 set_cpus_allowed_ptr(current, cpumask_of(boot_cpuid));
672 if (smp_ops && smp_ops->setup_cpu)
673 smp_ops->setup_cpu(boot_cpuid);
675 set_cpus_allowed_ptr(current, old_mask);
677 free_cpumask_var(old_mask);
679 if (smp_ops && smp_ops->bringup_done)
680 smp_ops->bringup_done();
682 dump_numa_cpu_topology();
686 int arch_sd_sibling_asym_packing(void)
688 if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
689 printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
690 return SD_ASYM_PACKING;
695 #ifdef CONFIG_HOTPLUG_CPU
696 int __cpu_disable(void)
698 struct device_node *l2_cache;
699 int cpu = smp_processor_id();
703 if (!smp_ops->cpu_disable)
706 err = smp_ops->cpu_disable();
710 /* Update sibling maps */
711 base = cpu_first_thread_sibling(cpu);
712 for (i = 0; i < threads_per_core; i++) {
713 cpumask_clear_cpu(cpu, cpu_sibling_mask(base + i));
714 cpumask_clear_cpu(base + i, cpu_sibling_mask(cpu));
715 cpumask_clear_cpu(cpu, cpu_core_mask(base + i));
716 cpumask_clear_cpu(base + i, cpu_core_mask(cpu));
719 l2_cache = cpu_to_l2cache(cpu);
720 for_each_present_cpu(i) {
721 struct device_node *np = cpu_to_l2cache(i);
724 if (np == l2_cache) {
725 cpumask_clear_cpu(cpu, cpu_core_mask(i));
726 cpumask_clear_cpu(i, cpu_core_mask(cpu));
730 of_node_put(l2_cache);
736 void __cpu_die(unsigned int cpu)
738 if (smp_ops->cpu_die)
739 smp_ops->cpu_die(cpu);
742 static DEFINE_MUTEX(powerpc_cpu_hotplug_driver_mutex);
744 void cpu_hotplug_driver_lock()
746 mutex_lock(&powerpc_cpu_hotplug_driver_mutex);
749 void cpu_hotplug_driver_unlock()
751 mutex_unlock(&powerpc_cpu_hotplug_driver_mutex);
759 /* If we return, we re-enter start_secondary */
760 start_secondary_resume();