kvm tools, setup: Create private directory
[linux-2.6/next.git] / arch / powerpc / kernel / smp.c
blob8ebc6700b98d18769f6e8afd1eb54f9f0363ec58
1 /*
2 * SMP support for ppc.
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.
18 #undef DEBUG
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>
37 #include <asm/irq.h>
38 #include <asm/page.h>
39 #include <asm/pgtable.h>
40 #include <asm/prom.h>
41 #include <asm/smp.h>
42 #include <asm/time.h>
43 #include <asm/machdep.h>
44 #include <asm/cputhreads.h>
45 #include <asm/cputable.h>
46 #include <asm/system.h>
47 #include <asm/mpic.h>
48 #include <asm/vdso_datapage.h>
49 #ifdef CONFIG_PPC64
50 #include <asm/paca.h>
51 #endif
53 #ifdef DEBUG
54 #include <asm/udbg.h>
55 #define DBG(fmt...) udbg_printf(fmt)
56 #else
57 #define DBG(fmt...)
58 #endif
61 /* Store all idle threads, this can be reused instead of creating
62 * a new thread. Also avoids complicated thread destroy functionality
63 * for idle threads.
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))
73 #else
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))
77 #endif
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;
97 #ifdef CONFIG_PPC64
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;
108 smp_mb();
110 return 0;
112 #endif
114 static irqreturn_t call_function_action(int irq, void *data)
116 generic_smp_call_function_interrupt();
117 return IRQ_HANDLED;
120 static irqreturn_t reschedule_action(int irq, void *data)
122 scheduler_ipi();
123 return IRQ_HANDLED;
126 static irqreturn_t call_function_single_action(int irq, void *data)
128 generic_smp_call_function_single_interrupt();
129 return IRQ_HANDLED;
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());
136 return IRQ_HANDLED;
139 #ifdef CONFIG_DEBUGGER
140 debugger_ipi(get_irq_regs());
141 #endif /* CONFIG_DEBUGGER */
143 return IRQ_HANDLED;
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)
163 int err;
165 if (msg < 0 || msg > PPC_MSG_DEBUGGER_BREAK) {
166 return -EINVAL;
168 #if !defined(CONFIG_DEBUGGER) && !defined(CONFIG_KEXEC)
169 if (msg == PPC_MSG_DEBUGGER_BREAK) {
170 return 1;
172 #endif
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);
178 return 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);
192 info->data = data;
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;
200 message[msg] = 1;
201 mb();
202 smp_ops->cause_ipi(cpu, info->data);
205 void smp_muxed_ipi_resend(void)
207 struct cpu_messages *info = &__get_cpu_var(ipi_message);
209 if (info->messages)
210 smp_ops->cause_ipi(smp_processor_id(), info->data);
213 irqreturn_t smp_ipi_demux(void)
215 struct cpu_messages *info = &__get_cpu_var(ipi_message);
216 unsigned int all;
218 mb(); /* order any irq clear */
220 do {
221 all = xchg_local(&info->messages, 0);
223 #ifdef __BIG_ENDIAN
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)))
227 scheduler_ipi();
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);
232 #else
233 #error Unsupported ENDIAN
234 #endif
235 } while (info->messages);
237 return IRQ_HANDLED;
239 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
241 void smp_send_reschedule(int cpu)
243 if (likely(smp_ops))
244 smp_ops->message_pass(cpu, PPC_MSG_RESCHEDULE);
247 void arch_send_call_function_single_ipi(int cpu)
249 smp_ops->message_pass(cpu, PPC_MSG_CALL_FUNC_SINGLE);
252 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
254 unsigned int cpu;
256 for_each_cpu(cpu, mask)
257 smp_ops->message_pass(cpu, PPC_MSG_CALL_FUNCTION);
260 #if defined(CONFIG_DEBUGGER) || defined(CONFIG_KEXEC)
261 void smp_send_debugger_break(void)
263 int cpu;
264 int me = raw_smp_processor_id();
266 if (unlikely(!smp_ops))
267 return;
269 for_each_online_cpu(cpu)
270 if (cpu != me)
271 smp_ops->message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
273 #endif
275 #ifdef CONFIG_KEXEC
276 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
278 crash_ipi_function_ptr = crash_ipi_callback;
279 if (crash_ipi_callback) {
280 mb();
281 smp_send_debugger_break();
284 #endif
286 static void stop_this_cpu(void *dummy)
288 /* Remove this CPU */
289 set_cpu_online(smp_processor_id(), false);
291 local_irq_disable();
292 while (1)
296 void smp_send_stop(void)
298 smp_call_function(stop_this_cpu, NULL, 0);
301 struct thread_info *current_set[NR_CPUS];
303 static void __devinit smp_store_cpu_info(int id)
305 per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
308 void __init smp_prepare_cpus(unsigned int max_cpus)
310 unsigned int cpu;
312 DBG("smp_prepare_cpus\n");
315 * setup_cpu may need to be called on the boot cpu. We havent
316 * spun any cpus up but lets be paranoid.
318 BUG_ON(boot_cpuid != smp_processor_id());
320 /* Fixup boot cpu */
321 smp_store_cpu_info(boot_cpuid);
322 cpu_callin_map[boot_cpuid] = 1;
324 for_each_possible_cpu(cpu) {
325 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
326 GFP_KERNEL, cpu_to_node(cpu));
327 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
328 GFP_KERNEL, cpu_to_node(cpu));
331 cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
332 cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
334 if (smp_ops)
335 if (smp_ops->probe)
336 max_cpus = smp_ops->probe();
337 else
338 max_cpus = NR_CPUS;
339 else
340 max_cpus = 1;
343 void __devinit smp_prepare_boot_cpu(void)
345 BUG_ON(smp_processor_id() != boot_cpuid);
346 #ifdef CONFIG_PPC64
347 paca[boot_cpuid].__current = current;
348 #endif
349 current_set[boot_cpuid] = task_thread_info(current);
352 #ifdef CONFIG_HOTPLUG_CPU
353 /* State of each CPU during hotplug phases */
354 static DEFINE_PER_CPU(int, cpu_state) = { 0 };
356 int generic_cpu_disable(void)
358 unsigned int cpu = smp_processor_id();
360 if (cpu == boot_cpuid)
361 return -EBUSY;
363 set_cpu_online(cpu, false);
364 #ifdef CONFIG_PPC64
365 vdso_data->processorCount--;
366 #endif
367 migrate_irqs();
368 return 0;
371 void generic_cpu_die(unsigned int cpu)
373 int i;
375 for (i = 0; i < 100; i++) {
376 smp_rmb();
377 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
378 return;
379 msleep(100);
381 printk(KERN_ERR "CPU%d didn't die...\n", cpu);
384 void generic_mach_cpu_die(void)
386 unsigned int cpu;
388 local_irq_disable();
389 idle_task_exit();
390 cpu = smp_processor_id();
391 printk(KERN_DEBUG "CPU%d offline\n", cpu);
392 __get_cpu_var(cpu_state) = CPU_DEAD;
393 smp_wmb();
394 while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
395 cpu_relax();
398 void generic_set_cpu_dead(unsigned int cpu)
400 per_cpu(cpu_state, cpu) = CPU_DEAD;
402 #endif
404 struct create_idle {
405 struct work_struct work;
406 struct task_struct *idle;
407 struct completion done;
408 int cpu;
411 static void __cpuinit do_fork_idle(struct work_struct *work)
413 struct create_idle *c_idle =
414 container_of(work, struct create_idle, work);
416 c_idle->idle = fork_idle(c_idle->cpu);
417 complete(&c_idle->done);
420 static int __cpuinit create_idle(unsigned int cpu)
422 struct thread_info *ti;
423 struct create_idle c_idle = {
424 .cpu = cpu,
425 .done = COMPLETION_INITIALIZER_ONSTACK(c_idle.done),
427 INIT_WORK_ONSTACK(&c_idle.work, do_fork_idle);
429 c_idle.idle = get_idle_for_cpu(cpu);
431 /* We can't use kernel_thread since we must avoid to
432 * reschedule the child. We use a workqueue because
433 * we want to fork from a kernel thread, not whatever
434 * userspace process happens to be trying to online us.
436 if (!c_idle.idle) {
437 schedule_work(&c_idle.work);
438 wait_for_completion(&c_idle.done);
439 } else
440 init_idle(c_idle.idle, cpu);
441 if (IS_ERR(c_idle.idle)) {
442 pr_err("Failed fork for CPU %u: %li", cpu, PTR_ERR(c_idle.idle));
443 return PTR_ERR(c_idle.idle);
445 ti = task_thread_info(c_idle.idle);
447 #ifdef CONFIG_PPC64
448 paca[cpu].__current = c_idle.idle;
449 paca[cpu].kstack = (unsigned long)ti + THREAD_SIZE - STACK_FRAME_OVERHEAD;
450 #endif
451 ti->cpu = cpu;
452 current_set[cpu] = ti;
454 return 0;
457 int __cpuinit __cpu_up(unsigned int cpu)
459 int rc, c;
461 if (smp_ops == NULL ||
462 (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
463 return -EINVAL;
465 /* Make sure we have an idle thread */
466 rc = create_idle(cpu);
467 if (rc)
468 return rc;
470 secondary_ti = current_set[cpu];
472 /* Make sure callin-map entry is 0 (can be leftover a CPU
473 * hotplug
475 cpu_callin_map[cpu] = 0;
477 /* The information for processor bringup must
478 * be written out to main store before we release
479 * the processor.
481 smp_mb();
483 /* wake up cpus */
484 DBG("smp: kicking cpu %d\n", cpu);
485 rc = smp_ops->kick_cpu(cpu);
486 if (rc) {
487 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
488 return rc;
492 * wait to see if the cpu made a callin (is actually up).
493 * use this value that I found through experimentation.
494 * -- Cort
496 if (system_state < SYSTEM_RUNNING)
497 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
498 udelay(100);
499 #ifdef CONFIG_HOTPLUG_CPU
500 else
502 * CPUs can take much longer to come up in the
503 * hotplug case. Wait five seconds.
505 for (c = 5000; c && !cpu_callin_map[cpu]; c--)
506 msleep(1);
507 #endif
509 if (!cpu_callin_map[cpu]) {
510 printk(KERN_ERR "Processor %u is stuck.\n", cpu);
511 return -ENOENT;
514 DBG("Processor %u found.\n", cpu);
516 if (smp_ops->give_timebase)
517 smp_ops->give_timebase();
519 /* Wait until cpu puts itself in the online map */
520 while (!cpu_online(cpu))
521 cpu_relax();
523 return 0;
526 /* Return the value of the reg property corresponding to the given
527 * logical cpu.
529 int cpu_to_core_id(int cpu)
531 struct device_node *np;
532 const int *reg;
533 int id = -1;
535 np = of_get_cpu_node(cpu, NULL);
536 if (!np)
537 goto out;
539 reg = of_get_property(np, "reg", NULL);
540 if (!reg)
541 goto out;
543 id = *reg;
544 out:
545 of_node_put(np);
546 return id;
549 /* Helper routines for cpu to core mapping */
550 int cpu_core_index_of_thread(int cpu)
552 return cpu >> threads_shift;
554 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
556 int cpu_first_thread_of_core(int core)
558 return core << threads_shift;
560 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
562 /* Must be called when no change can occur to cpu_present_mask,
563 * i.e. during cpu online or offline.
565 static struct device_node *cpu_to_l2cache(int cpu)
567 struct device_node *np;
568 struct device_node *cache;
570 if (!cpu_present(cpu))
571 return NULL;
573 np = of_get_cpu_node(cpu, NULL);
574 if (np == NULL)
575 return NULL;
577 cache = of_find_next_cache_node(np);
579 of_node_put(np);
581 return cache;
584 /* Activate a secondary processor. */
585 void __devinit start_secondary(void *unused)
587 unsigned int cpu = smp_processor_id();
588 struct device_node *l2_cache;
589 int i, base;
591 atomic_inc(&init_mm.mm_count);
592 current->active_mm = &init_mm;
594 smp_store_cpu_info(cpu);
595 set_dec(tb_ticks_per_jiffy);
596 preempt_disable();
597 cpu_callin_map[cpu] = 1;
599 if (smp_ops->setup_cpu)
600 smp_ops->setup_cpu(cpu);
601 if (smp_ops->take_timebase)
602 smp_ops->take_timebase();
604 secondary_cpu_time_init();
606 #ifdef CONFIG_PPC64
607 if (system_state == SYSTEM_RUNNING)
608 vdso_data->processorCount++;
609 #endif
610 ipi_call_lock();
611 notify_cpu_starting(cpu);
612 set_cpu_online(cpu, true);
613 /* Update sibling maps */
614 base = cpu_first_thread_sibling(cpu);
615 for (i = 0; i < threads_per_core; i++) {
616 if (cpu_is_offline(base + i))
617 continue;
618 cpumask_set_cpu(cpu, cpu_sibling_mask(base + i));
619 cpumask_set_cpu(base + i, cpu_sibling_mask(cpu));
621 /* cpu_core_map should be a superset of
622 * cpu_sibling_map even if we don't have cache
623 * information, so update the former here, too.
625 cpumask_set_cpu(cpu, cpu_core_mask(base + i));
626 cpumask_set_cpu(base + i, cpu_core_mask(cpu));
628 l2_cache = cpu_to_l2cache(cpu);
629 for_each_online_cpu(i) {
630 struct device_node *np = cpu_to_l2cache(i);
631 if (!np)
632 continue;
633 if (np == l2_cache) {
634 cpumask_set_cpu(cpu, cpu_core_mask(i));
635 cpumask_set_cpu(i, cpu_core_mask(cpu));
637 of_node_put(np);
639 of_node_put(l2_cache);
640 ipi_call_unlock();
642 local_irq_enable();
644 cpu_idle();
646 BUG();
649 int setup_profiling_timer(unsigned int multiplier)
651 return 0;
654 void __init smp_cpus_done(unsigned int max_cpus)
656 cpumask_var_t old_mask;
658 /* We want the setup_cpu() here to be called from CPU 0, but our
659 * init thread may have been "borrowed" by another CPU in the meantime
660 * se we pin us down to CPU 0 for a short while
662 alloc_cpumask_var(&old_mask, GFP_NOWAIT);
663 cpumask_copy(old_mask, tsk_cpus_allowed(current));
664 set_cpus_allowed_ptr(current, cpumask_of(boot_cpuid));
666 if (smp_ops && smp_ops->setup_cpu)
667 smp_ops->setup_cpu(boot_cpuid);
669 set_cpus_allowed_ptr(current, old_mask);
671 free_cpumask_var(old_mask);
673 if (smp_ops && smp_ops->bringup_done)
674 smp_ops->bringup_done();
676 dump_numa_cpu_topology();
680 int arch_sd_sibling_asym_packing(void)
682 if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
683 printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
684 return SD_ASYM_PACKING;
686 return 0;
689 #ifdef CONFIG_HOTPLUG_CPU
690 int __cpu_disable(void)
692 struct device_node *l2_cache;
693 int cpu = smp_processor_id();
694 int base, i;
695 int err;
697 if (!smp_ops->cpu_disable)
698 return -ENOSYS;
700 err = smp_ops->cpu_disable();
701 if (err)
702 return err;
704 /* Update sibling maps */
705 base = cpu_first_thread_sibling(cpu);
706 for (i = 0; i < threads_per_core; i++) {
707 cpumask_clear_cpu(cpu, cpu_sibling_mask(base + i));
708 cpumask_clear_cpu(base + i, cpu_sibling_mask(cpu));
709 cpumask_clear_cpu(cpu, cpu_core_mask(base + i));
710 cpumask_clear_cpu(base + i, cpu_core_mask(cpu));
713 l2_cache = cpu_to_l2cache(cpu);
714 for_each_present_cpu(i) {
715 struct device_node *np = cpu_to_l2cache(i);
716 if (!np)
717 continue;
718 if (np == l2_cache) {
719 cpumask_clear_cpu(cpu, cpu_core_mask(i));
720 cpumask_clear_cpu(i, cpu_core_mask(cpu));
722 of_node_put(np);
724 of_node_put(l2_cache);
727 return 0;
730 void __cpu_die(unsigned int cpu)
732 if (smp_ops->cpu_die)
733 smp_ops->cpu_die(cpu);
736 static DEFINE_MUTEX(powerpc_cpu_hotplug_driver_mutex);
738 void cpu_hotplug_driver_lock()
740 mutex_lock(&powerpc_cpu_hotplug_driver_mutex);
743 void cpu_hotplug_driver_unlock()
745 mutex_unlock(&powerpc_cpu_hotplug_driver_mutex);
748 void cpu_die(void)
750 if (ppc_md.cpu_die)
751 ppc_md.cpu_die();
753 /* If we return, we re-enter start_secondary */
754 start_secondary_resume();
757 #endif