Merge remote-tracking branch 'upstream/master' into kvm-devel
[linux-2.6/kvm.git] / arch / powerpc / kernel / smp.c
blob7bf2187dfd99e3135f78225a3217d53a3a099b6e
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 <linux/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 irqreturn_t smp_ipi_demux(void)
207 struct cpu_messages *info = &__get_cpu_var(ipi_message);
208 unsigned int all;
210 mb(); /* order any irq clear */
212 do {
213 all = xchg_local(&info->messages, 0);
215 #ifdef __BIG_ENDIAN
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)))
219 scheduler_ipi();
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);
224 #else
225 #error Unsupported ENDIAN
226 #endif
227 } while (info->messages);
229 return IRQ_HANDLED;
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
238 else
239 smp_muxed_ipi_message_pass(cpu, msg);
240 #endif
243 void smp_send_reschedule(int cpu)
245 if (likely(smp_ops))
246 do_message_pass(cpu, PPC_MSG_RESCHEDULE);
248 EXPORT_SYMBOL_GPL(smp_send_reschedule);
250 void arch_send_call_function_single_ipi(int cpu)
252 do_message_pass(cpu, PPC_MSG_CALL_FUNC_SINGLE);
255 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
257 unsigned int cpu;
259 for_each_cpu(cpu, mask)
260 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
263 #if defined(CONFIG_DEBUGGER) || defined(CONFIG_KEXEC)
264 void smp_send_debugger_break(void)
266 int cpu;
267 int me = raw_smp_processor_id();
269 if (unlikely(!smp_ops))
270 return;
272 for_each_online_cpu(cpu)
273 if (cpu != me)
274 do_message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
276 #endif
278 #ifdef CONFIG_KEXEC
279 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
281 crash_ipi_function_ptr = crash_ipi_callback;
282 if (crash_ipi_callback) {
283 mb();
284 smp_send_debugger_break();
287 #endif
289 static void stop_this_cpu(void *dummy)
291 /* Remove this CPU */
292 set_cpu_online(smp_processor_id(), false);
294 local_irq_disable();
295 while (1)
299 void smp_send_stop(void)
301 smp_call_function(stop_this_cpu, NULL, 0);
304 struct thread_info *current_set[NR_CPUS];
306 static void __devinit smp_store_cpu_info(int id)
308 per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
309 #ifdef CONFIG_PPC_FSL_BOOK3E
310 per_cpu(next_tlbcam_idx, id)
311 = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) - 1;
312 #endif
315 void __init smp_prepare_cpus(unsigned int max_cpus)
317 unsigned int cpu;
319 DBG("smp_prepare_cpus\n");
322 * setup_cpu may need to be called on the boot cpu. We havent
323 * spun any cpus up but lets be paranoid.
325 BUG_ON(boot_cpuid != smp_processor_id());
327 /* Fixup boot cpu */
328 smp_store_cpu_info(boot_cpuid);
329 cpu_callin_map[boot_cpuid] = 1;
331 for_each_possible_cpu(cpu) {
332 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
333 GFP_KERNEL, cpu_to_node(cpu));
334 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
335 GFP_KERNEL, cpu_to_node(cpu));
338 cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
339 cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
341 if (smp_ops)
342 if (smp_ops->probe)
343 max_cpus = smp_ops->probe();
344 else
345 max_cpus = NR_CPUS;
346 else
347 max_cpus = 1;
350 void __devinit smp_prepare_boot_cpu(void)
352 BUG_ON(smp_processor_id() != boot_cpuid);
353 #ifdef CONFIG_PPC64
354 paca[boot_cpuid].__current = current;
355 #endif
356 current_set[boot_cpuid] = task_thread_info(current);
359 #ifdef CONFIG_HOTPLUG_CPU
360 /* State of each CPU during hotplug phases */
361 static DEFINE_PER_CPU(int, cpu_state) = { 0 };
363 int generic_cpu_disable(void)
365 unsigned int cpu = smp_processor_id();
367 if (cpu == boot_cpuid)
368 return -EBUSY;
370 set_cpu_online(cpu, false);
371 #ifdef CONFIG_PPC64
372 vdso_data->processorCount--;
373 #endif
374 migrate_irqs();
375 return 0;
378 void generic_cpu_die(unsigned int cpu)
380 int i;
382 for (i = 0; i < 100; i++) {
383 smp_rmb();
384 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
385 return;
386 msleep(100);
388 printk(KERN_ERR "CPU%d didn't die...\n", cpu);
391 void generic_mach_cpu_die(void)
393 unsigned int cpu;
395 local_irq_disable();
396 idle_task_exit();
397 cpu = smp_processor_id();
398 printk(KERN_DEBUG "CPU%d offline\n", cpu);
399 __get_cpu_var(cpu_state) = CPU_DEAD;
400 smp_wmb();
401 while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
402 cpu_relax();
405 void generic_set_cpu_dead(unsigned int cpu)
407 per_cpu(cpu_state, cpu) = CPU_DEAD;
409 #endif
411 struct create_idle {
412 struct work_struct work;
413 struct task_struct *idle;
414 struct completion done;
415 int cpu;
418 static void __cpuinit do_fork_idle(struct work_struct *work)
420 struct create_idle *c_idle =
421 container_of(work, struct create_idle, work);
423 c_idle->idle = fork_idle(c_idle->cpu);
424 complete(&c_idle->done);
427 static int __cpuinit create_idle(unsigned int cpu)
429 struct thread_info *ti;
430 struct create_idle c_idle = {
431 .cpu = cpu,
432 .done = COMPLETION_INITIALIZER_ONSTACK(c_idle.done),
434 INIT_WORK_ONSTACK(&c_idle.work, do_fork_idle);
436 c_idle.idle = get_idle_for_cpu(cpu);
438 /* We can't use kernel_thread since we must avoid to
439 * reschedule the child. We use a workqueue because
440 * we want to fork from a kernel thread, not whatever
441 * userspace process happens to be trying to online us.
443 if (!c_idle.idle) {
444 schedule_work(&c_idle.work);
445 wait_for_completion(&c_idle.done);
446 } else
447 init_idle(c_idle.idle, cpu);
448 if (IS_ERR(c_idle.idle)) {
449 pr_err("Failed fork for CPU %u: %li", cpu, PTR_ERR(c_idle.idle));
450 return PTR_ERR(c_idle.idle);
452 ti = task_thread_info(c_idle.idle);
454 #ifdef CONFIG_PPC64
455 paca[cpu].__current = c_idle.idle;
456 paca[cpu].kstack = (unsigned long)ti + THREAD_SIZE - STACK_FRAME_OVERHEAD;
457 #endif
458 ti->cpu = cpu;
459 current_set[cpu] = ti;
461 return 0;
464 int __cpuinit __cpu_up(unsigned int cpu)
466 int rc, c;
468 if (smp_ops == NULL ||
469 (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
470 return -EINVAL;
472 /* Make sure we have an idle thread */
473 rc = create_idle(cpu);
474 if (rc)
475 return rc;
477 secondary_ti = current_set[cpu];
479 /* Make sure callin-map entry is 0 (can be leftover a CPU
480 * hotplug
482 cpu_callin_map[cpu] = 0;
484 /* The information for processor bringup must
485 * be written out to main store before we release
486 * the processor.
488 smp_mb();
490 /* wake up cpus */
491 DBG("smp: kicking cpu %d\n", cpu);
492 rc = smp_ops->kick_cpu(cpu);
493 if (rc) {
494 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
495 return rc;
499 * wait to see if the cpu made a callin (is actually up).
500 * use this value that I found through experimentation.
501 * -- Cort
503 if (system_state < SYSTEM_RUNNING)
504 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
505 udelay(100);
506 #ifdef CONFIG_HOTPLUG_CPU
507 else
509 * CPUs can take much longer to come up in the
510 * hotplug case. Wait five seconds.
512 for (c = 5000; c && !cpu_callin_map[cpu]; c--)
513 msleep(1);
514 #endif
516 if (!cpu_callin_map[cpu]) {
517 printk(KERN_ERR "Processor %u is stuck.\n", cpu);
518 return -ENOENT;
521 DBG("Processor %u found.\n", cpu);
523 if (smp_ops->give_timebase)
524 smp_ops->give_timebase();
526 /* Wait until cpu puts itself in the online map */
527 while (!cpu_online(cpu))
528 cpu_relax();
530 return 0;
533 /* Return the value of the reg property corresponding to the given
534 * logical cpu.
536 int cpu_to_core_id(int cpu)
538 struct device_node *np;
539 const int *reg;
540 int id = -1;
542 np = of_get_cpu_node(cpu, NULL);
543 if (!np)
544 goto out;
546 reg = of_get_property(np, "reg", NULL);
547 if (!reg)
548 goto out;
550 id = *reg;
551 out:
552 of_node_put(np);
553 return id;
556 /* Helper routines for cpu to core mapping */
557 int cpu_core_index_of_thread(int cpu)
559 return cpu >> threads_shift;
561 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
563 int cpu_first_thread_of_core(int core)
565 return core << threads_shift;
567 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
569 /* Must be called when no change can occur to cpu_present_mask,
570 * i.e. during cpu online or offline.
572 static struct device_node *cpu_to_l2cache(int cpu)
574 struct device_node *np;
575 struct device_node *cache;
577 if (!cpu_present(cpu))
578 return NULL;
580 np = of_get_cpu_node(cpu, NULL);
581 if (np == NULL)
582 return NULL;
584 cache = of_find_next_cache_node(np);
586 of_node_put(np);
588 return cache;
591 /* Activate a secondary processor. */
592 void __devinit start_secondary(void *unused)
594 unsigned int cpu = smp_processor_id();
595 struct device_node *l2_cache;
596 int i, base;
598 atomic_inc(&init_mm.mm_count);
599 current->active_mm = &init_mm;
601 smp_store_cpu_info(cpu);
602 set_dec(tb_ticks_per_jiffy);
603 preempt_disable();
604 cpu_callin_map[cpu] = 1;
606 if (smp_ops->setup_cpu)
607 smp_ops->setup_cpu(cpu);
608 if (smp_ops->take_timebase)
609 smp_ops->take_timebase();
611 secondary_cpu_time_init();
613 #ifdef CONFIG_PPC64
614 if (system_state == SYSTEM_RUNNING)
615 vdso_data->processorCount++;
616 #endif
617 ipi_call_lock();
618 notify_cpu_starting(cpu);
619 set_cpu_online(cpu, true);
620 /* Update sibling maps */
621 base = cpu_first_thread_sibling(cpu);
622 for (i = 0; i < threads_per_core; i++) {
623 if (cpu_is_offline(base + i))
624 continue;
625 cpumask_set_cpu(cpu, cpu_sibling_mask(base + i));
626 cpumask_set_cpu(base + i, cpu_sibling_mask(cpu));
628 /* cpu_core_map should be a superset of
629 * cpu_sibling_map even if we don't have cache
630 * information, so update the former here, too.
632 cpumask_set_cpu(cpu, cpu_core_mask(base + i));
633 cpumask_set_cpu(base + i, cpu_core_mask(cpu));
635 l2_cache = cpu_to_l2cache(cpu);
636 for_each_online_cpu(i) {
637 struct device_node *np = cpu_to_l2cache(i);
638 if (!np)
639 continue;
640 if (np == l2_cache) {
641 cpumask_set_cpu(cpu, cpu_core_mask(i));
642 cpumask_set_cpu(i, cpu_core_mask(cpu));
644 of_node_put(np);
646 of_node_put(l2_cache);
647 ipi_call_unlock();
649 local_irq_enable();
651 cpu_idle();
653 BUG();
656 int setup_profiling_timer(unsigned int multiplier)
658 return 0;
661 void __init smp_cpus_done(unsigned int max_cpus)
663 cpumask_var_t old_mask;
665 /* We want the setup_cpu() here to be called from CPU 0, but our
666 * init thread may have been "borrowed" by another CPU in the meantime
667 * se we pin us down to CPU 0 for a short while
669 alloc_cpumask_var(&old_mask, GFP_NOWAIT);
670 cpumask_copy(old_mask, tsk_cpus_allowed(current));
671 set_cpus_allowed_ptr(current, cpumask_of(boot_cpuid));
673 if (smp_ops && smp_ops->setup_cpu)
674 smp_ops->setup_cpu(boot_cpuid);
676 set_cpus_allowed_ptr(current, old_mask);
678 free_cpumask_var(old_mask);
680 if (smp_ops && smp_ops->bringup_done)
681 smp_ops->bringup_done();
683 dump_numa_cpu_topology();
687 int arch_sd_sibling_asym_packing(void)
689 if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
690 printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
691 return SD_ASYM_PACKING;
693 return 0;
696 #ifdef CONFIG_HOTPLUG_CPU
697 int __cpu_disable(void)
699 struct device_node *l2_cache;
700 int cpu = smp_processor_id();
701 int base, i;
702 int err;
704 if (!smp_ops->cpu_disable)
705 return -ENOSYS;
707 err = smp_ops->cpu_disable();
708 if (err)
709 return err;
711 /* Update sibling maps */
712 base = cpu_first_thread_sibling(cpu);
713 for (i = 0; i < threads_per_core; i++) {
714 cpumask_clear_cpu(cpu, cpu_sibling_mask(base + i));
715 cpumask_clear_cpu(base + i, cpu_sibling_mask(cpu));
716 cpumask_clear_cpu(cpu, cpu_core_mask(base + i));
717 cpumask_clear_cpu(base + i, cpu_core_mask(cpu));
720 l2_cache = cpu_to_l2cache(cpu);
721 for_each_present_cpu(i) {
722 struct device_node *np = cpu_to_l2cache(i);
723 if (!np)
724 continue;
725 if (np == l2_cache) {
726 cpumask_clear_cpu(cpu, cpu_core_mask(i));
727 cpumask_clear_cpu(i, cpu_core_mask(cpu));
729 of_node_put(np);
731 of_node_put(l2_cache);
734 return 0;
737 void __cpu_die(unsigned int cpu)
739 if (smp_ops->cpu_die)
740 smp_ops->cpu_die(cpu);
743 static DEFINE_MUTEX(powerpc_cpu_hotplug_driver_mutex);
745 void cpu_hotplug_driver_lock()
747 mutex_lock(&powerpc_cpu_hotplug_driver_mutex);
750 void cpu_hotplug_driver_unlock()
752 mutex_unlock(&powerpc_cpu_hotplug_driver_mutex);
755 void cpu_die(void)
757 if (ppc_md.cpu_die)
758 ppc_md.cpu_die();
760 /* If we return, we re-enter start_secondary */
761 start_secondary_resume();
764 #endif