Merge tag 'v3.3.7' into 3.3/master
[zen-stable.git] / arch / powerpc / kernel / smp.c
blob46695febc09f6e10b96c1b8edcb736f749f40db4
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/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>
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))
74 /* State of each CPU during hotplug phases */
75 static DEFINE_PER_CPU(int, cpu_state) = { 0 };
77 #else
78 static struct task_struct *idle_thread_array[NR_CPUS] __cpuinitdata ;
79 #define get_idle_for_cpu(x) (idle_thread_array[(x)])
80 #define set_idle_for_cpu(x, p) (idle_thread_array[(x)] = (p))
81 #endif
83 struct thread_info *secondary_ti;
85 DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
86 DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
88 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
89 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
91 /* SMP operations for this machine */
92 struct smp_ops_t *smp_ops;
94 /* Can't be static due to PowerMac hackery */
95 volatile unsigned int cpu_callin_map[NR_CPUS];
97 int smt_enabled_at_boot = 1;
99 static void (*crash_ipi_function_ptr)(struct pt_regs *) = NULL;
101 #ifdef CONFIG_PPC64
102 int __devinit smp_generic_kick_cpu(int nr)
104 BUG_ON(nr < 0 || nr >= NR_CPUS);
107 * The processor is currently spinning, waiting for the
108 * cpu_start field to become non-zero After we set cpu_start,
109 * the processor will continue on to secondary_start
111 if (!paca[nr].cpu_start) {
112 paca[nr].cpu_start = 1;
113 smp_mb();
114 return 0;
117 #ifdef CONFIG_HOTPLUG_CPU
119 * Ok it's not there, so it might be soft-unplugged, let's
120 * try to bring it back
122 per_cpu(cpu_state, nr) = CPU_UP_PREPARE;
123 smp_wmb();
124 smp_send_reschedule(nr);
125 #endif /* CONFIG_HOTPLUG_CPU */
127 return 0;
129 #endif /* CONFIG_PPC64 */
131 static irqreturn_t call_function_action(int irq, void *data)
133 generic_smp_call_function_interrupt();
134 return IRQ_HANDLED;
137 static irqreturn_t reschedule_action(int irq, void *data)
139 scheduler_ipi();
140 return IRQ_HANDLED;
143 static irqreturn_t call_function_single_action(int irq, void *data)
145 generic_smp_call_function_single_interrupt();
146 return IRQ_HANDLED;
149 static irqreturn_t debug_ipi_action(int irq, void *data)
151 if (crash_ipi_function_ptr) {
152 crash_ipi_function_ptr(get_irq_regs());
153 return IRQ_HANDLED;
156 #ifdef CONFIG_DEBUGGER
157 debugger_ipi(get_irq_regs());
158 #endif /* CONFIG_DEBUGGER */
160 return IRQ_HANDLED;
163 static irq_handler_t smp_ipi_action[] = {
164 [PPC_MSG_CALL_FUNCTION] = call_function_action,
165 [PPC_MSG_RESCHEDULE] = reschedule_action,
166 [PPC_MSG_CALL_FUNC_SINGLE] = call_function_single_action,
167 [PPC_MSG_DEBUGGER_BREAK] = debug_ipi_action,
170 const char *smp_ipi_name[] = {
171 [PPC_MSG_CALL_FUNCTION] = "ipi call function",
172 [PPC_MSG_RESCHEDULE] = "ipi reschedule",
173 [PPC_MSG_CALL_FUNC_SINGLE] = "ipi call function single",
174 [PPC_MSG_DEBUGGER_BREAK] = "ipi debugger",
177 /* optional function to request ipi, for controllers with >= 4 ipis */
178 int smp_request_message_ipi(int virq, int msg)
180 int err;
182 if (msg < 0 || msg > PPC_MSG_DEBUGGER_BREAK) {
183 return -EINVAL;
185 #if !defined(CONFIG_DEBUGGER) && !defined(CONFIG_KEXEC)
186 if (msg == PPC_MSG_DEBUGGER_BREAK) {
187 return 1;
189 #endif
190 err = request_irq(virq, smp_ipi_action[msg],
191 IRQF_PERCPU | IRQF_NO_THREAD,
192 smp_ipi_name[msg], 0);
193 WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n",
194 virq, smp_ipi_name[msg], err);
196 return err;
199 #ifdef CONFIG_PPC_SMP_MUXED_IPI
200 struct cpu_messages {
201 int messages; /* current messages */
202 unsigned long data; /* data for cause ipi */
204 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages, ipi_message);
206 void smp_muxed_ipi_set_data(int cpu, unsigned long data)
208 struct cpu_messages *info = &per_cpu(ipi_message, cpu);
210 info->data = data;
213 void smp_muxed_ipi_message_pass(int cpu, int msg)
215 struct cpu_messages *info = &per_cpu(ipi_message, cpu);
216 char *message = (char *)&info->messages;
218 message[msg] = 1;
219 mb();
220 smp_ops->cause_ipi(cpu, info->data);
223 irqreturn_t smp_ipi_demux(void)
225 struct cpu_messages *info = &__get_cpu_var(ipi_message);
226 unsigned int all;
228 mb(); /* order any irq clear */
230 do {
231 all = xchg_local(&info->messages, 0);
233 #ifdef __BIG_ENDIAN
234 if (all & (1 << (24 - 8 * PPC_MSG_CALL_FUNCTION)))
235 generic_smp_call_function_interrupt();
236 if (all & (1 << (24 - 8 * PPC_MSG_RESCHEDULE)))
237 scheduler_ipi();
238 if (all & (1 << (24 - 8 * PPC_MSG_CALL_FUNC_SINGLE)))
239 generic_smp_call_function_single_interrupt();
240 if (all & (1 << (24 - 8 * PPC_MSG_DEBUGGER_BREAK)))
241 debug_ipi_action(0, NULL);
242 #else
243 #error Unsupported ENDIAN
244 #endif
245 } while (info->messages);
247 return IRQ_HANDLED;
249 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
251 static inline void do_message_pass(int cpu, int msg)
253 if (smp_ops->message_pass)
254 smp_ops->message_pass(cpu, msg);
255 #ifdef CONFIG_PPC_SMP_MUXED_IPI
256 else
257 smp_muxed_ipi_message_pass(cpu, msg);
258 #endif
261 void smp_send_reschedule(int cpu)
263 if (likely(smp_ops))
264 do_message_pass(cpu, PPC_MSG_RESCHEDULE);
266 EXPORT_SYMBOL_GPL(smp_send_reschedule);
268 void arch_send_call_function_single_ipi(int cpu)
270 do_message_pass(cpu, PPC_MSG_CALL_FUNC_SINGLE);
273 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
275 unsigned int cpu;
277 for_each_cpu(cpu, mask)
278 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
281 #if defined(CONFIG_DEBUGGER) || defined(CONFIG_KEXEC)
282 void smp_send_debugger_break(void)
284 int cpu;
285 int me = raw_smp_processor_id();
287 if (unlikely(!smp_ops))
288 return;
290 for_each_online_cpu(cpu)
291 if (cpu != me)
292 do_message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
294 #endif
296 #ifdef CONFIG_KEXEC
297 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
299 crash_ipi_function_ptr = crash_ipi_callback;
300 if (crash_ipi_callback) {
301 mb();
302 smp_send_debugger_break();
305 #endif
307 static void stop_this_cpu(void *dummy)
309 /* Remove this CPU */
310 set_cpu_online(smp_processor_id(), false);
312 local_irq_disable();
313 while (1)
317 void smp_send_stop(void)
319 smp_call_function(stop_this_cpu, NULL, 0);
322 struct thread_info *current_set[NR_CPUS];
324 static void __devinit smp_store_cpu_info(int id)
326 per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
327 #ifdef CONFIG_PPC_FSL_BOOK3E
328 per_cpu(next_tlbcam_idx, id)
329 = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) - 1;
330 #endif
333 void __init smp_prepare_cpus(unsigned int max_cpus)
335 unsigned int cpu;
337 DBG("smp_prepare_cpus\n");
340 * setup_cpu may need to be called on the boot cpu. We havent
341 * spun any cpus up but lets be paranoid.
343 BUG_ON(boot_cpuid != smp_processor_id());
345 /* Fixup boot cpu */
346 smp_store_cpu_info(boot_cpuid);
347 cpu_callin_map[boot_cpuid] = 1;
349 for_each_possible_cpu(cpu) {
350 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
351 GFP_KERNEL, cpu_to_node(cpu));
352 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
353 GFP_KERNEL, cpu_to_node(cpu));
356 cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
357 cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
359 if (smp_ops)
360 if (smp_ops->probe)
361 max_cpus = smp_ops->probe();
362 else
363 max_cpus = NR_CPUS;
364 else
365 max_cpus = 1;
368 void __devinit smp_prepare_boot_cpu(void)
370 BUG_ON(smp_processor_id() != boot_cpuid);
371 #ifdef CONFIG_PPC64
372 paca[boot_cpuid].__current = current;
373 #endif
374 current_set[boot_cpuid] = task_thread_info(current);
377 #ifdef CONFIG_HOTPLUG_CPU
379 int generic_cpu_disable(void)
381 unsigned int cpu = smp_processor_id();
383 if (cpu == boot_cpuid)
384 return -EBUSY;
386 set_cpu_online(cpu, false);
387 #ifdef CONFIG_PPC64
388 vdso_data->processorCount--;
389 #endif
390 migrate_irqs();
391 return 0;
394 void generic_cpu_die(unsigned int cpu)
396 int i;
398 for (i = 0; i < 100; i++) {
399 smp_rmb();
400 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
401 return;
402 msleep(100);
404 printk(KERN_ERR "CPU%d didn't die...\n", cpu);
407 void generic_mach_cpu_die(void)
409 unsigned int cpu;
411 local_irq_disable();
412 idle_task_exit();
413 cpu = smp_processor_id();
414 printk(KERN_DEBUG "CPU%d offline\n", cpu);
415 __get_cpu_var(cpu_state) = CPU_DEAD;
416 smp_wmb();
417 while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
418 cpu_relax();
421 void generic_set_cpu_dead(unsigned int cpu)
423 per_cpu(cpu_state, cpu) = CPU_DEAD;
426 int generic_check_cpu_restart(unsigned int cpu)
428 return per_cpu(cpu_state, cpu) == CPU_UP_PREPARE;
430 #endif
432 struct create_idle {
433 struct work_struct work;
434 struct task_struct *idle;
435 struct completion done;
436 int cpu;
439 static void __cpuinit do_fork_idle(struct work_struct *work)
441 struct create_idle *c_idle =
442 container_of(work, struct create_idle, work);
444 c_idle->idle = fork_idle(c_idle->cpu);
445 complete(&c_idle->done);
448 static int __cpuinit create_idle(unsigned int cpu)
450 struct thread_info *ti;
451 struct create_idle c_idle = {
452 .cpu = cpu,
453 .done = COMPLETION_INITIALIZER_ONSTACK(c_idle.done),
455 INIT_WORK_ONSTACK(&c_idle.work, do_fork_idle);
457 c_idle.idle = get_idle_for_cpu(cpu);
459 /* We can't use kernel_thread since we must avoid to
460 * reschedule the child. We use a workqueue because
461 * we want to fork from a kernel thread, not whatever
462 * userspace process happens to be trying to online us.
464 if (!c_idle.idle) {
465 schedule_work(&c_idle.work);
466 wait_for_completion(&c_idle.done);
467 } else
468 init_idle(c_idle.idle, cpu);
469 if (IS_ERR(c_idle.idle)) {
470 pr_err("Failed fork for CPU %u: %li", cpu, PTR_ERR(c_idle.idle));
471 return PTR_ERR(c_idle.idle);
473 ti = task_thread_info(c_idle.idle);
475 #ifdef CONFIG_PPC64
476 paca[cpu].__current = c_idle.idle;
477 paca[cpu].kstack = (unsigned long)ti + THREAD_SIZE - STACK_FRAME_OVERHEAD;
478 #endif
479 ti->cpu = cpu;
480 current_set[cpu] = ti;
482 return 0;
485 int __cpuinit __cpu_up(unsigned int cpu)
487 int rc, c;
489 if (smp_ops == NULL ||
490 (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
491 return -EINVAL;
493 /* Make sure we have an idle thread */
494 rc = create_idle(cpu);
495 if (rc)
496 return rc;
498 secondary_ti = current_set[cpu];
500 /* Make sure callin-map entry is 0 (can be leftover a CPU
501 * hotplug
503 cpu_callin_map[cpu] = 0;
505 /* The information for processor bringup must
506 * be written out to main store before we release
507 * the processor.
509 smp_mb();
511 /* wake up cpus */
512 DBG("smp: kicking cpu %d\n", cpu);
513 rc = smp_ops->kick_cpu(cpu);
514 if (rc) {
515 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
516 return rc;
520 * wait to see if the cpu made a callin (is actually up).
521 * use this value that I found through experimentation.
522 * -- Cort
524 if (system_state < SYSTEM_RUNNING)
525 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
526 udelay(100);
527 #ifdef CONFIG_HOTPLUG_CPU
528 else
530 * CPUs can take much longer to come up in the
531 * hotplug case. Wait five seconds.
533 for (c = 5000; c && !cpu_callin_map[cpu]; c--)
534 msleep(1);
535 #endif
537 if (!cpu_callin_map[cpu]) {
538 printk(KERN_ERR "Processor %u is stuck.\n", cpu);
539 return -ENOENT;
542 DBG("Processor %u found.\n", cpu);
544 if (smp_ops->give_timebase)
545 smp_ops->give_timebase();
547 /* Wait until cpu puts itself in the online map */
548 while (!cpu_online(cpu))
549 cpu_relax();
551 return 0;
554 /* Return the value of the reg property corresponding to the given
555 * logical cpu.
557 int cpu_to_core_id(int cpu)
559 struct device_node *np;
560 const int *reg;
561 int id = -1;
563 np = of_get_cpu_node(cpu, NULL);
564 if (!np)
565 goto out;
567 reg = of_get_property(np, "reg", NULL);
568 if (!reg)
569 goto out;
571 id = *reg;
572 out:
573 of_node_put(np);
574 return id;
577 /* Helper routines for cpu to core mapping */
578 int cpu_core_index_of_thread(int cpu)
580 return cpu >> threads_shift;
582 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
584 int cpu_first_thread_of_core(int core)
586 return core << threads_shift;
588 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
590 /* Must be called when no change can occur to cpu_present_mask,
591 * i.e. during cpu online or offline.
593 static struct device_node *cpu_to_l2cache(int cpu)
595 struct device_node *np;
596 struct device_node *cache;
598 if (!cpu_present(cpu))
599 return NULL;
601 np = of_get_cpu_node(cpu, NULL);
602 if (np == NULL)
603 return NULL;
605 cache = of_find_next_cache_node(np);
607 of_node_put(np);
609 return cache;
612 /* Activate a secondary processor. */
613 void __devinit start_secondary(void *unused)
615 unsigned int cpu = smp_processor_id();
616 struct device_node *l2_cache;
617 int i, base;
619 atomic_inc(&init_mm.mm_count);
620 current->active_mm = &init_mm;
622 smp_store_cpu_info(cpu);
623 set_dec(tb_ticks_per_jiffy);
624 preempt_disable();
625 cpu_callin_map[cpu] = 1;
627 if (smp_ops->setup_cpu)
628 smp_ops->setup_cpu(cpu);
629 if (smp_ops->take_timebase)
630 smp_ops->take_timebase();
632 secondary_cpu_time_init();
634 #ifdef CONFIG_PPC64
635 if (system_state == SYSTEM_RUNNING)
636 vdso_data->processorCount++;
637 #endif
638 ipi_call_lock();
639 notify_cpu_starting(cpu);
640 set_cpu_online(cpu, true);
641 /* Update sibling maps */
642 base = cpu_first_thread_sibling(cpu);
643 for (i = 0; i < threads_per_core; i++) {
644 if (cpu_is_offline(base + i))
645 continue;
646 cpumask_set_cpu(cpu, cpu_sibling_mask(base + i));
647 cpumask_set_cpu(base + i, cpu_sibling_mask(cpu));
649 /* cpu_core_map should be a superset of
650 * cpu_sibling_map even if we don't have cache
651 * information, so update the former here, too.
653 cpumask_set_cpu(cpu, cpu_core_mask(base + i));
654 cpumask_set_cpu(base + i, cpu_core_mask(cpu));
656 l2_cache = cpu_to_l2cache(cpu);
657 for_each_online_cpu(i) {
658 struct device_node *np = cpu_to_l2cache(i);
659 if (!np)
660 continue;
661 if (np == l2_cache) {
662 cpumask_set_cpu(cpu, cpu_core_mask(i));
663 cpumask_set_cpu(i, cpu_core_mask(cpu));
665 of_node_put(np);
667 of_node_put(l2_cache);
668 ipi_call_unlock();
670 local_irq_enable();
672 cpu_idle();
674 BUG();
677 int setup_profiling_timer(unsigned int multiplier)
679 return 0;
682 void __init smp_cpus_done(unsigned int max_cpus)
684 cpumask_var_t old_mask;
686 /* We want the setup_cpu() here to be called from CPU 0, but our
687 * init thread may have been "borrowed" by another CPU in the meantime
688 * se we pin us down to CPU 0 for a short while
690 alloc_cpumask_var(&old_mask, GFP_NOWAIT);
691 cpumask_copy(old_mask, tsk_cpus_allowed(current));
692 set_cpus_allowed_ptr(current, cpumask_of(boot_cpuid));
694 if (smp_ops && smp_ops->setup_cpu)
695 smp_ops->setup_cpu(boot_cpuid);
697 set_cpus_allowed_ptr(current, old_mask);
699 free_cpumask_var(old_mask);
701 if (smp_ops && smp_ops->bringup_done)
702 smp_ops->bringup_done();
704 dump_numa_cpu_topology();
708 int arch_sd_sibling_asym_packing(void)
710 if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
711 printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
712 return SD_ASYM_PACKING;
714 return 0;
717 #ifdef CONFIG_HOTPLUG_CPU
718 int __cpu_disable(void)
720 struct device_node *l2_cache;
721 int cpu = smp_processor_id();
722 int base, i;
723 int err;
725 if (!smp_ops->cpu_disable)
726 return -ENOSYS;
728 err = smp_ops->cpu_disable();
729 if (err)
730 return err;
732 /* Update sibling maps */
733 base = cpu_first_thread_sibling(cpu);
734 for (i = 0; i < threads_per_core; i++) {
735 cpumask_clear_cpu(cpu, cpu_sibling_mask(base + i));
736 cpumask_clear_cpu(base + i, cpu_sibling_mask(cpu));
737 cpumask_clear_cpu(cpu, cpu_core_mask(base + i));
738 cpumask_clear_cpu(base + i, cpu_core_mask(cpu));
741 l2_cache = cpu_to_l2cache(cpu);
742 for_each_present_cpu(i) {
743 struct device_node *np = cpu_to_l2cache(i);
744 if (!np)
745 continue;
746 if (np == l2_cache) {
747 cpumask_clear_cpu(cpu, cpu_core_mask(i));
748 cpumask_clear_cpu(i, cpu_core_mask(cpu));
750 of_node_put(np);
752 of_node_put(l2_cache);
755 return 0;
758 void __cpu_die(unsigned int cpu)
760 if (smp_ops->cpu_die)
761 smp_ops->cpu_die(cpu);
764 static DEFINE_MUTEX(powerpc_cpu_hotplug_driver_mutex);
766 void cpu_hotplug_driver_lock()
768 mutex_lock(&powerpc_cpu_hotplug_driver_mutex);
771 void cpu_hotplug_driver_unlock()
773 mutex_unlock(&powerpc_cpu_hotplug_driver_mutex);
776 void cpu_die(void)
778 if (ppc_md.cpu_die)
779 ppc_md.cpu_die();
781 /* If we return, we re-enter start_secondary */
782 start_secondary_resume();
785 #endif