]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - arch/arm/kernel/smp.c
Merge branch 'cpuidle' into release
[karo-tx-linux.git] / arch / arm / kernel / smp.c
1 /*
2  *  linux/arch/arm/kernel/smp.c
3  *
4  *  Copyright (C) 2002 ARM Limited, All Rights Reserved.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 as
8  * published by the Free Software Foundation.
9  */
10 #include <linux/module.h>
11 #include <linux/delay.h>
12 #include <linux/init.h>
13 #include <linux/spinlock.h>
14 #include <linux/sched.h>
15 #include <linux/interrupt.h>
16 #include <linux/cache.h>
17 #include <linux/profile.h>
18 #include <linux/errno.h>
19 #include <linux/mm.h>
20 #include <linux/err.h>
21 #include <linux/cpu.h>
22 #include <linux/seq_file.h>
23 #include <linux/irq.h>
24 #include <linux/nmi.h>
25 #include <linux/percpu.h>
26 #include <linux/clockchips.h>
27 #include <linux/completion.h>
28 #include <linux/cpufreq.h>
29 #include <linux/irq_work.h>
30
31 #include <linux/atomic.h>
32 #include <asm/smp.h>
33 #include <asm/cacheflush.h>
34 #include <asm/cpu.h>
35 #include <asm/cputype.h>
36 #include <asm/exception.h>
37 #include <asm/idmap.h>
38 #include <asm/topology.h>
39 #include <asm/mmu_context.h>
40 #include <asm/pgtable.h>
41 #include <asm/pgalloc.h>
42 #include <asm/processor.h>
43 #include <asm/sections.h>
44 #include <asm/tlbflush.h>
45 #include <asm/ptrace.h>
46 #include <asm/smp_plat.h>
47 #include <asm/virt.h>
48 #include <asm/mach/arch.h>
49 #include <asm/mpu.h>
50
51 #define CREATE_TRACE_POINTS
52 #include <trace/events/ipi.h>
53
54 /*
55  * as from 2.5, kernels no longer have an init_tasks structure
56  * so we need some other way of telling a new secondary core
57  * where to place its SVC stack
58  */
59 struct secondary_data secondary_data;
60
61 /*
62  * control for which core is the next to come out of the secondary
63  * boot "holding pen"
64  */
65 volatile int pen_release = -1;
66
67 enum ipi_msg_type {
68         IPI_WAKEUP,
69         IPI_TIMER,
70         IPI_RESCHEDULE,
71         IPI_CALL_FUNC,
72         IPI_CALL_FUNC_SINGLE,
73         IPI_CPU_STOP,
74         IPI_IRQ_WORK,
75         IPI_COMPLETION,
76         IPI_CPU_BACKTRACE = 15,
77 };
78
79 static DECLARE_COMPLETION(cpu_running);
80
81 static struct smp_operations smp_ops;
82
83 void __init smp_set_ops(struct smp_operations *ops)
84 {
85         if (ops)
86                 smp_ops = *ops;
87 };
88
89 static unsigned long get_arch_pgd(pgd_t *pgd)
90 {
91 #ifdef CONFIG_ARM_LPAE
92         return __phys_to_pfn(virt_to_phys(pgd));
93 #else
94         return virt_to_phys(pgd);
95 #endif
96 }
97
98 int __cpu_up(unsigned int cpu, struct task_struct *idle)
99 {
100         int ret;
101
102         if (!smp_ops.smp_boot_secondary)
103                 return -ENOSYS;
104
105         /*
106          * We need to tell the secondary core where to find
107          * its stack and the page tables.
108          */
109         secondary_data.stack = task_stack_page(idle) + THREAD_START_SP;
110 #ifdef CONFIG_ARM_MPU
111         secondary_data.mpu_rgn_szr = mpu_rgn_info.rgns[MPU_RAM_REGION].drsr;
112 #endif
113
114 #ifdef CONFIG_MMU
115         secondary_data.pgdir = virt_to_phys(idmap_pgd);
116         secondary_data.swapper_pg_dir = get_arch_pgd(swapper_pg_dir);
117 #endif
118         sync_cache_w(&secondary_data);
119
120         /*
121          * Now bring the CPU into our world.
122          */
123         ret = smp_ops.smp_boot_secondary(cpu, idle);
124         if (ret == 0) {
125                 /*
126                  * CPU was successfully started, wait for it
127                  * to come online or time out.
128                  */
129                 wait_for_completion_timeout(&cpu_running,
130                                                  msecs_to_jiffies(1000));
131
132                 if (!cpu_online(cpu)) {
133                         pr_crit("CPU%u: failed to come online\n", cpu);
134                         ret = -EIO;
135                 }
136         } else {
137                 pr_err("CPU%u: failed to boot: %d\n", cpu, ret);
138         }
139
140
141         memset(&secondary_data, 0, sizeof(secondary_data));
142         return ret;
143 }
144
145 /* platform specific SMP operations */
146 void __init smp_init_cpus(void)
147 {
148         if (smp_ops.smp_init_cpus)
149                 smp_ops.smp_init_cpus();
150 }
151
152 int platform_can_secondary_boot(void)
153 {
154         return !!smp_ops.smp_boot_secondary;
155 }
156
157 int platform_can_cpu_hotplug(void)
158 {
159 #ifdef CONFIG_HOTPLUG_CPU
160         if (smp_ops.cpu_kill)
161                 return 1;
162 #endif
163
164         return 0;
165 }
166
167 #ifdef CONFIG_HOTPLUG_CPU
168 static int platform_cpu_kill(unsigned int cpu)
169 {
170         if (smp_ops.cpu_kill)
171                 return smp_ops.cpu_kill(cpu);
172         return 1;
173 }
174
175 static int platform_cpu_disable(unsigned int cpu)
176 {
177         if (smp_ops.cpu_disable)
178                 return smp_ops.cpu_disable(cpu);
179
180         return 0;
181 }
182
183 int platform_can_hotplug_cpu(unsigned int cpu)
184 {
185         /* cpu_die must be specified to support hotplug */
186         if (!smp_ops.cpu_die)
187                 return 0;
188
189         if (smp_ops.cpu_can_disable)
190                 return smp_ops.cpu_can_disable(cpu);
191
192         /*
193          * By default, allow disabling all CPUs except the first one,
194          * since this is special on a lot of platforms, e.g. because
195          * of clock tick interrupts.
196          */
197         return cpu != 0;
198 }
199
200 /*
201  * __cpu_disable runs on the processor to be shutdown.
202  */
203 int __cpu_disable(void)
204 {
205         unsigned int cpu = smp_processor_id();
206         int ret;
207
208         ret = platform_cpu_disable(cpu);
209         if (ret)
210                 return ret;
211
212         /*
213          * Take this CPU offline.  Once we clear this, we can't return,
214          * and we must not schedule until we're ready to give up the cpu.
215          */
216         set_cpu_online(cpu, false);
217
218         /*
219          * OK - migrate IRQs away from this CPU
220          */
221         migrate_irqs();
222
223         /*
224          * Flush user cache and TLB mappings, and then remove this CPU
225          * from the vm mask set of all processes.
226          *
227          * Caches are flushed to the Level of Unification Inner Shareable
228          * to write-back dirty lines to unified caches shared by all CPUs.
229          */
230         flush_cache_louis();
231         local_flush_tlb_all();
232
233         clear_tasks_mm_cpumask(cpu);
234
235         return 0;
236 }
237
238 static DECLARE_COMPLETION(cpu_died);
239
240 /*
241  * called on the thread which is asking for a CPU to be shutdown -
242  * waits until shutdown has completed, or it is timed out.
243  */
244 void __cpu_die(unsigned int cpu)
245 {
246         if (!wait_for_completion_timeout(&cpu_died, msecs_to_jiffies(5000))) {
247                 pr_err("CPU%u: cpu didn't die\n", cpu);
248                 return;
249         }
250         pr_notice("CPU%u: shutdown\n", cpu);
251
252         /*
253          * platform_cpu_kill() is generally expected to do the powering off
254          * and/or cutting of clocks to the dying CPU.  Optionally, this may
255          * be done by the CPU which is dying in preference to supporting
256          * this call, but that means there is _no_ synchronisation between
257          * the requesting CPU and the dying CPU actually losing power.
258          */
259         if (!platform_cpu_kill(cpu))
260                 pr_err("CPU%u: unable to kill\n", cpu);
261 }
262
263 /*
264  * Called from the idle thread for the CPU which has been shutdown.
265  *
266  * Note that we disable IRQs here, but do not re-enable them
267  * before returning to the caller. This is also the behaviour
268  * of the other hotplug-cpu capable cores, so presumably coming
269  * out of idle fixes this.
270  */
271 void arch_cpu_idle_dead(void)
272 {
273         unsigned int cpu = smp_processor_id();
274
275         idle_task_exit();
276
277         local_irq_disable();
278
279         /*
280          * Flush the data out of the L1 cache for this CPU.  This must be
281          * before the completion to ensure that data is safely written out
282          * before platform_cpu_kill() gets called - which may disable
283          * *this* CPU and power down its cache.
284          */
285         flush_cache_louis();
286
287         /*
288          * Tell __cpu_die() that this CPU is now safe to dispose of.  Once
289          * this returns, power and/or clocks can be removed at any point
290          * from this CPU and its cache by platform_cpu_kill().
291          */
292         complete(&cpu_died);
293
294         /*
295          * Ensure that the cache lines associated with that completion are
296          * written out.  This covers the case where _this_ CPU is doing the
297          * powering down, to ensure that the completion is visible to the
298          * CPU waiting for this one.
299          */
300         flush_cache_louis();
301
302         /*
303          * The actual CPU shutdown procedure is at least platform (if not
304          * CPU) specific.  This may remove power, or it may simply spin.
305          *
306          * Platforms are generally expected *NOT* to return from this call,
307          * although there are some which do because they have no way to
308          * power down the CPU.  These platforms are the _only_ reason we
309          * have a return path which uses the fragment of assembly below.
310          *
311          * The return path should not be used for platforms which can
312          * power off the CPU.
313          */
314         if (smp_ops.cpu_die)
315                 smp_ops.cpu_die(cpu);
316
317         pr_warn("CPU%u: smp_ops.cpu_die() returned, trying to resuscitate\n",
318                 cpu);
319
320         /*
321          * Do not return to the idle loop - jump back to the secondary
322          * cpu initialisation.  There's some initialisation which needs
323          * to be repeated to undo the effects of taking the CPU offline.
324          */
325         __asm__("mov    sp, %0\n"
326         "       mov     fp, #0\n"
327         "       b       secondary_start_kernel"
328                 :
329                 : "r" (task_stack_page(current) + THREAD_SIZE - 8));
330 }
331 #endif /* CONFIG_HOTPLUG_CPU */
332
333 /*
334  * Called by both boot and secondaries to move global data into
335  * per-processor storage.
336  */
337 static void smp_store_cpu_info(unsigned int cpuid)
338 {
339         struct cpuinfo_arm *cpu_info = &per_cpu(cpu_data, cpuid);
340
341         cpu_info->loops_per_jiffy = loops_per_jiffy;
342         cpu_info->cpuid = read_cpuid_id();
343
344         store_cpu_topology(cpuid);
345 }
346
347 /*
348  * This is the secondary CPU boot entry.  We're using this CPUs
349  * idle thread stack, but a set of temporary page tables.
350  */
351 asmlinkage void secondary_start_kernel(void)
352 {
353         struct mm_struct *mm = &init_mm;
354         unsigned int cpu;
355
356         /*
357          * The identity mapping is uncached (strongly ordered), so
358          * switch away from it before attempting any exclusive accesses.
359          */
360         cpu_switch_mm(mm->pgd, mm);
361         local_flush_bp_all();
362         enter_lazy_tlb(mm, current);
363         local_flush_tlb_all();
364
365         /*
366          * All kernel threads share the same mm context; grab a
367          * reference and switch to it.
368          */
369         cpu = smp_processor_id();
370         atomic_inc(&mm->mm_count);
371         current->active_mm = mm;
372         cpumask_set_cpu(cpu, mm_cpumask(mm));
373
374         cpu_init();
375
376         pr_debug("CPU%u: Booted secondary processor\n", cpu);
377
378         preempt_disable();
379         trace_hardirqs_off();
380
381         /*
382          * Give the platform a chance to do its own initialisation.
383          */
384         if (smp_ops.smp_secondary_init)
385                 smp_ops.smp_secondary_init(cpu);
386
387         notify_cpu_starting(cpu);
388
389         calibrate_delay();
390
391         smp_store_cpu_info(cpu);
392
393         /*
394          * OK, now it's safe to let the boot CPU continue.  Wait for
395          * the CPU migration code to notice that the CPU is online
396          * before we continue - which happens after __cpu_up returns.
397          */
398         set_cpu_online(cpu, true);
399         complete(&cpu_running);
400
401         local_irq_enable();
402         local_fiq_enable();
403
404         /*
405          * OK, it's off to the idle thread for us
406          */
407         cpu_startup_entry(CPUHP_ONLINE);
408 }
409
410 void __init smp_cpus_done(unsigned int max_cpus)
411 {
412         int cpu;
413         unsigned long bogosum = 0;
414
415         for_each_online_cpu(cpu)
416                 bogosum += per_cpu(cpu_data, cpu).loops_per_jiffy;
417
418         printk(KERN_INFO "SMP: Total of %d processors activated "
419                "(%lu.%02lu BogoMIPS).\n",
420                num_online_cpus(),
421                bogosum / (500000/HZ),
422                (bogosum / (5000/HZ)) % 100);
423
424         hyp_mode_check();
425 }
426
427 void __init smp_prepare_boot_cpu(void)
428 {
429         set_my_cpu_offset(per_cpu_offset(smp_processor_id()));
430 }
431
432 void __init smp_prepare_cpus(unsigned int max_cpus)
433 {
434         unsigned int ncores = num_possible_cpus();
435
436         init_cpu_topology();
437
438         smp_store_cpu_info(smp_processor_id());
439
440         /*
441          * are we trying to boot more cores than exist?
442          */
443         if (max_cpus > ncores)
444                 max_cpus = ncores;
445         if (ncores > 1 && max_cpus) {
446                 /*
447                  * Initialise the present map, which describes the set of CPUs
448                  * actually populated at the present time. A platform should
449                  * re-initialize the map in the platforms smp_prepare_cpus()
450                  * if present != possible (e.g. physical hotplug).
451                  */
452                 init_cpu_present(cpu_possible_mask);
453
454                 /*
455                  * Initialise the SCU if there are more than one CPU
456                  * and let them know where to start.
457                  */
458                 if (smp_ops.smp_prepare_cpus)
459                         smp_ops.smp_prepare_cpus(max_cpus);
460         }
461 }
462
463 static void (*__smp_cross_call)(const struct cpumask *, unsigned int);
464
465 void __init set_smp_cross_call(void (*fn)(const struct cpumask *, unsigned int))
466 {
467         if (!__smp_cross_call)
468                 __smp_cross_call = fn;
469 }
470
471 static const char *ipi_types[NR_IPI] __tracepoint_string = {
472 #define S(x,s)  [x] = s
473         S(IPI_WAKEUP, "CPU wakeup interrupts"),
474         S(IPI_TIMER, "Timer broadcast interrupts"),
475         S(IPI_RESCHEDULE, "Rescheduling interrupts"),
476         S(IPI_CALL_FUNC, "Function call interrupts"),
477         S(IPI_CALL_FUNC_SINGLE, "Single function call interrupts"),
478         S(IPI_CPU_STOP, "CPU stop interrupts"),
479         S(IPI_IRQ_WORK, "IRQ work interrupts"),
480         S(IPI_COMPLETION, "completion interrupts"),
481 };
482
483 static void smp_cross_call(const struct cpumask *target, unsigned int ipinr)
484 {
485         trace_ipi_raise(target, ipi_types[ipinr]);
486         __smp_cross_call(target, ipinr);
487 }
488
489 void show_ipi_list(struct seq_file *p, int prec)
490 {
491         unsigned int cpu, i;
492
493         for (i = 0; i < NR_IPI; i++) {
494                 seq_printf(p, "%*s%u: ", prec - 1, "IPI", i);
495
496                 for_each_online_cpu(cpu)
497                         seq_printf(p, "%10u ",
498                                    __get_irq_stat(cpu, ipi_irqs[i]));
499
500                 seq_printf(p, " %s\n", ipi_types[i]);
501         }
502 }
503
504 u64 smp_irq_stat_cpu(unsigned int cpu)
505 {
506         u64 sum = 0;
507         int i;
508
509         for (i = 0; i < NR_IPI; i++)
510                 sum += __get_irq_stat(cpu, ipi_irqs[i]);
511
512         return sum;
513 }
514
515 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
516 {
517         smp_cross_call(mask, IPI_CALL_FUNC);
518 }
519
520 void arch_send_wakeup_ipi_mask(const struct cpumask *mask)
521 {
522         smp_cross_call(mask, IPI_WAKEUP);
523 }
524
525 void arch_send_call_function_single_ipi(int cpu)
526 {
527         smp_cross_call(cpumask_of(cpu), IPI_CALL_FUNC_SINGLE);
528 }
529
530 #ifdef CONFIG_IRQ_WORK
531 void arch_irq_work_raise(void)
532 {
533         if (arch_irq_work_has_interrupt())
534                 smp_cross_call(cpumask_of(smp_processor_id()), IPI_IRQ_WORK);
535 }
536 #endif
537
538 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
539 void tick_broadcast(const struct cpumask *mask)
540 {
541         smp_cross_call(mask, IPI_TIMER);
542 }
543 #endif
544
545 static DEFINE_RAW_SPINLOCK(stop_lock);
546
547 /*
548  * ipi_cpu_stop - handle IPI from smp_send_stop()
549  */
550 static void ipi_cpu_stop(unsigned int cpu)
551 {
552         if (system_state == SYSTEM_BOOTING ||
553             system_state == SYSTEM_RUNNING) {
554                 raw_spin_lock(&stop_lock);
555                 pr_crit("CPU%u: stopping\n", cpu);
556                 dump_stack();
557                 raw_spin_unlock(&stop_lock);
558         }
559
560         set_cpu_online(cpu, false);
561
562         local_fiq_disable();
563         local_irq_disable();
564
565         while (1)
566                 cpu_relax();
567 }
568
569 static DEFINE_PER_CPU(struct completion *, cpu_completion);
570
571 int register_ipi_completion(struct completion *completion, int cpu)
572 {
573         per_cpu(cpu_completion, cpu) = completion;
574         return IPI_COMPLETION;
575 }
576
577 static void ipi_complete(unsigned int cpu)
578 {
579         complete(per_cpu(cpu_completion, cpu));
580 }
581
582 /*
583  * Main handler for inter-processor interrupts
584  */
585 asmlinkage void __exception_irq_entry do_IPI(int ipinr, struct pt_regs *regs)
586 {
587         handle_IPI(ipinr, regs);
588 }
589
590 void handle_IPI(int ipinr, struct pt_regs *regs)
591 {
592         unsigned int cpu = smp_processor_id();
593         struct pt_regs *old_regs = set_irq_regs(regs);
594
595         if ((unsigned)ipinr < NR_IPI) {
596                 trace_ipi_entry_rcuidle(ipi_types[ipinr]);
597                 __inc_irq_stat(cpu, ipi_irqs[ipinr]);
598         }
599
600         switch (ipinr) {
601         case IPI_WAKEUP:
602                 break;
603
604 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
605         case IPI_TIMER:
606                 irq_enter();
607                 tick_receive_broadcast();
608                 irq_exit();
609                 break;
610 #endif
611
612         case IPI_RESCHEDULE:
613                 scheduler_ipi();
614                 break;
615
616         case IPI_CALL_FUNC:
617                 irq_enter();
618                 generic_smp_call_function_interrupt();
619                 irq_exit();
620                 break;
621
622         case IPI_CALL_FUNC_SINGLE:
623                 irq_enter();
624                 generic_smp_call_function_single_interrupt();
625                 irq_exit();
626                 break;
627
628         case IPI_CPU_STOP:
629                 irq_enter();
630                 ipi_cpu_stop(cpu);
631                 irq_exit();
632                 break;
633
634 #ifdef CONFIG_IRQ_WORK
635         case IPI_IRQ_WORK:
636                 irq_enter();
637                 irq_work_run();
638                 irq_exit();
639                 break;
640 #endif
641
642         case IPI_COMPLETION:
643                 irq_enter();
644                 ipi_complete(cpu);
645                 irq_exit();
646                 break;
647
648         case IPI_CPU_BACKTRACE:
649                 irq_enter();
650                 nmi_cpu_backtrace(regs);
651                 irq_exit();
652                 break;
653
654         default:
655                 pr_crit("CPU%u: Unknown IPI message 0x%x\n",
656                         cpu, ipinr);
657                 break;
658         }
659
660         if ((unsigned)ipinr < NR_IPI)
661                 trace_ipi_exit_rcuidle(ipi_types[ipinr]);
662         set_irq_regs(old_regs);
663 }
664
665 void smp_send_reschedule(int cpu)
666 {
667         smp_cross_call(cpumask_of(cpu), IPI_RESCHEDULE);
668 }
669
670 void smp_send_stop(void)
671 {
672         unsigned long timeout;
673         struct cpumask mask;
674
675         cpumask_copy(&mask, cpu_online_mask);
676         cpumask_clear_cpu(smp_processor_id(), &mask);
677         if (!cpumask_empty(&mask))
678                 smp_cross_call(&mask, IPI_CPU_STOP);
679
680         /* Wait up to one second for other CPUs to stop */
681         timeout = USEC_PER_SEC;
682         while (num_online_cpus() > 1 && timeout--)
683                 udelay(1);
684
685         if (num_online_cpus() > 1)
686                 pr_warn("SMP: failed to stop secondary CPUs\n");
687 }
688
689 /*
690  * not supported here
691  */
692 int setup_profiling_timer(unsigned int multiplier)
693 {
694         return -EINVAL;
695 }
696
697 #ifdef CONFIG_CPU_FREQ
698
699 static DEFINE_PER_CPU(unsigned long, l_p_j_ref);
700 static DEFINE_PER_CPU(unsigned long, l_p_j_ref_freq);
701 static unsigned long global_l_p_j_ref;
702 static unsigned long global_l_p_j_ref_freq;
703
704 static int cpufreq_callback(struct notifier_block *nb,
705                                         unsigned long val, void *data)
706 {
707         struct cpufreq_freqs *freq = data;
708         int cpu = freq->cpu;
709
710         if (freq->flags & CPUFREQ_CONST_LOOPS)
711                 return NOTIFY_OK;
712
713         if (!per_cpu(l_p_j_ref, cpu)) {
714                 per_cpu(l_p_j_ref, cpu) =
715                         per_cpu(cpu_data, cpu).loops_per_jiffy;
716                 per_cpu(l_p_j_ref_freq, cpu) = freq->old;
717                 if (!global_l_p_j_ref) {
718                         global_l_p_j_ref = loops_per_jiffy;
719                         global_l_p_j_ref_freq = freq->old;
720                 }
721         }
722
723         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
724             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new)) {
725                 loops_per_jiffy = cpufreq_scale(global_l_p_j_ref,
726                                                 global_l_p_j_ref_freq,
727                                                 freq->new);
728                 per_cpu(cpu_data, cpu).loops_per_jiffy =
729                         cpufreq_scale(per_cpu(l_p_j_ref, cpu),
730                                         per_cpu(l_p_j_ref_freq, cpu),
731                                         freq->new);
732         }
733         return NOTIFY_OK;
734 }
735
736 static struct notifier_block cpufreq_notifier = {
737         .notifier_call  = cpufreq_callback,
738 };
739
740 static int __init register_cpufreq_notifier(void)
741 {
742         return cpufreq_register_notifier(&cpufreq_notifier,
743                                                 CPUFREQ_TRANSITION_NOTIFIER);
744 }
745 core_initcall(register_cpufreq_notifier);
746
747 #endif
748
749 static void raise_nmi(cpumask_t *mask)
750 {
751         smp_cross_call(mask, IPI_CPU_BACKTRACE);
752 }
753
754 void arch_trigger_all_cpu_backtrace(bool include_self)
755 {
756         nmi_trigger_all_cpu_backtrace(include_self, raise_nmi);
757 }