Kernel bump from 4.1.3-rt to 4.1.7-rt.
[kvmfornfv.git] / kernel / 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/percpu.h>
25 #include <linux/clockchips.h>
26 #include <linux/completion.h>
27 #include <linux/cpufreq.h>
28 #include <linux/irq_work.h>
29
30 #include <linux/atomic.h>
31 #include <asm/smp.h>
32 #include <asm/cacheflush.h>
33 #include <asm/cpu.h>
34 #include <asm/cputype.h>
35 #include <asm/exception.h>
36 #include <asm/idmap.h>
37 #include <asm/topology.h>
38 #include <asm/mmu_context.h>
39 #include <asm/pgtable.h>
40 #include <asm/pgalloc.h>
41 #include <asm/processor.h>
42 #include <asm/sections.h>
43 #include <asm/tlbflush.h>
44 #include <asm/ptrace.h>
45 #include <asm/smp_plat.h>
46 #include <asm/virt.h>
47 #include <asm/mach/arch.h>
48 #include <asm/mpu.h>
49
50 #define CREATE_TRACE_POINTS
51 #include <trace/events/ipi.h>
52
53 /*
54  * as from 2.5, kernels no longer have an init_tasks structure
55  * so we need some other way of telling a new secondary core
56  * where to place its SVC stack
57  */
58 struct secondary_data secondary_data;
59
60 /*
61  * control for which core is the next to come out of the secondary
62  * boot "holding pen"
63  */
64 volatile int pen_release = -1;
65
66 enum ipi_msg_type {
67         IPI_WAKEUP,
68         IPI_TIMER,
69         IPI_RESCHEDULE,
70         IPI_CALL_FUNC,
71         IPI_CALL_FUNC_SINGLE,
72         IPI_CPU_STOP,
73         IPI_IRQ_WORK,
74         IPI_COMPLETION,
75 };
76
77 static DECLARE_COMPLETION(cpu_running);
78
79 static struct smp_operations smp_ops;
80
81 void __init smp_set_ops(struct smp_operations *ops)
82 {
83         if (ops)
84                 smp_ops = *ops;
85 };
86
87 static unsigned long get_arch_pgd(pgd_t *pgd)
88 {
89         phys_addr_t pgdir = virt_to_idmap(pgd);
90         BUG_ON(pgdir & ARCH_PGD_MASK);
91         return pgdir >> ARCH_PGD_SHIFT;
92 }
93
94 int __cpu_up(unsigned int cpu, struct task_struct *idle)
95 {
96         int ret;
97
98         if (!smp_ops.smp_boot_secondary)
99                 return -ENOSYS;
100
101         /*
102          * We need to tell the secondary core where to find
103          * its stack and the page tables.
104          */
105         secondary_data.stack = task_stack_page(idle) + THREAD_START_SP;
106 #ifdef CONFIG_ARM_MPU
107         secondary_data.mpu_rgn_szr = mpu_rgn_info.rgns[MPU_RAM_REGION].drsr;
108 #endif
109
110 #ifdef CONFIG_MMU
111         secondary_data.pgdir = get_arch_pgd(idmap_pgd);
112         secondary_data.swapper_pg_dir = get_arch_pgd(swapper_pg_dir);
113 #endif
114         sync_cache_w(&secondary_data);
115
116         /*
117          * Now bring the CPU into our world.
118          */
119         ret = smp_ops.smp_boot_secondary(cpu, idle);
120         if (ret == 0) {
121                 /*
122                  * CPU was successfully started, wait for it
123                  * to come online or time out.
124                  */
125                 wait_for_completion_timeout(&cpu_running,
126                                                  msecs_to_jiffies(1000));
127
128                 if (!cpu_online(cpu)) {
129                         pr_crit("CPU%u: failed to come online\n", cpu);
130                         ret = -EIO;
131                 }
132         } else {
133                 pr_err("CPU%u: failed to boot: %d\n", cpu, ret);
134         }
135
136
137         memset(&secondary_data, 0, sizeof(secondary_data));
138         return ret;
139 }
140
141 /* platform specific SMP operations */
142 void __init smp_init_cpus(void)
143 {
144         if (smp_ops.smp_init_cpus)
145                 smp_ops.smp_init_cpus();
146 }
147
148 int platform_can_secondary_boot(void)
149 {
150         return !!smp_ops.smp_boot_secondary;
151 }
152
153 int platform_can_cpu_hotplug(void)
154 {
155 #ifdef CONFIG_HOTPLUG_CPU
156         if (smp_ops.cpu_kill)
157                 return 1;
158 #endif
159
160         return 0;
161 }
162
163 #ifdef CONFIG_HOTPLUG_CPU
164 static int platform_cpu_kill(unsigned int cpu)
165 {
166         if (smp_ops.cpu_kill)
167                 return smp_ops.cpu_kill(cpu);
168         return 1;
169 }
170
171 static int platform_cpu_disable(unsigned int cpu)
172 {
173         if (smp_ops.cpu_disable)
174                 return smp_ops.cpu_disable(cpu);
175
176         /*
177          * By default, allow disabling all CPUs except the first one,
178          * since this is special on a lot of platforms, e.g. because
179          * of clock tick interrupts.
180          */
181         return cpu == 0 ? -EPERM : 0;
182 }
183 /*
184  * __cpu_disable runs on the processor to be shutdown.
185  */
186 int __cpu_disable(void)
187 {
188         unsigned int cpu = smp_processor_id();
189         int ret;
190
191         ret = platform_cpu_disable(cpu);
192         if (ret)
193                 return ret;
194
195         /*
196          * Take this CPU offline.  Once we clear this, we can't return,
197          * and we must not schedule until we're ready to give up the cpu.
198          */
199         set_cpu_online(cpu, false);
200
201         /*
202          * OK - migrate IRQs away from this CPU
203          */
204         migrate_irqs();
205
206         /*
207          * Flush user cache and TLB mappings, and then remove this CPU
208          * from the vm mask set of all processes.
209          *
210          * Caches are flushed to the Level of Unification Inner Shareable
211          * to write-back dirty lines to unified caches shared by all CPUs.
212          */
213         flush_cache_louis();
214         local_flush_tlb_all();
215
216         return 0;
217 }
218
219 static DECLARE_COMPLETION(cpu_died);
220
221 /*
222  * called on the thread which is asking for a CPU to be shutdown -
223  * waits until shutdown has completed, or it is timed out.
224  */
225 void __cpu_die(unsigned int cpu)
226 {
227         if (!wait_for_completion_timeout(&cpu_died, msecs_to_jiffies(5000))) {
228                 pr_err("CPU%u: cpu didn't die\n", cpu);
229                 return;
230         }
231
232         clear_tasks_mm_cpumask(cpu);
233
234         pr_notice("CPU%u: shutdown\n", cpu);
235
236         /*
237          * platform_cpu_kill() is generally expected to do the powering off
238          * and/or cutting of clocks to the dying CPU.  Optionally, this may
239          * be done by the CPU which is dying in preference to supporting
240          * this call, but that means there is _no_ synchronisation between
241          * the requesting CPU and the dying CPU actually losing power.
242          */
243         if (!platform_cpu_kill(cpu))
244                 pr_err("CPU%u: unable to kill\n", cpu);
245 }
246
247 /*
248  * Called from the idle thread for the CPU which has been shutdown.
249  *
250  * Note that we disable IRQs here, but do not re-enable them
251  * before returning to the caller. This is also the behaviour
252  * of the other hotplug-cpu capable cores, so presumably coming
253  * out of idle fixes this.
254  */
255 void __ref cpu_die(void)
256 {
257         unsigned int cpu = smp_processor_id();
258
259         idle_task_exit();
260
261         local_irq_disable();
262
263         /*
264          * Flush the data out of the L1 cache for this CPU.  This must be
265          * before the completion to ensure that data is safely written out
266          * before platform_cpu_kill() gets called - which may disable
267          * *this* CPU and power down its cache.
268          */
269         flush_cache_louis();
270
271         /*
272          * Tell __cpu_die() that this CPU is now safe to dispose of.  Once
273          * this returns, power and/or clocks can be removed at any point
274          * from this CPU and its cache by platform_cpu_kill().
275          */
276         complete(&cpu_died);
277
278         /*
279          * Ensure that the cache lines associated with that completion are
280          * written out.  This covers the case where _this_ CPU is doing the
281          * powering down, to ensure that the completion is visible to the
282          * CPU waiting for this one.
283          */
284         flush_cache_louis();
285
286         /*
287          * The actual CPU shutdown procedure is at least platform (if not
288          * CPU) specific.  This may remove power, or it may simply spin.
289          *
290          * Platforms are generally expected *NOT* to return from this call,
291          * although there are some which do because they have no way to
292          * power down the CPU.  These platforms are the _only_ reason we
293          * have a return path which uses the fragment of assembly below.
294          *
295          * The return path should not be used for platforms which can
296          * power off the CPU.
297          */
298         if (smp_ops.cpu_die)
299                 smp_ops.cpu_die(cpu);
300
301         pr_warn("CPU%u: smp_ops.cpu_die() returned, trying to resuscitate\n",
302                 cpu);
303
304         /*
305          * Do not return to the idle loop - jump back to the secondary
306          * cpu initialisation.  There's some initialisation which needs
307          * to be repeated to undo the effects of taking the CPU offline.
308          */
309         __asm__("mov    sp, %0\n"
310         "       mov     fp, #0\n"
311         "       b       secondary_start_kernel"
312                 :
313                 : "r" (task_stack_page(current) + THREAD_SIZE - 8));
314 }
315 #endif /* CONFIG_HOTPLUG_CPU */
316
317 /*
318  * Called by both boot and secondaries to move global data into
319  * per-processor storage.
320  */
321 static void smp_store_cpu_info(unsigned int cpuid)
322 {
323         struct cpuinfo_arm *cpu_info = &per_cpu(cpu_data, cpuid);
324
325         cpu_info->loops_per_jiffy = loops_per_jiffy;
326         cpu_info->cpuid = read_cpuid_id();
327
328         store_cpu_topology(cpuid);
329 }
330
331 /*
332  * This is the secondary CPU boot entry.  We're using this CPUs
333  * idle thread stack, but a set of temporary page tables.
334  */
335 asmlinkage void secondary_start_kernel(void)
336 {
337         struct mm_struct *mm = &init_mm;
338         unsigned int cpu;
339
340         /*
341          * The identity mapping is uncached (strongly ordered), so
342          * switch away from it before attempting any exclusive accesses.
343          */
344         cpu_switch_mm(mm->pgd, mm);
345         local_flush_bp_all();
346         enter_lazy_tlb(mm, current);
347         local_flush_tlb_all();
348
349         /*
350          * All kernel threads share the same mm context; grab a
351          * reference and switch to it.
352          */
353         cpu = smp_processor_id();
354         atomic_inc(&mm->mm_count);
355         current->active_mm = mm;
356         cpumask_set_cpu(cpu, mm_cpumask(mm));
357
358         cpu_init();
359
360         pr_debug("CPU%u: Booted secondary processor\n", cpu);
361
362         preempt_disable();
363         trace_hardirqs_off();
364
365         /*
366          * Give the platform a chance to do its own initialisation.
367          */
368         if (smp_ops.smp_secondary_init)
369                 smp_ops.smp_secondary_init(cpu);
370
371         notify_cpu_starting(cpu);
372
373         calibrate_delay();
374
375         smp_store_cpu_info(cpu);
376
377         /*
378          * OK, now it's safe to let the boot CPU continue.  Wait for
379          * the CPU migration code to notice that the CPU is online
380          * before we continue - which happens after __cpu_up returns.
381          */
382         set_cpu_online(cpu, true);
383         complete(&cpu_running);
384
385         local_irq_enable();
386         local_fiq_enable();
387
388         /*
389          * OK, it's off to the idle thread for us
390          */
391         cpu_startup_entry(CPUHP_ONLINE);
392 }
393
394 void __init smp_cpus_done(unsigned int max_cpus)
395 {
396         int cpu;
397         unsigned long bogosum = 0;
398
399         for_each_online_cpu(cpu)
400                 bogosum += per_cpu(cpu_data, cpu).loops_per_jiffy;
401
402         printk(KERN_INFO "SMP: Total of %d processors activated "
403                "(%lu.%02lu BogoMIPS).\n",
404                num_online_cpus(),
405                bogosum / (500000/HZ),
406                (bogosum / (5000/HZ)) % 100);
407
408         hyp_mode_check();
409 }
410
411 void __init smp_prepare_boot_cpu(void)
412 {
413         set_my_cpu_offset(per_cpu_offset(smp_processor_id()));
414 }
415
416 void __init smp_prepare_cpus(unsigned int max_cpus)
417 {
418         unsigned int ncores = num_possible_cpus();
419
420         init_cpu_topology();
421
422         smp_store_cpu_info(smp_processor_id());
423
424         /*
425          * are we trying to boot more cores than exist?
426          */
427         if (max_cpus > ncores)
428                 max_cpus = ncores;
429         if (ncores > 1 && max_cpus) {
430                 /*
431                  * Initialise the present map, which describes the set of CPUs
432                  * actually populated at the present time. A platform should
433                  * re-initialize the map in the platforms smp_prepare_cpus()
434                  * if present != possible (e.g. physical hotplug).
435                  */
436                 init_cpu_present(cpu_possible_mask);
437
438                 /*
439                  * Initialise the SCU if there are more than one CPU
440                  * and let them know where to start.
441                  */
442                 if (smp_ops.smp_prepare_cpus)
443                         smp_ops.smp_prepare_cpus(max_cpus);
444         }
445 }
446
447 static void (*__smp_cross_call)(const struct cpumask *, unsigned int);
448
449 void __init set_smp_cross_call(void (*fn)(const struct cpumask *, unsigned int))
450 {
451         if (!__smp_cross_call)
452                 __smp_cross_call = fn;
453 }
454
455 static const char *ipi_types[NR_IPI] __tracepoint_string = {
456 #define S(x,s)  [x] = s
457         S(IPI_WAKEUP, "CPU wakeup interrupts"),
458         S(IPI_TIMER, "Timer broadcast interrupts"),
459         S(IPI_RESCHEDULE, "Rescheduling interrupts"),
460         S(IPI_CALL_FUNC, "Function call interrupts"),
461         S(IPI_CALL_FUNC_SINGLE, "Single function call interrupts"),
462         S(IPI_CPU_STOP, "CPU stop interrupts"),
463         S(IPI_IRQ_WORK, "IRQ work interrupts"),
464         S(IPI_COMPLETION, "completion interrupts"),
465 };
466
467 static void smp_cross_call(const struct cpumask *target, unsigned int ipinr)
468 {
469         trace_ipi_raise(target, ipi_types[ipinr]);
470         __smp_cross_call(target, ipinr);
471 }
472
473 void show_ipi_list(struct seq_file *p, int prec)
474 {
475         unsigned int cpu, i;
476
477         for (i = 0; i < NR_IPI; i++) {
478                 seq_printf(p, "%*s%u: ", prec - 1, "IPI", i);
479
480                 for_each_online_cpu(cpu)
481                         seq_printf(p, "%10u ",
482                                    __get_irq_stat(cpu, ipi_irqs[i]));
483
484                 seq_printf(p, " %s\n", ipi_types[i]);
485         }
486 }
487
488 u64 smp_irq_stat_cpu(unsigned int cpu)
489 {
490         u64 sum = 0;
491         int i;
492
493         for (i = 0; i < NR_IPI; i++)
494                 sum += __get_irq_stat(cpu, ipi_irqs[i]);
495
496         return sum;
497 }
498
499 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
500 {
501         smp_cross_call(mask, IPI_CALL_FUNC);
502 }
503
504 void arch_send_wakeup_ipi_mask(const struct cpumask *mask)
505 {
506         smp_cross_call(mask, IPI_WAKEUP);
507 }
508
509 void arch_send_call_function_single_ipi(int cpu)
510 {
511         smp_cross_call(cpumask_of(cpu), IPI_CALL_FUNC_SINGLE);
512 }
513
514 #ifdef CONFIG_IRQ_WORK
515 void arch_irq_work_raise(void)
516 {
517         if (arch_irq_work_has_interrupt())
518                 smp_cross_call(cpumask_of(smp_processor_id()), IPI_IRQ_WORK);
519 }
520 #endif
521
522 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
523 void tick_broadcast(const struct cpumask *mask)
524 {
525         smp_cross_call(mask, IPI_TIMER);
526 }
527 #endif
528
529 static DEFINE_RAW_SPINLOCK(stop_lock);
530
531 /*
532  * ipi_cpu_stop - handle IPI from smp_send_stop()
533  */
534 static void ipi_cpu_stop(unsigned int cpu)
535 {
536         if (system_state == SYSTEM_BOOTING ||
537             system_state == SYSTEM_RUNNING) {
538                 raw_spin_lock(&stop_lock);
539                 pr_crit("CPU%u: stopping\n", cpu);
540                 dump_stack();
541                 raw_spin_unlock(&stop_lock);
542         }
543
544         set_cpu_online(cpu, false);
545
546         local_fiq_disable();
547         local_irq_disable();
548
549         while (1)
550                 cpu_relax();
551 }
552
553 static DEFINE_PER_CPU(struct completion *, cpu_completion);
554
555 int register_ipi_completion(struct completion *completion, int cpu)
556 {
557         per_cpu(cpu_completion, cpu) = completion;
558         return IPI_COMPLETION;
559 }
560
561 static void ipi_complete(unsigned int cpu)
562 {
563         complete(per_cpu(cpu_completion, cpu));
564 }
565
566 /*
567  * Main handler for inter-processor interrupts
568  */
569 asmlinkage void __exception_irq_entry do_IPI(int ipinr, struct pt_regs *regs)
570 {
571         handle_IPI(ipinr, regs);
572 }
573
574 void handle_IPI(int ipinr, struct pt_regs *regs)
575 {
576         unsigned int cpu = smp_processor_id();
577         struct pt_regs *old_regs = set_irq_regs(regs);
578
579         if ((unsigned)ipinr < NR_IPI) {
580                 trace_ipi_entry_rcuidle(ipi_types[ipinr]);
581                 __inc_irq_stat(cpu, ipi_irqs[ipinr]);
582         }
583
584         switch (ipinr) {
585         case IPI_WAKEUP:
586                 break;
587
588 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
589         case IPI_TIMER:
590                 irq_enter();
591                 tick_receive_broadcast();
592                 irq_exit();
593                 break;
594 #endif
595
596         case IPI_RESCHEDULE:
597                 scheduler_ipi();
598                 break;
599
600         case IPI_CALL_FUNC:
601                 irq_enter();
602                 generic_smp_call_function_interrupt();
603                 irq_exit();
604                 break;
605
606         case IPI_CALL_FUNC_SINGLE:
607                 irq_enter();
608                 generic_smp_call_function_single_interrupt();
609                 irq_exit();
610                 break;
611
612         case IPI_CPU_STOP:
613                 irq_enter();
614                 ipi_cpu_stop(cpu);
615                 irq_exit();
616                 break;
617
618 #ifdef CONFIG_IRQ_WORK
619         case IPI_IRQ_WORK:
620                 irq_enter();
621                 irq_work_run();
622                 irq_exit();
623                 break;
624 #endif
625
626         case IPI_COMPLETION:
627                 irq_enter();
628                 ipi_complete(cpu);
629                 irq_exit();
630                 break;
631
632         default:
633                 pr_crit("CPU%u: Unknown IPI message 0x%x\n",
634                         cpu, ipinr);
635                 break;
636         }
637
638         if ((unsigned)ipinr < NR_IPI)
639                 trace_ipi_exit_rcuidle(ipi_types[ipinr]);
640         set_irq_regs(old_regs);
641 }
642
643 void smp_send_reschedule(int cpu)
644 {
645         smp_cross_call(cpumask_of(cpu), IPI_RESCHEDULE);
646 }
647
648 void smp_send_stop(void)
649 {
650         unsigned long timeout;
651         struct cpumask mask;
652
653         cpumask_copy(&mask, cpu_online_mask);
654         cpumask_clear_cpu(smp_processor_id(), &mask);
655         if (!cpumask_empty(&mask))
656                 smp_cross_call(&mask, IPI_CPU_STOP);
657
658         /* Wait up to one second for other CPUs to stop */
659         timeout = USEC_PER_SEC;
660         while (num_online_cpus() > 1 && timeout--)
661                 udelay(1);
662
663         if (num_online_cpus() > 1)
664                 pr_warn("SMP: failed to stop secondary CPUs\n");
665 }
666
667 /*
668  * not supported here
669  */
670 int setup_profiling_timer(unsigned int multiplier)
671 {
672         return -EINVAL;
673 }
674
675 #ifdef CONFIG_CPU_FREQ
676
677 static DEFINE_PER_CPU(unsigned long, l_p_j_ref);
678 static DEFINE_PER_CPU(unsigned long, l_p_j_ref_freq);
679 static unsigned long global_l_p_j_ref;
680 static unsigned long global_l_p_j_ref_freq;
681
682 static int cpufreq_callback(struct notifier_block *nb,
683                                         unsigned long val, void *data)
684 {
685         struct cpufreq_freqs *freq = data;
686         int cpu = freq->cpu;
687
688         if (freq->flags & CPUFREQ_CONST_LOOPS)
689                 return NOTIFY_OK;
690
691         if (!per_cpu(l_p_j_ref, cpu)) {
692                 per_cpu(l_p_j_ref, cpu) =
693                         per_cpu(cpu_data, cpu).loops_per_jiffy;
694                 per_cpu(l_p_j_ref_freq, cpu) = freq->old;
695                 if (!global_l_p_j_ref) {
696                         global_l_p_j_ref = loops_per_jiffy;
697                         global_l_p_j_ref_freq = freq->old;
698                 }
699         }
700
701         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
702             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new)) {
703                 loops_per_jiffy = cpufreq_scale(global_l_p_j_ref,
704                                                 global_l_p_j_ref_freq,
705                                                 freq->new);
706                 per_cpu(cpu_data, cpu).loops_per_jiffy =
707                         cpufreq_scale(per_cpu(l_p_j_ref, cpu),
708                                         per_cpu(l_p_j_ref_freq, cpu),
709                                         freq->new);
710         }
711         return NOTIFY_OK;
712 }
713
714 static struct notifier_block cpufreq_notifier = {
715         .notifier_call  = cpufreq_callback,
716 };
717
718 static int __init register_cpufreq_notifier(void)
719 {
720         return cpufreq_register_notifier(&cpufreq_notifier,
721                                                 CPUFREQ_TRANSITION_NOTIFIER);
722 }
723 core_initcall(register_cpufreq_notifier);
724
725 #endif