]> nv-tegra.nvidia Code Review - linux-3.10.git/blob - arch/arm/kernel/smp.c
Thumb-2: Implementation of the unified start-up and exceptions code
[linux-3.10.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/smp.h>
23 #include <linux/seq_file.h>
24 #include <linux/irq.h>
25 #include <linux/percpu.h>
26 #include <linux/clockchips.h>
27
28 #include <asm/atomic.h>
29 #include <asm/cacheflush.h>
30 #include <asm/cpu.h>
31 #include <asm/cputype.h>
32 #include <asm/mmu_context.h>
33 #include <asm/pgtable.h>
34 #include <asm/pgalloc.h>
35 #include <asm/processor.h>
36 #include <asm/tlbflush.h>
37 #include <asm/ptrace.h>
38 #include <asm/localtimer.h>
39
40 /*
41  * as from 2.5, kernels no longer have an init_tasks structure
42  * so we need some other way of telling a new secondary core
43  * where to place its SVC stack
44  */
45 struct secondary_data secondary_data;
46
47 /*
48  * structures for inter-processor calls
49  * - A collection of single bit ipi messages.
50  */
51 struct ipi_data {
52         spinlock_t lock;
53         unsigned long ipi_count;
54         unsigned long bits;
55 };
56
57 static DEFINE_PER_CPU(struct ipi_data, ipi_data) = {
58         .lock   = SPIN_LOCK_UNLOCKED,
59 };
60
61 enum ipi_msg_type {
62         IPI_TIMER,
63         IPI_RESCHEDULE,
64         IPI_CALL_FUNC,
65         IPI_CALL_FUNC_SINGLE,
66         IPI_CPU_STOP,
67 };
68
69 int __cpuinit __cpu_up(unsigned int cpu)
70 {
71         struct cpuinfo_arm *ci = &per_cpu(cpu_data, cpu);
72         struct task_struct *idle = ci->idle;
73         pgd_t *pgd;
74         pmd_t *pmd;
75         int ret;
76
77         /*
78          * Spawn a new process manually, if not already done.
79          * Grab a pointer to its task struct so we can mess with it
80          */
81         if (!idle) {
82                 idle = fork_idle(cpu);
83                 if (IS_ERR(idle)) {
84                         printk(KERN_ERR "CPU%u: fork() failed\n", cpu);
85                         return PTR_ERR(idle);
86                 }
87                 ci->idle = idle;
88         }
89
90         /*
91          * Allocate initial page tables to allow the new CPU to
92          * enable the MMU safely.  This essentially means a set
93          * of our "standard" page tables, with the addition of
94          * a 1:1 mapping for the physical address of the kernel.
95          */
96         pgd = pgd_alloc(&init_mm);
97         pmd = pmd_offset(pgd + pgd_index(PHYS_OFFSET), PHYS_OFFSET);
98         *pmd = __pmd((PHYS_OFFSET & PGDIR_MASK) |
99                      PMD_TYPE_SECT | PMD_SECT_AP_WRITE);
100         flush_pmd_entry(pmd);
101
102         /*
103          * We need to tell the secondary core where to find
104          * its stack and the page tables.
105          */
106         secondary_data.stack = task_stack_page(idle) + THREAD_START_SP;
107         secondary_data.pgdir = virt_to_phys(pgd);
108         wmb();
109
110         /*
111          * Now bring the CPU into our world.
112          */
113         ret = boot_secondary(cpu, idle);
114         if (ret == 0) {
115                 unsigned long timeout;
116
117                 /*
118                  * CPU was successfully started, wait for it
119                  * to come online or time out.
120                  */
121                 timeout = jiffies + HZ;
122                 while (time_before(jiffies, timeout)) {
123                         if (cpu_online(cpu))
124                                 break;
125
126                         udelay(10);
127                         barrier();
128                 }
129
130                 if (!cpu_online(cpu))
131                         ret = -EIO;
132         }
133
134         secondary_data.stack = NULL;
135         secondary_data.pgdir = 0;
136
137         *pmd = __pmd(0);
138         clean_pmd_entry(pmd);
139         pgd_free(&init_mm, pgd);
140
141         if (ret) {
142                 printk(KERN_CRIT "CPU%u: processor failed to boot\n", cpu);
143
144                 /*
145                  * FIXME: We need to clean up the new idle thread. --rmk
146                  */
147         }
148
149         return ret;
150 }
151
152 #ifdef CONFIG_HOTPLUG_CPU
153 /*
154  * __cpu_disable runs on the processor to be shutdown.
155  */
156 int __cpuexit __cpu_disable(void)
157 {
158         unsigned int cpu = smp_processor_id();
159         struct task_struct *p;
160         int ret;
161
162         ret = mach_cpu_disable(cpu);
163         if (ret)
164                 return ret;
165
166         /*
167          * Take this CPU offline.  Once we clear this, we can't return,
168          * and we must not schedule until we're ready to give up the cpu.
169          */
170         set_cpu_online(cpu, false);
171
172         /*
173          * OK - migrate IRQs away from this CPU
174          */
175         migrate_irqs();
176
177         /*
178          * Stop the local timer for this CPU.
179          */
180         local_timer_stop();
181
182         /*
183          * Flush user cache and TLB mappings, and then remove this CPU
184          * from the vm mask set of all processes.
185          */
186         flush_cache_all();
187         local_flush_tlb_all();
188
189         read_lock(&tasklist_lock);
190         for_each_process(p) {
191                 if (p->mm)
192                         cpu_clear(cpu, p->mm->cpu_vm_mask);
193         }
194         read_unlock(&tasklist_lock);
195
196         return 0;
197 }
198
199 /*
200  * called on the thread which is asking for a CPU to be shutdown -
201  * waits until shutdown has completed, or it is timed out.
202  */
203 void __cpuexit __cpu_die(unsigned int cpu)
204 {
205         if (!platform_cpu_kill(cpu))
206                 printk("CPU%u: unable to kill\n", cpu);
207 }
208
209 /*
210  * Called from the idle thread for the CPU which has been shutdown.
211  *
212  * Note that we disable IRQs here, but do not re-enable them
213  * before returning to the caller. This is also the behaviour
214  * of the other hotplug-cpu capable cores, so presumably coming
215  * out of idle fixes this.
216  */
217 void __cpuexit cpu_die(void)
218 {
219         unsigned int cpu = smp_processor_id();
220
221         local_irq_disable();
222         idle_task_exit();
223
224         /*
225          * actual CPU shutdown procedure is at least platform (if not
226          * CPU) specific
227          */
228         platform_cpu_die(cpu);
229
230         /*
231          * Do not return to the idle loop - jump back to the secondary
232          * cpu initialisation.  There's some initialisation which needs
233          * to be repeated to undo the effects of taking the CPU offline.
234          */
235         __asm__("mov    sp, %0\n"
236         "       b       secondary_start_kernel"
237                 :
238                 : "r" (task_stack_page(current) + THREAD_SIZE - 8));
239 }
240 #endif /* CONFIG_HOTPLUG_CPU */
241
242 /*
243  * This is the secondary CPU boot entry.  We're using this CPUs
244  * idle thread stack, but a set of temporary page tables.
245  */
246 asmlinkage void __cpuinit secondary_start_kernel(void)
247 {
248         struct mm_struct *mm = &init_mm;
249         unsigned int cpu = smp_processor_id();
250
251         printk("CPU%u: Booted secondary processor\n", cpu);
252
253         /*
254          * All kernel threads share the same mm context; grab a
255          * reference and switch to it.
256          */
257         atomic_inc(&mm->mm_users);
258         atomic_inc(&mm->mm_count);
259         current->active_mm = mm;
260         cpu_set(cpu, mm->cpu_vm_mask);
261         cpu_switch_mm(mm->pgd, mm);
262         enter_lazy_tlb(mm, current);
263         local_flush_tlb_all();
264
265         cpu_init();
266         preempt_disable();
267
268         /*
269          * Give the platform a chance to do its own initialisation.
270          */
271         platform_secondary_init(cpu);
272
273         /*
274          * Enable local interrupts.
275          */
276         notify_cpu_starting(cpu);
277         local_irq_enable();
278         local_fiq_enable();
279
280         /*
281          * Setup the percpu timer for this CPU.
282          */
283         percpu_timer_setup();
284
285         calibrate_delay();
286
287         smp_store_cpu_info(cpu);
288
289         /*
290          * OK, now it's safe to let the boot CPU continue
291          */
292         set_cpu_online(cpu, true);
293
294         /*
295          * OK, it's off to the idle thread for us
296          */
297         cpu_idle();
298 }
299
300 /*
301  * Called by both boot and secondaries to move global data into
302  * per-processor storage.
303  */
304 void __cpuinit smp_store_cpu_info(unsigned int cpuid)
305 {
306         struct cpuinfo_arm *cpu_info = &per_cpu(cpu_data, cpuid);
307
308         cpu_info->loops_per_jiffy = loops_per_jiffy;
309 }
310
311 void __init smp_cpus_done(unsigned int max_cpus)
312 {
313         int cpu;
314         unsigned long bogosum = 0;
315
316         for_each_online_cpu(cpu)
317                 bogosum += per_cpu(cpu_data, cpu).loops_per_jiffy;
318
319         printk(KERN_INFO "SMP: Total of %d processors activated "
320                "(%lu.%02lu BogoMIPS).\n",
321                num_online_cpus(),
322                bogosum / (500000/HZ),
323                (bogosum / (5000/HZ)) % 100);
324 }
325
326 void __init smp_prepare_boot_cpu(void)
327 {
328         unsigned int cpu = smp_processor_id();
329
330         per_cpu(cpu_data, cpu).idle = current;
331 }
332
333 static void send_ipi_message(const struct cpumask *mask, enum ipi_msg_type msg)
334 {
335         unsigned long flags;
336         unsigned int cpu;
337
338         local_irq_save(flags);
339
340         for_each_cpu(cpu, mask) {
341                 struct ipi_data *ipi = &per_cpu(ipi_data, cpu);
342
343                 spin_lock(&ipi->lock);
344                 ipi->bits |= 1 << msg;
345                 spin_unlock(&ipi->lock);
346         }
347
348         /*
349          * Call the platform specific cross-CPU call function.
350          */
351         smp_cross_call(mask);
352
353         local_irq_restore(flags);
354 }
355
356 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
357 {
358         send_ipi_message(mask, IPI_CALL_FUNC);
359 }
360
361 void arch_send_call_function_single_ipi(int cpu)
362 {
363         send_ipi_message(cpumask_of(cpu), IPI_CALL_FUNC_SINGLE);
364 }
365
366 void show_ipi_list(struct seq_file *p)
367 {
368         unsigned int cpu;
369
370         seq_puts(p, "IPI:");
371
372         for_each_present_cpu(cpu)
373                 seq_printf(p, " %10lu", per_cpu(ipi_data, cpu).ipi_count);
374
375         seq_putc(p, '\n');
376 }
377
378 void show_local_irqs(struct seq_file *p)
379 {
380         unsigned int cpu;
381
382         seq_printf(p, "LOC: ");
383
384         for_each_present_cpu(cpu)
385                 seq_printf(p, "%10u ", irq_stat[cpu].local_timer_irqs);
386
387         seq_putc(p, '\n');
388 }
389
390 /*
391  * Timer (local or broadcast) support
392  */
393 static DEFINE_PER_CPU(struct clock_event_device, percpu_clockevent);
394
395 static void ipi_timer(void)
396 {
397         struct clock_event_device *evt = &__get_cpu_var(percpu_clockevent);
398         irq_enter();
399         evt->event_handler(evt);
400         irq_exit();
401 }
402
403 #ifdef CONFIG_LOCAL_TIMERS
404 asmlinkage void __exception do_local_timer(struct pt_regs *regs)
405 {
406         struct pt_regs *old_regs = set_irq_regs(regs);
407         int cpu = smp_processor_id();
408
409         if (local_timer_ack()) {
410                 irq_stat[cpu].local_timer_irqs++;
411                 ipi_timer();
412         }
413
414         set_irq_regs(old_regs);
415 }
416 #endif
417
418 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
419 static void smp_timer_broadcast(const struct cpumask *mask)
420 {
421         send_ipi_message(mask, IPI_TIMER);
422 }
423
424 static void broadcast_timer_set_mode(enum clock_event_mode mode,
425         struct clock_event_device *evt)
426 {
427 }
428
429 static void local_timer_setup(struct clock_event_device *evt)
430 {
431         evt->name       = "dummy_timer";
432         evt->features   = CLOCK_EVT_FEAT_ONESHOT |
433                           CLOCK_EVT_FEAT_PERIODIC |
434                           CLOCK_EVT_FEAT_DUMMY;
435         evt->rating     = 400;
436         evt->mult       = 1;
437         evt->set_mode   = broadcast_timer_set_mode;
438         evt->broadcast  = smp_timer_broadcast;
439
440         clockevents_register_device(evt);
441 }
442 #endif
443
444 void __cpuinit percpu_timer_setup(void)
445 {
446         unsigned int cpu = smp_processor_id();
447         struct clock_event_device *evt = &per_cpu(percpu_clockevent, cpu);
448
449         evt->cpumask = cpumask_of(cpu);
450
451         local_timer_setup(evt);
452 }
453
454 static DEFINE_SPINLOCK(stop_lock);
455
456 /*
457  * ipi_cpu_stop - handle IPI from smp_send_stop()
458  */
459 static void ipi_cpu_stop(unsigned int cpu)
460 {
461         spin_lock(&stop_lock);
462         printk(KERN_CRIT "CPU%u: stopping\n", cpu);
463         dump_stack();
464         spin_unlock(&stop_lock);
465
466         set_cpu_online(cpu, false);
467
468         local_fiq_disable();
469         local_irq_disable();
470
471         while (1)
472                 cpu_relax();
473 }
474
475 /*
476  * Main handler for inter-processor interrupts
477  *
478  * For ARM, the ipimask now only identifies a single
479  * category of IPI (Bit 1 IPIs have been replaced by a
480  * different mechanism):
481  *
482  *  Bit 0 - Inter-processor function call
483  */
484 asmlinkage void __exception do_IPI(struct pt_regs *regs)
485 {
486         unsigned int cpu = smp_processor_id();
487         struct ipi_data *ipi = &per_cpu(ipi_data, cpu);
488         struct pt_regs *old_regs = set_irq_regs(regs);
489
490         ipi->ipi_count++;
491
492         for (;;) {
493                 unsigned long msgs;
494
495                 spin_lock(&ipi->lock);
496                 msgs = ipi->bits;
497                 ipi->bits = 0;
498                 spin_unlock(&ipi->lock);
499
500                 if (!msgs)
501                         break;
502
503                 do {
504                         unsigned nextmsg;
505
506                         nextmsg = msgs & -msgs;
507                         msgs &= ~nextmsg;
508                         nextmsg = ffz(~nextmsg);
509
510                         switch (nextmsg) {
511                         case IPI_TIMER:
512                                 ipi_timer();
513                                 break;
514
515                         case IPI_RESCHEDULE:
516                                 /*
517                                  * nothing more to do - eveything is
518                                  * done on the interrupt return path
519                                  */
520                                 break;
521
522                         case IPI_CALL_FUNC:
523                                 generic_smp_call_function_interrupt();
524                                 break;
525
526                         case IPI_CALL_FUNC_SINGLE:
527                                 generic_smp_call_function_single_interrupt();
528                                 break;
529
530                         case IPI_CPU_STOP:
531                                 ipi_cpu_stop(cpu);
532                                 break;
533
534                         default:
535                                 printk(KERN_CRIT "CPU%u: Unknown IPI message 0x%x\n",
536                                        cpu, nextmsg);
537                                 break;
538                         }
539                 } while (msgs);
540         }
541
542         set_irq_regs(old_regs);
543 }
544
545 void smp_send_reschedule(int cpu)
546 {
547         send_ipi_message(cpumask_of(cpu), IPI_RESCHEDULE);
548 }
549
550 void smp_send_stop(void)
551 {
552         cpumask_t mask = cpu_online_map;
553         cpu_clear(smp_processor_id(), mask);
554         send_ipi_message(&mask, IPI_CPU_STOP);
555 }
556
557 /*
558  * not supported here
559  */
560 int setup_profiling_timer(unsigned int multiplier)
561 {
562         return -EINVAL;
563 }
564
565 static void
566 on_each_cpu_mask(void (*func)(void *), void *info, int wait,
567                 const struct cpumask *mask)
568 {
569         preempt_disable();
570
571         smp_call_function_many(mask, func, info, wait);
572         if (cpumask_test_cpu(smp_processor_id(), mask))
573                 func(info);
574
575         preempt_enable();
576 }
577
578 /**********************************************************************/
579
580 /*
581  * TLB operations
582  */
583 struct tlb_args {
584         struct vm_area_struct *ta_vma;
585         unsigned long ta_start;
586         unsigned long ta_end;
587 };
588
589 /* all SMP configurations have the extended CPUID registers */
590 static inline int tlb_ops_need_broadcast(void)
591 {
592         return ((read_cpuid_ext(CPUID_EXT_MMFR3) >> 12) & 0xf) < 2;
593 }
594
595 static inline void ipi_flush_tlb_all(void *ignored)
596 {
597         local_flush_tlb_all();
598 }
599
600 static inline void ipi_flush_tlb_mm(void *arg)
601 {
602         struct mm_struct *mm = (struct mm_struct *)arg;
603
604         local_flush_tlb_mm(mm);
605 }
606
607 static inline void ipi_flush_tlb_page(void *arg)
608 {
609         struct tlb_args *ta = (struct tlb_args *)arg;
610
611         local_flush_tlb_page(ta->ta_vma, ta->ta_start);
612 }
613
614 static inline void ipi_flush_tlb_kernel_page(void *arg)
615 {
616         struct tlb_args *ta = (struct tlb_args *)arg;
617
618         local_flush_tlb_kernel_page(ta->ta_start);
619 }
620
621 static inline void ipi_flush_tlb_range(void *arg)
622 {
623         struct tlb_args *ta = (struct tlb_args *)arg;
624
625         local_flush_tlb_range(ta->ta_vma, ta->ta_start, ta->ta_end);
626 }
627
628 static inline void ipi_flush_tlb_kernel_range(void *arg)
629 {
630         struct tlb_args *ta = (struct tlb_args *)arg;
631
632         local_flush_tlb_kernel_range(ta->ta_start, ta->ta_end);
633 }
634
635 void flush_tlb_all(void)
636 {
637         if (tlb_ops_need_broadcast())
638                 on_each_cpu(ipi_flush_tlb_all, NULL, 1);
639         else
640                 local_flush_tlb_all();
641 }
642
643 void flush_tlb_mm(struct mm_struct *mm)
644 {
645         if (tlb_ops_need_broadcast())
646                 on_each_cpu_mask(ipi_flush_tlb_mm, mm, 1, &mm->cpu_vm_mask);
647         else
648                 local_flush_tlb_mm(mm);
649 }
650
651 void flush_tlb_page(struct vm_area_struct *vma, unsigned long uaddr)
652 {
653         if (tlb_ops_need_broadcast()) {
654                 struct tlb_args ta;
655                 ta.ta_vma = vma;
656                 ta.ta_start = uaddr;
657                 on_each_cpu_mask(ipi_flush_tlb_page, &ta, 1, &vma->vm_mm->cpu_vm_mask);
658         } else
659                 local_flush_tlb_page(vma, uaddr);
660 }
661
662 void flush_tlb_kernel_page(unsigned long kaddr)
663 {
664         if (tlb_ops_need_broadcast()) {
665                 struct tlb_args ta;
666                 ta.ta_start = kaddr;
667                 on_each_cpu(ipi_flush_tlb_kernel_page, &ta, 1);
668         } else
669                 local_flush_tlb_kernel_page(kaddr);
670 }
671
672 void flush_tlb_range(struct vm_area_struct *vma,
673                      unsigned long start, unsigned long end)
674 {
675         if (tlb_ops_need_broadcast()) {
676                 struct tlb_args ta;
677                 ta.ta_vma = vma;
678                 ta.ta_start = start;
679                 ta.ta_end = end;
680                 on_each_cpu_mask(ipi_flush_tlb_range, &ta, 1, &vma->vm_mm->cpu_vm_mask);
681         } else
682                 local_flush_tlb_range(vma, start, end);
683 }
684
685 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
686 {
687         if (tlb_ops_need_broadcast()) {
688                 struct tlb_args ta;
689                 ta.ta_start = start;
690                 ta.ta_end = end;
691                 on_each_cpu(ipi_flush_tlb_kernel_range, &ta, 1);
692         } else
693                 local_flush_tlb_kernel_range(start, end);
694 }