a521fb8a40ee2fcb206397cf490aab9550cf9c96
[linux-2.6.git] / arch / powerpc / kernel / smp.c
1 /*
2  * SMP support for ppc.
3  *
4  * Written by Cort Dougan (cort@cs.nmt.edu) borrowing a great
5  * deal of code from the sparc and intel versions.
6  *
7  * Copyright (C) 1999 Cort Dougan <cort@cs.nmt.edu>
8  *
9  * PowerPC-64 Support added by Dave Engebretsen, Peter Bergner, and
10  * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
11  *
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.
16  */
17
18 #undef DEBUG
19
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>
34
35 #include <asm/ptrace.h>
36 #include <asm/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
52
53 #ifdef DEBUG
54 #include <asm/udbg.h>
55 #define DBG(fmt...) udbg_printf(fmt)
56 #else
57 #define DBG(fmt...)
58 #endif
59
60 struct thread_info *secondary_ti;
61
62 DEFINE_PER_CPU(cpumask_t, cpu_sibling_map) = CPU_MASK_NONE;
63 DEFINE_PER_CPU(cpumask_t, cpu_core_map) = CPU_MASK_NONE;
64
65 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
66 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
67
68 /* SMP operations for this machine */
69 struct smp_ops_t *smp_ops;
70
71 /* Can't be static due to PowerMac hackery */
72 volatile unsigned int cpu_callin_map[NR_CPUS];
73
74 int smt_enabled_at_boot = 1;
75
76 static void (*crash_ipi_function_ptr)(struct pt_regs *) = NULL;
77
78 #ifdef CONFIG_PPC64
79 void __devinit smp_generic_kick_cpu(int nr)
80 {
81         BUG_ON(nr < 0 || nr >= NR_CPUS);
82
83         /*
84          * The processor is currently spinning, waiting for the
85          * cpu_start field to become non-zero After we set cpu_start,
86          * the processor will continue on to secondary_start
87          */
88         paca[nr].cpu_start = 1;
89         smp_mb();
90 }
91 #endif
92
93 void smp_message_recv(int msg)
94 {
95         switch(msg) {
96         case PPC_MSG_CALL_FUNCTION:
97                 generic_smp_call_function_interrupt();
98                 break;
99         case PPC_MSG_RESCHEDULE:
100                 /* we notice need_resched on exit */
101                 break;
102         case PPC_MSG_CALL_FUNC_SINGLE:
103                 generic_smp_call_function_single_interrupt();
104                 break;
105         case PPC_MSG_DEBUGGER_BREAK:
106                 if (crash_ipi_function_ptr) {
107                         crash_ipi_function_ptr(get_irq_regs());
108                         break;
109                 }
110 #ifdef CONFIG_DEBUGGER
111                 debugger_ipi(get_irq_regs());
112                 break;
113 #endif /* CONFIG_DEBUGGER */
114                 /* FALLTHROUGH */
115         default:
116                 printk("SMP %d: smp_message_recv(): unknown msg %d\n",
117                        smp_processor_id(), msg);
118                 break;
119         }
120 }
121
122 static irqreturn_t call_function_action(int irq, void *data)
123 {
124         generic_smp_call_function_interrupt();
125         return IRQ_HANDLED;
126 }
127
128 static irqreturn_t reschedule_action(int irq, void *data)
129 {
130         /* we just need the return path side effect of checking need_resched */
131         return IRQ_HANDLED;
132 }
133
134 static irqreturn_t call_function_single_action(int irq, void *data)
135 {
136         generic_smp_call_function_single_interrupt();
137         return IRQ_HANDLED;
138 }
139
140 static irqreturn_t debug_ipi_action(int irq, void *data)
141 {
142         smp_message_recv(PPC_MSG_DEBUGGER_BREAK);
143         return IRQ_HANDLED;
144 }
145
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,
151 };
152
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",
158 };
159
160 /* optional function to request ipi, for controllers with >= 4 ipis */
161 int smp_request_message_ipi(int virq, int msg)
162 {
163         int err;
164
165         if (msg < 0 || msg > PPC_MSG_DEBUGGER_BREAK) {
166                 return -EINVAL;
167         }
168 #if !defined(CONFIG_DEBUGGER) && !defined(CONFIG_KEXEC)
169         if (msg == PPC_MSG_DEBUGGER_BREAK) {
170                 return 1;
171         }
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);
177
178         return err;
179 }
180
181 void smp_send_reschedule(int cpu)
182 {
183         if (likely(smp_ops))
184                 smp_ops->message_pass(cpu, PPC_MSG_RESCHEDULE);
185 }
186
187 void arch_send_call_function_single_ipi(int cpu)
188 {
189         smp_ops->message_pass(cpu, PPC_MSG_CALL_FUNC_SINGLE);
190 }
191
192 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
193 {
194         unsigned int cpu;
195
196         for_each_cpu(cpu, mask)
197                 smp_ops->message_pass(cpu, PPC_MSG_CALL_FUNCTION);
198 }
199
200 #ifdef CONFIG_DEBUGGER
201 void smp_send_debugger_break(int cpu)
202 {
203         if (likely(smp_ops))
204                 smp_ops->message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
205 }
206 #endif
207
208 #ifdef CONFIG_KEXEC
209 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
210 {
211         crash_ipi_function_ptr = crash_ipi_callback;
212         if (crash_ipi_callback && smp_ops) {
213                 mb();
214                 smp_ops->message_pass(MSG_ALL_BUT_SELF, PPC_MSG_DEBUGGER_BREAK);
215         }
216 }
217 #endif
218
219 static void stop_this_cpu(void *dummy)
220 {
221         /* Remove this CPU */
222         set_cpu_online(smp_processor_id(), false);
223
224         local_irq_disable();
225         while (1)
226                 ;
227 }
228
229 void smp_send_stop(void)
230 {
231         smp_call_function(stop_this_cpu, NULL, 0);
232 }
233
234 struct thread_info *current_set[NR_CPUS];
235
236 static void __devinit smp_store_cpu_info(int id)
237 {
238         per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
239 }
240
241 static void __init smp_create_idle(unsigned int cpu)
242 {
243         struct task_struct *p;
244
245         /* create a process for the processor */
246         p = fork_idle(cpu);
247         if (IS_ERR(p))
248                 panic("failed fork for CPU %u: %li", cpu, PTR_ERR(p));
249 #ifdef CONFIG_PPC64
250         paca[cpu].__current = p;
251         paca[cpu].kstack = (unsigned long) task_thread_info(p)
252                 + THREAD_SIZE - STACK_FRAME_OVERHEAD;
253 #endif
254         current_set[cpu] = task_thread_info(p);
255         task_thread_info(p)->cpu = cpu;
256 }
257
258 void __init smp_prepare_cpus(unsigned int max_cpus)
259 {
260         unsigned int cpu;
261
262         DBG("smp_prepare_cpus\n");
263
264         /* 
265          * setup_cpu may need to be called on the boot cpu. We havent
266          * spun any cpus up but lets be paranoid.
267          */
268         BUG_ON(boot_cpuid != smp_processor_id());
269
270         /* Fixup boot cpu */
271         smp_store_cpu_info(boot_cpuid);
272         cpu_callin_map[boot_cpuid] = 1;
273
274         if (smp_ops)
275                 if (smp_ops->probe)
276                         max_cpus = smp_ops->probe();
277                 else
278                         max_cpus = NR_CPUS;
279         else
280                 max_cpus = 1;
281  
282         smp_space_timers(max_cpus);
283
284         for_each_possible_cpu(cpu)
285                 if (cpu != boot_cpuid)
286                         smp_create_idle(cpu);
287 }
288
289 void __devinit smp_prepare_boot_cpu(void)
290 {
291         BUG_ON(smp_processor_id() != boot_cpuid);
292
293         set_cpu_online(boot_cpuid, true);
294         cpu_set(boot_cpuid, per_cpu(cpu_sibling_map, boot_cpuid));
295         cpu_set(boot_cpuid, per_cpu(cpu_core_map, boot_cpuid));
296 #ifdef CONFIG_PPC64
297         paca[boot_cpuid].__current = current;
298 #endif
299         current_set[boot_cpuid] = task_thread_info(current);
300 }
301
302 #ifdef CONFIG_HOTPLUG_CPU
303 /* State of each CPU during hotplug phases */
304 DEFINE_PER_CPU(int, cpu_state) = { 0 };
305
306 int generic_cpu_disable(void)
307 {
308         unsigned int cpu = smp_processor_id();
309
310         if (cpu == boot_cpuid)
311                 return -EBUSY;
312
313         set_cpu_online(cpu, false);
314 #ifdef CONFIG_PPC64
315         vdso_data->processorCount--;
316         fixup_irqs(cpu_online_map);
317 #endif
318         return 0;
319 }
320
321 int generic_cpu_enable(unsigned int cpu)
322 {
323         /* Do the normal bootup if we haven't
324          * already bootstrapped. */
325         if (system_state != SYSTEM_RUNNING)
326                 return -ENOSYS;
327
328         /* get the target out of it's holding state */
329         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
330         smp_wmb();
331
332         while (!cpu_online(cpu))
333                 cpu_relax();
334
335 #ifdef CONFIG_PPC64
336         fixup_irqs(cpu_online_map);
337         /* counter the irq disable in fixup_irqs */
338         local_irq_enable();
339 #endif
340         return 0;
341 }
342
343 void generic_cpu_die(unsigned int cpu)
344 {
345         int i;
346
347         for (i = 0; i < 100; i++) {
348                 smp_rmb();
349                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
350                         return;
351                 msleep(100);
352         }
353         printk(KERN_ERR "CPU%d didn't die...\n", cpu);
354 }
355
356 void generic_mach_cpu_die(void)
357 {
358         unsigned int cpu;
359
360         local_irq_disable();
361         cpu = smp_processor_id();
362         printk(KERN_DEBUG "CPU%d offline\n", cpu);
363         __get_cpu_var(cpu_state) = CPU_DEAD;
364         smp_wmb();
365         while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
366                 cpu_relax();
367         set_cpu_online(cpu, true);
368         local_irq_enable();
369 }
370 #endif
371
372 static int __devinit cpu_enable(unsigned int cpu)
373 {
374         if (smp_ops && smp_ops->cpu_enable)
375                 return smp_ops->cpu_enable(cpu);
376
377         return -ENOSYS;
378 }
379
380 int __cpuinit __cpu_up(unsigned int cpu)
381 {
382         int c;
383
384         secondary_ti = current_set[cpu];
385         if (!cpu_enable(cpu))
386                 return 0;
387
388         if (smp_ops == NULL ||
389             (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
390                 return -EINVAL;
391
392         /* Make sure callin-map entry is 0 (can be leftover a CPU
393          * hotplug
394          */
395         cpu_callin_map[cpu] = 0;
396
397         /* The information for processor bringup must
398          * be written out to main store before we release
399          * the processor.
400          */
401         smp_mb();
402
403         /* wake up cpus */
404         DBG("smp: kicking cpu %d\n", cpu);
405         smp_ops->kick_cpu(cpu);
406
407         /*
408          * wait to see if the cpu made a callin (is actually up).
409          * use this value that I found through experimentation.
410          * -- Cort
411          */
412         if (system_state < SYSTEM_RUNNING)
413                 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
414                         udelay(100);
415 #ifdef CONFIG_HOTPLUG_CPU
416         else
417                 /*
418                  * CPUs can take much longer to come up in the
419                  * hotplug case.  Wait five seconds.
420                  */
421                 for (c = 5000; c && !cpu_callin_map[cpu]; c--)
422                         msleep(1);
423 #endif
424
425         if (!cpu_callin_map[cpu]) {
426                 printk("Processor %u is stuck.\n", cpu);
427                 return -ENOENT;
428         }
429
430         printk("Processor %u found.\n", cpu);
431
432         if (smp_ops->give_timebase)
433                 smp_ops->give_timebase();
434
435         /* Wait until cpu puts itself in the online map */
436         while (!cpu_online(cpu))
437                 cpu_relax();
438
439         return 0;
440 }
441
442 /* Return the value of the reg property corresponding to the given
443  * logical cpu.
444  */
445 int cpu_to_core_id(int cpu)
446 {
447         struct device_node *np;
448         const int *reg;
449         int id = -1;
450
451         np = of_get_cpu_node(cpu, NULL);
452         if (!np)
453                 goto out;
454
455         reg = of_get_property(np, "reg", NULL);
456         if (!reg)
457                 goto out;
458
459         id = *reg;
460 out:
461         of_node_put(np);
462         return id;
463 }
464
465 /* Must be called when no change can occur to cpu_present_map,
466  * i.e. during cpu online or offline.
467  */
468 static struct device_node *cpu_to_l2cache(int cpu)
469 {
470         struct device_node *np;
471         struct device_node *cache;
472
473         if (!cpu_present(cpu))
474                 return NULL;
475
476         np = of_get_cpu_node(cpu, NULL);
477         if (np == NULL)
478                 return NULL;
479
480         cache = of_find_next_cache_node(np);
481
482         of_node_put(np);
483
484         return cache;
485 }
486
487 /* Activate a secondary processor. */
488 int __devinit start_secondary(void *unused)
489 {
490         unsigned int cpu = smp_processor_id();
491         struct device_node *l2_cache;
492         int i, base;
493
494         atomic_inc(&init_mm.mm_count);
495         current->active_mm = &init_mm;
496
497         smp_store_cpu_info(cpu);
498         set_dec(tb_ticks_per_jiffy);
499         preempt_disable();
500         cpu_callin_map[cpu] = 1;
501
502         if (smp_ops->setup_cpu)
503                 smp_ops->setup_cpu(cpu);
504         if (smp_ops->take_timebase)
505                 smp_ops->take_timebase();
506
507         if (system_state > SYSTEM_BOOTING)
508                 snapshot_timebase();
509
510         secondary_cpu_time_init();
511
512         ipi_call_lock();
513         notify_cpu_starting(cpu);
514         set_cpu_online(cpu, true);
515         /* Update sibling maps */
516         base = cpu_first_thread_in_core(cpu);
517         for (i = 0; i < threads_per_core; i++) {
518                 if (cpu_is_offline(base + i))
519                         continue;
520                 cpu_set(cpu, per_cpu(cpu_sibling_map, base + i));
521                 cpu_set(base + i, per_cpu(cpu_sibling_map, cpu));
522
523                 /* cpu_core_map should be a superset of
524                  * cpu_sibling_map even if we don't have cache
525                  * information, so update the former here, too.
526                  */
527                 cpu_set(cpu, per_cpu(cpu_core_map, base +i));
528                 cpu_set(base + i, per_cpu(cpu_core_map, cpu));
529         }
530         l2_cache = cpu_to_l2cache(cpu);
531         for_each_online_cpu(i) {
532                 struct device_node *np = cpu_to_l2cache(i);
533                 if (!np)
534                         continue;
535                 if (np == l2_cache) {
536                         cpu_set(cpu, per_cpu(cpu_core_map, i));
537                         cpu_set(i, per_cpu(cpu_core_map, cpu));
538                 }
539                 of_node_put(np);
540         }
541         of_node_put(l2_cache);
542         ipi_call_unlock();
543
544         local_irq_enable();
545
546         cpu_idle();
547         return 0;
548 }
549
550 int setup_profiling_timer(unsigned int multiplier)
551 {
552         return 0;
553 }
554
555 void __init smp_cpus_done(unsigned int max_cpus)
556 {
557         cpumask_t old_mask;
558
559         /* We want the setup_cpu() here to be called from CPU 0, but our
560          * init thread may have been "borrowed" by another CPU in the meantime
561          * se we pin us down to CPU 0 for a short while
562          */
563         old_mask = current->cpus_allowed;
564         set_cpus_allowed(current, cpumask_of_cpu(boot_cpuid));
565         
566         if (smp_ops && smp_ops->setup_cpu)
567                 smp_ops->setup_cpu(boot_cpuid);
568
569         set_cpus_allowed(current, old_mask);
570
571         snapshot_timebases();
572
573         dump_numa_cpu_topology();
574 }
575
576 #ifdef CONFIG_HOTPLUG_CPU
577 int __cpu_disable(void)
578 {
579         struct device_node *l2_cache;
580         int cpu = smp_processor_id();
581         int base, i;
582         int err;
583
584         if (!smp_ops->cpu_disable)
585                 return -ENOSYS;
586
587         err = smp_ops->cpu_disable();
588         if (err)
589                 return err;
590
591         /* Update sibling maps */
592         base = cpu_first_thread_in_core(cpu);
593         for (i = 0; i < threads_per_core; i++) {
594                 cpu_clear(cpu, per_cpu(cpu_sibling_map, base + i));
595                 cpu_clear(base + i, per_cpu(cpu_sibling_map, cpu));
596                 cpu_clear(cpu, per_cpu(cpu_core_map, base +i));
597                 cpu_clear(base + i, per_cpu(cpu_core_map, cpu));
598         }
599
600         l2_cache = cpu_to_l2cache(cpu);
601         for_each_present_cpu(i) {
602                 struct device_node *np = cpu_to_l2cache(i);
603                 if (!np)
604                         continue;
605                 if (np == l2_cache) {
606                         cpu_clear(cpu, per_cpu(cpu_core_map, i));
607                         cpu_clear(i, per_cpu(cpu_core_map, cpu));
608                 }
609                 of_node_put(np);
610         }
611         of_node_put(l2_cache);
612
613
614         return 0;
615 }
616
617 void __cpu_die(unsigned int cpu)
618 {
619         if (smp_ops->cpu_die)
620                 smp_ops->cpu_die(cpu);
621 }
622 #endif