Merge commit 'origin'
[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 int smp_hw_index[NR_CPUS];
61 struct thread_info *secondary_ti;
62
63 cpumask_t cpu_possible_map = CPU_MASK_NONE;
64 cpumask_t cpu_online_map = CPU_MASK_NONE;
65 DEFINE_PER_CPU(cpumask_t, cpu_sibling_map) = CPU_MASK_NONE;
66 DEFINE_PER_CPU(cpumask_t, cpu_core_map) = CPU_MASK_NONE;
67
68 EXPORT_SYMBOL(cpu_online_map);
69 EXPORT_SYMBOL(cpu_possible_map);
70 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
71 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
72
73 /* SMP operations for this machine */
74 struct smp_ops_t *smp_ops;
75
76 static volatile unsigned int cpu_callin_map[NR_CPUS];
77
78 int smt_enabled_at_boot = 1;
79
80 static void (*crash_ipi_function_ptr)(struct pt_regs *) = NULL;
81
82 #ifdef CONFIG_PPC64
83 void __devinit smp_generic_kick_cpu(int nr)
84 {
85         BUG_ON(nr < 0 || nr >= NR_CPUS);
86
87         /*
88          * The processor is currently spinning, waiting for the
89          * cpu_start field to become non-zero After we set cpu_start,
90          * the processor will continue on to secondary_start
91          */
92         paca[nr].cpu_start = 1;
93         smp_mb();
94 }
95 #endif
96
97 void smp_message_recv(int msg)
98 {
99         switch(msg) {
100         case PPC_MSG_CALL_FUNCTION:
101                 generic_smp_call_function_interrupt();
102                 break;
103         case PPC_MSG_RESCHEDULE:
104                 /* we notice need_resched on exit */
105                 break;
106         case PPC_MSG_CALL_FUNC_SINGLE:
107                 generic_smp_call_function_single_interrupt();
108                 break;
109         case PPC_MSG_DEBUGGER_BREAK:
110                 if (crash_ipi_function_ptr) {
111                         crash_ipi_function_ptr(get_irq_regs());
112                         break;
113                 }
114 #ifdef CONFIG_DEBUGGER
115                 debugger_ipi(get_irq_regs());
116                 break;
117 #endif /* CONFIG_DEBUGGER */
118                 /* FALLTHROUGH */
119         default:
120                 printk("SMP %d: smp_message_recv(): unknown msg %d\n",
121                        smp_processor_id(), msg);
122                 break;
123         }
124 }
125
126 void smp_send_reschedule(int cpu)
127 {
128         if (likely(smp_ops))
129                 smp_ops->message_pass(cpu, PPC_MSG_RESCHEDULE);
130 }
131
132 void arch_send_call_function_single_ipi(int cpu)
133 {
134         smp_ops->message_pass(cpu, PPC_MSG_CALL_FUNC_SINGLE);
135 }
136
137 void arch_send_call_function_ipi(cpumask_t mask)
138 {
139         unsigned int cpu;
140
141         for_each_cpu_mask(cpu, mask)
142                 smp_ops->message_pass(cpu, PPC_MSG_CALL_FUNCTION);
143 }
144
145 #ifdef CONFIG_DEBUGGER
146 void smp_send_debugger_break(int cpu)
147 {
148         if (likely(smp_ops))
149                 smp_ops->message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
150 }
151 #endif
152
153 #ifdef CONFIG_KEXEC
154 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
155 {
156         crash_ipi_function_ptr = crash_ipi_callback;
157         if (crash_ipi_callback && smp_ops) {
158                 mb();
159                 smp_ops->message_pass(MSG_ALL_BUT_SELF, PPC_MSG_DEBUGGER_BREAK);
160         }
161 }
162 #endif
163
164 static void stop_this_cpu(void *dummy)
165 {
166         local_irq_disable();
167         while (1)
168                 ;
169 }
170
171 void smp_send_stop(void)
172 {
173         smp_call_function(stop_this_cpu, NULL, 0);
174 }
175
176 struct thread_info *current_set[NR_CPUS];
177
178 static void __devinit smp_store_cpu_info(int id)
179 {
180         per_cpu(pvr, id) = mfspr(SPRN_PVR);
181 }
182
183 static void __init smp_create_idle(unsigned int cpu)
184 {
185         struct task_struct *p;
186
187         /* create a process for the processor */
188         p = fork_idle(cpu);
189         if (IS_ERR(p))
190                 panic("failed fork for CPU %u: %li", cpu, PTR_ERR(p));
191 #ifdef CONFIG_PPC64
192         paca[cpu].__current = p;
193         paca[cpu].kstack = (unsigned long) task_thread_info(p)
194                 + THREAD_SIZE - STACK_FRAME_OVERHEAD;
195 #endif
196         current_set[cpu] = task_thread_info(p);
197         task_thread_info(p)->cpu = cpu;
198 }
199
200 void __init smp_prepare_cpus(unsigned int max_cpus)
201 {
202         unsigned int cpu;
203
204         DBG("smp_prepare_cpus\n");
205
206         /* 
207          * setup_cpu may need to be called on the boot cpu. We havent
208          * spun any cpus up but lets be paranoid.
209          */
210         BUG_ON(boot_cpuid != smp_processor_id());
211
212         /* Fixup boot cpu */
213         smp_store_cpu_info(boot_cpuid);
214         cpu_callin_map[boot_cpuid] = 1;
215
216         if (smp_ops)
217                 max_cpus = smp_ops->probe();
218         else
219                 max_cpus = 1;
220  
221         smp_space_timers(max_cpus);
222
223         for_each_possible_cpu(cpu)
224                 if (cpu != boot_cpuid)
225                         smp_create_idle(cpu);
226 }
227
228 void __devinit smp_prepare_boot_cpu(void)
229 {
230         BUG_ON(smp_processor_id() != boot_cpuid);
231
232         cpu_set(boot_cpuid, cpu_online_map);
233         cpu_set(boot_cpuid, per_cpu(cpu_sibling_map, boot_cpuid));
234         cpu_set(boot_cpuid, per_cpu(cpu_core_map, boot_cpuid));
235 #ifdef CONFIG_PPC64
236         paca[boot_cpuid].__current = current;
237 #endif
238         current_set[boot_cpuid] = task_thread_info(current);
239 }
240
241 #ifdef CONFIG_HOTPLUG_CPU
242 /* State of each CPU during hotplug phases */
243 DEFINE_PER_CPU(int, cpu_state) = { 0 };
244
245 int generic_cpu_disable(void)
246 {
247         unsigned int cpu = smp_processor_id();
248
249         if (cpu == boot_cpuid)
250                 return -EBUSY;
251
252         cpu_clear(cpu, cpu_online_map);
253 #ifdef CONFIG_PPC64
254         vdso_data->processorCount--;
255         fixup_irqs(cpu_online_map);
256 #endif
257         return 0;
258 }
259
260 int generic_cpu_enable(unsigned int cpu)
261 {
262         /* Do the normal bootup if we haven't
263          * already bootstrapped. */
264         if (system_state != SYSTEM_RUNNING)
265                 return -ENOSYS;
266
267         /* get the target out of it's holding state */
268         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
269         smp_wmb();
270
271         while (!cpu_online(cpu))
272                 cpu_relax();
273
274 #ifdef CONFIG_PPC64
275         fixup_irqs(cpu_online_map);
276         /* counter the irq disable in fixup_irqs */
277         local_irq_enable();
278 #endif
279         return 0;
280 }
281
282 void generic_cpu_die(unsigned int cpu)
283 {
284         int i;
285
286         for (i = 0; i < 100; i++) {
287                 smp_rmb();
288                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
289                         return;
290                 msleep(100);
291         }
292         printk(KERN_ERR "CPU%d didn't die...\n", cpu);
293 }
294
295 void generic_mach_cpu_die(void)
296 {
297         unsigned int cpu;
298
299         local_irq_disable();
300         cpu = smp_processor_id();
301         printk(KERN_DEBUG "CPU%d offline\n", cpu);
302         __get_cpu_var(cpu_state) = CPU_DEAD;
303         smp_wmb();
304         while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
305                 cpu_relax();
306         cpu_set(cpu, cpu_online_map);
307         local_irq_enable();
308 }
309 #endif
310
311 static int __devinit cpu_enable(unsigned int cpu)
312 {
313         if (smp_ops && smp_ops->cpu_enable)
314                 return smp_ops->cpu_enable(cpu);
315
316         return -ENOSYS;
317 }
318
319 int __cpuinit __cpu_up(unsigned int cpu)
320 {
321         int c;
322
323         secondary_ti = current_set[cpu];
324         if (!cpu_enable(cpu))
325                 return 0;
326
327         if (smp_ops == NULL ||
328             (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
329                 return -EINVAL;
330
331         /* Make sure callin-map entry is 0 (can be leftover a CPU
332          * hotplug
333          */
334         cpu_callin_map[cpu] = 0;
335
336         /* The information for processor bringup must
337          * be written out to main store before we release
338          * the processor.
339          */
340         smp_mb();
341
342         /* wake up cpus */
343         DBG("smp: kicking cpu %d\n", cpu);
344         smp_ops->kick_cpu(cpu);
345
346         /*
347          * wait to see if the cpu made a callin (is actually up).
348          * use this value that I found through experimentation.
349          * -- Cort
350          */
351         if (system_state < SYSTEM_RUNNING)
352                 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
353                         udelay(100);
354 #ifdef CONFIG_HOTPLUG_CPU
355         else
356                 /*
357                  * CPUs can take much longer to come up in the
358                  * hotplug case.  Wait five seconds.
359                  */
360                 for (c = 25; c && !cpu_callin_map[cpu]; c--) {
361                         msleep(200);
362                 }
363 #endif
364
365         if (!cpu_callin_map[cpu]) {
366                 printk("Processor %u is stuck.\n", cpu);
367                 return -ENOENT;
368         }
369
370         printk("Processor %u found.\n", cpu);
371
372         if (smp_ops->give_timebase)
373                 smp_ops->give_timebase();
374
375         /* Wait until cpu puts itself in the online map */
376         while (!cpu_online(cpu))
377                 cpu_relax();
378
379         return 0;
380 }
381
382 /* Return the value of the reg property corresponding to the given
383  * logical cpu.
384  */
385 int cpu_to_core_id(int cpu)
386 {
387         struct device_node *np;
388         const int *reg;
389         int id = -1;
390
391         np = of_get_cpu_node(cpu, NULL);
392         if (!np)
393                 goto out;
394
395         reg = of_get_property(np, "reg", NULL);
396         if (!reg)
397                 goto out;
398
399         id = *reg;
400 out:
401         of_node_put(np);
402         return id;
403 }
404
405 /* Must be called when no change can occur to cpu_present_map,
406  * i.e. during cpu online or offline.
407  */
408 static struct device_node *cpu_to_l2cache(int cpu)
409 {
410         struct device_node *np;
411         const phandle *php;
412         phandle ph;
413
414         if (!cpu_present(cpu))
415                 return NULL;
416
417         np = of_get_cpu_node(cpu, NULL);
418         if (np == NULL)
419                 return NULL;
420
421         php = of_get_property(np, "l2-cache", NULL);
422         if (php == NULL)
423                 return NULL;
424         ph = *php;
425         of_node_put(np);
426
427         return of_find_node_by_phandle(ph);
428 }
429
430 /* Activate a secondary processor. */
431 int __devinit start_secondary(void *unused)
432 {
433         unsigned int cpu = smp_processor_id();
434         struct device_node *l2_cache;
435         int i, base;
436
437         atomic_inc(&init_mm.mm_count);
438         current->active_mm = &init_mm;
439
440         smp_store_cpu_info(cpu);
441         set_dec(tb_ticks_per_jiffy);
442         preempt_disable();
443         cpu_callin_map[cpu] = 1;
444
445         smp_ops->setup_cpu(cpu);
446         if (smp_ops->take_timebase)
447                 smp_ops->take_timebase();
448
449         if (system_state > SYSTEM_BOOTING)
450                 snapshot_timebase();
451
452         secondary_cpu_time_init();
453
454         ipi_call_lock();
455         notify_cpu_starting(cpu);
456         cpu_set(cpu, cpu_online_map);
457         /* Update sibling maps */
458         base = cpu_first_thread_in_core(cpu);
459         for (i = 0; i < threads_per_core; i++) {
460                 if (cpu_is_offline(base + i))
461                         continue;
462                 cpu_set(cpu, per_cpu(cpu_sibling_map, base + i));
463                 cpu_set(base + i, per_cpu(cpu_sibling_map, cpu));
464
465                 /* cpu_core_map should be a superset of
466                  * cpu_sibling_map even if we don't have cache
467                  * information, so update the former here, too.
468                  */
469                 cpu_set(cpu, per_cpu(cpu_core_map, base +i));
470                 cpu_set(base + i, per_cpu(cpu_core_map, cpu));
471         }
472         l2_cache = cpu_to_l2cache(cpu);
473         for_each_online_cpu(i) {
474                 struct device_node *np = cpu_to_l2cache(i);
475                 if (!np)
476                         continue;
477                 if (np == l2_cache) {
478                         cpu_set(cpu, per_cpu(cpu_core_map, i));
479                         cpu_set(i, per_cpu(cpu_core_map, cpu));
480                 }
481                 of_node_put(np);
482         }
483         of_node_put(l2_cache);
484         ipi_call_unlock();
485
486         local_irq_enable();
487
488         cpu_idle();
489         return 0;
490 }
491
492 int setup_profiling_timer(unsigned int multiplier)
493 {
494         return 0;
495 }
496
497 void __init smp_cpus_done(unsigned int max_cpus)
498 {
499         cpumask_t old_mask;
500
501         /* We want the setup_cpu() here to be called from CPU 0, but our
502          * init thread may have been "borrowed" by another CPU in the meantime
503          * se we pin us down to CPU 0 for a short while
504          */
505         old_mask = current->cpus_allowed;
506         set_cpus_allowed(current, cpumask_of_cpu(boot_cpuid));
507         
508         if (smp_ops)
509                 smp_ops->setup_cpu(boot_cpuid);
510
511         set_cpus_allowed(current, old_mask);
512
513         snapshot_timebases();
514
515         dump_numa_cpu_topology();
516 }
517
518 #ifdef CONFIG_HOTPLUG_CPU
519 int __cpu_disable(void)
520 {
521         struct device_node *l2_cache;
522         int cpu = smp_processor_id();
523         int base, i;
524         int err;
525
526         if (!smp_ops->cpu_disable)
527                 return -ENOSYS;
528
529         err = smp_ops->cpu_disable();
530         if (err)
531                 return err;
532
533         /* Update sibling maps */
534         base = cpu_first_thread_in_core(cpu);
535         for (i = 0; i < threads_per_core; i++) {
536                 cpu_clear(cpu, per_cpu(cpu_sibling_map, base + i));
537                 cpu_clear(base + i, per_cpu(cpu_sibling_map, cpu));
538                 cpu_clear(cpu, per_cpu(cpu_core_map, base +i));
539                 cpu_clear(base + i, per_cpu(cpu_core_map, cpu));
540         }
541
542         l2_cache = cpu_to_l2cache(cpu);
543         for_each_present_cpu(i) {
544                 struct device_node *np = cpu_to_l2cache(i);
545                 if (!np)
546                         continue;
547                 if (np == l2_cache) {
548                         cpu_clear(cpu, per_cpu(cpu_core_map, i));
549                         cpu_clear(i, per_cpu(cpu_core_map, cpu));
550                 }
551                 of_node_put(np);
552         }
553         of_node_put(l2_cache);
554
555
556         return 0;
557 }
558
559 void __cpu_die(unsigned int cpu)
560 {
561         if (smp_ops->cpu_die)
562                 smp_ops->cpu_die(cpu);
563 }
564 #endif