Merge branches 'sched/devel', 'sched/cpu-hotplug', 'sched/cpusets' and 'sched/urgent...
[linux-2.6] / 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                 /* XXX Do we have to do this? */
105                 set_need_resched();
106                 break;
107         case PPC_MSG_CALL_FUNC_SINGLE:
108                 generic_smp_call_function_single_interrupt();
109                 break;
110         case PPC_MSG_DEBUGGER_BREAK:
111                 if (crash_ipi_function_ptr) {
112                         crash_ipi_function_ptr(get_irq_regs());
113                         break;
114                 }
115 #ifdef CONFIG_DEBUGGER
116                 debugger_ipi(get_irq_regs());
117                 break;
118 #endif /* CONFIG_DEBUGGER */
119                 /* FALLTHROUGH */
120         default:
121                 printk("SMP %d: smp_message_recv(): unknown msg %d\n",
122                        smp_processor_id(), msg);
123                 break;
124         }
125 }
126
127 void smp_send_reschedule(int cpu)
128 {
129         if (likely(smp_ops))
130                 smp_ops->message_pass(cpu, PPC_MSG_RESCHEDULE);
131 }
132
133 void arch_send_call_function_single_ipi(int cpu)
134 {
135         smp_ops->message_pass(cpu, PPC_MSG_CALL_FUNC_SINGLE);
136 }
137
138 void arch_send_call_function_ipi(cpumask_t mask)
139 {
140         unsigned int cpu;
141
142         for_each_cpu_mask(cpu, mask)
143                 smp_ops->message_pass(cpu, PPC_MSG_CALL_FUNCTION);
144 }
145
146 #ifdef CONFIG_DEBUGGER
147 void smp_send_debugger_break(int cpu)
148 {
149         if (likely(smp_ops))
150                 smp_ops->message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
151 }
152 #endif
153
154 #ifdef CONFIG_KEXEC
155 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
156 {
157         crash_ipi_function_ptr = crash_ipi_callback;
158         if (crash_ipi_callback && smp_ops) {
159                 mb();
160                 smp_ops->message_pass(MSG_ALL_BUT_SELF, PPC_MSG_DEBUGGER_BREAK);
161         }
162 }
163 #endif
164
165 static void stop_this_cpu(void *dummy)
166 {
167         local_irq_disable();
168         while (1)
169                 ;
170 }
171
172 void smp_send_stop(void)
173 {
174         smp_call_function(stop_this_cpu, NULL, 0);
175 }
176
177 struct thread_info *current_set[NR_CPUS];
178
179 static void __devinit smp_store_cpu_info(int id)
180 {
181         per_cpu(pvr, id) = mfspr(SPRN_PVR);
182 }
183
184 static void __init smp_create_idle(unsigned int cpu)
185 {
186         struct task_struct *p;
187
188         /* create a process for the processor */
189         p = fork_idle(cpu);
190         if (IS_ERR(p))
191                 panic("failed fork for CPU %u: %li", cpu, PTR_ERR(p));
192 #ifdef CONFIG_PPC64
193         paca[cpu].__current = p;
194         paca[cpu].kstack = (unsigned long) task_thread_info(p)
195                 + THREAD_SIZE - STACK_FRAME_OVERHEAD;
196 #endif
197         current_set[cpu] = task_thread_info(p);
198         task_thread_info(p)->cpu = cpu;
199 }
200
201 void __init smp_prepare_cpus(unsigned int max_cpus)
202 {
203         unsigned int cpu;
204
205         DBG("smp_prepare_cpus\n");
206
207         /* 
208          * setup_cpu may need to be called on the boot cpu. We havent
209          * spun any cpus up but lets be paranoid.
210          */
211         BUG_ON(boot_cpuid != smp_processor_id());
212
213         /* Fixup boot cpu */
214         smp_store_cpu_info(boot_cpuid);
215         cpu_callin_map[boot_cpuid] = 1;
216
217         if (smp_ops)
218                 max_cpus = smp_ops->probe();
219         else
220                 max_cpus = 1;
221  
222         smp_space_timers(max_cpus);
223
224         for_each_possible_cpu(cpu)
225                 if (cpu != boot_cpuid)
226                         smp_create_idle(cpu);
227 }
228
229 void __devinit smp_prepare_boot_cpu(void)
230 {
231         BUG_ON(smp_processor_id() != boot_cpuid);
232
233         cpu_set(boot_cpuid, cpu_online_map);
234         cpu_set(boot_cpuid, per_cpu(cpu_sibling_map, boot_cpuid));
235         cpu_set(boot_cpuid, per_cpu(cpu_core_map, boot_cpuid));
236 #ifdef CONFIG_PPC64
237         paca[boot_cpuid].__current = current;
238 #endif
239         current_set[boot_cpuid] = task_thread_info(current);
240 }
241
242 #ifdef CONFIG_HOTPLUG_CPU
243 /* State of each CPU during hotplug phases */
244 DEFINE_PER_CPU(int, cpu_state) = { 0 };
245
246 int generic_cpu_disable(void)
247 {
248         unsigned int cpu = smp_processor_id();
249
250         if (cpu == boot_cpuid)
251                 return -EBUSY;
252
253         cpu_clear(cpu, cpu_online_map);
254 #ifdef CONFIG_PPC64
255         vdso_data->processorCount--;
256         fixup_irqs(cpu_online_map);
257 #endif
258         return 0;
259 }
260
261 int generic_cpu_enable(unsigned int cpu)
262 {
263         /* Do the normal bootup if we haven't
264          * already bootstrapped. */
265         if (system_state != SYSTEM_RUNNING)
266                 return -ENOSYS;
267
268         /* get the target out of it's holding state */
269         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
270         smp_wmb();
271
272         while (!cpu_online(cpu))
273                 cpu_relax();
274
275 #ifdef CONFIG_PPC64
276         fixup_irqs(cpu_online_map);
277         /* counter the irq disable in fixup_irqs */
278         local_irq_enable();
279 #endif
280         return 0;
281 }
282
283 void generic_cpu_die(unsigned int cpu)
284 {
285         int i;
286
287         for (i = 0; i < 100; i++) {
288                 smp_rmb();
289                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
290                         return;
291                 msleep(100);
292         }
293         printk(KERN_ERR "CPU%d didn't die...\n", cpu);
294 }
295
296 void generic_mach_cpu_die(void)
297 {
298         unsigned int cpu;
299
300         local_irq_disable();
301         cpu = smp_processor_id();
302         printk(KERN_DEBUG "CPU%d offline\n", cpu);
303         __get_cpu_var(cpu_state) = CPU_DEAD;
304         smp_wmb();
305         while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
306                 cpu_relax();
307         cpu_set(cpu, cpu_online_map);
308         local_irq_enable();
309 }
310 #endif
311
312 static int __devinit cpu_enable(unsigned int cpu)
313 {
314         if (smp_ops && smp_ops->cpu_enable)
315                 return smp_ops->cpu_enable(cpu);
316
317         return -ENOSYS;
318 }
319
320 int __cpuinit __cpu_up(unsigned int cpu)
321 {
322         int c;
323
324         secondary_ti = current_set[cpu];
325         if (!cpu_enable(cpu))
326                 return 0;
327
328         if (smp_ops == NULL ||
329             (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
330                 return -EINVAL;
331
332         /* Make sure callin-map entry is 0 (can be leftover a CPU
333          * hotplug
334          */
335         cpu_callin_map[cpu] = 0;
336
337         /* The information for processor bringup must
338          * be written out to main store before we release
339          * the processor.
340          */
341         smp_mb();
342
343         /* wake up cpus */
344         DBG("smp: kicking cpu %d\n", cpu);
345         smp_ops->kick_cpu(cpu);
346
347         /*
348          * wait to see if the cpu made a callin (is actually up).
349          * use this value that I found through experimentation.
350          * -- Cort
351          */
352         if (system_state < SYSTEM_RUNNING)
353                 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
354                         udelay(100);
355 #ifdef CONFIG_HOTPLUG_CPU
356         else
357                 /*
358                  * CPUs can take much longer to come up in the
359                  * hotplug case.  Wait five seconds.
360                  */
361                 for (c = 25; c && !cpu_callin_map[cpu]; c--) {
362                         msleep(200);
363                 }
364 #endif
365
366         if (!cpu_callin_map[cpu]) {
367                 printk("Processor %u is stuck.\n", cpu);
368                 return -ENOENT;
369         }
370
371         printk("Processor %u found.\n", cpu);
372
373         if (smp_ops->give_timebase)
374                 smp_ops->give_timebase();
375
376         /* Wait until cpu puts itself in the online map */
377         while (!cpu_online(cpu))
378                 cpu_relax();
379
380         return 0;
381 }
382
383 /* Return the value of the reg property corresponding to the given
384  * logical cpu.
385  */
386 int cpu_to_core_id(int cpu)
387 {
388         struct device_node *np;
389         const int *reg;
390         int id = -1;
391
392         np = of_get_cpu_node(cpu, NULL);
393         if (!np)
394                 goto out;
395
396         reg = of_get_property(np, "reg", NULL);
397         if (!reg)
398                 goto out;
399
400         id = *reg;
401 out:
402         of_node_put(np);
403         return id;
404 }
405
406 /* Must be called when no change can occur to cpu_present_map,
407  * i.e. during cpu online or offline.
408  */
409 static struct device_node *cpu_to_l2cache(int cpu)
410 {
411         struct device_node *np;
412         const phandle *php;
413         phandle ph;
414
415         if (!cpu_present(cpu))
416                 return NULL;
417
418         np = of_get_cpu_node(cpu, NULL);
419         if (np == NULL)
420                 return NULL;
421
422         php = of_get_property(np, "l2-cache", NULL);
423         if (php == NULL)
424                 return NULL;
425         ph = *php;
426         of_node_put(np);
427
428         return of_find_node_by_phandle(ph);
429 }
430
431 /* Activate a secondary processor. */
432 int __devinit start_secondary(void *unused)
433 {
434         unsigned int cpu = smp_processor_id();
435         struct device_node *l2_cache;
436         int i, base;
437
438         atomic_inc(&init_mm.mm_count);
439         current->active_mm = &init_mm;
440
441         smp_store_cpu_info(cpu);
442         set_dec(tb_ticks_per_jiffy);
443         preempt_disable();
444         cpu_callin_map[cpu] = 1;
445
446         smp_ops->setup_cpu(cpu);
447         if (smp_ops->take_timebase)
448                 smp_ops->take_timebase();
449
450         if (system_state > SYSTEM_BOOTING)
451                 snapshot_timebase();
452
453         secondary_cpu_time_init();
454
455         ipi_call_lock();
456         notify_cpu_starting(cpu);
457         cpu_set(cpu, cpu_online_map);
458         /* Update sibling maps */
459         base = cpu_first_thread_in_core(cpu);
460         for (i = 0; i < threads_per_core; i++) {
461                 if (cpu_is_offline(base + i))
462                         continue;
463                 cpu_set(cpu, per_cpu(cpu_sibling_map, base + i));
464                 cpu_set(base + i, per_cpu(cpu_sibling_map, cpu));
465
466                 /* cpu_core_map should be a superset of
467                  * cpu_sibling_map even if we don't have cache
468                  * information, so update the former here, too.
469                  */
470                 cpu_set(cpu, per_cpu(cpu_core_map, base +i));
471                 cpu_set(base + i, per_cpu(cpu_core_map, cpu));
472         }
473         l2_cache = cpu_to_l2cache(cpu);
474         for_each_online_cpu(i) {
475                 struct device_node *np = cpu_to_l2cache(i);
476                 if (!np)
477                         continue;
478                 if (np == l2_cache) {
479                         cpu_set(cpu, per_cpu(cpu_core_map, i));
480                         cpu_set(i, per_cpu(cpu_core_map, cpu));
481                 }
482                 of_node_put(np);
483         }
484         of_node_put(l2_cache);
485         ipi_call_unlock();
486
487         local_irq_enable();
488
489         cpu_idle();
490         return 0;
491 }
492
493 int setup_profiling_timer(unsigned int multiplier)
494 {
495         return 0;
496 }
497
498 void __init smp_cpus_done(unsigned int max_cpus)
499 {
500         cpumask_t old_mask;
501
502         /* We want the setup_cpu() here to be called from CPU 0, but our
503          * init thread may have been "borrowed" by another CPU in the meantime
504          * se we pin us down to CPU 0 for a short while
505          */
506         old_mask = current->cpus_allowed;
507         set_cpus_allowed(current, cpumask_of_cpu(boot_cpuid));
508         
509         if (smp_ops)
510                 smp_ops->setup_cpu(boot_cpuid);
511
512         set_cpus_allowed(current, old_mask);
513
514         snapshot_timebases();
515
516         dump_numa_cpu_topology();
517 }
518
519 #ifdef CONFIG_HOTPLUG_CPU
520 int __cpu_disable(void)
521 {
522         struct device_node *l2_cache;
523         int cpu = smp_processor_id();
524         int base, i;
525         int err;
526
527         if (!smp_ops->cpu_disable)
528                 return -ENOSYS;
529
530         err = smp_ops->cpu_disable();
531         if (err)
532                 return err;
533
534         /* Update sibling maps */
535         base = cpu_first_thread_in_core(cpu);
536         for (i = 0; i < threads_per_core; i++) {
537                 cpu_clear(cpu, per_cpu(cpu_sibling_map, base + i));
538                 cpu_clear(base + i, per_cpu(cpu_sibling_map, cpu));
539                 cpu_clear(cpu, per_cpu(cpu_core_map, base +i));
540                 cpu_clear(base + i, per_cpu(cpu_core_map, cpu));
541         }
542
543         l2_cache = cpu_to_l2cache(cpu);
544         for_each_present_cpu(i) {
545                 struct device_node *np = cpu_to_l2cache(i);
546                 if (!np)
547                         continue;
548                 if (np == l2_cache) {
549                         cpu_clear(cpu, per_cpu(cpu_core_map, i));
550                         cpu_clear(i, per_cpu(cpu_core_map, cpu));
551                 }
552                 of_node_put(np);
553         }
554         of_node_put(l2_cache);
555
556
557         return 0;
558 }
559
560 void __cpu_die(unsigned int cpu)
561 {
562         if (smp_ops->cpu_die)
563                 smp_ops->cpu_die(cpu);
564 }
565 #endif