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