3c0fab5e1e16c033f15928eae5217da0e30e6931
[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_var_t, cpu_sibling_map);
63 DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
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         for_each_possible_cpu(cpu) {
275                 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
276                                         GFP_KERNEL, cpu_to_node(cpu));
277                 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
278                                         GFP_KERNEL, cpu_to_node(cpu));
279         }
280
281         cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
282         cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
283
284         if (smp_ops)
285                 if (smp_ops->probe)
286                         max_cpus = smp_ops->probe();
287                 else
288                         max_cpus = NR_CPUS;
289         else
290                 max_cpus = 1;
291
292         for_each_possible_cpu(cpu)
293                 if (cpu != boot_cpuid)
294                         smp_create_idle(cpu);
295 }
296
297 void __devinit smp_prepare_boot_cpu(void)
298 {
299         BUG_ON(smp_processor_id() != boot_cpuid);
300 #ifdef CONFIG_PPC64
301         paca[boot_cpuid].__current = current;
302 #endif
303         current_set[boot_cpuid] = task_thread_info(current);
304 }
305
306 #ifdef CONFIG_HOTPLUG_CPU
307 /* State of each CPU during hotplug phases */
308 DEFINE_PER_CPU(int, cpu_state) = { 0 };
309
310 int generic_cpu_disable(void)
311 {
312         unsigned int cpu = smp_processor_id();
313
314         if (cpu == boot_cpuid)
315                 return -EBUSY;
316
317         set_cpu_online(cpu, false);
318 #ifdef CONFIG_PPC64
319         vdso_data->processorCount--;
320         fixup_irqs(cpu_online_mask);
321 #endif
322         return 0;
323 }
324
325 int generic_cpu_enable(unsigned int cpu)
326 {
327         /* Do the normal bootup if we haven't
328          * already bootstrapped. */
329         if (system_state != SYSTEM_RUNNING)
330                 return -ENOSYS;
331
332         /* get the target out of it's holding state */
333         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
334         smp_wmb();
335
336         while (!cpu_online(cpu))
337                 cpu_relax();
338
339 #ifdef CONFIG_PPC64
340         fixup_irqs(cpu_online_mask);
341         /* counter the irq disable in fixup_irqs */
342         local_irq_enable();
343 #endif
344         return 0;
345 }
346
347 void generic_cpu_die(unsigned int cpu)
348 {
349         int i;
350
351         for (i = 0; i < 100; i++) {
352                 smp_rmb();
353                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
354                         return;
355                 msleep(100);
356         }
357         printk(KERN_ERR "CPU%d didn't die...\n", cpu);
358 }
359
360 void generic_mach_cpu_die(void)
361 {
362         unsigned int cpu;
363
364         local_irq_disable();
365         idle_task_exit();
366         cpu = smp_processor_id();
367         printk(KERN_DEBUG "CPU%d offline\n", cpu);
368         __get_cpu_var(cpu_state) = CPU_DEAD;
369         smp_wmb();
370         while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
371                 cpu_relax();
372 }
373 #endif
374
375 static int __devinit cpu_enable(unsigned int cpu)
376 {
377         if (smp_ops && smp_ops->cpu_enable)
378                 return smp_ops->cpu_enable(cpu);
379
380         return -ENOSYS;
381 }
382
383 int __cpuinit __cpu_up(unsigned int cpu)
384 {
385         int c;
386
387         secondary_ti = current_set[cpu];
388         if (!cpu_enable(cpu))
389                 return 0;
390
391         if (smp_ops == NULL ||
392             (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
393                 return -EINVAL;
394
395         /* Make sure callin-map entry is 0 (can be leftover a CPU
396          * hotplug
397          */
398         cpu_callin_map[cpu] = 0;
399
400         /* The information for processor bringup must
401          * be written out to main store before we release
402          * the processor.
403          */
404         smp_mb();
405
406         /* wake up cpus */
407         DBG("smp: kicking cpu %d\n", cpu);
408         smp_ops->kick_cpu(cpu);
409
410         /*
411          * wait to see if the cpu made a callin (is actually up).
412          * use this value that I found through experimentation.
413          * -- Cort
414          */
415         if (system_state < SYSTEM_RUNNING)
416                 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
417                         udelay(100);
418 #ifdef CONFIG_HOTPLUG_CPU
419         else
420                 /*
421                  * CPUs can take much longer to come up in the
422                  * hotplug case.  Wait five seconds.
423                  */
424                 for (c = 5000; c && !cpu_callin_map[cpu]; c--)
425                         msleep(1);
426 #endif
427
428         if (!cpu_callin_map[cpu]) {
429                 printk(KERN_ERR "Processor %u is stuck.\n", cpu);
430                 return -ENOENT;
431         }
432
433         DBG("Processor %u found.\n", cpu);
434
435         if (smp_ops->give_timebase)
436                 smp_ops->give_timebase();
437
438         /* Wait until cpu puts itself in the online map */
439         while (!cpu_online(cpu))
440                 cpu_relax();
441
442         return 0;
443 }
444
445 /* Return the value of the reg property corresponding to the given
446  * logical cpu.
447  */
448 int cpu_to_core_id(int cpu)
449 {
450         struct device_node *np;
451         const int *reg;
452         int id = -1;
453
454         np = of_get_cpu_node(cpu, NULL);
455         if (!np)
456                 goto out;
457
458         reg = of_get_property(np, "reg", NULL);
459         if (!reg)
460                 goto out;
461
462         id = *reg;
463 out:
464         of_node_put(np);
465         return id;
466 }
467
468 /* Helper routines for cpu to core mapping */
469 int cpu_core_index_of_thread(int cpu)
470 {
471         return cpu >> threads_shift;
472 }
473 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
474
475 int cpu_first_thread_of_core(int core)
476 {
477         return core << threads_shift;
478 }
479 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
480
481 /* Must be called when no change can occur to cpu_present_map,
482  * i.e. during cpu online or offline.
483  */
484 static struct device_node *cpu_to_l2cache(int cpu)
485 {
486         struct device_node *np;
487         struct device_node *cache;
488
489         if (!cpu_present(cpu))
490                 return NULL;
491
492         np = of_get_cpu_node(cpu, NULL);
493         if (np == NULL)
494                 return NULL;
495
496         cache = of_find_next_cache_node(np);
497
498         of_node_put(np);
499
500         return cache;
501 }
502
503 /* Activate a secondary processor. */
504 void __devinit start_secondary(void *unused)
505 {
506         unsigned int cpu = smp_processor_id();
507         struct device_node *l2_cache;
508         int i, base;
509
510         atomic_inc(&init_mm.mm_count);
511         current->active_mm = &init_mm;
512
513         smp_store_cpu_info(cpu);
514         set_dec(tb_ticks_per_jiffy);
515         preempt_disable();
516         cpu_callin_map[cpu] = 1;
517
518         if (smp_ops->setup_cpu)
519                 smp_ops->setup_cpu(cpu);
520         if (smp_ops->take_timebase)
521                 smp_ops->take_timebase();
522
523         secondary_cpu_time_init();
524
525         ipi_call_lock();
526         notify_cpu_starting(cpu);
527         set_cpu_online(cpu, true);
528         /* Update sibling maps */
529         base = cpu_first_thread_sibling(cpu);
530         for (i = 0; i < threads_per_core; i++) {
531                 if (cpu_is_offline(base + i))
532                         continue;
533                 cpumask_set_cpu(cpu, cpu_sibling_mask(base + i));
534                 cpumask_set_cpu(base + i, cpu_sibling_mask(cpu));
535
536                 /* cpu_core_map should be a superset of
537                  * cpu_sibling_map even if we don't have cache
538                  * information, so update the former here, too.
539                  */
540                 cpumask_set_cpu(cpu, cpu_core_mask(base + i));
541                 cpumask_set_cpu(base + i, cpu_core_mask(cpu));
542         }
543         l2_cache = cpu_to_l2cache(cpu);
544         for_each_online_cpu(i) {
545                 struct device_node *np = cpu_to_l2cache(i);
546                 if (!np)
547                         continue;
548                 if (np == l2_cache) {
549                         cpumask_set_cpu(cpu, cpu_core_mask(i));
550                         cpumask_set_cpu(i, cpu_core_mask(cpu));
551                 }
552                 of_node_put(np);
553         }
554         of_node_put(l2_cache);
555         ipi_call_unlock();
556
557         local_irq_enable();
558
559         cpu_idle();
560
561         BUG();
562 }
563
564 int setup_profiling_timer(unsigned int multiplier)
565 {
566         return 0;
567 }
568
569 void __init smp_cpus_done(unsigned int max_cpus)
570 {
571         cpumask_var_t old_mask;
572
573         /* We want the setup_cpu() here to be called from CPU 0, but our
574          * init thread may have been "borrowed" by another CPU in the meantime
575          * se we pin us down to CPU 0 for a short while
576          */
577         alloc_cpumask_var(&old_mask, GFP_NOWAIT);
578         cpumask_copy(old_mask, &current->cpus_allowed);
579         set_cpus_allowed_ptr(current, cpumask_of(boot_cpuid));
580         
581         if (smp_ops && smp_ops->setup_cpu)
582                 smp_ops->setup_cpu(boot_cpuid);
583
584         set_cpus_allowed_ptr(current, old_mask);
585
586         free_cpumask_var(old_mask);
587
588         dump_numa_cpu_topology();
589 }
590
591 int arch_sd_sibling_asym_packing(void)
592 {
593         if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
594                 printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
595                 return SD_ASYM_PACKING;
596         }
597         return 0;
598 }
599
600 #ifdef CONFIG_HOTPLUG_CPU
601 int __cpu_disable(void)
602 {
603         struct device_node *l2_cache;
604         int cpu = smp_processor_id();
605         int base, i;
606         int err;
607
608         if (!smp_ops->cpu_disable)
609                 return -ENOSYS;
610
611         err = smp_ops->cpu_disable();
612         if (err)
613                 return err;
614
615         /* Update sibling maps */
616         base = cpu_first_thread_sibling(cpu);
617         for (i = 0; i < threads_per_core; i++) {
618                 cpumask_clear_cpu(cpu, cpu_sibling_mask(base + i));
619                 cpumask_clear_cpu(base + i, cpu_sibling_mask(cpu));
620                 cpumask_clear_cpu(cpu, cpu_core_mask(base + i));
621                 cpumask_clear_cpu(base + i, cpu_core_mask(cpu));
622         }
623
624         l2_cache = cpu_to_l2cache(cpu);
625         for_each_present_cpu(i) {
626                 struct device_node *np = cpu_to_l2cache(i);
627                 if (!np)
628                         continue;
629                 if (np == l2_cache) {
630                         cpumask_clear_cpu(cpu, cpu_core_mask(i));
631                         cpumask_clear_cpu(i, cpu_core_mask(cpu));
632                 }
633                 of_node_put(np);
634         }
635         of_node_put(l2_cache);
636
637
638         return 0;
639 }
640
641 void __cpu_die(unsigned int cpu)
642 {
643         if (smp_ops->cpu_die)
644                 smp_ops->cpu_die(cpu);
645 }
646
647 static DEFINE_MUTEX(powerpc_cpu_hotplug_driver_mutex);
648
649 void cpu_hotplug_driver_lock()
650 {
651         mutex_lock(&powerpc_cpu_hotplug_driver_mutex);
652 }
653
654 void cpu_hotplug_driver_unlock()
655 {
656         mutex_unlock(&powerpc_cpu_hotplug_driver_mutex);
657 }
658
659 void cpu_die(void)
660 {
661         if (ppc_md.cpu_die)
662                 ppc_md.cpu_die();
663
664         /* If we return, we re-enter start_secondary */
665         start_secondary_resume();
666 }
667
668 #endif