Merge branch 'x86-debug-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[cascardo/linux.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/export.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/device.h>
31 #include <linux/cpu.h>
32 #include <linux/notifier.h>
33 #include <linux/topology.h>
34
35 #include <asm/ptrace.h>
36 #include <linux/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/mpic.h>
47 #include <asm/vdso_datapage.h>
48 #ifdef CONFIG_PPC64
49 #include <asm/paca.h>
50 #endif
51 #include <asm/debug.h>
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 #ifdef CONFIG_HOTPLUG_CPU
61 /* State of each CPU during hotplug phases */
62 static DEFINE_PER_CPU(int, cpu_state) = { 0 };
63 #endif
64
65 struct thread_info *secondary_ti;
66
67 DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
68 DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
69
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 /* Can't be static due to PowerMac hackery */
77 volatile unsigned int cpu_callin_map[NR_CPUS];
78
79 int smt_enabled_at_boot = 1;
80
81 static void (*crash_ipi_function_ptr)(struct pt_regs *) = NULL;
82
83 #ifdef CONFIG_PPC64
84 int __devinit smp_generic_kick_cpu(int nr)
85 {
86         BUG_ON(nr < 0 || nr >= NR_CPUS);
87
88         /*
89          * The processor is currently spinning, waiting for the
90          * cpu_start field to become non-zero After we set cpu_start,
91          * the processor will continue on to secondary_start
92          */
93         if (!paca[nr].cpu_start) {
94                 paca[nr].cpu_start = 1;
95                 smp_mb();
96                 return 0;
97         }
98
99 #ifdef CONFIG_HOTPLUG_CPU
100         /*
101          * Ok it's not there, so it might be soft-unplugged, let's
102          * try to bring it back
103          */
104         per_cpu(cpu_state, nr) = CPU_UP_PREPARE;
105         smp_wmb();
106         smp_send_reschedule(nr);
107 #endif /* CONFIG_HOTPLUG_CPU */
108
109         return 0;
110 }
111 #endif /* CONFIG_PPC64 */
112
113 static irqreturn_t call_function_action(int irq, void *data)
114 {
115         generic_smp_call_function_interrupt();
116         return IRQ_HANDLED;
117 }
118
119 static irqreturn_t reschedule_action(int irq, void *data)
120 {
121         scheduler_ipi();
122         return IRQ_HANDLED;
123 }
124
125 static irqreturn_t call_function_single_action(int irq, void *data)
126 {
127         generic_smp_call_function_single_interrupt();
128         return IRQ_HANDLED;
129 }
130
131 static irqreturn_t debug_ipi_action(int irq, void *data)
132 {
133         if (crash_ipi_function_ptr) {
134                 crash_ipi_function_ptr(get_irq_regs());
135                 return IRQ_HANDLED;
136         }
137
138 #ifdef CONFIG_DEBUGGER
139         debugger_ipi(get_irq_regs());
140 #endif /* CONFIG_DEBUGGER */
141
142         return IRQ_HANDLED;
143 }
144
145 static irq_handler_t smp_ipi_action[] = {
146         [PPC_MSG_CALL_FUNCTION] =  call_function_action,
147         [PPC_MSG_RESCHEDULE] = reschedule_action,
148         [PPC_MSG_CALL_FUNC_SINGLE] = call_function_single_action,
149         [PPC_MSG_DEBUGGER_BREAK] = debug_ipi_action,
150 };
151
152 const char *smp_ipi_name[] = {
153         [PPC_MSG_CALL_FUNCTION] =  "ipi call function",
154         [PPC_MSG_RESCHEDULE] = "ipi reschedule",
155         [PPC_MSG_CALL_FUNC_SINGLE] = "ipi call function single",
156         [PPC_MSG_DEBUGGER_BREAK] = "ipi debugger",
157 };
158
159 /* optional function to request ipi, for controllers with >= 4 ipis */
160 int smp_request_message_ipi(int virq, int msg)
161 {
162         int err;
163
164         if (msg < 0 || msg > PPC_MSG_DEBUGGER_BREAK) {
165                 return -EINVAL;
166         }
167 #if !defined(CONFIG_DEBUGGER) && !defined(CONFIG_KEXEC)
168         if (msg == PPC_MSG_DEBUGGER_BREAK) {
169                 return 1;
170         }
171 #endif
172         err = request_irq(virq, smp_ipi_action[msg],
173                           IRQF_PERCPU | IRQF_NO_THREAD,
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 #ifdef CONFIG_PPC_SMP_MUXED_IPI
182 struct cpu_messages {
183         int messages;                   /* current messages */
184         unsigned long data;             /* data for cause ipi */
185 };
186 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages, ipi_message);
187
188 void smp_muxed_ipi_set_data(int cpu, unsigned long data)
189 {
190         struct cpu_messages *info = &per_cpu(ipi_message, cpu);
191
192         info->data = data;
193 }
194
195 void smp_muxed_ipi_message_pass(int cpu, int msg)
196 {
197         struct cpu_messages *info = &per_cpu(ipi_message, cpu);
198         char *message = (char *)&info->messages;
199
200         message[msg] = 1;
201         mb();
202         smp_ops->cause_ipi(cpu, info->data);
203 }
204
205 irqreturn_t smp_ipi_demux(void)
206 {
207         struct cpu_messages *info = &__get_cpu_var(ipi_message);
208         unsigned int all;
209
210         mb();   /* order any irq clear */
211
212         do {
213                 all = xchg_local(&info->messages, 0);
214
215 #ifdef __BIG_ENDIAN
216                 if (all & (1 << (24 - 8 * PPC_MSG_CALL_FUNCTION)))
217                         generic_smp_call_function_interrupt();
218                 if (all & (1 << (24 - 8 * PPC_MSG_RESCHEDULE)))
219                         scheduler_ipi();
220                 if (all & (1 << (24 - 8 * PPC_MSG_CALL_FUNC_SINGLE)))
221                         generic_smp_call_function_single_interrupt();
222                 if (all & (1 << (24 - 8 * PPC_MSG_DEBUGGER_BREAK)))
223                         debug_ipi_action(0, NULL);
224 #else
225 #error Unsupported ENDIAN
226 #endif
227         } while (info->messages);
228
229         return IRQ_HANDLED;
230 }
231 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
232
233 static inline void do_message_pass(int cpu, int msg)
234 {
235         if (smp_ops->message_pass)
236                 smp_ops->message_pass(cpu, msg);
237 #ifdef CONFIG_PPC_SMP_MUXED_IPI
238         else
239                 smp_muxed_ipi_message_pass(cpu, msg);
240 #endif
241 }
242
243 void smp_send_reschedule(int cpu)
244 {
245         if (likely(smp_ops))
246                 do_message_pass(cpu, PPC_MSG_RESCHEDULE);
247 }
248 EXPORT_SYMBOL_GPL(smp_send_reschedule);
249
250 void arch_send_call_function_single_ipi(int cpu)
251 {
252         do_message_pass(cpu, PPC_MSG_CALL_FUNC_SINGLE);
253 }
254
255 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
256 {
257         unsigned int cpu;
258
259         for_each_cpu(cpu, mask)
260                 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
261 }
262
263 #if defined(CONFIG_DEBUGGER) || defined(CONFIG_KEXEC)
264 void smp_send_debugger_break(void)
265 {
266         int cpu;
267         int me = raw_smp_processor_id();
268
269         if (unlikely(!smp_ops))
270                 return;
271
272         for_each_online_cpu(cpu)
273                 if (cpu != me)
274                         do_message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
275 }
276 #endif
277
278 #ifdef CONFIG_KEXEC
279 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
280 {
281         crash_ipi_function_ptr = crash_ipi_callback;
282         if (crash_ipi_callback) {
283                 mb();
284                 smp_send_debugger_break();
285         }
286 }
287 #endif
288
289 static void stop_this_cpu(void *dummy)
290 {
291         /* Remove this CPU */
292         set_cpu_online(smp_processor_id(), false);
293
294         local_irq_disable();
295         while (1)
296                 ;
297 }
298
299 void smp_send_stop(void)
300 {
301         smp_call_function(stop_this_cpu, NULL, 0);
302 }
303
304 struct thread_info *current_set[NR_CPUS];
305
306 static void __devinit smp_store_cpu_info(int id)
307 {
308         per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
309 #ifdef CONFIG_PPC_FSL_BOOK3E
310         per_cpu(next_tlbcam_idx, id)
311                 = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) - 1;
312 #endif
313 }
314
315 void __init smp_prepare_cpus(unsigned int max_cpus)
316 {
317         unsigned int cpu;
318
319         DBG("smp_prepare_cpus\n");
320
321         /* 
322          * setup_cpu may need to be called on the boot cpu. We havent
323          * spun any cpus up but lets be paranoid.
324          */
325         BUG_ON(boot_cpuid != smp_processor_id());
326
327         /* Fixup boot cpu */
328         smp_store_cpu_info(boot_cpuid);
329         cpu_callin_map[boot_cpuid] = 1;
330
331         for_each_possible_cpu(cpu) {
332                 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
333                                         GFP_KERNEL, cpu_to_node(cpu));
334                 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
335                                         GFP_KERNEL, cpu_to_node(cpu));
336         }
337
338         cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
339         cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
340
341         if (smp_ops)
342                 if (smp_ops->probe)
343                         max_cpus = smp_ops->probe();
344                 else
345                         max_cpus = NR_CPUS;
346         else
347                 max_cpus = 1;
348 }
349
350 void __devinit smp_prepare_boot_cpu(void)
351 {
352         BUG_ON(smp_processor_id() != boot_cpuid);
353 #ifdef CONFIG_PPC64
354         paca[boot_cpuid].__current = current;
355 #endif
356         current_set[boot_cpuid] = task_thread_info(current);
357 }
358
359 #ifdef CONFIG_HOTPLUG_CPU
360
361 int generic_cpu_disable(void)
362 {
363         unsigned int cpu = smp_processor_id();
364
365         if (cpu == boot_cpuid)
366                 return -EBUSY;
367
368         set_cpu_online(cpu, false);
369 #ifdef CONFIG_PPC64
370         vdso_data->processorCount--;
371 #endif
372         migrate_irqs();
373         return 0;
374 }
375
376 void generic_cpu_die(unsigned int cpu)
377 {
378         int i;
379
380         for (i = 0; i < 100; i++) {
381                 smp_rmb();
382                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
383                         return;
384                 msleep(100);
385         }
386         printk(KERN_ERR "CPU%d didn't die...\n", cpu);
387 }
388
389 void generic_mach_cpu_die(void)
390 {
391         unsigned int cpu;
392
393         local_irq_disable();
394         idle_task_exit();
395         cpu = smp_processor_id();
396         printk(KERN_DEBUG "CPU%d offline\n", cpu);
397         __get_cpu_var(cpu_state) = CPU_DEAD;
398         smp_wmb();
399         while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
400                 cpu_relax();
401 }
402
403 void generic_set_cpu_dead(unsigned int cpu)
404 {
405         per_cpu(cpu_state, cpu) = CPU_DEAD;
406 }
407
408 int generic_check_cpu_restart(unsigned int cpu)
409 {
410         return per_cpu(cpu_state, cpu) == CPU_UP_PREPARE;
411 }
412 #endif
413
414 static void cpu_idle_thread_init(unsigned int cpu, struct task_struct *idle)
415 {
416         struct thread_info *ti = task_thread_info(idle);
417
418 #ifdef CONFIG_PPC64
419         paca[cpu].__current = idle;
420         paca[cpu].kstack = (unsigned long)ti + THREAD_SIZE - STACK_FRAME_OVERHEAD;
421 #endif
422         ti->cpu = cpu;
423         secondary_ti = current_set[cpu] = ti;
424 }
425
426 int __cpuinit __cpu_up(unsigned int cpu, struct task_struct *tidle)
427 {
428         int rc, c;
429
430         if (smp_ops == NULL ||
431             (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
432                 return -EINVAL;
433
434         cpu_idle_thread_init(cpu, tidle);
435
436         /* Make sure callin-map entry is 0 (can be leftover a CPU
437          * hotplug
438          */
439         cpu_callin_map[cpu] = 0;
440
441         /* The information for processor bringup must
442          * be written out to main store before we release
443          * the processor.
444          */
445         smp_mb();
446
447         /* wake up cpus */
448         DBG("smp: kicking cpu %d\n", cpu);
449         rc = smp_ops->kick_cpu(cpu);
450         if (rc) {
451                 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
452                 return rc;
453         }
454
455         /*
456          * wait to see if the cpu made a callin (is actually up).
457          * use this value that I found through experimentation.
458          * -- Cort
459          */
460         if (system_state < SYSTEM_RUNNING)
461                 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
462                         udelay(100);
463 #ifdef CONFIG_HOTPLUG_CPU
464         else
465                 /*
466                  * CPUs can take much longer to come up in the
467                  * hotplug case.  Wait five seconds.
468                  */
469                 for (c = 5000; c && !cpu_callin_map[cpu]; c--)
470                         msleep(1);
471 #endif
472
473         if (!cpu_callin_map[cpu]) {
474                 printk(KERN_ERR "Processor %u is stuck.\n", cpu);
475                 return -ENOENT;
476         }
477
478         DBG("Processor %u found.\n", cpu);
479
480         if (smp_ops->give_timebase)
481                 smp_ops->give_timebase();
482
483         /* Wait until cpu puts itself in the online map */
484         while (!cpu_online(cpu))
485                 cpu_relax();
486
487         return 0;
488 }
489
490 /* Return the value of the reg property corresponding to the given
491  * logical cpu.
492  */
493 int cpu_to_core_id(int cpu)
494 {
495         struct device_node *np;
496         const int *reg;
497         int id = -1;
498
499         np = of_get_cpu_node(cpu, NULL);
500         if (!np)
501                 goto out;
502
503         reg = of_get_property(np, "reg", NULL);
504         if (!reg)
505                 goto out;
506
507         id = *reg;
508 out:
509         of_node_put(np);
510         return id;
511 }
512
513 /* Helper routines for cpu to core mapping */
514 int cpu_core_index_of_thread(int cpu)
515 {
516         return cpu >> threads_shift;
517 }
518 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
519
520 int cpu_first_thread_of_core(int core)
521 {
522         return core << threads_shift;
523 }
524 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
525
526 /* Must be called when no change can occur to cpu_present_mask,
527  * i.e. during cpu online or offline.
528  */
529 static struct device_node *cpu_to_l2cache(int cpu)
530 {
531         struct device_node *np;
532         struct device_node *cache;
533
534         if (!cpu_present(cpu))
535                 return NULL;
536
537         np = of_get_cpu_node(cpu, NULL);
538         if (np == NULL)
539                 return NULL;
540
541         cache = of_find_next_cache_node(np);
542
543         of_node_put(np);
544
545         return cache;
546 }
547
548 /* Activate a secondary processor. */
549 void __devinit start_secondary(void *unused)
550 {
551         unsigned int cpu = smp_processor_id();
552         struct device_node *l2_cache;
553         int i, base;
554
555         atomic_inc(&init_mm.mm_count);
556         current->active_mm = &init_mm;
557
558         smp_store_cpu_info(cpu);
559         set_dec(tb_ticks_per_jiffy);
560         preempt_disable();
561         cpu_callin_map[cpu] = 1;
562
563         if (smp_ops->setup_cpu)
564                 smp_ops->setup_cpu(cpu);
565         if (smp_ops->take_timebase)
566                 smp_ops->take_timebase();
567
568         secondary_cpu_time_init();
569
570 #ifdef CONFIG_PPC64
571         if (system_state == SYSTEM_RUNNING)
572                 vdso_data->processorCount++;
573 #endif
574         notify_cpu_starting(cpu);
575         set_cpu_online(cpu, true);
576         /* Update sibling maps */
577         base = cpu_first_thread_sibling(cpu);
578         for (i = 0; i < threads_per_core; i++) {
579                 if (cpu_is_offline(base + i))
580                         continue;
581                 cpumask_set_cpu(cpu, cpu_sibling_mask(base + i));
582                 cpumask_set_cpu(base + i, cpu_sibling_mask(cpu));
583
584                 /* cpu_core_map should be a superset of
585                  * cpu_sibling_map even if we don't have cache
586                  * information, so update the former here, too.
587                  */
588                 cpumask_set_cpu(cpu, cpu_core_mask(base + i));
589                 cpumask_set_cpu(base + i, cpu_core_mask(cpu));
590         }
591         l2_cache = cpu_to_l2cache(cpu);
592         for_each_online_cpu(i) {
593                 struct device_node *np = cpu_to_l2cache(i);
594                 if (!np)
595                         continue;
596                 if (np == l2_cache) {
597                         cpumask_set_cpu(cpu, cpu_core_mask(i));
598                         cpumask_set_cpu(i, cpu_core_mask(cpu));
599                 }
600                 of_node_put(np);
601         }
602         of_node_put(l2_cache);
603
604         local_irq_enable();
605
606         cpu_idle();
607
608         BUG();
609 }
610
611 int setup_profiling_timer(unsigned int multiplier)
612 {
613         return 0;
614 }
615
616 void __init smp_cpus_done(unsigned int max_cpus)
617 {
618         cpumask_var_t old_mask;
619
620         /* We want the setup_cpu() here to be called from CPU 0, but our
621          * init thread may have been "borrowed" by another CPU in the meantime
622          * se we pin us down to CPU 0 for a short while
623          */
624         alloc_cpumask_var(&old_mask, GFP_NOWAIT);
625         cpumask_copy(old_mask, tsk_cpus_allowed(current));
626         set_cpus_allowed_ptr(current, cpumask_of(boot_cpuid));
627         
628         if (smp_ops && smp_ops->setup_cpu)
629                 smp_ops->setup_cpu(boot_cpuid);
630
631         set_cpus_allowed_ptr(current, old_mask);
632
633         free_cpumask_var(old_mask);
634
635         if (smp_ops && smp_ops->bringup_done)
636                 smp_ops->bringup_done();
637
638         dump_numa_cpu_topology();
639
640 }
641
642 int arch_sd_sibling_asym_packing(void)
643 {
644         if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
645                 printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
646                 return SD_ASYM_PACKING;
647         }
648         return 0;
649 }
650
651 #ifdef CONFIG_HOTPLUG_CPU
652 int __cpu_disable(void)
653 {
654         struct device_node *l2_cache;
655         int cpu = smp_processor_id();
656         int base, i;
657         int err;
658
659         if (!smp_ops->cpu_disable)
660                 return -ENOSYS;
661
662         err = smp_ops->cpu_disable();
663         if (err)
664                 return err;
665
666         /* Update sibling maps */
667         base = cpu_first_thread_sibling(cpu);
668         for (i = 0; i < threads_per_core; i++) {
669                 cpumask_clear_cpu(cpu, cpu_sibling_mask(base + i));
670                 cpumask_clear_cpu(base + i, cpu_sibling_mask(cpu));
671                 cpumask_clear_cpu(cpu, cpu_core_mask(base + i));
672                 cpumask_clear_cpu(base + i, cpu_core_mask(cpu));
673         }
674
675         l2_cache = cpu_to_l2cache(cpu);
676         for_each_present_cpu(i) {
677                 struct device_node *np = cpu_to_l2cache(i);
678                 if (!np)
679                         continue;
680                 if (np == l2_cache) {
681                         cpumask_clear_cpu(cpu, cpu_core_mask(i));
682                         cpumask_clear_cpu(i, cpu_core_mask(cpu));
683                 }
684                 of_node_put(np);
685         }
686         of_node_put(l2_cache);
687
688
689         return 0;
690 }
691
692 void __cpu_die(unsigned int cpu)
693 {
694         if (smp_ops->cpu_die)
695                 smp_ops->cpu_die(cpu);
696 }
697
698 static DEFINE_MUTEX(powerpc_cpu_hotplug_driver_mutex);
699
700 void cpu_hotplug_driver_lock()
701 {
702         mutex_lock(&powerpc_cpu_hotplug_driver_mutex);
703 }
704
705 void cpu_hotplug_driver_unlock()
706 {
707         mutex_unlock(&powerpc_cpu_hotplug_driver_mutex);
708 }
709
710 void cpu_die(void)
711 {
712         if (ppc_md.cpu_die)
713                 ppc_md.cpu_die();
714
715         /* If we return, we re-enter start_secondary */
716         start_secondary_resume();
717 }
718
719 #endif