]> Pileus Git - ~andy/linux/blob - arch/powerpc/kernel/smp.c
8d4214afc21d6934fef8d12b582073c64a6ff7f3
[~andy/linux] / 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/vdso.h>
52 #include <asm/debug.h>
53
54 #ifdef DEBUG
55 #include <asm/udbg.h>
56 #define DBG(fmt...) udbg_printf(fmt)
57 #else
58 #define DBG(fmt...)
59 #endif
60
61 #ifdef CONFIG_HOTPLUG_CPU
62 /* State of each CPU during hotplug phases */
63 static DEFINE_PER_CPU(int, cpu_state) = { 0 };
64 #endif
65
66 struct thread_info *secondary_ti;
67
68 DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
69 DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
70
71 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
72 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
73
74 /* SMP operations for this machine */
75 struct smp_ops_t *smp_ops;
76
77 /* Can't be static due to PowerMac hackery */
78 volatile unsigned int cpu_callin_map[NR_CPUS];
79
80 int smt_enabled_at_boot = 1;
81
82 static void (*crash_ipi_function_ptr)(struct pt_regs *) = NULL;
83
84 #ifdef CONFIG_PPC64
85 int __devinit smp_generic_kick_cpu(int nr)
86 {
87         BUG_ON(nr < 0 || nr >= NR_CPUS);
88
89         /*
90          * The processor is currently spinning, waiting for the
91          * cpu_start field to become non-zero After we set cpu_start,
92          * the processor will continue on to secondary_start
93          */
94         if (!paca[nr].cpu_start) {
95                 paca[nr].cpu_start = 1;
96                 smp_mb();
97                 return 0;
98         }
99
100 #ifdef CONFIG_HOTPLUG_CPU
101         /*
102          * Ok it's not there, so it might be soft-unplugged, let's
103          * try to bring it back
104          */
105         per_cpu(cpu_state, nr) = CPU_UP_PREPARE;
106         smp_wmb();
107         smp_send_reschedule(nr);
108 #endif /* CONFIG_HOTPLUG_CPU */
109
110         return 0;
111 }
112 #endif /* CONFIG_PPC64 */
113
114 static irqreturn_t call_function_action(int irq, void *data)
115 {
116         generic_smp_call_function_interrupt();
117         return IRQ_HANDLED;
118 }
119
120 static irqreturn_t reschedule_action(int irq, void *data)
121 {
122         scheduler_ipi();
123         return IRQ_HANDLED;
124 }
125
126 static irqreturn_t call_function_single_action(int irq, void *data)
127 {
128         generic_smp_call_function_single_interrupt();
129         return IRQ_HANDLED;
130 }
131
132 static irqreturn_t debug_ipi_action(int irq, void *data)
133 {
134         if (crash_ipi_function_ptr) {
135                 crash_ipi_function_ptr(get_irq_regs());
136                 return IRQ_HANDLED;
137         }
138
139 #ifdef CONFIG_DEBUGGER
140         debugger_ipi(get_irq_regs());
141 #endif /* CONFIG_DEBUGGER */
142
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],
174                           IRQF_PERCPU | IRQF_NO_THREAD,
175                           smp_ipi_name[msg], 0);
176         WARN(err < 0, "unable to request_irq %d for %s (rc %d)\n",
177                 virq, smp_ipi_name[msg], err);
178
179         return err;
180 }
181
182 #ifdef CONFIG_PPC_SMP_MUXED_IPI
183 struct cpu_messages {
184         int messages;                   /* current messages */
185         unsigned long data;             /* data for cause ipi */
186 };
187 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cpu_messages, ipi_message);
188
189 void smp_muxed_ipi_set_data(int cpu, unsigned long data)
190 {
191         struct cpu_messages *info = &per_cpu(ipi_message, cpu);
192
193         info->data = data;
194 }
195
196 void smp_muxed_ipi_message_pass(int cpu, int msg)
197 {
198         struct cpu_messages *info = &per_cpu(ipi_message, cpu);
199         char *message = (char *)&info->messages;
200
201         /*
202          * Order previous accesses before accesses in the IPI handler.
203          */
204         smp_mb();
205         message[msg] = 1;
206         /*
207          * cause_ipi functions are required to include a full barrier
208          * before doing whatever causes the IPI.
209          */
210         smp_ops->cause_ipi(cpu, info->data);
211 }
212
213 irqreturn_t smp_ipi_demux(void)
214 {
215         struct cpu_messages *info = &__get_cpu_var(ipi_message);
216         unsigned int all;
217
218         mb();   /* order any irq clear */
219
220         do {
221                 all = xchg(&info->messages, 0);
222
223 #ifdef __BIG_ENDIAN
224                 if (all & (1 << (24 - 8 * PPC_MSG_CALL_FUNCTION)))
225                         generic_smp_call_function_interrupt();
226                 if (all & (1 << (24 - 8 * PPC_MSG_RESCHEDULE)))
227                         scheduler_ipi();
228                 if (all & (1 << (24 - 8 * PPC_MSG_CALL_FUNC_SINGLE)))
229                         generic_smp_call_function_single_interrupt();
230                 if (all & (1 << (24 - 8 * PPC_MSG_DEBUGGER_BREAK)))
231                         debug_ipi_action(0, NULL);
232 #else
233 #error Unsupported ENDIAN
234 #endif
235         } while (info->messages);
236
237         return IRQ_HANDLED;
238 }
239 #endif /* CONFIG_PPC_SMP_MUXED_IPI */
240
241 static inline void do_message_pass(int cpu, int msg)
242 {
243         if (smp_ops->message_pass)
244                 smp_ops->message_pass(cpu, msg);
245 #ifdef CONFIG_PPC_SMP_MUXED_IPI
246         else
247                 smp_muxed_ipi_message_pass(cpu, msg);
248 #endif
249 }
250
251 void smp_send_reschedule(int cpu)
252 {
253         if (likely(smp_ops))
254                 do_message_pass(cpu, PPC_MSG_RESCHEDULE);
255 }
256 EXPORT_SYMBOL_GPL(smp_send_reschedule);
257
258 void arch_send_call_function_single_ipi(int cpu)
259 {
260         do_message_pass(cpu, PPC_MSG_CALL_FUNC_SINGLE);
261 }
262
263 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
264 {
265         unsigned int cpu;
266
267         for_each_cpu(cpu, mask)
268                 do_message_pass(cpu, PPC_MSG_CALL_FUNCTION);
269 }
270
271 #if defined(CONFIG_DEBUGGER) || defined(CONFIG_KEXEC)
272 void smp_send_debugger_break(void)
273 {
274         int cpu;
275         int me = raw_smp_processor_id();
276
277         if (unlikely(!smp_ops))
278                 return;
279
280         for_each_online_cpu(cpu)
281                 if (cpu != me)
282                         do_message_pass(cpu, PPC_MSG_DEBUGGER_BREAK);
283 }
284 #endif
285
286 #ifdef CONFIG_KEXEC
287 void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *))
288 {
289         crash_ipi_function_ptr = crash_ipi_callback;
290         if (crash_ipi_callback) {
291                 mb();
292                 smp_send_debugger_break();
293         }
294 }
295 #endif
296
297 static void stop_this_cpu(void *dummy)
298 {
299         /* Remove this CPU */
300         set_cpu_online(smp_processor_id(), false);
301
302         local_irq_disable();
303         while (1)
304                 ;
305 }
306
307 void smp_send_stop(void)
308 {
309         smp_call_function(stop_this_cpu, NULL, 0);
310 }
311
312 struct thread_info *current_set[NR_CPUS];
313
314 static void __devinit smp_store_cpu_info(int id)
315 {
316         per_cpu(cpu_pvr, id) = mfspr(SPRN_PVR);
317 #ifdef CONFIG_PPC_FSL_BOOK3E
318         per_cpu(next_tlbcam_idx, id)
319                 = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) - 1;
320 #endif
321 }
322
323 void __init smp_prepare_cpus(unsigned int max_cpus)
324 {
325         unsigned int cpu;
326
327         DBG("smp_prepare_cpus\n");
328
329         /* 
330          * setup_cpu may need to be called on the boot cpu. We havent
331          * spun any cpus up but lets be paranoid.
332          */
333         BUG_ON(boot_cpuid != smp_processor_id());
334
335         /* Fixup boot cpu */
336         smp_store_cpu_info(boot_cpuid);
337         cpu_callin_map[boot_cpuid] = 1;
338
339         for_each_possible_cpu(cpu) {
340                 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu),
341                                         GFP_KERNEL, cpu_to_node(cpu));
342                 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
343                                         GFP_KERNEL, cpu_to_node(cpu));
344         }
345
346         cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
347         cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
348
349         if (smp_ops)
350                 if (smp_ops->probe)
351                         max_cpus = smp_ops->probe();
352                 else
353                         max_cpus = NR_CPUS;
354         else
355                 max_cpus = 1;
356 }
357
358 void __devinit smp_prepare_boot_cpu(void)
359 {
360         BUG_ON(smp_processor_id() != boot_cpuid);
361 #ifdef CONFIG_PPC64
362         paca[boot_cpuid].__current = current;
363 #endif
364         current_set[boot_cpuid] = task_thread_info(current);
365 }
366
367 #ifdef CONFIG_HOTPLUG_CPU
368
369 int generic_cpu_disable(void)
370 {
371         unsigned int cpu = smp_processor_id();
372
373         if (cpu == boot_cpuid)
374                 return -EBUSY;
375
376         set_cpu_online(cpu, false);
377 #ifdef CONFIG_PPC64
378         vdso_data->processorCount--;
379 #endif
380         migrate_irqs();
381         return 0;
382 }
383
384 void generic_cpu_die(unsigned int cpu)
385 {
386         int i;
387
388         for (i = 0; i < 100; i++) {
389                 smp_rmb();
390                 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
391                         return;
392                 msleep(100);
393         }
394         printk(KERN_ERR "CPU%d didn't die...\n", cpu);
395 }
396
397 void generic_mach_cpu_die(void)
398 {
399         unsigned int cpu;
400
401         local_irq_disable();
402         idle_task_exit();
403         cpu = smp_processor_id();
404         printk(KERN_DEBUG "CPU%d offline\n", cpu);
405         __get_cpu_var(cpu_state) = CPU_DEAD;
406         smp_wmb();
407         while (__get_cpu_var(cpu_state) != CPU_UP_PREPARE)
408                 cpu_relax();
409 }
410
411 void generic_set_cpu_dead(unsigned int cpu)
412 {
413         per_cpu(cpu_state, cpu) = CPU_DEAD;
414 }
415
416 int generic_check_cpu_restart(unsigned int cpu)
417 {
418         return per_cpu(cpu_state, cpu) == CPU_UP_PREPARE;
419 }
420 #endif
421
422 static void cpu_idle_thread_init(unsigned int cpu, struct task_struct *idle)
423 {
424         struct thread_info *ti = task_thread_info(idle);
425
426 #ifdef CONFIG_PPC64
427         paca[cpu].__current = idle;
428         paca[cpu].kstack = (unsigned long)ti + THREAD_SIZE - STACK_FRAME_OVERHEAD;
429 #endif
430         ti->cpu = cpu;
431         secondary_ti = current_set[cpu] = ti;
432 }
433
434 int __cpuinit __cpu_up(unsigned int cpu, struct task_struct *tidle)
435 {
436         int rc, c;
437
438         if (smp_ops == NULL ||
439             (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
440                 return -EINVAL;
441
442         cpu_idle_thread_init(cpu, tidle);
443
444         /* Make sure callin-map entry is 0 (can be leftover a CPU
445          * hotplug
446          */
447         cpu_callin_map[cpu] = 0;
448
449         /* The information for processor bringup must
450          * be written out to main store before we release
451          * the processor.
452          */
453         smp_mb();
454
455         /* wake up cpus */
456         DBG("smp: kicking cpu %d\n", cpu);
457         rc = smp_ops->kick_cpu(cpu);
458         if (rc) {
459                 pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
460                 return rc;
461         }
462
463         /*
464          * wait to see if the cpu made a callin (is actually up).
465          * use this value that I found through experimentation.
466          * -- Cort
467          */
468         if (system_state < SYSTEM_RUNNING)
469                 for (c = 50000; c && !cpu_callin_map[cpu]; c--)
470                         udelay(100);
471 #ifdef CONFIG_HOTPLUG_CPU
472         else
473                 /*
474                  * CPUs can take much longer to come up in the
475                  * hotplug case.  Wait five seconds.
476                  */
477                 for (c = 5000; c && !cpu_callin_map[cpu]; c--)
478                         msleep(1);
479 #endif
480
481         if (!cpu_callin_map[cpu]) {
482                 printk(KERN_ERR "Processor %u is stuck.\n", cpu);
483                 return -ENOENT;
484         }
485
486         DBG("Processor %u found.\n", cpu);
487
488         if (smp_ops->give_timebase)
489                 smp_ops->give_timebase();
490
491         /* Wait until cpu puts itself in the online map */
492         while (!cpu_online(cpu))
493                 cpu_relax();
494
495         return 0;
496 }
497
498 /* Return the value of the reg property corresponding to the given
499  * logical cpu.
500  */
501 int cpu_to_core_id(int cpu)
502 {
503         struct device_node *np;
504         const int *reg;
505         int id = -1;
506
507         np = of_get_cpu_node(cpu, NULL);
508         if (!np)
509                 goto out;
510
511         reg = of_get_property(np, "reg", NULL);
512         if (!reg)
513                 goto out;
514
515         id = *reg;
516 out:
517         of_node_put(np);
518         return id;
519 }
520
521 /* Helper routines for cpu to core mapping */
522 int cpu_core_index_of_thread(int cpu)
523 {
524         return cpu >> threads_shift;
525 }
526 EXPORT_SYMBOL_GPL(cpu_core_index_of_thread);
527
528 int cpu_first_thread_of_core(int core)
529 {
530         return core << threads_shift;
531 }
532 EXPORT_SYMBOL_GPL(cpu_first_thread_of_core);
533
534 /* Must be called when no change can occur to cpu_present_mask,
535  * i.e. during cpu online or offline.
536  */
537 static struct device_node *cpu_to_l2cache(int cpu)
538 {
539         struct device_node *np;
540         struct device_node *cache;
541
542         if (!cpu_present(cpu))
543                 return NULL;
544
545         np = of_get_cpu_node(cpu, NULL);
546         if (np == NULL)
547                 return NULL;
548
549         cache = of_find_next_cache_node(np);
550
551         of_node_put(np);
552
553         return cache;
554 }
555
556 /* Activate a secondary processor. */
557 void __devinit start_secondary(void *unused)
558 {
559         unsigned int cpu = smp_processor_id();
560         struct device_node *l2_cache;
561         int i, base;
562
563         atomic_inc(&init_mm.mm_count);
564         current->active_mm = &init_mm;
565
566         smp_store_cpu_info(cpu);
567         set_dec(tb_ticks_per_jiffy);
568         preempt_disable();
569         cpu_callin_map[cpu] = 1;
570
571         if (smp_ops->setup_cpu)
572                 smp_ops->setup_cpu(cpu);
573         if (smp_ops->take_timebase)
574                 smp_ops->take_timebase();
575
576         secondary_cpu_time_init();
577
578 #ifdef CONFIG_PPC64
579         if (system_state == SYSTEM_RUNNING)
580                 vdso_data->processorCount++;
581
582         vdso_getcpu_init();
583 #endif
584         notify_cpu_starting(cpu);
585         set_cpu_online(cpu, true);
586         /* Update sibling maps */
587         base = cpu_first_thread_sibling(cpu);
588         for (i = 0; i < threads_per_core; i++) {
589                 if (cpu_is_offline(base + i))
590                         continue;
591                 cpumask_set_cpu(cpu, cpu_sibling_mask(base + i));
592                 cpumask_set_cpu(base + i, cpu_sibling_mask(cpu));
593
594                 /* cpu_core_map should be a superset of
595                  * cpu_sibling_map even if we don't have cache
596                  * information, so update the former here, too.
597                  */
598                 cpumask_set_cpu(cpu, cpu_core_mask(base + i));
599                 cpumask_set_cpu(base + i, cpu_core_mask(cpu));
600         }
601         l2_cache = cpu_to_l2cache(cpu);
602         for_each_online_cpu(i) {
603                 struct device_node *np = cpu_to_l2cache(i);
604                 if (!np)
605                         continue;
606                 if (np == l2_cache) {
607                         cpumask_set_cpu(cpu, cpu_core_mask(i));
608                         cpumask_set_cpu(i, cpu_core_mask(cpu));
609                 }
610                 of_node_put(np);
611         }
612         of_node_put(l2_cache);
613
614         local_irq_enable();
615
616         cpu_idle();
617
618         BUG();
619 }
620
621 int setup_profiling_timer(unsigned int multiplier)
622 {
623         return 0;
624 }
625
626 void __init smp_cpus_done(unsigned int max_cpus)
627 {
628         cpumask_var_t old_mask;
629
630         /* We want the setup_cpu() here to be called from CPU 0, but our
631          * init thread may have been "borrowed" by another CPU in the meantime
632          * se we pin us down to CPU 0 for a short while
633          */
634         alloc_cpumask_var(&old_mask, GFP_NOWAIT);
635         cpumask_copy(old_mask, tsk_cpus_allowed(current));
636         set_cpus_allowed_ptr(current, cpumask_of(boot_cpuid));
637         
638         if (smp_ops && smp_ops->setup_cpu)
639                 smp_ops->setup_cpu(boot_cpuid);
640
641         set_cpus_allowed_ptr(current, old_mask);
642
643         free_cpumask_var(old_mask);
644
645         if (smp_ops && smp_ops->bringup_done)
646                 smp_ops->bringup_done();
647
648         dump_numa_cpu_topology();
649
650 }
651
652 int arch_sd_sibling_asym_packing(void)
653 {
654         if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
655                 printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
656                 return SD_ASYM_PACKING;
657         }
658         return 0;
659 }
660
661 #ifdef CONFIG_HOTPLUG_CPU
662 int __cpu_disable(void)
663 {
664         struct device_node *l2_cache;
665         int cpu = smp_processor_id();
666         int base, i;
667         int err;
668
669         if (!smp_ops->cpu_disable)
670                 return -ENOSYS;
671
672         err = smp_ops->cpu_disable();
673         if (err)
674                 return err;
675
676         /* Update sibling maps */
677         base = cpu_first_thread_sibling(cpu);
678         for (i = 0; i < threads_per_core; i++) {
679                 cpumask_clear_cpu(cpu, cpu_sibling_mask(base + i));
680                 cpumask_clear_cpu(base + i, cpu_sibling_mask(cpu));
681                 cpumask_clear_cpu(cpu, cpu_core_mask(base + i));
682                 cpumask_clear_cpu(base + i, cpu_core_mask(cpu));
683         }
684
685         l2_cache = cpu_to_l2cache(cpu);
686         for_each_present_cpu(i) {
687                 struct device_node *np = cpu_to_l2cache(i);
688                 if (!np)
689                         continue;
690                 if (np == l2_cache) {
691                         cpumask_clear_cpu(cpu, cpu_core_mask(i));
692                         cpumask_clear_cpu(i, cpu_core_mask(cpu));
693                 }
694                 of_node_put(np);
695         }
696         of_node_put(l2_cache);
697
698
699         return 0;
700 }
701
702 void __cpu_die(unsigned int cpu)
703 {
704         if (smp_ops->cpu_die)
705                 smp_ops->cpu_die(cpu);
706 }
707
708 static DEFINE_MUTEX(powerpc_cpu_hotplug_driver_mutex);
709
710 void cpu_hotplug_driver_lock()
711 {
712         mutex_lock(&powerpc_cpu_hotplug_driver_mutex);
713 }
714
715 void cpu_hotplug_driver_unlock()
716 {
717         mutex_unlock(&powerpc_cpu_hotplug_driver_mutex);
718 }
719
720 void cpu_die(void)
721 {
722         if (ppc_md.cpu_die)
723                 ppc_md.cpu_die();
724
725         /* If we return, we re-enter start_secondary */
726         start_secondary_resume();
727 }
728
729 #endif