Linux Kernel  3.7.1
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
smp.c
Go to the documentation of this file.
1 /*
2  * SMP support for ppc.
3  *
4  * Written by Cort Dougan ([email protected]) borrowing a great
5  * deal of code from the sparc and intel versions.
6  *
7  * Copyright (C) 1999 Cort Dougan <[email protected]>
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 
67 
68 DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
70 
71 EXPORT_PER_CPU_SYMBOL(cpu_sibling_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 
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  generic_set_cpu_up(nr);
106  smp_wmb();
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],
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 }
257 
259 {
260  do_message_pass(cpu, PPC_MSG_CALL_FUNC_SINGLE);
261 }
262 
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 
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 */
301 
303  while (1)
304  ;
305 }
306 
307 void smp_send_stop(void)
308 {
310 }
311 
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  */
334 
335  /* Fixup boot cpu */
336  smp_store_cpu_info(boot_cpuid);
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 
359 {
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 
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 /*
417  * The cpu_state should be set to CPU_UP_PREPARE in kick_cpu(), otherwise
418  * the cpu_state is always CPU_DEAD after calling generic_set_cpu_dead(),
419  * which makes the delay in generic_cpu_die() not happen.
420  */
421 void generic_set_cpu_up(unsigned int cpu)
422 {
423  per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
424 }
425 
426 int generic_check_cpu_restart(unsigned int cpu)
427 {
428  return per_cpu(cpu_state, cpu) == CPU_UP_PREPARE;
429 }
430 #endif
431 
432 static void cpu_idle_thread_init(unsigned int cpu, struct task_struct *idle)
433 {
434  struct thread_info *ti = task_thread_info(idle);
435 
436 #ifdef CONFIG_PPC64
437  paca[cpu].__current = idle;
438  paca[cpu].kstack = (unsigned long)ti + THREAD_SIZE - STACK_FRAME_OVERHEAD;
439 #endif
440  ti->cpu = cpu;
441  secondary_ti = current_set[cpu] = ti;
442 }
443 
444 int __cpuinit __cpu_up(unsigned int cpu, struct task_struct *tidle)
445 {
446  int rc, c;
447 
448  if (smp_ops == NULL ||
449  (smp_ops->cpu_bootable && !smp_ops->cpu_bootable(cpu)))
450  return -EINVAL;
451 
452  cpu_idle_thread_init(cpu, tidle);
453 
454  /* Make sure callin-map entry is 0 (can be leftover a CPU
455  * hotplug
456  */
457  cpu_callin_map[cpu] = 0;
458 
459  /* The information for processor bringup must
460  * be written out to main store before we release
461  * the processor.
462  */
463  smp_mb();
464 
465  /* wake up cpus */
466  DBG("smp: kicking cpu %d\n", cpu);
467  rc = smp_ops->kick_cpu(cpu);
468  if (rc) {
469  pr_err("smp: failed starting cpu %d (rc %d)\n", cpu, rc);
470  return rc;
471  }
472 
473  /*
474  * wait to see if the cpu made a callin (is actually up).
475  * use this value that I found through experimentation.
476  * -- Cort
477  */
479  for (c = 50000; c && !cpu_callin_map[cpu]; c--)
480  udelay(100);
481 #ifdef CONFIG_HOTPLUG_CPU
482  else
483  /*
484  * CPUs can take much longer to come up in the
485  * hotplug case. Wait five seconds.
486  */
487  for (c = 5000; c && !cpu_callin_map[cpu]; c--)
488  msleep(1);
489 #endif
490 
491  if (!cpu_callin_map[cpu]) {
492  printk(KERN_ERR "Processor %u is stuck.\n", cpu);
493  return -ENOENT;
494  }
495 
496  DBG("Processor %u found.\n", cpu);
497 
498  if (smp_ops->give_timebase)
499  smp_ops->give_timebase();
500 
501  /* Wait until cpu puts itself in the online map */
502  while (!cpu_online(cpu))
503  cpu_relax();
504 
505  return 0;
506 }
507 
508 /* Return the value of the reg property corresponding to the given
509  * logical cpu.
510  */
511 int cpu_to_core_id(int cpu)
512 {
513  struct device_node *np;
514  const int *reg;
515  int id = -1;
516 
517  np = of_get_cpu_node(cpu, NULL);
518  if (!np)
519  goto out;
520 
521  reg = of_get_property(np, "reg", NULL);
522  if (!reg)
523  goto out;
524 
525  id = *reg;
526 out:
527  of_node_put(np);
528  return id;
529 }
530 
531 /* Helper routines for cpu to core mapping */
533 {
534  return cpu >> threads_shift;
535 }
537 
539 {
540  return core << threads_shift;
541 }
543 
544 /* Must be called when no change can occur to cpu_present_mask,
545  * i.e. during cpu online or offline.
546  */
547 static struct device_node *cpu_to_l2cache(int cpu)
548 {
549  struct device_node *np;
550  struct device_node *cache;
551 
552  if (!cpu_present(cpu))
553  return NULL;
554 
555  np = of_get_cpu_node(cpu, NULL);
556  if (np == NULL)
557  return NULL;
558 
559  cache = of_find_next_cache_node(np);
560 
561  of_node_put(np);
562 
563  return cache;
564 }
565 
566 /* Activate a secondary processor. */
568 {
569  unsigned int cpu = smp_processor_id();
570  struct device_node *l2_cache;
571  int i, base;
572 
573  atomic_inc(&init_mm.mm_count);
574  current->active_mm = &init_mm;
575 
576  smp_store_cpu_info(cpu);
577  set_dec(tb_ticks_per_jiffy);
578  preempt_disable();
579  cpu_callin_map[cpu] = 1;
580 
581  if (smp_ops->setup_cpu)
582  smp_ops->setup_cpu(cpu);
583  if (smp_ops->take_timebase)
584  smp_ops->take_timebase();
585 
587 
588 #ifdef CONFIG_PPC64
590  vdso_data->processorCount++;
591 
592  vdso_getcpu_init();
593 #endif
594  notify_cpu_starting(cpu);
595  set_cpu_online(cpu, true);
596  /* Update sibling maps */
597  base = cpu_first_thread_sibling(cpu);
598  for (i = 0; i < threads_per_core; i++) {
599  if (cpu_is_offline(base + i))
600  continue;
601  cpumask_set_cpu(cpu, cpu_sibling_mask(base + i));
602  cpumask_set_cpu(base + i, cpu_sibling_mask(cpu));
603 
604  /* cpu_core_map should be a superset of
605  * cpu_sibling_map even if we don't have cache
606  * information, so update the former here, too.
607  */
608  cpumask_set_cpu(cpu, cpu_core_mask(base + i));
609  cpumask_set_cpu(base + i, cpu_core_mask(cpu));
610  }
611  l2_cache = cpu_to_l2cache(cpu);
613  struct device_node *np = cpu_to_l2cache(i);
614  if (!np)
615  continue;
616  if (np == l2_cache) {
617  cpumask_set_cpu(cpu, cpu_core_mask(i));
618  cpumask_set_cpu(i, cpu_core_mask(cpu));
619  }
620  of_node_put(np);
621  }
622  of_node_put(l2_cache);
623 
625 
626  cpu_idle();
627 
628  BUG();
629 }
630 
631 int setup_profiling_timer(unsigned int multiplier)
632 {
633  return 0;
634 }
635 
636 void __init smp_cpus_done(unsigned int max_cpus)
637 {
638  cpumask_var_t old_mask;
639 
640  /* We want the setup_cpu() here to be called from CPU 0, but our
641  * init thread may have been "borrowed" by another CPU in the meantime
642  * se we pin us down to CPU 0 for a short while
643  */
644  alloc_cpumask_var(&old_mask, GFP_NOWAIT);
645  cpumask_copy(old_mask, tsk_cpus_allowed(current));
646  set_cpus_allowed_ptr(current, cpumask_of(boot_cpuid));
647 
648  if (smp_ops && smp_ops->setup_cpu)
649  smp_ops->setup_cpu(boot_cpuid);
650 
651  set_cpus_allowed_ptr(current, old_mask);
652 
653  free_cpumask_var(old_mask);
654 
655  if (smp_ops && smp_ops->bringup_done)
656  smp_ops->bringup_done();
657 
659 
660 }
661 
663 {
665  printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
666  return SD_ASYM_PACKING;
667  }
668  return 0;
669 }
670 
671 #ifdef CONFIG_HOTPLUG_CPU
672 int __cpu_disable(void)
673 {
674  struct device_node *l2_cache;
675  int cpu = smp_processor_id();
676  int base, i;
677  int err;
678 
679  if (!smp_ops->cpu_disable)
680  return -ENOSYS;
681 
682  err = smp_ops->cpu_disable();
683  if (err)
684  return err;
685 
686  /* Update sibling maps */
687  base = cpu_first_thread_sibling(cpu);
688  for (i = 0; i < threads_per_core; i++) {
689  cpumask_clear_cpu(cpu, cpu_sibling_mask(base + i));
690  cpumask_clear_cpu(base + i, cpu_sibling_mask(cpu));
691  cpumask_clear_cpu(cpu, cpu_core_mask(base + i));
692  cpumask_clear_cpu(base + i, cpu_core_mask(cpu));
693  }
694 
695  l2_cache = cpu_to_l2cache(cpu);
697  struct device_node *np = cpu_to_l2cache(i);
698  if (!np)
699  continue;
700  if (np == l2_cache) {
701  cpumask_clear_cpu(cpu, cpu_core_mask(i));
702  cpumask_clear_cpu(i, cpu_core_mask(cpu));
703  }
704  of_node_put(np);
705  }
706  of_node_put(l2_cache);
707 
708 
709  return 0;
710 }
711 
712 void __cpu_die(unsigned int cpu)
713 {
714  if (smp_ops->cpu_die)
715  smp_ops->cpu_die(cpu);
716 }
717 
718 static DEFINE_MUTEX(powerpc_cpu_hotplug_driver_mutex);
719 
720 void cpu_hotplug_driver_lock()
721 {
722  mutex_lock(&powerpc_cpu_hotplug_driver_mutex);
723 }
724 
725 void cpu_hotplug_driver_unlock()
726 {
727  mutex_unlock(&powerpc_cpu_hotplug_driver_mutex);
728 }
729 
730 void cpu_die(void)
731 {
732  if (ppc_md.cpu_die)
733  ppc_md.cpu_die();
734 
735  /* If we return, we re-enter start_secondary */
736  start_secondary_resume();
737 }
738 
739 #endif