nmi.c 6.0 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281
  1. /* Pseudo NMI support on sparc64 systems.
  2. *
  3. * Copyright (C) 2009 David S. Miller <davem@davemloft.net>
  4. *
  5. * The NMI watchdog support and infrastructure is based almost
  6. * entirely upon the x86 NMI support code.
  7. */
  8. #include <linux/kernel.h>
  9. #include <linux/param.h>
  10. #include <linux/init.h>
  11. #include <linux/percpu.h>
  12. #include <linux/nmi.h>
  13. #include <linux/module.h>
  14. #include <linux/kprobes.h>
  15. #include <linux/kernel_stat.h>
  16. #include <linux/reboot.h>
  17. #include <linux/slab.h>
  18. #include <linux/kdebug.h>
  19. #include <linux/delay.h>
  20. #include <linux/smp.h>
  21. #include <asm/perf_event.h>
  22. #include <asm/ptrace.h>
  23. #include <asm/local.h>
  24. #include <asm/pcr.h>
  25. /* We don't have a real NMI on sparc64, but we can fake one
  26. * up using profiling counter overflow interrupts and interrupt
  27. * levels.
  28. *
  29. * The profile overflow interrupts at level 15, so we use
  30. * level 14 as our IRQ off level.
  31. */
  32. static int panic_on_timeout;
  33. /* nmi_active:
  34. * >0: the NMI watchdog is active, but can be disabled
  35. * <0: the NMI watchdog has not been set up, and cannot be enabled
  36. * 0: the NMI watchdog is disabled, but can be enabled
  37. */
  38. atomic_t nmi_active = ATOMIC_INIT(0); /* oprofile uses this */
  39. EXPORT_SYMBOL(nmi_active);
  40. static unsigned int nmi_hz = HZ;
  41. static DEFINE_PER_CPU(short, wd_enabled);
  42. static int endflag __initdata;
  43. static DEFINE_PER_CPU(unsigned int, last_irq_sum);
  44. static DEFINE_PER_CPU(long, alert_counter);
  45. static DEFINE_PER_CPU(int, nmi_touch);
  46. void touch_nmi_watchdog(void)
  47. {
  48. if (atomic_read(&nmi_active)) {
  49. int cpu;
  50. for_each_present_cpu(cpu) {
  51. if (per_cpu(nmi_touch, cpu) != 1)
  52. per_cpu(nmi_touch, cpu) = 1;
  53. }
  54. }
  55. touch_softlockup_watchdog();
  56. }
  57. EXPORT_SYMBOL(touch_nmi_watchdog);
  58. static void die_nmi(const char *str, struct pt_regs *regs, int do_panic)
  59. {
  60. if (notify_die(DIE_NMIWATCHDOG, str, regs, 0,
  61. pt_regs_trap_type(regs), SIGINT) == NOTIFY_STOP)
  62. return;
  63. console_verbose();
  64. bust_spinlocks(1);
  65. printk(KERN_EMERG "%s", str);
  66. printk(" on CPU%d, ip %08lx, registers:\n",
  67. smp_processor_id(), regs->tpc);
  68. show_regs(regs);
  69. dump_stack();
  70. bust_spinlocks(0);
  71. if (do_panic || panic_on_oops)
  72. panic("Non maskable interrupt");
  73. nmi_exit();
  74. local_irq_enable();
  75. do_exit(SIGBUS);
  76. }
  77. notrace __kprobes void perfctr_irq(int irq, struct pt_regs *regs)
  78. {
  79. unsigned int sum, touched = 0;
  80. int cpu = smp_processor_id();
  81. clear_softint(1 << irq);
  82. local_cpu_data().__nmi_count++;
  83. nmi_enter();
  84. if (notify_die(DIE_NMI, "nmi", regs, 0,
  85. pt_regs_trap_type(regs), SIGINT) == NOTIFY_STOP)
  86. touched = 1;
  87. else
  88. pcr_ops->write(PCR_PIC_PRIV);
  89. sum = kstat_irqs_cpu(0, cpu);
  90. if (__get_cpu_var(nmi_touch)) {
  91. __get_cpu_var(nmi_touch) = 0;
  92. touched = 1;
  93. }
  94. if (!touched && __get_cpu_var(last_irq_sum) == sum) {
  95. __this_cpu_inc(per_cpu_var(alert_counter));
  96. if (__this_cpu_read(per_cpu_var(alert_counter)) == 30 * nmi_hz)
  97. die_nmi("BUG: NMI Watchdog detected LOCKUP",
  98. regs, panic_on_timeout);
  99. } else {
  100. __get_cpu_var(last_irq_sum) = sum;
  101. __this_cpu_write(per_cpu_var(alert_counter), 0);
  102. }
  103. if (__get_cpu_var(wd_enabled)) {
  104. write_pic(picl_value(nmi_hz));
  105. pcr_ops->write(pcr_enable);
  106. }
  107. nmi_exit();
  108. }
  109. static inline unsigned int get_nmi_count(int cpu)
  110. {
  111. return cpu_data(cpu).__nmi_count;
  112. }
  113. static __init void nmi_cpu_busy(void *data)
  114. {
  115. local_irq_enable_in_hardirq();
  116. while (endflag == 0)
  117. mb();
  118. }
  119. static void report_broken_nmi(int cpu, int *prev_nmi_count)
  120. {
  121. printk(KERN_CONT "\n");
  122. printk(KERN_WARNING
  123. "WARNING: CPU#%d: NMI appears to be stuck (%d->%d)!\n",
  124. cpu, prev_nmi_count[cpu], get_nmi_count(cpu));
  125. printk(KERN_WARNING
  126. "Please report this to bugzilla.kernel.org,\n");
  127. printk(KERN_WARNING
  128. "and attach the output of the 'dmesg' command.\n");
  129. per_cpu(wd_enabled, cpu) = 0;
  130. atomic_dec(&nmi_active);
  131. }
  132. void stop_nmi_watchdog(void *unused)
  133. {
  134. pcr_ops->write(PCR_PIC_PRIV);
  135. __get_cpu_var(wd_enabled) = 0;
  136. atomic_dec(&nmi_active);
  137. }
  138. static int __init check_nmi_watchdog(void)
  139. {
  140. unsigned int *prev_nmi_count;
  141. int cpu, err;
  142. if (!atomic_read(&nmi_active))
  143. return 0;
  144. prev_nmi_count = kmalloc(nr_cpu_ids * sizeof(unsigned int), GFP_KERNEL);
  145. if (!prev_nmi_count) {
  146. err = -ENOMEM;
  147. goto error;
  148. }
  149. printk(KERN_INFO "Testing NMI watchdog ... ");
  150. smp_call_function(nmi_cpu_busy, (void *)&endflag, 0);
  151. for_each_possible_cpu(cpu)
  152. prev_nmi_count[cpu] = get_nmi_count(cpu);
  153. local_irq_enable();
  154. mdelay((20 * 1000) / nmi_hz); /* wait 20 ticks */
  155. for_each_online_cpu(cpu) {
  156. if (!per_cpu(wd_enabled, cpu))
  157. continue;
  158. if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5)
  159. report_broken_nmi(cpu, prev_nmi_count);
  160. }
  161. endflag = 1;
  162. if (!atomic_read(&nmi_active)) {
  163. kfree(prev_nmi_count);
  164. atomic_set(&nmi_active, -1);
  165. err = -ENODEV;
  166. goto error;
  167. }
  168. printk("OK.\n");
  169. nmi_hz = 1;
  170. kfree(prev_nmi_count);
  171. return 0;
  172. error:
  173. on_each_cpu(stop_nmi_watchdog, NULL, 1);
  174. return err;
  175. }
  176. void start_nmi_watchdog(void *unused)
  177. {
  178. __get_cpu_var(wd_enabled) = 1;
  179. atomic_inc(&nmi_active);
  180. pcr_ops->write(PCR_PIC_PRIV);
  181. write_pic(picl_value(nmi_hz));
  182. pcr_ops->write(pcr_enable);
  183. }
  184. static void nmi_adjust_hz_one(void *unused)
  185. {
  186. if (!__get_cpu_var(wd_enabled))
  187. return;
  188. pcr_ops->write(PCR_PIC_PRIV);
  189. write_pic(picl_value(nmi_hz));
  190. pcr_ops->write(pcr_enable);
  191. }
  192. void nmi_adjust_hz(unsigned int new_hz)
  193. {
  194. nmi_hz = new_hz;
  195. on_each_cpu(nmi_adjust_hz_one, NULL, 1);
  196. }
  197. EXPORT_SYMBOL_GPL(nmi_adjust_hz);
  198. static int nmi_shutdown(struct notifier_block *nb, unsigned long cmd, void *p)
  199. {
  200. on_each_cpu(stop_nmi_watchdog, NULL, 1);
  201. return 0;
  202. }
  203. static struct notifier_block nmi_reboot_notifier = {
  204. .notifier_call = nmi_shutdown,
  205. };
  206. int __init nmi_init(void)
  207. {
  208. int err;
  209. on_each_cpu(start_nmi_watchdog, NULL, 1);
  210. err = check_nmi_watchdog();
  211. if (!err) {
  212. err = register_reboot_notifier(&nmi_reboot_notifier);
  213. if (err) {
  214. on_each_cpu(stop_nmi_watchdog, NULL, 1);
  215. atomic_set(&nmi_active, -1);
  216. }
  217. }
  218. if (!err)
  219. init_hw_perf_events();
  220. return err;
  221. }
  222. static int __init setup_nmi_watchdog(char *str)
  223. {
  224. if (!strncmp(str, "panic", 5))
  225. panic_on_timeout = 1;
  226. return 0;
  227. }
  228. __setup("nmi_watchdog=", setup_nmi_watchdog);