Linux kernel & device driver programming

Cross-Referenced Linux and Device Driver Code

[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ]
Version: [ 2.6.11.8 ] [ 2.6.25 ] [ 2.6.25.8 ] [ 2.6.31.13 ] Architecture: [ i386 ]
  1 /*
  2  * Machine check handler.
  3  *
  4  * K8 parts Copyright 2002,2003 Andi Kleen, SuSE Labs.
  5  * Rest from unknown author(s).
  6  * 2004 Andi Kleen. Rewrote most of it.
  7  * Copyright 2008 Intel Corporation
  8  * Author: Andi Kleen
  9  */
 10 #include <linux/thread_info.h>
 11 #include <linux/capability.h>
 12 #include <linux/miscdevice.h>
 13 #include <linux/interrupt.h>
 14 #include <linux/ratelimit.h>
 15 #include <linux/kallsyms.h>
 16 #include <linux/rcupdate.h>
 17 #include <linux/kobject.h>
 18 #include <linux/uaccess.h>
 19 #include <linux/kdebug.h>
 20 #include <linux/kernel.h>
 21 #include <linux/percpu.h>
 22 #include <linux/string.h>
 23 #include <linux/sysdev.h>
 24 #include <linux/delay.h>
 25 #include <linux/ctype.h>
 26 #include <linux/sched.h>
 27 #include <linux/sysfs.h>
 28 #include <linux/types.h>
 29 #include <linux/init.h>
 30 #include <linux/kmod.h>
 31 #include <linux/poll.h>
 32 #include <linux/nmi.h>
 33 #include <linux/cpu.h>
 34 #include <linux/smp.h>
 35 #include <linux/fs.h>
 36 #include <linux/mm.h>
 37 
 38 #include <asm/processor.h>
 39 #include <asm/hw_irq.h>
 40 #include <asm/apic.h>
 41 #include <asm/idle.h>
 42 #include <asm/ipi.h>
 43 #include <asm/mce.h>
 44 #include <asm/msr.h>
 45 
 46 #include "mce-internal.h"
 47 
 48 /* Handle unconfigured int18 (should never happen) */
 49 static void unexpected_machine_check(struct pt_regs *regs, long error_code)
 50 {
 51         printk(KERN_ERR "CPU#%d: Unexpected int18 (Machine Check).\n",
 52                smp_processor_id());
 53 }
 54 
 55 /* Call the installed machine check handler for this CPU setup. */
 56 void (*machine_check_vector)(struct pt_regs *, long error_code) =
 57                                                 unexpected_machine_check;
 58 
 59 int mce_disabled __read_mostly;
 60 
 61 #ifdef CONFIG_X86_NEW_MCE
 62 
 63 #define MISC_MCELOG_MINOR       227
 64 
 65 #define SPINUNIT 100    /* 100ns */
 66 
 67 atomic_t mce_entry;
 68 
 69 DEFINE_PER_CPU(unsigned, mce_exception_count);
 70 
 71 /*
 72  * Tolerant levels:
 73  *   0: always panic on uncorrected errors, log corrected errors
 74  *   1: panic or SIGBUS on uncorrected errors, log corrected errors
 75  *   2: SIGBUS or log uncorrected errors (if possible), log corrected errors
 76  *   3: never panic or SIGBUS, log all errors (for testing only)
 77  */
 78 static int                      tolerant                __read_mostly = 1;
 79 static int                      banks                   __read_mostly;
 80 static u64                      *bank                   __read_mostly;
 81 static int                      rip_msr                 __read_mostly;
 82 static int                      mce_bootlog             __read_mostly = -1;
 83 static int                      monarch_timeout         __read_mostly = -1;
 84 static int                      mce_panic_timeout       __read_mostly;
 85 static int                      mce_dont_log_ce         __read_mostly;
 86 int                             mce_cmci_disabled       __read_mostly;
 87 int                             mce_ignore_ce           __read_mostly;
 88 int                             mce_ser                 __read_mostly;
 89 
 90 /* User mode helper program triggered by machine check event */
 91 static unsigned long            mce_need_notify;
 92 static char                     mce_helper[128];
 93 static char                     *mce_helper_argv[2] = { mce_helper, NULL };
 94 
 95 static unsigned long            dont_init_banks;
 96 
 97 static DECLARE_WAIT_QUEUE_HEAD(mce_wait);
 98 static DEFINE_PER_CPU(struct mce, mces_seen);
 99 static int                      cpu_missing;
100 
101 
102 /* MCA banks polled by the period polling timer for corrected events */
103 DEFINE_PER_CPU(mce_banks_t, mce_poll_banks) = {
104         [0 ... BITS_TO_LONGS(MAX_NR_BANKS)-1] = ~0UL
105 };
106 
107 static inline int skip_bank_init(int i)
108 {
109         return i < BITS_PER_LONG && test_bit(i, &dont_init_banks);
110 }
111 
112 static DEFINE_PER_CPU(struct work_struct, mce_work);
113 
114 /* Do initial initialization of a struct mce */
115 void mce_setup(struct mce *m)
116 {
117         memset(m, 0, sizeof(struct mce));
118         m->cpu = m->extcpu = smp_processor_id();
119         rdtscll(m->tsc);
120         /* We hope get_seconds stays lockless */
121         m->time = get_seconds();
122         m->cpuvendor = boot_cpu_data.x86_vendor;
123         m->cpuid = cpuid_eax(1);
124 #ifdef CONFIG_SMP
125         m->socketid = cpu_data(m->extcpu).phys_proc_id;
126 #endif
127         m->apicid = cpu_data(m->extcpu).initial_apicid;
128         rdmsrl(MSR_IA32_MCG_CAP, m->mcgcap);
129 }
130 
131 DEFINE_PER_CPU(struct mce, injectm);
132 EXPORT_PER_CPU_SYMBOL_GPL(injectm);
133 
134 /*
135  * Lockless MCE logging infrastructure.
136  * This avoids deadlocks on printk locks without having to break locks. Also
137  * separate MCEs from kernel messages to avoid bogus bug reports.
138  */
139 
140 static struct mce_log mcelog = {
141         .signature      = MCE_LOG_SIGNATURE,
142         .len            = MCE_LOG_LEN,
143         .recordlen      = sizeof(struct mce),
144 };
145 
146 void mce_log(struct mce *mce)
147 {
148         unsigned next, entry;
149 
150         mce->finished = 0;
151         wmb();
152         for (;;) {
153                 entry = rcu_dereference(mcelog.next);
154                 for (;;) {
155                         /*
156                          * When the buffer fills up discard new entries.
157                          * Assume that the earlier errors are the more
158                          * interesting ones:
159                          */
160                         if (entry >= MCE_LOG_LEN) {
161                                 set_bit(MCE_OVERFLOW,
162                                         (unsigned long *)&mcelog.flags);
163                                 return;
164                         }
165                         /* Old left over entry. Skip: */
166                         if (mcelog.entry[entry].finished) {
167                                 entry++;
168                                 continue;
169                         }
170                         break;
171                 }
172                 smp_rmb();
173                 next = entry + 1;
174                 if (cmpxchg(&mcelog.next, entry, next) == entry)
175                         break;
176         }
177         memcpy(mcelog.entry + entry, mce, sizeof(struct mce));
178         wmb();
179         mcelog.entry[entry].finished = 1;
180         wmb();
181 
182         mce->finished = 1;
183         set_bit(0, &mce_need_notify);
184 }
185 
186 static void print_mce(struct mce *m)
187 {
188         printk(KERN_EMERG
189                "CPU %d: Machine Check Exception: %16Lx Bank %d: %016Lx\n",
190                m->extcpu, m->mcgstatus, m->bank, m->status);
191         if (m->ip) {
192                 printk(KERN_EMERG "RIP%s %02x:<%016Lx> ",
193                        !(m->mcgstatus & MCG_STATUS_EIPV) ? " !INEXACT!" : "",
194                        m->cs, m->ip);
195                 if (m->cs == __KERNEL_CS)
196                         print_symbol("{%s}", m->ip);
197                 printk(KERN_CONT "\n");
198         }
199         printk(KERN_EMERG "TSC %llx ", m->tsc);
200         if (m->addr)
201                 printk(KERN_CONT "ADDR %llx ", m->addr);
202         if (m->misc)
203                 printk(KERN_CONT "MISC %llx ", m->misc);
204         printk(KERN_CONT "\n");
205         printk(KERN_EMERG "PROCESSOR %u:%x TIME %llu SOCKET %u APIC %x\n",
206                         m->cpuvendor, m->cpuid, m->time, m->socketid,
207                         m->apicid);
208 }
209 
210 static void print_mce_head(void)
211 {
212         printk(KERN_EMERG "\nHARDWARE ERROR\n");
213 }
214 
215 static void print_mce_tail(void)
216 {
217         printk(KERN_EMERG "This is not a software problem!\n"
218                "Run through mcelog --ascii to decode and contact your hardware vendor\n");
219 }
220 
221 #define PANIC_TIMEOUT 5 /* 5 seconds */
222 
223 static atomic_t mce_paniced;
224 
225 /* Panic in progress. Enable interrupts and wait for final IPI */
226 static void wait_for_panic(void)
227 {
228         long timeout = PANIC_TIMEOUT*USEC_PER_SEC;
229         preempt_disable();
230         local_irq_enable();
231         while (timeout-- > 0)
232                 udelay(1);
233         if (panic_timeout == 0)
234                 panic_timeout = mce_panic_timeout;
235         panic("Panicing machine check CPU died");
236 }
237 
238 static void mce_panic(char *msg, struct mce *final, char *exp)
239 {
240         int i;
241 
242         /*
243          * Make sure only one CPU runs in machine check panic
244          */
245         if (atomic_add_return(1, &mce_paniced) > 1)
246                 wait_for_panic();
247         barrier();
248 
249         bust_spinlocks(1);
250         console_verbose();
251         print_mce_head();
252         /* First print corrected ones that are still unlogged */
253         for (i = 0; i < MCE_LOG_LEN; i++) {
254                 struct mce *m = &mcelog.entry[i];
255                 if (!(m->status & MCI_STATUS_VAL))
256                         continue;
257                 if (!(m->status & MCI_STATUS_UC))
258                         print_mce(m);
259         }
260         /* Now print uncorrected but with the final one last */
261         for (i = 0; i < MCE_LOG_LEN; i++) {
262                 struct mce *m = &mcelog.entry[i];
263                 if (!(m->status & MCI_STATUS_VAL))
264                         continue;
265                 if (!(m->status & MCI_STATUS_UC))
266                         continue;
267                 if (!final || memcmp(m, final, sizeof(struct mce)))
268                         print_mce(m);
269         }
270         if (final)
271                 print_mce(final);
272         if (cpu_missing)
273                 printk(KERN_EMERG "Some CPUs didn't answer in synchronization\n");
274         print_mce_tail();
275         if (exp)
276                 printk(KERN_EMERG "Machine check: %s\n", exp);
277         if (panic_timeout == 0)
278                 panic_timeout = mce_panic_timeout;
279         panic(msg);
280 }
281 
282 /* Support code for software error injection */
283 
284 static int msr_to_offset(u32 msr)
285 {
286         unsigned bank = __get_cpu_var(injectm.bank);
287         if (msr == rip_msr)
288                 return offsetof(struct mce, ip);
289         if (msr == MSR_IA32_MC0_STATUS + bank*4)
290                 return offsetof(struct mce, status);
291         if (msr == MSR_IA32_MC0_ADDR + bank*4)
292                 return offsetof(struct mce, addr);
293         if (msr == MSR_IA32_MC0_MISC + bank*4)
294                 return offsetof(struct mce, misc);
295         if (msr == MSR_IA32_MCG_STATUS)
296                 return offsetof(struct mce, mcgstatus);
297         return -1;
298 }
299 
300 /* MSR access wrappers used for error injection */
301 static u64 mce_rdmsrl(u32 msr)
302 {
303         u64 v;
304         if (__get_cpu_var(injectm).finished) {
305                 int offset = msr_to_offset(msr);
306                 if (offset < 0)
307                         return 0;
308                 return *(u64 *)((char *)&__get_cpu_var(injectm) + offset);
309         }
310         rdmsrl(msr, v);
311         return v;
312 }
313 
314 static void mce_wrmsrl(u32 msr, u64 v)
315 {
316         if (__get_cpu_var(injectm).finished) {
317                 int offset = msr_to_offset(msr);
318                 if (offset >= 0)
319                         *(u64 *)((char *)&__get_cpu_var(injectm) + offset) = v;
320                 return;
321         }
322         wrmsrl(msr, v);
323 }
324 
325 /*
326  * Simple lockless ring to communicate PFNs from the exception handler with the
327  * process context work function. This is vastly simplified because there's
328  * only a single reader and a single writer.
329  */
330 #define MCE_RING_SIZE 16        /* we use one entry less */
331 
332 struct mce_ring {
333         unsigned short start;
334         unsigned short end;
335         unsigned long ring[MCE_RING_SIZE];
336 };
337 static DEFINE_PER_CPU(struct mce_ring, mce_ring);
338 
339 /* Runs with CPU affinity in workqueue */
340 static int mce_ring_empty(void)
341 {
342         struct mce_ring *r = &__get_cpu_var(mce_ring);
343 
344         return r->start == r->end;
345 }
346 
347 static int mce_ring_get(unsigned long *pfn)
348 {
349         struct mce_ring *r;
350         int ret = 0;
351 
352         *pfn = 0;
353         get_cpu();
354         r = &__get_cpu_var(mce_ring);
355         if (r->start == r->end)
356                 goto out;
357         *pfn = r->ring[r->start];
358         r->start = (r->start + 1) % MCE_RING_SIZE;
359         ret = 1;
360 out:
361         put_cpu();
362         return ret;
363 }
364 
365 /* Always runs in MCE context with preempt off */
366 static int mce_ring_add(unsigned long pfn)
367 {
368         struct mce_ring *r = &__get_cpu_var(mce_ring);
369         unsigned next;
370 
371         next = (r->end + 1) % MCE_RING_SIZE;
372         if (next == r->start)
373                 return -1;
374         r->ring[r->end] = pfn;
375         wmb();
376         r->end = next;
377         return 0;
378 }
379 
380 int mce_available(struct cpuinfo_x86 *c)
381 {
382         if (mce_disabled)
383                 return 0;
384         return cpu_has(c, X86_FEATURE_MCE) && cpu_has(c, X86_FEATURE_MCA);
385 }
386 
387 static void mce_schedule_work(void)
388 {
389         if (!mce_ring_empty()) {
390                 struct work_struct *work = &__get_cpu_var(mce_work);
391                 if (!work_pending(work))
392                         schedule_work(work);
393         }
394 }
395 
396 /*
397  * Get the address of the instruction at the time of the machine check
398  * error.
399  */
400 static inline void mce_get_rip(struct mce *m, struct pt_regs *regs)
401 {
402 
403         if (regs && (m->mcgstatus & (MCG_STATUS_RIPV|MCG_STATUS_EIPV))) {
404                 m->ip = regs->ip;
405                 m->cs = regs->cs;
406         } else {
407                 m->ip = 0;
408                 m->cs = 0;
409         }
410         if (rip_msr)
411                 m->ip = mce_rdmsrl(rip_msr);
412 }
413 
414 #ifdef CONFIG_X86_LOCAL_APIC 
415 /*
416  * Called after interrupts have been reenabled again
417  * when a MCE happened during an interrupts off region
418  * in the kernel.
419  */
420 asmlinkage void smp_mce_self_interrupt(struct pt_regs *regs)
421 {
422         ack_APIC_irq();
423         exit_idle();
424         irq_enter();
425         mce_notify_irq();
426         mce_schedule_work();
427         irq_exit();
428 }
429 #endif
430 
431 static void mce_report_event(struct pt_regs *regs)
432 {
433         if (regs->flags & (X86_VM_MASK|X86_EFLAGS_IF)) {
434                 mce_notify_irq();
435                 /*
436                  * Triggering the work queue here is just an insurance
437                  * policy in case the syscall exit notify handler
438                  * doesn't run soon enough or ends up running on the
439                  * wrong CPU (can happen when audit sleeps)
440                  */
441                 mce_schedule_work();
442                 return;
443         }
444 
445 #ifdef CONFIG_X86_LOCAL_APIC
446         /*
447          * Without APIC do not notify. The event will be picked
448          * up eventually.
449          */
450         if (!cpu_has_apic)
451                 return;
452 
453         /*
454          * When interrupts are disabled we cannot use
455          * kernel services safely. Trigger an self interrupt
456          * through the APIC to instead do the notification
457          * after interrupts are reenabled again.
458          */
459         apic->send_IPI_self(MCE_SELF_VECTOR);
460 
461         /*
462          * Wait for idle afterwards again so that we don't leave the
463          * APIC in a non idle state because the normal APIC writes
464          * cannot exclude us.
465          */
466         apic_wait_icr_idle();
467 #endif
468 }
469 
470 DEFINE_PER_CPU(unsigned, mce_poll_count);
471 
472 /*
473  * Poll for corrected events or events that happened before reset.
474  * Those are just logged through /dev/mcelog.
475  *
476  * This is executed in standard interrupt context.
477  *
478  * Note: spec recommends to panic for fatal unsignalled
479  * errors here. However this would be quite problematic --
480  * we would need to reimplement the Monarch handling and
481  * it would mess up the exclusion between exception handler
482  * and poll hander -- * so we skip this for now.
483  * These cases should not happen anyways, or only when the CPU
484  * is already totally * confused. In this case it's likely it will
485  * not fully execute the machine check handler either.
486  */
487 void machine_check_poll(enum mcp_flags flags, mce_banks_t *b)
488 {
489         struct mce m;
490         int i;
491 
492         __get_cpu_var(mce_poll_count)++;
493 
494         mce_setup(&m);
495 
496         m.mcgstatus = mce_rdmsrl(MSR_IA32_MCG_STATUS);
497         for (i = 0; i < banks; i++) {
498                 if (!bank[i] || !test_bit(i, *b))
499                         continue;
500 
501                 m.misc = 0;
502                 m.addr = 0;
503                 m.bank = i;
504                 m.tsc = 0;
505 
506                 barrier();
507                 m.status = mce_rdmsrl(MSR_IA32_MC0_STATUS + i*4);
508                 if (!(m.status & MCI_STATUS_VAL))
509                         continue;
510 
511                 /*
512                  * Uncorrected or signalled events are handled by the exception
513                  * handler when it is enabled, so don't process those here.
514                  *
515                  * TBD do the same check for MCI_STATUS_EN here?
516                  */
517                 if (!(flags & MCP_UC) &&
518                     (m.status & (mce_ser ? MCI_STATUS_S : MCI_STATUS_UC)))
519                         continue;
520 
521                 if (m.status & MCI_STATUS_MISCV)
522                         m.misc = mce_rdmsrl(MSR_IA32_MC0_MISC + i*4);
523                 if (m.status & MCI_STATUS_ADDRV)
524                         m.addr = mce_rdmsrl(MSR_IA32_MC0_ADDR + i*4);
525 
526                 if (!(flags & MCP_TIMESTAMP))
527                         m.tsc = 0;
528                 /*
529                  * Don't get the IP here because it's unlikely to
530                  * have anything to do with the actual error location.
531                  */
532                 if (!(flags & MCP_DONTLOG) && !mce_dont_log_ce) {
533                         mce_log(&m);
534                         add_taint(TAINT_MACHINE_CHECK);
535                 }
536 
537                 /*
538                  * Clear state for this bank.
539                  */
540                 mce_wrmsrl(MSR_IA32_MC0_STATUS+4*i, 0);
541         }
542 
543         /*
544          * Don't clear MCG_STATUS here because it's only defined for
545          * exceptions.
546          */
547 
548         sync_core();
549 }
550 EXPORT_SYMBOL_GPL(machine_check_poll);
551 
552 /*
553  * Do a quick check if any of the events requires a panic.
554  * This decides if we keep the events around or clear them.
555  */
556 static int mce_no_way_out(struct mce *m, char **msg)
557 {
558         int i;
559 
560         for (i = 0; i < banks; i++) {
561                 m->status = mce_rdmsrl(MSR_IA32_MC0_STATUS + i*4);
562                 if (mce_severity(m, tolerant, msg) >= MCE_PANIC_SEVERITY)
563                         return 1;
564         }
565         return 0;
566 }
567 
568 /*
569  * Variable to establish order between CPUs while scanning.
570  * Each CPU spins initially until executing is equal its number.
571  */
572 static atomic_t mce_executing;
573 
574 /*
575  * Defines order of CPUs on entry. First CPU becomes Monarch.
576  */
577 static atomic_t mce_callin;
578 
579 /*
580  * Check if a timeout waiting for other CPUs happened.
581  */
582 static int mce_timed_out(u64 *t)
583 {
584         /*
585          * The others already did panic for some reason.
586          * Bail out like in a timeout.
587          * rmb() to tell the compiler that system_state
588          * might have been modified by someone else.
589          */
590         rmb();
591         if (atomic_read(&mce_paniced))
592                 wait_for_panic();
593         if (!monarch_timeout)
594                 goto out;
595         if ((s64)*t < SPINUNIT) {
596                 /* CHECKME: Make panic default for 1 too? */
597                 if (tolerant < 1)
598                         mce_panic("Timeout synchronizing machine check over CPUs",
599                                   NULL, NULL);
600                 cpu_missing = 1;
601                 return 1;
602         }
603         *t -= SPINUNIT;
604 out:
605         touch_nmi_watchdog();
606         return 0;
607 }
608 
609 /*
610  * The Monarch's reign.  The Monarch is the CPU who entered
611  * the machine check handler first. It waits for the others to
612  * raise the exception too and then grades them. When any
613  * error is fatal panic. Only then let the others continue.
614  *
615  * The other CPUs entering the MCE handler will be controlled by the
616  * Monarch. They are called Subjects.
617  *
618  * This way we prevent any potential data corruption in a unrecoverable case
619  * and also makes sure always all CPU's errors are examined.
620  *
621  * Also this detects the case of an machine check event coming from outer
622  * space (not detected by any CPUs) In this case some external agent wants
623  * us to shut down, so panic too.
624  *
625  * The other CPUs might still decide to panic if the handler happens
626  * in a unrecoverable place, but in this case the system is in a semi-stable
627  * state and won't corrupt anything by itself. It's ok to let the others
628  * continue for a bit first.
629  *
630  * All the spin loops have timeouts; when a timeout happens a CPU
631  * typically elects itself to be Monarch.
632  */
633 static void mce_reign(void)
634 {
635         int cpu;
636         struct mce *m = NULL;
637         int global_worst = 0;
638         char *msg = NULL;
639         char *nmsg = NULL;
640 
641         /*
642          * This CPU is the Monarch and the other CPUs have run
643          * through their handlers.
644          * Grade the severity of the errors of all the CPUs.
645          */
646         for_each_possible_cpu(cpu) {
647                 int severity = mce_severity(&per_cpu(mces_seen, cpu), tolerant,
648                                             &nmsg);
649                 if (severity > global_worst) {
650                         msg = nmsg;
651                         global_worst = severity;
652                         m = &per_cpu(mces_seen, cpu);
653                 }
654         }
655 
656         /*
657          * Cannot recover? Panic here then.
658          * This dumps all the mces in the log buffer and stops the
659          * other CPUs.
660          */
661         if (m && global_worst >= MCE_PANIC_SEVERITY && tolerant < 3)
662                 mce_panic("Fatal Machine check", m, msg);
663 
664         /*
665          * For UC somewhere we let the CPU who detects it handle it.
666          * Also must let continue the others, otherwise the handling
667          * CPU could deadlock on a lock.
668          */
669 
670         /*
671          * No machine check event found. Must be some external
672          * source or one CPU is hung. Panic.
673          */
674         if (!m && tolerant < 3)
675                 mce_panic("Machine check from unknown source", NULL, NULL);
676 
677         /*
678          * Now clear all the mces_seen so that they don't reappear on
679          * the next mce.
680          */
681         for_each_possible_cpu(cpu)
682                 memset(&per_cpu(mces_seen, cpu), 0, sizeof(struct mce));
683 }
684 
685 static atomic_t global_nwo;
686 
687 /*
688  * Start of Monarch synchronization. This waits until all CPUs have
689  * entered the exception handler and then determines if any of them
690  * saw a fatal event that requires panic. Then it executes them
691  * in the entry order.
692  * TBD double check parallel CPU hotunplug
693  */
694 static int mce_start(int *no_way_out)
695 {
696         int order;
697         int cpus = num_online_cpus();
698         u64 timeout = (u64)monarch_timeout * NSEC_PER_USEC;
699 
700         if (!timeout)
701                 return -1;
702 
703         atomic_add(*no_way_out, &global_nwo);
704         /*
705          * global_nwo should be updated before mce_callin
706          */
707         smp_wmb();
708         order = atomic_add_return(1, &mce_callin);
709 
710         /*
711          * Wait for everyone.
712          */
713         while (atomic_read(&mce_callin) != cpus) {
714                 if (mce_timed_out(&timeout)) {
715                         atomic_set(&global_nwo, 0);
716                         return -1;
717                 }
718                 ndelay(SPINUNIT);
719         }
720 
721         /*
722          * mce_callin should be read before global_nwo
723          */
724         smp_rmb();
725 
726         if (order == 1) {
727                 /*
728                  * Monarch: Starts executing now, the others wait.
729                  */
730                 atomic_set(&mce_executing, 1);
731         } else {
732                 /*
733                  * Subject: Now start the scanning loop one by one in
734                  * the original callin order.
735                  * This way when there are any shared banks it will be
736                  * only seen by one CPU before cleared, avoiding duplicates.
737                  */
738                 while (atomic_read(&mce_executing) < order) {
739                         if (mce_timed_out(&timeout)) {
740                                 atomic_set(&global_nwo, 0);
741                                 return -1;
742                         }
743                         ndelay(SPINUNIT);
744                 }
745         }
746 
747         /*
748          * Cache the global no_way_out state.
749          */
750         *no_way_out = atomic_read(&global_nwo);
751 
752         return order;
753 }
754 
755 /*
756  * Synchronize between CPUs after main scanning loop.
757  * This invokes the bulk of the Monarch processing.
758  */
759 static int mce_end(int order)
760 {
761         int ret = -1;
762         u64 timeout = (u64)monarch_timeout * NSEC_PER_USEC;
763 
764         if (!timeout)
765                 goto reset;
766         if (order < 0)
767                 goto reset;
768 
769         /*
770          * Allow others to run.
771          */
772         atomic_inc(&mce_executing);
773 
774         if (order == 1) {
775                 /* CHECKME: Can this race with a parallel hotplug? */
776                 int cpus = num_online_cpus();
777 
778                 /*
779                  * Monarch: Wait for everyone to go through their scanning
780                  * loops.
781                  */
782                 while (atomic_read(&mce_executing) <= cpus) {
783                         if (mce_timed_out(&timeout))
784                                 goto reset;
785                         ndelay(SPINUNIT);
786                 }
787 
788                 mce_reign();
789                 barrier();
790                 ret = 0;
791         } else {
792                 /*
793                  * Subject: Wait for Monarch to finish.
794                  */
795                 while (atomic_read(&mce_executing) != 0) {
796                         if (mce_timed_out(&timeout))
797                                 goto reset;
798                         ndelay(SPINUNIT);
799                 }
800 
801                 /*
802                  * Don't reset anything. That's done by the Monarch.
803                  */
804                 return 0;
805         }
806 
807         /*
808          * Reset all global state.
809          */
810 reset:
811         atomic_set(&global_nwo, 0);
812         atomic_set(&mce_callin, 0);
813         barrier();
814 
815         /*
816          * Let others run again.
817          */
818         atomic_set(&mce_executing, 0);
819         return ret;
820 }
821 
822 /*
823  * Check if the address reported by the CPU is in a format we can parse.
824  * It would be possible to add code for most other cases, but all would
825  * be somewhat complicated (e.g. segment offset would require an instruction
826  * parser). So only support physical addresses upto page granuality for now.
827  */
828 static int mce_usable_address(struct mce *m)
829 {
830         if (!(m->status & MCI_STATUS_MISCV) || !(m->status & MCI_STATUS_ADDRV))
831                 return 0;
832         if ((m->misc & 0x3f) > PAGE_SHIFT)
833                 return 0;
834         if (((m->misc >> 6) & 7) != MCM_ADDR_PHYS)
835                 return 0;
836         return 1;
837 }
838 
839 static void mce_clear_state(unsigned long *toclear)
840 {
841         int i;
842 
843         for (i = 0; i < banks; i++) {
844                 if (test_bit(i, toclear))
845                         mce_wrmsrl(MSR_IA32_MC0_STATUS+4*i, 0);
846         }
847 }
848 
849 /*
850  * The actual machine check handler. This only handles real
851  * exceptions when something got corrupted coming in through int 18.
852  *
853  * This is executed in NMI context not subject to normal locking rules. This
854  * implies that most kernel services cannot be safely used. Don't even
855  * think about putting a printk in there!
856  *
857  * On Intel systems this is entered on all CPUs in parallel through
858  * MCE broadcast. However some CPUs might be broken beyond repair,
859  * so be always careful when synchronizing with others.
860  */
861 void do_machine_check(struct pt_regs *regs, long error_code)
862 {
863         struct mce m, *final;
864         int i;
865         int worst = 0;
866         int severity;
867         /*
868          * Establish sequential order between the CPUs entering the machine
869          * check handler.
870          */
871         int order;
872         /*
873          * If no_way_out gets set, there is no safe way to recover from this
874          * MCE.  If tolerant is cranked up, we'll try anyway.
875          */
876         int no_way_out = 0;
877         /*
878          * If kill_it gets set, there might be a way to recover from this
879          * error.
880          */
881         int kill_it = 0;
882         DECLARE_BITMAP(toclear, MAX_NR_BANKS);
883         char *msg = "Unknown";
884 
885         atomic_inc(&mce_entry);
886 
887         __get_cpu_var(mce_exception_count)++;
888 
889         if (notify_die(DIE_NMI, "machine check", regs, error_code,
890                            18, SIGKILL) == NOTIFY_STOP)
891                 goto out;
892         if (!banks)
893                 goto out;
894 
895         mce_setup(&m);
896 
897         m.mcgstatus = mce_rdmsrl(MSR_IA32_MCG_STATUS);
898         no_way_out = mce_no_way_out(&m, &msg);
899 
900         final = &__get_cpu_var(mces_seen);
901         *final = m;
902 
903         barrier();
904 
905         /*
906          * When no restart IP must always kill or panic.
907          */
908         if (!(m.mcgstatus & MCG_STATUS_RIPV))
909                 kill_it = 1;
910 
911         /*
912          * Go through all the banks in exclusion of the other CPUs.
913          * This way we don't report duplicated events on shared banks
914          * because the first one to see it will clear it.
915          */
916         order = mce_start(&no_way_out);
917         for (i = 0; i < banks; i++) {
918                 __clear_bit(i, toclear);
919                 if (!bank[i])
920                         continue;
921 
922                 m.misc = 0;
923                 m.addr = 0;
924                 m.bank = i;
925 
926                 m.status = mce_rdmsrl(MSR_IA32_MC0_STATUS + i*4);
927                 if ((m.status & MCI_STATUS_VAL) == 0)
928                         continue;
929 
930                 /*
931                  * Non uncorrected or non signaled errors are handled by
932                  * machine_check_poll. Leave them alone, unless this panics.
933                  */
934                 if (!(m.status & (mce_ser ? MCI_STATUS_S : MCI_STATUS_UC)) &&
935                         !no_way_out)
936                         continue;
937 
938                 /*
939                  * Set taint even when machine check was not enabled.
940                  */
941                 add_taint(TAINT_MACHINE_CHECK);
942 
943                 severity = mce_severity(&m, tolerant, NULL);
944 
945                 /*
946                  * When machine check was for corrected handler don't touch,
947                  * unless we're panicing.
948                  */
949                 if (severity == MCE_KEEP_SEVERITY && !no_way_out)
950                         continue;
951                 __set_bit(i, toclear);
952                 if (severity == MCE_NO_SEVERITY) {
953                         /*
954                          * Machine check event was not enabled. Clear, but
955                          * ignore.
956                          */
957                         continue;
958                 }
959 
960                 /*
961                  * Kill on action required.
962                  */
963                 if (severity == MCE_AR_SEVERITY)
964                         kill_it = 1;
965 
966                 if (m.status & MCI_STATUS_MISCV)
967                         m.misc = mce_rdmsrl(MSR_IA32_MC0_MISC + i*4);
968                 if (m.status & MCI_STATUS_ADDRV)
969                         m.addr = mce_rdmsrl(MSR_IA32_MC0_ADDR + i*4);
970 
971                 /*
972                  * Action optional error. Queue address for later processing.
973                  * When the ring overflows we just ignore the AO error.
974                  * RED-PEN add some logging mechanism when
975                  * usable_address or mce_add_ring fails.
976                  * RED-PEN don't ignore overflow for tolerant == 0
977                  */
978                 if (severity == MCE_AO_SEVERITY && mce_usable_address(&m))
979                         mce_ring_add(m.addr >> PAGE_SHIFT);
980 
981                 mce_get_rip(&m, regs);
982                 mce_log(&m);
983 
984                 if (severity > worst) {
985                         *final = m;
986                         worst = severity;
987                 }
988         }
989 
990         if (!no_way_out)
991                 mce_clear_state(toclear);
992 
993         /*
994          * Do most of the synchronization with other CPUs.
995          * When there's any problem use only local no_way_out state.
996          */
997         if (mce_end(order) < 0)
998                 no_way_out = worst >= MCE_PANIC_SEVERITY;
999 
1000         /*
1001          * If we have decided that we just CAN'T continue, and the user
1002          * has not set tolerant to an insane level, give up and die.
1003          *
1004          * This is mainly used in the case when the system doesn't
1005          * support MCE broadcasting or it has been disabled.
1006          */
1007         if (no_way_out && tolerant < 3)
1008                 mce_panic("Fatal machine check on current CPU", final, msg);
1009 
1010         /*
1011          * If the error seems to be unrecoverable, something should be
1012          * done.  Try to kill as little as possible.  If we can kill just
1013          * one task, do that.  If the user has set the tolerance very
1014          * high, don't try to do anything at all.
1015          */
1016 
1017         if (kill_it && tolerant < 3)
1018                 force_sig(SIGBUS, current);
1019 
1020         /* notify userspace ASAP */
1021         set_thread_flag(TIF_MCE_NOTIFY);
1022 
1023         if (worst > 0)
1024                 mce_report_event(regs);
1025         mce_wrmsrl(MSR_IA32_MCG_STATUS, 0);
1026 out:
1027         atomic_dec(&mce_entry);
1028         sync_core();
1029 }
1030 EXPORT_SYMBOL_GPL(do_machine_check);
1031 
1032 /* dummy to break dependency. actual code is in mm/memory-failure.c */
1033 void __attribute__((weak)) memory_failure(unsigned long pfn, int vector)
1034 {
1035         printk(KERN_ERR "Action optional memory failure at %lx ignored\n", pfn);
1036 }
1037 
1038 /*
1039  * Called after mce notification in process context. This code
1040  * is allowed to sleep. Call the high level VM handler to process
1041  * any corrupted pages.
1042  * Assume that the work queue code only calls this one at a time
1043  * per CPU.
1044  * Note we don't disable preemption, so this code might run on the wrong
1045  * CPU. In this case the event is picked up by the scheduled work queue.
1046  * This is merely a fast path to expedite processing in some common
1047  * cases.
1048  */
1049 void mce_notify_process(void)
1050 {
1051         unsigned long pfn;
1052         mce_notify_irq();
1053         while (mce_ring_get(&pfn))
1054                 memory_failure(pfn, MCE_VECTOR);
1055 }
1056 
1057 static void mce_process_work(struct work_struct *dummy)
1058 {
1059         mce_notify_process();
1060 }
1061 
1062 #ifdef CONFIG_X86_MCE_INTEL
1063 /***
1064  * mce_log_therm_throt_event - Logs the thermal throttling event to mcelog
1065  * @cpu: The CPU on which the event occurred.
1066  * @status: Event status information
1067  *
1068  * This function should be called by the thermal interrupt after the
1069  * event has been processed and the decision was made to log the event
1070  * further.
1071  *
1072  * The status parameter will be saved to the 'status' field of 'struct mce'
1073  * and historically has been the register value of the
1074  * MSR_IA32_THERMAL_STATUS (Intel) msr.
1075  */
1076 void mce_log_therm_throt_event(__u64 status)
1077 {
1078         struct mce m;
1079 
1080         mce_setup(&m);
1081         m.bank = MCE_THERMAL_BANK;
1082         m.status = status;
1083         mce_log(&m);
1084 }
1085 #endif /* CONFIG_X86_MCE_INTEL */
1086 
1087 /*
1088  * Periodic polling timer for "silent" machine check errors.  If the
1089  * poller finds an MCE, poll 2x faster.  When the poller finds no more
1090  * errors, poll 2x slower (up to check_interval seconds).
1091  */
1092 static int check_interval = 5 * 60; /* 5 minutes */
1093 
1094 static DEFINE_PER_CPU(int, next_interval); /* in jiffies */
1095 static DEFINE_PER_CPU(struct timer_list, mce_timer);
1096 
1097 static void mcheck_timer(unsigned long data)
1098 {
1099         struct timer_list *t = &per_cpu(mce_timer, data);
1100         int *n;
1101 
1102         WARN_ON(smp_processor_id() != data);
1103 
1104         if (mce_available(&current_cpu_data)) {
1105                 machine_check_poll(MCP_TIMESTAMP,
1106                                 &__get_cpu_var(mce_poll_banks));
1107         }
1108 
1109         /*
1110          * Alert userspace if needed.  If we logged an MCE, reduce the
1111          * polling interval, otherwise increase the polling interval.
1112          */
1113         n = &__get_cpu_var(next_interval);
1114         if (mce_notify_irq())
1115                 *n = max(*n/2, HZ/100);
1116         else
1117                 *n = min(*n*2, (int)round_jiffies_relative(check_interval*HZ));
1118 
1119         t->expires = jiffies + *n;
1120         add_timer_on(t, smp_processor_id());
1121 }
1122 
1123 static void mce_do_trigger(struct work_struct *work)
1124 {
1125         call_usermodehelper(mce_helper, mce_helper_argv, NULL, UMH_NO_WAIT);
1126 }
1127 
1128 static DECLARE_WORK(mce_trigger_work, mce_do_trigger);
1129 
1130 /*
1131  * Notify the user(s) about new machine check events.
1132  * Can be called from interrupt context, but not from machine check/NMI
1133  * context.
1134  */
1135 int mce_notify_irq(void)
1136 {
1137         /* Not more than two messages every minute */
1138         static DEFINE_RATELIMIT_STATE(ratelimit, 60*HZ, 2);
1139 
1140         clear_thread_flag(TIF_MCE_NOTIFY);
1141 
1142         if (test_and_clear_bit(0, &mce_need_notify)) {
1143                 wake_up_interruptible(&mce_wait);
1144 
1145                 /*
1146                  * There is no risk of missing notifications because
1147                  * work_pending is always cleared before the function is
1148                  * executed.
1149                  */
1150                 if (mce_helper[0] && !work_pending(&mce_trigger_work))
1151                         schedule_work(&mce_trigger_work);
1152 
1153                 if (__ratelimit(&ratelimit))
1154                         printk(KERN_INFO "Machine check events logged\n");
1155 
1156                 return 1;
1157         }
1158         return 0;
1159 }
1160 EXPORT_SYMBOL_GPL(mce_notify_irq);
1161 
1162 /*
1163  * Initialize Machine Checks for a CPU.
1164  */
1165 static int mce_cap_init(void)
1166 {
1167         unsigned b;
1168         u64 cap;
1169 
1170         rdmsrl(MSR_IA32_MCG_CAP, cap);
1171 
1172         b = cap & MCG_BANKCNT_MASK;
1173         printk(KERN_INFO "mce: CPU supports %d MCE banks\n", b);
1174 
1175         if (b > MAX_NR_BANKS) {
1176                 printk(KERN_WARNING
1177                        "MCE: Using only %u machine check banks out of %u\n",
1178                         MAX_NR_BANKS, b);
1179                 b = MAX_NR_BANKS;
1180         }
1181 
1182         /* Don't support asymmetric configurations today */
1183         WARN_ON(banks != 0 && b != banks);
1184         banks = b;
1185         if (!bank) {
1186                 bank = kmalloc(banks * sizeof(u64), GFP_KERNEL);
1187                 if (!bank)
1188                         return -ENOMEM;
1189                 memset(bank, 0xff, banks * sizeof(u64));
1190         }
1191 
1192         /* Use accurate RIP reporting if available. */
1193         if ((cap & MCG_EXT_P) && MCG_EXT_CNT(cap) >= 9)
1194                 rip_msr = MSR_IA32_MCG_EIP;
1195 
1196         if (cap & MCG_SER_P)
1197                 mce_ser = 1;
1198 
1199         return 0;
1200 }
1201 
1202 static void mce_init(void)
1203 {
1204         mce_banks_t all_banks;
1205         u64 cap;
1206         int i;
1207 
1208         /*
1209          * Log the machine checks left over from the previous reset.
1210          */
1211         bitmap_fill(all_banks, MAX_NR_BANKS);
1212         machine_check_poll(MCP_UC|(!mce_bootlog ? MCP_DONTLOG : 0), &all_banks);
1213 
1214         set_in_cr4(X86_CR4_MCE);
1215 
1216         rdmsrl(MSR_IA32_MCG_CAP, cap);
1217         if (cap & MCG_CTL_P)
1218                 wrmsr(MSR_IA32_MCG_CTL, 0xffffffff, 0xffffffff);
1219 
1220         for (i = 0; i < banks; i++) {
1221                 if (skip_bank_init(i))
1222                         continue;
1223                 wrmsrl(MSR_IA32_MC0_CTL+4*i, bank[i]);
1224                 wrmsrl(MSR_IA32_MC0_STATUS+4*i, 0);
1225         }
1226 }
1227 
1228 /* Add per CPU specific workarounds here */
1229 static int mce_cpu_quirks(struct cpuinfo_x86 *c)
1230 {
1231         if (c->x86_vendor == X86_VENDOR_UNKNOWN) {
1232                 pr_info("MCE: unknown CPU type - not enabling MCE support.\n");
1233                 return -EOPNOTSUPP;
1234         }
1235 
1236         /* This should be disabled by the BIOS, but isn't always */
1237         if (c->x86_vendor == X86_VENDOR_AMD) {
1238                 if (c->x86 == 15 && banks > 4) {
1239                         /*
1240                          * disable GART TBL walk error reporting, which
1241                          * trips off incorrectly with the IOMMU & 3ware
1242                          * & Cerberus:
1243                          */
1244                         clear_bit(10, (unsigned long *)&bank[4]);
1245                 }
1246                 if (c->x86 <= 17 && mce_bootlog < 0) {
1247                         /*
1248                          * Lots of broken BIOS around that don't clear them
1249                          * by default and leave crap in there. Don't log:
1250                          */
1251                         mce_bootlog = 0;
1252                 }
1253                 /*
1254                  * Various K7s with broken bank 0 around. Always disable
1255                  * by default.
1256                  */
1257                  if (c->x86 == 6 && banks > 0)
1258                         bank[0] = 0;
1259         }
1260 
1261         if (c->x86_vendor == X86_VENDOR_INTEL) {
1262                 /*
1263                  * SDM documents that on family 6 bank 0 should not be written
1264                  * because it aliases to another special BIOS controlled
1265                  * register.
1266                  * But it's not aliased anymore on model 0x1a+
1267                  * Don't ignore bank 0 completely because there could be a
1268                  * valid event later, merely don't write CTL0.
1269                  */
1270 
1271                 if (c->x86 == 6 && c->x86_model < 0x1A)
1272                         __set_bit(0, &dont_init_banks);
1273 
1274                 /*
1275                  * All newer Intel systems support MCE broadcasting. Enable
1276                  * synchronization with a one second timeout.
1277                  */
1278                 if ((c->x86 > 6 || (c->x86 == 6 && c->x86_model >= 0xe)) &&
1279                         monarch_timeout < 0)
1280                         monarch_timeout = USEC_PER_SEC;
1281 
1282                 /*
1283                  * There are also broken BIOSes on some Pentium M and
1284                  * earlier systems:
1285                  */
1286                 if (c->x86 == 6 && c->x86_model <= 13 && mce_bootlog < 0)
1287                         mce_bootlog = 0;
1288         }
1289         if (monarch_timeout < 0)
1290                 monarch_timeout = 0;
1291         if (mce_bootlog != 0)
1292                 mce_panic_timeout = 30;
1293 
1294         return 0;
1295 }
1296 
1297 static void __cpuinit mce_ancient_init(struct cpuinfo_x86 *c)
1298 {
1299         if (c->x86 != 5)
1300                 return;
1301         switch (c->x86_vendor) {
1302         case X86_VENDOR_INTEL:
1303                 intel_p5_mcheck_init(c);
1304                 break;
1305         case X86_VENDOR_CENTAUR:
1306                 winchip_mcheck_init(c);
1307                 break;
1308         }
1309 }
1310 
1311 static void mce_cpu_features(struct cpuinfo_x86 *c)
1312 {
1313         switch (c->x86_vendor) {
1314         case X86_VENDOR_INTEL:
1315                 mce_intel_feature_init(c);
1316                 break;
1317         case X86_VENDOR_AMD:
1318                 mce_amd_feature_init(c);
1319                 break;
1320         default:
1321                 break;
1322         }
1323 }
1324 
1325 static void mce_init_timer(void)
1326 {
1327         struct timer_list *t = &__get_cpu_var(mce_timer);
1328         int *n = &__get_cpu_var(next_interval);
1329 
1330         setup_timer(t, mcheck_timer, smp_processor_id());
1331 
1332         if (mce_ignore_ce)
1333                 return;
1334 
1335         *n = check_interval * HZ;
1336         if (!*n)
1337                 return;
1338         t->expires = round_jiffies(jiffies + *n);
1339         add_timer_on(t, smp_processor_id());
1340 }
1341 
1342 /*
1343  * Called for each booted CPU to set up machine checks.
1344  * Must be called with preempt off:
1345  */
1346 void __cpuinit mcheck_init(struct cpuinfo_x86 *c)
1347 {
1348         if (mce_disabled)
1349                 return;
1350 
1351         mce_ancient_init(c);
1352 
1353         if (!mce_available(c))
1354                 return;
1355 
1356         if (mce_cap_init() < 0 || mce_cpu_quirks(c) < 0) {
1357                 mce_disabled = 1;
1358                 return;
1359         }
1360 
1361         machine_check_vector = do_machine_check;
1362 
1363         mce_init();
1364         mce_cpu_features(c);
1365         mce_init_timer();
1366         INIT_WORK(&__get_cpu_var(mce_work), mce_process_work);
1367 }
1368 
1369 /*
1370  * Character device to read and clear the MCE log.
1371  */
1372 
1373 static DEFINE_SPINLOCK(mce_state_lock);
1374 static int              open_count;             /* #times opened */
1375 static int              open_exclu;             /* already open exclusive? */
1376 
1377 static int mce_open(struct inode *inode, struct file *file)
1378 {
1379         spin_lock(&mce_state_lock);
1380 
1381         if (open_exclu || (open_count && (file->f_flags & O_EXCL))) {
1382                 spin_unlock(&mce_state_lock);
1383 
1384                 return -EBUSY;
1385         }
1386 
1387         if (file->f_flags & O_EXCL)
1388                 open_exclu = 1;
1389         open_count++;
1390 
1391         spin_unlock(&mce_state_lock);
1392 
1393         return nonseekable_open(inode, file);
1394 }
1395 
1396 static int mce_release(struct inode *inode, struct file *file)
1397 {
1398         spin_lock(&mce_state_lock);
1399 
1400         open_count--;
1401         open_exclu = 0;
1402 
1403         spin_unlock(&mce_state_lock);
1404 
1405         return 0;
1406 }
1407 
1408 static void collect_tscs(void *data)
1409 {
1410         unsigned long *cpu_tsc = (unsigned long *)data;
1411 
1412         rdtscll(cpu_tsc[smp_processor_id()]);
1413 }
1414 
1415 static DEFINE_MUTEX(mce_read_mutex);
1416 
1417 static ssize_t mce_read(struct file *filp, char __user *ubuf, size_t usize,
1418                         loff_t *off)
1419 {
1420         char __user *buf = ubuf;
1421         unsigned long *cpu_tsc;
1422         unsigned prev, next;
1423         int i, err;
1424 
1425         cpu_tsc = kmalloc(nr_cpu_ids * sizeof(long), GFP_KERNEL);
1426         if (!cpu_tsc)
1427                 return -ENOMEM;
1428 
1429         mutex_lock(&mce_read_mutex);
1430         next = rcu_dereference(mcelog.next);
1431 
1432         /* Only supports full reads right now */
1433         if (*off != 0 || usize < MCE_LOG_LEN*sizeof(struct mce)) {
1434                 mutex_unlock(&mce_read_mutex);
1435                 kfree(cpu_tsc);
1436 
1437                 return -EINVAL;
1438         }
1439 
1440         err = 0;
1441         prev = 0;
1442         do {
1443                 for (i = prev; i < next; i++) {
1444                         unsigned long start = jiffies;
1445 
1446                         while (!mcelog.entry[i].finished) {
1447                                 if (time_after_eq(jiffies, start + 2)) {
1448                                         memset(mcelog.entry + i, 0,
1449                                                sizeof(struct mce));
1450                                         goto timeout;
1451                                 }
1452                                 cpu_relax();
1453                         }
1454                         smp_rmb();
1455                         err |= copy_to_user(buf, mcelog.entry + i,
1456                                             sizeof(struct mce));
1457                         buf += sizeof(struct mce);
1458 timeout:
1459                         ;
1460                 }
1461 
1462                 memset(mcelog.entry + prev, 0,
1463                        (next - prev) * sizeof(struct mce));
1464                 prev = next;
1465                 next = cmpxchg(&mcelog.next, prev, 0);
1466         } while (next != prev);
1467 
1468         synchronize_sched();
1469 
1470         /*
1471          * Collect entries that were still getting written before the
1472          * synchronize.
1473          */
1474         on_each_cpu(collect_tscs, cpu_tsc, 1);
1475 
1476         for (i = next; i < MCE_LOG_LEN; i++) {
1477                 if (mcelog.entry[i].finished &&
1478                     mcelog.entry[i].tsc < cpu_tsc[mcelog.entry[i].cpu]) {
1479                         err |= copy_to_user(buf, mcelog.entry+i,
1480                                             sizeof(struct mce));
1481                         smp_rmb();
1482                         buf += sizeof(struct mce);
1483                         memset(&mcelog.entry[i], 0, sizeof(struct mce));
1484                 }
1485         }
1486         mutex_unlock(&mce_read_mutex);
1487         kfree(cpu_tsc);
1488 
1489         return err ? -EFAULT : buf - ubuf;
1490 }
1491 
1492 static unsigned int mce_poll(struct file *file, poll_table *wait)
1493 {
1494         poll_wait(file, &mce_wait, wait);
1495         if (rcu_dereference(mcelog.next))
1496                 return POLLIN | POLLRDNORM;
1497         return 0;
1498 }
1499 
1500 static long mce_ioctl(struct file *f, unsigned int cmd, unsigned long arg)
1501 {
1502         int __user *p = (int __user *)arg;
1503 
1504         if (!capable(CAP_SYS_ADMIN))
1505                 return -EPERM;
1506 
1507         switch (cmd) {
1508         case MCE_GET_RECORD_LEN:
1509                 return put_user(sizeof(struct mce), p);
1510         case MCE_GET_LOG_LEN:
1511                 return put_user(MCE_LOG_LEN, p);
1512         case MCE_GETCLEAR_FLAGS: {
1513                 unsigned flags;
1514 
1515                 do {
1516                         flags = mcelog.flags;
1517                 } while (cmpxchg(&mcelog.flags, flags, 0) != flags);
1518 
1519                 return put_user(flags, p);
1520         }
1521         default:
1522                 return -ENOTTY;
1523         }
1524 }
1525 
1526 /* Modified in mce-inject.c, so not static or const */
1527 struct file_operations mce_chrdev_ops = {
1528         .open                   = mce_open,
1529         .release                = mce_release,
1530         .read                   = mce_read,
1531         .poll                   = mce_poll,
1532         .unlocked_ioctl         = mce_ioctl,
1533 };
1534 EXPORT_SYMBOL_GPL(mce_chrdev_ops);
1535 
1536 static struct miscdevice mce_log_device = {
1537         MISC_MCELOG_MINOR,
1538         "mcelog",
1539         &mce_chrdev_ops,
1540 };
1541 
1542 /*
1543  * mce=off Disables machine check
1544  * mce=no_cmci Disables CMCI
1545  * mce=dont_log_ce Clears corrected events silently, no log created for CEs.
1546  * mce=ignore_ce Disables polling and CMCI, corrected events are not cleared.
1547  * mce=TOLERANCELEVEL[,monarchtimeout] (number, see above)
1548  *      monarchtimeout is how long to wait for other CPUs on machine
1549  *      check, or 0 to not wait
1550  * mce=bootlog Log MCEs from before booting. Disabled by default on AMD.
1551  * mce=nobootlog Don't log MCEs from before booting.
1552  */
1553 static int __init mcheck_enable(char *str)
1554 {
1555         if (*str == 0)
1556                 enable_p5_mce();
1557         if (*str == '=')
1558                 str++;
1559         if (!strcmp(str, "off"))
1560                 mce_disabled = 1;
1561         else if (!strcmp(str, "no_cmci"))
1562                 mce_cmci_disabled = 1;
1563         else if (!strcmp(str, "dont_log_ce"))
1564                 mce_dont_log_ce = 1;
1565         else if (!strcmp(str, "ignore_ce"))
1566                 mce_ignore_ce = 1;
1567         else if (!strcmp(str, "bootlog") || !strcmp(str, "nobootlog"))
1568                 mce_bootlog = (str[0] == 'b');
1569         else if (isdigit(str[0])) {
1570                 get_option(&str, &tolerant);
1571                 if (*str == ',') {
1572                         ++str;
1573                         get_option(&str, &monarch_timeout);
1574                 }
1575         } else {
1576                 printk(KERN_INFO "mce argument %s ignored. Please use /sys\n",
1577                        str);
1578                 return 0;
1579         }
1580         return 1;
1581 }
1582 __setup("mce", mcheck_enable);
1583 
1584 /*
1585  * Sysfs support
1586  */
1587 
1588 /*
1589  * Disable machine checks on suspend and shutdown. We can't really handle
1590  * them later.
1591  */
1592 static int mce_disable(void)
1593 {
1594         int i;
1595 
1596         for (i = 0; i < banks; i++) {
1597                 if (!skip_bank_init(i))
1598                         wrmsrl(MSR_IA32_MC0_CTL + i*4, 0);
1599         }
1600         return 0;
1601 }
1602 
1603 static int mce_suspend(struct sys_device *dev, pm_message_t state)
1604 {
1605         return mce_disable();
1606 }
1607 
1608 static int mce_shutdown(struct sys_device *dev)
1609 {
1610         return mce_disable();
1611 }
1612 
1613 /*
1614  * On resume clear all MCE state. Don't want to see leftovers from the BIOS.
1615  * Only one CPU is active at this time, the others get re-added later using
1616  * CPU hotplug:
1617  */
1618 static int mce_resume(struct sys_device *dev)
1619 {
1620         mce_init();
1621         mce_cpu_features(&current_cpu_data);
1622 
1623         return 0;
1624 }
1625 
1626 static void mce_cpu_restart(void *data)
1627 {
1628         del_timer_sync(&__get_cpu_var(mce_timer));
1629         if (!mce_available(&current_cpu_data))
1630                 return;
1631         mce_init();
1632         mce_init_timer();
1633 }
1634 
1635 /* Reinit MCEs after user configuration changes */
1636 static void mce_restart(void)
1637 {
1638         on_each_cpu(mce_cpu_restart, NULL, 1);
1639 }
1640 
1641 /* Toggle features for corrected errors */
1642 static void mce_disable_ce(void *all)
1643 {
1644         if (!mce_available(&current_cpu_data))
1645                 return;
1646         if (all)
1647                 del_timer_sync(&__get_cpu_var(mce_timer));
1648         cmci_clear();
1649 }
1650 
1651 static void mce_enable_ce(void *all)
1652 {
1653         if (!mce_available(&current_cpu_data))
1654                 return;
1655         cmci_reenable();
1656         cmci_recheck();
1657         if (all)
1658                 mce_init_timer();
1659 }
1660 
1661 static struct sysdev_class mce_sysclass = {
1662         .suspend        = mce_suspend,
1663         .shutdown       = mce_shutdown,
1664         .resume         = mce_resume,
1665         .name           = "machinecheck",
1666 };
1667 
1668 DEFINE_PER_CPU(struct sys_device, mce_dev);
1669 
1670 __cpuinitdata
1671 void (*threshold_cpu_callback)(unsigned long action, unsigned int cpu);
1672 
1673 static struct sysdev_attribute *bank_attrs;
1674 
1675 static ssize_t show_bank(struct sys_device *s, struct sysdev_attribute *attr,
1676                          char *buf)
1677 {
1678         u64 b = bank[attr - bank_attrs];
1679 
1680         return sprintf(buf, "%llx\n", b);
1681 }
1682 
1683 static ssize_t set_bank(struct sys_device *s, struct sysdev_attribute *attr,
1684                         const char *buf, size_t size)
1685 {
1686         u64 new;
1687 
1688         if (strict_strtoull(buf, 0, &new) < 0)
1689                 return -EINVAL;
1690 
1691         bank[attr - bank_attrs] = new;
1692         mce_restart();
1693 
1694         return size;
1695 }
1696 
1697 static ssize_t
1698 show_trigger(struct sys_device *s, struct sysdev_attribute *attr, char *buf)
1699 {
1700         strcpy(buf, mce_helper);
1701         strcat(buf, "\n");
1702         return strlen(mce_helper) + 1;
1703 }
1704 
1705 static ssize_t set_trigger(struct sys_device *s, struct sysdev_attribute *attr,
1706                                 const char *buf, size_t siz)
1707 {
1708         char *p;
1709 
1710         strncpy(mce_helper, buf, sizeof(mce_helper));
1711         mce_helper[sizeof(mce_helper)-1] = 0;
1712         p = strchr(mce_helper, '\n');
1713 
1714         if (p)
1715                 *p = 0;
1716 
1717         return strlen(mce_helper) + !!p;
1718 }
1719 
1720 static ssize_t set_ignore_ce(struct sys_device *s,
1721                              struct sysdev_attribute *attr,
1722                              const char *buf, size_t size)
1723 {
1724         u64 new;
1725 
1726         if (strict_strtoull(buf, 0, &new) < 0)
1727                 return -EINVAL;
1728 
1729         if (mce_ignore_ce ^ !!new) {
1730                 if (new) {
1731                         /* disable ce features */
1732                         on_each_cpu(mce_disable_ce, (void *)1, 1);
1733                         mce_ignore_ce = 1;
1734                 } else {
1735                         /* enable ce features */
1736                         mce_ignore_ce = 0;
1737                         on_each_cpu(mce_enable_ce, (void *)1, 1);
1738                 }
1739         }
1740         return size;
1741 }
1742 
1743 static ssize_t set_cmci_disabled(struct sys_device *s,
1744                                  struct sysdev_attribute *attr,
1745                                  const char *buf, size_t size)
1746 {
1747         u64 new;
1748 
1749         if (strict_strtoull(buf, 0, &new) < 0)
1750                 return -EINVAL;
1751 
1752         if (mce_cmci_disabled ^ !!new) {
1753                 if (new) {
1754                         /* disable cmci */
1755                         on_each_cpu(mce_disable_ce, NULL, 1);
1756                         mce_cmci_disabled = 1;
1757                 } else {
1758                         /* enable cmci */
1759                         mce_cmci_disabled = 0;
1760                         on_each_cpu(mce_enable_ce, NULL, 1);
1761                 }
1762         }
1763         return size;
1764 }
1765 
1766 static ssize_t store_int_with_restart(struct sys_device *s,
1767                                       struct sysdev_attribute *attr,
1768                                       const char *buf, size_t size)
1769 {
1770         ssize_t ret = sysdev_store_int(s, attr, buf, size);
1771         mce_restart();
1772         return ret;
1773 }
1774 
1775 static SYSDEV_ATTR(trigger, 0644, show_trigger, set_trigger);
1776 static SYSDEV_INT_ATTR(tolerant, 0644, tolerant);
1777 static SYSDEV_INT_ATTR(monarch_timeout, 0644, monarch_timeout);
1778 static SYSDEV_INT_ATTR(dont_log_ce, 0644, mce_dont_log_ce);
1779 
1780 static struct sysdev_ext_attribute attr_check_interval = {
1781         _SYSDEV_ATTR(check_interval, 0644, sysdev_show_int,
1782                      store_int_with_restart),
1783         &check_interval
1784 };
1785 
1786 static struct sysdev_ext_attribute attr_ignore_ce = {
1787         _SYSDEV_ATTR(ignore_ce, 0644, sysdev_show_int, set_ignore_ce),
1788         &mce_ignore_ce
1789 };
1790 
1791 static struct sysdev_ext_attribute attr_cmci_disabled = {
1792         _SYSDEV_ATTR(cmci_disabled, 0644, sysdev_show_int, set_cmci_disabled),
1793         &mce_cmci_disabled
1794 };
1795 
1796 static struct sysdev_attribute *mce_attrs[] = {
1797         &attr_tolerant.attr,
1798         &attr_check_interval.attr,
1799         &attr_trigger,
1800         &attr_monarch_timeout.attr,
1801         &attr_dont_log_ce.attr,
1802         &attr_ignore_ce.attr,
1803         &attr_cmci_disabled.attr,
1804         NULL
1805 };
1806 
1807 static cpumask_var_t mce_dev_initialized;
1808 
1809 /* Per cpu sysdev init. All of the cpus still share the same ctrl bank: */
1810 static __cpuinit int mce_create_device(unsigned int cpu)
1811 {
1812         int err;
1813         int i, j;
1814 
1815         if (!mce_available(&boot_cpu_data))
1816                 return -EIO;
1817 
1818         memset(&per_cpu(mce_dev, cpu).kobj, 0, sizeof(struct kobject));
1819         per_cpu(mce_dev, cpu).id        = cpu;
1820         per_cpu(mce_dev, cpu).cls       = &mce_sysclass;
1821 
1822         err = sysdev_register(&per_cpu(mce_dev, cpu));
1823         if (err)
1824                 return err;
1825 
1826         for (i = 0; mce_attrs[i]; i++) {
1827                 err = sysdev_create_file(&per_cpu(mce_dev, cpu), mce_attrs[i]);
1828                 if (err)
1829                         goto error;
1830         }
1831         for (j = 0; j < banks; j++) {
1832                 err = sysdev_create_file(&per_cpu(mce_dev, cpu),
1833                                         &bank_attrs[j]);
1834                 if (err)
1835                         goto error2;
1836         }
1837         cpumask_set_cpu(cpu, mce_dev_initialized);
1838 
1839         return 0;
1840 error2:
1841         while (--j >= 0)
1842                 sysdev_remove_file(&per_cpu(mce_dev, cpu), &bank_attrs[j]);
1843 error:
1844         while (--i >= 0)
1845                 sysdev_remove_file(&per_cpu(mce_dev, cpu), mce_attrs[i]);
1846 
1847         sysdev_unregister(&per_cpu(mce_dev, cpu));
1848 
1849         return err;
1850 }
1851 
1852 static __cpuinit void mce_remove_device(unsigned int cpu)
1853 {
1854         int i;
1855 
1856         if (!cpumask_test_cpu(cpu, mce_dev_initialized))
1857                 return;
1858 
1859         for (i = 0; mce_attrs[i]; i++)
1860                 sysdev_remove_file(&per_cpu(mce_dev, cpu), mce_attrs[i]);
1861 
1862         for (i = 0; i < banks; i++)
1863                 sysdev_remove_file(&per_cpu(mce_dev, cpu), &bank_attrs[i]);
1864 
1865         sysdev_unregister(&per_cpu(mce_dev, cpu));
1866         cpumask_clear_cpu(cpu, mce_dev_initialized);
1867 }
1868 
1869 /* Make sure there are no machine checks on offlined CPUs. */
1870 static void mce_disable_cpu(void *h)
1871 {
1872         unsigned long action = *(unsigned long *)h;
1873         int i;
1874 
1875         if (!mce_available(&current_cpu_data))
1876                 return;
1877         if (!(action & CPU_TASKS_FROZEN))
1878                 cmci_clear();
1879         for (i = 0; i < banks; i++) {
1880                 if (!skip_bank_init(i))
1881                         wrmsrl(MSR_IA32_MC0_CTL + i*4, 0);
1882         }
1883 }
1884 
1885 static void mce_reenable_cpu(void *h)
1886 {
1887         unsigned long action = *(unsigned long *)h;
1888         int i;
1889 
1890         if (!mce_available(&current_cpu_data))
1891                 return;
1892 
1893         if (!(action & CPU_TASKS_FROZEN))
1894                 cmci_reenable();
1895         for (i = 0; i < banks; i++) {
1896                 if (!skip_bank_init(i))
1897                         wrmsrl(MSR_IA32_MC0_CTL + i*4, bank[i]);
1898         }
1899 }
1900 
1901 /* Get notified when a cpu comes on/off. Be hotplug friendly. */
1902 static int __cpuinit
1903 mce_cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
1904 {
1905         unsigned int cpu = (unsigned long)hcpu;
1906         struct timer_list *t = &per_cpu(mce_timer, cpu);
1907 
1908         switch (action) {
1909         case CPU_ONLINE:
1910         case CPU_ONLINE_FROZEN:
1911                 mce_create_device(cpu);
1912                 if (threshold_cpu_callback)
1913                         threshold_cpu_callback(action, cpu);
1914                 break;
1915         case CPU_DEAD:
1916         case CPU_DEAD_FROZEN:
1917                 if (threshold_cpu_callback)
1918                         threshold_cpu_callback(action, cpu);
1919                 mce_remove_device(cpu);
1920                 break;
1921         case CPU_DOWN_PREPARE:
1922         case CPU_DOWN_PREPARE_FROZEN:
1923                 del_timer_sync(t);
1924                 smp_call_function_single(cpu, mce_disable_cpu, &action, 1);
1925                 break;
1926         case CPU_DOWN_FAILED:
1927         case CPU_DOWN_FAILED_FROZEN:
1928                 t->expires = round_jiffies(jiffies +
1929                                                 __get_cpu_var(next_interval));
1930                 add_timer_on(t, cpu);
1931                 smp_call_function_single(cpu, mce_reenable_cpu, &action, 1);
1932                 break;
1933         case CPU_POST_DEAD:
1934                 /* intentionally ignoring frozen here */
1935                 cmci_rediscover(cpu);
1936                 break;
1937         }
1938         return NOTIFY_OK;
1939 }
1940 
1941 static struct notifier_block mce_cpu_notifier __cpuinitdata = {
1942         .notifier_call = mce_cpu_callback,
1943 };
1944 
1945 static __init int mce_init_banks(void)
1946 {
1947         int i;
1948 
1949         bank_attrs = kzalloc(sizeof(struct sysdev_attribute) * banks,
1950                                 GFP_KERNEL);
1951         if (!bank_attrs)
1952                 return -ENOMEM;
1953 
1954         for (i = 0; i < banks; i++) {
1955                 struct sysdev_attribute *a = &bank_attrs[i];
1956 
1957                 a->attr.name    = kasprintf(GFP_KERNEL, "bank%d", i);
1958                 if (!a->attr.name)
1959                         goto nomem;
1960 
1961                 a->attr.mode    = 0644;
1962                 a->show         = show_bank;
1963                 a->store        = set_bank;
1964         }
1965         return 0;
1966 
1967 nomem:
1968         while (--i >= 0)
1969                 kfree(bank_attrs[i].attr.name);
1970         kfree(bank_attrs);
1971         bank_attrs = NULL;
1972 
1973         return -ENOMEM;
1974 }
1975 
1976 static __init int mce_init_device(void)
1977 {
1978         int err;
1979         int i = 0;
1980 
1981         if (!mce_available(&boot_cpu_data))
1982                 return -EIO;
1983 
1984         zalloc_cpumask_var(&mce_dev_initialized, GFP_KERNEL);
1985 
1986         err = mce_init_banks();
1987         if (err)
1988                 return err;
1989 
1990         err = sysdev_class_register(&mce_sysclass);
1991         if (err)
1992                 return err;
1993 
1994         for_each_online_cpu(i) {
1995                 err = mce_create_device(i);
1996                 if (err)
1997                         return err;
1998         }
1999 
2000         register_hotcpu_notifier(&mce_cpu_notifier);
2001         misc_register(&mce_log_device);
2002 
2003         return err;
2004 }
2005 
2006 device_initcall(mce_init_device);
2007 
2008 #else /* CONFIG_X86_OLD_MCE: */
2009 
2010 int nr_mce_banks;
2011 EXPORT_SYMBOL_GPL(nr_mce_banks);        /* non-fatal.o */
2012 
2013 /* This has to be run for each processor */
2014 void mcheck_init(struct cpuinfo_x86 *c)
2015 {
2016         if (mce_disabled)
2017                 return;
2018 
2019         switch (c->x86_vendor) {
2020         case X86_VENDOR_AMD:
2021                 amd_mcheck_init(c);
2022                 break;
2023 
2024         case X86_VENDOR_INTEL:
2025                 if (c->x86 == 5)
2026                         intel_p5_mcheck_init(c);
2027                 if (c->x86 == 6)
2028                         intel_p6_mcheck_init(c);
2029                 if (c->x86 == 15)
2030                         intel_p4_mcheck_init(c);
2031                 break;
2032 
2033         case X86_VENDOR_CENTAUR:
2034                 if (c->x86 == 5)
2035                         winchip_mcheck_init(c);
2036                 break;
2037 
2038         default:
2039                 break;
2040         }
2041         printk(KERN_INFO "mce: CPU supports %d MCE banks\n", nr_mce_banks);
2042 }
2043 
2044 static int __init mcheck_enable(char *str)
2045 {
2046         mce_p5_enabled = 1;
2047         return 1;
2048 }
2049 __setup("mce", mcheck_enable);
2050 
2051 #endif /* CONFIG_X86_OLD_MCE */
2052 
2053 /*
2054  * Old style boot options parsing. Only for compatibility.
2055  */
2056 static int __init mcheck_disable(char *str)
2057 {
2058         mce_disabled = 1;
2059         return 1;
2060 }
2061 __setup("nomce", mcheck_disable);
2062 
  This page was automatically generated by the LXR engine.