git.oblomov.eu Git - linux-2.6/blob - arch/x86_64/kernel/nmi.c

   1 /*
   2  *  linux/arch/x86_64/nmi.c
   3  *
   4  *  NMI watchdog support on APIC systems
   5  *
   6  *  Started by Ingo Molnar <mingo@redhat.com>
   7  *
   8  *  Fixes:
   9  *  Mikael Pettersson   : AMD K7 support for local APIC NMI watchdog.
  10  *  Mikael Pettersson   : Power Management for local APIC NMI watchdog.
  11  *  Pavel Machek and
  12  *  Mikael Pettersson   : PM converted to driver model. Disable/enable API.
  13  */
  14
  15 #include <linux/mm.h>
  16 #include <linux/delay.h>
  17 #include <linux/interrupt.h>
  18 #include <linux/module.h>
  19 #include <linux/sysdev.h>
  20 #include <linux/nmi.h>
  21 #include <linux/sysctl.h>
  22 #include <linux/kprobes.h>
  23
  24 #include <asm/smp.h>
  25 #include <asm/nmi.h>
  26 #include <asm/proto.h>
  27 #include <asm/kdebug.h>
  28 #include <asm/mce.h>
  29 #include <asm/intel_arch_perfmon.h>
  30
  31 int unknown_nmi_panic;
  32 int nmi_watchdog_enabled;
  33 int panic_on_unrecovered_nmi;
  34
  35 /* perfctr_nmi_owner tracks the ownership of the perfctr registers:
  36  * evtsel_nmi_owner tracks the ownership of the event selection
  37  * - different performance counters/ event selection may be reserved for
  38  *   different subsystems this reservation system just tries to coordinate
  39  *   things a little
  40  */
  41 static DEFINE_PER_CPU(unsigned, perfctr_nmi_owner);
  42 static DEFINE_PER_CPU(unsigned, evntsel_nmi_owner[2]);
  43
  44 /* this number is calculated from Intel's MSR_P4_CRU_ESCR5 register and it's
  45  * offset from MSR_P4_BSU_ESCR0.  It will be the max for all platforms (for now)
  46  */
  47 #define NMI_MAX_COUNTER_BITS 66
  48
  49 /* nmi_active:
  50  * >0: the lapic NMI watchdog is active, but can be disabled
  51  * <0: the lapic NMI watchdog has not been set up, and cannot
  52  *     be enabled
  53  *  0: the lapic NMI watchdog is disabled, but can be enabled
  54  */
  55 atomic_t nmi_active = ATOMIC_INIT(0);           /* oprofile uses this */
  56 int panic_on_timeout;
  57
  58 unsigned int nmi_watchdog = NMI_DEFAULT;
  59 static unsigned int nmi_hz = HZ;
  60
  61 struct nmi_watchdog_ctlblk {
  62         int enabled;
  63         u64 check_bit;
  64         unsigned int cccr_msr;
  65         unsigned int perfctr_msr;  /* the MSR to reset in NMI handler */
  66         unsigned int evntsel_msr;  /* the MSR to select the events to handle */
  67 };
  68 static DEFINE_PER_CPU(struct nmi_watchdog_ctlblk, nmi_watchdog_ctlblk);
  69
  70 /* local prototypes */
  71 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu);
  72
  73 /* converts an msr to an appropriate reservation bit */
  74 static inline unsigned int nmi_perfctr_msr_to_bit(unsigned int msr)
  75 {
  76         /* returns the bit offset of the performance counter register */
  77         switch (boot_cpu_data.x86_vendor) {
  78         case X86_VENDOR_AMD:
  79                 return (msr - MSR_K7_PERFCTR0);
  80         case X86_VENDOR_INTEL:
  81                 if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
  82                         return (msr - MSR_ARCH_PERFMON_PERFCTR0);
  83                 else
  84                         return (msr - MSR_P4_BPU_PERFCTR0);
  85         }
  86         return 0;
  87 }
  88
  89 /* converts an msr to an appropriate reservation bit */
  90 static inline unsigned int nmi_evntsel_msr_to_bit(unsigned int msr)
  91 {
  92         /* returns the bit offset of the event selection register */
  93         switch (boot_cpu_data.x86_vendor) {
  94         case X86_VENDOR_AMD:
  95                 return (msr - MSR_K7_EVNTSEL0);
  96         case X86_VENDOR_INTEL:
  97                 if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
  98                         return (msr - MSR_ARCH_PERFMON_EVENTSEL0);
  99                 else
 100                         return (msr - MSR_P4_BSU_ESCR0);
 101         }
 102         return 0;
 103 }
 104
 105 /* checks for a bit availability (hack for oprofile) */
 106 int avail_to_resrv_perfctr_nmi_bit(unsigned int counter)
 107 {
 108         BUG_ON(counter > NMI_MAX_COUNTER_BITS);
 109
 110         return (!test_bit(counter, &__get_cpu_var(perfctr_nmi_owner)));
 111 }
 112
 113 /* checks the an msr for availability */
 114 int avail_to_resrv_perfctr_nmi(unsigned int msr)
 115 {
 116         unsigned int counter;
 117
 118         counter = nmi_perfctr_msr_to_bit(msr);
 119         BUG_ON(counter > NMI_MAX_COUNTER_BITS);
 120
 121         return (!test_bit(counter, &__get_cpu_var(perfctr_nmi_owner)));
 122 }
 123
 124 int reserve_perfctr_nmi(unsigned int msr)
 125 {
 126         unsigned int counter;
 127
 128         counter = nmi_perfctr_msr_to_bit(msr);
 129         BUG_ON(counter > NMI_MAX_COUNTER_BITS);
 130
 131         if (!test_and_set_bit(counter, &__get_cpu_var(perfctr_nmi_owner)))
 132                 return 1;
 133         return 0;
 134 }
 135
 136 void release_perfctr_nmi(unsigned int msr)
 137 {
 138         unsigned int counter;
 139
 140         counter = nmi_perfctr_msr_to_bit(msr);
 141         BUG_ON(counter > NMI_MAX_COUNTER_BITS);
 142
 143         clear_bit(counter, &__get_cpu_var(perfctr_nmi_owner));
 144 }
 145
 146 int reserve_evntsel_nmi(unsigned int msr)
 147 {
 148         unsigned int counter;
 149
 150         counter = nmi_evntsel_msr_to_bit(msr);
 151         BUG_ON(counter > NMI_MAX_COUNTER_BITS);
 152
 153         if (!test_and_set_bit(counter, &__get_cpu_var(evntsel_nmi_owner)))
 154                 return 1;
 155         return 0;
 156 }
 157
 158 void release_evntsel_nmi(unsigned int msr)
 159 {
 160         unsigned int counter;
 161
 162         counter = nmi_evntsel_msr_to_bit(msr);
 163         BUG_ON(counter > NMI_MAX_COUNTER_BITS);
 164
 165         clear_bit(counter, &__get_cpu_var(evntsel_nmi_owner));
 166 }
 167
 168 static __cpuinit inline int nmi_known_cpu(void)
 169 {
 170         switch (boot_cpu_data.x86_vendor) {
 171         case X86_VENDOR_AMD:
 172                 return boot_cpu_data.x86 == 15;
 173         case X86_VENDOR_INTEL:
 174                 if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
 175                         return 1;
 176                 else
 177                         return (boot_cpu_data.x86 == 15);
 178         }
 179         return 0;
 180 }
 181
 182 /* Run after command line and cpu_init init, but before all other checks */
 183 void nmi_watchdog_default(void)
 184 {
 185         if (nmi_watchdog != NMI_DEFAULT)
 186                 return;
 187         if (nmi_known_cpu())
 188                 nmi_watchdog = NMI_LOCAL_APIC;
 189         else
 190                 nmi_watchdog = NMI_IO_APIC;
 191 }
 192
 193 #ifdef CONFIG_SMP
 194 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
 195  * the CPU is idle. To make sure the NMI watchdog really ticks on all
 196  * CPUs during the test make them busy.
 197  */
 198 static __init void nmi_cpu_busy(void *data)
 199 {
 200         volatile int *endflag = data;
 201         local_irq_enable_in_hardirq();
 202         /* Intentionally don't use cpu_relax here. This is
 203            to make sure that the performance counter really ticks,
 204            even if there is a simulator or similar that catches the
 205            pause instruction. On a real HT machine this is fine because
 206            all other CPUs are busy with "useless" delay loops and don't
 207            care if they get somewhat less cycles. */
 208         while (*endflag == 0)
 209                 barrier();
 210 }
 211 #endif
 212
 213 int __init check_nmi_watchdog (void)
 214 {
 215         volatile int endflag = 0;
 216         int *counts;
 217         int cpu;
 218
 219         if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DEFAULT))
 220                 return 0;
 221
 222         if (!atomic_read(&nmi_active))
 223                 return 0;
 224
 225         counts = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
 226         if (!counts)
 227                 return -1;
 228
 229         printk(KERN_INFO "testing NMI watchdog ... ");
 230
 231 #ifdef CONFIG_SMP
 232         if (nmi_watchdog == NMI_LOCAL_APIC)
 233                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
 234 #endif
 235
 236         for (cpu = 0; cpu < NR_CPUS; cpu++)
 237                 counts[cpu] = cpu_pda(cpu)->__nmi_count;
 238         local_irq_enable();
 239         mdelay((10*1000)/nmi_hz); // wait 10 ticks
 240
 241         for_each_online_cpu(cpu) {
 242                 if (!per_cpu(nmi_watchdog_ctlblk, cpu).enabled)
 243                         continue;
 244                 if (cpu_pda(cpu)->__nmi_count - counts[cpu] <= 5) {
 245                         printk("CPU#%d: NMI appears to be stuck (%d->%d)!\n",
 246                                cpu,
 247                                counts[cpu],
 248                                cpu_pda(cpu)->__nmi_count);
 249                         per_cpu(nmi_watchdog_ctlblk, cpu).enabled = 0;
 250                         atomic_dec(&nmi_active);
 251                 }
 252         }
 253         if (!atomic_read(&nmi_active)) {
 254                 kfree(counts);
 255                 atomic_set(&nmi_active, -1);
 256                 return -1;
 257         }
 258         endflag = 1;
 259         printk("OK.\n");
 260
 261         /* now that we know it works we can reduce NMI frequency to
 262            something more reasonable; makes a difference in some configs */
 263         if (nmi_watchdog == NMI_LOCAL_APIC) {
 264                 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 265
 266                 nmi_hz = 1;
 267                 /*
 268                  * On Intel CPUs with ARCH_PERFMON only 32 bits in the counter
 269                  * are writable, with higher bits sign extending from bit 31.
 270                  * So, we can only program the counter with 31 bit values and
 271                  * 32nd bit should be 1, for 33.. to be 1.
 272                  * Find the appropriate nmi_hz
 273                  */
 274                 if (wd->perfctr_msr == MSR_ARCH_PERFMON_PERFCTR0 &&
 275                         ((u64)cpu_khz * 1000) > 0x7fffffffULL) {
 276                         nmi_hz = ((u64)cpu_khz * 1000) / 0x7fffffffUL + 1;
 277                 }
 278         }
 279
 280         kfree(counts);
 281         return 0;
 282 }
 283
 284 int __init setup_nmi_watchdog(char *str)
 285 {
 286         int nmi;
 287
 288         if (!strncmp(str,"panic",5)) {
 289                 panic_on_timeout = 1;
 290                 str = strchr(str, ',');
 291                 if (!str)
 292                         return 1;
 293                 ++str;
 294         }
 295
 296         get_option(&str, &nmi);
 297
 298         if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
 299                 return 0;
 300
 301         if ((nmi == NMI_LOCAL_APIC) && (nmi_known_cpu() == 0))
 302                 return 0;  /* no lapic support */
 303         nmi_watchdog = nmi;
 304         return 1;
 305 }
 306
 307 __setup("nmi_watchdog=", setup_nmi_watchdog);
 308
 309 static void disable_lapic_nmi_watchdog(void)
 310 {
 311         BUG_ON(nmi_watchdog != NMI_LOCAL_APIC);
 312
 313         if (atomic_read(&nmi_active) <= 0)
 314                 return;
 315
 316         on_each_cpu(stop_apic_nmi_watchdog, NULL, 0, 1);
 317
 318         BUG_ON(atomic_read(&nmi_active) != 0);
 319 }
 320
 321 static void enable_lapic_nmi_watchdog(void)
 322 {
 323         BUG_ON(nmi_watchdog != NMI_LOCAL_APIC);
 324
 325         /* are we already enabled */
 326         if (atomic_read(&nmi_active) != 0)
 327                 return;
 328
 329         /* are we lapic aware */
 330         if (nmi_known_cpu() <= 0)
 331                 return;
 332
 333         on_each_cpu(setup_apic_nmi_watchdog, NULL, 0, 1);
 334         touch_nmi_watchdog();
 335 }
 336
 337 void disable_timer_nmi_watchdog(void)
 338 {
 339         BUG_ON(nmi_watchdog != NMI_IO_APIC);
 340
 341         if (atomic_read(&nmi_active) <= 0)
 342                 return;
 343
 344         disable_irq(0);
 345         on_each_cpu(stop_apic_nmi_watchdog, NULL, 0, 1);
 346
 347         BUG_ON(atomic_read(&nmi_active) != 0);
 348 }
 349
 350 void enable_timer_nmi_watchdog(void)
 351 {
 352         BUG_ON(nmi_watchdog != NMI_IO_APIC);
 353
 354         if (atomic_read(&nmi_active) == 0) {
 355                 touch_nmi_watchdog();
 356                 on_each_cpu(setup_apic_nmi_watchdog, NULL, 0, 1);
 357                 enable_irq(0);
 358         }
 359 }
 360
 361 #ifdef CONFIG_PM
 362
 363 static int nmi_pm_active; /* nmi_active before suspend */
 364
 365 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 366 {
 367         /* only CPU0 goes here, other CPUs should be offline */
 368         nmi_pm_active = atomic_read(&nmi_active);
 369         stop_apic_nmi_watchdog(NULL);
 370         BUG_ON(atomic_read(&nmi_active) != 0);
 371         return 0;
 372 }
 373
 374 static int lapic_nmi_resume(struct sys_device *dev)
 375 {
 376         /* only CPU0 goes here, other CPUs should be offline */
 377         if (nmi_pm_active > 0) {
 378                 setup_apic_nmi_watchdog(NULL);
 379                 touch_nmi_watchdog();
 380         }
 381         return 0;
 382 }
 383
 384 static struct sysdev_class nmi_sysclass = {
 385         set_kset_name("lapic_nmi"),
 386         .resume         = lapic_nmi_resume,
 387         .suspend        = lapic_nmi_suspend,
 388 };
 389
 390 static struct sys_device device_lapic_nmi = {
 391         .id             = 0,
 392         .cls    = &nmi_sysclass,
 393 };
 394
 395 static int __init init_lapic_nmi_sysfs(void)
 396 {
 397         int error;
 398
 399         /* should really be a BUG_ON but b/c this is an
 400          * init call, it just doesn't work.  -dcz
 401          */
 402         if (nmi_watchdog != NMI_LOCAL_APIC)
 403                 return 0;
 404
 405         if ( atomic_read(&nmi_active) < 0 )
 406                 return 0;
 407
 408         error = sysdev_class_register(&nmi_sysclass);
 409         if (!error)
 410                 error = sysdev_register(&device_lapic_nmi);
 411         return error;
 412 }
 413 /* must come after the local APIC's device_initcall() */
 414 late_initcall(init_lapic_nmi_sysfs);
 415
 416 #endif  /* CONFIG_PM */
 417
 418 /*
 419  * Activate the NMI watchdog via the local APIC.
 420  * Original code written by Keith Owens.
 421  */
 422
 423 /* Note that these events don't tick when the CPU idles. This means
 424    the frequency varies with CPU load. */
 425
 426 #define K7_EVNTSEL_ENABLE       (1 << 22)
 427 #define K7_EVNTSEL_INT          (1 << 20)
 428 #define K7_EVNTSEL_OS           (1 << 17)
 429 #define K7_EVNTSEL_USR          (1 << 16)
 430 #define K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING    0x76
 431 #define K7_NMI_EVENT            K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING
 432
 433 static int setup_k7_watchdog(void)
 434 {
 435         unsigned int perfctr_msr, evntsel_msr;
 436         unsigned int evntsel;
 437         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 438
 439         perfctr_msr = MSR_K7_PERFCTR0;
 440         evntsel_msr = MSR_K7_EVNTSEL0;
 441         if (!reserve_perfctr_nmi(perfctr_msr))
 442                 goto fail;
 443
 444         if (!reserve_evntsel_nmi(evntsel_msr))
 445                 goto fail1;
 446
 447         /* Simulator may not support it */
 448         if (checking_wrmsrl(evntsel_msr, 0UL))
 449                 goto fail2;
 450         wrmsrl(perfctr_msr, 0UL);
 451
 452         evntsel = K7_EVNTSEL_INT
 453                 | K7_EVNTSEL_OS
 454                 | K7_EVNTSEL_USR
 455                 | K7_NMI_EVENT;
 456
 457         /* setup the timer */
 458         wrmsr(evntsel_msr, evntsel, 0);
 459         wrmsrl(perfctr_msr, -((u64)cpu_khz * 1000 / nmi_hz));
 460         apic_write(APIC_LVTPC, APIC_DM_NMI);
 461         evntsel |= K7_EVNTSEL_ENABLE;
 462         wrmsr(evntsel_msr, evntsel, 0);
 463
 464         wd->perfctr_msr = perfctr_msr;
 465         wd->evntsel_msr = evntsel_msr;
 466         wd->cccr_msr = 0;  //unused
 467         wd->check_bit = 1ULL<<63;
 468         return 1;
 469 fail2:
 470         release_evntsel_nmi(evntsel_msr);
 471 fail1:
 472         release_perfctr_nmi(perfctr_msr);
 473 fail:
 474         return 0;
 475 }
 476
 477 static void stop_k7_watchdog(void)
 478 {
 479         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 480
 481         wrmsr(wd->evntsel_msr, 0, 0);
 482
 483         release_evntsel_nmi(wd->evntsel_msr);
 484         release_perfctr_nmi(wd->perfctr_msr);
 485 }
 486
 487 /* Note that these events don't tick when the CPU idles. This means
 488    the frequency varies with CPU load. */
 489
 490 #define MSR_P4_MISC_ENABLE_PERF_AVAIL   (1<<7)
 491 #define P4_ESCR_EVENT_SELECT(N) ((N)<<25)
 492 #define P4_ESCR_OS              (1<<3)
 493 #define P4_ESCR_USR             (1<<2)
 494 #define P4_CCCR_OVF_PMI0        (1<<26)
 495 #define P4_CCCR_OVF_PMI1        (1<<27)
 496 #define P4_CCCR_THRESHOLD(N)    ((N)<<20)
 497 #define P4_CCCR_COMPLEMENT      (1<<19)
 498 #define P4_CCCR_COMPARE         (1<<18)
 499 #define P4_CCCR_REQUIRED        (3<<16)
 500 #define P4_CCCR_ESCR_SELECT(N)  ((N)<<13)
 501 #define P4_CCCR_ENABLE          (1<<12)
 502 #define P4_CCCR_OVF             (1<<31)
 503 /* Set up IQ_COUNTER0 to behave like a clock, by having IQ_CCCR0 filter
 504    CRU_ESCR0 (with any non-null event selector) through a complemented
 505    max threshold. [IA32-Vol3, Section 14.9.9] */
 506
 507 static int setup_p4_watchdog(void)
 508 {
 509         unsigned int perfctr_msr, evntsel_msr, cccr_msr;
 510         unsigned int evntsel, cccr_val;
 511         unsigned int misc_enable, dummy;
 512         unsigned int ht_num;
 513         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 514
 515         rdmsr(MSR_IA32_MISC_ENABLE, misc_enable, dummy);
 516         if (!(misc_enable & MSR_P4_MISC_ENABLE_PERF_AVAIL))
 517                 return 0;
 518
 519 #ifdef CONFIG_SMP
 520         /* detect which hyperthread we are on */
 521         if (smp_num_siblings == 2) {
 522                 unsigned int ebx, apicid;
 523
 524                 ebx = cpuid_ebx(1);
 525                 apicid = (ebx >> 24) & 0xff;
 526                 ht_num = apicid & 1;
 527         } else
 528 #endif
 529                 ht_num = 0;
 530
 531         /* performance counters are shared resources
 532          * assign each hyperthread its own set
 533          * (re-use the ESCR0 register, seems safe
 534          * and keeps the cccr_val the same)
 535          */
 536         if (!ht_num) {
 537                 /* logical cpu 0 */
 538                 perfctr_msr = MSR_P4_IQ_PERFCTR0;
 539                 evntsel_msr = MSR_P4_CRU_ESCR0;
 540                 cccr_msr = MSR_P4_IQ_CCCR0;
 541                 cccr_val = P4_CCCR_OVF_PMI0 | P4_CCCR_ESCR_SELECT(4);
 542         } else {
 543                 /* logical cpu 1 */
 544                 perfctr_msr = MSR_P4_IQ_PERFCTR1;
 545                 evntsel_msr = MSR_P4_CRU_ESCR0;
 546                 cccr_msr = MSR_P4_IQ_CCCR1;
 547                 cccr_val = P4_CCCR_OVF_PMI1 | P4_CCCR_ESCR_SELECT(4);
 548         }
 549
 550         if (!reserve_perfctr_nmi(perfctr_msr))
 551                 goto fail;
 552
 553         if (!reserve_evntsel_nmi(evntsel_msr))
 554                 goto fail1;
 555
 556         evntsel = P4_ESCR_EVENT_SELECT(0x3F)
 557                 | P4_ESCR_OS
 558                 | P4_ESCR_USR;
 559
 560         cccr_val |= P4_CCCR_THRESHOLD(15)
 561                  | P4_CCCR_COMPLEMENT
 562                  | P4_CCCR_COMPARE
 563                  | P4_CCCR_REQUIRED;
 564
 565         wrmsr(evntsel_msr, evntsel, 0);
 566         wrmsr(cccr_msr, cccr_val, 0);
 567         wrmsrl(perfctr_msr, -((u64)cpu_khz * 1000 / nmi_hz));
 568         apic_write(APIC_LVTPC, APIC_DM_NMI);
 569         cccr_val |= P4_CCCR_ENABLE;
 570         wrmsr(cccr_msr, cccr_val, 0);
 571
 572         wd->perfctr_msr = perfctr_msr;
 573         wd->evntsel_msr = evntsel_msr;
 574         wd->cccr_msr = cccr_msr;
 575         wd->check_bit = 1ULL<<39;
 576         return 1;
 577 fail1:
 578         release_perfctr_nmi(perfctr_msr);
 579 fail:
 580         return 0;
 581 }
 582
 583 static void stop_p4_watchdog(void)
 584 {
 585         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 586
 587         wrmsr(wd->cccr_msr, 0, 0);
 588         wrmsr(wd->evntsel_msr, 0, 0);
 589
 590         release_evntsel_nmi(wd->evntsel_msr);
 591         release_perfctr_nmi(wd->perfctr_msr);
 592 }
 593
 594 #define ARCH_PERFMON_NMI_EVENT_SEL      ARCH_PERFMON_UNHALTED_CORE_CYCLES_SEL
 595 #define ARCH_PERFMON_NMI_EVENT_UMASK    ARCH_PERFMON_UNHALTED_CORE_CYCLES_UMASK
 596
 597 static int setup_intel_arch_watchdog(void)
 598 {
 599         unsigned int ebx;
 600         union cpuid10_eax eax;
 601         unsigned int unused;
 602         unsigned int perfctr_msr, evntsel_msr;
 603         unsigned int evntsel;
 604         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 605
 606         /*
 607          * Check whether the Architectural PerfMon supports
 608          * Unhalted Core Cycles Event or not.
 609          * NOTE: Corresponding bit = 0 in ebx indicates event present.
 610          */
 611         cpuid(10, &(eax.full), &ebx, &unused, &unused);
 612         if ((eax.split.mask_length < (ARCH_PERFMON_UNHALTED_CORE_CYCLES_INDEX+1)) ||
 613             (ebx & ARCH_PERFMON_UNHALTED_CORE_CYCLES_PRESENT))
 614                 goto fail;
 615
 616         perfctr_msr = MSR_ARCH_PERFMON_PERFCTR0;
 617         evntsel_msr = MSR_ARCH_PERFMON_EVENTSEL0;
 618
 619         if (!reserve_perfctr_nmi(perfctr_msr))
 620                 goto fail;
 621
 622         if (!reserve_evntsel_nmi(evntsel_msr))
 623                 goto fail1;
 624
 625         wrmsrl(perfctr_msr, 0UL);
 626
 627         evntsel = ARCH_PERFMON_EVENTSEL_INT
 628                 | ARCH_PERFMON_EVENTSEL_OS
 629                 | ARCH_PERFMON_EVENTSEL_USR
 630                 | ARCH_PERFMON_NMI_EVENT_SEL
 631                 | ARCH_PERFMON_NMI_EVENT_UMASK;
 632
 633         /* setup the timer */
 634         wrmsr(evntsel_msr, evntsel, 0);
 635         wrmsrl(perfctr_msr, -((u64)cpu_khz * 1000 / nmi_hz));
 636
 637         apic_write(APIC_LVTPC, APIC_DM_NMI);
 638         evntsel |= ARCH_PERFMON_EVENTSEL0_ENABLE;
 639         wrmsr(evntsel_msr, evntsel, 0);
 640
 641         wd->perfctr_msr = perfctr_msr;
 642         wd->evntsel_msr = evntsel_msr;
 643         wd->cccr_msr = 0;  //unused
 644         wd->check_bit = 1ULL << (eax.split.bit_width - 1);
 645         return 1;
 646 fail1:
 647         release_perfctr_nmi(perfctr_msr);
 648 fail:
 649         return 0;
 650 }
 651
 652 static void stop_intel_arch_watchdog(void)
 653 {
 654         unsigned int ebx;
 655         union cpuid10_eax eax;
 656         unsigned int unused;
 657         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 658
 659         /*
 660          * Check whether the Architectural PerfMon supports
 661          * Unhalted Core Cycles Event or not.
 662          * NOTE: Corresponding bit = 0 in ebx indicates event present.
 663          */
 664         cpuid(10, &(eax.full), &ebx, &unused, &unused);
 665         if ((eax.split.mask_length < (ARCH_PERFMON_UNHALTED_CORE_CYCLES_INDEX+1)) ||
 666             (ebx & ARCH_PERFMON_UNHALTED_CORE_CYCLES_PRESENT))
 667                 return;
 668
 669         wrmsr(wd->evntsel_msr, 0, 0);
 670
 671         release_evntsel_nmi(wd->evntsel_msr);
 672         release_perfctr_nmi(wd->perfctr_msr);
 673 }
 674
 675 void setup_apic_nmi_watchdog(void *unused)
 676 {
 677         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 678
 679         /* only support LOCAL and IO APICs for now */
 680         if ((nmi_watchdog != NMI_LOCAL_APIC) &&
 681             (nmi_watchdog != NMI_IO_APIC))
 682                 return;
 683
 684         if (wd->enabled == 1)
 685                 return;
 686
 687         /* cheap hack to support suspend/resume */
 688         /* if cpu0 is not active neither should the other cpus */
 689         if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
 690                 return;
 691
 692         if (nmi_watchdog == NMI_LOCAL_APIC) {
 693                 switch (boot_cpu_data.x86_vendor) {
 694                 case X86_VENDOR_AMD:
 695                         if (strstr(boot_cpu_data.x86_model_id, "Screwdriver"))
 696                                 return;
 697                         if (!setup_k7_watchdog())
 698                                 return;
 699                         break;
 700                 case X86_VENDOR_INTEL:
 701                         if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
 702                                 if (!setup_intel_arch_watchdog())
 703                                         return;
 704                                 break;
 705                         }
 706                         if (!setup_p4_watchdog())
 707                                 return;
 708                         break;
 709                 default:
 710                         return;
 711                 }
 712         }
 713         wd->enabled = 1;
 714         atomic_inc(&nmi_active);
 715 }
 716
 717 void stop_apic_nmi_watchdog(void *unused)
 718 {
 719         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 720
 721         /* only support LOCAL and IO APICs for now */
 722         if ((nmi_watchdog != NMI_LOCAL_APIC) &&
 723             (nmi_watchdog != NMI_IO_APIC))
 724                 return;
 725
 726         if (wd->enabled == 0)
 727                 return;
 728
 729         if (nmi_watchdog == NMI_LOCAL_APIC) {
 730                 switch (boot_cpu_data.x86_vendor) {
 731                 case X86_VENDOR_AMD:
 732                         if (strstr(boot_cpu_data.x86_model_id, "Screwdriver"))
 733                                 return;
 734                         stop_k7_watchdog();
 735                         break;
 736                 case X86_VENDOR_INTEL:
 737                         if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
 738                                 stop_intel_arch_watchdog();
 739                                 break;
 740                         }
 741                         stop_p4_watchdog();
 742                         break;
 743                 default:
 744                         return;
 745                 }
 746         }
 747         wd->enabled = 0;
 748         atomic_dec(&nmi_active);
 749 }
 750
 751 /*
 752  * the best way to detect whether a CPU has a 'hard lockup' problem
 753  * is to check it's local APIC timer IRQ counts. If they are not
 754  * changing then that CPU has some problem.
 755  *
 756  * as these watchdog NMI IRQs are generated on every CPU, we only
 757  * have to check the current processor.
 758  */
 759
 760 static DEFINE_PER_CPU(unsigned, last_irq_sum);
 761 static DEFINE_PER_CPU(local_t, alert_counter);
 762 static DEFINE_PER_CPU(int, nmi_touch);
 763
 764 void touch_nmi_watchdog (void)
 765 {
 766         if (nmi_watchdog > 0) {
 767                 unsigned cpu;
 768
 769                 /*
 770                  * Tell other CPUs to reset their alert counters. We cannot
 771                  * do it ourselves because the alert count increase is not
 772                  * atomic.
 773                  */
 774                 for_each_present_cpu (cpu)
 775                         per_cpu(nmi_touch, cpu) = 1;
 776         }
 777
 778         touch_softlockup_watchdog();
 779 }
 780
 781 int __kprobes nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
 782 {
 783         int sum;
 784         int touched = 0;
 785         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 786         u64 dummy;
 787         int rc=0;
 788
 789         /* check for other users first */
 790         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 791                         == NOTIFY_STOP) {
 792                 rc = 1;
 793                 touched = 1;
 794         }
 795
 796         sum = read_pda(apic_timer_irqs);
 797         if (__get_cpu_var(nmi_touch)) {
 798                 __get_cpu_var(nmi_touch) = 0;
 799                 touched = 1;
 800         }
 801
 802 #ifdef CONFIG_X86_MCE
 803         /* Could check oops_in_progress here too, but it's safer
 804            not too */
 805         if (atomic_read(&mce_entry) > 0)
 806                 touched = 1;
 807 #endif
 808         /* if the apic timer isn't firing, this cpu isn't doing much */
 809         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 810                 /*
 811                  * Ayiee, looks like this CPU is stuck ...
 812                  * wait a few IRQs (5 seconds) before doing the oops ...
 813                  */
 814                 local_inc(&__get_cpu_var(alert_counter));
 815                 if (local_read(&__get_cpu_var(alert_counter)) == 5*nmi_hz)
 816                         die_nmi("NMI Watchdog detected LOCKUP on CPU %d\n", regs,
 817                                 panic_on_timeout);
 818         } else {
 819                 __get_cpu_var(last_irq_sum) = sum;
 820                 local_set(&__get_cpu_var(alert_counter), 0);
 821         }
 822
 823         /* see if the nmi watchdog went off */
 824         if (wd->enabled) {
 825                 if (nmi_watchdog == NMI_LOCAL_APIC) {
 826                         rdmsrl(wd->perfctr_msr, dummy);
 827                         if (dummy & wd->check_bit){
 828                                 /* this wasn't a watchdog timer interrupt */
 829                                 goto done;
 830                         }
 831
 832                         /* only Intel uses the cccr msr */
 833                         if (wd->cccr_msr != 0) {
 834                                 /*
 835                                  * P4 quirks:
 836                                  * - An overflown perfctr will assert its interrupt
 837                                  *   until the OVF flag in its CCCR is cleared.
 838                                  * - LVTPC is masked on interrupt and must be
 839                                  *   unmasked by the LVTPC handler.
 840                                  */
 841                                 rdmsrl(wd->cccr_msr, dummy);
 842                                 dummy &= ~P4_CCCR_OVF;
 843                                 wrmsrl(wd->cccr_msr, dummy);
 844                                 apic_write(APIC_LVTPC, APIC_DM_NMI);
 845                         } else if (wd->perfctr_msr == MSR_ARCH_PERFMON_PERFCTR0) {
 846                                 /*
 847                                  * ArchPerfom/Core Duo needs to re-unmask
 848                                  * the apic vector
 849                                  */
 850                                 apic_write(APIC_LVTPC, APIC_DM_NMI);
 851                         }
 852                         /* start the cycle over again */
 853                         wrmsrl(wd->perfctr_msr, -((u64)cpu_khz * 1000 / nmi_hz));
 854                         rc = 1;
 855                 } else  if (nmi_watchdog == NMI_IO_APIC) {
 856                         /* don't know how to accurately check for this.
 857                          * just assume it was a watchdog timer interrupt
 858                          * This matches the old behaviour.
 859                          */
 860                         rc = 1;
 861                 } else
 862                         printk(KERN_WARNING "Unknown enabled NMI hardware?!\n");
 863         }
 864 done:
 865         return rc;
 866 }
 867
 868 asmlinkage __kprobes void do_nmi(struct pt_regs * regs, long error_code)
 869 {
 870         nmi_enter();
 871         add_pda(__nmi_count,1);
 872         default_do_nmi(regs);
 873         nmi_exit();
 874 }
 875
 876 int do_nmi_callback(struct pt_regs * regs, int cpu)
 877 {
 878 #ifdef CONFIG_SYSCTL
 879         if (unknown_nmi_panic)
 880                 return unknown_nmi_panic_callback(regs, cpu);
 881 #endif
 882         return 0;
 883 }
 884
 885 #ifdef CONFIG_SYSCTL
 886
 887 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 888 {
 889         unsigned char reason = get_nmi_reason();
 890         char buf[64];
 891
 892         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 893         die_nmi(buf, regs, 1);  /* Always panic here */
 894         return 0;
 895 }
 896
 897 /*
 898  * proc handler for /proc/sys/kernel/nmi
 899  */
 900 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 901                         void __user *buffer, size_t *length, loff_t *ppos)
 902 {
 903         int old_state;
 904
 905         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 906         old_state = nmi_watchdog_enabled;
 907         proc_dointvec(table, write, file, buffer, length, ppos);
 908         if (!!old_state == !!nmi_watchdog_enabled)
 909                 return 0;
 910
 911         if (atomic_read(&nmi_active) < 0) {
 912                 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
 913                 return -EIO;
 914         }
 915
 916         /* if nmi_watchdog is not set yet, then set it */
 917         nmi_watchdog_default();
 918
 919         if (nmi_watchdog == NMI_LOCAL_APIC) {
 920                 if (nmi_watchdog_enabled)
 921                         enable_lapic_nmi_watchdog();
 922                 else
 923                         disable_lapic_nmi_watchdog();
 924         } else {
 925                 printk( KERN_WARNING
 926                         "NMI watchdog doesn't know what hardware to touch\n");
 927                 return -EIO;
 928         }
 929         return 0;
 930 }
 931
 932 #endif
 933
 934 EXPORT_SYMBOL(nmi_active);
 935 EXPORT_SYMBOL(nmi_watchdog);
 936 EXPORT_SYMBOL(avail_to_resrv_perfctr_nmi);
 937 EXPORT_SYMBOL(avail_to_resrv_perfctr_nmi_bit);
 938 EXPORT_SYMBOL(reserve_perfctr_nmi);
 939 EXPORT_SYMBOL(release_perfctr_nmi);
 940 EXPORT_SYMBOL(reserve_evntsel_nmi);
 941 EXPORT_SYMBOL(release_evntsel_nmi);
 942 EXPORT_SYMBOL(disable_timer_nmi_watchdog);
 943 EXPORT_SYMBOL(enable_timer_nmi_watchdog);
 944 EXPORT_SYMBOL(touch_nmi_watchdog);