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/nmi.h>
  16 #include <linux/mm.h>
  17 #include <linux/delay.h>
  18 #include <linux/interrupt.h>
  19 #include <linux/module.h>
  20 #include <linux/sysdev.h>
  21 #include <linux/sysctl.h>
  22 #include <linux/kprobes.h>
  23 #include <linux/cpumask.h>
  24
  25 #include <asm/smp.h>
  26 #include <asm/nmi.h>
  27 #include <asm/proto.h>
  28 #include <asm/kdebug.h>
  29 #include <asm/mce.h>
  30 #include <asm/intel_arch_perfmon.h>
  31
  32 int unknown_nmi_panic;
  33 int nmi_watchdog_enabled;
  34 int panic_on_unrecovered_nmi;
  35
  36 /* perfctr_nmi_owner tracks the ownership of the perfctr registers:
  37  * evtsel_nmi_owner tracks the ownership of the event selection
  38  * - different performance counters/ event selection may be reserved for
  39  *   different subsystems this reservation system just tries to coordinate
  40  *   things a little
  41  */
  42 static DEFINE_PER_CPU(unsigned, perfctr_nmi_owner);
  43 static DEFINE_PER_CPU(unsigned, evntsel_nmi_owner[2]);
  44
  45 static cpumask_t backtrace_mask = CPU_MASK_NONE;
  46
  47 /* this number is calculated from Intel's MSR_P4_CRU_ESCR5 register and it's
  48  * offset from MSR_P4_BSU_ESCR0.  It will be the max for all platforms (for now)
  49  */
  50 #define NMI_MAX_COUNTER_BITS 66
  51
  52 /* nmi_active:
  53  * >0: the lapic NMI watchdog is active, but can be disabled
  54  * <0: the lapic NMI watchdog has not been set up, and cannot
  55  *     be enabled
  56  *  0: the lapic NMI watchdog is disabled, but can be enabled
  57  */
  58 atomic_t nmi_active = ATOMIC_INIT(0);           /* oprofile uses this */
  59 int panic_on_timeout;
  60
  61 unsigned int nmi_watchdog = NMI_DEFAULT;
  62 static unsigned int nmi_hz = HZ;
  63
  64 struct nmi_watchdog_ctlblk {
  65         int enabled;
  66         u64 check_bit;
  67         unsigned int cccr_msr;
  68         unsigned int perfctr_msr;  /* the MSR to reset in NMI handler */
  69         unsigned int evntsel_msr;  /* the MSR to select the events to handle */
  70 };
  71 static DEFINE_PER_CPU(struct nmi_watchdog_ctlblk, nmi_watchdog_ctlblk);
  72
  73 /* local prototypes */
  74 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu);
  75
  76 /* converts an msr to an appropriate reservation bit */
  77 static inline unsigned int nmi_perfctr_msr_to_bit(unsigned int msr)
  78 {
  79         /* returns the bit offset of the performance counter register */
  80         switch (boot_cpu_data.x86_vendor) {
  81         case X86_VENDOR_AMD:
  82                 return (msr - MSR_K7_PERFCTR0);
  83         case X86_VENDOR_INTEL:
  84                 if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
  85                         return (msr - MSR_ARCH_PERFMON_PERFCTR0);
  86                 else
  87                         return (msr - MSR_P4_BPU_PERFCTR0);
  88         }
  89         return 0;
  90 }
  91
  92 /* converts an msr to an appropriate reservation bit */
  93 static inline unsigned int nmi_evntsel_msr_to_bit(unsigned int msr)
  94 {
  95         /* returns the bit offset of the event selection register */
  96         switch (boot_cpu_data.x86_vendor) {
  97         case X86_VENDOR_AMD:
  98                 return (msr - MSR_K7_EVNTSEL0);
  99         case X86_VENDOR_INTEL:
 100                 if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
 101                         return (msr - MSR_ARCH_PERFMON_EVENTSEL0);
 102                 else
 103                         return (msr - MSR_P4_BSU_ESCR0);
 104         }
 105         return 0;
 106 }
 107
 108 /* checks for a bit availability (hack for oprofile) */
 109 int avail_to_resrv_perfctr_nmi_bit(unsigned int counter)
 110 {
 111         BUG_ON(counter > NMI_MAX_COUNTER_BITS);
 112
 113         return (!test_bit(counter, &__get_cpu_var(perfctr_nmi_owner)));
 114 }
 115
 116 /* checks the an msr for availability */
 117 int avail_to_resrv_perfctr_nmi(unsigned int msr)
 118 {
 119         unsigned int counter;
 120
 121         counter = nmi_perfctr_msr_to_bit(msr);
 122         BUG_ON(counter > NMI_MAX_COUNTER_BITS);
 123
 124         return (!test_bit(counter, &__get_cpu_var(perfctr_nmi_owner)));
 125 }
 126
 127 int reserve_perfctr_nmi(unsigned int msr)
 128 {
 129         unsigned int counter;
 130
 131         counter = nmi_perfctr_msr_to_bit(msr);
 132         BUG_ON(counter > NMI_MAX_COUNTER_BITS);
 133
 134         if (!test_and_set_bit(counter, &__get_cpu_var(perfctr_nmi_owner)))
 135                 return 1;
 136         return 0;
 137 }
 138
 139 void release_perfctr_nmi(unsigned int msr)
 140 {
 141         unsigned int counter;
 142
 143         counter = nmi_perfctr_msr_to_bit(msr);
 144         BUG_ON(counter > NMI_MAX_COUNTER_BITS);
 145
 146         clear_bit(counter, &__get_cpu_var(perfctr_nmi_owner));
 147 }
 148
 149 int reserve_evntsel_nmi(unsigned int msr)
 150 {
 151         unsigned int counter;
 152
 153         counter = nmi_evntsel_msr_to_bit(msr);
 154         BUG_ON(counter > NMI_MAX_COUNTER_BITS);
 155
 156         if (!test_and_set_bit(counter, &__get_cpu_var(evntsel_nmi_owner)))
 157                 return 1;
 158         return 0;
 159 }
 160
 161 void release_evntsel_nmi(unsigned int msr)
 162 {
 163         unsigned int counter;
 164
 165         counter = nmi_evntsel_msr_to_bit(msr);
 166         BUG_ON(counter > NMI_MAX_COUNTER_BITS);
 167
 168         clear_bit(counter, &__get_cpu_var(evntsel_nmi_owner));
 169 }
 170
 171 static __cpuinit inline int nmi_known_cpu(void)
 172 {
 173         switch (boot_cpu_data.x86_vendor) {
 174         case X86_VENDOR_AMD:
 175                 return boot_cpu_data.x86 == 15;
 176         case X86_VENDOR_INTEL:
 177                 if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
 178                         return 1;
 179                 else
 180                         return (boot_cpu_data.x86 == 15);
 181         }
 182         return 0;
 183 }
 184
 185 /* Run after command line and cpu_init init, but before all other checks */
 186 void nmi_watchdog_default(void)
 187 {
 188         if (nmi_watchdog != NMI_DEFAULT)
 189                 return;
 190         if (nmi_known_cpu())
 191                 nmi_watchdog = NMI_LOCAL_APIC;
 192         else
 193                 nmi_watchdog = NMI_IO_APIC;
 194 }
 195
 196 static int endflag __initdata = 0;
 197
 198 #ifdef CONFIG_SMP
 199 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
 200  * the CPU is idle. To make sure the NMI watchdog really ticks on all
 201  * CPUs during the test make them busy.
 202  */
 203 static __init void nmi_cpu_busy(void *data)
 204 {
 205         local_irq_enable_in_hardirq();
 206         /* Intentionally don't use cpu_relax here. This is
 207            to make sure that the performance counter really ticks,
 208            even if there is a simulator or similar that catches the
 209            pause instruction. On a real HT machine this is fine because
 210            all other CPUs are busy with "useless" delay loops and don't
 211            care if they get somewhat less cycles. */
 212         while (endflag == 0)
 213                 mb();
 214 }
 215 #endif
 216
 217 int __init check_nmi_watchdog (void)
 218 {
 219         int *counts;
 220         int cpu;
 221
 222         if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DEFAULT))
 223                 return 0;
 224
 225         if (!atomic_read(&nmi_active))
 226                 return 0;
 227
 228         counts = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
 229         if (!counts)
 230                 return -1;
 231
 232         printk(KERN_INFO "testing NMI watchdog ... ");
 233
 234 #ifdef CONFIG_SMP
 235         if (nmi_watchdog == NMI_LOCAL_APIC)
 236                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
 237 #endif
 238
 239         for (cpu = 0; cpu < NR_CPUS; cpu++)
 240                 counts[cpu] = cpu_pda(cpu)->__nmi_count;
 241         local_irq_enable();
 242         mdelay((10*1000)/nmi_hz); // wait 10 ticks
 243
 244         for_each_online_cpu(cpu) {
 245                 if (!per_cpu(nmi_watchdog_ctlblk, cpu).enabled)
 246                         continue;
 247                 if (cpu_pda(cpu)->__nmi_count - counts[cpu] <= 5) {
 248                         printk("CPU#%d: NMI appears to be stuck (%d->%d)!\n",
 249                                cpu,
 250                                counts[cpu],
 251                                cpu_pda(cpu)->__nmi_count);
 252                         per_cpu(nmi_watchdog_ctlblk, cpu).enabled = 0;
 253                         atomic_dec(&nmi_active);
 254                 }
 255         }
 256         if (!atomic_read(&nmi_active)) {
 257                 kfree(counts);
 258                 atomic_set(&nmi_active, -1);
 259                 endflag = 1;
 260                 return -1;
 261         }
 262         endflag = 1;
 263         printk("OK.\n");
 264
 265         /* now that we know it works we can reduce NMI frequency to
 266            something more reasonable; makes a difference in some configs */
 267         if (nmi_watchdog == NMI_LOCAL_APIC) {
 268                 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 269
 270                 nmi_hz = 1;
 271                 /*
 272                  * On Intel CPUs with ARCH_PERFMON only 32 bits in the counter
 273                  * are writable, with higher bits sign extending from bit 31.
 274                  * So, we can only program the counter with 31 bit values and
 275                  * 32nd bit should be 1, for 33.. to be 1.
 276                  * Find the appropriate nmi_hz
 277                  */
 278                 if (wd->perfctr_msr == MSR_ARCH_PERFMON_PERFCTR0 &&
 279                         ((u64)cpu_khz * 1000) > 0x7fffffffULL) {
 280                         nmi_hz = ((u64)cpu_khz * 1000) / 0x7fffffffUL + 1;
 281                 }
 282         }
 283
 284         kfree(counts);
 285         return 0;
 286 }
 287
 288 int __init setup_nmi_watchdog(char *str)
 289 {
 290         int nmi;
 291
 292         if (!strncmp(str,"panic",5)) {
 293                 panic_on_timeout = 1;
 294                 str = strchr(str, ',');
 295                 if (!str)
 296                         return 1;
 297                 ++str;
 298         }
 299
 300         get_option(&str, &nmi);
 301
 302         if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
 303                 return 0;
 304
 305         if ((nmi == NMI_LOCAL_APIC) && (nmi_known_cpu() == 0))
 306                 return 0;  /* no lapic support */
 307         nmi_watchdog = nmi;
 308         return 1;
 309 }
 310
 311 __setup("nmi_watchdog=", setup_nmi_watchdog);
 312
 313 static void disable_lapic_nmi_watchdog(void)
 314 {
 315         BUG_ON(nmi_watchdog != NMI_LOCAL_APIC);
 316
 317         if (atomic_read(&nmi_active) <= 0)
 318                 return;
 319
 320         on_each_cpu(stop_apic_nmi_watchdog, NULL, 0, 1);
 321
 322         BUG_ON(atomic_read(&nmi_active) != 0);
 323 }
 324
 325 static void enable_lapic_nmi_watchdog(void)
 326 {
 327         BUG_ON(nmi_watchdog != NMI_LOCAL_APIC);
 328
 329         /* are we already enabled */
 330         if (atomic_read(&nmi_active) != 0)
 331                 return;
 332
 333         /* are we lapic aware */
 334         if (nmi_known_cpu() <= 0)
 335                 return;
 336
 337         on_each_cpu(setup_apic_nmi_watchdog, NULL, 0, 1);
 338         touch_nmi_watchdog();
 339 }
 340
 341 void disable_timer_nmi_watchdog(void)
 342 {
 343         BUG_ON(nmi_watchdog != NMI_IO_APIC);
 344
 345         if (atomic_read(&nmi_active) <= 0)
 346                 return;
 347
 348         disable_irq(0);
 349         on_each_cpu(stop_apic_nmi_watchdog, NULL, 0, 1);
 350
 351         BUG_ON(atomic_read(&nmi_active) != 0);
 352 }
 353
 354 void enable_timer_nmi_watchdog(void)
 355 {
 356         BUG_ON(nmi_watchdog != NMI_IO_APIC);
 357
 358         if (atomic_read(&nmi_active) == 0) {
 359                 touch_nmi_watchdog();
 360                 on_each_cpu(setup_apic_nmi_watchdog, NULL, 0, 1);
 361                 enable_irq(0);
 362         }
 363 }
 364
 365 #ifdef CONFIG_PM
 366
 367 static int nmi_pm_active; /* nmi_active before suspend */
 368
 369 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 370 {
 371         /* only CPU0 goes here, other CPUs should be offline */
 372         nmi_pm_active = atomic_read(&nmi_active);
 373         stop_apic_nmi_watchdog(NULL);
 374         BUG_ON(atomic_read(&nmi_active) != 0);
 375         return 0;
 376 }
 377
 378 static int lapic_nmi_resume(struct sys_device *dev)
 379 {
 380         /* only CPU0 goes here, other CPUs should be offline */
 381         if (nmi_pm_active > 0) {
 382                 setup_apic_nmi_watchdog(NULL);
 383                 touch_nmi_watchdog();
 384         }
 385         return 0;
 386 }
 387
 388 static struct sysdev_class nmi_sysclass = {
 389         set_kset_name("lapic_nmi"),
 390         .resume         = lapic_nmi_resume,
 391         .suspend        = lapic_nmi_suspend,
 392 };
 393
 394 static struct sys_device device_lapic_nmi = {
 395         .id             = 0,
 396         .cls    = &nmi_sysclass,
 397 };
 398
 399 static int __init init_lapic_nmi_sysfs(void)
 400 {
 401         int error;
 402
 403         /* should really be a BUG_ON but b/c this is an
 404          * init call, it just doesn't work.  -dcz
 405          */
 406         if (nmi_watchdog != NMI_LOCAL_APIC)
 407                 return 0;
 408
 409         if ( atomic_read(&nmi_active) < 0 )
 410                 return 0;
 411
 412         error = sysdev_class_register(&nmi_sysclass);
 413         if (!error)
 414                 error = sysdev_register(&device_lapic_nmi);
 415         return error;
 416 }
 417 /* must come after the local APIC's device_initcall() */
 418 late_initcall(init_lapic_nmi_sysfs);
 419
 420 #endif  /* CONFIG_PM */
 421
 422 /*
 423  * Activate the NMI watchdog via the local APIC.
 424  * Original code written by Keith Owens.
 425  */
 426
 427 /* Note that these events don't tick when the CPU idles. This means
 428    the frequency varies with CPU load. */
 429
 430 #define K7_EVNTSEL_ENABLE       (1 << 22)
 431 #define K7_EVNTSEL_INT          (1 << 20)
 432 #define K7_EVNTSEL_OS           (1 << 17)
 433 #define K7_EVNTSEL_USR          (1 << 16)
 434 #define K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING    0x76
 435 #define K7_NMI_EVENT            K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING
 436
 437 static int setup_k7_watchdog(void)
 438 {
 439         unsigned int perfctr_msr, evntsel_msr;
 440         unsigned int evntsel;
 441         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 442
 443         perfctr_msr = MSR_K7_PERFCTR0;
 444         evntsel_msr = MSR_K7_EVNTSEL0;
 445         if (!reserve_perfctr_nmi(perfctr_msr))
 446                 goto fail;
 447
 448         if (!reserve_evntsel_nmi(evntsel_msr))
 449                 goto fail1;
 450
 451         /* Simulator may not support it */
 452         if (checking_wrmsrl(evntsel_msr, 0UL))
 453                 goto fail2;
 454         wrmsrl(perfctr_msr, 0UL);
 455
 456         evntsel = K7_EVNTSEL_INT
 457                 | K7_EVNTSEL_OS
 458                 | K7_EVNTSEL_USR
 459                 | K7_NMI_EVENT;
 460
 461         /* setup the timer */
 462         wrmsr(evntsel_msr, evntsel, 0);
 463         wrmsrl(perfctr_msr, -((u64)cpu_khz * 1000 / nmi_hz));
 464         apic_write(APIC_LVTPC, APIC_DM_NMI);
 465         evntsel |= K7_EVNTSEL_ENABLE;
 466         wrmsr(evntsel_msr, evntsel, 0);
 467
 468         wd->perfctr_msr = perfctr_msr;
 469         wd->evntsel_msr = evntsel_msr;
 470         wd->cccr_msr = 0;  //unused
 471         wd->check_bit = 1ULL<<63;
 472         return 1;
 473 fail2:
 474         release_evntsel_nmi(evntsel_msr);
 475 fail1:
 476         release_perfctr_nmi(perfctr_msr);
 477 fail:
 478         return 0;
 479 }
 480
 481 static void stop_k7_watchdog(void)
 482 {
 483         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 484
 485         wrmsr(wd->evntsel_msr, 0, 0);
 486
 487         release_evntsel_nmi(wd->evntsel_msr);
 488         release_perfctr_nmi(wd->perfctr_msr);
 489 }
 490
 491 /* Note that these events don't tick when the CPU idles. This means
 492    the frequency varies with CPU load. */
 493
 494 #define MSR_P4_MISC_ENABLE_PERF_AVAIL   (1<<7)
 495 #define P4_ESCR_EVENT_SELECT(N) ((N)<<25)
 496 #define P4_ESCR_OS              (1<<3)
 497 #define P4_ESCR_USR             (1<<2)
 498 #define P4_CCCR_OVF_PMI0        (1<<26)
 499 #define P4_CCCR_OVF_PMI1        (1<<27)
 500 #define P4_CCCR_THRESHOLD(N)    ((N)<<20)
 501 #define P4_CCCR_COMPLEMENT      (1<<19)
 502 #define P4_CCCR_COMPARE         (1<<18)
 503 #define P4_CCCR_REQUIRED        (3<<16)
 504 #define P4_CCCR_ESCR_SELECT(N)  ((N)<<13)
 505 #define P4_CCCR_ENABLE          (1<<12)
 506 #define P4_CCCR_OVF             (1<<31)
 507 /* Set up IQ_COUNTER0 to behave like a clock, by having IQ_CCCR0 filter
 508    CRU_ESCR0 (with any non-null event selector) through a complemented
 509    max threshold. [IA32-Vol3, Section 14.9.9] */
 510
 511 static int setup_p4_watchdog(void)
 512 {
 513         unsigned int perfctr_msr, evntsel_msr, cccr_msr;
 514         unsigned int evntsel, cccr_val;
 515         unsigned int misc_enable, dummy;
 516         unsigned int ht_num;
 517         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 518
 519         rdmsr(MSR_IA32_MISC_ENABLE, misc_enable, dummy);
 520         if (!(misc_enable & MSR_P4_MISC_ENABLE_PERF_AVAIL))
 521                 return 0;
 522
 523 #ifdef CONFIG_SMP
 524         /* detect which hyperthread we are on */
 525         if (smp_num_siblings == 2) {
 526                 unsigned int ebx, apicid;
 527
 528                 ebx = cpuid_ebx(1);
 529                 apicid = (ebx >> 24) & 0xff;
 530                 ht_num = apicid & 1;
 531         } else
 532 #endif
 533                 ht_num = 0;
 534
 535         /* performance counters are shared resources
 536          * assign each hyperthread its own set
 537          * (re-use the ESCR0 register, seems safe
 538          * and keeps the cccr_val the same)
 539          */
 540         if (!ht_num) {
 541                 /* logical cpu 0 */
 542                 perfctr_msr = MSR_P4_IQ_PERFCTR0;
 543                 evntsel_msr = MSR_P4_CRU_ESCR0;
 544                 cccr_msr = MSR_P4_IQ_CCCR0;
 545                 cccr_val = P4_CCCR_OVF_PMI0 | P4_CCCR_ESCR_SELECT(4);
 546         } else {
 547                 /* logical cpu 1 */
 548                 perfctr_msr = MSR_P4_IQ_PERFCTR1;
 549                 evntsel_msr = MSR_P4_CRU_ESCR0;
 550                 cccr_msr = MSR_P4_IQ_CCCR1;
 551                 cccr_val = P4_CCCR_OVF_PMI1 | P4_CCCR_ESCR_SELECT(4);
 552         }
 553
 554         if (!reserve_perfctr_nmi(perfctr_msr))
 555                 goto fail;
 556
 557         if (!reserve_evntsel_nmi(evntsel_msr))
 558                 goto fail1;
 559
 560         evntsel = P4_ESCR_EVENT_SELECT(0x3F)
 561                 | P4_ESCR_OS
 562                 | P4_ESCR_USR;
 563
 564         cccr_val |= P4_CCCR_THRESHOLD(15)
 565                  | P4_CCCR_COMPLEMENT
 566                  | P4_CCCR_COMPARE
 567                  | P4_CCCR_REQUIRED;
 568
 569         wrmsr(evntsel_msr, evntsel, 0);
 570         wrmsr(cccr_msr, cccr_val, 0);
 571         wrmsrl(perfctr_msr, -((u64)cpu_khz * 1000 / nmi_hz));
 572         apic_write(APIC_LVTPC, APIC_DM_NMI);
 573         cccr_val |= P4_CCCR_ENABLE;
 574         wrmsr(cccr_msr, cccr_val, 0);
 575
 576         wd->perfctr_msr = perfctr_msr;
 577         wd->evntsel_msr = evntsel_msr;
 578         wd->cccr_msr = cccr_msr;
 579         wd->check_bit = 1ULL<<39;
 580         return 1;
 581 fail1:
 582         release_perfctr_nmi(perfctr_msr);
 583 fail:
 584         return 0;
 585 }
 586
 587 static void stop_p4_watchdog(void)
 588 {
 589         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 590
 591         wrmsr(wd->cccr_msr, 0, 0);
 592         wrmsr(wd->evntsel_msr, 0, 0);
 593
 594         release_evntsel_nmi(wd->evntsel_msr);
 595         release_perfctr_nmi(wd->perfctr_msr);
 596 }
 597
 598 #define ARCH_PERFMON_NMI_EVENT_SEL      ARCH_PERFMON_UNHALTED_CORE_CYCLES_SEL
 599 #define ARCH_PERFMON_NMI_EVENT_UMASK    ARCH_PERFMON_UNHALTED_CORE_CYCLES_UMASK
 600
 601 static int setup_intel_arch_watchdog(void)
 602 {
 603         unsigned int ebx;
 604         union cpuid10_eax eax;
 605         unsigned int unused;
 606         unsigned int perfctr_msr, evntsel_msr;
 607         unsigned int evntsel;
 608         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 609
 610         /*
 611          * Check whether the Architectural PerfMon supports
 612          * Unhalted Core Cycles Event or not.
 613          * NOTE: Corresponding bit = 0 in ebx indicates event present.
 614          */
 615         cpuid(10, &(eax.full), &ebx, &unused, &unused);
 616         if ((eax.split.mask_length < (ARCH_PERFMON_UNHALTED_CORE_CYCLES_INDEX+1)) ||
 617             (ebx & ARCH_PERFMON_UNHALTED_CORE_CYCLES_PRESENT))
 618                 goto fail;
 619
 620         perfctr_msr = MSR_ARCH_PERFMON_PERFCTR0;
 621         evntsel_msr = MSR_ARCH_PERFMON_EVENTSEL0;
 622
 623         if (!reserve_perfctr_nmi(perfctr_msr))
 624                 goto fail;
 625
 626         if (!reserve_evntsel_nmi(evntsel_msr))
 627                 goto fail1;
 628
 629         wrmsrl(perfctr_msr, 0UL);
 630
 631         evntsel = ARCH_PERFMON_EVENTSEL_INT
 632                 | ARCH_PERFMON_EVENTSEL_OS
 633                 | ARCH_PERFMON_EVENTSEL_USR
 634                 | ARCH_PERFMON_NMI_EVENT_SEL
 635                 | ARCH_PERFMON_NMI_EVENT_UMASK;
 636
 637         /* setup the timer */
 638         wrmsr(evntsel_msr, evntsel, 0);
 639         wrmsrl(perfctr_msr, -((u64)cpu_khz * 1000 / nmi_hz));
 640
 641         apic_write(APIC_LVTPC, APIC_DM_NMI);
 642         evntsel |= ARCH_PERFMON_EVENTSEL0_ENABLE;
 643         wrmsr(evntsel_msr, evntsel, 0);
 644
 645         wd->perfctr_msr = perfctr_msr;
 646         wd->evntsel_msr = evntsel_msr;
 647         wd->cccr_msr = 0;  //unused
 648         wd->check_bit = 1ULL << (eax.split.bit_width - 1);
 649         return 1;
 650 fail1:
 651         release_perfctr_nmi(perfctr_msr);
 652 fail:
 653         return 0;
 654 }
 655
 656 static void stop_intel_arch_watchdog(void)
 657 {
 658         unsigned int ebx;
 659         union cpuid10_eax eax;
 660         unsigned int unused;
 661         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 662
 663         /*
 664          * Check whether the Architectural PerfMon supports
 665          * Unhalted Core Cycles Event or not.
 666          * NOTE: Corresponding bit = 0 in ebx indicates event present.
 667          */
 668         cpuid(10, &(eax.full), &ebx, &unused, &unused);
 669         if ((eax.split.mask_length < (ARCH_PERFMON_UNHALTED_CORE_CYCLES_INDEX+1)) ||
 670             (ebx & ARCH_PERFMON_UNHALTED_CORE_CYCLES_PRESENT))
 671                 return;
 672
 673         wrmsr(wd->evntsel_msr, 0, 0);
 674
 675         release_evntsel_nmi(wd->evntsel_msr);
 676         release_perfctr_nmi(wd->perfctr_msr);
 677 }
 678
 679 void setup_apic_nmi_watchdog(void *unused)
 680 {
 681         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 682
 683         /* only support LOCAL and IO APICs for now */
 684         if ((nmi_watchdog != NMI_LOCAL_APIC) &&
 685             (nmi_watchdog != NMI_IO_APIC))
 686                 return;
 687
 688         if (wd->enabled == 1)
 689                 return;
 690
 691         /* cheap hack to support suspend/resume */
 692         /* if cpu0 is not active neither should the other cpus */
 693         if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
 694                 return;
 695
 696         if (nmi_watchdog == NMI_LOCAL_APIC) {
 697                 switch (boot_cpu_data.x86_vendor) {
 698                 case X86_VENDOR_AMD:
 699                         if (strstr(boot_cpu_data.x86_model_id, "Screwdriver"))
 700                                 return;
 701                         if (!setup_k7_watchdog())
 702                                 return;
 703                         break;
 704                 case X86_VENDOR_INTEL:
 705                         if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
 706                                 if (!setup_intel_arch_watchdog())
 707                                         return;
 708                                 break;
 709                         }
 710                         if (!setup_p4_watchdog())
 711                                 return;
 712                         break;
 713                 default:
 714                         return;
 715                 }
 716         }
 717         wd->enabled = 1;
 718         atomic_inc(&nmi_active);
 719 }
 720
 721 void stop_apic_nmi_watchdog(void *unused)
 722 {
 723         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 724
 725         /* only support LOCAL and IO APICs for now */
 726         if ((nmi_watchdog != NMI_LOCAL_APIC) &&
 727             (nmi_watchdog != NMI_IO_APIC))
 728                 return;
 729
 730         if (wd->enabled == 0)
 731                 return;
 732
 733         if (nmi_watchdog == NMI_LOCAL_APIC) {
 734                 switch (boot_cpu_data.x86_vendor) {
 735                 case X86_VENDOR_AMD:
 736                         if (strstr(boot_cpu_data.x86_model_id, "Screwdriver"))
 737                                 return;
 738                         stop_k7_watchdog();
 739                         break;
 740                 case X86_VENDOR_INTEL:
 741                         if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
 742                                 stop_intel_arch_watchdog();
 743                                 break;
 744                         }
 745                         stop_p4_watchdog();
 746                         break;
 747                 default:
 748                         return;
 749                 }
 750         }
 751         wd->enabled = 0;
 752         atomic_dec(&nmi_active);
 753 }
 754
 755 /*
 756  * the best way to detect whether a CPU has a 'hard lockup' problem
 757  * is to check it's local APIC timer IRQ counts. If they are not
 758  * changing then that CPU has some problem.
 759  *
 760  * as these watchdog NMI IRQs are generated on every CPU, we only
 761  * have to check the current processor.
 762  */
 763
 764 static DEFINE_PER_CPU(unsigned, last_irq_sum);
 765 static DEFINE_PER_CPU(local_t, alert_counter);
 766 static DEFINE_PER_CPU(int, nmi_touch);
 767
 768 void touch_nmi_watchdog (void)
 769 {
 770         if (nmi_watchdog > 0) {
 771                 unsigned cpu;
 772
 773                 /*
 774                  * Tell other CPUs to reset their alert counters. We cannot
 775                  * do it ourselves because the alert count increase is not
 776                  * atomic.
 777                  */
 778                 for_each_present_cpu (cpu)
 779                         per_cpu(nmi_touch, cpu) = 1;
 780         }
 781
 782         touch_softlockup_watchdog();
 783 }
 784
 785 int __kprobes nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
 786 {
 787         int sum;
 788         int touched = 0;
 789         int cpu = smp_processor_id();
 790         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 791         u64 dummy;
 792         int rc=0;
 793
 794         /* check for other users first */
 795         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 796                         == NOTIFY_STOP) {
 797                 rc = 1;
 798                 touched = 1;
 799         }
 800
 801         sum = read_pda(apic_timer_irqs);
 802         if (__get_cpu_var(nmi_touch)) {
 803                 __get_cpu_var(nmi_touch) = 0;
 804                 touched = 1;
 805         }
 806
 807         if (cpu_isset(cpu, backtrace_mask)) {
 808                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 809
 810                 spin_lock(&lock);
 811                 printk("NMI backtrace for cpu %d\n", cpu);
 812                 dump_stack();
 813                 spin_unlock(&lock);
 814                 cpu_clear(cpu, backtrace_mask);
 815         }
 816
 817 #ifdef CONFIG_X86_MCE
 818         /* Could check oops_in_progress here too, but it's safer
 819            not too */
 820         if (atomic_read(&mce_entry) > 0)
 821                 touched = 1;
 822 #endif
 823         /* if the apic timer isn't firing, this cpu isn't doing much */
 824         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
 825                 /*
 826                  * Ayiee, looks like this CPU is stuck ...
 827                  * wait a few IRQs (5 seconds) before doing the oops ...
 828                  */
 829                 local_inc(&__get_cpu_var(alert_counter));
 830                 if (local_read(&__get_cpu_var(alert_counter)) == 5*nmi_hz)
 831                         die_nmi("NMI Watchdog detected LOCKUP on CPU %d\n", regs,
 832                                 panic_on_timeout);
 833         } else {
 834                 __get_cpu_var(last_irq_sum) = sum;
 835                 local_set(&__get_cpu_var(alert_counter), 0);
 836         }
 837
 838         /* see if the nmi watchdog went off */
 839         if (wd->enabled) {
 840                 if (nmi_watchdog == NMI_LOCAL_APIC) {
 841                         rdmsrl(wd->perfctr_msr, dummy);
 842                         if (dummy & wd->check_bit){
 843                                 /* this wasn't a watchdog timer interrupt */
 844                                 goto done;
 845                         }
 846
 847                         /* only Intel uses the cccr msr */
 848                         if (wd->cccr_msr != 0) {
 849                                 /*
 850                                  * P4 quirks:
 851                                  * - An overflown perfctr will assert its interrupt
 852                                  *   until the OVF flag in its CCCR is cleared.
 853                                  * - LVTPC is masked on interrupt and must be
 854                                  *   unmasked by the LVTPC handler.
 855                                  */
 856                                 rdmsrl(wd->cccr_msr, dummy);
 857                                 dummy &= ~P4_CCCR_OVF;
 858                                 wrmsrl(wd->cccr_msr, dummy);
 859                                 apic_write(APIC_LVTPC, APIC_DM_NMI);
 860                         } else if (wd->perfctr_msr == MSR_ARCH_PERFMON_PERFCTR0) {
 861                                 /*
 862                                  * ArchPerfom/Core Duo needs to re-unmask
 863                                  * the apic vector
 864                                  */
 865                                 apic_write(APIC_LVTPC, APIC_DM_NMI);
 866                         }
 867                         /* start the cycle over again */
 868                         wrmsrl(wd->perfctr_msr, -((u64)cpu_khz * 1000 / nmi_hz));
 869                         rc = 1;
 870                 } else  if (nmi_watchdog == NMI_IO_APIC) {
 871                         /* don't know how to accurately check for this.
 872                          * just assume it was a watchdog timer interrupt
 873                          * This matches the old behaviour.
 874                          */
 875                         rc = 1;
 876                 } else
 877                         printk(KERN_WARNING "Unknown enabled NMI hardware?!\n");
 878         }
 879 done:
 880         return rc;
 881 }
 882
 883 asmlinkage __kprobes void do_nmi(struct pt_regs * regs, long error_code)
 884 {
 885         nmi_enter();
 886         add_pda(__nmi_count,1);
 887         default_do_nmi(regs);
 888         nmi_exit();
 889 }
 890
 891 int do_nmi_callback(struct pt_regs * regs, int cpu)
 892 {
 893 #ifdef CONFIG_SYSCTL
 894         if (unknown_nmi_panic)
 895                 return unknown_nmi_panic_callback(regs, cpu);
 896 #endif
 897         return 0;
 898 }
 899
 900 #ifdef CONFIG_SYSCTL
 901
 902 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 903 {
 904         unsigned char reason = get_nmi_reason();
 905         char buf[64];
 906
 907         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 908         die_nmi(buf, regs, 1);  /* Always panic here */
 909         return 0;
 910 }
 911
 912 /*
 913  * proc handler for /proc/sys/kernel/nmi
 914  */
 915 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 916                         void __user *buffer, size_t *length, loff_t *ppos)
 917 {
 918         int old_state;
 919
 920         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 921         old_state = nmi_watchdog_enabled;
 922         proc_dointvec(table, write, file, buffer, length, ppos);
 923         if (!!old_state == !!nmi_watchdog_enabled)
 924                 return 0;
 925
 926         if (atomic_read(&nmi_active) < 0) {
 927                 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
 928                 return -EIO;
 929         }
 930
 931         /* if nmi_watchdog is not set yet, then set it */
 932         nmi_watchdog_default();
 933
 934         if (nmi_watchdog == NMI_LOCAL_APIC) {
 935                 if (nmi_watchdog_enabled)
 936                         enable_lapic_nmi_watchdog();
 937                 else
 938                         disable_lapic_nmi_watchdog();
 939         } else {
 940                 printk( KERN_WARNING
 941                         "NMI watchdog doesn't know what hardware to touch\n");
 942                 return -EIO;
 943         }
 944         return 0;
 945 }
 946
 947 #endif
 948
 949 void __trigger_all_cpu_backtrace(void)
 950 {
 951         int i;
 952
 953         backtrace_mask = cpu_online_map;
 954         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 955         for (i = 0; i < 10 * 1000; i++) {
 956                 if (cpus_empty(backtrace_mask))
 957                         break;
 958                 mdelay(1);
 959         }
 960 }
 961
 962 EXPORT_SYMBOL(nmi_active);
 963 EXPORT_SYMBOL(nmi_watchdog);
 964 EXPORT_SYMBOL(avail_to_resrv_perfctr_nmi);
 965 EXPORT_SYMBOL(avail_to_resrv_perfctr_nmi_bit);
 966 EXPORT_SYMBOL(reserve_perfctr_nmi);
 967 EXPORT_SYMBOL(release_perfctr_nmi);
 968 EXPORT_SYMBOL(reserve_evntsel_nmi);
 969 EXPORT_SYMBOL(release_evntsel_nmi);
 970 EXPORT_SYMBOL(disable_timer_nmi_watchdog);
 971 EXPORT_SYMBOL(enable_timer_nmi_watchdog);
 972 EXPORT_SYMBOL(touch_nmi_watchdog);