git.oblomov.eu Git - linux-2.6/blob - arch/x86/kernel/cpu/perfctr-watchdog.c

   1 /* local apic based NMI watchdog for various CPUs.
   2    This file also handles reservation of performance counters for coordination
   3    with other users (like oprofile).
   4
   5    Note that these events normally don't tick when the CPU idles. This means
   6    the frequency varies with CPU load.
   7
   8    Original code for K7/P6 written by Keith Owens */
   9
  10 #include <linux/percpu.h>
  11 #include <linux/module.h>
  12 #include <linux/kernel.h>
  13 #include <linux/bitops.h>
  14 #include <linux/smp.h>
  15 #include <linux/nmi.h>
  16 #include <asm/apic.h>
  17 #include <asm/intel_arch_perfmon.h>
  18
  19 struct nmi_watchdog_ctlblk {
  20         unsigned int cccr_msr;
  21         unsigned int perfctr_msr;  /* the MSR to reset in NMI handler */
  22         unsigned int evntsel_msr;  /* the MSR to select the events to handle */
  23 };
  24
  25 /* Interface defining a CPU specific perfctr watchdog */
  26 struct wd_ops {
  27         int (*reserve)(void);
  28         void (*unreserve)(void);
  29         int (*setup)(unsigned nmi_hz);
  30         void (*rearm)(struct nmi_watchdog_ctlblk *wd, unsigned nmi_hz);
  31         void (*stop)(void);
  32         unsigned perfctr;
  33         unsigned evntsel;
  34         u64 checkbit;
  35 };
  36
  37 static const struct wd_ops *wd_ops;
  38
  39 /* this number is calculated from Intel's MSR_P4_CRU_ESCR5 register and it's
  40  * offset from MSR_P4_BSU_ESCR0.  It will be the max for all platforms (for now)
  41  */
  42 #define NMI_MAX_COUNTER_BITS 66
  43
  44 /* perfctr_nmi_owner tracks the ownership of the perfctr registers:
  45  * evtsel_nmi_owner tracks the ownership of the event selection
  46  * - different performance counters/ event selection may be reserved for
  47  *   different subsystems this reservation system just tries to coordinate
  48  *   things a little
  49  */
  50 static DECLARE_BITMAP(perfctr_nmi_owner, NMI_MAX_COUNTER_BITS);
  51 static DECLARE_BITMAP(evntsel_nmi_owner, NMI_MAX_COUNTER_BITS);
  52
  53 static DEFINE_PER_CPU(struct nmi_watchdog_ctlblk, nmi_watchdog_ctlblk);
  54
  55 /* converts an msr to an appropriate reservation bit */
  56 static inline unsigned int nmi_perfctr_msr_to_bit(unsigned int msr)
  57 {
  58         /* returns the bit offset of the performance counter register */
  59         switch (boot_cpu_data.x86_vendor) {
  60         case X86_VENDOR_AMD:
  61                 return (msr - MSR_K7_PERFCTR0);
  62         case X86_VENDOR_INTEL:
  63                 if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
  64                         return (msr - MSR_ARCH_PERFMON_PERFCTR0);
  65
  66                 switch (boot_cpu_data.x86) {
  67                 case 6:
  68                         return (msr - MSR_P6_PERFCTR0);
  69                 case 15:
  70                         return (msr - MSR_P4_BPU_PERFCTR0);
  71                 }
  72         }
  73         return 0;
  74 }
  75
  76 /* converts an msr to an appropriate reservation bit */
  77 /* returns the bit offset of the event selection register */
  78 static inline unsigned int nmi_evntsel_msr_to_bit(unsigned int msr)
  79 {
  80         /* returns the bit offset of the event selection register */
  81         switch (boot_cpu_data.x86_vendor) {
  82         case X86_VENDOR_AMD:
  83                 return (msr - MSR_K7_EVNTSEL0);
  84         case X86_VENDOR_INTEL:
  85                 if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
  86                         return (msr - MSR_ARCH_PERFMON_EVENTSEL0);
  87
  88                 switch (boot_cpu_data.x86) {
  89                 case 6:
  90                         return (msr - MSR_P6_EVNTSEL0);
  91                 case 15:
  92                         return (msr - MSR_P4_BSU_ESCR0);
  93                 }
  94         }
  95         return 0;
  96
  97 }
  98
  99 /* checks for a bit availability (hack for oprofile) */
 100 int avail_to_resrv_perfctr_nmi_bit(unsigned int counter)
 101 {
 102         BUG_ON(counter > NMI_MAX_COUNTER_BITS);
 103
 104         return (!test_bit(counter, perfctr_nmi_owner));
 105 }
 106
 107 /* checks the an msr for availability */
 108 int avail_to_resrv_perfctr_nmi(unsigned int msr)
 109 {
 110         unsigned int counter;
 111
 112         counter = nmi_perfctr_msr_to_bit(msr);
 113         BUG_ON(counter > NMI_MAX_COUNTER_BITS);
 114
 115         return (!test_bit(counter, perfctr_nmi_owner));
 116 }
 117
 118 int reserve_perfctr_nmi(unsigned int msr)
 119 {
 120         unsigned int counter;
 121
 122         counter = nmi_perfctr_msr_to_bit(msr);
 123         /* register not managed by the allocator? */
 124         if (counter > NMI_MAX_COUNTER_BITS)
 125                 return 1;
 126
 127         if (!test_and_set_bit(counter, perfctr_nmi_owner))
 128                 return 1;
 129         return 0;
 130 }
 131
 132 void release_perfctr_nmi(unsigned int msr)
 133 {
 134         unsigned int counter;
 135
 136         counter = nmi_perfctr_msr_to_bit(msr);
 137         /* register not managed by the allocator? */
 138         if (counter > NMI_MAX_COUNTER_BITS)
 139                 return;
 140
 141         clear_bit(counter, perfctr_nmi_owner);
 142 }
 143
 144 int reserve_evntsel_nmi(unsigned int msr)
 145 {
 146         unsigned int counter;
 147
 148         counter = nmi_evntsel_msr_to_bit(msr);
 149         /* register not managed by the allocator? */
 150         if (counter > NMI_MAX_COUNTER_BITS)
 151                 return 1;
 152
 153         if (!test_and_set_bit(counter, 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         /* register not managed by the allocator? */
 164         if (counter > NMI_MAX_COUNTER_BITS)
 165                 return;
 166
 167         clear_bit(counter, evntsel_nmi_owner);
 168 }
 169
 170 EXPORT_SYMBOL(avail_to_resrv_perfctr_nmi_bit);
 171 EXPORT_SYMBOL(reserve_perfctr_nmi);
 172 EXPORT_SYMBOL(release_perfctr_nmi);
 173 EXPORT_SYMBOL(reserve_evntsel_nmi);
 174 EXPORT_SYMBOL(release_evntsel_nmi);
 175
 176 void disable_lapic_nmi_watchdog(void)
 177 {
 178         BUG_ON(nmi_watchdog != NMI_LOCAL_APIC);
 179
 180         if (atomic_read(&nmi_active) <= 0)
 181                 return;
 182
 183         on_each_cpu(stop_apic_nmi_watchdog, NULL, 0, 1);
 184         wd_ops->unreserve();
 185
 186         BUG_ON(atomic_read(&nmi_active) != 0);
 187 }
 188
 189 void enable_lapic_nmi_watchdog(void)
 190 {
 191         BUG_ON(nmi_watchdog != NMI_LOCAL_APIC);
 192
 193         /* are we already enabled */
 194         if (atomic_read(&nmi_active) != 0)
 195                 return;
 196
 197         /* are we lapic aware */
 198         if (!wd_ops)
 199                 return;
 200         if (!wd_ops->reserve()) {
 201                 printk(KERN_ERR "NMI watchdog: cannot reserve perfctrs\n");
 202                 return;
 203         }
 204
 205         on_each_cpu(setup_apic_nmi_watchdog, NULL, 0, 1);
 206         touch_nmi_watchdog();
 207 }
 208
 209 /*
 210  * Activate the NMI watchdog via the local APIC.
 211  */
 212
 213 static unsigned int adjust_for_32bit_ctr(unsigned int hz)
 214 {
 215         u64 counter_val;
 216         unsigned int retval = hz;
 217
 218         /*
 219          * On Intel CPUs with P6/ARCH_PERFMON only 32 bits in the counter
 220          * are writable, with higher bits sign extending from bit 31.
 221          * So, we can only program the counter with 31 bit values and
 222          * 32nd bit should be 1, for 33.. to be 1.
 223          * Find the appropriate nmi_hz
 224          */
 225         counter_val = (u64)cpu_khz * 1000;
 226         do_div(counter_val, retval);
 227         if (counter_val > 0x7fffffffULL) {
 228                 u64 count = (u64)cpu_khz * 1000;
 229                 do_div(count, 0x7fffffffUL);
 230                 retval = count + 1;
 231         }
 232         return retval;
 233 }
 234
 235 static void
 236 write_watchdog_counter(unsigned int perfctr_msr, const char *descr, unsigned nmi_hz)
 237 {
 238         u64 count = (u64)cpu_khz * 1000;
 239
 240         do_div(count, nmi_hz);
 241         if(descr)
 242                 Dprintk("setting %s to -0x%08Lx\n", descr, count);
 243         wrmsrl(perfctr_msr, 0 - count);
 244 }
 245
 246 static void write_watchdog_counter32(unsigned int perfctr_msr,
 247                 const char *descr, unsigned nmi_hz)
 248 {
 249         u64 count = (u64)cpu_khz * 1000;
 250
 251         do_div(count, nmi_hz);
 252         if(descr)
 253                 Dprintk("setting %s to -0x%08Lx\n", descr, count);
 254         wrmsr(perfctr_msr, (u32)(-count), 0);
 255 }
 256
 257 /* AMD K7/K8/Family10h/Family11h support. AMD keeps this interface
 258    nicely stable so there is not much variety */
 259
 260 #define K7_EVNTSEL_ENABLE       (1 << 22)
 261 #define K7_EVNTSEL_INT          (1 << 20)
 262 #define K7_EVNTSEL_OS           (1 << 17)
 263 #define K7_EVNTSEL_USR          (1 << 16)
 264 #define K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING    0x76
 265 #define K7_NMI_EVENT            K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING
 266
 267 static int setup_k7_watchdog(unsigned nmi_hz)
 268 {
 269         unsigned int perfctr_msr, evntsel_msr;
 270         unsigned int evntsel;
 271         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 272
 273         perfctr_msr = wd_ops->perfctr;
 274         evntsel_msr = wd_ops->evntsel;
 275
 276         wrmsrl(perfctr_msr, 0UL);
 277
 278         evntsel = K7_EVNTSEL_INT
 279                 | K7_EVNTSEL_OS
 280                 | K7_EVNTSEL_USR
 281                 | K7_NMI_EVENT;
 282
 283         /* setup the timer */
 284         wrmsr(evntsel_msr, evntsel, 0);
 285         write_watchdog_counter(perfctr_msr, "K7_PERFCTR0",nmi_hz);
 286         apic_write(APIC_LVTPC, APIC_DM_NMI);
 287         evntsel |= K7_EVNTSEL_ENABLE;
 288         wrmsr(evntsel_msr, evntsel, 0);
 289
 290         wd->perfctr_msr = perfctr_msr;
 291         wd->evntsel_msr = evntsel_msr;
 292         wd->cccr_msr = 0;  //unused
 293         return 1;
 294 }
 295
 296 static void single_msr_stop_watchdog(void)
 297 {
 298         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 299
 300         wrmsr(wd->evntsel_msr, 0, 0);
 301 }
 302
 303 static int single_msr_reserve(void)
 304 {
 305         if (!reserve_perfctr_nmi(wd_ops->perfctr))
 306                 return 0;
 307
 308         if (!reserve_evntsel_nmi(wd_ops->evntsel)) {
 309                 release_perfctr_nmi(wd_ops->perfctr);
 310                 return 0;
 311         }
 312         return 1;
 313 }
 314
 315 static void single_msr_unreserve(void)
 316 {
 317         release_evntsel_nmi(wd_ops->evntsel);
 318         release_perfctr_nmi(wd_ops->perfctr);
 319 }
 320
 321 static void single_msr_rearm(struct nmi_watchdog_ctlblk *wd, unsigned nmi_hz)
 322 {
 323         /* start the cycle over again */
 324         write_watchdog_counter(wd->perfctr_msr, NULL, nmi_hz);
 325 }
 326
 327 static const struct wd_ops k7_wd_ops = {
 328         .reserve = single_msr_reserve,
 329         .unreserve = single_msr_unreserve,
 330         .setup = setup_k7_watchdog,
 331         .rearm = single_msr_rearm,
 332         .stop = single_msr_stop_watchdog,
 333         .perfctr = MSR_K7_PERFCTR0,
 334         .evntsel = MSR_K7_EVNTSEL0,
 335         .checkbit = 1ULL<<47,
 336 };
 337
 338 /* Intel Model 6 (PPro+,P2,P3,P-M,Core1) */
 339
 340 #define P6_EVNTSEL0_ENABLE      (1 << 22)
 341 #define P6_EVNTSEL_INT          (1 << 20)
 342 #define P6_EVNTSEL_OS           (1 << 17)
 343 #define P6_EVNTSEL_USR          (1 << 16)
 344 #define P6_EVENT_CPU_CLOCKS_NOT_HALTED  0x79
 345 #define P6_NMI_EVENT            P6_EVENT_CPU_CLOCKS_NOT_HALTED
 346
 347 static int setup_p6_watchdog(unsigned nmi_hz)
 348 {
 349         unsigned int perfctr_msr, evntsel_msr;
 350         unsigned int evntsel;
 351         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 352
 353         perfctr_msr = wd_ops->perfctr;
 354         evntsel_msr = wd_ops->evntsel;
 355
 356         /* KVM doesn't implement this MSR */
 357         if (wrmsr_safe(perfctr_msr, 0, 0) < 0)
 358                 return 0;
 359
 360         evntsel = P6_EVNTSEL_INT
 361                 | P6_EVNTSEL_OS
 362                 | P6_EVNTSEL_USR
 363                 | P6_NMI_EVENT;
 364
 365         /* setup the timer */
 366         wrmsr(evntsel_msr, evntsel, 0);
 367         nmi_hz = adjust_for_32bit_ctr(nmi_hz);
 368         write_watchdog_counter32(perfctr_msr, "P6_PERFCTR0",nmi_hz);
 369         apic_write(APIC_LVTPC, APIC_DM_NMI);
 370         evntsel |= P6_EVNTSEL0_ENABLE;
 371         wrmsr(evntsel_msr, evntsel, 0);
 372
 373         wd->perfctr_msr = perfctr_msr;
 374         wd->evntsel_msr = evntsel_msr;
 375         wd->cccr_msr = 0;  //unused
 376         return 1;
 377 }
 378
 379 static void p6_rearm(struct nmi_watchdog_ctlblk *wd, unsigned nmi_hz)
 380 {
 381         /* P6 based Pentium M need to re-unmask
 382          * the apic vector but it doesn't hurt
 383          * other P6 variant.
 384          * ArchPerfom/Core Duo also needs this */
 385         apic_write(APIC_LVTPC, APIC_DM_NMI);
 386         /* P6/ARCH_PERFMON has 32 bit counter write */
 387         write_watchdog_counter32(wd->perfctr_msr, NULL,nmi_hz);
 388 }
 389
 390 static const struct wd_ops p6_wd_ops = {
 391         .reserve = single_msr_reserve,
 392         .unreserve = single_msr_unreserve,
 393         .setup = setup_p6_watchdog,
 394         .rearm = p6_rearm,
 395         .stop = single_msr_stop_watchdog,
 396         .perfctr = MSR_P6_PERFCTR0,
 397         .evntsel = MSR_P6_EVNTSEL0,
 398         .checkbit = 1ULL<<39,
 399 };
 400
 401 /* Intel P4 performance counters. By far the most complicated of all. */
 402
 403 #define MSR_P4_MISC_ENABLE_PERF_AVAIL   (1<<7)
 404 #define P4_ESCR_EVENT_SELECT(N) ((N)<<25)
 405 #define P4_ESCR_OS              (1<<3)
 406 #define P4_ESCR_USR             (1<<2)
 407 #define P4_CCCR_OVF_PMI0        (1<<26)
 408 #define P4_CCCR_OVF_PMI1        (1<<27)
 409 #define P4_CCCR_THRESHOLD(N)    ((N)<<20)
 410 #define P4_CCCR_COMPLEMENT      (1<<19)
 411 #define P4_CCCR_COMPARE         (1<<18)
 412 #define P4_CCCR_REQUIRED        (3<<16)
 413 #define P4_CCCR_ESCR_SELECT(N)  ((N)<<13)
 414 #define P4_CCCR_ENABLE          (1<<12)
 415 #define P4_CCCR_OVF             (1<<31)
 416
 417 /* Set up IQ_COUNTER0 to behave like a clock, by having IQ_CCCR0 filter
 418    CRU_ESCR0 (with any non-null event selector) through a complemented
 419    max threshold. [IA32-Vol3, Section 14.9.9] */
 420
 421 static int setup_p4_watchdog(unsigned nmi_hz)
 422 {
 423         unsigned int perfctr_msr, evntsel_msr, cccr_msr;
 424         unsigned int evntsel, cccr_val;
 425         unsigned int misc_enable, dummy;
 426         unsigned int ht_num;
 427         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 428
 429         rdmsr(MSR_IA32_MISC_ENABLE, misc_enable, dummy);
 430         if (!(misc_enable & MSR_P4_MISC_ENABLE_PERF_AVAIL))
 431                 return 0;
 432
 433 #ifdef CONFIG_SMP
 434         /* detect which hyperthread we are on */
 435         if (smp_num_siblings == 2) {
 436                 unsigned int ebx, apicid;
 437
 438                 ebx = cpuid_ebx(1);
 439                 apicid = (ebx >> 24) & 0xff;
 440                 ht_num = apicid & 1;
 441         } else
 442 #endif
 443                 ht_num = 0;
 444
 445         /* performance counters are shared resources
 446          * assign each hyperthread its own set
 447          * (re-use the ESCR0 register, seems safe
 448          * and keeps the cccr_val the same)
 449          */
 450         if (!ht_num) {
 451                 /* logical cpu 0 */
 452                 perfctr_msr = MSR_P4_IQ_PERFCTR0;
 453                 evntsel_msr = MSR_P4_CRU_ESCR0;
 454                 cccr_msr = MSR_P4_IQ_CCCR0;
 455                 cccr_val = P4_CCCR_OVF_PMI0 | P4_CCCR_ESCR_SELECT(4);
 456         } else {
 457                 /* logical cpu 1 */
 458                 perfctr_msr = MSR_P4_IQ_PERFCTR1;
 459                 evntsel_msr = MSR_P4_CRU_ESCR0;
 460                 cccr_msr = MSR_P4_IQ_CCCR1;
 461                 cccr_val = P4_CCCR_OVF_PMI1 | P4_CCCR_ESCR_SELECT(4);
 462         }
 463
 464         evntsel = P4_ESCR_EVENT_SELECT(0x3F)
 465                 | P4_ESCR_OS
 466                 | P4_ESCR_USR;
 467
 468         cccr_val |= P4_CCCR_THRESHOLD(15)
 469                  | P4_CCCR_COMPLEMENT
 470                  | P4_CCCR_COMPARE
 471                  | P4_CCCR_REQUIRED;
 472
 473         wrmsr(evntsel_msr, evntsel, 0);
 474         wrmsr(cccr_msr, cccr_val, 0);
 475         write_watchdog_counter(perfctr_msr, "P4_IQ_COUNTER0", nmi_hz);
 476         apic_write(APIC_LVTPC, APIC_DM_NMI);
 477         cccr_val |= P4_CCCR_ENABLE;
 478         wrmsr(cccr_msr, cccr_val, 0);
 479         wd->perfctr_msr = perfctr_msr;
 480         wd->evntsel_msr = evntsel_msr;
 481         wd->cccr_msr = cccr_msr;
 482         return 1;
 483 }
 484
 485 static void stop_p4_watchdog(void)
 486 {
 487         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 488         wrmsr(wd->cccr_msr, 0, 0);
 489         wrmsr(wd->evntsel_msr, 0, 0);
 490 }
 491
 492 static int p4_reserve(void)
 493 {
 494         if (!reserve_perfctr_nmi(MSR_P4_IQ_PERFCTR0))
 495                 return 0;
 496 #ifdef CONFIG_SMP
 497         if (smp_num_siblings > 1 && !reserve_perfctr_nmi(MSR_P4_IQ_PERFCTR1))
 498                 goto fail1;
 499 #endif
 500         if (!reserve_evntsel_nmi(MSR_P4_CRU_ESCR0))
 501                 goto fail2;
 502         /* RED-PEN why is ESCR1 not reserved here? */
 503         return 1;
 504  fail2:
 505 #ifdef CONFIG_SMP
 506         if (smp_num_siblings > 1)
 507                 release_perfctr_nmi(MSR_P4_IQ_PERFCTR1);
 508  fail1:
 509 #endif
 510         release_perfctr_nmi(MSR_P4_IQ_PERFCTR0);
 511         return 0;
 512 }
 513
 514 static void p4_unreserve(void)
 515 {
 516 #ifdef CONFIG_SMP
 517         if (smp_num_siblings > 1)
 518                 release_perfctr_nmi(MSR_P4_IQ_PERFCTR1);
 519 #endif
 520         release_evntsel_nmi(MSR_P4_CRU_ESCR0);
 521         release_perfctr_nmi(MSR_P4_IQ_PERFCTR0);
 522 }
 523
 524 static void p4_rearm(struct nmi_watchdog_ctlblk *wd, unsigned nmi_hz)
 525 {
 526         unsigned dummy;
 527         /*
 528          * P4 quirks:
 529          * - An overflown perfctr will assert its interrupt
 530          *   until the OVF flag in its CCCR is cleared.
 531          * - LVTPC is masked on interrupt and must be
 532          *   unmasked by the LVTPC handler.
 533          */
 534         rdmsrl(wd->cccr_msr, dummy);
 535         dummy &= ~P4_CCCR_OVF;
 536         wrmsrl(wd->cccr_msr, dummy);
 537         apic_write(APIC_LVTPC, APIC_DM_NMI);
 538         /* start the cycle over again */
 539         write_watchdog_counter(wd->perfctr_msr, NULL, nmi_hz);
 540 }
 541
 542 static const struct wd_ops p4_wd_ops = {
 543         .reserve = p4_reserve,
 544         .unreserve = p4_unreserve,
 545         .setup = setup_p4_watchdog,
 546         .rearm = p4_rearm,
 547         .stop = stop_p4_watchdog,
 548         /* RED-PEN this is wrong for the other sibling */
 549         .perfctr = MSR_P4_BPU_PERFCTR0,
 550         .evntsel = MSR_P4_BSU_ESCR0,
 551         .checkbit = 1ULL<<39,
 552 };
 553
 554 /* Watchdog using the Intel architected PerfMon. Used for Core2 and hopefully
 555    all future Intel CPUs. */
 556
 557 #define ARCH_PERFMON_NMI_EVENT_SEL      ARCH_PERFMON_UNHALTED_CORE_CYCLES_SEL
 558 #define ARCH_PERFMON_NMI_EVENT_UMASK    ARCH_PERFMON_UNHALTED_CORE_CYCLES_UMASK
 559
 560 static struct wd_ops intel_arch_wd_ops;
 561
 562 static int setup_intel_arch_watchdog(unsigned nmi_hz)
 563 {
 564         unsigned int ebx;
 565         union cpuid10_eax eax;
 566         unsigned int unused;
 567         unsigned int perfctr_msr, evntsel_msr;
 568         unsigned int evntsel;
 569         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 570
 571         /*
 572          * Check whether the Architectural PerfMon supports
 573          * Unhalted Core Cycles Event or not.
 574          * NOTE: Corresponding bit = 0 in ebx indicates event present.
 575          */
 576         cpuid(10, &(eax.full), &ebx, &unused, &unused);
 577         if ((eax.split.mask_length < (ARCH_PERFMON_UNHALTED_CORE_CYCLES_INDEX+1)) ||
 578             (ebx & ARCH_PERFMON_UNHALTED_CORE_CYCLES_PRESENT))
 579                 return 0;
 580
 581         perfctr_msr = wd_ops->perfctr;
 582         evntsel_msr = wd_ops->evntsel;
 583
 584         wrmsrl(perfctr_msr, 0UL);
 585
 586         evntsel = ARCH_PERFMON_EVENTSEL_INT
 587                 | ARCH_PERFMON_EVENTSEL_OS
 588                 | ARCH_PERFMON_EVENTSEL_USR
 589                 | ARCH_PERFMON_NMI_EVENT_SEL
 590                 | ARCH_PERFMON_NMI_EVENT_UMASK;
 591
 592         /* setup the timer */
 593         wrmsr(evntsel_msr, evntsel, 0);
 594         nmi_hz = adjust_for_32bit_ctr(nmi_hz);
 595         write_watchdog_counter32(perfctr_msr, "INTEL_ARCH_PERFCTR0", nmi_hz);
 596         apic_write(APIC_LVTPC, APIC_DM_NMI);
 597         evntsel |= ARCH_PERFMON_EVENTSEL0_ENABLE;
 598         wrmsr(evntsel_msr, evntsel, 0);
 599
 600         wd->perfctr_msr = perfctr_msr;
 601         wd->evntsel_msr = evntsel_msr;
 602         wd->cccr_msr = 0;  //unused
 603         intel_arch_wd_ops.checkbit = 1ULL << (eax.split.bit_width - 1);
 604         return 1;
 605 }
 606
 607 static struct wd_ops intel_arch_wd_ops __read_mostly = {
 608         .reserve = single_msr_reserve,
 609         .unreserve = single_msr_unreserve,
 610         .setup = setup_intel_arch_watchdog,
 611         .rearm = p6_rearm,
 612         .stop = single_msr_stop_watchdog,
 613         .perfctr = MSR_ARCH_PERFMON_PERFCTR1,
 614         .evntsel = MSR_ARCH_PERFMON_EVENTSEL1,
 615 };
 616
 617 static struct wd_ops coreduo_wd_ops = {
 618         .reserve = single_msr_reserve,
 619         .unreserve = single_msr_unreserve,
 620         .setup = setup_intel_arch_watchdog,
 621         .rearm = p6_rearm,
 622         .stop = single_msr_stop_watchdog,
 623         .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
 624         .evntsel = MSR_ARCH_PERFMON_EVENTSEL0,
 625 };
 626
 627 static void probe_nmi_watchdog(void)
 628 {
 629         switch (boot_cpu_data.x86_vendor) {
 630         case X86_VENDOR_AMD:
 631                 if (boot_cpu_data.x86 != 6 && boot_cpu_data.x86 != 15 &&
 632                     boot_cpu_data.x86 != 16)
 633                         return;
 634                 wd_ops = &k7_wd_ops;
 635                 break;
 636         case X86_VENDOR_INTEL:
 637                 /* Work around Core Duo (Yonah) errata AE49 where perfctr1
 638                    doesn't have a working enable bit. */
 639                 if (boot_cpu_data.x86 == 6 && boot_cpu_data.x86_model == 14) {
 640                         wd_ops = &coreduo_wd_ops;
 641                         break;
 642                 }
 643                 if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
 644                         wd_ops = &intel_arch_wd_ops;
 645                         break;
 646                 }
 647                 switch (boot_cpu_data.x86) {
 648                 case 6:
 649                         if (boot_cpu_data.x86_model > 0xd)
 650                                 return;
 651
 652                         wd_ops = &p6_wd_ops;
 653                         break;
 654                 case 15:
 655                         if (boot_cpu_data.x86_model > 0x4)
 656                                 return;
 657
 658                         wd_ops = &p4_wd_ops;
 659                         break;
 660                 default:
 661                         return;
 662                 }
 663                 break;
 664         }
 665 }
 666
 667 /* Interface to nmi.c */
 668
 669 int lapic_watchdog_init(unsigned nmi_hz)
 670 {
 671         if (!wd_ops) {
 672                 probe_nmi_watchdog();
 673                 if (!wd_ops)
 674                         return -1;
 675
 676                 if (!wd_ops->reserve()) {
 677                         printk(KERN_ERR
 678                                 "NMI watchdog: cannot reserve perfctrs\n");
 679                         return -1;
 680                 }
 681         }
 682
 683         if (!(wd_ops->setup(nmi_hz))) {
 684                 printk(KERN_ERR "Cannot setup NMI watchdog on CPU %d\n",
 685                        raw_smp_processor_id());
 686                 return -1;
 687         }
 688
 689         return 0;
 690 }
 691
 692 void lapic_watchdog_stop(void)
 693 {
 694         if (wd_ops)
 695                 wd_ops->stop();
 696 }
 697
 698 unsigned lapic_adjust_nmi_hz(unsigned hz)
 699 {
 700         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 701         if (wd->perfctr_msr == MSR_P6_PERFCTR0 ||
 702             wd->perfctr_msr == MSR_ARCH_PERFMON_PERFCTR1)
 703                 hz = adjust_for_32bit_ctr(hz);
 704         return hz;
 705 }
 706
 707 int lapic_wd_event(unsigned nmi_hz)
 708 {
 709         struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
 710         u64 ctr;
 711         rdmsrl(wd->perfctr_msr, ctr);
 712         if (ctr & wd_ops->checkbit) { /* perfctr still running? */
 713                 return 0;
 714         }
 715         wd_ops->rearm(wd, nmi_hz);
 716         return 1;
 717 }
 718
 719 int lapic_watchdog_ok(void)
 720 {
 721         return wd_ops != NULL;
 722 }