arch/x86/oprofile/nmi_int.c

   1 /**
   2  * @file nmi_int.c
   3  *
   4  * @remark Copyright 2002-2008 OProfile authors
   5  * @remark Read the file COPYING
   6  *
   7  * @author John Levon <levon@movementarian.org>
   8  * @author Robert Richter <robert.richter@amd.com>
   9  */
  10
  11 #include <linux/init.h>
  12 #include <linux/notifier.h>
  13 #include <linux/smp.h>
  14 #include <linux/oprofile.h>
  15 #include <linux/sysdev.h>
  16 #include <linux/slab.h>
  17 #include <linux/moduleparam.h>
  18 #include <linux/kdebug.h>
  19 #include <linux/cpu.h>
  20 #include <asm/nmi.h>
  21 #include <asm/msr.h>
  22 #include <asm/apic.h>
  23
  24 #include "op_counter.h"
  25 #include "op_x86_model.h"
  26
  27 static struct op_x86_model_spec const *model;
  28 static DEFINE_PER_CPU(struct op_msrs, cpu_msrs);
  29 static DEFINE_PER_CPU(unsigned long, saved_lvtpc);
  30
  31 /* 0 == registered but off, 1 == registered and on */
  32 static int nmi_enabled = 0;
  33
  34 static int profile_exceptions_notify(struct notifier_block *self,
  35                                      unsigned long val, void *data)
  36 {
  37         struct die_args *args = (struct die_args *)data;
  38         int ret = NOTIFY_DONE;
  39         int cpu = smp_processor_id();
  40
  41         switch (val) {
  42         case DIE_NMI:
  43                 if (model->check_ctrs(args->regs, &per_cpu(cpu_msrs, cpu)))
  44                         ret = NOTIFY_STOP;
  45                 break;
  46         default:
  47                 break;
  48         }
  49         return ret;
  50 }
  51
  52 static void nmi_cpu_save_registers(struct op_msrs *msrs)
  53 {
  54         unsigned int const nr_ctrs = model->num_counters;
  55         unsigned int const nr_ctrls = model->num_controls;
  56         struct op_msr *counters = msrs->counters;
  57         struct op_msr *controls = msrs->controls;
  58         unsigned int i;
  59
  60         for (i = 0; i < nr_ctrs; ++i) {
  61                 if (counters[i].addr) {
  62                         rdmsr(counters[i].addr,
  63                                 counters[i].saved.low,
  64                                 counters[i].saved.high);
  65                 }
  66         }
  67
  68         for (i = 0; i < nr_ctrls; ++i) {
  69                 if (controls[i].addr) {
  70                         rdmsr(controls[i].addr,
  71                                 controls[i].saved.low,
  72                                 controls[i].saved.high);
  73                 }
  74         }
  75 }
  76
  77 static void nmi_save_registers(void *dummy)
  78 {
  79         int cpu = smp_processor_id();
  80         struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu);
  81         nmi_cpu_save_registers(msrs);
  82 }
  83
  84 static void free_msrs(void)
  85 {
  86         int i;
  87         for_each_possible_cpu(i) {
  88                 kfree(per_cpu(cpu_msrs, i).counters);
  89                 per_cpu(cpu_msrs, i).counters = NULL;
  90                 kfree(per_cpu(cpu_msrs, i).controls);
  91                 per_cpu(cpu_msrs, i).controls = NULL;
  92         }
  93 }
  94
  95 static int allocate_msrs(void)
  96 {
  97         int success = 1;
  98         size_t controls_size = sizeof(struct op_msr) * model->num_controls;
  99         size_t counters_size = sizeof(struct op_msr) * model->num_counters;
 100
 101         int i;
 102         for_each_possible_cpu(i) {
 103                 per_cpu(cpu_msrs, i).counters = kmalloc(counters_size,
 104                                                                 GFP_KERNEL);
 105                 if (!per_cpu(cpu_msrs, i).counters) {
 106                         success = 0;
 107                         break;
 108                 }
 109                 per_cpu(cpu_msrs, i).controls = kmalloc(controls_size,
 110                                                                 GFP_KERNEL);
 111                 if (!per_cpu(cpu_msrs, i).controls) {
 112                         success = 0;
 113                         break;
 114                 }
 115         }
 116
 117         if (!success)
 118                 free_msrs();
 119
 120         return success;
 121 }
 122
 123 static void nmi_cpu_setup(void *dummy)
 124 {
 125         int cpu = smp_processor_id();
 126         struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu);
 127         spin_lock(&oprofilefs_lock);
 128         model->setup_ctrs(msrs);
 129         spin_unlock(&oprofilefs_lock);
 130         per_cpu(saved_lvtpc, cpu) = apic_read(APIC_LVTPC);
 131         apic_write(APIC_LVTPC, APIC_DM_NMI);
 132 }
 133
 134 static struct notifier_block profile_exceptions_nb = {
 135         .notifier_call = profile_exceptions_notify,
 136         .next = NULL,
 137         .priority = 0
 138 };
 139
 140 static int nmi_setup(void)
 141 {
 142         int err = 0;
 143         int cpu;
 144
 145         if (!allocate_msrs())
 146                 return -ENOMEM;
 147
 148         err = register_die_notifier(&profile_exceptions_nb);
 149         if (err) {
 150                 free_msrs();
 151                 return err;
 152         }
 153
 154         /* We need to serialize save and setup for HT because the subset
 155          * of msrs are distinct for save and setup operations
 156          */
 157
 158         /* Assume saved/restored counters are the same on all CPUs */
 159         model->fill_in_addresses(&per_cpu(cpu_msrs, 0));
 160         for_each_possible_cpu(cpu) {
 161                 if (cpu != 0) {
 162                         memcpy(per_cpu(cpu_msrs, cpu).counters,
 163                                 per_cpu(cpu_msrs, 0).counters,
 164                                 sizeof(struct op_msr) * model->num_counters);
 165
 166                         memcpy(per_cpu(cpu_msrs, cpu).controls,
 167                                 per_cpu(cpu_msrs, 0).controls,
 168                                 sizeof(struct op_msr) * model->num_controls);
 169                 }
 170
 171         }
 172         on_each_cpu(nmi_save_registers, NULL, 1);
 173         on_each_cpu(nmi_cpu_setup, NULL, 1);
 174         nmi_enabled = 1;
 175         return 0;
 176 }
 177
 178 static void nmi_restore_registers(struct op_msrs *msrs)
 179 {
 180         unsigned int const nr_ctrs = model->num_counters;
 181         unsigned int const nr_ctrls = model->num_controls;
 182         struct op_msr *counters = msrs->counters;
 183         struct op_msr *controls = msrs->controls;
 184         unsigned int i;
 185
 186         for (i = 0; i < nr_ctrls; ++i) {
 187                 if (controls[i].addr) {
 188                         wrmsr(controls[i].addr,
 189                                 controls[i].saved.low,
 190                                 controls[i].saved.high);
 191                 }
 192         }
 193
 194         for (i = 0; i < nr_ctrs; ++i) {
 195                 if (counters[i].addr) {
 196                         wrmsr(counters[i].addr,
 197                                 counters[i].saved.low,
 198                                 counters[i].saved.high);
 199                 }
 200         }
 201 }
 202
 203 static void nmi_cpu_shutdown(void *dummy)
 204 {
 205         unsigned int v;
 206         int cpu = smp_processor_id();
 207         struct op_msrs *msrs = &__get_cpu_var(cpu_msrs);
 208
 209         /* restoring APIC_LVTPC can trigger an apic error because the delivery
 210          * mode and vector nr combination can be illegal. That's by design: on
 211          * power on apic lvt contain a zero vector nr which are legal only for
 212          * NMI delivery mode. So inhibit apic err before restoring lvtpc
 213          */
 214         v = apic_read(APIC_LVTERR);
 215         apic_write(APIC_LVTERR, v | APIC_LVT_MASKED);
 216         apic_write(APIC_LVTPC, per_cpu(saved_lvtpc, cpu));
 217         apic_write(APIC_LVTERR, v);
 218         nmi_restore_registers(msrs);
 219 }
 220
 221 static void nmi_shutdown(void)
 222 {
 223         struct op_msrs *msrs;
 224
 225         nmi_enabled = 0;
 226         on_each_cpu(nmi_cpu_shutdown, NULL, 1);
 227         unregister_die_notifier(&profile_exceptions_nb);
 228         msrs = &get_cpu_var(cpu_msrs);
 229         model->shutdown(msrs);
 230         free_msrs();
 231         put_cpu_var(cpu_msrs);
 232 }
 233
 234 static void nmi_cpu_start(void *dummy)
 235 {
 236         struct op_msrs const *msrs = &__get_cpu_var(cpu_msrs);
 237         model->start(msrs);
 238 }
 239
 240 static int nmi_start(void)
 241 {
 242         on_each_cpu(nmi_cpu_start, NULL, 1);
 243         return 0;
 244 }
 245
 246 static void nmi_cpu_stop(void *dummy)
 247 {
 248         struct op_msrs const *msrs = &__get_cpu_var(cpu_msrs);
 249         model->stop(msrs);
 250 }
 251
 252 static void nmi_stop(void)
 253 {
 254         on_each_cpu(nmi_cpu_stop, NULL, 1);
 255 }
 256
 257 struct op_counter_config counter_config[OP_MAX_COUNTER];
 258
 259 static int nmi_create_files(struct super_block *sb, struct dentry *root)
 260 {
 261         unsigned int i;
 262
 263         for (i = 0; i < model->num_counters; ++i) {
 264                 struct dentry *dir;
 265                 char buf[4];
 266
 267                 /* quick little hack to _not_ expose a counter if it is not
 268                  * available for use.  This should protect userspace app.
 269                  * NOTE:  assumes 1:1 mapping here (that counters are organized
 270                  *        sequentially in their struct assignment).
 271                  */
 272                 if (unlikely(!avail_to_resrv_perfctr_nmi_bit(i)))
 273                         continue;
 274
 275                 snprintf(buf,  sizeof(buf), "%d", i);
 276                 dir = oprofilefs_mkdir(sb, root, buf);
 277                 oprofilefs_create_ulong(sb, dir, "enabled", &counter_config[i].enabled);
 278                 oprofilefs_create_ulong(sb, dir, "event", &counter_config[i].event);
 279                 oprofilefs_create_ulong(sb, dir, "count", &counter_config[i].count);
 280                 oprofilefs_create_ulong(sb, dir, "unit_mask", &counter_config[i].unit_mask);
 281                 oprofilefs_create_ulong(sb, dir, "kernel", &counter_config[i].kernel);
 282                 oprofilefs_create_ulong(sb, dir, "user", &counter_config[i].user);
 283         }
 284
 285         return 0;
 286 }
 287
 288 #ifdef CONFIG_SMP
 289 static int oprofile_cpu_notifier(struct notifier_block *b, unsigned long action,
 290                                  void *data)
 291 {
 292         int cpu = (unsigned long)data;
 293         switch (action) {
 294         case CPU_DOWN_FAILED:
 295         case CPU_ONLINE:
 296                 smp_call_function_single(cpu, nmi_cpu_start, NULL, 0);
 297                 break;
 298         case CPU_DOWN_PREPARE:
 299                 smp_call_function_single(cpu, nmi_cpu_stop, NULL, 1);
 300                 break;
 301         }
 302         return NOTIFY_DONE;
 303 }
 304
 305 static struct notifier_block oprofile_cpu_nb = {
 306         .notifier_call = oprofile_cpu_notifier
 307 };
 308 #endif
 309
 310 #ifdef CONFIG_PM
 311
 312 static int nmi_suspend(struct sys_device *dev, pm_message_t state)
 313 {
 314         /* Only one CPU left, just stop that one */
 315         if (nmi_enabled == 1)
 316                 nmi_cpu_stop(NULL);
 317         return 0;
 318 }
 319
 320 static int nmi_resume(struct sys_device *dev)
 321 {
 322         if (nmi_enabled == 1)
 323                 nmi_cpu_start(NULL);
 324         return 0;
 325 }
 326
 327 static struct sysdev_class oprofile_sysclass = {
 328         .name           = "oprofile",
 329         .resume         = nmi_resume,
 330         .suspend        = nmi_suspend,
 331 };
 332
 333 static struct sys_device device_oprofile = {
 334         .id     = 0,
 335         .cls    = &oprofile_sysclass,
 336 };
 337
 338 static int __init init_sysfs(void)
 339 {
 340         int error;
 341
 342         error = sysdev_class_register(&oprofile_sysclass);
 343         if (!error)
 344                 error = sysdev_register(&device_oprofile);
 345         return error;
 346 }
 347
 348 static void exit_sysfs(void)
 349 {
 350         sysdev_unregister(&device_oprofile);
 351         sysdev_class_unregister(&oprofile_sysclass);
 352 }
 353
 354 #else
 355 #define init_sysfs() do { } while (0)
 356 #define exit_sysfs() do { } while (0)
 357 #endif /* CONFIG_PM */
 358
 359 static int __init p4_init(char **cpu_type)
 360 {
 361         __u8 cpu_model = boot_cpu_data.x86_model;
 362
 363         if (cpu_model > 6 || cpu_model == 5)
 364                 return 0;
 365
 366 #ifndef CONFIG_SMP
 367         *cpu_type = "i386/p4";
 368         model = &op_p4_spec;
 369         return 1;
 370 #else
 371         switch (smp_num_siblings) {
 372         case 1:
 373                 *cpu_type = "i386/p4";
 374                 model = &op_p4_spec;
 375                 return 1;
 376
 377         case 2:
 378                 *cpu_type = "i386/p4-ht";
 379                 model = &op_p4_ht2_spec;
 380                 return 1;
 381         }
 382 #endif
 383
 384         printk(KERN_INFO "oprofile: P4 HyperThreading detected with > 2 threads\n");
 385         printk(KERN_INFO "oprofile: Reverting to timer mode.\n");
 386         return 0;
 387 }
 388
 389 int force_arch_perfmon;
 390 module_param(force_arch_perfmon, int, 0);
 391
 392 static int __init ppro_init(char **cpu_type)
 393 {
 394         __u8 cpu_model = boot_cpu_data.x86_model;
 395
 396         if (force_arch_perfmon && cpu_has_arch_perfmon)
 397                 return 0;
 398
 399         switch (cpu_model) {
 400         case 0 ... 2:
 401                 *cpu_type = "i386/ppro";
 402                 break;
 403         case 3 ... 5:
 404                 *cpu_type = "i386/pii";
 405                 break;
 406         case 6 ... 8:
 407         case 10 ... 11:
 408                 *cpu_type = "i386/piii";
 409                 break;
 410         case 9:
 411         case 13:
 412                 *cpu_type = "i386/p6_mobile";
 413                 break;
 414         case 14:
 415                 *cpu_type = "i386/core";
 416                 break;
 417         case 15: case 23:
 418                 *cpu_type = "i386/core_2";
 419                 break;
 420         case 26:
 421                 arch_perfmon_setup_counters();
 422                 *cpu_type = "i386/core_i7";
 423                 break;
 424         case 28:
 425                 *cpu_type = "i386/atom";
 426                 break;
 427         default:
 428                 /* Unknown */
 429                 return 0;
 430         }
 431
 432         model = &op_ppro_spec;
 433         return 1;
 434 }
 435
 436 static int __init arch_perfmon_init(char **cpu_type)
 437 {
 438         if (!cpu_has_arch_perfmon)
 439                 return 0;
 440         *cpu_type = "i386/arch_perfmon";
 441         model = &op_arch_perfmon_spec;
 442         arch_perfmon_setup_counters();
 443         return 1;
 444 }
 445
 446 /* in order to get sysfs right */
 447 static int using_nmi;
 448
 449 int __init op_nmi_init(struct oprofile_operations *ops)
 450 {
 451         __u8 vendor = boot_cpu_data.x86_vendor;
 452         __u8 family = boot_cpu_data.x86;
 453         char *cpu_type = NULL;
 454         int ret = 0;
 455
 456         if (!cpu_has_apic)
 457                 return -ENODEV;
 458
 459         switch (vendor) {
 460         case X86_VENDOR_AMD:
 461                 /* Needs to be at least an Athlon (or hammer in 32bit mode) */
 462
 463                 switch (family) {
 464                 default:
 465                         return -ENODEV;
 466                 case 6:
 467                         model = &op_amd_spec;
 468                         cpu_type = "i386/athlon";
 469                         break;
 470                 case 0xf:
 471                         model = &op_amd_spec;
 472                         /* Actually it could be i386/hammer too, but give
 473                          user space an consistent name. */
 474                         cpu_type = "x86-64/hammer";
 475                         break;
 476                 case 0x10:
 477                         model = &op_amd_spec;
 478                         cpu_type = "x86-64/family10";
 479                         break;
 480                 case 0x11:
 481                         model = &op_amd_spec;
 482                         cpu_type = "x86-64/family11h";
 483                         break;
 484                 }
 485                 break;
 486
 487         case X86_VENDOR_INTEL:
 488                 switch (family) {
 489                         /* Pentium IV */
 490                 case 0xf:
 491                         p4_init(&cpu_type);
 492                         break;
 493
 494                         /* A P6-class processor */
 495                 case 6:
 496                         ppro_init(&cpu_type);
 497                         break;
 498
 499                 default:
 500                         break;
 501                 }
 502
 503                 if (!cpu_type && !arch_perfmon_init(&cpu_type))
 504                         return -ENODEV;
 505                 break;
 506
 507         default:
 508                 return -ENODEV;
 509         }
 510
 511 #ifdef CONFIG_SMP
 512         register_cpu_notifier(&oprofile_cpu_nb);
 513 #endif
 514         /* default values, can be overwritten by model */
 515         ops->create_files = nmi_create_files;
 516         ops->setup = nmi_setup;
 517         ops->shutdown = nmi_shutdown;
 518         ops->start = nmi_start;
 519         ops->stop = nmi_stop;
 520         ops->cpu_type = cpu_type;
 521
 522         if (model->init)
 523                 ret = model->init(ops);
 524         if (ret)
 525                 return ret;
 526
 527         init_sysfs();
 528         using_nmi = 1;
 529         printk(KERN_INFO "oprofile: using NMI interrupt.\n");
 530         return 0;
 531 }
 532
 533 void op_nmi_exit(void)
 534 {
 535         if (using_nmi) {
 536                 exit_sysfs();
 537 #ifdef CONFIG_SMP
 538                 unregister_cpu_notifier(&oprofile_cpu_nb);
 539 #endif
 540         }
 541         if (model->exit)
 542                 model->exit();
 543 }