arch/i386/oprofile/nmi_int.c

   1 /**
   2  * @file nmi_int.c
   3  *
   4  * @remark Copyright 2002 OProfile authors
   5  * @remark Read the file COPYING
   6  *
   7  * @author John Levon <levon@movementarian.org>
   8  */
   9
  10 #include <linux/init.h>
  11 #include <linux/notifier.h>
  12 #include <linux/smp.h>
  13 #include <linux/oprofile.h>
  14 #include <linux/sysdev.h>
  15 #include <linux/slab.h>
  16 #include <linux/moduleparam.h>
  17 #include <asm/nmi.h>
  18 #include <asm/msr.h>
  19 #include <asm/apic.h>
  20 #include <asm/kdebug.h>
  21
  22 #include "op_counter.h"
  23 #include "op_x86_model.h"
  24
  25 static struct op_x86_model_spec const * model;
  26 static struct op_msrs cpu_msrs[NR_CPUS];
  27 static unsigned long saved_lvtpc[NR_CPUS];
  28
  29 static int nmi_start(void);
  30 static void nmi_stop(void);
  31
  32 /* 0 == registered but off, 1 == registered and on */
  33 static int nmi_enabled = 0;
  34
  35 #ifdef CONFIG_PM
  36
  37 static int nmi_suspend(struct sys_device *dev, pm_message_t state)
  38 {
  39         if (nmi_enabled == 1)
  40                 nmi_stop();
  41         return 0;
  42 }
  43
  44
  45 static int nmi_resume(struct sys_device *dev)
  46 {
  47         if (nmi_enabled == 1)
  48                 nmi_start();
  49         return 0;
  50 }
  51
  52
  53 static struct sysdev_class oprofile_sysclass = {
  54         set_kset_name("oprofile"),
  55         .resume         = nmi_resume,
  56         .suspend        = nmi_suspend,
  57 };
  58
  59
  60 static struct sys_device device_oprofile = {
  61         .id     = 0,
  62         .cls    = &oprofile_sysclass,
  63 };
  64
  65
  66 static int __init init_driverfs(void)
  67 {
  68         int error;
  69         if (!(error = sysdev_class_register(&oprofile_sysclass)))
  70                 error = sysdev_register(&device_oprofile);
  71         return error;
  72 }
  73
  74
  75 static void exit_driverfs(void)
  76 {
  77         sysdev_unregister(&device_oprofile);
  78         sysdev_class_unregister(&oprofile_sysclass);
  79 }
  80
  81 #else
  82 #define init_driverfs() do { } while (0)
  83 #define exit_driverfs() do { } while (0)
  84 #endif /* CONFIG_PM */
  85
  86 static int profile_exceptions_notify(struct notifier_block *self,
  87                                      unsigned long val, void *data)
  88 {
  89         struct die_args *args = (struct die_args *)data;
  90         int ret = NOTIFY_DONE;
  91         int cpu = smp_processor_id();
  92
  93         switch(val) {
  94         case DIE_NMI:
  95                 if (model->check_ctrs(args->regs, &cpu_msrs[cpu]))
  96                         ret = NOTIFY_STOP;
  97                 break;
  98         default:
  99                 break;
 100         }
 101         return ret;
 102 }
 103
 104 static void nmi_cpu_save_registers(struct op_msrs * msrs)
 105 {
 106         unsigned int const nr_ctrs = model->num_counters;
 107         unsigned int const nr_ctrls = model->num_controls;
 108         struct op_msr * counters = msrs->counters;
 109         struct op_msr * controls = msrs->controls;
 110         unsigned int i;
 111
 112         for (i = 0; i < nr_ctrs; ++i) {
 113                 if (counters[i].addr){
 114                         rdmsr(counters[i].addr,
 115                                 counters[i].saved.low,
 116                                 counters[i].saved.high);
 117                 }
 118         }
 119
 120         for (i = 0; i < nr_ctrls; ++i) {
 121                 if (controls[i].addr){
 122                         rdmsr(controls[i].addr,
 123                                 controls[i].saved.low,
 124                                 controls[i].saved.high);
 125                 }
 126         }
 127 }
 128
 129
 130 static void nmi_save_registers(void * dummy)
 131 {
 132         int cpu = smp_processor_id();
 133         struct op_msrs * msrs = &cpu_msrs[cpu];
 134         model->fill_in_addresses(msrs);
 135         nmi_cpu_save_registers(msrs);
 136 }
 137
 138
 139 static void free_msrs(void)
 140 {
 141         int i;
 142         for_each_possible_cpu(i) {
 143                 kfree(cpu_msrs[i].counters);
 144                 cpu_msrs[i].counters = NULL;
 145                 kfree(cpu_msrs[i].controls);
 146                 cpu_msrs[i].controls = NULL;
 147         }
 148 }
 149
 150
 151 static int allocate_msrs(void)
 152 {
 153         int success = 1;
 154         size_t controls_size = sizeof(struct op_msr) * model->num_controls;
 155         size_t counters_size = sizeof(struct op_msr) * model->num_counters;
 156
 157         int i;
 158         for_each_online_cpu(i) {
 159                 cpu_msrs[i].counters = kmalloc(counters_size, GFP_KERNEL);
 160                 if (!cpu_msrs[i].counters) {
 161                         success = 0;
 162                         break;
 163                 }
 164                 cpu_msrs[i].controls = kmalloc(controls_size, GFP_KERNEL);
 165                 if (!cpu_msrs[i].controls) {
 166                         success = 0;
 167                         break;
 168                 }
 169         }
 170
 171         if (!success)
 172                 free_msrs();
 173
 174         return success;
 175 }
 176
 177
 178 static void nmi_cpu_setup(void * dummy)
 179 {
 180         int cpu = smp_processor_id();
 181         struct op_msrs * msrs = &cpu_msrs[cpu];
 182         spin_lock(&oprofilefs_lock);
 183         model->setup_ctrs(msrs);
 184         spin_unlock(&oprofilefs_lock);
 185         saved_lvtpc[cpu] = apic_read(APIC_LVTPC);
 186         apic_write(APIC_LVTPC, APIC_DM_NMI);
 187 }
 188
 189 static struct notifier_block profile_exceptions_nb = {
 190         .notifier_call = profile_exceptions_notify,
 191         .next = NULL,
 192         .priority = 0
 193 };
 194
 195 static int nmi_setup(void)
 196 {
 197         int err=0;
 198
 199         if (!allocate_msrs())
 200                 return -ENOMEM;
 201
 202         if ((err = register_die_notifier(&profile_exceptions_nb))){
 203                 free_msrs();
 204                 return err;
 205         }
 206
 207         /* We need to serialize save and setup for HT because the subset
 208          * of msrs are distinct for save and setup operations
 209          */
 210         on_each_cpu(nmi_save_registers, NULL, 0, 1);
 211         on_each_cpu(nmi_cpu_setup, NULL, 0, 1);
 212         nmi_enabled = 1;
 213         return 0;
 214 }
 215
 216
 217 static void nmi_restore_registers(struct op_msrs * msrs)
 218 {
 219         unsigned int const nr_ctrs = model->num_counters;
 220         unsigned int const nr_ctrls = model->num_controls;
 221         struct op_msr * counters = msrs->counters;
 222         struct op_msr * controls = msrs->controls;
 223         unsigned int i;
 224
 225         for (i = 0; i < nr_ctrls; ++i) {
 226                 if (controls[i].addr){
 227                         wrmsr(controls[i].addr,
 228                                 controls[i].saved.low,
 229                                 controls[i].saved.high);
 230                 }
 231         }
 232
 233         for (i = 0; i < nr_ctrs; ++i) {
 234                 if (counters[i].addr){
 235                         wrmsr(counters[i].addr,
 236                                 counters[i].saved.low,
 237                                 counters[i].saved.high);
 238                 }
 239         }
 240 }
 241
 242
 243 static void nmi_cpu_shutdown(void * dummy)
 244 {
 245         unsigned int v;
 246         int cpu = smp_processor_id();
 247         struct op_msrs * msrs = &cpu_msrs[cpu];
 248
 249         /* restoring APIC_LVTPC can trigger an apic error because the delivery
 250          * mode and vector nr combination can be illegal. That's by design: on
 251          * power on apic lvt contain a zero vector nr which are legal only for
 252          * NMI delivery mode. So inhibit apic err before restoring lvtpc
 253          */
 254         v = apic_read(APIC_LVTERR);
 255         apic_write(APIC_LVTERR, v | APIC_LVT_MASKED);
 256         apic_write(APIC_LVTPC, saved_lvtpc[cpu]);
 257         apic_write(APIC_LVTERR, v);
 258         nmi_restore_registers(msrs);
 259         model->shutdown(msrs);
 260 }
 261
 262
 263 static void nmi_shutdown(void)
 264 {
 265         nmi_enabled = 0;
 266         on_each_cpu(nmi_cpu_shutdown, NULL, 0, 1);
 267         unregister_die_notifier(&profile_exceptions_nb);
 268         free_msrs();
 269 }
 270
 271
 272 static void nmi_cpu_start(void * dummy)
 273 {
 274         struct op_msrs const * msrs = &cpu_msrs[smp_processor_id()];
 275         model->start(msrs);
 276 }
 277
 278
 279 static int nmi_start(void)
 280 {
 281         on_each_cpu(nmi_cpu_start, NULL, 0, 1);
 282         return 0;
 283 }
 284
 285
 286 static void nmi_cpu_stop(void * dummy)
 287 {
 288         struct op_msrs const * msrs = &cpu_msrs[smp_processor_id()];
 289         model->stop(msrs);
 290 }
 291
 292
 293 static void nmi_stop(void)
 294 {
 295         on_each_cpu(nmi_cpu_stop, NULL, 0, 1);
 296 }
 297
 298
 299 struct op_counter_config counter_config[OP_MAX_COUNTER];
 300
 301 static int nmi_create_files(struct super_block * sb, struct dentry * root)
 302 {
 303         unsigned int i;
 304
 305         for (i = 0; i < model->num_counters; ++i) {
 306                 struct dentry * dir;
 307                 char buf[4];
 308
 309                 /* quick little hack to _not_ expose a counter if it is not
 310                  * available for use.  This should protect userspace app.
 311                  * NOTE:  assumes 1:1 mapping here (that counters are organized
 312                  *        sequentially in their struct assignment).
 313                  */
 314                 if (unlikely(!avail_to_resrv_perfctr_nmi_bit(i)))
 315                         continue;
 316
 317                 snprintf(buf,  sizeof(buf), "%d", i);
 318                 dir = oprofilefs_mkdir(sb, root, buf);
 319                 oprofilefs_create_ulong(sb, dir, "enabled", &counter_config[i].enabled);
 320                 oprofilefs_create_ulong(sb, dir, "event", &counter_config[i].event);
 321                 oprofilefs_create_ulong(sb, dir, "count", &counter_config[i].count);
 322                 oprofilefs_create_ulong(sb, dir, "unit_mask", &counter_config[i].unit_mask);
 323                 oprofilefs_create_ulong(sb, dir, "kernel", &counter_config[i].kernel);
 324                 oprofilefs_create_ulong(sb, dir, "user", &counter_config[i].user);
 325         }
 326
 327         return 0;
 328 }
 329
 330 static int p4force;
 331 module_param(p4force, int, 0);
 332
 333 static int __init p4_init(char ** cpu_type)
 334 {
 335         __u8 cpu_model = boot_cpu_data.x86_model;
 336
 337         if (!p4force && (cpu_model > 6 || cpu_model == 5))
 338                 return 0;
 339
 340 #ifndef CONFIG_SMP
 341         *cpu_type = "i386/p4";
 342         model = &op_p4_spec;
 343         return 1;
 344 #else
 345         switch (smp_num_siblings) {
 346                 case 1:
 347                         *cpu_type = "i386/p4";
 348                         model = &op_p4_spec;
 349                         return 1;
 350
 351                 case 2:
 352                         *cpu_type = "i386/p4-ht";
 353                         model = &op_p4_ht2_spec;
 354                         return 1;
 355         }
 356 #endif
 357
 358         printk(KERN_INFO "oprofile: P4 HyperThreading detected with > 2 threads\n");
 359         printk(KERN_INFO "oprofile: Reverting to timer mode.\n");
 360         return 0;
 361 }
 362
 363
 364 static int __init ppro_init(char ** cpu_type)
 365 {
 366         __u8 cpu_model = boot_cpu_data.x86_model;
 367
 368         if (cpu_model == 14)
 369                 *cpu_type = "i386/core";
 370         else if (cpu_model == 15)
 371                 *cpu_type = "i386/core_2";
 372         else if (cpu_model > 0xd)
 373                 return 0;
 374         else if (cpu_model == 9) {
 375                 *cpu_type = "i386/p6_mobile";
 376         } else if (cpu_model > 5) {
 377                 *cpu_type = "i386/piii";
 378         } else if (cpu_model > 2) {
 379                 *cpu_type = "i386/pii";
 380         } else {
 381                 *cpu_type = "i386/ppro";
 382         }
 383
 384         model = &op_ppro_spec;
 385         return 1;
 386 }
 387
 388 /* in order to get driverfs right */
 389 static int using_nmi;
 390
 391 int __init op_nmi_init(struct oprofile_operations *ops)
 392 {
 393         __u8 vendor = boot_cpu_data.x86_vendor;
 394         __u8 family = boot_cpu_data.x86;
 395         char *cpu_type;
 396
 397         if (!cpu_has_apic)
 398                 return -ENODEV;
 399
 400         switch (vendor) {
 401                 case X86_VENDOR_AMD:
 402                         /* Needs to be at least an Athlon (or hammer in 32bit mode) */
 403
 404                         switch (family) {
 405                         default:
 406                                 return -ENODEV;
 407                         case 6:
 408                                 model = &op_athlon_spec;
 409                                 cpu_type = "i386/athlon";
 410                                 break;
 411                         case 0xf:
 412                                 model = &op_athlon_spec;
 413                                 /* Actually it could be i386/hammer too, but give
 414                                    user space an consistent name. */
 415                                 cpu_type = "x86-64/hammer";
 416                                 break;
 417                         }
 418                         break;
 419
 420                 case X86_VENDOR_INTEL:
 421                         switch (family) {
 422                                 /* Pentium IV */
 423                                 case 0xf:
 424                                         if (!p4_init(&cpu_type))
 425                                                 return -ENODEV;
 426                                         break;
 427
 428                                 /* A P6-class processor */
 429                                 case 6:
 430                                         if (!ppro_init(&cpu_type))
 431                                                 return -ENODEV;
 432                                         break;
 433
 434                                 default:
 435                                         return -ENODEV;
 436                         }
 437                         break;
 438
 439                 default:
 440                         return -ENODEV;
 441         }
 442
 443         init_driverfs();
 444         using_nmi = 1;
 445         ops->create_files = nmi_create_files;
 446         ops->setup = nmi_setup;
 447         ops->shutdown = nmi_shutdown;
 448         ops->start = nmi_start;
 449         ops->stop = nmi_stop;
 450         ops->cpu_type = cpu_type;
 451         printk(KERN_INFO "oprofile: using NMI interrupt.\n");
 452         return 0;
 453 }
 454
 455
 456 void op_nmi_exit(void)
 457 {
 458         if (using_nmi)
 459                 exit_driverfs();
 460 }