arch/x86/oprofile/op_model_ppro.c

   1 /*
   2  * @file op_model_ppro.h
   3  * Family 6 perfmon and architectural perfmon MSR operations
   4  *
   5  * @remark Copyright 2002 OProfile authors
   6  * @remark Copyright 2008 Intel Corporation
   7  * @remark Read the file COPYING
   8  *
   9  * @author John Levon
  10  * @author Philippe Elie
  11  * @author Graydon Hoare
  12  * @author Andi Kleen
  13  */
  14
  15 #include <linux/oprofile.h>
  16 #include <linux/slab.h>
  17 #include <asm/ptrace.h>
  18 #include <asm/msr.h>
  19 #include <asm/apic.h>
  20 #include <asm/nmi.h>
  21 #include <asm/intel_arch_perfmon.h>
  22
  23 #include "op_x86_model.h"
  24 #include "op_counter.h"
  25
  26 static int num_counters = 2;
  27 static int counter_width = 32;
  28
  29 #define CTR_IS_RESERVED(msrs, c) (msrs->counters[(c)].addr ? 1 : 0)
  30 #define CTR_READ(l, h, msrs, c) do {rdmsr(msrs->counters[(c)].addr, (l), (h)); } while (0)
  31 #define CTR_OVERFLOWED(n) (!((n) & (1U<<(counter_width-1))))
  32
  33 #define CTRL_IS_RESERVED(msrs, c) (msrs->controls[(c)].addr ? 1 : 0)
  34 #define CTRL_READ(l, h, msrs, c) do {rdmsr((msrs->controls[(c)].addr), (l), (h)); } while (0)
  35 #define CTRL_WRITE(l, h, msrs, c) do {wrmsr((msrs->controls[(c)].addr), (l), (h)); } while (0)
  36 #define CTRL_SET_ACTIVE(n) (n |= (1<<22))
  37 #define CTRL_SET_INACTIVE(n) (n &= ~(1<<22))
  38 #define CTRL_CLEAR(x) (x &= (1<<21))
  39 #define CTRL_SET_ENABLE(val) (val |= 1<<20)
  40 #define CTRL_SET_USR(val, u) (val |= ((u & 1) << 16))
  41 #define CTRL_SET_KERN(val, k) (val |= ((k & 1) << 17))
  42 #define CTRL_SET_UM(val, m) (val |= (m << 8))
  43 #define CTRL_SET_EVENT(val, e) (val |= e)
  44
  45 static u64 *reset_value;
  46
  47 static void ppro_fill_in_addresses(struct op_msrs * const msrs)
  48 {
  49         int i;
  50
  51         for (i = 0; i < num_counters; i++) {
  52                 if (reserve_perfctr_nmi(MSR_P6_PERFCTR0 + i))
  53                         msrs->counters[i].addr = MSR_P6_PERFCTR0 + i;
  54                 else
  55                         msrs->counters[i].addr = 0;
  56         }
  57
  58         for (i = 0; i < num_counters; i++) {
  59                 if (reserve_evntsel_nmi(MSR_P6_EVNTSEL0 + i))
  60                         msrs->controls[i].addr = MSR_P6_EVNTSEL0 + i;
  61                 else
  62                         msrs->controls[i].addr = 0;
  63         }
  64 }
  65
  66
  67 static void ppro_setup_ctrs(struct op_msrs const * const msrs)
  68 {
  69         unsigned int low, high;
  70         int i;
  71
  72         if (!reset_value) {
  73                 reset_value = kmalloc(sizeof(unsigned) * num_counters,
  74                                         GFP_ATOMIC);
  75                 if (!reset_value)
  76                         return;
  77         }
  78
  79         if (cpu_has_arch_perfmon) {
  80                 union cpuid10_eax eax;
  81                 eax.full = cpuid_eax(0xa);
  82                 if (counter_width < eax.split.bit_width)
  83                         counter_width = eax.split.bit_width;
  84         }
  85
  86         /* clear all counters */
  87         for (i = 0 ; i < num_counters; ++i) {
  88                 if (unlikely(!CTRL_IS_RESERVED(msrs, i)))
  89                         continue;
  90                 CTRL_READ(low, high, msrs, i);
  91                 CTRL_CLEAR(low);
  92                 CTRL_WRITE(low, high, msrs, i);
  93         }
  94
  95         /* avoid a false detection of ctr overflows in NMI handler */
  96         for (i = 0; i < num_counters; ++i) {
  97                 if (unlikely(!CTR_IS_RESERVED(msrs, i)))
  98                         continue;
  99                 wrmsrl(msrs->counters[i].addr, -1LL);
 100         }
 101
 102         /* enable active counters */
 103         for (i = 0; i < num_counters; ++i) {
 104                 if ((counter_config[i].enabled) && (CTR_IS_RESERVED(msrs, i))) {
 105                         reset_value[i] = counter_config[i].count;
 106
 107                         wrmsrl(msrs->counters[i].addr, -reset_value[i]);
 108
 109                         CTRL_READ(low, high, msrs, i);
 110                         CTRL_CLEAR(low);
 111                         CTRL_SET_ENABLE(low);
 112                         CTRL_SET_USR(low, counter_config[i].user);
 113                         CTRL_SET_KERN(low, counter_config[i].kernel);
 114                         CTRL_SET_UM(low, counter_config[i].unit_mask);
 115                         CTRL_SET_EVENT(low, counter_config[i].event);
 116                         CTRL_WRITE(low, high, msrs, i);
 117                 } else {
 118                         reset_value[i] = 0;
 119                 }
 120         }
 121 }
 122
 123
 124 static int ppro_check_ctrs(struct pt_regs * const regs,
 125                            struct op_msrs const * const msrs)
 126 {
 127         unsigned int low, high;
 128         int i;
 129
 130         for (i = 0 ; i < num_counters; ++i) {
 131                 if (!reset_value[i])
 132                         continue;
 133                 CTR_READ(low, high, msrs, i);
 134                 if (CTR_OVERFLOWED(low)) {
 135                         oprofile_add_sample(regs, i);
 136                         wrmsrl(msrs->counters[i].addr, -reset_value[i]);
 137                 }
 138         }
 139
 140         /* Only P6 based Pentium M need to re-unmask the apic vector but it
 141          * doesn't hurt other P6 variant */
 142         apic_write(APIC_LVTPC, apic_read(APIC_LVTPC) & ~APIC_LVT_MASKED);
 143
 144         /* We can't work out if we really handled an interrupt. We
 145          * might have caught a *second* counter just after overflowing
 146          * the interrupt for this counter then arrives
 147          * and we don't find a counter that's overflowed, so we
 148          * would return 0 and get dazed + confused. Instead we always
 149          * assume we found an overflow. This sucks.
 150          */
 151         return 1;
 152 }
 153
 154
 155 static void ppro_start(struct op_msrs const * const msrs)
 156 {
 157         unsigned int low, high;
 158         int i;
 159
 160         for (i = 0; i < num_counters; ++i) {
 161                 if (reset_value[i]) {
 162                         CTRL_READ(low, high, msrs, i);
 163                         CTRL_SET_ACTIVE(low);
 164                         CTRL_WRITE(low, high, msrs, i);
 165                 }
 166         }
 167 }
 168
 169
 170 static void ppro_stop(struct op_msrs const * const msrs)
 171 {
 172         unsigned int low, high;
 173         int i;
 174
 175         for (i = 0; i < num_counters; ++i) {
 176                 if (!reset_value[i])
 177                         continue;
 178                 CTRL_READ(low, high, msrs, i);
 179                 CTRL_SET_INACTIVE(low);
 180                 CTRL_WRITE(low, high, msrs, i);
 181         }
 182 }
 183
 184 static void ppro_shutdown(struct op_msrs const * const msrs)
 185 {
 186         int i;
 187
 188         for (i = 0 ; i < num_counters ; ++i) {
 189                 if (CTR_IS_RESERVED(msrs, i))
 190                         release_perfctr_nmi(MSR_P6_PERFCTR0 + i);
 191         }
 192         for (i = 0 ; i < num_counters ; ++i) {
 193                 if (CTRL_IS_RESERVED(msrs, i))
 194                         release_evntsel_nmi(MSR_P6_EVNTSEL0 + i);
 195         }
 196         if (reset_value) {
 197                 kfree(reset_value);
 198                 reset_value = NULL;
 199         }
 200 }
 201
 202
 203 struct op_x86_model_spec op_ppro_spec = {
 204         .num_counters           = 2,    /* can be overriden */
 205         .num_controls           = 2,    /* dito */
 206         .fill_in_addresses      = &ppro_fill_in_addresses,
 207         .setup_ctrs             = &ppro_setup_ctrs,
 208         .check_ctrs             = &ppro_check_ctrs,
 209         .start                  = &ppro_start,
 210         .stop                   = &ppro_stop,
 211         .shutdown               = &ppro_shutdown
 212 };
 213
 214 /*
 215  * Architectural performance monitoring.
 216  *
 217  * Newer Intel CPUs (Core1+) have support for architectural
 218  * events described in CPUID 0xA. See the IA32 SDM Vol3b.18 for details.
 219  * The advantage of this is that it can be done without knowing about
 220  * the specific CPU.
 221  */
 222
 223 void arch_perfmon_setup_counters(void)
 224 {
 225         union cpuid10_eax eax;
 226
 227         eax.full = cpuid_eax(0xa);
 228
 229         /* Workaround for BIOS bugs in 6/15. Taken from perfmon2 */
 230         if (eax.split.version_id == 0 && current_cpu_data.x86 == 6 &&
 231                 current_cpu_data.x86_model == 15) {
 232                 eax.split.version_id = 2;
 233                 eax.split.num_counters = 2;
 234                 eax.split.bit_width = 40;
 235         }
 236
 237         num_counters = eax.split.num_counters;
 238
 239         op_arch_perfmon_spec.num_counters = num_counters;
 240         op_arch_perfmon_spec.num_controls = num_counters;
 241         op_ppro_spec.num_counters = num_counters;
 242         op_ppro_spec.num_controls = num_counters;
 243 }
 244
 245 struct op_x86_model_spec op_arch_perfmon_spec = {
 246         /* num_counters/num_controls filled in at runtime */
 247         .fill_in_addresses      = &ppro_fill_in_addresses,
 248         /* user space does the cpuid check for available events */
 249         .setup_ctrs             = &ppro_setup_ctrs,
 250         .check_ctrs             = &ppro_check_ctrs,
 251         .start                  = &ppro_start,
 252         .stop                   = &ppro_stop,
 253         .shutdown               = &ppro_shutdown
 254 };