arch/i386/oprofile/op_model_ppro.c

   1 /**
   2  * @file op_model_ppro.h
   3  * pentium pro / P6 model-specific MSR operations
   4  *
   5  * @remark Copyright 2002 OProfile authors
   6  * @remark Read the file COPYING
   7  *
   8  * @author John Levon
   9  * @author Philippe Elie
  10  * @author Graydon Hoare
  11  */
  12
  13 #include <linux/oprofile.h>
  14 #include <asm/ptrace.h>
  15 #include <asm/msr.h>
  16 #include <asm/apic.h>
  17
  18 #include "op_x86_model.h"
  19 #include "op_counter.h"
  20
  21 #define NUM_COUNTERS 2
  22 #define NUM_CONTROLS 2
  23
  24 #define CTR_READ(l,h,msrs,c) do {rdmsr(msrs->counters[(c)].addr, (l), (h));} while (0)
  25 #define CTR_WRITE(l,msrs,c) do {wrmsr(msrs->counters[(c)].addr, -(u32)(l), -1);} while (0)
  26 #define CTR_OVERFLOWED(n) (!((n) & (1U<<31)))
  27
  28 #define CTRL_READ(l,h,msrs,c) do {rdmsr((msrs->controls[(c)].addr), (l), (h));} while (0)
  29 #define CTRL_WRITE(l,h,msrs,c) do {wrmsr((msrs->controls[(c)].addr), (l), (h));} while (0)
  30 #define CTRL_SET_ACTIVE(n) (n |= (1<<22))
  31 #define CTRL_SET_INACTIVE(n) (n &= ~(1<<22))
  32 #define CTRL_CLEAR(x) (x &= (1<<21))
  33 #define CTRL_SET_ENABLE(val) (val |= 1<<20)
  34 #define CTRL_SET_USR(val,u) (val |= ((u & 1) << 16))
  35 #define CTRL_SET_KERN(val,k) (val |= ((k & 1) << 17))
  36 #define CTRL_SET_UM(val, m) (val |= (m << 8))
  37 #define CTRL_SET_EVENT(val, e) (val |= e)
  38
  39 static unsigned long reset_value[NUM_COUNTERS];
  40
  41 static void ppro_fill_in_addresses(struct op_msrs * const msrs)
  42 {
  43         msrs->counters[0].addr = MSR_P6_PERFCTR0;
  44         msrs->counters[1].addr = MSR_P6_PERFCTR1;
  45
  46         msrs->controls[0].addr = MSR_P6_EVNTSEL0;
  47         msrs->controls[1].addr = MSR_P6_EVNTSEL1;
  48 }
  49
  50
  51 static void ppro_setup_ctrs(struct op_msrs const * const msrs)
  52 {
  53         unsigned int low, high;
  54         int i;
  55
  56         /* clear all counters */
  57         for (i = 0 ; i < NUM_CONTROLS; ++i) {
  58                 CTRL_READ(low, high, msrs, i);
  59                 CTRL_CLEAR(low);
  60                 CTRL_WRITE(low, high, msrs, i);
  61         }
  62
  63         /* avoid a false detection of ctr overflows in NMI handler */
  64         for (i = 0; i < NUM_COUNTERS; ++i) {
  65                 CTR_WRITE(1, msrs, i);
  66         }
  67
  68         /* enable active counters */
  69         for (i = 0; i < NUM_COUNTERS; ++i) {
  70                 if (counter_config[i].enabled) {
  71                         reset_value[i] = counter_config[i].count;
  72
  73                         CTR_WRITE(counter_config[i].count, msrs, i);
  74
  75                         CTRL_READ(low, high, msrs, i);
  76                         CTRL_CLEAR(low);
  77                         CTRL_SET_ENABLE(low);
  78                         CTRL_SET_USR(low, counter_config[i].user);
  79                         CTRL_SET_KERN(low, counter_config[i].kernel);
  80                         CTRL_SET_UM(low, counter_config[i].unit_mask);
  81                         CTRL_SET_EVENT(low, counter_config[i].event);
  82                         CTRL_WRITE(low, high, msrs, i);
  83                 }
  84         }
  85 }
  86
  87
  88 static int ppro_check_ctrs(struct pt_regs * const regs,
  89                            struct op_msrs const * const msrs)
  90 {
  91         unsigned int low, high;
  92         int i;
  93
  94         for (i = 0 ; i < NUM_COUNTERS; ++i) {
  95                 CTR_READ(low, high, msrs, i);
  96                 if (CTR_OVERFLOWED(low)) {
  97                         oprofile_add_sample(regs, i);
  98                         CTR_WRITE(reset_value[i], msrs, i);
  99                 }
 100         }
 101
 102         /* Only P6 based Pentium M need to re-unmask the apic vector but it
 103          * doesn't hurt other P6 variant */
 104         apic_write(APIC_LVTPC, apic_read(APIC_LVTPC) & ~APIC_LVT_MASKED);
 105
 106         /* We can't work out if we really handled an interrupt. We
 107          * might have caught a *second* counter just after overflowing
 108          * the interrupt for this counter then arrives
 109          * and we don't find a counter that's overflowed, so we
 110          * would return 0 and get dazed + confused. Instead we always
 111          * assume we found an overflow. This sucks.
 112          */
 113         return 1;
 114 }
 115
 116
 117 static void ppro_start(struct op_msrs const * const msrs)
 118 {
 119         unsigned int low,high;
 120         CTRL_READ(low, high, msrs, 0);
 121         CTRL_SET_ACTIVE(low);
 122         CTRL_WRITE(low, high, msrs, 0);
 123 }
 124
 125
 126 static void ppro_stop(struct op_msrs const * const msrs)
 127 {
 128         unsigned int low,high;
 129         CTRL_READ(low, high, msrs, 0);
 130         CTRL_SET_INACTIVE(low);
 131         CTRL_WRITE(low, high, msrs, 0);
 132 }
 133
 134
 135 struct op_x86_model_spec const op_ppro_spec = {
 136         .num_counters = NUM_COUNTERS,
 137         .num_controls = NUM_CONTROLS,
 138         .fill_in_addresses = &ppro_fill_in_addresses,
 139         .setup_ctrs = &ppro_setup_ctrs,
 140         .check_ctrs = &ppro_check_ctrs,
 141         .start = &ppro_start,
 142         .stop = &ppro_stop
 143 };