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  * @author Robert Richter <robert.richter@amd.com>
  14  */
  15
  16 #include <linux/oprofile.h>
  17 #include <linux/slab.h>
  18 #include <asm/ptrace.h>
  19 #include <asm/msr.h>
  20 #include <asm/apic.h>
  21 #include <asm/nmi.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 MSR_PPRO_EVENTSEL_RESERVED      ((0xFFFFFFFFULL<<32)|(1ULL<<21))
  30
  31 static u64 *reset_value;
  32
  33 static void ppro_fill_in_addresses(struct op_msrs * const msrs)
  34 {
  35         int i;
  36
  37         for (i = 0; i < num_counters; i++) {
  38                 if (reserve_perfctr_nmi(MSR_P6_PERFCTR0 + i))
  39                         msrs->counters[i].addr = MSR_P6_PERFCTR0 + i;
  40         }
  41
  42         for (i = 0; i < num_counters; i++) {
  43                 if (reserve_evntsel_nmi(MSR_P6_EVNTSEL0 + i))
  44                         msrs->controls[i].addr = MSR_P6_EVNTSEL0 + i;
  45         }
  46 }
  47
  48
  49 static void ppro_setup_ctrs(struct op_x86_model_spec const *model,
  50                             struct op_msrs const * const msrs)
  51 {
  52         u64 val;
  53         int i;
  54
  55         if (!reset_value) {
  56                 reset_value = kzalloc(sizeof(reset_value[0]) * num_counters,
  57                                         GFP_ATOMIC);
  58                 if (!reset_value)
  59                         return;
  60         }
  61
  62         if (cpu_has_arch_perfmon) {
  63                 union cpuid10_eax eax;
  64                 eax.full = cpuid_eax(0xa);
  65
  66                 /*
  67                  * For Core2 (family 6, model 15), don't reset the
  68                  * counter width:
  69                  */
  70                 if (!(eax.split.version_id == 0 &&
  71                         current_cpu_data.x86 == 6 &&
  72                                 current_cpu_data.x86_model == 15)) {
  73
  74                         if (counter_width < eax.split.bit_width)
  75                                 counter_width = eax.split.bit_width;
  76                 }
  77         }
  78
  79         /* clear all counters */
  80         for (i = 0; i < num_counters; ++i) {
  81                 if (unlikely(!msrs->controls[i].addr)) {
  82                         if (counter_config[i].enabled && !smp_processor_id())
  83                                 /*
  84                                  * counter is reserved, this is on all
  85                                  * cpus, so report only for cpu #0
  86                                  */
  87                                 op_x86_warn_reserved(i);
  88                         continue;
  89                 }
  90                 rdmsrl(msrs->controls[i].addr, val);
  91                 if (val & ARCH_PERFMON_EVENTSEL0_ENABLE)
  92                         op_x86_warn_in_use(i);
  93                 val &= model->reserved;
  94                 wrmsrl(msrs->controls[i].addr, val);
  95         }
  96
  97         /* avoid a false detection of ctr overflows in NMI handler */
  98         for (i = 0; i < num_counters; ++i) {
  99                 if (unlikely(!msrs->counters[i].addr))
 100                         continue;
 101                 wrmsrl(msrs->counters[i].addr, -1LL);
 102         }
 103
 104         /* enable active counters */
 105         for (i = 0; i < num_counters; ++i) {
 106                 if (counter_config[i].enabled && msrs->counters[i].addr) {
 107                         reset_value[i] = counter_config[i].count;
 108                         wrmsrl(msrs->counters[i].addr, -reset_value[i]);
 109                         rdmsrl(msrs->controls[i].addr, val);
 110                         val &= model->reserved;
 111                         val |= op_x86_get_ctrl(model, &counter_config[i]);
 112                         wrmsrl(msrs->controls[i].addr, val);
 113                 } else {
 114                         reset_value[i] = 0;
 115                 }
 116         }
 117 }
 118
 119
 120 static int ppro_check_ctrs(struct pt_regs * const regs,
 121                            struct op_msrs const * const msrs)
 122 {
 123         u64 val;
 124         int i;
 125
 126         /*
 127          * This can happen if perf counters are in use when
 128          * we steal the die notifier NMI.
 129          */
 130         if (unlikely(!reset_value))
 131                 goto out;
 132
 133         for (i = 0; i < num_counters; ++i) {
 134                 if (!reset_value[i])
 135                         continue;
 136                 rdmsrl(msrs->counters[i].addr, val);
 137                 if (val & (1ULL << (counter_width - 1)))
 138                         continue;
 139                 oprofile_add_sample(regs, i);
 140                 wrmsrl(msrs->counters[i].addr, -reset_value[i]);
 141         }
 142
 143 out:
 144         /* Only P6 based Pentium M need to re-unmask the apic vector but it
 145          * doesn't hurt other P6 variant */
 146         apic_write(APIC_LVTPC, apic_read(APIC_LVTPC) & ~APIC_LVT_MASKED);
 147
 148         /* We can't work out if we really handled an interrupt. We
 149          * might have caught a *second* counter just after overflowing
 150          * the interrupt for this counter then arrives
 151          * and we don't find a counter that's overflowed, so we
 152          * would return 0 and get dazed + confused. Instead we always
 153          * assume we found an overflow. This sucks.
 154          */
 155         return 1;
 156 }
 157
 158
 159 static void ppro_start(struct op_msrs const * const msrs)
 160 {
 161         u64 val;
 162         int i;
 163
 164         if (!reset_value)
 165                 return;
 166         for (i = 0; i < num_counters; ++i) {
 167                 if (reset_value[i]) {
 168                         rdmsrl(msrs->controls[i].addr, val);
 169                         val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
 170                         wrmsrl(msrs->controls[i].addr, val);
 171                 }
 172         }
 173 }
 174
 175
 176 static void ppro_stop(struct op_msrs const * const msrs)
 177 {
 178         u64 val;
 179         int i;
 180
 181         if (!reset_value)
 182                 return;
 183         for (i = 0; i < num_counters; ++i) {
 184                 if (!reset_value[i])
 185                         continue;
 186                 rdmsrl(msrs->controls[i].addr, val);
 187                 val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
 188                 wrmsrl(msrs->controls[i].addr, val);
 189         }
 190 }
 191
 192 static void ppro_shutdown(struct op_msrs const * const msrs)
 193 {
 194         int i;
 195
 196         for (i = 0; i < num_counters; ++i) {
 197                 if (msrs->counters[i].addr)
 198                         release_perfctr_nmi(MSR_P6_PERFCTR0 + i);
 199         }
 200         for (i = 0; i < num_counters; ++i) {
 201                 if (msrs->controls[i].addr)
 202                         release_evntsel_nmi(MSR_P6_EVNTSEL0 + i);
 203         }
 204         if (reset_value) {
 205                 kfree(reset_value);
 206                 reset_value = NULL;
 207         }
 208 }
 209
 210
 211 struct op_x86_model_spec op_ppro_spec = {
 212         .num_counters           = 2,
 213         .num_controls           = 2,
 214         .reserved               = MSR_PPRO_EVENTSEL_RESERVED,
 215         .fill_in_addresses      = &ppro_fill_in_addresses,
 216         .setup_ctrs             = &ppro_setup_ctrs,
 217         .check_ctrs             = &ppro_check_ctrs,
 218         .start                  = &ppro_start,
 219         .stop                   = &ppro_stop,
 220         .shutdown               = &ppro_shutdown
 221 };
 222
 223 /*
 224  * Architectural performance monitoring.
 225  *
 226  * Newer Intel CPUs (Core1+) have support for architectural
 227  * events described in CPUID 0xA. See the IA32 SDM Vol3b.18 for details.
 228  * The advantage of this is that it can be done without knowing about
 229  * the specific CPU.
 230  */
 231
 232 static void arch_perfmon_setup_counters(void)
 233 {
 234         union cpuid10_eax eax;
 235
 236         eax.full = cpuid_eax(0xa);
 237
 238         /* Workaround for BIOS bugs in 6/15. Taken from perfmon2 */
 239         if (eax.split.version_id == 0 && current_cpu_data.x86 == 6 &&
 240                 current_cpu_data.x86_model == 15) {
 241                 eax.split.version_id = 2;
 242                 eax.split.num_events = 2;
 243                 eax.split.bit_width = 40;
 244         }
 245
 246         num_counters = eax.split.num_events;
 247
 248         op_arch_perfmon_spec.num_counters = num_counters;
 249         op_arch_perfmon_spec.num_controls = num_counters;
 250 }
 251
 252 static int arch_perfmon_init(struct oprofile_operations *ignore)
 253 {
 254         arch_perfmon_setup_counters();
 255         return 0;
 256 }
 257
 258 struct op_x86_model_spec op_arch_perfmon_spec = {
 259         .reserved               = MSR_PPRO_EVENTSEL_RESERVED,
 260         .init                   = &arch_perfmon_init,
 261         /* num_counters/num_controls filled in at runtime */
 262         .fill_in_addresses      = &ppro_fill_in_addresses,
 263         /* user space does the cpuid check for available events */
 264         .setup_ctrs             = &ppro_setup_ctrs,
 265         .check_ctrs             = &ppro_check_ctrs,
 266         .start                  = &ppro_start,
 267         .stop                   = &ppro_stop,
 268         .shutdown               = &ppro_shutdown
 269 };