2 * @file op_model_ppro.h
3 * Family 6 perfmon and architectural perfmon MSR operations
5 * @remark Copyright 2002 OProfile authors
6 * @remark Copyright 2008 Intel Corporation
7 * @remark Read the file COPYING
10 * @author Philippe Elie
11 * @author Graydon Hoare
13 * @author Robert Richter <robert.richter@amd.com>
16 #include <linux/oprofile.h>
17 #include <linux/slab.h>
18 #include <asm/ptrace.h>
23 #include "op_x86_model.h"
24 #include "op_counter.h"
26 static int num_counters = 2;
27 static int counter_width = 32;
29 #define MSR_PPRO_EVENTSEL_RESERVED ((0xFFFFFFFFULL<<32)|(1ULL<<21))
31 static u64 *reset_value;
33 static void ppro_fill_in_addresses(struct op_msrs * const msrs)
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;
41 msrs->counters[i].addr = 0;
44 for (i = 0; i < num_counters; i++) {
45 if (reserve_evntsel_nmi(MSR_P6_EVNTSEL0 + i))
46 msrs->controls[i].addr = MSR_P6_EVNTSEL0 + i;
48 msrs->controls[i].addr = 0;
53 static void ppro_setup_ctrs(struct op_x86_model_spec const *model,
54 struct op_msrs const * const msrs)
60 reset_value = kmalloc(sizeof(reset_value[0]) * num_counters,
66 if (cpu_has_arch_perfmon) {
67 union cpuid10_eax eax;
68 eax.full = cpuid_eax(0xa);
71 * For Core2 (family 6, model 15), don't reset the
74 if (!(eax.split.version_id == 0 &&
75 current_cpu_data.x86 == 6 &&
76 current_cpu_data.x86_model == 15)) {
78 if (counter_width < eax.split.bit_width)
79 counter_width = eax.split.bit_width;
83 /* clear all counters */
84 for (i = 0 ; i < num_counters; ++i) {
85 if (unlikely(!CTRL_IS_RESERVED(msrs, i)))
87 rdmsrl(msrs->controls[i].addr, val);
88 val &= model->reserved;
89 wrmsrl(msrs->controls[i].addr, val);
92 /* avoid a false detection of ctr overflows in NMI handler */
93 for (i = 0; i < num_counters; ++i) {
94 if (unlikely(!CTR_IS_RESERVED(msrs, i)))
96 wrmsrl(msrs->counters[i].addr, -1LL);
99 /* enable active counters */
100 for (i = 0; i < num_counters; ++i) {
101 if ((counter_config[i].enabled) && (CTR_IS_RESERVED(msrs, i))) {
102 reset_value[i] = counter_config[i].count;
103 wrmsrl(msrs->counters[i].addr, -reset_value[i]);
104 rdmsrl(msrs->controls[i].addr, val);
105 val &= model->reserved;
106 val |= op_x86_get_ctrl(model, &counter_config[i]);
107 wrmsrl(msrs->controls[i].addr, val);
115 static int ppro_check_ctrs(struct pt_regs * const regs,
116 struct op_msrs const * const msrs)
121 for (i = 0 ; i < num_counters; ++i) {
124 rdmsrl(msrs->counters[i].addr, val);
125 if (val & (1ULL << (counter_width - 1)))
127 oprofile_add_sample(regs, i);
128 wrmsrl(msrs->counters[i].addr, -reset_value[i]);
131 /* Only P6 based Pentium M need to re-unmask the apic vector but it
132 * doesn't hurt other P6 variant */
133 apic_write(APIC_LVTPC, apic_read(APIC_LVTPC) & ~APIC_LVT_MASKED);
135 /* We can't work out if we really handled an interrupt. We
136 * might have caught a *second* counter just after overflowing
137 * the interrupt for this counter then arrives
138 * and we don't find a counter that's overflowed, so we
139 * would return 0 and get dazed + confused. Instead we always
140 * assume we found an overflow. This sucks.
146 static void ppro_start(struct op_msrs const * const msrs)
148 unsigned int low, high;
153 for (i = 0; i < num_counters; ++i) {
154 if (reset_value[i]) {
155 rdmsr(msrs->controls[i].addr, low, high);
156 CTRL_SET_ACTIVE(low);
157 wrmsr(msrs->controls[i].addr, low, high);
163 static void ppro_stop(struct op_msrs const * const msrs)
165 unsigned int low, high;
170 for (i = 0; i < num_counters; ++i) {
173 rdmsr(msrs->controls[i].addr, low, high);
174 CTRL_SET_INACTIVE(low);
175 wrmsr(msrs->controls[i].addr, low, high);
179 static void ppro_shutdown(struct op_msrs const * const msrs)
183 for (i = 0 ; i < num_counters ; ++i) {
184 if (CTR_IS_RESERVED(msrs, i))
185 release_perfctr_nmi(MSR_P6_PERFCTR0 + i);
187 for (i = 0 ; i < num_counters ; ++i) {
188 if (CTRL_IS_RESERVED(msrs, i))
189 release_evntsel_nmi(MSR_P6_EVNTSEL0 + i);
198 struct op_x86_model_spec const op_ppro_spec = {
201 .reserved = MSR_PPRO_EVENTSEL_RESERVED,
202 .fill_in_addresses = &ppro_fill_in_addresses,
203 .setup_ctrs = &ppro_setup_ctrs,
204 .check_ctrs = &ppro_check_ctrs,
205 .start = &ppro_start,
207 .shutdown = &ppro_shutdown
211 * Architectural performance monitoring.
213 * Newer Intel CPUs (Core1+) have support for architectural
214 * events described in CPUID 0xA. See the IA32 SDM Vol3b.18 for details.
215 * The advantage of this is that it can be done without knowing about
219 static void arch_perfmon_setup_counters(void)
221 union cpuid10_eax eax;
223 eax.full = cpuid_eax(0xa);
225 /* Workaround for BIOS bugs in 6/15. Taken from perfmon2 */
226 if (eax.split.version_id == 0 && current_cpu_data.x86 == 6 &&
227 current_cpu_data.x86_model == 15) {
228 eax.split.version_id = 2;
229 eax.split.num_counters = 2;
230 eax.split.bit_width = 40;
233 num_counters = eax.split.num_counters;
235 op_arch_perfmon_spec.num_counters = num_counters;
236 op_arch_perfmon_spec.num_controls = num_counters;
239 static int arch_perfmon_init(struct oprofile_operations *ignore)
241 arch_perfmon_setup_counters();
245 struct op_x86_model_spec op_arch_perfmon_spec = {
246 .reserved = MSR_PPRO_EVENTSEL_RESERVED,
247 .init = &arch_perfmon_init,
248 /* num_counters/num_controls filled in at runtime */
249 .fill_in_addresses = &ppro_fill_in_addresses,
250 /* user space does the cpuid check for available events */
251 .setup_ctrs = &ppro_setup_ctrs,
252 .check_ctrs = &ppro_check_ctrs,
253 .start = &ppro_start,
255 .shutdown = &ppro_shutdown