Merge git://git.infradead.org/mtd-2.6
[pandora-kernel.git] / 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/perf_counter.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_OVERFLOWED(n) (!((n) & (1ULL<<(counter_width-1))))
31
32 #define CTRL_IS_RESERVED(msrs, c) (msrs->controls[(c)].addr ? 1 : 0)
33 #define CTRL_READ(l, h, msrs, c) do {rdmsr((msrs->controls[(c)].addr), (l), (h)); } while (0)
34 #define CTRL_WRITE(l, h, msrs, c) do {wrmsr((msrs->controls[(c)].addr), (l), (h)); } while (0)
35 #define CTRL_SET_ACTIVE(n) (n |= (1<<22))
36 #define CTRL_SET_INACTIVE(n) (n &= ~(1<<22))
37 #define CTRL_CLEAR(x) (x &= (1<<21))
38 #define CTRL_SET_ENABLE(val) (val |= 1<<20)
39 #define CTRL_SET_USR(val, u) (val |= ((u & 1) << 16))
40 #define CTRL_SET_KERN(val, k) (val |= ((k & 1) << 17))
41 #define CTRL_SET_UM(val, m) (val |= (m << 8))
42 #define CTRL_SET_EVENT(val, e) (val |= e)
43
44 static u64 *reset_value;
45
46 static void ppro_fill_in_addresses(struct op_msrs * const msrs)
47 {
48         int i;
49
50         for (i = 0; i < num_counters; i++) {
51                 if (reserve_perfctr_nmi(MSR_P6_PERFCTR0 + i))
52                         msrs->counters[i].addr = MSR_P6_PERFCTR0 + i;
53                 else
54                         msrs->counters[i].addr = 0;
55         }
56
57         for (i = 0; i < num_counters; i++) {
58                 if (reserve_evntsel_nmi(MSR_P6_EVNTSEL0 + i))
59                         msrs->controls[i].addr = MSR_P6_EVNTSEL0 + i;
60                 else
61                         msrs->controls[i].addr = 0;
62         }
63 }
64
65
66 static void ppro_setup_ctrs(struct op_msrs const * const msrs)
67 {
68         unsigned int low, high;
69         int i;
70
71         if (!reset_value) {
72                 reset_value = kmalloc(sizeof(reset_value[0]) * num_counters,
73                                         GFP_ATOMIC);
74                 if (!reset_value)
75                         return;
76         }
77
78         if (cpu_has_arch_perfmon) {
79                 union cpuid10_eax eax;
80                 eax.full = cpuid_eax(0xa);
81
82                 /*
83                  * For Core2 (family 6, model 15), don't reset the
84                  * counter width:
85                  */
86                 if (!(eax.split.version_id == 0 &&
87                         current_cpu_data.x86 == 6 &&
88                                 current_cpu_data.x86_model == 15)) {
89
90                         if (counter_width < eax.split.bit_width)
91                                 counter_width = eax.split.bit_width;
92                 }
93         }
94
95         /* clear all counters */
96         for (i = 0 ; i < num_counters; ++i) {
97                 if (unlikely(!CTRL_IS_RESERVED(msrs, i)))
98                         continue;
99                 CTRL_READ(low, high, msrs, i);
100                 CTRL_CLEAR(low);
101                 CTRL_WRITE(low, high, msrs, i);
102         }
103
104         /* avoid a false detection of ctr overflows in NMI handler */
105         for (i = 0; i < num_counters; ++i) {
106                 if (unlikely(!CTR_IS_RESERVED(msrs, i)))
107                         continue;
108                 wrmsrl(msrs->counters[i].addr, -1LL);
109         }
110
111         /* enable active counters */
112         for (i = 0; i < num_counters; ++i) {
113                 if ((counter_config[i].enabled) && (CTR_IS_RESERVED(msrs, i))) {
114                         reset_value[i] = counter_config[i].count;
115
116                         wrmsrl(msrs->counters[i].addr, -reset_value[i]);
117
118                         CTRL_READ(low, high, msrs, i);
119                         CTRL_CLEAR(low);
120                         CTRL_SET_ENABLE(low);
121                         CTRL_SET_USR(low, counter_config[i].user);
122                         CTRL_SET_KERN(low, counter_config[i].kernel);
123                         CTRL_SET_UM(low, counter_config[i].unit_mask);
124                         CTRL_SET_EVENT(low, counter_config[i].event);
125                         CTRL_WRITE(low, high, msrs, i);
126                 } else {
127                         reset_value[i] = 0;
128                 }
129         }
130 }
131
132
133 static int ppro_check_ctrs(struct pt_regs * const regs,
134                            struct op_msrs const * const msrs)
135 {
136         u64 val;
137         int i;
138
139         /*
140          * This can happen if perf counters are in use when
141          * we steal the die notifier NMI.
142          */
143         if (unlikely(!reset_value))
144                 goto out;
145
146         for (i = 0 ; i < num_counters; ++i) {
147                 if (!reset_value[i])
148                         continue;
149                 rdmsrl(msrs->counters[i].addr, val);
150                 if (CTR_OVERFLOWED(val)) {
151                         oprofile_add_sample(regs, i);
152                         wrmsrl(msrs->counters[i].addr, -reset_value[i]);
153                 }
154         }
155
156 out:
157         /* Only P6 based Pentium M need to re-unmask the apic vector but it
158          * doesn't hurt other P6 variant */
159         apic_write(APIC_LVTPC, apic_read(APIC_LVTPC) & ~APIC_LVT_MASKED);
160
161         /* We can't work out if we really handled an interrupt. We
162          * might have caught a *second* counter just after overflowing
163          * the interrupt for this counter then arrives
164          * and we don't find a counter that's overflowed, so we
165          * would return 0 and get dazed + confused. Instead we always
166          * assume we found an overflow. This sucks.
167          */
168         return 1;
169 }
170
171
172 static void ppro_start(struct op_msrs const * const msrs)
173 {
174         unsigned int low, high;
175         int i;
176
177         if (!reset_value)
178                 return;
179         for (i = 0; i < num_counters; ++i) {
180                 if (reset_value[i]) {
181                         CTRL_READ(low, high, msrs, i);
182                         CTRL_SET_ACTIVE(low);
183                         CTRL_WRITE(low, high, msrs, i);
184                 }
185         }
186 }
187
188
189 static void ppro_stop(struct op_msrs const * const msrs)
190 {
191         unsigned int low, high;
192         int i;
193
194         if (!reset_value)
195                 return;
196         for (i = 0; i < num_counters; ++i) {
197                 if (!reset_value[i])
198                         continue;
199                 CTRL_READ(low, high, msrs, i);
200                 CTRL_SET_INACTIVE(low);
201                 CTRL_WRITE(low, high, msrs, i);
202         }
203 }
204
205 static void ppro_shutdown(struct op_msrs const * const msrs)
206 {
207         int i;
208
209         for (i = 0 ; i < num_counters ; ++i) {
210                 if (CTR_IS_RESERVED(msrs, i))
211                         release_perfctr_nmi(MSR_P6_PERFCTR0 + i);
212         }
213         for (i = 0 ; i < num_counters ; ++i) {
214                 if (CTRL_IS_RESERVED(msrs, i))
215                         release_evntsel_nmi(MSR_P6_EVNTSEL0 + i);
216         }
217         if (reset_value) {
218                 kfree(reset_value);
219                 reset_value = NULL;
220         }
221 }
222
223
224 struct op_x86_model_spec op_ppro_spec = {
225         .num_counters           = 2,    /* can be overriden */
226         .num_controls           = 2,    /* dito */
227         .fill_in_addresses      = &ppro_fill_in_addresses,
228         .setup_ctrs             = &ppro_setup_ctrs,
229         .check_ctrs             = &ppro_check_ctrs,
230         .start                  = &ppro_start,
231         .stop                   = &ppro_stop,
232         .shutdown               = &ppro_shutdown
233 };
234
235 /*
236  * Architectural performance monitoring.
237  *
238  * Newer Intel CPUs (Core1+) have support for architectural
239  * events described in CPUID 0xA. See the IA32 SDM Vol3b.18 for details.
240  * The advantage of this is that it can be done without knowing about
241  * the specific CPU.
242  */
243
244 void arch_perfmon_setup_counters(void)
245 {
246         union cpuid10_eax eax;
247
248         eax.full = cpuid_eax(0xa);
249
250         /* Workaround for BIOS bugs in 6/15. Taken from perfmon2 */
251         if (eax.split.version_id == 0 && current_cpu_data.x86 == 6 &&
252                 current_cpu_data.x86_model == 15) {
253                 eax.split.version_id = 2;
254                 eax.split.num_counters = 2;
255                 eax.split.bit_width = 40;
256         }
257
258         num_counters = eax.split.num_counters;
259
260         op_arch_perfmon_spec.num_counters = num_counters;
261         op_arch_perfmon_spec.num_controls = num_counters;
262         op_ppro_spec.num_counters = num_counters;
263         op_ppro_spec.num_controls = num_counters;
264 }
265
266 struct op_x86_model_spec op_arch_perfmon_spec = {
267         /* num_counters/num_controls filled in at runtime */
268         .fill_in_addresses      = &ppro_fill_in_addresses,
269         /* user space does the cpuid check for available events */
270         .setup_ctrs             = &ppro_setup_ctrs,
271         .check_ctrs             = &ppro_check_ctrs,
272         .start                  = &ppro_start,
273         .stop                   = &ppro_stop,
274         .shutdown               = &ppro_shutdown
275 };