1 #include <linux/perf_event.h>
2 #include <linux/types.h>
3
4 #include "perf_event.h"
5
6 /*
7 * Not sure about some of these
8 */
9 static const u64 p6_perfmon_event_map[] =
10 {
11 [PERF_COUNT_HW_CPU_CYCLES] = 0x0079, /* CPU_CLK_UNHALTED */
12 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0, /* INST_RETIRED */
13 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x0f2e, /* L2_RQSTS:M:E:S:I */
14 [PERF_COUNT_HW_CACHE_MISSES] = 0x012e, /* L2_RQSTS:I */
15 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4, /* BR_INST_RETIRED */
16 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5, /* BR_MISS_PRED_RETIRED */
17 [PERF_COUNT_HW_BUS_CYCLES] = 0x0062, /* BUS_DRDY_CLOCKS */
18 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x00a2, /* RESOURCE_STALLS */
19
20 };
21
22 static const u64 __initconst p6_hw_cache_event_ids
23 [PERF_COUNT_HW_CACHE_MAX]
24 [PERF_COUNT_HW_CACHE_OP_MAX]
25 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
26 {
27 [ C(L1D) ] = {
28 [ C(OP_READ) ] = {
29 [ C(RESULT_ACCESS) ] = 0x0043, /* DATA_MEM_REFS */
30 [ C(RESULT_MISS) ] = 0x0045, /* DCU_LINES_IN */
31 },
32 [ C(OP_WRITE) ] = {
33 [ C(RESULT_ACCESS) ] = 0,
34 [ C(RESULT_MISS) ] = 0x0f29, /* L2_LD:M:E:S:I */
35 },
36 [ C(OP_PREFETCH) ] = {
37 [ C(RESULT_ACCESS) ] = 0,
38 [ C(RESULT_MISS) ] = 0,
39 },
40 },
41 [ C(L1I ) ] = {
42 [ C(OP_READ) ] = {
43 [ C(RESULT_ACCESS) ] = 0x0080, /* IFU_IFETCH */
44 [ C(RESULT_MISS) ] = 0x0f28, /* L2_IFETCH:M:E:S:I */
45 },
46 [ C(OP_WRITE) ] = {
47 [ C(RESULT_ACCESS) ] = -1,
48 [ C(RESULT_MISS) ] = -1,
49 },
50 [ C(OP_PREFETCH) ] = {
51 [ C(RESULT_ACCESS) ] = 0,
52 [ C(RESULT_MISS) ] = 0,
53 },
54 },
55 [ C(LL ) ] = {
56 [ C(OP_READ) ] = {
57 [ C(RESULT_ACCESS) ] = 0,
58 [ C(RESULT_MISS) ] = 0,
59 },
60 [ C(OP_WRITE) ] = {
61 [ C(RESULT_ACCESS) ] = 0,
62 [ C(RESULT_MISS) ] = 0x0025, /* L2_M_LINES_INM */
63 },
64 [ C(OP_PREFETCH) ] = {
65 [ C(RESULT_ACCESS) ] = 0,
66 [ C(RESULT_MISS) ] = 0,
67 },
68 },
69 [ C(DTLB) ] = {
70 [ C(OP_READ) ] = {
71 [ C(RESULT_ACCESS) ] = 0x0043, /* DATA_MEM_REFS */
72 [ C(RESULT_MISS) ] = 0,
73 },
74 [ C(OP_WRITE) ] = {
75 [ C(RESULT_ACCESS) ] = 0,
76 [ C(RESULT_MISS) ] = 0,
77 },
78 [ C(OP_PREFETCH) ] = {
79 [ C(RESULT_ACCESS) ] = 0,
80 [ C(RESULT_MISS) ] = 0,
81 },
82 },
83 [ C(ITLB) ] = {
84 [ C(OP_READ) ] = {
85 [ C(RESULT_ACCESS) ] = 0x0080, /* IFU_IFETCH */
86 [ C(RESULT_MISS) ] = 0x0085, /* ITLB_MISS */
87 },
88 [ C(OP_WRITE) ] = {
89 [ C(RESULT_ACCESS) ] = -1,
90 [ C(RESULT_MISS) ] = -1,
91 },
92 [ C(OP_PREFETCH) ] = {
93 [ C(RESULT_ACCESS) ] = -1,
94 [ C(RESULT_MISS) ] = -1,
95 },
96 },
97 [ C(BPU ) ] = {
98 [ C(OP_READ) ] = {
99 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED */
100 [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISS_PRED_RETIRED */
101 },
102 [ C(OP_WRITE) ] = {
103 [ C(RESULT_ACCESS) ] = -1,
104 [ C(RESULT_MISS) ] = -1,
105 },
106 [ C(OP_PREFETCH) ] = {
107 [ C(RESULT_ACCESS) ] = -1,
108 [ C(RESULT_MISS) ] = -1,
109 },
110 },
111 };
112
p6_pmu_event_map(int hw_event)113 static u64 p6_pmu_event_map(int hw_event)
114 {
115 return p6_perfmon_event_map[hw_event];
116 }
117
118 /*
119 * Event setting that is specified not to count anything.
120 * We use this to effectively disable a counter.
121 *
122 * L2_RQSTS with 0 MESI unit mask.
123 */
124 #define P6_NOP_EVENT 0x0000002EULL
125
126 static struct event_constraint p6_event_constraints[] =
127 {
128 INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FLOPS */
129 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
130 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
131 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
132 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
133 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
134 EVENT_CONSTRAINT_END
135 };
136
p6_pmu_disable_all(void)137 static void p6_pmu_disable_all(void)
138 {
139 u64 val;
140
141 /* p6 only has one enable register */
142 rdmsrl(MSR_P6_EVNTSEL0, val);
143 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
144 wrmsrl(MSR_P6_EVNTSEL0, val);
145 }
146
p6_pmu_enable_all(int added)147 static void p6_pmu_enable_all(int added)
148 {
149 unsigned long val;
150
151 /* p6 only has one enable register */
152 rdmsrl(MSR_P6_EVNTSEL0, val);
153 val |= ARCH_PERFMON_EVENTSEL_ENABLE;
154 wrmsrl(MSR_P6_EVNTSEL0, val);
155 }
156
157 static inline void
p6_pmu_disable_event(struct perf_event * event)158 p6_pmu_disable_event(struct perf_event *event)
159 {
160 struct hw_perf_event *hwc = &event->hw;
161 u64 val = P6_NOP_EVENT;
162
163 (void)wrmsrl_safe(hwc->config_base, val);
164 }
165
p6_pmu_enable_event(struct perf_event * event)166 static void p6_pmu_enable_event(struct perf_event *event)
167 {
168 struct hw_perf_event *hwc = &event->hw;
169 u64 val;
170
171 val = hwc->config;
172
173 /*
174 * p6 only has a global event enable, set on PerfEvtSel0
175 * We "disable" events by programming P6_NOP_EVENT
176 * and we rely on p6_pmu_enable_all() being called
177 * to actually enable the events.
178 */
179
180 (void)wrmsrl_safe(hwc->config_base, val);
181 }
182
183 PMU_FORMAT_ATTR(event, "config:0-7" );
184 PMU_FORMAT_ATTR(umask, "config:8-15" );
185 PMU_FORMAT_ATTR(edge, "config:18" );
186 PMU_FORMAT_ATTR(pc, "config:19" );
187 PMU_FORMAT_ATTR(inv, "config:23" );
188 PMU_FORMAT_ATTR(cmask, "config:24-31" );
189
190 static struct attribute *intel_p6_formats_attr[] = {
191 &format_attr_event.attr,
192 &format_attr_umask.attr,
193 &format_attr_edge.attr,
194 &format_attr_pc.attr,
195 &format_attr_inv.attr,
196 &format_attr_cmask.attr,
197 NULL,
198 };
199
200 static __initconst const struct x86_pmu p6_pmu = {
201 .name = "p6",
202 .handle_irq = x86_pmu_handle_irq,
203 .disable_all = p6_pmu_disable_all,
204 .enable_all = p6_pmu_enable_all,
205 .enable = p6_pmu_enable_event,
206 .disable = p6_pmu_disable_event,
207 .hw_config = x86_pmu_hw_config,
208 .schedule_events = x86_schedule_events,
209 .eventsel = MSR_P6_EVNTSEL0,
210 .perfctr = MSR_P6_PERFCTR0,
211 .event_map = p6_pmu_event_map,
212 .max_events = ARRAY_SIZE(p6_perfmon_event_map),
213 .apic = 1,
214 .max_period = (1ULL << 31) - 1,
215 .version = 0,
216 .num_counters = 2,
217 /*
218 * Events have 40 bits implemented. However they are designed such
219 * that bits [32-39] are sign extensions of bit 31. As such the
220 * effective width of a event for P6-like PMU is 32 bits only.
221 *
222 * See IA-32 Intel Architecture Software developer manual Vol 3B
223 */
224 .cntval_bits = 32,
225 .cntval_mask = (1ULL << 32) - 1,
226 .get_event_constraints = x86_get_event_constraints,
227 .event_constraints = p6_event_constraints,
228
229 .format_attrs = intel_p6_formats_attr,
230 .events_sysfs_show = intel_event_sysfs_show,
231
232 };
233
p6_pmu_rdpmc_quirk(void)234 static __init void p6_pmu_rdpmc_quirk(void)
235 {
236 if (boot_cpu_data.x86_stepping < 9) {
237 /*
238 * PPro erratum 26; fixed in stepping 9 and above.
239 */
240 pr_warn("Userspace RDPMC support disabled due to a CPU erratum\n");
241 x86_pmu.attr_rdpmc_broken = 1;
242 x86_pmu.attr_rdpmc = 0;
243 }
244 }
245
p6_pmu_init(void)246 __init int p6_pmu_init(void)
247 {
248 x86_pmu = p6_pmu;
249
250 switch (boot_cpu_data.x86_model) {
251 case 1: /* Pentium Pro */
252 x86_add_quirk(p6_pmu_rdpmc_quirk);
253 break;
254
255 case 3: /* Pentium II - Klamath */
256 case 5: /* Pentium II - Deschutes */
257 case 6: /* Pentium II - Mendocino */
258 break;
259
260 case 7: /* Pentium III - Katmai */
261 case 8: /* Pentium III - Coppermine */
262 case 10: /* Pentium III Xeon */
263 case 11: /* Pentium III - Tualatin */
264 break;
265
266 case 9: /* Pentium M - Banias */
267 case 13: /* Pentium M - Dothan */
268 break;
269
270 default:
271 pr_cont("unsupported p6 CPU model %d ", boot_cpu_data.x86_model);
272 return -ENODEV;
273 }
274
275 memcpy(hw_cache_event_ids, p6_hw_cache_event_ids,
276 sizeof(hw_cache_event_ids));
277
278 return 0;
279 }
280