1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/perf_event.h>
3 #include <linux/types.h>
4
5 #include "../perf_event.h"
6
7 /*
8 * Not sure about some of these
9 */
10 static const u64 p6_perfmon_event_map[] =
11 {
12 [PERF_COUNT_HW_CPU_CYCLES] = 0x0079, /* CPU_CLK_UNHALTED */
13 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0, /* INST_RETIRED */
14 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x0f2e, /* L2_RQSTS:M:E:S:I */
15 [PERF_COUNT_HW_CACHE_MISSES] = 0x012e, /* L2_RQSTS:I */
16 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4, /* BR_INST_RETIRED */
17 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5, /* BR_MISS_PRED_RETIRED */
18 [PERF_COUNT_HW_BUS_CYCLES] = 0x0062, /* BUS_DRDY_CLOCKS */
19 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x00a2, /* RESOURCE_STALLS */
20
21 };
22
23 static const u64 __initconst p6_hw_cache_event_ids
24 [PERF_COUNT_HW_CACHE_MAX]
25 [PERF_COUNT_HW_CACHE_OP_MAX]
26 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
27 {
28 [ C(L1D) ] = {
29 [ C(OP_READ) ] = {
30 [ C(RESULT_ACCESS) ] = 0x0043, /* DATA_MEM_REFS */
31 [ C(RESULT_MISS) ] = 0x0045, /* DCU_LINES_IN */
32 },
33 [ C(OP_WRITE) ] = {
34 [ C(RESULT_ACCESS) ] = 0,
35 [ C(RESULT_MISS) ] = 0x0f29, /* L2_LD:M:E:S:I */
36 },
37 [ C(OP_PREFETCH) ] = {
38 [ C(RESULT_ACCESS) ] = 0,
39 [ C(RESULT_MISS) ] = 0,
40 },
41 },
42 [ C(L1I ) ] = {
43 [ C(OP_READ) ] = {
44 [ C(RESULT_ACCESS) ] = 0x0080, /* IFU_IFETCH */
45 [ C(RESULT_MISS) ] = 0x0f28, /* L2_IFETCH:M:E:S:I */
46 },
47 [ C(OP_WRITE) ] = {
48 [ C(RESULT_ACCESS) ] = -1,
49 [ C(RESULT_MISS) ] = -1,
50 },
51 [ C(OP_PREFETCH) ] = {
52 [ C(RESULT_ACCESS) ] = 0,
53 [ C(RESULT_MISS) ] = 0,
54 },
55 },
56 [ C(LL ) ] = {
57 [ C(OP_READ) ] = {
58 [ C(RESULT_ACCESS) ] = 0,
59 [ C(RESULT_MISS) ] = 0,
60 },
61 [ C(OP_WRITE) ] = {
62 [ C(RESULT_ACCESS) ] = 0,
63 [ C(RESULT_MISS) ] = 0x0025, /* L2_M_LINES_INM */
64 },
65 [ C(OP_PREFETCH) ] = {
66 [ C(RESULT_ACCESS) ] = 0,
67 [ C(RESULT_MISS) ] = 0,
68 },
69 },
70 [ C(DTLB) ] = {
71 [ C(OP_READ) ] = {
72 [ C(RESULT_ACCESS) ] = 0x0043, /* DATA_MEM_REFS */
73 [ C(RESULT_MISS) ] = 0,
74 },
75 [ C(OP_WRITE) ] = {
76 [ C(RESULT_ACCESS) ] = 0,
77 [ C(RESULT_MISS) ] = 0,
78 },
79 [ C(OP_PREFETCH) ] = {
80 [ C(RESULT_ACCESS) ] = 0,
81 [ C(RESULT_MISS) ] = 0,
82 },
83 },
84 [ C(ITLB) ] = {
85 [ C(OP_READ) ] = {
86 [ C(RESULT_ACCESS) ] = 0x0080, /* IFU_IFETCH */
87 [ C(RESULT_MISS) ] = 0x0085, /* ITLB_MISS */
88 },
89 [ C(OP_WRITE) ] = {
90 [ C(RESULT_ACCESS) ] = -1,
91 [ C(RESULT_MISS) ] = -1,
92 },
93 [ C(OP_PREFETCH) ] = {
94 [ C(RESULT_ACCESS) ] = -1,
95 [ C(RESULT_MISS) ] = -1,
96 },
97 },
98 [ C(BPU ) ] = {
99 [ C(OP_READ) ] = {
100 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED */
101 [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISS_PRED_RETIRED */
102 },
103 [ C(OP_WRITE) ] = {
104 [ C(RESULT_ACCESS) ] = -1,
105 [ C(RESULT_MISS) ] = -1,
106 },
107 [ C(OP_PREFETCH) ] = {
108 [ C(RESULT_ACCESS) ] = -1,
109 [ C(RESULT_MISS) ] = -1,
110 },
111 },
112 };
113
p6_pmu_event_map(int hw_event)114 static u64 p6_pmu_event_map(int hw_event)
115 {
116 return p6_perfmon_event_map[hw_event];
117 }
118
119 /*
120 * Event setting that is specified not to count anything.
121 * We use this to effectively disable a counter.
122 *
123 * L2_RQSTS with 0 MESI unit mask.
124 */
125 #define P6_NOP_EVENT 0x0000002EULL
126
127 static struct event_constraint p6_event_constraints[] =
128 {
129 INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FLOPS */
130 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
131 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
132 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
133 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
134 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
135 EVENT_CONSTRAINT_END
136 };
137
p6_pmu_disable_all(void)138 static void p6_pmu_disable_all(void)
139 {
140 u64 val;
141
142 /* p6 only has one enable register */
143 rdmsrl(MSR_P6_EVNTSEL0, val);
144 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
145 wrmsrl(MSR_P6_EVNTSEL0, val);
146 }
147
p6_pmu_enable_all(int added)148 static void p6_pmu_enable_all(int added)
149 {
150 unsigned long val;
151
152 /* p6 only has one enable register */
153 rdmsrl(MSR_P6_EVNTSEL0, val);
154 val |= ARCH_PERFMON_EVENTSEL_ENABLE;
155 wrmsrl(MSR_P6_EVNTSEL0, val);
156 }
157
158 static inline void
p6_pmu_disable_event(struct perf_event * event)159 p6_pmu_disable_event(struct perf_event *event)
160 {
161 struct hw_perf_event *hwc = &event->hw;
162 u64 val = P6_NOP_EVENT;
163
164 (void)wrmsrl_safe(hwc->config_base, val);
165 }
166
p6_pmu_enable_event(struct perf_event * event)167 static void p6_pmu_enable_event(struct perf_event *event)
168 {
169 struct hw_perf_event *hwc = &event->hw;
170 u64 val;
171
172 val = hwc->config;
173
174 /*
175 * p6 only has a global event enable, set on PerfEvtSel0
176 * We "disable" events by programming P6_NOP_EVENT
177 * and we rely on p6_pmu_enable_all() being called
178 * to actually enable the events.
179 */
180
181 (void)wrmsrl_safe(hwc->config_base, val);
182 }
183
184 PMU_FORMAT_ATTR(event, "config:0-7" );
185 PMU_FORMAT_ATTR(umask, "config:8-15" );
186 PMU_FORMAT_ATTR(edge, "config:18" );
187 PMU_FORMAT_ATTR(pc, "config:19" );
188 PMU_FORMAT_ATTR(inv, "config:23" );
189 PMU_FORMAT_ATTR(cmask, "config:24-31" );
190
191 static struct attribute *intel_p6_formats_attr[] = {
192 &format_attr_event.attr,
193 &format_attr_umask.attr,
194 &format_attr_edge.attr,
195 &format_attr_pc.attr,
196 &format_attr_inv.attr,
197 &format_attr_cmask.attr,
198 NULL,
199 };
200
201 static __initconst const struct x86_pmu p6_pmu = {
202 .name = "p6",
203 .handle_irq = x86_pmu_handle_irq,
204 .disable_all = p6_pmu_disable_all,
205 .enable_all = p6_pmu_enable_all,
206 .enable = p6_pmu_enable_event,
207 .disable = p6_pmu_disable_event,
208 .hw_config = x86_pmu_hw_config,
209 .schedule_events = x86_schedule_events,
210 .eventsel = MSR_P6_EVNTSEL0,
211 .perfctr = MSR_P6_PERFCTR0,
212 .event_map = p6_pmu_event_map,
213 .max_events = ARRAY_SIZE(p6_perfmon_event_map),
214 .apic = 1,
215 .max_period = (1ULL << 31) - 1,
216 .version = 0,
217 .num_counters = 2,
218 /*
219 * Events have 40 bits implemented. However they are designed such
220 * that bits [32-39] are sign extensions of bit 31. As such the
221 * effective width of a event for P6-like PMU is 32 bits only.
222 *
223 * See IA-32 Intel Architecture Software developer manual Vol 3B
224 */
225 .cntval_bits = 32,
226 .cntval_mask = (1ULL << 32) - 1,
227 .get_event_constraints = x86_get_event_constraints,
228 .event_constraints = p6_event_constraints,
229
230 .format_attrs = intel_p6_formats_attr,
231 .events_sysfs_show = intel_event_sysfs_show,
232
233 };
234
p6_pmu_rdpmc_quirk(void)235 static __init void p6_pmu_rdpmc_quirk(void)
236 {
237 if (boot_cpu_data.x86_stepping < 9) {
238 /*
239 * PPro erratum 26; fixed in stepping 9 and above.
240 */
241 pr_warn("Userspace RDPMC support disabled due to a CPU erratum\n");
242 x86_pmu.attr_rdpmc_broken = 1;
243 x86_pmu.attr_rdpmc = 0;
244 }
245 }
246
p6_pmu_init(void)247 __init int p6_pmu_init(void)
248 {
249 x86_pmu = p6_pmu;
250
251 switch (boot_cpu_data.x86_model) {
252 case 1: /* Pentium Pro */
253 x86_add_quirk(p6_pmu_rdpmc_quirk);
254 break;
255
256 case 3: /* Pentium II - Klamath */
257 case 5: /* Pentium II - Deschutes */
258 case 6: /* Pentium II - Mendocino */
259 break;
260
261 case 7: /* Pentium III - Katmai */
262 case 8: /* Pentium III - Coppermine */
263 case 10: /* Pentium III Xeon */
264 case 11: /* Pentium III - Tualatin */
265 break;
266
267 case 9: /* Pentium M - Banias */
268 case 13: /* Pentium M - Dothan */
269 break;
270
271 default:
272 pr_cont("unsupported p6 CPU model %d ", boot_cpu_data.x86_model);
273 return -ENODEV;
274 }
275
276 memcpy(hw_cache_event_ids, p6_hw_cache_event_ids,
277 sizeof(hw_cache_event_ids));
278
279 return 0;
280 }
281