1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Per core/cpu state
4  *
5  * Used to coordinate shared registers between HT threads or
6  * among events on a single PMU.
7  */
8 
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10 
11 #include <linux/stddef.h>
12 #include <linux/types.h>
13 #include <linux/init.h>
14 #include <linux/slab.h>
15 #include <linux/export.h>
16 #include <linux/nmi.h>
17 #include <linux/kvm_host.h>
18 
19 #include <asm/cpufeature.h>
20 #include <asm/debugreg.h>
21 #include <asm/hardirq.h>
22 #include <asm/intel-family.h>
23 #include <asm/intel_pt.h>
24 #include <asm/apic.h>
25 #include <asm/cpu_device_id.h>
26 
27 #include "../perf_event.h"
28 
29 /*
30  * Intel PerfMon, used on Core and later.
31  */
32 static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly =
33 {
34 	[PERF_COUNT_HW_CPU_CYCLES]		= 0x003c,
35 	[PERF_COUNT_HW_INSTRUCTIONS]		= 0x00c0,
36 	[PERF_COUNT_HW_CACHE_REFERENCES]	= 0x4f2e,
37 	[PERF_COUNT_HW_CACHE_MISSES]		= 0x412e,
38 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0x00c4,
39 	[PERF_COUNT_HW_BRANCH_MISSES]		= 0x00c5,
40 	[PERF_COUNT_HW_BUS_CYCLES]		= 0x013c,
41 	[PERF_COUNT_HW_REF_CPU_CYCLES]		= 0x0300, /* pseudo-encoding */
42 };
43 
44 static struct event_constraint intel_core_event_constraints[] __read_mostly =
45 {
46 	INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
47 	INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
48 	INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
49 	INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
50 	INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
51 	INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */
52 	EVENT_CONSTRAINT_END
53 };
54 
55 static struct event_constraint intel_core2_event_constraints[] __read_mostly =
56 {
57 	FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
58 	FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
59 	FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
60 	INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
61 	INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
62 	INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
63 	INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
64 	INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
65 	INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */
66 	INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
67 	INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */
68 	INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */
69 	INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */
70 	EVENT_CONSTRAINT_END
71 };
72 
73 static struct event_constraint intel_nehalem_event_constraints[] __read_mostly =
74 {
75 	FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
76 	FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
77 	FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
78 	INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */
79 	INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */
80 	INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */
81 	INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */
82 	INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */
83 	INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */
84 	INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
85 	INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
86 	EVENT_CONSTRAINT_END
87 };
88 
89 static struct extra_reg intel_nehalem_extra_regs[] __read_mostly =
90 {
91 	/* must define OFFCORE_RSP_X first, see intel_fixup_er() */
92 	INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0),
93 	INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b),
94 	EVENT_EXTRA_END
95 };
96 
97 static struct event_constraint intel_westmere_event_constraints[] __read_mostly =
98 {
99 	FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
100 	FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
101 	FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
102 	INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
103 	INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */
104 	INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
105 	INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */
106 	EVENT_CONSTRAINT_END
107 };
108 
109 static struct event_constraint intel_snb_event_constraints[] __read_mostly =
110 {
111 	FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
112 	FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
113 	FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
114 	INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */
115 	INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */
116 	INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
117 	INTEL_UEVENT_CONSTRAINT(0x06a3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
118 	INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */
119 	INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
120 	INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
121 	INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */
122 	INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
123 
124 	/*
125 	 * When HT is off these events can only run on the bottom 4 counters
126 	 * When HT is on, they are impacted by the HT bug and require EXCL access
127 	 */
128 	INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
129 	INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
130 	INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
131 	INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
132 
133 	EVENT_CONSTRAINT_END
134 };
135 
136 static struct event_constraint intel_ivb_event_constraints[] __read_mostly =
137 {
138 	FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
139 	FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
140 	FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
141 	INTEL_UEVENT_CONSTRAINT(0x0148, 0x4), /* L1D_PEND_MISS.PENDING */
142 	INTEL_UEVENT_CONSTRAINT(0x0279, 0xf), /* IDQ.EMPTY */
143 	INTEL_UEVENT_CONSTRAINT(0x019c, 0xf), /* IDQ_UOPS_NOT_DELIVERED.CORE */
144 	INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_LDM_PENDING */
145 	INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */
146 	INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */
147 	INTEL_UEVENT_CONSTRAINT(0x06a3, 0xf), /* CYCLE_ACTIVITY.STALLS_LDM_PENDING */
148 	INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
149 	INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
150 	INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
151 
152 	/*
153 	 * When HT is off these events can only run on the bottom 4 counters
154 	 * When HT is on, they are impacted by the HT bug and require EXCL access
155 	 */
156 	INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
157 	INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
158 	INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
159 	INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
160 
161 	EVENT_CONSTRAINT_END
162 };
163 
164 static struct extra_reg intel_westmere_extra_regs[] __read_mostly =
165 {
166 	/* must define OFFCORE_RSP_X first, see intel_fixup_er() */
167 	INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0),
168 	INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0xffff, RSP_1),
169 	INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b),
170 	EVENT_EXTRA_END
171 };
172 
173 static struct event_constraint intel_v1_event_constraints[] __read_mostly =
174 {
175 	EVENT_CONSTRAINT_END
176 };
177 
178 static struct event_constraint intel_gen_event_constraints[] __read_mostly =
179 {
180 	FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
181 	FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
182 	FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
183 	EVENT_CONSTRAINT_END
184 };
185 
186 static struct event_constraint intel_v5_gen_event_constraints[] __read_mostly =
187 {
188 	FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
189 	FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
190 	FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
191 	FIXED_EVENT_CONSTRAINT(0x0400, 3), /* SLOTS */
192 	FIXED_EVENT_CONSTRAINT(0x0500, 4),
193 	FIXED_EVENT_CONSTRAINT(0x0600, 5),
194 	FIXED_EVENT_CONSTRAINT(0x0700, 6),
195 	FIXED_EVENT_CONSTRAINT(0x0800, 7),
196 	FIXED_EVENT_CONSTRAINT(0x0900, 8),
197 	FIXED_EVENT_CONSTRAINT(0x0a00, 9),
198 	FIXED_EVENT_CONSTRAINT(0x0b00, 10),
199 	FIXED_EVENT_CONSTRAINT(0x0c00, 11),
200 	FIXED_EVENT_CONSTRAINT(0x0d00, 12),
201 	FIXED_EVENT_CONSTRAINT(0x0e00, 13),
202 	FIXED_EVENT_CONSTRAINT(0x0f00, 14),
203 	FIXED_EVENT_CONSTRAINT(0x1000, 15),
204 	EVENT_CONSTRAINT_END
205 };
206 
207 static struct event_constraint intel_slm_event_constraints[] __read_mostly =
208 {
209 	FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
210 	FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
211 	FIXED_EVENT_CONSTRAINT(0x0300, 2), /* pseudo CPU_CLK_UNHALTED.REF */
212 	EVENT_CONSTRAINT_END
213 };
214 
215 static struct event_constraint intel_grt_event_constraints[] __read_mostly = {
216 	FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
217 	FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
218 	FIXED_EVENT_CONSTRAINT(0x0300, 2), /* pseudo CPU_CLK_UNHALTED.REF */
219 	FIXED_EVENT_CONSTRAINT(0x013c, 2), /* CPU_CLK_UNHALTED.REF_TSC_P */
220 	EVENT_CONSTRAINT_END
221 };
222 
223 static struct event_constraint intel_skt_event_constraints[] __read_mostly = {
224 	FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
225 	FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
226 	FIXED_EVENT_CONSTRAINT(0x0300, 2), /* pseudo CPU_CLK_UNHALTED.REF */
227 	FIXED_EVENT_CONSTRAINT(0x013c, 2), /* CPU_CLK_UNHALTED.REF_TSC_P */
228 	FIXED_EVENT_CONSTRAINT(0x0073, 4), /* TOPDOWN_BAD_SPECULATION.ALL */
229 	FIXED_EVENT_CONSTRAINT(0x019c, 5), /* TOPDOWN_FE_BOUND.ALL */
230 	FIXED_EVENT_CONSTRAINT(0x02c2, 6), /* TOPDOWN_RETIRING.ALL */
231 	EVENT_CONSTRAINT_END
232 };
233 
234 static struct event_constraint intel_skl_event_constraints[] = {
235 	FIXED_EVENT_CONSTRAINT(0x00c0, 0),	/* INST_RETIRED.ANY */
236 	FIXED_EVENT_CONSTRAINT(0x003c, 1),	/* CPU_CLK_UNHALTED.CORE */
237 	FIXED_EVENT_CONSTRAINT(0x0300, 2),	/* CPU_CLK_UNHALTED.REF */
238 	INTEL_UEVENT_CONSTRAINT(0x1c0, 0x2),	/* INST_RETIRED.PREC_DIST */
239 
240 	/*
241 	 * when HT is off, these can only run on the bottom 4 counters
242 	 */
243 	INTEL_EVENT_CONSTRAINT(0xd0, 0xf),	/* MEM_INST_RETIRED.* */
244 	INTEL_EVENT_CONSTRAINT(0xd1, 0xf),	/* MEM_LOAD_RETIRED.* */
245 	INTEL_EVENT_CONSTRAINT(0xd2, 0xf),	/* MEM_LOAD_L3_HIT_RETIRED.* */
246 	INTEL_EVENT_CONSTRAINT(0xcd, 0xf),	/* MEM_TRANS_RETIRED.* */
247 	INTEL_EVENT_CONSTRAINT(0xc6, 0xf),	/* FRONTEND_RETIRED.* */
248 
249 	EVENT_CONSTRAINT_END
250 };
251 
252 static struct extra_reg intel_knl_extra_regs[] __read_mostly = {
253 	INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x799ffbb6e7ull, RSP_0),
254 	INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x399ffbffe7ull, RSP_1),
255 	EVENT_EXTRA_END
256 };
257 
258 static struct extra_reg intel_snb_extra_regs[] __read_mostly = {
259 	/* must define OFFCORE_RSP_X first, see intel_fixup_er() */
260 	INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3f807f8fffull, RSP_0),
261 	INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3f807f8fffull, RSP_1),
262 	INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
263 	EVENT_EXTRA_END
264 };
265 
266 static struct extra_reg intel_snbep_extra_regs[] __read_mostly = {
267 	/* must define OFFCORE_RSP_X first, see intel_fixup_er() */
268 	INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0),
269 	INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1),
270 	INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
271 	EVENT_EXTRA_END
272 };
273 
274 static struct extra_reg intel_skl_extra_regs[] __read_mostly = {
275 	INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0),
276 	INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1),
277 	INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
278 	/*
279 	 * Note the low 8 bits eventsel code is not a continuous field, containing
280 	 * some #GPing bits. These are masked out.
281 	 */
282 	INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE),
283 	EVENT_EXTRA_END
284 };
285 
286 static struct event_constraint intel_icl_event_constraints[] = {
287 	FIXED_EVENT_CONSTRAINT(0x00c0, 0),	/* INST_RETIRED.ANY */
288 	FIXED_EVENT_CONSTRAINT(0x01c0, 0),	/* old INST_RETIRED.PREC_DIST */
289 	FIXED_EVENT_CONSTRAINT(0x0100, 0),	/* INST_RETIRED.PREC_DIST */
290 	FIXED_EVENT_CONSTRAINT(0x003c, 1),	/* CPU_CLK_UNHALTED.CORE */
291 	FIXED_EVENT_CONSTRAINT(0x0300, 2),	/* CPU_CLK_UNHALTED.REF */
292 	FIXED_EVENT_CONSTRAINT(0x0400, 3),	/* SLOTS */
293 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_RETIRING, 0),
294 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BAD_SPEC, 1),
295 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FE_BOUND, 2),
296 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BE_BOUND, 3),
297 	INTEL_EVENT_CONSTRAINT_RANGE(0x03, 0x0a, 0xf),
298 	INTEL_EVENT_CONSTRAINT_RANGE(0x1f, 0x28, 0xf),
299 	INTEL_EVENT_CONSTRAINT(0x32, 0xf),	/* SW_PREFETCH_ACCESS.* */
300 	INTEL_EVENT_CONSTRAINT_RANGE(0x48, 0x56, 0xf),
301 	INTEL_EVENT_CONSTRAINT_RANGE(0x60, 0x8b, 0xf),
302 	INTEL_UEVENT_CONSTRAINT(0x04a3, 0xff),  /* CYCLE_ACTIVITY.STALLS_TOTAL */
303 	INTEL_UEVENT_CONSTRAINT(0x10a3, 0xff),  /* CYCLE_ACTIVITY.CYCLES_MEM_ANY */
304 	INTEL_UEVENT_CONSTRAINT(0x14a3, 0xff),  /* CYCLE_ACTIVITY.STALLS_MEM_ANY */
305 	INTEL_EVENT_CONSTRAINT(0xa3, 0xf),      /* CYCLE_ACTIVITY.* */
306 	INTEL_EVENT_CONSTRAINT_RANGE(0xa8, 0xb0, 0xf),
307 	INTEL_EVENT_CONSTRAINT_RANGE(0xb7, 0xbd, 0xf),
308 	INTEL_EVENT_CONSTRAINT_RANGE(0xd0, 0xe6, 0xf),
309 	INTEL_EVENT_CONSTRAINT(0xef, 0xf),
310 	INTEL_EVENT_CONSTRAINT_RANGE(0xf0, 0xf4, 0xf),
311 	EVENT_CONSTRAINT_END
312 };
313 
314 static struct extra_reg intel_icl_extra_regs[] __read_mostly = {
315 	INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffffbfffull, RSP_0),
316 	INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffffbfffull, RSP_1),
317 	INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
318 	INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE),
319 	EVENT_EXTRA_END
320 };
321 
322 static struct extra_reg intel_glc_extra_regs[] __read_mostly = {
323 	INTEL_UEVENT_EXTRA_REG(0x012a, MSR_OFFCORE_RSP_0, 0x3fffffffffull, RSP_0),
324 	INTEL_UEVENT_EXTRA_REG(0x012b, MSR_OFFCORE_RSP_1, 0x3fffffffffull, RSP_1),
325 	INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
326 	INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff1f, FE),
327 	INTEL_UEVENT_EXTRA_REG(0x40ad, MSR_PEBS_FRONTEND, 0x7, FE),
328 	INTEL_UEVENT_EXTRA_REG(0x04c2, MSR_PEBS_FRONTEND, 0x8, FE),
329 	EVENT_EXTRA_END
330 };
331 
332 static struct event_constraint intel_glc_event_constraints[] = {
333 	FIXED_EVENT_CONSTRAINT(0x00c0, 0),	/* INST_RETIRED.ANY */
334 	FIXED_EVENT_CONSTRAINT(0x0100, 0),	/* INST_RETIRED.PREC_DIST */
335 	FIXED_EVENT_CONSTRAINT(0x003c, 1),	/* CPU_CLK_UNHALTED.CORE */
336 	FIXED_EVENT_CONSTRAINT(0x0300, 2),	/* CPU_CLK_UNHALTED.REF */
337 	FIXED_EVENT_CONSTRAINT(0x013c, 2),	/* CPU_CLK_UNHALTED.REF_TSC_P */
338 	FIXED_EVENT_CONSTRAINT(0x0400, 3),	/* SLOTS */
339 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_RETIRING, 0),
340 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BAD_SPEC, 1),
341 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FE_BOUND, 2),
342 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BE_BOUND, 3),
343 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_HEAVY_OPS, 4),
344 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BR_MISPREDICT, 5),
345 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FETCH_LAT, 6),
346 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_MEM_BOUND, 7),
347 
348 	INTEL_EVENT_CONSTRAINT(0x2e, 0xff),
349 	INTEL_EVENT_CONSTRAINT(0x3c, 0xff),
350 	/*
351 	 * Generally event codes < 0x90 are restricted to counters 0-3.
352 	 * The 0x2E and 0x3C are exception, which has no restriction.
353 	 */
354 	INTEL_EVENT_CONSTRAINT_RANGE(0x01, 0x8f, 0xf),
355 
356 	INTEL_UEVENT_CONSTRAINT(0x01a3, 0xf),
357 	INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf),
358 	INTEL_UEVENT_CONSTRAINT(0x08a3, 0xf),
359 	INTEL_UEVENT_CONSTRAINT(0x04a4, 0x1),
360 	INTEL_UEVENT_CONSTRAINT(0x08a4, 0x1),
361 	INTEL_UEVENT_CONSTRAINT(0x02cd, 0x1),
362 	INTEL_EVENT_CONSTRAINT(0xce, 0x1),
363 	INTEL_EVENT_CONSTRAINT_RANGE(0xd0, 0xdf, 0xf),
364 	/*
365 	 * Generally event codes >= 0x90 are likely to have no restrictions.
366 	 * The exception are defined as above.
367 	 */
368 	INTEL_EVENT_CONSTRAINT_RANGE(0x90, 0xfe, 0xff),
369 
370 	EVENT_CONSTRAINT_END
371 };
372 
373 static struct extra_reg intel_rwc_extra_regs[] __read_mostly = {
374 	INTEL_UEVENT_EXTRA_REG(0x012a, MSR_OFFCORE_RSP_0, 0x3fffffffffull, RSP_0),
375 	INTEL_UEVENT_EXTRA_REG(0x012b, MSR_OFFCORE_RSP_1, 0x3fffffffffull, RSP_1),
376 	INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
377 	INTEL_UEVENT_EXTRA_REG(0x02c6, MSR_PEBS_FRONTEND, 0x9, FE),
378 	INTEL_UEVENT_EXTRA_REG(0x03c6, MSR_PEBS_FRONTEND, 0x7fff1f, FE),
379 	INTEL_UEVENT_EXTRA_REG(0x40ad, MSR_PEBS_FRONTEND, 0x7, FE),
380 	INTEL_UEVENT_EXTRA_REG(0x04c2, MSR_PEBS_FRONTEND, 0x8, FE),
381 	EVENT_EXTRA_END
382 };
383 
384 static struct event_constraint intel_lnc_event_constraints[] = {
385 	FIXED_EVENT_CONSTRAINT(0x00c0, 0),	/* INST_RETIRED.ANY */
386 	FIXED_EVENT_CONSTRAINT(0x0100, 0),	/* INST_RETIRED.PREC_DIST */
387 	FIXED_EVENT_CONSTRAINT(0x003c, 1),	/* CPU_CLK_UNHALTED.CORE */
388 	FIXED_EVENT_CONSTRAINT(0x0300, 2),	/* CPU_CLK_UNHALTED.REF */
389 	FIXED_EVENT_CONSTRAINT(0x013c, 2),	/* CPU_CLK_UNHALTED.REF_TSC_P */
390 	FIXED_EVENT_CONSTRAINT(0x0400, 3),	/* SLOTS */
391 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_RETIRING, 0),
392 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BAD_SPEC, 1),
393 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FE_BOUND, 2),
394 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BE_BOUND, 3),
395 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_HEAVY_OPS, 4),
396 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BR_MISPREDICT, 5),
397 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FETCH_LAT, 6),
398 	METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_MEM_BOUND, 7),
399 
400 	INTEL_EVENT_CONSTRAINT(0x20, 0xf),
401 
402 	INTEL_UEVENT_CONSTRAINT(0x012a, 0xf),
403 	INTEL_UEVENT_CONSTRAINT(0x012b, 0xf),
404 	INTEL_UEVENT_CONSTRAINT(0x0148, 0x4),
405 	INTEL_UEVENT_CONSTRAINT(0x0175, 0x4),
406 
407 	INTEL_EVENT_CONSTRAINT(0x2e, 0x3ff),
408 	INTEL_EVENT_CONSTRAINT(0x3c, 0x3ff),
409 
410 	INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4),
411 	INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4),
412 	INTEL_UEVENT_CONSTRAINT(0x04a4, 0x1),
413 	INTEL_UEVENT_CONSTRAINT(0x08a4, 0x1),
414 	INTEL_UEVENT_CONSTRAINT(0x10a4, 0x1),
415 	INTEL_UEVENT_CONSTRAINT(0x01b1, 0x8),
416 	INTEL_UEVENT_CONSTRAINT(0x01cd, 0x3fc),
417 	INTEL_UEVENT_CONSTRAINT(0x02cd, 0x3),
418 
419 	INTEL_EVENT_CONSTRAINT_RANGE(0xd0, 0xdf, 0xf),
420 
421 	INTEL_UEVENT_CONSTRAINT(0x00e0, 0xf),
422 
423 	EVENT_CONSTRAINT_END
424 };
425 
426 static struct extra_reg intel_lnc_extra_regs[] __read_mostly = {
427 	INTEL_UEVENT_EXTRA_REG(0x012a, MSR_OFFCORE_RSP_0, 0xfffffffffffull, RSP_0),
428 	INTEL_UEVENT_EXTRA_REG(0x012b, MSR_OFFCORE_RSP_1, 0xfffffffffffull, RSP_1),
429 	INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd),
430 	INTEL_UEVENT_EXTRA_REG(0x02c6, MSR_PEBS_FRONTEND, 0x9, FE),
431 	INTEL_UEVENT_EXTRA_REG(0x03c6, MSR_PEBS_FRONTEND, 0x7fff1f, FE),
432 	INTEL_UEVENT_EXTRA_REG(0x40ad, MSR_PEBS_FRONTEND, 0xf, FE),
433 	INTEL_UEVENT_EXTRA_REG(0x04c2, MSR_PEBS_FRONTEND, 0x8, FE),
434 	EVENT_EXTRA_END
435 };
436 
437 EVENT_ATTR_STR(mem-loads,	mem_ld_nhm,	"event=0x0b,umask=0x10,ldlat=3");
438 EVENT_ATTR_STR(mem-loads,	mem_ld_snb,	"event=0xcd,umask=0x1,ldlat=3");
439 EVENT_ATTR_STR(mem-stores,	mem_st_snb,	"event=0xcd,umask=0x2");
440 
441 static struct attribute *nhm_mem_events_attrs[] = {
442 	EVENT_PTR(mem_ld_nhm),
443 	NULL,
444 };
445 
446 /*
447  * topdown events for Intel Core CPUs.
448  *
449  * The events are all in slots, which is a free slot in a 4 wide
450  * pipeline. Some events are already reported in slots, for cycle
451  * events we multiply by the pipeline width (4).
452  *
453  * With Hyper Threading on, topdown metrics are either summed or averaged
454  * between the threads of a core: (count_t0 + count_t1).
455  *
456  * For the average case the metric is always scaled to pipeline width,
457  * so we use factor 2 ((count_t0 + count_t1) / 2 * 4)
458  */
459 
460 EVENT_ATTR_STR_HT(topdown-total-slots, td_total_slots,
461 	"event=0x3c,umask=0x0",			/* cpu_clk_unhalted.thread */
462 	"event=0x3c,umask=0x0,any=1");		/* cpu_clk_unhalted.thread_any */
463 EVENT_ATTR_STR_HT(topdown-total-slots.scale, td_total_slots_scale, "4", "2");
464 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued,
465 	"event=0xe,umask=0x1");			/* uops_issued.any */
466 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired,
467 	"event=0xc2,umask=0x2");		/* uops_retired.retire_slots */
468 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles,
469 	"event=0x9c,umask=0x1");		/* idq_uops_not_delivered_core */
470 EVENT_ATTR_STR_HT(topdown-recovery-bubbles, td_recovery_bubbles,
471 	"event=0xd,umask=0x3,cmask=1",		/* int_misc.recovery_cycles */
472 	"event=0xd,umask=0x3,cmask=1,any=1");	/* int_misc.recovery_cycles_any */
473 EVENT_ATTR_STR_HT(topdown-recovery-bubbles.scale, td_recovery_bubbles_scale,
474 	"4", "2");
475 
476 EVENT_ATTR_STR(slots,			slots,			"event=0x00,umask=0x4");
477 EVENT_ATTR_STR(topdown-retiring,	td_retiring,		"event=0x00,umask=0x80");
478 EVENT_ATTR_STR(topdown-bad-spec,	td_bad_spec,		"event=0x00,umask=0x81");
479 EVENT_ATTR_STR(topdown-fe-bound,	td_fe_bound,		"event=0x00,umask=0x82");
480 EVENT_ATTR_STR(topdown-be-bound,	td_be_bound,		"event=0x00,umask=0x83");
481 EVENT_ATTR_STR(topdown-heavy-ops,	td_heavy_ops,		"event=0x00,umask=0x84");
482 EVENT_ATTR_STR(topdown-br-mispredict,	td_br_mispredict,	"event=0x00,umask=0x85");
483 EVENT_ATTR_STR(topdown-fetch-lat,	td_fetch_lat,		"event=0x00,umask=0x86");
484 EVENT_ATTR_STR(topdown-mem-bound,	td_mem_bound,		"event=0x00,umask=0x87");
485 
486 static struct attribute *snb_events_attrs[] = {
487 	EVENT_PTR(td_slots_issued),
488 	EVENT_PTR(td_slots_retired),
489 	EVENT_PTR(td_fetch_bubbles),
490 	EVENT_PTR(td_total_slots),
491 	EVENT_PTR(td_total_slots_scale),
492 	EVENT_PTR(td_recovery_bubbles),
493 	EVENT_PTR(td_recovery_bubbles_scale),
494 	NULL,
495 };
496 
497 static struct attribute *snb_mem_events_attrs[] = {
498 	EVENT_PTR(mem_ld_snb),
499 	EVENT_PTR(mem_st_snb),
500 	NULL,
501 };
502 
503 static struct event_constraint intel_hsw_event_constraints[] = {
504 	FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
505 	FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
506 	FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */
507 	INTEL_UEVENT_CONSTRAINT(0x148, 0x4),	/* L1D_PEND_MISS.PENDING */
508 	INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
509 	INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
510 	/* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */
511 	INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4),
512 	/* CYCLE_ACTIVITY.STALLS_L1D_PENDING */
513 	INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4),
514 	/* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */
515 	INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf),
516 
517 	/*
518 	 * When HT is off these events can only run on the bottom 4 counters
519 	 * When HT is on, they are impacted by the HT bug and require EXCL access
520 	 */
521 	INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */
522 	INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
523 	INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
524 	INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
525 
526 	EVENT_CONSTRAINT_END
527 };
528 
529 static struct event_constraint intel_bdw_event_constraints[] = {
530 	FIXED_EVENT_CONSTRAINT(0x00c0, 0),	/* INST_RETIRED.ANY */
531 	FIXED_EVENT_CONSTRAINT(0x003c, 1),	/* CPU_CLK_UNHALTED.CORE */
532 	FIXED_EVENT_CONSTRAINT(0x0300, 2),	/* CPU_CLK_UNHALTED.REF */
533 	INTEL_UEVENT_CONSTRAINT(0x148, 0x4),	/* L1D_PEND_MISS.PENDING */
534 	INTEL_UBIT_EVENT_CONSTRAINT(0x8a3, 0x4),	/* CYCLE_ACTIVITY.CYCLES_L1D_MISS */
535 	/*
536 	 * when HT is off, these can only run on the bottom 4 counters
537 	 */
538 	INTEL_EVENT_CONSTRAINT(0xd0, 0xf),	/* MEM_INST_RETIRED.* */
539 	INTEL_EVENT_CONSTRAINT(0xd1, 0xf),	/* MEM_LOAD_RETIRED.* */
540 	INTEL_EVENT_CONSTRAINT(0xd2, 0xf),	/* MEM_LOAD_L3_HIT_RETIRED.* */
541 	INTEL_EVENT_CONSTRAINT(0xcd, 0xf),	/* MEM_TRANS_RETIRED.* */
542 	EVENT_CONSTRAINT_END
543 };
544 
intel_pmu_event_map(int hw_event)545 static u64 intel_pmu_event_map(int hw_event)
546 {
547 	return intel_perfmon_event_map[hw_event];
548 }
549 
550 static __initconst const u64 glc_hw_cache_event_ids
551 				[PERF_COUNT_HW_CACHE_MAX]
552 				[PERF_COUNT_HW_CACHE_OP_MAX]
553 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
554 {
555  [ C(L1D ) ] = {
556 	[ C(OP_READ) ] = {
557 		[ C(RESULT_ACCESS) ] = 0x81d0,
558 		[ C(RESULT_MISS)   ] = 0xe124,
559 	},
560 	[ C(OP_WRITE) ] = {
561 		[ C(RESULT_ACCESS) ] = 0x82d0,
562 	},
563  },
564  [ C(L1I ) ] = {
565 	[ C(OP_READ) ] = {
566 		[ C(RESULT_MISS)   ] = 0xe424,
567 	},
568 	[ C(OP_WRITE) ] = {
569 		[ C(RESULT_ACCESS) ] = -1,
570 		[ C(RESULT_MISS)   ] = -1,
571 	},
572  },
573  [ C(LL  ) ] = {
574 	[ C(OP_READ) ] = {
575 		[ C(RESULT_ACCESS) ] = 0x12a,
576 		[ C(RESULT_MISS)   ] = 0x12a,
577 	},
578 	[ C(OP_WRITE) ] = {
579 		[ C(RESULT_ACCESS) ] = 0x12a,
580 		[ C(RESULT_MISS)   ] = 0x12a,
581 	},
582  },
583  [ C(DTLB) ] = {
584 	[ C(OP_READ) ] = {
585 		[ C(RESULT_ACCESS) ] = 0x81d0,
586 		[ C(RESULT_MISS)   ] = 0xe12,
587 	},
588 	[ C(OP_WRITE) ] = {
589 		[ C(RESULT_ACCESS) ] = 0x82d0,
590 		[ C(RESULT_MISS)   ] = 0xe13,
591 	},
592  },
593  [ C(ITLB) ] = {
594 	[ C(OP_READ) ] = {
595 		[ C(RESULT_ACCESS) ] = -1,
596 		[ C(RESULT_MISS)   ] = 0xe11,
597 	},
598 	[ C(OP_WRITE) ] = {
599 		[ C(RESULT_ACCESS) ] = -1,
600 		[ C(RESULT_MISS)   ] = -1,
601 	},
602 	[ C(OP_PREFETCH) ] = {
603 		[ C(RESULT_ACCESS) ] = -1,
604 		[ C(RESULT_MISS)   ] = -1,
605 	},
606  },
607  [ C(BPU ) ] = {
608 	[ C(OP_READ) ] = {
609 		[ C(RESULT_ACCESS) ] = 0x4c4,
610 		[ C(RESULT_MISS)   ] = 0x4c5,
611 	},
612 	[ C(OP_WRITE) ] = {
613 		[ C(RESULT_ACCESS) ] = -1,
614 		[ C(RESULT_MISS)   ] = -1,
615 	},
616 	[ C(OP_PREFETCH) ] = {
617 		[ C(RESULT_ACCESS) ] = -1,
618 		[ C(RESULT_MISS)   ] = -1,
619 	},
620  },
621  [ C(NODE) ] = {
622 	[ C(OP_READ) ] = {
623 		[ C(RESULT_ACCESS) ] = 0x12a,
624 		[ C(RESULT_MISS)   ] = 0x12a,
625 	},
626  },
627 };
628 
629 static __initconst const u64 glc_hw_cache_extra_regs
630 				[PERF_COUNT_HW_CACHE_MAX]
631 				[PERF_COUNT_HW_CACHE_OP_MAX]
632 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
633 {
634  [ C(LL  ) ] = {
635 	[ C(OP_READ) ] = {
636 		[ C(RESULT_ACCESS) ] = 0x10001,
637 		[ C(RESULT_MISS)   ] = 0x3fbfc00001,
638 	},
639 	[ C(OP_WRITE) ] = {
640 		[ C(RESULT_ACCESS) ] = 0x3f3ffc0002,
641 		[ C(RESULT_MISS)   ] = 0x3f3fc00002,
642 	},
643  },
644  [ C(NODE) ] = {
645 	[ C(OP_READ) ] = {
646 		[ C(RESULT_ACCESS) ] = 0x10c000001,
647 		[ C(RESULT_MISS)   ] = 0x3fb3000001,
648 	},
649  },
650 };
651 
652 /*
653  * Notes on the events:
654  * - data reads do not include code reads (comparable to earlier tables)
655  * - data counts include speculative execution (except L1 write, dtlb, bpu)
656  * - remote node access includes remote memory, remote cache, remote mmio.
657  * - prefetches are not included in the counts.
658  * - icache miss does not include decoded icache
659  */
660 
661 #define SKL_DEMAND_DATA_RD		BIT_ULL(0)
662 #define SKL_DEMAND_RFO			BIT_ULL(1)
663 #define SKL_ANY_RESPONSE		BIT_ULL(16)
664 #define SKL_SUPPLIER_NONE		BIT_ULL(17)
665 #define SKL_L3_MISS_LOCAL_DRAM		BIT_ULL(26)
666 #define SKL_L3_MISS_REMOTE_HOP0_DRAM	BIT_ULL(27)
667 #define SKL_L3_MISS_REMOTE_HOP1_DRAM	BIT_ULL(28)
668 #define SKL_L3_MISS_REMOTE_HOP2P_DRAM	BIT_ULL(29)
669 #define SKL_L3_MISS			(SKL_L3_MISS_LOCAL_DRAM| \
670 					 SKL_L3_MISS_REMOTE_HOP0_DRAM| \
671 					 SKL_L3_MISS_REMOTE_HOP1_DRAM| \
672 					 SKL_L3_MISS_REMOTE_HOP2P_DRAM)
673 #define SKL_SPL_HIT			BIT_ULL(30)
674 #define SKL_SNOOP_NONE			BIT_ULL(31)
675 #define SKL_SNOOP_NOT_NEEDED		BIT_ULL(32)
676 #define SKL_SNOOP_MISS			BIT_ULL(33)
677 #define SKL_SNOOP_HIT_NO_FWD		BIT_ULL(34)
678 #define SKL_SNOOP_HIT_WITH_FWD		BIT_ULL(35)
679 #define SKL_SNOOP_HITM			BIT_ULL(36)
680 #define SKL_SNOOP_NON_DRAM		BIT_ULL(37)
681 #define SKL_ANY_SNOOP			(SKL_SPL_HIT|SKL_SNOOP_NONE| \
682 					 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \
683 					 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \
684 					 SKL_SNOOP_HITM|SKL_SNOOP_NON_DRAM)
685 #define SKL_DEMAND_READ			SKL_DEMAND_DATA_RD
686 #define SKL_SNOOP_DRAM			(SKL_SNOOP_NONE| \
687 					 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \
688 					 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \
689 					 SKL_SNOOP_HITM|SKL_SPL_HIT)
690 #define SKL_DEMAND_WRITE		SKL_DEMAND_RFO
691 #define SKL_LLC_ACCESS			SKL_ANY_RESPONSE
692 #define SKL_L3_MISS_REMOTE		(SKL_L3_MISS_REMOTE_HOP0_DRAM| \
693 					 SKL_L3_MISS_REMOTE_HOP1_DRAM| \
694 					 SKL_L3_MISS_REMOTE_HOP2P_DRAM)
695 
696 static __initconst const u64 skl_hw_cache_event_ids
697 				[PERF_COUNT_HW_CACHE_MAX]
698 				[PERF_COUNT_HW_CACHE_OP_MAX]
699 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
700 {
701  [ C(L1D ) ] = {
702 	[ C(OP_READ) ] = {
703 		[ C(RESULT_ACCESS) ] = 0x81d0,	/* MEM_INST_RETIRED.ALL_LOADS */
704 		[ C(RESULT_MISS)   ] = 0x151,	/* L1D.REPLACEMENT */
705 	},
706 	[ C(OP_WRITE) ] = {
707 		[ C(RESULT_ACCESS) ] = 0x82d0,	/* MEM_INST_RETIRED.ALL_STORES */
708 		[ C(RESULT_MISS)   ] = 0x0,
709 	},
710 	[ C(OP_PREFETCH) ] = {
711 		[ C(RESULT_ACCESS) ] = 0x0,
712 		[ C(RESULT_MISS)   ] = 0x0,
713 	},
714  },
715  [ C(L1I ) ] = {
716 	[ C(OP_READ) ] = {
717 		[ C(RESULT_ACCESS) ] = 0x0,
718 		[ C(RESULT_MISS)   ] = 0x283,	/* ICACHE_64B.MISS */
719 	},
720 	[ C(OP_WRITE) ] = {
721 		[ C(RESULT_ACCESS) ] = -1,
722 		[ C(RESULT_MISS)   ] = -1,
723 	},
724 	[ C(OP_PREFETCH) ] = {
725 		[ C(RESULT_ACCESS) ] = 0x0,
726 		[ C(RESULT_MISS)   ] = 0x0,
727 	},
728  },
729  [ C(LL  ) ] = {
730 	[ C(OP_READ) ] = {
731 		[ C(RESULT_ACCESS) ] = 0x1b7,	/* OFFCORE_RESPONSE */
732 		[ C(RESULT_MISS)   ] = 0x1b7,	/* OFFCORE_RESPONSE */
733 	},
734 	[ C(OP_WRITE) ] = {
735 		[ C(RESULT_ACCESS) ] = 0x1b7,	/* OFFCORE_RESPONSE */
736 		[ C(RESULT_MISS)   ] = 0x1b7,	/* OFFCORE_RESPONSE */
737 	},
738 	[ C(OP_PREFETCH) ] = {
739 		[ C(RESULT_ACCESS) ] = 0x0,
740 		[ C(RESULT_MISS)   ] = 0x0,
741 	},
742  },
743  [ C(DTLB) ] = {
744 	[ C(OP_READ) ] = {
745 		[ C(RESULT_ACCESS) ] = 0x81d0,	/* MEM_INST_RETIRED.ALL_LOADS */
746 		[ C(RESULT_MISS)   ] = 0xe08,	/* DTLB_LOAD_MISSES.WALK_COMPLETED */
747 	},
748 	[ C(OP_WRITE) ] = {
749 		[ C(RESULT_ACCESS) ] = 0x82d0,	/* MEM_INST_RETIRED.ALL_STORES */
750 		[ C(RESULT_MISS)   ] = 0xe49,	/* DTLB_STORE_MISSES.WALK_COMPLETED */
751 	},
752 	[ C(OP_PREFETCH) ] = {
753 		[ C(RESULT_ACCESS) ] = 0x0,
754 		[ C(RESULT_MISS)   ] = 0x0,
755 	},
756  },
757  [ C(ITLB) ] = {
758 	[ C(OP_READ) ] = {
759 		[ C(RESULT_ACCESS) ] = 0x2085,	/* ITLB_MISSES.STLB_HIT */
760 		[ C(RESULT_MISS)   ] = 0xe85,	/* ITLB_MISSES.WALK_COMPLETED */
761 	},
762 	[ C(OP_WRITE) ] = {
763 		[ C(RESULT_ACCESS) ] = -1,
764 		[ C(RESULT_MISS)   ] = -1,
765 	},
766 	[ C(OP_PREFETCH) ] = {
767 		[ C(RESULT_ACCESS) ] = -1,
768 		[ C(RESULT_MISS)   ] = -1,
769 	},
770  },
771  [ C(BPU ) ] = {
772 	[ C(OP_READ) ] = {
773 		[ C(RESULT_ACCESS) ] = 0xc4,	/* BR_INST_RETIRED.ALL_BRANCHES */
774 		[ C(RESULT_MISS)   ] = 0xc5,	/* BR_MISP_RETIRED.ALL_BRANCHES */
775 	},
776 	[ C(OP_WRITE) ] = {
777 		[ C(RESULT_ACCESS) ] = -1,
778 		[ C(RESULT_MISS)   ] = -1,
779 	},
780 	[ C(OP_PREFETCH) ] = {
781 		[ C(RESULT_ACCESS) ] = -1,
782 		[ C(RESULT_MISS)   ] = -1,
783 	},
784  },
785  [ C(NODE) ] = {
786 	[ C(OP_READ) ] = {
787 		[ C(RESULT_ACCESS) ] = 0x1b7,	/* OFFCORE_RESPONSE */
788 		[ C(RESULT_MISS)   ] = 0x1b7,	/* OFFCORE_RESPONSE */
789 	},
790 	[ C(OP_WRITE) ] = {
791 		[ C(RESULT_ACCESS) ] = 0x1b7,	/* OFFCORE_RESPONSE */
792 		[ C(RESULT_MISS)   ] = 0x1b7,	/* OFFCORE_RESPONSE */
793 	},
794 	[ C(OP_PREFETCH) ] = {
795 		[ C(RESULT_ACCESS) ] = 0x0,
796 		[ C(RESULT_MISS)   ] = 0x0,
797 	},
798  },
799 };
800 
801 static __initconst const u64 skl_hw_cache_extra_regs
802 				[PERF_COUNT_HW_CACHE_MAX]
803 				[PERF_COUNT_HW_CACHE_OP_MAX]
804 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
805 {
806  [ C(LL  ) ] = {
807 	[ C(OP_READ) ] = {
808 		[ C(RESULT_ACCESS) ] = SKL_DEMAND_READ|
809 				       SKL_LLC_ACCESS|SKL_ANY_SNOOP,
810 		[ C(RESULT_MISS)   ] = SKL_DEMAND_READ|
811 				       SKL_L3_MISS|SKL_ANY_SNOOP|
812 				       SKL_SUPPLIER_NONE,
813 	},
814 	[ C(OP_WRITE) ] = {
815 		[ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE|
816 				       SKL_LLC_ACCESS|SKL_ANY_SNOOP,
817 		[ C(RESULT_MISS)   ] = SKL_DEMAND_WRITE|
818 				       SKL_L3_MISS|SKL_ANY_SNOOP|
819 				       SKL_SUPPLIER_NONE,
820 	},
821 	[ C(OP_PREFETCH) ] = {
822 		[ C(RESULT_ACCESS) ] = 0x0,
823 		[ C(RESULT_MISS)   ] = 0x0,
824 	},
825  },
826  [ C(NODE) ] = {
827 	[ C(OP_READ) ] = {
828 		[ C(RESULT_ACCESS) ] = SKL_DEMAND_READ|
829 				       SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM,
830 		[ C(RESULT_MISS)   ] = SKL_DEMAND_READ|
831 				       SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM,
832 	},
833 	[ C(OP_WRITE) ] = {
834 		[ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE|
835 				       SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM,
836 		[ C(RESULT_MISS)   ] = SKL_DEMAND_WRITE|
837 				       SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM,
838 	},
839 	[ C(OP_PREFETCH) ] = {
840 		[ C(RESULT_ACCESS) ] = 0x0,
841 		[ C(RESULT_MISS)   ] = 0x0,
842 	},
843  },
844 };
845 
846 #define SNB_DMND_DATA_RD	(1ULL << 0)
847 #define SNB_DMND_RFO		(1ULL << 1)
848 #define SNB_DMND_IFETCH		(1ULL << 2)
849 #define SNB_DMND_WB		(1ULL << 3)
850 #define SNB_PF_DATA_RD		(1ULL << 4)
851 #define SNB_PF_RFO		(1ULL << 5)
852 #define SNB_PF_IFETCH		(1ULL << 6)
853 #define SNB_LLC_DATA_RD		(1ULL << 7)
854 #define SNB_LLC_RFO		(1ULL << 8)
855 #define SNB_LLC_IFETCH		(1ULL << 9)
856 #define SNB_BUS_LOCKS		(1ULL << 10)
857 #define SNB_STRM_ST		(1ULL << 11)
858 #define SNB_OTHER		(1ULL << 15)
859 #define SNB_RESP_ANY		(1ULL << 16)
860 #define SNB_NO_SUPP		(1ULL << 17)
861 #define SNB_LLC_HITM		(1ULL << 18)
862 #define SNB_LLC_HITE		(1ULL << 19)
863 #define SNB_LLC_HITS		(1ULL << 20)
864 #define SNB_LLC_HITF		(1ULL << 21)
865 #define SNB_LOCAL		(1ULL << 22)
866 #define SNB_REMOTE		(0xffULL << 23)
867 #define SNB_SNP_NONE		(1ULL << 31)
868 #define SNB_SNP_NOT_NEEDED	(1ULL << 32)
869 #define SNB_SNP_MISS		(1ULL << 33)
870 #define SNB_NO_FWD		(1ULL << 34)
871 #define SNB_SNP_FWD		(1ULL << 35)
872 #define SNB_HITM		(1ULL << 36)
873 #define SNB_NON_DRAM		(1ULL << 37)
874 
875 #define SNB_DMND_READ		(SNB_DMND_DATA_RD|SNB_LLC_DATA_RD)
876 #define SNB_DMND_WRITE		(SNB_DMND_RFO|SNB_LLC_RFO)
877 #define SNB_DMND_PREFETCH	(SNB_PF_DATA_RD|SNB_PF_RFO)
878 
879 #define SNB_SNP_ANY		(SNB_SNP_NONE|SNB_SNP_NOT_NEEDED| \
880 				 SNB_SNP_MISS|SNB_NO_FWD|SNB_SNP_FWD| \
881 				 SNB_HITM)
882 
883 #define SNB_DRAM_ANY		(SNB_LOCAL|SNB_REMOTE|SNB_SNP_ANY)
884 #define SNB_DRAM_REMOTE		(SNB_REMOTE|SNB_SNP_ANY)
885 
886 #define SNB_L3_ACCESS		SNB_RESP_ANY
887 #define SNB_L3_MISS		(SNB_DRAM_ANY|SNB_NON_DRAM)
888 
889 static __initconst const u64 snb_hw_cache_extra_regs
890 				[PERF_COUNT_HW_CACHE_MAX]
891 				[PERF_COUNT_HW_CACHE_OP_MAX]
892 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
893 {
894  [ C(LL  ) ] = {
895 	[ C(OP_READ) ] = {
896 		[ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_L3_ACCESS,
897 		[ C(RESULT_MISS)   ] = SNB_DMND_READ|SNB_L3_MISS,
898 	},
899 	[ C(OP_WRITE) ] = {
900 		[ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_L3_ACCESS,
901 		[ C(RESULT_MISS)   ] = SNB_DMND_WRITE|SNB_L3_MISS,
902 	},
903 	[ C(OP_PREFETCH) ] = {
904 		[ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_L3_ACCESS,
905 		[ C(RESULT_MISS)   ] = SNB_DMND_PREFETCH|SNB_L3_MISS,
906 	},
907  },
908  [ C(NODE) ] = {
909 	[ C(OP_READ) ] = {
910 		[ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_DRAM_ANY,
911 		[ C(RESULT_MISS)   ] = SNB_DMND_READ|SNB_DRAM_REMOTE,
912 	},
913 	[ C(OP_WRITE) ] = {
914 		[ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_DRAM_ANY,
915 		[ C(RESULT_MISS)   ] = SNB_DMND_WRITE|SNB_DRAM_REMOTE,
916 	},
917 	[ C(OP_PREFETCH) ] = {
918 		[ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_DRAM_ANY,
919 		[ C(RESULT_MISS)   ] = SNB_DMND_PREFETCH|SNB_DRAM_REMOTE,
920 	},
921  },
922 };
923 
924 static __initconst const u64 snb_hw_cache_event_ids
925 				[PERF_COUNT_HW_CACHE_MAX]
926 				[PERF_COUNT_HW_CACHE_OP_MAX]
927 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
928 {
929  [ C(L1D) ] = {
930 	[ C(OP_READ) ] = {
931 		[ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS        */
932 		[ C(RESULT_MISS)   ] = 0x0151, /* L1D.REPLACEMENT              */
933 	},
934 	[ C(OP_WRITE) ] = {
935 		[ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES       */
936 		[ C(RESULT_MISS)   ] = 0x0851, /* L1D.ALL_M_REPLACEMENT        */
937 	},
938 	[ C(OP_PREFETCH) ] = {
939 		[ C(RESULT_ACCESS) ] = 0x0,
940 		[ C(RESULT_MISS)   ] = 0x024e, /* HW_PRE_REQ.DL1_MISS          */
941 	},
942  },
943  [ C(L1I ) ] = {
944 	[ C(OP_READ) ] = {
945 		[ C(RESULT_ACCESS) ] = 0x0,
946 		[ C(RESULT_MISS)   ] = 0x0280, /* ICACHE.MISSES */
947 	},
948 	[ C(OP_WRITE) ] = {
949 		[ C(RESULT_ACCESS) ] = -1,
950 		[ C(RESULT_MISS)   ] = -1,
951 	},
952 	[ C(OP_PREFETCH) ] = {
953 		[ C(RESULT_ACCESS) ] = 0x0,
954 		[ C(RESULT_MISS)   ] = 0x0,
955 	},
956  },
957  [ C(LL  ) ] = {
958 	[ C(OP_READ) ] = {
959 		/* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
960 		[ C(RESULT_ACCESS) ] = 0x01b7,
961 		/* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
962 		[ C(RESULT_MISS)   ] = 0x01b7,
963 	},
964 	[ C(OP_WRITE) ] = {
965 		/* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
966 		[ C(RESULT_ACCESS) ] = 0x01b7,
967 		/* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
968 		[ C(RESULT_MISS)   ] = 0x01b7,
969 	},
970 	[ C(OP_PREFETCH) ] = {
971 		/* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
972 		[ C(RESULT_ACCESS) ] = 0x01b7,
973 		/* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
974 		[ C(RESULT_MISS)   ] = 0x01b7,
975 	},
976  },
977  [ C(DTLB) ] = {
978 	[ C(OP_READ) ] = {
979 		[ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */
980 		[ C(RESULT_MISS)   ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */
981 	},
982 	[ C(OP_WRITE) ] = {
983 		[ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */
984 		[ C(RESULT_MISS)   ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
985 	},
986 	[ C(OP_PREFETCH) ] = {
987 		[ C(RESULT_ACCESS) ] = 0x0,
988 		[ C(RESULT_MISS)   ] = 0x0,
989 	},
990  },
991  [ C(ITLB) ] = {
992 	[ C(OP_READ) ] = {
993 		[ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT         */
994 		[ C(RESULT_MISS)   ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK    */
995 	},
996 	[ C(OP_WRITE) ] = {
997 		[ C(RESULT_ACCESS) ] = -1,
998 		[ C(RESULT_MISS)   ] = -1,
999 	},
1000 	[ C(OP_PREFETCH) ] = {
1001 		[ C(RESULT_ACCESS) ] = -1,
1002 		[ C(RESULT_MISS)   ] = -1,
1003 	},
1004  },
1005  [ C(BPU ) ] = {
1006 	[ C(OP_READ) ] = {
1007 		[ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1008 		[ C(RESULT_MISS)   ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
1009 	},
1010 	[ C(OP_WRITE) ] = {
1011 		[ C(RESULT_ACCESS) ] = -1,
1012 		[ C(RESULT_MISS)   ] = -1,
1013 	},
1014 	[ C(OP_PREFETCH) ] = {
1015 		[ C(RESULT_ACCESS) ] = -1,
1016 		[ C(RESULT_MISS)   ] = -1,
1017 	},
1018  },
1019  [ C(NODE) ] = {
1020 	[ C(OP_READ) ] = {
1021 		[ C(RESULT_ACCESS) ] = 0x01b7,
1022 		[ C(RESULT_MISS)   ] = 0x01b7,
1023 	},
1024 	[ C(OP_WRITE) ] = {
1025 		[ C(RESULT_ACCESS) ] = 0x01b7,
1026 		[ C(RESULT_MISS)   ] = 0x01b7,
1027 	},
1028 	[ C(OP_PREFETCH) ] = {
1029 		[ C(RESULT_ACCESS) ] = 0x01b7,
1030 		[ C(RESULT_MISS)   ] = 0x01b7,
1031 	},
1032  },
1033 
1034 };
1035 
1036 /*
1037  * Notes on the events:
1038  * - data reads do not include code reads (comparable to earlier tables)
1039  * - data counts include speculative execution (except L1 write, dtlb, bpu)
1040  * - remote node access includes remote memory, remote cache, remote mmio.
1041  * - prefetches are not included in the counts because they are not
1042  *   reliably counted.
1043  */
1044 
1045 #define HSW_DEMAND_DATA_RD		BIT_ULL(0)
1046 #define HSW_DEMAND_RFO			BIT_ULL(1)
1047 #define HSW_ANY_RESPONSE		BIT_ULL(16)
1048 #define HSW_SUPPLIER_NONE		BIT_ULL(17)
1049 #define HSW_L3_MISS_LOCAL_DRAM		BIT_ULL(22)
1050 #define HSW_L3_MISS_REMOTE_HOP0		BIT_ULL(27)
1051 #define HSW_L3_MISS_REMOTE_HOP1		BIT_ULL(28)
1052 #define HSW_L3_MISS_REMOTE_HOP2P	BIT_ULL(29)
1053 #define HSW_L3_MISS			(HSW_L3_MISS_LOCAL_DRAM| \
1054 					 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \
1055 					 HSW_L3_MISS_REMOTE_HOP2P)
1056 #define HSW_SNOOP_NONE			BIT_ULL(31)
1057 #define HSW_SNOOP_NOT_NEEDED		BIT_ULL(32)
1058 #define HSW_SNOOP_MISS			BIT_ULL(33)
1059 #define HSW_SNOOP_HIT_NO_FWD		BIT_ULL(34)
1060 #define HSW_SNOOP_HIT_WITH_FWD		BIT_ULL(35)
1061 #define HSW_SNOOP_HITM			BIT_ULL(36)
1062 #define HSW_SNOOP_NON_DRAM		BIT_ULL(37)
1063 #define HSW_ANY_SNOOP			(HSW_SNOOP_NONE| \
1064 					 HSW_SNOOP_NOT_NEEDED|HSW_SNOOP_MISS| \
1065 					 HSW_SNOOP_HIT_NO_FWD|HSW_SNOOP_HIT_WITH_FWD| \
1066 					 HSW_SNOOP_HITM|HSW_SNOOP_NON_DRAM)
1067 #define HSW_SNOOP_DRAM			(HSW_ANY_SNOOP & ~HSW_SNOOP_NON_DRAM)
1068 #define HSW_DEMAND_READ			HSW_DEMAND_DATA_RD
1069 #define HSW_DEMAND_WRITE		HSW_DEMAND_RFO
1070 #define HSW_L3_MISS_REMOTE		(HSW_L3_MISS_REMOTE_HOP0|\
1071 					 HSW_L3_MISS_REMOTE_HOP1|HSW_L3_MISS_REMOTE_HOP2P)
1072 #define HSW_LLC_ACCESS			HSW_ANY_RESPONSE
1073 
1074 #define BDW_L3_MISS_LOCAL		BIT(26)
1075 #define BDW_L3_MISS			(BDW_L3_MISS_LOCAL| \
1076 					 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \
1077 					 HSW_L3_MISS_REMOTE_HOP2P)
1078 
1079 
1080 static __initconst const u64 hsw_hw_cache_event_ids
1081 				[PERF_COUNT_HW_CACHE_MAX]
1082 				[PERF_COUNT_HW_CACHE_OP_MAX]
1083 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
1084 {
1085  [ C(L1D ) ] = {
1086 	[ C(OP_READ) ] = {
1087 		[ C(RESULT_ACCESS) ] = 0x81d0,	/* MEM_UOPS_RETIRED.ALL_LOADS */
1088 		[ C(RESULT_MISS)   ] = 0x151,	/* L1D.REPLACEMENT */
1089 	},
1090 	[ C(OP_WRITE) ] = {
1091 		[ C(RESULT_ACCESS) ] = 0x82d0,	/* MEM_UOPS_RETIRED.ALL_STORES */
1092 		[ C(RESULT_MISS)   ] = 0x0,
1093 	},
1094 	[ C(OP_PREFETCH) ] = {
1095 		[ C(RESULT_ACCESS) ] = 0x0,
1096 		[ C(RESULT_MISS)   ] = 0x0,
1097 	},
1098  },
1099  [ C(L1I ) ] = {
1100 	[ C(OP_READ) ] = {
1101 		[ C(RESULT_ACCESS) ] = 0x0,
1102 		[ C(RESULT_MISS)   ] = 0x280,	/* ICACHE.MISSES */
1103 	},
1104 	[ C(OP_WRITE) ] = {
1105 		[ C(RESULT_ACCESS) ] = -1,
1106 		[ C(RESULT_MISS)   ] = -1,
1107 	},
1108 	[ C(OP_PREFETCH) ] = {
1109 		[ C(RESULT_ACCESS) ] = 0x0,
1110 		[ C(RESULT_MISS)   ] = 0x0,
1111 	},
1112  },
1113  [ C(LL  ) ] = {
1114 	[ C(OP_READ) ] = {
1115 		[ C(RESULT_ACCESS) ] = 0x1b7,	/* OFFCORE_RESPONSE */
1116 		[ C(RESULT_MISS)   ] = 0x1b7,	/* OFFCORE_RESPONSE */
1117 	},
1118 	[ C(OP_WRITE) ] = {
1119 		[ C(RESULT_ACCESS) ] = 0x1b7,	/* OFFCORE_RESPONSE */
1120 		[ C(RESULT_MISS)   ] = 0x1b7,	/* OFFCORE_RESPONSE */
1121 	},
1122 	[ C(OP_PREFETCH) ] = {
1123 		[ C(RESULT_ACCESS) ] = 0x0,
1124 		[ C(RESULT_MISS)   ] = 0x0,
1125 	},
1126  },
1127  [ C(DTLB) ] = {
1128 	[ C(OP_READ) ] = {
1129 		[ C(RESULT_ACCESS) ] = 0x81d0,	/* MEM_UOPS_RETIRED.ALL_LOADS */
1130 		[ C(RESULT_MISS)   ] = 0x108,	/* DTLB_LOAD_MISSES.MISS_CAUSES_A_WALK */
1131 	},
1132 	[ C(OP_WRITE) ] = {
1133 		[ C(RESULT_ACCESS) ] = 0x82d0,	/* MEM_UOPS_RETIRED.ALL_STORES */
1134 		[ C(RESULT_MISS)   ] = 0x149,	/* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
1135 	},
1136 	[ C(OP_PREFETCH) ] = {
1137 		[ C(RESULT_ACCESS) ] = 0x0,
1138 		[ C(RESULT_MISS)   ] = 0x0,
1139 	},
1140  },
1141  [ C(ITLB) ] = {
1142 	[ C(OP_READ) ] = {
1143 		[ C(RESULT_ACCESS) ] = 0x6085,	/* ITLB_MISSES.STLB_HIT */
1144 		[ C(RESULT_MISS)   ] = 0x185,	/* ITLB_MISSES.MISS_CAUSES_A_WALK */
1145 	},
1146 	[ C(OP_WRITE) ] = {
1147 		[ C(RESULT_ACCESS) ] = -1,
1148 		[ C(RESULT_MISS)   ] = -1,
1149 	},
1150 	[ C(OP_PREFETCH) ] = {
1151 		[ C(RESULT_ACCESS) ] = -1,
1152 		[ C(RESULT_MISS)   ] = -1,
1153 	},
1154  },
1155  [ C(BPU ) ] = {
1156 	[ C(OP_READ) ] = {
1157 		[ C(RESULT_ACCESS) ] = 0xc4,	/* BR_INST_RETIRED.ALL_BRANCHES */
1158 		[ C(RESULT_MISS)   ] = 0xc5,	/* BR_MISP_RETIRED.ALL_BRANCHES */
1159 	},
1160 	[ C(OP_WRITE) ] = {
1161 		[ C(RESULT_ACCESS) ] = -1,
1162 		[ C(RESULT_MISS)   ] = -1,
1163 	},
1164 	[ C(OP_PREFETCH) ] = {
1165 		[ C(RESULT_ACCESS) ] = -1,
1166 		[ C(RESULT_MISS)   ] = -1,
1167 	},
1168  },
1169  [ C(NODE) ] = {
1170 	[ C(OP_READ) ] = {
1171 		[ C(RESULT_ACCESS) ] = 0x1b7,	/* OFFCORE_RESPONSE */
1172 		[ C(RESULT_MISS)   ] = 0x1b7,	/* OFFCORE_RESPONSE */
1173 	},
1174 	[ C(OP_WRITE) ] = {
1175 		[ C(RESULT_ACCESS) ] = 0x1b7,	/* OFFCORE_RESPONSE */
1176 		[ C(RESULT_MISS)   ] = 0x1b7,	/* OFFCORE_RESPONSE */
1177 	},
1178 	[ C(OP_PREFETCH) ] = {
1179 		[ C(RESULT_ACCESS) ] = 0x0,
1180 		[ C(RESULT_MISS)   ] = 0x0,
1181 	},
1182  },
1183 };
1184 
1185 static __initconst const u64 hsw_hw_cache_extra_regs
1186 				[PERF_COUNT_HW_CACHE_MAX]
1187 				[PERF_COUNT_HW_CACHE_OP_MAX]
1188 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
1189 {
1190  [ C(LL  ) ] = {
1191 	[ C(OP_READ) ] = {
1192 		[ C(RESULT_ACCESS) ] = HSW_DEMAND_READ|
1193 				       HSW_LLC_ACCESS,
1194 		[ C(RESULT_MISS)   ] = HSW_DEMAND_READ|
1195 				       HSW_L3_MISS|HSW_ANY_SNOOP,
1196 	},
1197 	[ C(OP_WRITE) ] = {
1198 		[ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE|
1199 				       HSW_LLC_ACCESS,
1200 		[ C(RESULT_MISS)   ] = HSW_DEMAND_WRITE|
1201 				       HSW_L3_MISS|HSW_ANY_SNOOP,
1202 	},
1203 	[ C(OP_PREFETCH) ] = {
1204 		[ C(RESULT_ACCESS) ] = 0x0,
1205 		[ C(RESULT_MISS)   ] = 0x0,
1206 	},
1207  },
1208  [ C(NODE) ] = {
1209 	[ C(OP_READ) ] = {
1210 		[ C(RESULT_ACCESS) ] = HSW_DEMAND_READ|
1211 				       HSW_L3_MISS_LOCAL_DRAM|
1212 				       HSW_SNOOP_DRAM,
1213 		[ C(RESULT_MISS)   ] = HSW_DEMAND_READ|
1214 				       HSW_L3_MISS_REMOTE|
1215 				       HSW_SNOOP_DRAM,
1216 	},
1217 	[ C(OP_WRITE) ] = {
1218 		[ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE|
1219 				       HSW_L3_MISS_LOCAL_DRAM|
1220 				       HSW_SNOOP_DRAM,
1221 		[ C(RESULT_MISS)   ] = HSW_DEMAND_WRITE|
1222 				       HSW_L3_MISS_REMOTE|
1223 				       HSW_SNOOP_DRAM,
1224 	},
1225 	[ C(OP_PREFETCH) ] = {
1226 		[ C(RESULT_ACCESS) ] = 0x0,
1227 		[ C(RESULT_MISS)   ] = 0x0,
1228 	},
1229  },
1230 };
1231 
1232 static __initconst const u64 westmere_hw_cache_event_ids
1233 				[PERF_COUNT_HW_CACHE_MAX]
1234 				[PERF_COUNT_HW_CACHE_OP_MAX]
1235 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
1236 {
1237  [ C(L1D) ] = {
1238 	[ C(OP_READ) ] = {
1239 		[ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS       */
1240 		[ C(RESULT_MISS)   ] = 0x0151, /* L1D.REPL                     */
1241 	},
1242 	[ C(OP_WRITE) ] = {
1243 		[ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES      */
1244 		[ C(RESULT_MISS)   ] = 0x0251, /* L1D.M_REPL                   */
1245 	},
1246 	[ C(OP_PREFETCH) ] = {
1247 		[ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS        */
1248 		[ C(RESULT_MISS)   ] = 0x024e, /* L1D_PREFETCH.MISS            */
1249 	},
1250  },
1251  [ C(L1I ) ] = {
1252 	[ C(OP_READ) ] = {
1253 		[ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS                    */
1254 		[ C(RESULT_MISS)   ] = 0x0280, /* L1I.MISSES                   */
1255 	},
1256 	[ C(OP_WRITE) ] = {
1257 		[ C(RESULT_ACCESS) ] = -1,
1258 		[ C(RESULT_MISS)   ] = -1,
1259 	},
1260 	[ C(OP_PREFETCH) ] = {
1261 		[ C(RESULT_ACCESS) ] = 0x0,
1262 		[ C(RESULT_MISS)   ] = 0x0,
1263 	},
1264  },
1265  [ C(LL  ) ] = {
1266 	[ C(OP_READ) ] = {
1267 		/* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
1268 		[ C(RESULT_ACCESS) ] = 0x01b7,
1269 		/* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
1270 		[ C(RESULT_MISS)   ] = 0x01b7,
1271 	},
1272 	/*
1273 	 * Use RFO, not WRITEBACK, because a write miss would typically occur
1274 	 * on RFO.
1275 	 */
1276 	[ C(OP_WRITE) ] = {
1277 		/* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
1278 		[ C(RESULT_ACCESS) ] = 0x01b7,
1279 		/* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
1280 		[ C(RESULT_MISS)   ] = 0x01b7,
1281 	},
1282 	[ C(OP_PREFETCH) ] = {
1283 		/* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
1284 		[ C(RESULT_ACCESS) ] = 0x01b7,
1285 		/* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
1286 		[ C(RESULT_MISS)   ] = 0x01b7,
1287 	},
1288  },
1289  [ C(DTLB) ] = {
1290 	[ C(OP_READ) ] = {
1291 		[ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS       */
1292 		[ C(RESULT_MISS)   ] = 0x0108, /* DTLB_LOAD_MISSES.ANY         */
1293 	},
1294 	[ C(OP_WRITE) ] = {
1295 		[ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES      */
1296 		[ C(RESULT_MISS)   ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS  */
1297 	},
1298 	[ C(OP_PREFETCH) ] = {
1299 		[ C(RESULT_ACCESS) ] = 0x0,
1300 		[ C(RESULT_MISS)   ] = 0x0,
1301 	},
1302  },
1303  [ C(ITLB) ] = {
1304 	[ C(OP_READ) ] = {
1305 		[ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P           */
1306 		[ C(RESULT_MISS)   ] = 0x0185, /* ITLB_MISSES.ANY              */
1307 	},
1308 	[ C(OP_WRITE) ] = {
1309 		[ C(RESULT_ACCESS) ] = -1,
1310 		[ C(RESULT_MISS)   ] = -1,
1311 	},
1312 	[ C(OP_PREFETCH) ] = {
1313 		[ C(RESULT_ACCESS) ] = -1,
1314 		[ C(RESULT_MISS)   ] = -1,
1315 	},
1316  },
1317  [ C(BPU ) ] = {
1318 	[ C(OP_READ) ] = {
1319 		[ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1320 		[ C(RESULT_MISS)   ] = 0x03e8, /* BPU_CLEARS.ANY               */
1321 	},
1322 	[ C(OP_WRITE) ] = {
1323 		[ C(RESULT_ACCESS) ] = -1,
1324 		[ C(RESULT_MISS)   ] = -1,
1325 	},
1326 	[ C(OP_PREFETCH) ] = {
1327 		[ C(RESULT_ACCESS) ] = -1,
1328 		[ C(RESULT_MISS)   ] = -1,
1329 	},
1330  },
1331  [ C(NODE) ] = {
1332 	[ C(OP_READ) ] = {
1333 		[ C(RESULT_ACCESS) ] = 0x01b7,
1334 		[ C(RESULT_MISS)   ] = 0x01b7,
1335 	},
1336 	[ C(OP_WRITE) ] = {
1337 		[ C(RESULT_ACCESS) ] = 0x01b7,
1338 		[ C(RESULT_MISS)   ] = 0x01b7,
1339 	},
1340 	[ C(OP_PREFETCH) ] = {
1341 		[ C(RESULT_ACCESS) ] = 0x01b7,
1342 		[ C(RESULT_MISS)   ] = 0x01b7,
1343 	},
1344  },
1345 };
1346 
1347 /*
1348  * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits;
1349  * See IA32 SDM Vol 3B 30.6.1.3
1350  */
1351 
1352 #define NHM_DMND_DATA_RD	(1 << 0)
1353 #define NHM_DMND_RFO		(1 << 1)
1354 #define NHM_DMND_IFETCH		(1 << 2)
1355 #define NHM_DMND_WB		(1 << 3)
1356 #define NHM_PF_DATA_RD		(1 << 4)
1357 #define NHM_PF_DATA_RFO		(1 << 5)
1358 #define NHM_PF_IFETCH		(1 << 6)
1359 #define NHM_OFFCORE_OTHER	(1 << 7)
1360 #define NHM_UNCORE_HIT		(1 << 8)
1361 #define NHM_OTHER_CORE_HIT_SNP	(1 << 9)
1362 #define NHM_OTHER_CORE_HITM	(1 << 10)
1363         			/* reserved */
1364 #define NHM_REMOTE_CACHE_FWD	(1 << 12)
1365 #define NHM_REMOTE_DRAM		(1 << 13)
1366 #define NHM_LOCAL_DRAM		(1 << 14)
1367 #define NHM_NON_DRAM		(1 << 15)
1368 
1369 #define NHM_LOCAL		(NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD)
1370 #define NHM_REMOTE		(NHM_REMOTE_DRAM)
1371 
1372 #define NHM_DMND_READ		(NHM_DMND_DATA_RD)
1373 #define NHM_DMND_WRITE		(NHM_DMND_RFO|NHM_DMND_WB)
1374 #define NHM_DMND_PREFETCH	(NHM_PF_DATA_RD|NHM_PF_DATA_RFO)
1375 
1376 #define NHM_L3_HIT	(NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM)
1377 #define NHM_L3_MISS	(NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD)
1378 #define NHM_L3_ACCESS	(NHM_L3_HIT|NHM_L3_MISS)
1379 
1380 static __initconst const u64 nehalem_hw_cache_extra_regs
1381 				[PERF_COUNT_HW_CACHE_MAX]
1382 				[PERF_COUNT_HW_CACHE_OP_MAX]
1383 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
1384 {
1385  [ C(LL  ) ] = {
1386 	[ C(OP_READ) ] = {
1387 		[ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS,
1388 		[ C(RESULT_MISS)   ] = NHM_DMND_READ|NHM_L3_MISS,
1389 	},
1390 	[ C(OP_WRITE) ] = {
1391 		[ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS,
1392 		[ C(RESULT_MISS)   ] = NHM_DMND_WRITE|NHM_L3_MISS,
1393 	},
1394 	[ C(OP_PREFETCH) ] = {
1395 		[ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS,
1396 		[ C(RESULT_MISS)   ] = NHM_DMND_PREFETCH|NHM_L3_MISS,
1397 	},
1398  },
1399  [ C(NODE) ] = {
1400 	[ C(OP_READ) ] = {
1401 		[ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_LOCAL|NHM_REMOTE,
1402 		[ C(RESULT_MISS)   ] = NHM_DMND_READ|NHM_REMOTE,
1403 	},
1404 	[ C(OP_WRITE) ] = {
1405 		[ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_LOCAL|NHM_REMOTE,
1406 		[ C(RESULT_MISS)   ] = NHM_DMND_WRITE|NHM_REMOTE,
1407 	},
1408 	[ C(OP_PREFETCH) ] = {
1409 		[ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_LOCAL|NHM_REMOTE,
1410 		[ C(RESULT_MISS)   ] = NHM_DMND_PREFETCH|NHM_REMOTE,
1411 	},
1412  },
1413 };
1414 
1415 static __initconst const u64 nehalem_hw_cache_event_ids
1416 				[PERF_COUNT_HW_CACHE_MAX]
1417 				[PERF_COUNT_HW_CACHE_OP_MAX]
1418 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
1419 {
1420  [ C(L1D) ] = {
1421 	[ C(OP_READ) ] = {
1422 		[ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS       */
1423 		[ C(RESULT_MISS)   ] = 0x0151, /* L1D.REPL                     */
1424 	},
1425 	[ C(OP_WRITE) ] = {
1426 		[ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES      */
1427 		[ C(RESULT_MISS)   ] = 0x0251, /* L1D.M_REPL                   */
1428 	},
1429 	[ C(OP_PREFETCH) ] = {
1430 		[ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS        */
1431 		[ C(RESULT_MISS)   ] = 0x024e, /* L1D_PREFETCH.MISS            */
1432 	},
1433  },
1434  [ C(L1I ) ] = {
1435 	[ C(OP_READ) ] = {
1436 		[ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS                    */
1437 		[ C(RESULT_MISS)   ] = 0x0280, /* L1I.MISSES                   */
1438 	},
1439 	[ C(OP_WRITE) ] = {
1440 		[ C(RESULT_ACCESS) ] = -1,
1441 		[ C(RESULT_MISS)   ] = -1,
1442 	},
1443 	[ C(OP_PREFETCH) ] = {
1444 		[ C(RESULT_ACCESS) ] = 0x0,
1445 		[ C(RESULT_MISS)   ] = 0x0,
1446 	},
1447  },
1448  [ C(LL  ) ] = {
1449 	[ C(OP_READ) ] = {
1450 		/* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
1451 		[ C(RESULT_ACCESS) ] = 0x01b7,
1452 		/* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
1453 		[ C(RESULT_MISS)   ] = 0x01b7,
1454 	},
1455 	/*
1456 	 * Use RFO, not WRITEBACK, because a write miss would typically occur
1457 	 * on RFO.
1458 	 */
1459 	[ C(OP_WRITE) ] = {
1460 		/* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
1461 		[ C(RESULT_ACCESS) ] = 0x01b7,
1462 		/* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
1463 		[ C(RESULT_MISS)   ] = 0x01b7,
1464 	},
1465 	[ C(OP_PREFETCH) ] = {
1466 		/* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
1467 		[ C(RESULT_ACCESS) ] = 0x01b7,
1468 		/* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
1469 		[ C(RESULT_MISS)   ] = 0x01b7,
1470 	},
1471  },
1472  [ C(DTLB) ] = {
1473 	[ C(OP_READ) ] = {
1474 		[ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI   (alias)  */
1475 		[ C(RESULT_MISS)   ] = 0x0108, /* DTLB_LOAD_MISSES.ANY         */
1476 	},
1477 	[ C(OP_WRITE) ] = {
1478 		[ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI   (alias)  */
1479 		[ C(RESULT_MISS)   ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS  */
1480 	},
1481 	[ C(OP_PREFETCH) ] = {
1482 		[ C(RESULT_ACCESS) ] = 0x0,
1483 		[ C(RESULT_MISS)   ] = 0x0,
1484 	},
1485  },
1486  [ C(ITLB) ] = {
1487 	[ C(OP_READ) ] = {
1488 		[ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P           */
1489 		[ C(RESULT_MISS)   ] = 0x20c8, /* ITLB_MISS_RETIRED            */
1490 	},
1491 	[ C(OP_WRITE) ] = {
1492 		[ C(RESULT_ACCESS) ] = -1,
1493 		[ C(RESULT_MISS)   ] = -1,
1494 	},
1495 	[ C(OP_PREFETCH) ] = {
1496 		[ C(RESULT_ACCESS) ] = -1,
1497 		[ C(RESULT_MISS)   ] = -1,
1498 	},
1499  },
1500  [ C(BPU ) ] = {
1501 	[ C(OP_READ) ] = {
1502 		[ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
1503 		[ C(RESULT_MISS)   ] = 0x03e8, /* BPU_CLEARS.ANY               */
1504 	},
1505 	[ C(OP_WRITE) ] = {
1506 		[ C(RESULT_ACCESS) ] = -1,
1507 		[ C(RESULT_MISS)   ] = -1,
1508 	},
1509 	[ C(OP_PREFETCH) ] = {
1510 		[ C(RESULT_ACCESS) ] = -1,
1511 		[ C(RESULT_MISS)   ] = -1,
1512 	},
1513  },
1514  [ C(NODE) ] = {
1515 	[ C(OP_READ) ] = {
1516 		[ C(RESULT_ACCESS) ] = 0x01b7,
1517 		[ C(RESULT_MISS)   ] = 0x01b7,
1518 	},
1519 	[ C(OP_WRITE) ] = {
1520 		[ C(RESULT_ACCESS) ] = 0x01b7,
1521 		[ C(RESULT_MISS)   ] = 0x01b7,
1522 	},
1523 	[ C(OP_PREFETCH) ] = {
1524 		[ C(RESULT_ACCESS) ] = 0x01b7,
1525 		[ C(RESULT_MISS)   ] = 0x01b7,
1526 	},
1527  },
1528 };
1529 
1530 static __initconst const u64 core2_hw_cache_event_ids
1531 				[PERF_COUNT_HW_CACHE_MAX]
1532 				[PERF_COUNT_HW_CACHE_OP_MAX]
1533 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
1534 {
1535  [ C(L1D) ] = {
1536 	[ C(OP_READ) ] = {
1537 		[ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI          */
1538 		[ C(RESULT_MISS)   ] = 0x0140, /* L1D_CACHE_LD.I_STATE       */
1539 	},
1540 	[ C(OP_WRITE) ] = {
1541 		[ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI          */
1542 		[ C(RESULT_MISS)   ] = 0x0141, /* L1D_CACHE_ST.I_STATE       */
1543 	},
1544 	[ C(OP_PREFETCH) ] = {
1545 		[ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS      */
1546 		[ C(RESULT_MISS)   ] = 0,
1547 	},
1548  },
1549  [ C(L1I ) ] = {
1550 	[ C(OP_READ) ] = {
1551 		[ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS                  */
1552 		[ C(RESULT_MISS)   ] = 0x0081, /* L1I.MISSES                 */
1553 	},
1554 	[ C(OP_WRITE) ] = {
1555 		[ C(RESULT_ACCESS) ] = -1,
1556 		[ C(RESULT_MISS)   ] = -1,
1557 	},
1558 	[ C(OP_PREFETCH) ] = {
1559 		[ C(RESULT_ACCESS) ] = 0,
1560 		[ C(RESULT_MISS)   ] = 0,
1561 	},
1562  },
1563  [ C(LL  ) ] = {
1564 	[ C(OP_READ) ] = {
1565 		[ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI                 */
1566 		[ C(RESULT_MISS)   ] = 0x4129, /* L2_LD.ISTATE               */
1567 	},
1568 	[ C(OP_WRITE) ] = {
1569 		[ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI                 */
1570 		[ C(RESULT_MISS)   ] = 0x412A, /* L2_ST.ISTATE               */
1571 	},
1572 	[ C(OP_PREFETCH) ] = {
1573 		[ C(RESULT_ACCESS) ] = 0,
1574 		[ C(RESULT_MISS)   ] = 0,
1575 	},
1576  },
1577  [ C(DTLB) ] = {
1578 	[ C(OP_READ) ] = {
1579 		[ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI  (alias) */
1580 		[ C(RESULT_MISS)   ] = 0x0208, /* DTLB_MISSES.MISS_LD        */
1581 	},
1582 	[ C(OP_WRITE) ] = {
1583 		[ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI  (alias) */
1584 		[ C(RESULT_MISS)   ] = 0x0808, /* DTLB_MISSES.MISS_ST        */
1585 	},
1586 	[ C(OP_PREFETCH) ] = {
1587 		[ C(RESULT_ACCESS) ] = 0,
1588 		[ C(RESULT_MISS)   ] = 0,
1589 	},
1590  },
1591  [ C(ITLB) ] = {
1592 	[ C(OP_READ) ] = {
1593 		[ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P         */
1594 		[ C(RESULT_MISS)   ] = 0x1282, /* ITLBMISSES                 */
1595 	},
1596 	[ C(OP_WRITE) ] = {
1597 		[ C(RESULT_ACCESS) ] = -1,
1598 		[ C(RESULT_MISS)   ] = -1,
1599 	},
1600 	[ C(OP_PREFETCH) ] = {
1601 		[ C(RESULT_ACCESS) ] = -1,
1602 		[ C(RESULT_MISS)   ] = -1,
1603 	},
1604  },
1605  [ C(BPU ) ] = {
1606 	[ C(OP_READ) ] = {
1607 		[ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY        */
1608 		[ C(RESULT_MISS)   ] = 0x00c5, /* BP_INST_RETIRED.MISPRED    */
1609 	},
1610 	[ C(OP_WRITE) ] = {
1611 		[ C(RESULT_ACCESS) ] = -1,
1612 		[ C(RESULT_MISS)   ] = -1,
1613 	},
1614 	[ C(OP_PREFETCH) ] = {
1615 		[ C(RESULT_ACCESS) ] = -1,
1616 		[ C(RESULT_MISS)   ] = -1,
1617 	},
1618  },
1619 };
1620 
1621 static __initconst const u64 atom_hw_cache_event_ids
1622 				[PERF_COUNT_HW_CACHE_MAX]
1623 				[PERF_COUNT_HW_CACHE_OP_MAX]
1624 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
1625 {
1626  [ C(L1D) ] = {
1627 	[ C(OP_READ) ] = {
1628 		[ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD               */
1629 		[ C(RESULT_MISS)   ] = 0,
1630 	},
1631 	[ C(OP_WRITE) ] = {
1632 		[ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST               */
1633 		[ C(RESULT_MISS)   ] = 0,
1634 	},
1635 	[ C(OP_PREFETCH) ] = {
1636 		[ C(RESULT_ACCESS) ] = 0x0,
1637 		[ C(RESULT_MISS)   ] = 0,
1638 	},
1639  },
1640  [ C(L1I ) ] = {
1641 	[ C(OP_READ) ] = {
1642 		[ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS                  */
1643 		[ C(RESULT_MISS)   ] = 0x0280, /* L1I.MISSES                 */
1644 	},
1645 	[ C(OP_WRITE) ] = {
1646 		[ C(RESULT_ACCESS) ] = -1,
1647 		[ C(RESULT_MISS)   ] = -1,
1648 	},
1649 	[ C(OP_PREFETCH) ] = {
1650 		[ C(RESULT_ACCESS) ] = 0,
1651 		[ C(RESULT_MISS)   ] = 0,
1652 	},
1653  },
1654  [ C(LL  ) ] = {
1655 	[ C(OP_READ) ] = {
1656 		[ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI                 */
1657 		[ C(RESULT_MISS)   ] = 0x4129, /* L2_LD.ISTATE               */
1658 	},
1659 	[ C(OP_WRITE) ] = {
1660 		[ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI                 */
1661 		[ C(RESULT_MISS)   ] = 0x412A, /* L2_ST.ISTATE               */
1662 	},
1663 	[ C(OP_PREFETCH) ] = {
1664 		[ C(RESULT_ACCESS) ] = 0,
1665 		[ C(RESULT_MISS)   ] = 0,
1666 	},
1667  },
1668  [ C(DTLB) ] = {
1669 	[ C(OP_READ) ] = {
1670 		[ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI  (alias) */
1671 		[ C(RESULT_MISS)   ] = 0x0508, /* DTLB_MISSES.MISS_LD        */
1672 	},
1673 	[ C(OP_WRITE) ] = {
1674 		[ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI  (alias) */
1675 		[ C(RESULT_MISS)   ] = 0x0608, /* DTLB_MISSES.MISS_ST        */
1676 	},
1677 	[ C(OP_PREFETCH) ] = {
1678 		[ C(RESULT_ACCESS) ] = 0,
1679 		[ C(RESULT_MISS)   ] = 0,
1680 	},
1681  },
1682  [ C(ITLB) ] = {
1683 	[ C(OP_READ) ] = {
1684 		[ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P         */
1685 		[ C(RESULT_MISS)   ] = 0x0282, /* ITLB.MISSES                */
1686 	},
1687 	[ C(OP_WRITE) ] = {
1688 		[ C(RESULT_ACCESS) ] = -1,
1689 		[ C(RESULT_MISS)   ] = -1,
1690 	},
1691 	[ C(OP_PREFETCH) ] = {
1692 		[ C(RESULT_ACCESS) ] = -1,
1693 		[ C(RESULT_MISS)   ] = -1,
1694 	},
1695  },
1696  [ C(BPU ) ] = {
1697 	[ C(OP_READ) ] = {
1698 		[ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY        */
1699 		[ C(RESULT_MISS)   ] = 0x00c5, /* BP_INST_RETIRED.MISPRED    */
1700 	},
1701 	[ C(OP_WRITE) ] = {
1702 		[ C(RESULT_ACCESS) ] = -1,
1703 		[ C(RESULT_MISS)   ] = -1,
1704 	},
1705 	[ C(OP_PREFETCH) ] = {
1706 		[ C(RESULT_ACCESS) ] = -1,
1707 		[ C(RESULT_MISS)   ] = -1,
1708 	},
1709  },
1710 };
1711 
1712 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_slm, "event=0x3c");
1713 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_slm, "2");
1714 /* no_alloc_cycles.not_delivered */
1715 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_slm,
1716 	       "event=0xca,umask=0x50");
1717 EVENT_ATTR_STR(topdown-fetch-bubbles.scale, td_fetch_bubbles_scale_slm, "2");
1718 /* uops_retired.all */
1719 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_slm,
1720 	       "event=0xc2,umask=0x10");
1721 /* uops_retired.all */
1722 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_slm,
1723 	       "event=0xc2,umask=0x10");
1724 
1725 static struct attribute *slm_events_attrs[] = {
1726 	EVENT_PTR(td_total_slots_slm),
1727 	EVENT_PTR(td_total_slots_scale_slm),
1728 	EVENT_PTR(td_fetch_bubbles_slm),
1729 	EVENT_PTR(td_fetch_bubbles_scale_slm),
1730 	EVENT_PTR(td_slots_issued_slm),
1731 	EVENT_PTR(td_slots_retired_slm),
1732 	NULL
1733 };
1734 
1735 static struct extra_reg intel_slm_extra_regs[] __read_mostly =
1736 {
1737 	/* must define OFFCORE_RSP_X first, see intel_fixup_er() */
1738 	INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x768005ffffull, RSP_0),
1739 	INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x368005ffffull, RSP_1),
1740 	EVENT_EXTRA_END
1741 };
1742 
1743 #define SLM_DMND_READ		SNB_DMND_DATA_RD
1744 #define SLM_DMND_WRITE		SNB_DMND_RFO
1745 #define SLM_DMND_PREFETCH	(SNB_PF_DATA_RD|SNB_PF_RFO)
1746 
1747 #define SLM_SNP_ANY		(SNB_SNP_NONE|SNB_SNP_MISS|SNB_NO_FWD|SNB_HITM)
1748 #define SLM_LLC_ACCESS		SNB_RESP_ANY
1749 #define SLM_LLC_MISS		(SLM_SNP_ANY|SNB_NON_DRAM)
1750 
1751 static __initconst const u64 slm_hw_cache_extra_regs
1752 				[PERF_COUNT_HW_CACHE_MAX]
1753 				[PERF_COUNT_HW_CACHE_OP_MAX]
1754 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
1755 {
1756  [ C(LL  ) ] = {
1757 	[ C(OP_READ) ] = {
1758 		[ C(RESULT_ACCESS) ] = SLM_DMND_READ|SLM_LLC_ACCESS,
1759 		[ C(RESULT_MISS)   ] = 0,
1760 	},
1761 	[ C(OP_WRITE) ] = {
1762 		[ C(RESULT_ACCESS) ] = SLM_DMND_WRITE|SLM_LLC_ACCESS,
1763 		[ C(RESULT_MISS)   ] = SLM_DMND_WRITE|SLM_LLC_MISS,
1764 	},
1765 	[ C(OP_PREFETCH) ] = {
1766 		[ C(RESULT_ACCESS) ] = SLM_DMND_PREFETCH|SLM_LLC_ACCESS,
1767 		[ C(RESULT_MISS)   ] = SLM_DMND_PREFETCH|SLM_LLC_MISS,
1768 	},
1769  },
1770 };
1771 
1772 static __initconst const u64 slm_hw_cache_event_ids
1773 				[PERF_COUNT_HW_CACHE_MAX]
1774 				[PERF_COUNT_HW_CACHE_OP_MAX]
1775 				[PERF_COUNT_HW_CACHE_RESULT_MAX] =
1776 {
1777  [ C(L1D) ] = {
1778 	[ C(OP_READ) ] = {
1779 		[ C(RESULT_ACCESS) ] = 0,
1780 		[ C(RESULT_MISS)   ] = 0x0104, /* LD_DCU_MISS */
1781 	},
1782 	[ C(OP_WRITE) ] = {
1783 		[ C(RESULT_ACCESS) ] = 0,
1784 		[ C(RESULT_MISS)   ] = 0,
1785 	},
1786 	[ C(OP_PREFETCH) ] = {
1787 		[ C(RESULT_ACCESS) ] = 0,
1788 		[ C(RESULT_MISS)   ] = 0,
1789 	},
1790  },
1791  [ C(L1I ) ] = {
1792 	[ C(OP_READ) ] = {
1793 		[ C(RESULT_ACCESS) ] = 0x0380, /* ICACHE.ACCESSES */
1794 		[ C(RESULT_MISS)   ] = 0x0280, /* ICACGE.MISSES */
1795 	},
1796 	[ C(OP_WRITE) ] = {
1797 		[ C(RESULT_ACCESS) ] = -1,
1798 		[ C(RESULT_MISS)   ] = -1,
1799 	},
1800 	[ C(OP_PREFETCH) ] = {
1801 		[ C(RESULT_ACCESS) ] = 0,
1802 		[ C(RESULT_MISS)   ] = 0,
1803 	},
1804  },
1805  [ C(LL  ) ] = {
1806 	[ C(OP_READ) ] = {
1807 		/* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
1808 		[ C(RESULT_ACCESS) ] = 0x01b7,
1809 		[ C(RESULT_MISS)   ] = 0,
1810 	},
1811 	[ C(OP_WRITE) ] = {
1812 		/* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
1813 		[ C(RESULT_ACCESS) ] = 0x01b7,
1814 		/* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
1815 		[ C(RESULT_MISS)   ] = 0x01b7,
1816 	},
1817 	[ C(OP_PREFETCH) ] = {
1818 		/* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
1819 		[ C(RESULT_ACCESS) ] = 0x01b7,
1820 		/* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
1821 		[ C(RESULT_MISS)   ] = 0x01b7,
1822 	},
1823  },
1824  [ C(DTLB) ] = {
1825 	[ C(OP_READ) ] = {
1826 		[ C(RESULT_ACCESS) ] = 0,
1827 		[ C(RESULT_MISS)   ] = 0x0804, /* LD_DTLB_MISS */
1828 	},
1829 	[ C(OP_WRITE) ] = {
1830 		[ C(RESULT_ACCESS) ] = 0,
1831 		[ C(RESULT_MISS)   ] = 0,
1832 	},
1833 	[ C(OP_PREFETCH) ] = {
1834 		[ C(RESULT_ACCESS) ] = 0,
1835 		[ C(RESULT_MISS)   ] = 0,
1836 	},
1837  },
1838  [ C(ITLB) ] = {
1839 	[ C(OP_READ) ] = {
1840 		[ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */
1841 		[ C(RESULT_MISS)   ] = 0x40205, /* PAGE_WALKS.I_SIDE_WALKS */
1842 	},
1843 	[ C(OP_WRITE) ] = {
1844 		[ C(RESULT_ACCESS) ] = -1,
1845 		[ C(RESULT_MISS)   ] = -1,
1846 	},
1847 	[ C(OP_PREFETCH) ] = {
1848 		[ C(RESULT_ACCESS) ] = -1,
1849 		[ C(RESULT_MISS)   ] = -1,
1850 	},
1851  },
1852  [ C(BPU ) ] = {
1853 	[ C(OP_READ) ] = {
1854 		[ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */
1855 		[ C(RESULT_MISS)   ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */
1856 	},
1857 	[ C(OP_WRITE) ] = {
1858 		[ C(RESULT_ACCESS) ] = -1,
1859 		[ C(RESULT_MISS)   ] = -1,
1860 	},
1861 	[ C(OP_PREFETCH) ] = {
1862 		[ C(RESULT_ACCESS) ] = -1,
1863 		[ C(RESULT_MISS)   ] = -1,
1864 	},
1865  },
1866 };
1867 
1868 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_glm, "event=0x3c");
1869 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_glm, "3");
1870 /* UOPS_NOT_DELIVERED.ANY */
1871 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_glm, "event=0x9c");
1872 /* ISSUE_SLOTS_NOT_CONSUMED.RECOVERY */
1873 EVENT_ATTR_STR(topdown-recovery-bubbles, td_recovery_bubbles_glm, "event=0xca,umask=0x02");
1874 /* UOPS_RETIRED.ANY */
1875 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_glm, "event=0xc2");
1876 /* UOPS_ISSUED.ANY */
1877 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_glm, "event=0x0e");
1878 
1879 static struct attribute *glm_events_attrs[] = {
1880 	EVENT_PTR(td_total_slots_glm),
1881 	EVENT_PTR(td_total_slots_scale_glm),
1882 	EVENT_PTR(td_fetch_bubbles_glm),
1883 	EVENT_PTR(td_recovery_bubbles_glm),
1884 	EVENT_PTR(td_slots_issued_glm),
1885 	EVENT_PTR(td_slots_retired_glm),
1886 	NULL
1887 };
1888 
1889 static struct extra_reg intel_glm_extra_regs[] __read_mostly = {
1890 	/* must define OFFCORE_RSP_X first, see intel_fixup_er() */
1891 	INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x760005ffbfull, RSP_0),
1892 	INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x360005ffbfull, RSP_1),
1893 	EVENT_EXTRA_END
1894 };
1895 
1896 #define GLM_DEMAND_DATA_RD		BIT_ULL(0)
1897 #define GLM_DEMAND_RFO			BIT_ULL(1)
1898 #define GLM_ANY_RESPONSE		BIT_ULL(16)
1899 #define GLM_SNP_NONE_OR_MISS		BIT_ULL(33)
1900 #define GLM_DEMAND_READ			GLM_DEMAND_DATA_RD
1901 #define GLM_DEMAND_WRITE		GLM_DEMAND_RFO
1902 #define GLM_DEMAND_PREFETCH		(SNB_PF_DATA_RD|SNB_PF_RFO)
1903 #define GLM_LLC_ACCESS			GLM_ANY_RESPONSE
1904 #define GLM_SNP_ANY			(GLM_SNP_NONE_OR_MISS|SNB_NO_FWD|SNB_HITM)
1905 #define GLM_LLC_MISS			(GLM_SNP_ANY|SNB_NON_DRAM)
1906 
1907 static __initconst const u64 glm_hw_cache_event_ids
1908 				[PERF_COUNT_HW_CACHE_MAX]
1909 				[PERF_COUNT_HW_CACHE_OP_MAX]
1910 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1911 	[C(L1D)] = {
1912 		[C(OP_READ)] = {
1913 			[C(RESULT_ACCESS)]	= 0x81d0,	/* MEM_UOPS_RETIRED.ALL_LOADS */
1914 			[C(RESULT_MISS)]	= 0x0,
1915 		},
1916 		[C(OP_WRITE)] = {
1917 			[C(RESULT_ACCESS)]	= 0x82d0,	/* MEM_UOPS_RETIRED.ALL_STORES */
1918 			[C(RESULT_MISS)]	= 0x0,
1919 		},
1920 		[C(OP_PREFETCH)] = {
1921 			[C(RESULT_ACCESS)]	= 0x0,
1922 			[C(RESULT_MISS)]	= 0x0,
1923 		},
1924 	},
1925 	[C(L1I)] = {
1926 		[C(OP_READ)] = {
1927 			[C(RESULT_ACCESS)]	= 0x0380,	/* ICACHE.ACCESSES */
1928 			[C(RESULT_MISS)]	= 0x0280,	/* ICACHE.MISSES */
1929 		},
1930 		[C(OP_WRITE)] = {
1931 			[C(RESULT_ACCESS)]	= -1,
1932 			[C(RESULT_MISS)]	= -1,
1933 		},
1934 		[C(OP_PREFETCH)] = {
1935 			[C(RESULT_ACCESS)]	= 0x0,
1936 			[C(RESULT_MISS)]	= 0x0,
1937 		},
1938 	},
1939 	[C(LL)] = {
1940 		[C(OP_READ)] = {
1941 			[C(RESULT_ACCESS)]	= 0x1b7,	/* OFFCORE_RESPONSE */
1942 			[C(RESULT_MISS)]	= 0x1b7,	/* OFFCORE_RESPONSE */
1943 		},
1944 		[C(OP_WRITE)] = {
1945 			[C(RESULT_ACCESS)]	= 0x1b7,	/* OFFCORE_RESPONSE */
1946 			[C(RESULT_MISS)]	= 0x1b7,	/* OFFCORE_RESPONSE */
1947 		},
1948 		[C(OP_PREFETCH)] = {
1949 			[C(RESULT_ACCESS)]	= 0x1b7,	/* OFFCORE_RESPONSE */
1950 			[C(RESULT_MISS)]	= 0x1b7,	/* OFFCORE_RESPONSE */
1951 		},
1952 	},
1953 	[C(DTLB)] = {
1954 		[C(OP_READ)] = {
1955 			[C(RESULT_ACCESS)]	= 0x81d0,	/* MEM_UOPS_RETIRED.ALL_LOADS */
1956 			[C(RESULT_MISS)]	= 0x0,
1957 		},
1958 		[C(OP_WRITE)] = {
1959 			[C(RESULT_ACCESS)]	= 0x82d0,	/* MEM_UOPS_RETIRED.ALL_STORES */
1960 			[C(RESULT_MISS)]	= 0x0,
1961 		},
1962 		[C(OP_PREFETCH)] = {
1963 			[C(RESULT_ACCESS)]	= 0x0,
1964 			[C(RESULT_MISS)]	= 0x0,
1965 		},
1966 	},
1967 	[C(ITLB)] = {
1968 		[C(OP_READ)] = {
1969 			[C(RESULT_ACCESS)]	= 0x00c0,	/* INST_RETIRED.ANY_P */
1970 			[C(RESULT_MISS)]	= 0x0481,	/* ITLB.MISS */
1971 		},
1972 		[C(OP_WRITE)] = {
1973 			[C(RESULT_ACCESS)]	= -1,
1974 			[C(RESULT_MISS)]	= -1,
1975 		},
1976 		[C(OP_PREFETCH)] = {
1977 			[C(RESULT_ACCESS)]	= -1,
1978 			[C(RESULT_MISS)]	= -1,
1979 		},
1980 	},
1981 	[C(BPU)] = {
1982 		[C(OP_READ)] = {
1983 			[C(RESULT_ACCESS)]	= 0x00c4,	/* BR_INST_RETIRED.ALL_BRANCHES */
1984 			[C(RESULT_MISS)]	= 0x00c5,	/* BR_MISP_RETIRED.ALL_BRANCHES */
1985 		},
1986 		[C(OP_WRITE)] = {
1987 			[C(RESULT_ACCESS)]	= -1,
1988 			[C(RESULT_MISS)]	= -1,
1989 		},
1990 		[C(OP_PREFETCH)] = {
1991 			[C(RESULT_ACCESS)]	= -1,
1992 			[C(RESULT_MISS)]	= -1,
1993 		},
1994 	},
1995 };
1996 
1997 static __initconst const u64 glm_hw_cache_extra_regs
1998 				[PERF_COUNT_HW_CACHE_MAX]
1999 				[PERF_COUNT_HW_CACHE_OP_MAX]
2000 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
2001 	[C(LL)] = {
2002 		[C(OP_READ)] = {
2003 			[C(RESULT_ACCESS)]	= GLM_DEMAND_READ|
2004 						  GLM_LLC_ACCESS,
2005 			[C(RESULT_MISS)]	= GLM_DEMAND_READ|
2006 						  GLM_LLC_MISS,
2007 		},
2008 		[C(OP_WRITE)] = {
2009 			[C(RESULT_ACCESS)]	= GLM_DEMAND_WRITE|
2010 						  GLM_LLC_ACCESS,
2011 			[C(RESULT_MISS)]	= GLM_DEMAND_WRITE|
2012 						  GLM_LLC_MISS,
2013 		},
2014 		[C(OP_PREFETCH)] = {
2015 			[C(RESULT_ACCESS)]	= GLM_DEMAND_PREFETCH|
2016 						  GLM_LLC_ACCESS,
2017 			[C(RESULT_MISS)]	= GLM_DEMAND_PREFETCH|
2018 						  GLM_LLC_MISS,
2019 		},
2020 	},
2021 };
2022 
2023 static __initconst const u64 glp_hw_cache_event_ids
2024 				[PERF_COUNT_HW_CACHE_MAX]
2025 				[PERF_COUNT_HW_CACHE_OP_MAX]
2026 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
2027 	[C(L1D)] = {
2028 		[C(OP_READ)] = {
2029 			[C(RESULT_ACCESS)]	= 0x81d0,	/* MEM_UOPS_RETIRED.ALL_LOADS */
2030 			[C(RESULT_MISS)]	= 0x0,
2031 		},
2032 		[C(OP_WRITE)] = {
2033 			[C(RESULT_ACCESS)]	= 0x82d0,	/* MEM_UOPS_RETIRED.ALL_STORES */
2034 			[C(RESULT_MISS)]	= 0x0,
2035 		},
2036 		[C(OP_PREFETCH)] = {
2037 			[C(RESULT_ACCESS)]	= 0x0,
2038 			[C(RESULT_MISS)]	= 0x0,
2039 		},
2040 	},
2041 	[C(L1I)] = {
2042 		[C(OP_READ)] = {
2043 			[C(RESULT_ACCESS)]	= 0x0380,	/* ICACHE.ACCESSES */
2044 			[C(RESULT_MISS)]	= 0x0280,	/* ICACHE.MISSES */
2045 		},
2046 		[C(OP_WRITE)] = {
2047 			[C(RESULT_ACCESS)]	= -1,
2048 			[C(RESULT_MISS)]	= -1,
2049 		},
2050 		[C(OP_PREFETCH)] = {
2051 			[C(RESULT_ACCESS)]	= 0x0,
2052 			[C(RESULT_MISS)]	= 0x0,
2053 		},
2054 	},
2055 	[C(LL)] = {
2056 		[C(OP_READ)] = {
2057 			[C(RESULT_ACCESS)]	= 0x1b7,	/* OFFCORE_RESPONSE */
2058 			[C(RESULT_MISS)]	= 0x1b7,	/* OFFCORE_RESPONSE */
2059 		},
2060 		[C(OP_WRITE)] = {
2061 			[C(RESULT_ACCESS)]	= 0x1b7,	/* OFFCORE_RESPONSE */
2062 			[C(RESULT_MISS)]	= 0x1b7,	/* OFFCORE_RESPONSE */
2063 		},
2064 		[C(OP_PREFETCH)] = {
2065 			[C(RESULT_ACCESS)]	= 0x0,
2066 			[C(RESULT_MISS)]	= 0x0,
2067 		},
2068 	},
2069 	[C(DTLB)] = {
2070 		[C(OP_READ)] = {
2071 			[C(RESULT_ACCESS)]	= 0x81d0,	/* MEM_UOPS_RETIRED.ALL_LOADS */
2072 			[C(RESULT_MISS)]	= 0xe08,	/* DTLB_LOAD_MISSES.WALK_COMPLETED */
2073 		},
2074 		[C(OP_WRITE)] = {
2075 			[C(RESULT_ACCESS)]	= 0x82d0,	/* MEM_UOPS_RETIRED.ALL_STORES */
2076 			[C(RESULT_MISS)]	= 0xe49,	/* DTLB_STORE_MISSES.WALK_COMPLETED */
2077 		},
2078 		[C(OP_PREFETCH)] = {
2079 			[C(RESULT_ACCESS)]	= 0x0,
2080 			[C(RESULT_MISS)]	= 0x0,
2081 		},
2082 	},
2083 	[C(ITLB)] = {
2084 		[C(OP_READ)] = {
2085 			[C(RESULT_ACCESS)]	= 0x00c0,	/* INST_RETIRED.ANY_P */
2086 			[C(RESULT_MISS)]	= 0x0481,	/* ITLB.MISS */
2087 		},
2088 		[C(OP_WRITE)] = {
2089 			[C(RESULT_ACCESS)]	= -1,
2090 			[C(RESULT_MISS)]	= -1,
2091 		},
2092 		[C(OP_PREFETCH)] = {
2093 			[C(RESULT_ACCESS)]	= -1,
2094 			[C(RESULT_MISS)]	= -1,
2095 		},
2096 	},
2097 	[C(BPU)] = {
2098 		[C(OP_READ)] = {
2099 			[C(RESULT_ACCESS)]	= 0x00c4,	/* BR_INST_RETIRED.ALL_BRANCHES */
2100 			[C(RESULT_MISS)]	= 0x00c5,	/* BR_MISP_RETIRED.ALL_BRANCHES */
2101 		},
2102 		[C(OP_WRITE)] = {
2103 			[C(RESULT_ACCESS)]	= -1,
2104 			[C(RESULT_MISS)]	= -1,
2105 		},
2106 		[C(OP_PREFETCH)] = {
2107 			[C(RESULT_ACCESS)]	= -1,
2108 			[C(RESULT_MISS)]	= -1,
2109 		},
2110 	},
2111 };
2112 
2113 static __initconst const u64 glp_hw_cache_extra_regs
2114 				[PERF_COUNT_HW_CACHE_MAX]
2115 				[PERF_COUNT_HW_CACHE_OP_MAX]
2116 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
2117 	[C(LL)] = {
2118 		[C(OP_READ)] = {
2119 			[C(RESULT_ACCESS)]	= GLM_DEMAND_READ|
2120 						  GLM_LLC_ACCESS,
2121 			[C(RESULT_MISS)]	= GLM_DEMAND_READ|
2122 						  GLM_LLC_MISS,
2123 		},
2124 		[C(OP_WRITE)] = {
2125 			[C(RESULT_ACCESS)]	= GLM_DEMAND_WRITE|
2126 						  GLM_LLC_ACCESS,
2127 			[C(RESULT_MISS)]	= GLM_DEMAND_WRITE|
2128 						  GLM_LLC_MISS,
2129 		},
2130 		[C(OP_PREFETCH)] = {
2131 			[C(RESULT_ACCESS)]	= 0x0,
2132 			[C(RESULT_MISS)]	= 0x0,
2133 		},
2134 	},
2135 };
2136 
2137 #define TNT_LOCAL_DRAM			BIT_ULL(26)
2138 #define TNT_DEMAND_READ			GLM_DEMAND_DATA_RD
2139 #define TNT_DEMAND_WRITE		GLM_DEMAND_RFO
2140 #define TNT_LLC_ACCESS			GLM_ANY_RESPONSE
2141 #define TNT_SNP_ANY			(SNB_SNP_NOT_NEEDED|SNB_SNP_MISS| \
2142 					 SNB_NO_FWD|SNB_SNP_FWD|SNB_HITM)
2143 #define TNT_LLC_MISS			(TNT_SNP_ANY|SNB_NON_DRAM|TNT_LOCAL_DRAM)
2144 
2145 static __initconst const u64 tnt_hw_cache_extra_regs
2146 				[PERF_COUNT_HW_CACHE_MAX]
2147 				[PERF_COUNT_HW_CACHE_OP_MAX]
2148 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
2149 	[C(LL)] = {
2150 		[C(OP_READ)] = {
2151 			[C(RESULT_ACCESS)]	= TNT_DEMAND_READ|
2152 						  TNT_LLC_ACCESS,
2153 			[C(RESULT_MISS)]	= TNT_DEMAND_READ|
2154 						  TNT_LLC_MISS,
2155 		},
2156 		[C(OP_WRITE)] = {
2157 			[C(RESULT_ACCESS)]	= TNT_DEMAND_WRITE|
2158 						  TNT_LLC_ACCESS,
2159 			[C(RESULT_MISS)]	= TNT_DEMAND_WRITE|
2160 						  TNT_LLC_MISS,
2161 		},
2162 		[C(OP_PREFETCH)] = {
2163 			[C(RESULT_ACCESS)]	= 0x0,
2164 			[C(RESULT_MISS)]	= 0x0,
2165 		},
2166 	},
2167 };
2168 
2169 EVENT_ATTR_STR(topdown-fe-bound,       td_fe_bound_tnt,        "event=0x71,umask=0x0");
2170 EVENT_ATTR_STR(topdown-retiring,       td_retiring_tnt,        "event=0xc2,umask=0x0");
2171 EVENT_ATTR_STR(topdown-bad-spec,       td_bad_spec_tnt,        "event=0x73,umask=0x6");
2172 EVENT_ATTR_STR(topdown-be-bound,       td_be_bound_tnt,        "event=0x74,umask=0x0");
2173 
2174 static struct attribute *tnt_events_attrs[] = {
2175 	EVENT_PTR(td_fe_bound_tnt),
2176 	EVENT_PTR(td_retiring_tnt),
2177 	EVENT_PTR(td_bad_spec_tnt),
2178 	EVENT_PTR(td_be_bound_tnt),
2179 	NULL,
2180 };
2181 
2182 static struct extra_reg intel_tnt_extra_regs[] __read_mostly = {
2183 	/* must define OFFCORE_RSP_X first, see intel_fixup_er() */
2184 	INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x800ff0ffffff9fffull, RSP_0),
2185 	INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0xff0ffffff9fffull, RSP_1),
2186 	EVENT_EXTRA_END
2187 };
2188 
2189 EVENT_ATTR_STR(mem-loads,	mem_ld_grt,	"event=0xd0,umask=0x5,ldlat=3");
2190 EVENT_ATTR_STR(mem-stores,	mem_st_grt,	"event=0xd0,umask=0x6");
2191 
2192 static struct attribute *grt_mem_attrs[] = {
2193 	EVENT_PTR(mem_ld_grt),
2194 	EVENT_PTR(mem_st_grt),
2195 	NULL
2196 };
2197 
2198 static struct extra_reg intel_grt_extra_regs[] __read_mostly = {
2199 	/* must define OFFCORE_RSP_X first, see intel_fixup_er() */
2200 	INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffffffffull, RSP_0),
2201 	INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x3fffffffffull, RSP_1),
2202 	INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x5d0),
2203 	EVENT_EXTRA_END
2204 };
2205 
2206 EVENT_ATTR_STR(topdown-retiring,       td_retiring_cmt,        "event=0x72,umask=0x0");
2207 EVENT_ATTR_STR(topdown-bad-spec,       td_bad_spec_cmt,        "event=0x73,umask=0x0");
2208 
2209 static struct attribute *cmt_events_attrs[] = {
2210 	EVENT_PTR(td_fe_bound_tnt),
2211 	EVENT_PTR(td_retiring_cmt),
2212 	EVENT_PTR(td_bad_spec_cmt),
2213 	EVENT_PTR(td_be_bound_tnt),
2214 	NULL
2215 };
2216 
2217 static struct extra_reg intel_cmt_extra_regs[] __read_mostly = {
2218 	/* must define OFFCORE_RSP_X first, see intel_fixup_er() */
2219 	INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x800ff3ffffffffffull, RSP_0),
2220 	INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0xff3ffffffffffull, RSP_1),
2221 	INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x5d0),
2222 	INTEL_UEVENT_EXTRA_REG(0x0127, MSR_SNOOP_RSP_0, 0xffffffffffffffffull, SNOOP_0),
2223 	INTEL_UEVENT_EXTRA_REG(0x0227, MSR_SNOOP_RSP_1, 0xffffffffffffffffull, SNOOP_1),
2224 	EVENT_EXTRA_END
2225 };
2226 
2227 #define KNL_OT_L2_HITE		BIT_ULL(19) /* Other Tile L2 Hit */
2228 #define KNL_OT_L2_HITF		BIT_ULL(20) /* Other Tile L2 Hit */
2229 #define KNL_MCDRAM_LOCAL	BIT_ULL(21)
2230 #define KNL_MCDRAM_FAR		BIT_ULL(22)
2231 #define KNL_DDR_LOCAL		BIT_ULL(23)
2232 #define KNL_DDR_FAR		BIT_ULL(24)
2233 #define KNL_DRAM_ANY		(KNL_MCDRAM_LOCAL | KNL_MCDRAM_FAR | \
2234 				    KNL_DDR_LOCAL | KNL_DDR_FAR)
2235 #define KNL_L2_READ		SLM_DMND_READ
2236 #define KNL_L2_WRITE		SLM_DMND_WRITE
2237 #define KNL_L2_PREFETCH		SLM_DMND_PREFETCH
2238 #define KNL_L2_ACCESS		SLM_LLC_ACCESS
2239 #define KNL_L2_MISS		(KNL_OT_L2_HITE | KNL_OT_L2_HITF | \
2240 				   KNL_DRAM_ANY | SNB_SNP_ANY | \
2241 						  SNB_NON_DRAM)
2242 
2243 static __initconst const u64 knl_hw_cache_extra_regs
2244 				[PERF_COUNT_HW_CACHE_MAX]
2245 				[PERF_COUNT_HW_CACHE_OP_MAX]
2246 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
2247 	[C(LL)] = {
2248 		[C(OP_READ)] = {
2249 			[C(RESULT_ACCESS)] = KNL_L2_READ | KNL_L2_ACCESS,
2250 			[C(RESULT_MISS)]   = 0,
2251 		},
2252 		[C(OP_WRITE)] = {
2253 			[C(RESULT_ACCESS)] = KNL_L2_WRITE | KNL_L2_ACCESS,
2254 			[C(RESULT_MISS)]   = KNL_L2_WRITE | KNL_L2_MISS,
2255 		},
2256 		[C(OP_PREFETCH)] = {
2257 			[C(RESULT_ACCESS)] = KNL_L2_PREFETCH | KNL_L2_ACCESS,
2258 			[C(RESULT_MISS)]   = KNL_L2_PREFETCH | KNL_L2_MISS,
2259 		},
2260 	},
2261 };
2262 
2263 /*
2264  * Used from PMIs where the LBRs are already disabled.
2265  *
2266  * This function could be called consecutively. It is required to remain in
2267  * disabled state if called consecutively.
2268  *
2269  * During consecutive calls, the same disable value will be written to related
2270  * registers, so the PMU state remains unchanged.
2271  *
2272  * intel_bts events don't coexist with intel PMU's BTS events because of
2273  * x86_add_exclusive(x86_lbr_exclusive_lbr); there's no need to keep them
2274  * disabled around intel PMU's event batching etc, only inside the PMI handler.
2275  *
2276  * Avoid PEBS_ENABLE MSR access in PMIs.
2277  * The GLOBAL_CTRL has been disabled. All the counters do not count anymore.
2278  * It doesn't matter if the PEBS is enabled or not.
2279  * Usually, the PEBS status are not changed in PMIs. It's unnecessary to
2280  * access PEBS_ENABLE MSR in disable_all()/enable_all().
2281  * However, there are some cases which may change PEBS status, e.g. PMI
2282  * throttle. The PEBS_ENABLE should be updated where the status changes.
2283  */
__intel_pmu_disable_all(bool bts)2284 static __always_inline void __intel_pmu_disable_all(bool bts)
2285 {
2286 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2287 
2288 	wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
2289 
2290 	if (bts && test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask))
2291 		intel_pmu_disable_bts();
2292 }
2293 
intel_pmu_disable_all(void)2294 static __always_inline void intel_pmu_disable_all(void)
2295 {
2296 	__intel_pmu_disable_all(true);
2297 	intel_pmu_pebs_disable_all();
2298 	intel_pmu_lbr_disable_all();
2299 }
2300 
__intel_pmu_enable_all(int added,bool pmi)2301 static void __intel_pmu_enable_all(int added, bool pmi)
2302 {
2303 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2304 	u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl);
2305 
2306 	intel_pmu_lbr_enable_all(pmi);
2307 
2308 	if (cpuc->fixed_ctrl_val != cpuc->active_fixed_ctrl_val) {
2309 		wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, cpuc->fixed_ctrl_val);
2310 		cpuc->active_fixed_ctrl_val = cpuc->fixed_ctrl_val;
2311 	}
2312 
2313 	wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL,
2314 	       intel_ctrl & ~cpuc->intel_ctrl_guest_mask);
2315 
2316 	if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
2317 		struct perf_event *event =
2318 			cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
2319 
2320 		if (WARN_ON_ONCE(!event))
2321 			return;
2322 
2323 		intel_pmu_enable_bts(event->hw.config);
2324 	}
2325 }
2326 
intel_pmu_enable_all(int added)2327 static void intel_pmu_enable_all(int added)
2328 {
2329 	intel_pmu_pebs_enable_all();
2330 	__intel_pmu_enable_all(added, false);
2331 }
2332 
2333 static noinline int
__intel_pmu_snapshot_branch_stack(struct perf_branch_entry * entries,unsigned int cnt,unsigned long flags)2334 __intel_pmu_snapshot_branch_stack(struct perf_branch_entry *entries,
2335 				  unsigned int cnt, unsigned long flags)
2336 {
2337 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2338 
2339 	intel_pmu_lbr_read();
2340 	cnt = min_t(unsigned int, cnt, x86_pmu.lbr_nr);
2341 
2342 	memcpy(entries, cpuc->lbr_entries, sizeof(struct perf_branch_entry) * cnt);
2343 	intel_pmu_enable_all(0);
2344 	local_irq_restore(flags);
2345 	return cnt;
2346 }
2347 
2348 static int
intel_pmu_snapshot_branch_stack(struct perf_branch_entry * entries,unsigned int cnt)2349 intel_pmu_snapshot_branch_stack(struct perf_branch_entry *entries, unsigned int cnt)
2350 {
2351 	unsigned long flags;
2352 
2353 	/* must not have branches... */
2354 	local_irq_save(flags);
2355 	__intel_pmu_disable_all(false); /* we don't care about BTS */
2356 	__intel_pmu_lbr_disable();
2357 	/*            ... until here */
2358 	return __intel_pmu_snapshot_branch_stack(entries, cnt, flags);
2359 }
2360 
2361 static int
intel_pmu_snapshot_arch_branch_stack(struct perf_branch_entry * entries,unsigned int cnt)2362 intel_pmu_snapshot_arch_branch_stack(struct perf_branch_entry *entries, unsigned int cnt)
2363 {
2364 	unsigned long flags;
2365 
2366 	/* must not have branches... */
2367 	local_irq_save(flags);
2368 	__intel_pmu_disable_all(false); /* we don't care about BTS */
2369 	__intel_pmu_arch_lbr_disable();
2370 	/*            ... until here */
2371 	return __intel_pmu_snapshot_branch_stack(entries, cnt, flags);
2372 }
2373 
2374 /*
2375  * Workaround for:
2376  *   Intel Errata AAK100 (model 26)
2377  *   Intel Errata AAP53  (model 30)
2378  *   Intel Errata BD53   (model 44)
2379  *
2380  * The official story:
2381  *   These chips need to be 'reset' when adding counters by programming the
2382  *   magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either
2383  *   in sequence on the same PMC or on different PMCs.
2384  *
2385  * In practice it appears some of these events do in fact count, and
2386  * we need to program all 4 events.
2387  */
intel_pmu_nhm_workaround(void)2388 static void intel_pmu_nhm_workaround(void)
2389 {
2390 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2391 	static const unsigned long nhm_magic[4] = {
2392 		0x4300B5,
2393 		0x4300D2,
2394 		0x4300B1,
2395 		0x4300B1
2396 	};
2397 	struct perf_event *event;
2398 	int i;
2399 
2400 	/*
2401 	 * The Errata requires below steps:
2402 	 * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL;
2403 	 * 2) Configure 4 PERFEVTSELx with the magic events and clear
2404 	 *    the corresponding PMCx;
2405 	 * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL;
2406 	 * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL;
2407 	 * 5) Clear 4 pairs of ERFEVTSELx and PMCx;
2408 	 */
2409 
2410 	/*
2411 	 * The real steps we choose are a little different from above.
2412 	 * A) To reduce MSR operations, we don't run step 1) as they
2413 	 *    are already cleared before this function is called;
2414 	 * B) Call x86_perf_event_update to save PMCx before configuring
2415 	 *    PERFEVTSELx with magic number;
2416 	 * C) With step 5), we do clear only when the PERFEVTSELx is
2417 	 *    not used currently.
2418 	 * D) Call x86_perf_event_set_period to restore PMCx;
2419 	 */
2420 
2421 	/* We always operate 4 pairs of PERF Counters */
2422 	for (i = 0; i < 4; i++) {
2423 		event = cpuc->events[i];
2424 		if (event)
2425 			static_call(x86_pmu_update)(event);
2426 	}
2427 
2428 	for (i = 0; i < 4; i++) {
2429 		wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]);
2430 		wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0);
2431 	}
2432 
2433 	wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf);
2434 	wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0);
2435 
2436 	for (i = 0; i < 4; i++) {
2437 		event = cpuc->events[i];
2438 
2439 		if (event) {
2440 			static_call(x86_pmu_set_period)(event);
2441 			__x86_pmu_enable_event(&event->hw,
2442 					ARCH_PERFMON_EVENTSEL_ENABLE);
2443 		} else
2444 			wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0);
2445 	}
2446 }
2447 
intel_pmu_nhm_enable_all(int added)2448 static void intel_pmu_nhm_enable_all(int added)
2449 {
2450 	if (added)
2451 		intel_pmu_nhm_workaround();
2452 	intel_pmu_enable_all(added);
2453 }
2454 
intel_set_tfa(struct cpu_hw_events * cpuc,bool on)2455 static void intel_set_tfa(struct cpu_hw_events *cpuc, bool on)
2456 {
2457 	u64 val = on ? MSR_TFA_RTM_FORCE_ABORT : 0;
2458 
2459 	if (cpuc->tfa_shadow != val) {
2460 		cpuc->tfa_shadow = val;
2461 		wrmsrl(MSR_TSX_FORCE_ABORT, val);
2462 	}
2463 }
2464 
intel_tfa_commit_scheduling(struct cpu_hw_events * cpuc,int idx,int cntr)2465 static void intel_tfa_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr)
2466 {
2467 	/*
2468 	 * We're going to use PMC3, make sure TFA is set before we touch it.
2469 	 */
2470 	if (cntr == 3)
2471 		intel_set_tfa(cpuc, true);
2472 }
2473 
intel_tfa_pmu_enable_all(int added)2474 static void intel_tfa_pmu_enable_all(int added)
2475 {
2476 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2477 
2478 	/*
2479 	 * If we find PMC3 is no longer used when we enable the PMU, we can
2480 	 * clear TFA.
2481 	 */
2482 	if (!test_bit(3, cpuc->active_mask))
2483 		intel_set_tfa(cpuc, false);
2484 
2485 	intel_pmu_enable_all(added);
2486 }
2487 
intel_pmu_get_status(void)2488 static inline u64 intel_pmu_get_status(void)
2489 {
2490 	u64 status;
2491 
2492 	rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
2493 
2494 	return status;
2495 }
2496 
intel_pmu_ack_status(u64 ack)2497 static inline void intel_pmu_ack_status(u64 ack)
2498 {
2499 	wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
2500 }
2501 
event_is_checkpointed(struct perf_event * event)2502 static inline bool event_is_checkpointed(struct perf_event *event)
2503 {
2504 	return unlikely(event->hw.config & HSW_IN_TX_CHECKPOINTED) != 0;
2505 }
2506 
intel_set_masks(struct perf_event * event,int idx)2507 static inline void intel_set_masks(struct perf_event *event, int idx)
2508 {
2509 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2510 
2511 	if (event->attr.exclude_host)
2512 		__set_bit(idx, (unsigned long *)&cpuc->intel_ctrl_guest_mask);
2513 	if (event->attr.exclude_guest)
2514 		__set_bit(idx, (unsigned long *)&cpuc->intel_ctrl_host_mask);
2515 	if (event_is_checkpointed(event))
2516 		__set_bit(idx, (unsigned long *)&cpuc->intel_cp_status);
2517 }
2518 
intel_clear_masks(struct perf_event * event,int idx)2519 static inline void intel_clear_masks(struct perf_event *event, int idx)
2520 {
2521 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2522 
2523 	__clear_bit(idx, (unsigned long *)&cpuc->intel_ctrl_guest_mask);
2524 	__clear_bit(idx, (unsigned long *)&cpuc->intel_ctrl_host_mask);
2525 	__clear_bit(idx, (unsigned long *)&cpuc->intel_cp_status);
2526 }
2527 
intel_pmu_disable_fixed(struct perf_event * event)2528 static void intel_pmu_disable_fixed(struct perf_event *event)
2529 {
2530 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2531 	struct hw_perf_event *hwc = &event->hw;
2532 	int idx = hwc->idx;
2533 	u64 mask;
2534 
2535 	if (is_topdown_idx(idx)) {
2536 		struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2537 
2538 		/*
2539 		 * When there are other active TopDown events,
2540 		 * don't disable the fixed counter 3.
2541 		 */
2542 		if (*(u64 *)cpuc->active_mask & INTEL_PMC_OTHER_TOPDOWN_BITS(idx))
2543 			return;
2544 		idx = INTEL_PMC_IDX_FIXED_SLOTS;
2545 	}
2546 
2547 	intel_clear_masks(event, idx);
2548 
2549 	mask = intel_fixed_bits_by_idx(idx - INTEL_PMC_IDX_FIXED, INTEL_FIXED_BITS_MASK);
2550 	cpuc->fixed_ctrl_val &= ~mask;
2551 }
2552 
intel_pmu_disable_event(struct perf_event * event)2553 static void intel_pmu_disable_event(struct perf_event *event)
2554 {
2555 	struct hw_perf_event *hwc = &event->hw;
2556 	int idx = hwc->idx;
2557 
2558 	switch (idx) {
2559 	case 0 ... INTEL_PMC_IDX_FIXED - 1:
2560 		intel_clear_masks(event, idx);
2561 		x86_pmu_disable_event(event);
2562 		break;
2563 	case INTEL_PMC_IDX_FIXED ... INTEL_PMC_IDX_FIXED_BTS - 1:
2564 	case INTEL_PMC_IDX_METRIC_BASE ... INTEL_PMC_IDX_METRIC_END:
2565 		intel_pmu_disable_fixed(event);
2566 		break;
2567 	case INTEL_PMC_IDX_FIXED_BTS:
2568 		intel_pmu_disable_bts();
2569 		intel_pmu_drain_bts_buffer();
2570 		return;
2571 	case INTEL_PMC_IDX_FIXED_VLBR:
2572 		intel_clear_masks(event, idx);
2573 		break;
2574 	default:
2575 		intel_clear_masks(event, idx);
2576 		pr_warn("Failed to disable the event with invalid index %d\n",
2577 			idx);
2578 		return;
2579 	}
2580 
2581 	/*
2582 	 * Needs to be called after x86_pmu_disable_event,
2583 	 * so we don't trigger the event without PEBS bit set.
2584 	 */
2585 	if (unlikely(event->attr.precise_ip))
2586 		intel_pmu_pebs_disable(event);
2587 }
2588 
intel_pmu_assign_event(struct perf_event * event,int idx)2589 static void intel_pmu_assign_event(struct perf_event *event, int idx)
2590 {
2591 	if (is_pebs_pt(event))
2592 		perf_report_aux_output_id(event, idx);
2593 }
2594 
intel_pmu_needs_branch_stack(struct perf_event * event)2595 static __always_inline bool intel_pmu_needs_branch_stack(struct perf_event *event)
2596 {
2597 	return event->hw.flags & PERF_X86_EVENT_NEEDS_BRANCH_STACK;
2598 }
2599 
intel_pmu_del_event(struct perf_event * event)2600 static void intel_pmu_del_event(struct perf_event *event)
2601 {
2602 	if (intel_pmu_needs_branch_stack(event))
2603 		intel_pmu_lbr_del(event);
2604 	if (event->attr.precise_ip)
2605 		intel_pmu_pebs_del(event);
2606 }
2607 
icl_set_topdown_event_period(struct perf_event * event)2608 static int icl_set_topdown_event_period(struct perf_event *event)
2609 {
2610 	struct hw_perf_event *hwc = &event->hw;
2611 	s64 left = local64_read(&hwc->period_left);
2612 
2613 	/*
2614 	 * The values in PERF_METRICS MSR are derived from fixed counter 3.
2615 	 * Software should start both registers, PERF_METRICS and fixed
2616 	 * counter 3, from zero.
2617 	 * Clear PERF_METRICS and Fixed counter 3 in initialization.
2618 	 * After that, both MSRs will be cleared for each read.
2619 	 * Don't need to clear them again.
2620 	 */
2621 	if (left == x86_pmu.max_period) {
2622 		wrmsrl(MSR_CORE_PERF_FIXED_CTR3, 0);
2623 		wrmsrl(MSR_PERF_METRICS, 0);
2624 		hwc->saved_slots = 0;
2625 		hwc->saved_metric = 0;
2626 	}
2627 
2628 	if ((hwc->saved_slots) && is_slots_event(event)) {
2629 		wrmsrl(MSR_CORE_PERF_FIXED_CTR3, hwc->saved_slots);
2630 		wrmsrl(MSR_PERF_METRICS, hwc->saved_metric);
2631 	}
2632 
2633 	perf_event_update_userpage(event);
2634 
2635 	return 0;
2636 }
2637 
2638 DEFINE_STATIC_CALL(intel_pmu_set_topdown_event_period, x86_perf_event_set_period);
2639 
icl_get_metrics_event_value(u64 metric,u64 slots,int idx)2640 static inline u64 icl_get_metrics_event_value(u64 metric, u64 slots, int idx)
2641 {
2642 	u32 val;
2643 
2644 	/*
2645 	 * The metric is reported as an 8bit integer fraction
2646 	 * summing up to 0xff.
2647 	 * slots-in-metric = (Metric / 0xff) * slots
2648 	 */
2649 	val = (metric >> ((idx - INTEL_PMC_IDX_METRIC_BASE) * 8)) & 0xff;
2650 	return  mul_u64_u32_div(slots, val, 0xff);
2651 }
2652 
icl_get_topdown_value(struct perf_event * event,u64 slots,u64 metrics)2653 static u64 icl_get_topdown_value(struct perf_event *event,
2654 				       u64 slots, u64 metrics)
2655 {
2656 	int idx = event->hw.idx;
2657 	u64 delta;
2658 
2659 	if (is_metric_idx(idx))
2660 		delta = icl_get_metrics_event_value(metrics, slots, idx);
2661 	else
2662 		delta = slots;
2663 
2664 	return delta;
2665 }
2666 
__icl_update_topdown_event(struct perf_event * event,u64 slots,u64 metrics,u64 last_slots,u64 last_metrics)2667 static void __icl_update_topdown_event(struct perf_event *event,
2668 				       u64 slots, u64 metrics,
2669 				       u64 last_slots, u64 last_metrics)
2670 {
2671 	u64 delta, last = 0;
2672 
2673 	delta = icl_get_topdown_value(event, slots, metrics);
2674 	if (last_slots)
2675 		last = icl_get_topdown_value(event, last_slots, last_metrics);
2676 
2677 	/*
2678 	 * The 8bit integer fraction of metric may be not accurate,
2679 	 * especially when the changes is very small.
2680 	 * For example, if only a few bad_spec happens, the fraction
2681 	 * may be reduced from 1 to 0. If so, the bad_spec event value
2682 	 * will be 0 which is definitely less than the last value.
2683 	 * Avoid update event->count for this case.
2684 	 */
2685 	if (delta > last) {
2686 		delta -= last;
2687 		local64_add(delta, &event->count);
2688 	}
2689 }
2690 
update_saved_topdown_regs(struct perf_event * event,u64 slots,u64 metrics,int metric_end)2691 static void update_saved_topdown_regs(struct perf_event *event, u64 slots,
2692 				      u64 metrics, int metric_end)
2693 {
2694 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2695 	struct perf_event *other;
2696 	int idx;
2697 
2698 	event->hw.saved_slots = slots;
2699 	event->hw.saved_metric = metrics;
2700 
2701 	for_each_set_bit(idx, cpuc->active_mask, metric_end + 1) {
2702 		if (!is_topdown_idx(idx))
2703 			continue;
2704 		other = cpuc->events[idx];
2705 		other->hw.saved_slots = slots;
2706 		other->hw.saved_metric = metrics;
2707 	}
2708 }
2709 
2710 /*
2711  * Update all active Topdown events.
2712  *
2713  * The PERF_METRICS and Fixed counter 3 are read separately. The values may be
2714  * modify by a NMI. PMU has to be disabled before calling this function.
2715  */
2716 
intel_update_topdown_event(struct perf_event * event,int metric_end)2717 static u64 intel_update_topdown_event(struct perf_event *event, int metric_end)
2718 {
2719 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2720 	struct perf_event *other;
2721 	u64 slots, metrics;
2722 	bool reset = true;
2723 	int idx;
2724 
2725 	/* read Fixed counter 3 */
2726 	rdpmcl((3 | INTEL_PMC_FIXED_RDPMC_BASE), slots);
2727 	if (!slots)
2728 		return 0;
2729 
2730 	/* read PERF_METRICS */
2731 	rdpmcl(INTEL_PMC_FIXED_RDPMC_METRICS, metrics);
2732 
2733 	for_each_set_bit(idx, cpuc->active_mask, metric_end + 1) {
2734 		if (!is_topdown_idx(idx))
2735 			continue;
2736 		other = cpuc->events[idx];
2737 		__icl_update_topdown_event(other, slots, metrics,
2738 					   event ? event->hw.saved_slots : 0,
2739 					   event ? event->hw.saved_metric : 0);
2740 	}
2741 
2742 	/*
2743 	 * Check and update this event, which may have been cleared
2744 	 * in active_mask e.g. x86_pmu_stop()
2745 	 */
2746 	if (event && !test_bit(event->hw.idx, cpuc->active_mask)) {
2747 		__icl_update_topdown_event(event, slots, metrics,
2748 					   event->hw.saved_slots,
2749 					   event->hw.saved_metric);
2750 
2751 		/*
2752 		 * In x86_pmu_stop(), the event is cleared in active_mask first,
2753 		 * then drain the delta, which indicates context switch for
2754 		 * counting.
2755 		 * Save metric and slots for context switch.
2756 		 * Don't need to reset the PERF_METRICS and Fixed counter 3.
2757 		 * Because the values will be restored in next schedule in.
2758 		 */
2759 		update_saved_topdown_regs(event, slots, metrics, metric_end);
2760 		reset = false;
2761 	}
2762 
2763 	if (reset) {
2764 		/* The fixed counter 3 has to be written before the PERF_METRICS. */
2765 		wrmsrl(MSR_CORE_PERF_FIXED_CTR3, 0);
2766 		wrmsrl(MSR_PERF_METRICS, 0);
2767 		if (event)
2768 			update_saved_topdown_regs(event, 0, 0, metric_end);
2769 	}
2770 
2771 	return slots;
2772 }
2773 
icl_update_topdown_event(struct perf_event * event)2774 static u64 icl_update_topdown_event(struct perf_event *event)
2775 {
2776 	return intel_update_topdown_event(event, INTEL_PMC_IDX_METRIC_BASE +
2777 						 x86_pmu.num_topdown_events - 1);
2778 }
2779 
2780 DEFINE_STATIC_CALL(intel_pmu_update_topdown_event, x86_perf_event_update);
2781 
intel_pmu_read_event(struct perf_event * event)2782 static void intel_pmu_read_event(struct perf_event *event)
2783 {
2784 	if (event->hw.flags & (PERF_X86_EVENT_AUTO_RELOAD | PERF_X86_EVENT_TOPDOWN)) {
2785 		struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2786 		bool pmu_enabled = cpuc->enabled;
2787 
2788 		/* Only need to call update_topdown_event() once for group read. */
2789 		if (is_metric_event(event) && (cpuc->txn_flags & PERF_PMU_TXN_READ))
2790 			return;
2791 
2792 		cpuc->enabled = 0;
2793 		if (pmu_enabled)
2794 			intel_pmu_disable_all();
2795 
2796 		if (is_topdown_event(event))
2797 			static_call(intel_pmu_update_topdown_event)(event);
2798 		else
2799 			intel_pmu_drain_pebs_buffer();
2800 
2801 		cpuc->enabled = pmu_enabled;
2802 		if (pmu_enabled)
2803 			intel_pmu_enable_all(0);
2804 
2805 		return;
2806 	}
2807 
2808 	x86_perf_event_update(event);
2809 }
2810 
intel_pmu_enable_fixed(struct perf_event * event)2811 static void intel_pmu_enable_fixed(struct perf_event *event)
2812 {
2813 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2814 	struct hw_perf_event *hwc = &event->hw;
2815 	u64 mask, bits = 0;
2816 	int idx = hwc->idx;
2817 
2818 	if (is_topdown_idx(idx)) {
2819 		struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2820 		/*
2821 		 * When there are other active TopDown events,
2822 		 * don't enable the fixed counter 3 again.
2823 		 */
2824 		if (*(u64 *)cpuc->active_mask & INTEL_PMC_OTHER_TOPDOWN_BITS(idx))
2825 			return;
2826 
2827 		idx = INTEL_PMC_IDX_FIXED_SLOTS;
2828 
2829 		if (event->attr.config1 & INTEL_TD_CFG_METRIC_CLEAR)
2830 			bits |= INTEL_FIXED_3_METRICS_CLEAR;
2831 	}
2832 
2833 	intel_set_masks(event, idx);
2834 
2835 	/*
2836 	 * Enable IRQ generation (0x8), if not PEBS,
2837 	 * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
2838 	 * if requested:
2839 	 */
2840 	if (!event->attr.precise_ip)
2841 		bits |= INTEL_FIXED_0_ENABLE_PMI;
2842 	if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
2843 		bits |= INTEL_FIXED_0_USER;
2844 	if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
2845 		bits |= INTEL_FIXED_0_KERNEL;
2846 
2847 	/*
2848 	 * ANY bit is supported in v3 and up
2849 	 */
2850 	if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY)
2851 		bits |= INTEL_FIXED_0_ANYTHREAD;
2852 
2853 	idx -= INTEL_PMC_IDX_FIXED;
2854 	bits = intel_fixed_bits_by_idx(idx, bits);
2855 	mask = intel_fixed_bits_by_idx(idx, INTEL_FIXED_BITS_MASK);
2856 
2857 	if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip) {
2858 		bits |= intel_fixed_bits_by_idx(idx, ICL_FIXED_0_ADAPTIVE);
2859 		mask |= intel_fixed_bits_by_idx(idx, ICL_FIXED_0_ADAPTIVE);
2860 	}
2861 
2862 	cpuc->fixed_ctrl_val &= ~mask;
2863 	cpuc->fixed_ctrl_val |= bits;
2864 }
2865 
intel_pmu_enable_event(struct perf_event * event)2866 static void intel_pmu_enable_event(struct perf_event *event)
2867 {
2868 	u64 enable_mask = ARCH_PERFMON_EVENTSEL_ENABLE;
2869 	struct hw_perf_event *hwc = &event->hw;
2870 	int idx = hwc->idx;
2871 
2872 	if (unlikely(event->attr.precise_ip))
2873 		intel_pmu_pebs_enable(event);
2874 
2875 	switch (idx) {
2876 	case 0 ... INTEL_PMC_IDX_FIXED - 1:
2877 		if (branch_sample_counters(event))
2878 			enable_mask |= ARCH_PERFMON_EVENTSEL_BR_CNTR;
2879 		intel_set_masks(event, idx);
2880 		__x86_pmu_enable_event(hwc, enable_mask);
2881 		break;
2882 	case INTEL_PMC_IDX_FIXED ... INTEL_PMC_IDX_FIXED_BTS - 1:
2883 	case INTEL_PMC_IDX_METRIC_BASE ... INTEL_PMC_IDX_METRIC_END:
2884 		intel_pmu_enable_fixed(event);
2885 		break;
2886 	case INTEL_PMC_IDX_FIXED_BTS:
2887 		if (!__this_cpu_read(cpu_hw_events.enabled))
2888 			return;
2889 		intel_pmu_enable_bts(hwc->config);
2890 		break;
2891 	case INTEL_PMC_IDX_FIXED_VLBR:
2892 		intel_set_masks(event, idx);
2893 		break;
2894 	default:
2895 		pr_warn("Failed to enable the event with invalid index %d\n",
2896 			idx);
2897 	}
2898 }
2899 
intel_pmu_add_event(struct perf_event * event)2900 static void intel_pmu_add_event(struct perf_event *event)
2901 {
2902 	if (event->attr.precise_ip)
2903 		intel_pmu_pebs_add(event);
2904 	if (intel_pmu_needs_branch_stack(event))
2905 		intel_pmu_lbr_add(event);
2906 }
2907 
2908 /*
2909  * Save and restart an expired event. Called by NMI contexts,
2910  * so it has to be careful about preempting normal event ops:
2911  */
intel_pmu_save_and_restart(struct perf_event * event)2912 int intel_pmu_save_and_restart(struct perf_event *event)
2913 {
2914 	static_call(x86_pmu_update)(event);
2915 	/*
2916 	 * For a checkpointed counter always reset back to 0.  This
2917 	 * avoids a situation where the counter overflows, aborts the
2918 	 * transaction and is then set back to shortly before the
2919 	 * overflow, and overflows and aborts again.
2920 	 */
2921 	if (unlikely(event_is_checkpointed(event))) {
2922 		/* No race with NMIs because the counter should not be armed */
2923 		wrmsrl(event->hw.event_base, 0);
2924 		local64_set(&event->hw.prev_count, 0);
2925 	}
2926 	return static_call(x86_pmu_set_period)(event);
2927 }
2928 
intel_pmu_set_period(struct perf_event * event)2929 static int intel_pmu_set_period(struct perf_event *event)
2930 {
2931 	if (unlikely(is_topdown_count(event)))
2932 		return static_call(intel_pmu_set_topdown_event_period)(event);
2933 
2934 	return x86_perf_event_set_period(event);
2935 }
2936 
intel_pmu_update(struct perf_event * event)2937 static u64 intel_pmu_update(struct perf_event *event)
2938 {
2939 	if (unlikely(is_topdown_count(event)))
2940 		return static_call(intel_pmu_update_topdown_event)(event);
2941 
2942 	return x86_perf_event_update(event);
2943 }
2944 
intel_pmu_reset(void)2945 static void intel_pmu_reset(void)
2946 {
2947 	struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
2948 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2949 	unsigned long *cntr_mask = hybrid(cpuc->pmu, cntr_mask);
2950 	unsigned long *fixed_cntr_mask = hybrid(cpuc->pmu, fixed_cntr_mask);
2951 	unsigned long flags;
2952 	int idx;
2953 
2954 	if (!*(u64 *)cntr_mask)
2955 		return;
2956 
2957 	local_irq_save(flags);
2958 
2959 	pr_info("clearing PMU state on CPU#%d\n", smp_processor_id());
2960 
2961 	for_each_set_bit(idx, cntr_mask, INTEL_PMC_MAX_GENERIC) {
2962 		wrmsrl_safe(x86_pmu_config_addr(idx), 0ull);
2963 		wrmsrl_safe(x86_pmu_event_addr(idx),  0ull);
2964 	}
2965 	for_each_set_bit(idx, fixed_cntr_mask, INTEL_PMC_MAX_FIXED) {
2966 		if (fixed_counter_disabled(idx, cpuc->pmu))
2967 			continue;
2968 		wrmsrl_safe(x86_pmu_fixed_ctr_addr(idx), 0ull);
2969 	}
2970 
2971 	if (ds)
2972 		ds->bts_index = ds->bts_buffer_base;
2973 
2974 	/* Ack all overflows and disable fixed counters */
2975 	if (x86_pmu.version >= 2) {
2976 		intel_pmu_ack_status(intel_pmu_get_status());
2977 		wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
2978 	}
2979 
2980 	/* Reset LBRs and LBR freezing */
2981 	if (x86_pmu.lbr_nr) {
2982 		update_debugctlmsr(get_debugctlmsr() &
2983 			~(DEBUGCTLMSR_FREEZE_LBRS_ON_PMI|DEBUGCTLMSR_LBR));
2984 	}
2985 
2986 	local_irq_restore(flags);
2987 }
2988 
2989 /*
2990  * We may be running with guest PEBS events created by KVM, and the
2991  * PEBS records are logged into the guest's DS and invisible to host.
2992  *
2993  * In the case of guest PEBS overflow, we only trigger a fake event
2994  * to emulate the PEBS overflow PMI for guest PEBS counters in KVM.
2995  * The guest will then vm-entry and check the guest DS area to read
2996  * the guest PEBS records.
2997  *
2998  * The contents and other behavior of the guest event do not matter.
2999  */
x86_pmu_handle_guest_pebs(struct pt_regs * regs,struct perf_sample_data * data)3000 static void x86_pmu_handle_guest_pebs(struct pt_regs *regs,
3001 				      struct perf_sample_data *data)
3002 {
3003 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3004 	u64 guest_pebs_idxs = cpuc->pebs_enabled & ~cpuc->intel_ctrl_host_mask;
3005 	struct perf_event *event = NULL;
3006 	int bit;
3007 
3008 	if (!unlikely(perf_guest_state()))
3009 		return;
3010 
3011 	if (!x86_pmu.pebs_ept || !x86_pmu.pebs_active ||
3012 	    !guest_pebs_idxs)
3013 		return;
3014 
3015 	for_each_set_bit(bit, (unsigned long *)&guest_pebs_idxs, X86_PMC_IDX_MAX) {
3016 		event = cpuc->events[bit];
3017 		if (!event->attr.precise_ip)
3018 			continue;
3019 
3020 		perf_sample_data_init(data, 0, event->hw.last_period);
3021 		if (perf_event_overflow(event, data, regs))
3022 			x86_pmu_stop(event, 0);
3023 
3024 		/* Inject one fake event is enough. */
3025 		break;
3026 	}
3027 }
3028 
handle_pmi_common(struct pt_regs * regs,u64 status)3029 static int handle_pmi_common(struct pt_regs *regs, u64 status)
3030 {
3031 	struct perf_sample_data data;
3032 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3033 	int bit;
3034 	int handled = 0;
3035 	u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl);
3036 
3037 	inc_irq_stat(apic_perf_irqs);
3038 
3039 	/*
3040 	 * Ignore a range of extra bits in status that do not indicate
3041 	 * overflow by themselves.
3042 	 */
3043 	status &= ~(GLOBAL_STATUS_COND_CHG |
3044 		    GLOBAL_STATUS_ASIF |
3045 		    GLOBAL_STATUS_LBRS_FROZEN);
3046 	if (!status)
3047 		return 0;
3048 	/*
3049 	 * In case multiple PEBS events are sampled at the same time,
3050 	 * it is possible to have GLOBAL_STATUS bit 62 set indicating
3051 	 * PEBS buffer overflow and also seeing at most 3 PEBS counters
3052 	 * having their bits set in the status register. This is a sign
3053 	 * that there was at least one PEBS record pending at the time
3054 	 * of the PMU interrupt. PEBS counters must only be processed
3055 	 * via the drain_pebs() calls and not via the regular sample
3056 	 * processing loop coming after that the function, otherwise
3057 	 * phony regular samples may be generated in the sampling buffer
3058 	 * not marked with the EXACT tag. Another possibility is to have
3059 	 * one PEBS event and at least one non-PEBS event which overflows
3060 	 * while PEBS has armed. In this case, bit 62 of GLOBAL_STATUS will
3061 	 * not be set, yet the overflow status bit for the PEBS counter will
3062 	 * be on Skylake.
3063 	 *
3064 	 * To avoid this problem, we systematically ignore the PEBS-enabled
3065 	 * counters from the GLOBAL_STATUS mask and we always process PEBS
3066 	 * events via drain_pebs().
3067 	 */
3068 	status &= ~(cpuc->pebs_enabled & x86_pmu.pebs_capable);
3069 
3070 	/*
3071 	 * PEBS overflow sets bit 62 in the global status register
3072 	 */
3073 	if (__test_and_clear_bit(GLOBAL_STATUS_BUFFER_OVF_BIT, (unsigned long *)&status)) {
3074 		u64 pebs_enabled = cpuc->pebs_enabled;
3075 
3076 		handled++;
3077 		x86_pmu_handle_guest_pebs(regs, &data);
3078 		static_call(x86_pmu_drain_pebs)(regs, &data);
3079 		status &= intel_ctrl | GLOBAL_STATUS_TRACE_TOPAPMI;
3080 
3081 		/*
3082 		 * PMI throttle may be triggered, which stops the PEBS event.
3083 		 * Although cpuc->pebs_enabled is updated accordingly, the
3084 		 * MSR_IA32_PEBS_ENABLE is not updated. Because the
3085 		 * cpuc->enabled has been forced to 0 in PMI.
3086 		 * Update the MSR if pebs_enabled is changed.
3087 		 */
3088 		if (pebs_enabled != cpuc->pebs_enabled)
3089 			wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
3090 	}
3091 
3092 	/*
3093 	 * Intel PT
3094 	 */
3095 	if (__test_and_clear_bit(GLOBAL_STATUS_TRACE_TOPAPMI_BIT, (unsigned long *)&status)) {
3096 		handled++;
3097 		if (!perf_guest_handle_intel_pt_intr())
3098 			intel_pt_interrupt();
3099 	}
3100 
3101 	/*
3102 	 * Intel Perf metrics
3103 	 */
3104 	if (__test_and_clear_bit(GLOBAL_STATUS_PERF_METRICS_OVF_BIT, (unsigned long *)&status)) {
3105 		handled++;
3106 		static_call(intel_pmu_update_topdown_event)(NULL);
3107 	}
3108 
3109 	/*
3110 	 * Checkpointed counters can lead to 'spurious' PMIs because the
3111 	 * rollback caused by the PMI will have cleared the overflow status
3112 	 * bit. Therefore always force probe these counters.
3113 	 */
3114 	status |= cpuc->intel_cp_status;
3115 
3116 	for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
3117 		struct perf_event *event = cpuc->events[bit];
3118 
3119 		handled++;
3120 
3121 		if (!test_bit(bit, cpuc->active_mask))
3122 			continue;
3123 
3124 		if (!intel_pmu_save_and_restart(event))
3125 			continue;
3126 
3127 		perf_sample_data_init(&data, 0, event->hw.last_period);
3128 
3129 		if (has_branch_stack(event))
3130 			intel_pmu_lbr_save_brstack(&data, cpuc, event);
3131 
3132 		if (perf_event_overflow(event, &data, regs))
3133 			x86_pmu_stop(event, 0);
3134 	}
3135 
3136 	return handled;
3137 }
3138 
3139 /*
3140  * This handler is triggered by the local APIC, so the APIC IRQ handling
3141  * rules apply:
3142  */
intel_pmu_handle_irq(struct pt_regs * regs)3143 static int intel_pmu_handle_irq(struct pt_regs *regs)
3144 {
3145 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3146 	bool late_ack = hybrid_bit(cpuc->pmu, late_ack);
3147 	bool mid_ack = hybrid_bit(cpuc->pmu, mid_ack);
3148 	int loops;
3149 	u64 status;
3150 	int handled;
3151 	int pmu_enabled;
3152 
3153 	/*
3154 	 * Save the PMU state.
3155 	 * It needs to be restored when leaving the handler.
3156 	 */
3157 	pmu_enabled = cpuc->enabled;
3158 	/*
3159 	 * In general, the early ACK is only applied for old platforms.
3160 	 * For the big core starts from Haswell, the late ACK should be
3161 	 * applied.
3162 	 * For the small core after Tremont, we have to do the ACK right
3163 	 * before re-enabling counters, which is in the middle of the
3164 	 * NMI handler.
3165 	 */
3166 	if (!late_ack && !mid_ack)
3167 		apic_write(APIC_LVTPC, APIC_DM_NMI);
3168 	intel_bts_disable_local();
3169 	cpuc->enabled = 0;
3170 	__intel_pmu_disable_all(true);
3171 	handled = intel_pmu_drain_bts_buffer();
3172 	handled += intel_bts_interrupt();
3173 	status = intel_pmu_get_status();
3174 	if (!status)
3175 		goto done;
3176 
3177 	loops = 0;
3178 again:
3179 	intel_pmu_lbr_read();
3180 	intel_pmu_ack_status(status);
3181 	if (++loops > 100) {
3182 		static bool warned;
3183 
3184 		if (!warned) {
3185 			WARN(1, "perfevents: irq loop stuck!\n");
3186 			perf_event_print_debug();
3187 			warned = true;
3188 		}
3189 		intel_pmu_reset();
3190 		goto done;
3191 	}
3192 
3193 	handled += handle_pmi_common(regs, status);
3194 
3195 	/*
3196 	 * Repeat if there is more work to be done:
3197 	 */
3198 	status = intel_pmu_get_status();
3199 	if (status)
3200 		goto again;
3201 
3202 done:
3203 	if (mid_ack)
3204 		apic_write(APIC_LVTPC, APIC_DM_NMI);
3205 	/* Only restore PMU state when it's active. See x86_pmu_disable(). */
3206 	cpuc->enabled = pmu_enabled;
3207 	if (pmu_enabled)
3208 		__intel_pmu_enable_all(0, true);
3209 	intel_bts_enable_local();
3210 
3211 	/*
3212 	 * Only unmask the NMI after the overflow counters
3213 	 * have been reset. This avoids spurious NMIs on
3214 	 * Haswell CPUs.
3215 	 */
3216 	if (late_ack)
3217 		apic_write(APIC_LVTPC, APIC_DM_NMI);
3218 	return handled;
3219 }
3220 
3221 static struct event_constraint *
intel_bts_constraints(struct perf_event * event)3222 intel_bts_constraints(struct perf_event *event)
3223 {
3224 	if (unlikely(intel_pmu_has_bts(event)))
3225 		return &bts_constraint;
3226 
3227 	return NULL;
3228 }
3229 
3230 /*
3231  * Note: matches a fake event, like Fixed2.
3232  */
3233 static struct event_constraint *
intel_vlbr_constraints(struct perf_event * event)3234 intel_vlbr_constraints(struct perf_event *event)
3235 {
3236 	struct event_constraint *c = &vlbr_constraint;
3237 
3238 	if (unlikely(constraint_match(c, event->hw.config))) {
3239 		event->hw.flags |= c->flags;
3240 		return c;
3241 	}
3242 
3243 	return NULL;
3244 }
3245 
intel_alt_er(struct cpu_hw_events * cpuc,int idx,u64 config)3246 static int intel_alt_er(struct cpu_hw_events *cpuc,
3247 			int idx, u64 config)
3248 {
3249 	struct extra_reg *extra_regs = hybrid(cpuc->pmu, extra_regs);
3250 	int alt_idx = idx;
3251 
3252 	if (!(x86_pmu.flags & PMU_FL_HAS_RSP_1))
3253 		return idx;
3254 
3255 	if (idx == EXTRA_REG_RSP_0)
3256 		alt_idx = EXTRA_REG_RSP_1;
3257 
3258 	if (idx == EXTRA_REG_RSP_1)
3259 		alt_idx = EXTRA_REG_RSP_0;
3260 
3261 	if (config & ~extra_regs[alt_idx].valid_mask)
3262 		return idx;
3263 
3264 	return alt_idx;
3265 }
3266 
intel_fixup_er(struct perf_event * event,int idx)3267 static void intel_fixup_er(struct perf_event *event, int idx)
3268 {
3269 	struct extra_reg *extra_regs = hybrid(event->pmu, extra_regs);
3270 	event->hw.extra_reg.idx = idx;
3271 
3272 	if (idx == EXTRA_REG_RSP_0) {
3273 		event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
3274 		event->hw.config |= extra_regs[EXTRA_REG_RSP_0].event;
3275 		event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0;
3276 	} else if (idx == EXTRA_REG_RSP_1) {
3277 		event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
3278 		event->hw.config |= extra_regs[EXTRA_REG_RSP_1].event;
3279 		event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1;
3280 	}
3281 }
3282 
3283 /*
3284  * manage allocation of shared extra msr for certain events
3285  *
3286  * sharing can be:
3287  * per-cpu: to be shared between the various events on a single PMU
3288  * per-core: per-cpu + shared by HT threads
3289  */
3290 static struct event_constraint *
__intel_shared_reg_get_constraints(struct cpu_hw_events * cpuc,struct perf_event * event,struct hw_perf_event_extra * reg)3291 __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc,
3292 				   struct perf_event *event,
3293 				   struct hw_perf_event_extra *reg)
3294 {
3295 	struct event_constraint *c = &emptyconstraint;
3296 	struct er_account *era;
3297 	unsigned long flags;
3298 	int idx = reg->idx;
3299 
3300 	/*
3301 	 * reg->alloc can be set due to existing state, so for fake cpuc we
3302 	 * need to ignore this, otherwise we might fail to allocate proper fake
3303 	 * state for this extra reg constraint. Also see the comment below.
3304 	 */
3305 	if (reg->alloc && !cpuc->is_fake)
3306 		return NULL; /* call x86_get_event_constraint() */
3307 
3308 again:
3309 	era = &cpuc->shared_regs->regs[idx];
3310 	/*
3311 	 * we use spin_lock_irqsave() to avoid lockdep issues when
3312 	 * passing a fake cpuc
3313 	 */
3314 	raw_spin_lock_irqsave(&era->lock, flags);
3315 
3316 	if (!atomic_read(&era->ref) || era->config == reg->config) {
3317 
3318 		/*
3319 		 * If its a fake cpuc -- as per validate_{group,event}() we
3320 		 * shouldn't touch event state and we can avoid doing so
3321 		 * since both will only call get_event_constraints() once
3322 		 * on each event, this avoids the need for reg->alloc.
3323 		 *
3324 		 * Not doing the ER fixup will only result in era->reg being
3325 		 * wrong, but since we won't actually try and program hardware
3326 		 * this isn't a problem either.
3327 		 */
3328 		if (!cpuc->is_fake) {
3329 			if (idx != reg->idx)
3330 				intel_fixup_er(event, idx);
3331 
3332 			/*
3333 			 * x86_schedule_events() can call get_event_constraints()
3334 			 * multiple times on events in the case of incremental
3335 			 * scheduling(). reg->alloc ensures we only do the ER
3336 			 * allocation once.
3337 			 */
3338 			reg->alloc = 1;
3339 		}
3340 
3341 		/* lock in msr value */
3342 		era->config = reg->config;
3343 		era->reg = reg->reg;
3344 
3345 		/* one more user */
3346 		atomic_inc(&era->ref);
3347 
3348 		/*
3349 		 * need to call x86_get_event_constraint()
3350 		 * to check if associated event has constraints
3351 		 */
3352 		c = NULL;
3353 	} else {
3354 		idx = intel_alt_er(cpuc, idx, reg->config);
3355 		if (idx != reg->idx) {
3356 			raw_spin_unlock_irqrestore(&era->lock, flags);
3357 			goto again;
3358 		}
3359 	}
3360 	raw_spin_unlock_irqrestore(&era->lock, flags);
3361 
3362 	return c;
3363 }
3364 
3365 static void
__intel_shared_reg_put_constraints(struct cpu_hw_events * cpuc,struct hw_perf_event_extra * reg)3366 __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc,
3367 				   struct hw_perf_event_extra *reg)
3368 {
3369 	struct er_account *era;
3370 
3371 	/*
3372 	 * Only put constraint if extra reg was actually allocated. Also takes
3373 	 * care of event which do not use an extra shared reg.
3374 	 *
3375 	 * Also, if this is a fake cpuc we shouldn't touch any event state
3376 	 * (reg->alloc) and we don't care about leaving inconsistent cpuc state
3377 	 * either since it'll be thrown out.
3378 	 */
3379 	if (!reg->alloc || cpuc->is_fake)
3380 		return;
3381 
3382 	era = &cpuc->shared_regs->regs[reg->idx];
3383 
3384 	/* one fewer user */
3385 	atomic_dec(&era->ref);
3386 
3387 	/* allocate again next time */
3388 	reg->alloc = 0;
3389 }
3390 
3391 static struct event_constraint *
intel_shared_regs_constraints(struct cpu_hw_events * cpuc,struct perf_event * event)3392 intel_shared_regs_constraints(struct cpu_hw_events *cpuc,
3393 			      struct perf_event *event)
3394 {
3395 	struct event_constraint *c = NULL, *d;
3396 	struct hw_perf_event_extra *xreg, *breg;
3397 
3398 	xreg = &event->hw.extra_reg;
3399 	if (xreg->idx != EXTRA_REG_NONE) {
3400 		c = __intel_shared_reg_get_constraints(cpuc, event, xreg);
3401 		if (c == &emptyconstraint)
3402 			return c;
3403 	}
3404 	breg = &event->hw.branch_reg;
3405 	if (breg->idx != EXTRA_REG_NONE) {
3406 		d = __intel_shared_reg_get_constraints(cpuc, event, breg);
3407 		if (d == &emptyconstraint) {
3408 			__intel_shared_reg_put_constraints(cpuc, xreg);
3409 			c = d;
3410 		}
3411 	}
3412 	return c;
3413 }
3414 
3415 struct event_constraint *
x86_get_event_constraints(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)3416 x86_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3417 			  struct perf_event *event)
3418 {
3419 	struct event_constraint *event_constraints = hybrid(cpuc->pmu, event_constraints);
3420 	struct event_constraint *c;
3421 
3422 	if (event_constraints) {
3423 		for_each_event_constraint(c, event_constraints) {
3424 			if (constraint_match(c, event->hw.config)) {
3425 				event->hw.flags |= c->flags;
3426 				return c;
3427 			}
3428 		}
3429 	}
3430 
3431 	return &hybrid_var(cpuc->pmu, unconstrained);
3432 }
3433 
3434 static struct event_constraint *
__intel_get_event_constraints(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)3435 __intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3436 			    struct perf_event *event)
3437 {
3438 	struct event_constraint *c;
3439 
3440 	c = intel_vlbr_constraints(event);
3441 	if (c)
3442 		return c;
3443 
3444 	c = intel_bts_constraints(event);
3445 	if (c)
3446 		return c;
3447 
3448 	c = intel_shared_regs_constraints(cpuc, event);
3449 	if (c)
3450 		return c;
3451 
3452 	c = intel_pebs_constraints(event);
3453 	if (c)
3454 		return c;
3455 
3456 	return x86_get_event_constraints(cpuc, idx, event);
3457 }
3458 
3459 static void
intel_start_scheduling(struct cpu_hw_events * cpuc)3460 intel_start_scheduling(struct cpu_hw_events *cpuc)
3461 {
3462 	struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
3463 	struct intel_excl_states *xl;
3464 	int tid = cpuc->excl_thread_id;
3465 
3466 	/*
3467 	 * nothing needed if in group validation mode
3468 	 */
3469 	if (cpuc->is_fake || !is_ht_workaround_enabled())
3470 		return;
3471 
3472 	/*
3473 	 * no exclusion needed
3474 	 */
3475 	if (WARN_ON_ONCE(!excl_cntrs))
3476 		return;
3477 
3478 	xl = &excl_cntrs->states[tid];
3479 
3480 	xl->sched_started = true;
3481 	/*
3482 	 * lock shared state until we are done scheduling
3483 	 * in stop_event_scheduling()
3484 	 * makes scheduling appear as a transaction
3485 	 */
3486 	raw_spin_lock(&excl_cntrs->lock);
3487 }
3488 
intel_commit_scheduling(struct cpu_hw_events * cpuc,int idx,int cntr)3489 static void intel_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr)
3490 {
3491 	struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
3492 	struct event_constraint *c = cpuc->event_constraint[idx];
3493 	struct intel_excl_states *xl;
3494 	int tid = cpuc->excl_thread_id;
3495 
3496 	if (cpuc->is_fake || !is_ht_workaround_enabled())
3497 		return;
3498 
3499 	if (WARN_ON_ONCE(!excl_cntrs))
3500 		return;
3501 
3502 	if (!(c->flags & PERF_X86_EVENT_DYNAMIC))
3503 		return;
3504 
3505 	xl = &excl_cntrs->states[tid];
3506 
3507 	lockdep_assert_held(&excl_cntrs->lock);
3508 
3509 	if (c->flags & PERF_X86_EVENT_EXCL)
3510 		xl->state[cntr] = INTEL_EXCL_EXCLUSIVE;
3511 	else
3512 		xl->state[cntr] = INTEL_EXCL_SHARED;
3513 }
3514 
3515 static void
intel_stop_scheduling(struct cpu_hw_events * cpuc)3516 intel_stop_scheduling(struct cpu_hw_events *cpuc)
3517 {
3518 	struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
3519 	struct intel_excl_states *xl;
3520 	int tid = cpuc->excl_thread_id;
3521 
3522 	/*
3523 	 * nothing needed if in group validation mode
3524 	 */
3525 	if (cpuc->is_fake || !is_ht_workaround_enabled())
3526 		return;
3527 	/*
3528 	 * no exclusion needed
3529 	 */
3530 	if (WARN_ON_ONCE(!excl_cntrs))
3531 		return;
3532 
3533 	xl = &excl_cntrs->states[tid];
3534 
3535 	xl->sched_started = false;
3536 	/*
3537 	 * release shared state lock (acquired in intel_start_scheduling())
3538 	 */
3539 	raw_spin_unlock(&excl_cntrs->lock);
3540 }
3541 
3542 static struct event_constraint *
dyn_constraint(struct cpu_hw_events * cpuc,struct event_constraint * c,int idx)3543 dyn_constraint(struct cpu_hw_events *cpuc, struct event_constraint *c, int idx)
3544 {
3545 	WARN_ON_ONCE(!cpuc->constraint_list);
3546 
3547 	if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) {
3548 		struct event_constraint *cx;
3549 
3550 		/*
3551 		 * grab pre-allocated constraint entry
3552 		 */
3553 		cx = &cpuc->constraint_list[idx];
3554 
3555 		/*
3556 		 * initialize dynamic constraint
3557 		 * with static constraint
3558 		 */
3559 		*cx = *c;
3560 
3561 		/*
3562 		 * mark constraint as dynamic
3563 		 */
3564 		cx->flags |= PERF_X86_EVENT_DYNAMIC;
3565 		c = cx;
3566 	}
3567 
3568 	return c;
3569 }
3570 
3571 static struct event_constraint *
intel_get_excl_constraints(struct cpu_hw_events * cpuc,struct perf_event * event,int idx,struct event_constraint * c)3572 intel_get_excl_constraints(struct cpu_hw_events *cpuc, struct perf_event *event,
3573 			   int idx, struct event_constraint *c)
3574 {
3575 	struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
3576 	struct intel_excl_states *xlo;
3577 	int tid = cpuc->excl_thread_id;
3578 	int is_excl, i, w;
3579 
3580 	/*
3581 	 * validating a group does not require
3582 	 * enforcing cross-thread  exclusion
3583 	 */
3584 	if (cpuc->is_fake || !is_ht_workaround_enabled())
3585 		return c;
3586 
3587 	/*
3588 	 * no exclusion needed
3589 	 */
3590 	if (WARN_ON_ONCE(!excl_cntrs))
3591 		return c;
3592 
3593 	/*
3594 	 * because we modify the constraint, we need
3595 	 * to make a copy. Static constraints come
3596 	 * from static const tables.
3597 	 *
3598 	 * only needed when constraint has not yet
3599 	 * been cloned (marked dynamic)
3600 	 */
3601 	c = dyn_constraint(cpuc, c, idx);
3602 
3603 	/*
3604 	 * From here on, the constraint is dynamic.
3605 	 * Either it was just allocated above, or it
3606 	 * was allocated during a earlier invocation
3607 	 * of this function
3608 	 */
3609 
3610 	/*
3611 	 * state of sibling HT
3612 	 */
3613 	xlo = &excl_cntrs->states[tid ^ 1];
3614 
3615 	/*
3616 	 * event requires exclusive counter access
3617 	 * across HT threads
3618 	 */
3619 	is_excl = c->flags & PERF_X86_EVENT_EXCL;
3620 	if (is_excl && !(event->hw.flags & PERF_X86_EVENT_EXCL_ACCT)) {
3621 		event->hw.flags |= PERF_X86_EVENT_EXCL_ACCT;
3622 		if (!cpuc->n_excl++)
3623 			WRITE_ONCE(excl_cntrs->has_exclusive[tid], 1);
3624 	}
3625 
3626 	/*
3627 	 * Modify static constraint with current dynamic
3628 	 * state of thread
3629 	 *
3630 	 * EXCLUSIVE: sibling counter measuring exclusive event
3631 	 * SHARED   : sibling counter measuring non-exclusive event
3632 	 * UNUSED   : sibling counter unused
3633 	 */
3634 	w = c->weight;
3635 	for_each_set_bit(i, c->idxmsk, X86_PMC_IDX_MAX) {
3636 		/*
3637 		 * exclusive event in sibling counter
3638 		 * our corresponding counter cannot be used
3639 		 * regardless of our event
3640 		 */
3641 		if (xlo->state[i] == INTEL_EXCL_EXCLUSIVE) {
3642 			__clear_bit(i, c->idxmsk);
3643 			w--;
3644 			continue;
3645 		}
3646 		/*
3647 		 * if measuring an exclusive event, sibling
3648 		 * measuring non-exclusive, then counter cannot
3649 		 * be used
3650 		 */
3651 		if (is_excl && xlo->state[i] == INTEL_EXCL_SHARED) {
3652 			__clear_bit(i, c->idxmsk);
3653 			w--;
3654 			continue;
3655 		}
3656 	}
3657 
3658 	/*
3659 	 * if we return an empty mask, then switch
3660 	 * back to static empty constraint to avoid
3661 	 * the cost of freeing later on
3662 	 */
3663 	if (!w)
3664 		c = &emptyconstraint;
3665 
3666 	c->weight = w;
3667 
3668 	return c;
3669 }
3670 
3671 static struct event_constraint *
intel_get_event_constraints(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)3672 intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
3673 			    struct perf_event *event)
3674 {
3675 	struct event_constraint *c1, *c2;
3676 
3677 	c1 = cpuc->event_constraint[idx];
3678 
3679 	/*
3680 	 * first time only
3681 	 * - static constraint: no change across incremental scheduling calls
3682 	 * - dynamic constraint: handled by intel_get_excl_constraints()
3683 	 */
3684 	c2 = __intel_get_event_constraints(cpuc, idx, event);
3685 	if (c1) {
3686 	        WARN_ON_ONCE(!(c1->flags & PERF_X86_EVENT_DYNAMIC));
3687 		bitmap_copy(c1->idxmsk, c2->idxmsk, X86_PMC_IDX_MAX);
3688 		c1->weight = c2->weight;
3689 		c2 = c1;
3690 	}
3691 
3692 	if (cpuc->excl_cntrs)
3693 		return intel_get_excl_constraints(cpuc, event, idx, c2);
3694 
3695 	/* Not all counters support the branch counter feature. */
3696 	if (branch_sample_counters(event)) {
3697 		c2 = dyn_constraint(cpuc, c2, idx);
3698 		c2->idxmsk64 &= x86_pmu.lbr_counters;
3699 		c2->weight = hweight64(c2->idxmsk64);
3700 	}
3701 
3702 	return c2;
3703 }
3704 
intel_put_excl_constraints(struct cpu_hw_events * cpuc,struct perf_event * event)3705 static void intel_put_excl_constraints(struct cpu_hw_events *cpuc,
3706 		struct perf_event *event)
3707 {
3708 	struct hw_perf_event *hwc = &event->hw;
3709 	struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs;
3710 	int tid = cpuc->excl_thread_id;
3711 	struct intel_excl_states *xl;
3712 
3713 	/*
3714 	 * nothing needed if in group validation mode
3715 	 */
3716 	if (cpuc->is_fake)
3717 		return;
3718 
3719 	if (WARN_ON_ONCE(!excl_cntrs))
3720 		return;
3721 
3722 	if (hwc->flags & PERF_X86_EVENT_EXCL_ACCT) {
3723 		hwc->flags &= ~PERF_X86_EVENT_EXCL_ACCT;
3724 		if (!--cpuc->n_excl)
3725 			WRITE_ONCE(excl_cntrs->has_exclusive[tid], 0);
3726 	}
3727 
3728 	/*
3729 	 * If event was actually assigned, then mark the counter state as
3730 	 * unused now.
3731 	 */
3732 	if (hwc->idx >= 0) {
3733 		xl = &excl_cntrs->states[tid];
3734 
3735 		/*
3736 		 * put_constraint may be called from x86_schedule_events()
3737 		 * which already has the lock held so here make locking
3738 		 * conditional.
3739 		 */
3740 		if (!xl->sched_started)
3741 			raw_spin_lock(&excl_cntrs->lock);
3742 
3743 		xl->state[hwc->idx] = INTEL_EXCL_UNUSED;
3744 
3745 		if (!xl->sched_started)
3746 			raw_spin_unlock(&excl_cntrs->lock);
3747 	}
3748 }
3749 
3750 static void
intel_put_shared_regs_event_constraints(struct cpu_hw_events * cpuc,struct perf_event * event)3751 intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc,
3752 					struct perf_event *event)
3753 {
3754 	struct hw_perf_event_extra *reg;
3755 
3756 	reg = &event->hw.extra_reg;
3757 	if (reg->idx != EXTRA_REG_NONE)
3758 		__intel_shared_reg_put_constraints(cpuc, reg);
3759 
3760 	reg = &event->hw.branch_reg;
3761 	if (reg->idx != EXTRA_REG_NONE)
3762 		__intel_shared_reg_put_constraints(cpuc, reg);
3763 }
3764 
intel_put_event_constraints(struct cpu_hw_events * cpuc,struct perf_event * event)3765 static void intel_put_event_constraints(struct cpu_hw_events *cpuc,
3766 					struct perf_event *event)
3767 {
3768 	intel_put_shared_regs_event_constraints(cpuc, event);
3769 
3770 	/*
3771 	 * is PMU has exclusive counter restrictions, then
3772 	 * all events are subject to and must call the
3773 	 * put_excl_constraints() routine
3774 	 */
3775 	if (cpuc->excl_cntrs)
3776 		intel_put_excl_constraints(cpuc, event);
3777 }
3778 
intel_pebs_aliases_core2(struct perf_event * event)3779 static void intel_pebs_aliases_core2(struct perf_event *event)
3780 {
3781 	if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
3782 		/*
3783 		 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
3784 		 * (0x003c) so that we can use it with PEBS.
3785 		 *
3786 		 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
3787 		 * PEBS capable. However we can use INST_RETIRED.ANY_P
3788 		 * (0x00c0), which is a PEBS capable event, to get the same
3789 		 * count.
3790 		 *
3791 		 * INST_RETIRED.ANY_P counts the number of cycles that retires
3792 		 * CNTMASK instructions. By setting CNTMASK to a value (16)
3793 		 * larger than the maximum number of instructions that can be
3794 		 * retired per cycle (4) and then inverting the condition, we
3795 		 * count all cycles that retire 16 or less instructions, which
3796 		 * is every cycle.
3797 		 *
3798 		 * Thereby we gain a PEBS capable cycle counter.
3799 		 */
3800 		u64 alt_config = X86_CONFIG(.event=0xc0, .inv=1, .cmask=16);
3801 
3802 		alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
3803 		event->hw.config = alt_config;
3804 	}
3805 }
3806 
intel_pebs_aliases_snb(struct perf_event * event)3807 static void intel_pebs_aliases_snb(struct perf_event *event)
3808 {
3809 	if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
3810 		/*
3811 		 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
3812 		 * (0x003c) so that we can use it with PEBS.
3813 		 *
3814 		 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
3815 		 * PEBS capable. However we can use UOPS_RETIRED.ALL
3816 		 * (0x01c2), which is a PEBS capable event, to get the same
3817 		 * count.
3818 		 *
3819 		 * UOPS_RETIRED.ALL counts the number of cycles that retires
3820 		 * CNTMASK micro-ops. By setting CNTMASK to a value (16)
3821 		 * larger than the maximum number of micro-ops that can be
3822 		 * retired per cycle (4) and then inverting the condition, we
3823 		 * count all cycles that retire 16 or less micro-ops, which
3824 		 * is every cycle.
3825 		 *
3826 		 * Thereby we gain a PEBS capable cycle counter.
3827 		 */
3828 		u64 alt_config = X86_CONFIG(.event=0xc2, .umask=0x01, .inv=1, .cmask=16);
3829 
3830 		alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
3831 		event->hw.config = alt_config;
3832 	}
3833 }
3834 
intel_pebs_aliases_precdist(struct perf_event * event)3835 static void intel_pebs_aliases_precdist(struct perf_event *event)
3836 {
3837 	if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
3838 		/*
3839 		 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
3840 		 * (0x003c) so that we can use it with PEBS.
3841 		 *
3842 		 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
3843 		 * PEBS capable. However we can use INST_RETIRED.PREC_DIST
3844 		 * (0x01c0), which is a PEBS capable event, to get the same
3845 		 * count.
3846 		 *
3847 		 * The PREC_DIST event has special support to minimize sample
3848 		 * shadowing effects. One drawback is that it can be
3849 		 * only programmed on counter 1, but that seems like an
3850 		 * acceptable trade off.
3851 		 */
3852 		u64 alt_config = X86_CONFIG(.event=0xc0, .umask=0x01, .inv=1, .cmask=16);
3853 
3854 		alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
3855 		event->hw.config = alt_config;
3856 	}
3857 }
3858 
intel_pebs_aliases_ivb(struct perf_event * event)3859 static void intel_pebs_aliases_ivb(struct perf_event *event)
3860 {
3861 	if (event->attr.precise_ip < 3)
3862 		return intel_pebs_aliases_snb(event);
3863 	return intel_pebs_aliases_precdist(event);
3864 }
3865 
intel_pebs_aliases_skl(struct perf_event * event)3866 static void intel_pebs_aliases_skl(struct perf_event *event)
3867 {
3868 	if (event->attr.precise_ip < 3)
3869 		return intel_pebs_aliases_core2(event);
3870 	return intel_pebs_aliases_precdist(event);
3871 }
3872 
intel_pmu_large_pebs_flags(struct perf_event * event)3873 static unsigned long intel_pmu_large_pebs_flags(struct perf_event *event)
3874 {
3875 	unsigned long flags = x86_pmu.large_pebs_flags;
3876 
3877 	if (event->attr.use_clockid)
3878 		flags &= ~PERF_SAMPLE_TIME;
3879 	if (!event->attr.exclude_kernel)
3880 		flags &= ~PERF_SAMPLE_REGS_USER;
3881 	if (event->attr.sample_regs_user & ~PEBS_GP_REGS)
3882 		flags &= ~(PERF_SAMPLE_REGS_USER | PERF_SAMPLE_REGS_INTR);
3883 	return flags;
3884 }
3885 
intel_pmu_bts_config(struct perf_event * event)3886 static int intel_pmu_bts_config(struct perf_event *event)
3887 {
3888 	struct perf_event_attr *attr = &event->attr;
3889 
3890 	if (unlikely(intel_pmu_has_bts(event))) {
3891 		/* BTS is not supported by this architecture. */
3892 		if (!x86_pmu.bts_active)
3893 			return -EOPNOTSUPP;
3894 
3895 		/* BTS is currently only allowed for user-mode. */
3896 		if (!attr->exclude_kernel)
3897 			return -EOPNOTSUPP;
3898 
3899 		/* BTS is not allowed for precise events. */
3900 		if (attr->precise_ip)
3901 			return -EOPNOTSUPP;
3902 
3903 		/* disallow bts if conflicting events are present */
3904 		if (x86_add_exclusive(x86_lbr_exclusive_lbr))
3905 			return -EBUSY;
3906 
3907 		event->destroy = hw_perf_lbr_event_destroy;
3908 	}
3909 
3910 	return 0;
3911 }
3912 
core_pmu_hw_config(struct perf_event * event)3913 static int core_pmu_hw_config(struct perf_event *event)
3914 {
3915 	int ret = x86_pmu_hw_config(event);
3916 
3917 	if (ret)
3918 		return ret;
3919 
3920 	return intel_pmu_bts_config(event);
3921 }
3922 
3923 #define INTEL_TD_METRIC_AVAILABLE_MAX	(INTEL_TD_METRIC_RETIRING + \
3924 					 ((x86_pmu.num_topdown_events - 1) << 8))
3925 
is_available_metric_event(struct perf_event * event)3926 static bool is_available_metric_event(struct perf_event *event)
3927 {
3928 	return is_metric_event(event) &&
3929 		event->attr.config <= INTEL_TD_METRIC_AVAILABLE_MAX;
3930 }
3931 
is_mem_loads_event(struct perf_event * event)3932 static inline bool is_mem_loads_event(struct perf_event *event)
3933 {
3934 	return (event->attr.config & INTEL_ARCH_EVENT_MASK) == X86_CONFIG(.event=0xcd, .umask=0x01);
3935 }
3936 
is_mem_loads_aux_event(struct perf_event * event)3937 static inline bool is_mem_loads_aux_event(struct perf_event *event)
3938 {
3939 	return (event->attr.config & INTEL_ARCH_EVENT_MASK) == X86_CONFIG(.event=0x03, .umask=0x82);
3940 }
3941 
require_mem_loads_aux_event(struct perf_event * event)3942 static inline bool require_mem_loads_aux_event(struct perf_event *event)
3943 {
3944 	if (!(x86_pmu.flags & PMU_FL_MEM_LOADS_AUX))
3945 		return false;
3946 
3947 	if (is_hybrid())
3948 		return hybrid_pmu(event->pmu)->pmu_type == hybrid_big;
3949 
3950 	return true;
3951 }
3952 
intel_pmu_has_cap(struct perf_event * event,int idx)3953 static inline bool intel_pmu_has_cap(struct perf_event *event, int idx)
3954 {
3955 	union perf_capabilities *intel_cap = &hybrid(event->pmu, intel_cap);
3956 
3957 	return test_bit(idx, (unsigned long *)&intel_cap->capabilities);
3958 }
3959 
intel_pmu_freq_start_period(struct perf_event * event)3960 static u64 intel_pmu_freq_start_period(struct perf_event *event)
3961 {
3962 	int type = event->attr.type;
3963 	u64 config, factor;
3964 	s64 start;
3965 
3966 	/*
3967 	 * The 127 is the lowest possible recommended SAV (sample after value)
3968 	 * for a 4000 freq (default freq), according to the event list JSON file.
3969 	 * Also, assume the workload is idle 50% time.
3970 	 */
3971 	factor = 64 * 4000;
3972 	if (type != PERF_TYPE_HARDWARE && type != PERF_TYPE_HW_CACHE)
3973 		goto end;
3974 
3975 	/*
3976 	 * The estimation of the start period in the freq mode is
3977 	 * based on the below assumption.
3978 	 *
3979 	 * For a cycles or an instructions event, 1GHZ of the
3980 	 * underlying platform, 1 IPC. The workload is idle 50% time.
3981 	 * The start period = 1,000,000,000 * 1 / freq / 2.
3982 	 *		    = 500,000,000 / freq
3983 	 *
3984 	 * Usually, the branch-related events occur less than the
3985 	 * instructions event. According to the Intel event list JSON
3986 	 * file, the SAV (sample after value) of a branch-related event
3987 	 * is usually 1/4 of an instruction event.
3988 	 * The start period of branch-related events = 125,000,000 / freq.
3989 	 *
3990 	 * The cache-related events occurs even less. The SAV is usually
3991 	 * 1/20 of an instruction event.
3992 	 * The start period of cache-related events = 25,000,000 / freq.
3993 	 */
3994 	config = event->attr.config & PERF_HW_EVENT_MASK;
3995 	if (type == PERF_TYPE_HARDWARE) {
3996 		switch (config) {
3997 		case PERF_COUNT_HW_CPU_CYCLES:
3998 		case PERF_COUNT_HW_INSTRUCTIONS:
3999 		case PERF_COUNT_HW_BUS_CYCLES:
4000 		case PERF_COUNT_HW_STALLED_CYCLES_FRONTEND:
4001 		case PERF_COUNT_HW_STALLED_CYCLES_BACKEND:
4002 		case PERF_COUNT_HW_REF_CPU_CYCLES:
4003 			factor = 500000000;
4004 			break;
4005 		case PERF_COUNT_HW_BRANCH_INSTRUCTIONS:
4006 		case PERF_COUNT_HW_BRANCH_MISSES:
4007 			factor = 125000000;
4008 			break;
4009 		case PERF_COUNT_HW_CACHE_REFERENCES:
4010 		case PERF_COUNT_HW_CACHE_MISSES:
4011 			factor = 25000000;
4012 			break;
4013 		default:
4014 			goto end;
4015 		}
4016 	}
4017 
4018 	if (type == PERF_TYPE_HW_CACHE)
4019 		factor = 25000000;
4020 end:
4021 	/*
4022 	 * Usually, a prime or a number with less factors (close to prime)
4023 	 * is chosen as an SAV, which makes it less likely that the sampling
4024 	 * period synchronizes with some periodic event in the workload.
4025 	 * Minus 1 to make it at least avoiding values near power of twos
4026 	 * for the default freq.
4027 	 */
4028 	start = DIV_ROUND_UP_ULL(factor, event->attr.sample_freq) - 1;
4029 
4030 	if (start > x86_pmu.max_period)
4031 		start = x86_pmu.max_period;
4032 
4033 	if (x86_pmu.limit_period)
4034 		x86_pmu.limit_period(event, &start);
4035 
4036 	return start;
4037 }
4038 
intel_pmu_hw_config(struct perf_event * event)4039 static int intel_pmu_hw_config(struct perf_event *event)
4040 {
4041 	int ret = x86_pmu_hw_config(event);
4042 
4043 	if (ret)
4044 		return ret;
4045 
4046 	ret = intel_pmu_bts_config(event);
4047 	if (ret)
4048 		return ret;
4049 
4050 	if (event->attr.freq && event->attr.sample_freq) {
4051 		event->hw.sample_period = intel_pmu_freq_start_period(event);
4052 		event->hw.last_period = event->hw.sample_period;
4053 		local64_set(&event->hw.period_left, event->hw.sample_period);
4054 	}
4055 
4056 	if (event->attr.precise_ip) {
4057 		if ((event->attr.config & INTEL_ARCH_EVENT_MASK) == INTEL_FIXED_VLBR_EVENT)
4058 			return -EINVAL;
4059 
4060 		if (!(event->attr.freq || (event->attr.wakeup_events && !event->attr.watermark))) {
4061 			event->hw.flags |= PERF_X86_EVENT_AUTO_RELOAD;
4062 			if (!(event->attr.sample_type & ~intel_pmu_large_pebs_flags(event)) &&
4063 			    !has_aux_action(event)) {
4064 				event->hw.flags |= PERF_X86_EVENT_LARGE_PEBS;
4065 				event->attach_state |= PERF_ATTACH_SCHED_CB;
4066 			}
4067 		}
4068 		if (x86_pmu.pebs_aliases)
4069 			x86_pmu.pebs_aliases(event);
4070 	}
4071 
4072 	if (needs_branch_stack(event)) {
4073 		/* Avoid branch stack setup for counting events in SAMPLE READ */
4074 		if (is_sampling_event(event) ||
4075 		    !(event->attr.sample_type & PERF_SAMPLE_READ))
4076 			event->hw.flags |= PERF_X86_EVENT_NEEDS_BRANCH_STACK;
4077 	}
4078 
4079 	if (branch_sample_counters(event)) {
4080 		struct perf_event *leader, *sibling;
4081 		int num = 0;
4082 
4083 		if (!(x86_pmu.flags & PMU_FL_BR_CNTR) ||
4084 		    (event->attr.config & ~INTEL_ARCH_EVENT_MASK))
4085 			return -EINVAL;
4086 
4087 		/*
4088 		 * The branch counter logging is not supported in the call stack
4089 		 * mode yet, since we cannot simply flush the LBR during e.g.,
4090 		 * multiplexing. Also, there is no obvious usage with the call
4091 		 * stack mode. Simply forbids it for now.
4092 		 *
4093 		 * If any events in the group enable the branch counter logging
4094 		 * feature, the group is treated as a branch counter logging
4095 		 * group, which requires the extra space to store the counters.
4096 		 */
4097 		leader = event->group_leader;
4098 		if (branch_sample_call_stack(leader))
4099 			return -EINVAL;
4100 		if (branch_sample_counters(leader))
4101 			num++;
4102 		leader->hw.flags |= PERF_X86_EVENT_BRANCH_COUNTERS;
4103 
4104 		for_each_sibling_event(sibling, leader) {
4105 			if (branch_sample_call_stack(sibling))
4106 				return -EINVAL;
4107 			if (branch_sample_counters(sibling))
4108 				num++;
4109 		}
4110 
4111 		if (num > fls(x86_pmu.lbr_counters))
4112 			return -EINVAL;
4113 		/*
4114 		 * Only applying the PERF_SAMPLE_BRANCH_COUNTERS doesn't
4115 		 * require any branch stack setup.
4116 		 * Clear the bit to avoid unnecessary branch stack setup.
4117 		 */
4118 		if (0 == (event->attr.branch_sample_type &
4119 			  ~(PERF_SAMPLE_BRANCH_PLM_ALL |
4120 			    PERF_SAMPLE_BRANCH_COUNTERS)))
4121 			event->hw.flags  &= ~PERF_X86_EVENT_NEEDS_BRANCH_STACK;
4122 
4123 		/*
4124 		 * Force the leader to be a LBR event. So LBRs can be reset
4125 		 * with the leader event. See intel_pmu_lbr_del() for details.
4126 		 */
4127 		if (!intel_pmu_needs_branch_stack(leader))
4128 			return -EINVAL;
4129 	}
4130 
4131 	if (intel_pmu_needs_branch_stack(event)) {
4132 		ret = intel_pmu_setup_lbr_filter(event);
4133 		if (ret)
4134 			return ret;
4135 		event->attach_state |= PERF_ATTACH_SCHED_CB;
4136 
4137 		/*
4138 		 * BTS is set up earlier in this path, so don't account twice
4139 		 */
4140 		if (!unlikely(intel_pmu_has_bts(event))) {
4141 			/* disallow lbr if conflicting events are present */
4142 			if (x86_add_exclusive(x86_lbr_exclusive_lbr))
4143 				return -EBUSY;
4144 
4145 			event->destroy = hw_perf_lbr_event_destroy;
4146 		}
4147 	}
4148 
4149 	if (event->attr.aux_output) {
4150 		if (!event->attr.precise_ip)
4151 			return -EINVAL;
4152 
4153 		event->hw.flags |= PERF_X86_EVENT_PEBS_VIA_PT;
4154 	}
4155 
4156 	if ((event->attr.type == PERF_TYPE_HARDWARE) ||
4157 	    (event->attr.type == PERF_TYPE_HW_CACHE))
4158 		return 0;
4159 
4160 	/*
4161 	 * Config Topdown slots and metric events
4162 	 *
4163 	 * The slots event on Fixed Counter 3 can support sampling,
4164 	 * which will be handled normally in x86_perf_event_update().
4165 	 *
4166 	 * Metric events don't support sampling and require being paired
4167 	 * with a slots event as group leader. When the slots event
4168 	 * is used in a metrics group, it too cannot support sampling.
4169 	 */
4170 	if (intel_pmu_has_cap(event, PERF_CAP_METRICS_IDX) && is_topdown_event(event)) {
4171 		/* The metrics_clear can only be set for the slots event */
4172 		if (event->attr.config1 &&
4173 		    (!is_slots_event(event) || (event->attr.config1 & ~INTEL_TD_CFG_METRIC_CLEAR)))
4174 			return -EINVAL;
4175 
4176 		if (event->attr.config2)
4177 			return -EINVAL;
4178 
4179 		/*
4180 		 * The TopDown metrics events and slots event don't
4181 		 * support any filters.
4182 		 */
4183 		if (event->attr.config & X86_ALL_EVENT_FLAGS)
4184 			return -EINVAL;
4185 
4186 		if (is_available_metric_event(event)) {
4187 			struct perf_event *leader = event->group_leader;
4188 
4189 			/* The metric events don't support sampling. */
4190 			if (is_sampling_event(event))
4191 				return -EINVAL;
4192 
4193 			/* The metric events require a slots group leader. */
4194 			if (!is_slots_event(leader))
4195 				return -EINVAL;
4196 
4197 			/*
4198 			 * The leader/SLOTS must not be a sampling event for
4199 			 * metric use; hardware requires it starts at 0 when used
4200 			 * in conjunction with MSR_PERF_METRICS.
4201 			 */
4202 			if (is_sampling_event(leader))
4203 				return -EINVAL;
4204 
4205 			event->event_caps |= PERF_EV_CAP_SIBLING;
4206 			/*
4207 			 * Only once we have a METRICs sibling do we
4208 			 * need TopDown magic.
4209 			 */
4210 			leader->hw.flags |= PERF_X86_EVENT_TOPDOWN;
4211 			event->hw.flags  |= PERF_X86_EVENT_TOPDOWN;
4212 		}
4213 	}
4214 
4215 	/*
4216 	 * The load latency event X86_CONFIG(.event=0xcd, .umask=0x01) on SPR
4217 	 * doesn't function quite right. As a work-around it needs to always be
4218 	 * co-scheduled with a auxiliary event X86_CONFIG(.event=0x03, .umask=0x82).
4219 	 * The actual count of this second event is irrelevant it just needs
4220 	 * to be active to make the first event function correctly.
4221 	 *
4222 	 * In a group, the auxiliary event must be in front of the load latency
4223 	 * event. The rule is to simplify the implementation of the check.
4224 	 * That's because perf cannot have a complete group at the moment.
4225 	 */
4226 	if (require_mem_loads_aux_event(event) &&
4227 	    (event->attr.sample_type & PERF_SAMPLE_DATA_SRC) &&
4228 	    is_mem_loads_event(event)) {
4229 		struct perf_event *leader = event->group_leader;
4230 		struct perf_event *sibling = NULL;
4231 
4232 		/*
4233 		 * When this memload event is also the first event (no group
4234 		 * exists yet), then there is no aux event before it.
4235 		 */
4236 		if (leader == event)
4237 			return -ENODATA;
4238 
4239 		if (!is_mem_loads_aux_event(leader)) {
4240 			for_each_sibling_event(sibling, leader) {
4241 				if (is_mem_loads_aux_event(sibling))
4242 					break;
4243 			}
4244 			if (list_entry_is_head(sibling, &leader->sibling_list, sibling_list))
4245 				return -ENODATA;
4246 		}
4247 	}
4248 
4249 	if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY))
4250 		return 0;
4251 
4252 	if (x86_pmu.version < 3)
4253 		return -EINVAL;
4254 
4255 	ret = perf_allow_cpu(&event->attr);
4256 	if (ret)
4257 		return ret;
4258 
4259 	event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY;
4260 
4261 	return 0;
4262 }
4263 
4264 /*
4265  * Currently, the only caller of this function is the atomic_switch_perf_msrs().
4266  * The host perf context helps to prepare the values of the real hardware for
4267  * a set of msrs that need to be switched atomically in a vmx transaction.
4268  *
4269  * For example, the pseudocode needed to add a new msr should look like:
4270  *
4271  * arr[(*nr)++] = (struct perf_guest_switch_msr){
4272  *	.msr = the hardware msr address,
4273  *	.host = the value the hardware has when it doesn't run a guest,
4274  *	.guest = the value the hardware has when it runs a guest,
4275  * };
4276  *
4277  * These values have nothing to do with the emulated values the guest sees
4278  * when it uses {RD,WR}MSR, which should be handled by the KVM context,
4279  * specifically in the intel_pmu_{get,set}_msr().
4280  */
intel_guest_get_msrs(int * nr,void * data)4281 static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr, void *data)
4282 {
4283 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
4284 	struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
4285 	struct kvm_pmu *kvm_pmu = (struct kvm_pmu *)data;
4286 	u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl);
4287 	u64 pebs_mask = cpuc->pebs_enabled & x86_pmu.pebs_capable;
4288 	int global_ctrl, pebs_enable;
4289 
4290 	/*
4291 	 * In addition to obeying exclude_guest/exclude_host, remove bits being
4292 	 * used for PEBS when running a guest, because PEBS writes to virtual
4293 	 * addresses (not physical addresses).
4294 	 */
4295 	*nr = 0;
4296 	global_ctrl = (*nr)++;
4297 	arr[global_ctrl] = (struct perf_guest_switch_msr){
4298 		.msr = MSR_CORE_PERF_GLOBAL_CTRL,
4299 		.host = intel_ctrl & ~cpuc->intel_ctrl_guest_mask,
4300 		.guest = intel_ctrl & ~cpuc->intel_ctrl_host_mask & ~pebs_mask,
4301 	};
4302 
4303 	if (!x86_pmu.pebs)
4304 		return arr;
4305 
4306 	/*
4307 	 * If PMU counter has PEBS enabled it is not enough to
4308 	 * disable counter on a guest entry since PEBS memory
4309 	 * write can overshoot guest entry and corrupt guest
4310 	 * memory. Disabling PEBS solves the problem.
4311 	 *
4312 	 * Don't do this if the CPU already enforces it.
4313 	 */
4314 	if (x86_pmu.pebs_no_isolation) {
4315 		arr[(*nr)++] = (struct perf_guest_switch_msr){
4316 			.msr = MSR_IA32_PEBS_ENABLE,
4317 			.host = cpuc->pebs_enabled,
4318 			.guest = 0,
4319 		};
4320 		return arr;
4321 	}
4322 
4323 	if (!kvm_pmu || !x86_pmu.pebs_ept)
4324 		return arr;
4325 
4326 	arr[(*nr)++] = (struct perf_guest_switch_msr){
4327 		.msr = MSR_IA32_DS_AREA,
4328 		.host = (unsigned long)cpuc->ds,
4329 		.guest = kvm_pmu->ds_area,
4330 	};
4331 
4332 	if (x86_pmu.intel_cap.pebs_baseline) {
4333 		arr[(*nr)++] = (struct perf_guest_switch_msr){
4334 			.msr = MSR_PEBS_DATA_CFG,
4335 			.host = cpuc->active_pebs_data_cfg,
4336 			.guest = kvm_pmu->pebs_data_cfg,
4337 		};
4338 	}
4339 
4340 	pebs_enable = (*nr)++;
4341 	arr[pebs_enable] = (struct perf_guest_switch_msr){
4342 		.msr = MSR_IA32_PEBS_ENABLE,
4343 		.host = cpuc->pebs_enabled & ~cpuc->intel_ctrl_guest_mask,
4344 		.guest = pebs_mask & ~cpuc->intel_ctrl_host_mask,
4345 	};
4346 
4347 	if (arr[pebs_enable].host) {
4348 		/* Disable guest PEBS if host PEBS is enabled. */
4349 		arr[pebs_enable].guest = 0;
4350 	} else {
4351 		/* Disable guest PEBS thoroughly for cross-mapped PEBS counters. */
4352 		arr[pebs_enable].guest &= ~kvm_pmu->host_cross_mapped_mask;
4353 		arr[global_ctrl].guest &= ~kvm_pmu->host_cross_mapped_mask;
4354 		/* Set hw GLOBAL_CTRL bits for PEBS counter when it runs for guest */
4355 		arr[global_ctrl].guest |= arr[pebs_enable].guest;
4356 	}
4357 
4358 	return arr;
4359 }
4360 
core_guest_get_msrs(int * nr,void * data)4361 static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr, void *data)
4362 {
4363 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
4364 	struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
4365 	int idx;
4366 
4367 	for_each_set_bit(idx, x86_pmu.cntr_mask, X86_PMC_IDX_MAX) {
4368 		struct perf_event *event = cpuc->events[idx];
4369 
4370 		arr[idx].msr = x86_pmu_config_addr(idx);
4371 		arr[idx].host = arr[idx].guest = 0;
4372 
4373 		if (!test_bit(idx, cpuc->active_mask))
4374 			continue;
4375 
4376 		arr[idx].host = arr[idx].guest =
4377 			event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE;
4378 
4379 		if (event->attr.exclude_host)
4380 			arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
4381 		else if (event->attr.exclude_guest)
4382 			arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
4383 	}
4384 
4385 	*nr = x86_pmu_max_num_counters(cpuc->pmu);
4386 	return arr;
4387 }
4388 
core_pmu_enable_event(struct perf_event * event)4389 static void core_pmu_enable_event(struct perf_event *event)
4390 {
4391 	if (!event->attr.exclude_host)
4392 		x86_pmu_enable_event(event);
4393 }
4394 
core_pmu_enable_all(int added)4395 static void core_pmu_enable_all(int added)
4396 {
4397 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
4398 	int idx;
4399 
4400 	for_each_set_bit(idx, x86_pmu.cntr_mask, X86_PMC_IDX_MAX) {
4401 		struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
4402 
4403 		if (!test_bit(idx, cpuc->active_mask) ||
4404 				cpuc->events[idx]->attr.exclude_host)
4405 			continue;
4406 
4407 		__x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
4408 	}
4409 }
4410 
hsw_hw_config(struct perf_event * event)4411 static int hsw_hw_config(struct perf_event *event)
4412 {
4413 	int ret = intel_pmu_hw_config(event);
4414 
4415 	if (ret)
4416 		return ret;
4417 	if (!boot_cpu_has(X86_FEATURE_RTM) && !boot_cpu_has(X86_FEATURE_HLE))
4418 		return 0;
4419 	event->hw.config |= event->attr.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED);
4420 
4421 	/*
4422 	 * IN_TX/IN_TX-CP filters are not supported by the Haswell PMU with
4423 	 * PEBS or in ANY thread mode. Since the results are non-sensical forbid
4424 	 * this combination.
4425 	 */
4426 	if ((event->hw.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)) &&
4427 	     ((event->hw.config & ARCH_PERFMON_EVENTSEL_ANY) ||
4428 	      event->attr.precise_ip > 0))
4429 		return -EOPNOTSUPP;
4430 
4431 	if (event_is_checkpointed(event)) {
4432 		/*
4433 		 * Sampling of checkpointed events can cause situations where
4434 		 * the CPU constantly aborts because of a overflow, which is
4435 		 * then checkpointed back and ignored. Forbid checkpointing
4436 		 * for sampling.
4437 		 *
4438 		 * But still allow a long sampling period, so that perf stat
4439 		 * from KVM works.
4440 		 */
4441 		if (event->attr.sample_period > 0 &&
4442 		    event->attr.sample_period < 0x7fffffff)
4443 			return -EOPNOTSUPP;
4444 	}
4445 	return 0;
4446 }
4447 
4448 static struct event_constraint counter0_constraint =
4449 			INTEL_ALL_EVENT_CONSTRAINT(0, 0x1);
4450 
4451 static struct event_constraint counter1_constraint =
4452 			INTEL_ALL_EVENT_CONSTRAINT(0, 0x2);
4453 
4454 static struct event_constraint counter0_1_constraint =
4455 			INTEL_ALL_EVENT_CONSTRAINT(0, 0x3);
4456 
4457 static struct event_constraint counter2_constraint =
4458 			EVENT_CONSTRAINT(0, 0x4, 0);
4459 
4460 static struct event_constraint fixed0_constraint =
4461 			FIXED_EVENT_CONSTRAINT(0x00c0, 0);
4462 
4463 static struct event_constraint fixed0_counter0_constraint =
4464 			INTEL_ALL_EVENT_CONSTRAINT(0, 0x100000001ULL);
4465 
4466 static struct event_constraint fixed0_counter0_1_constraint =
4467 			INTEL_ALL_EVENT_CONSTRAINT(0, 0x100000003ULL);
4468 
4469 static struct event_constraint counters_1_7_constraint =
4470 			INTEL_ALL_EVENT_CONSTRAINT(0, 0xfeULL);
4471 
4472 static struct event_constraint *
hsw_get_event_constraints(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)4473 hsw_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4474 			  struct perf_event *event)
4475 {
4476 	struct event_constraint *c;
4477 
4478 	c = intel_get_event_constraints(cpuc, idx, event);
4479 
4480 	/* Handle special quirk on in_tx_checkpointed only in counter 2 */
4481 	if (event->hw.config & HSW_IN_TX_CHECKPOINTED) {
4482 		if (c->idxmsk64 & (1U << 2))
4483 			return &counter2_constraint;
4484 		return &emptyconstraint;
4485 	}
4486 
4487 	return c;
4488 }
4489 
4490 static struct event_constraint *
icl_get_event_constraints(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)4491 icl_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4492 			  struct perf_event *event)
4493 {
4494 	/*
4495 	 * Fixed counter 0 has less skid.
4496 	 * Force instruction:ppp in Fixed counter 0
4497 	 */
4498 	if ((event->attr.precise_ip == 3) &&
4499 	    constraint_match(&fixed0_constraint, event->hw.config))
4500 		return &fixed0_constraint;
4501 
4502 	return hsw_get_event_constraints(cpuc, idx, event);
4503 }
4504 
4505 static struct event_constraint *
glc_get_event_constraints(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)4506 glc_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4507 			  struct perf_event *event)
4508 {
4509 	struct event_constraint *c;
4510 
4511 	c = icl_get_event_constraints(cpuc, idx, event);
4512 
4513 	/*
4514 	 * The :ppp indicates the Precise Distribution (PDist) facility, which
4515 	 * is only supported on the GP counter 0. If a :ppp event which is not
4516 	 * available on the GP counter 0, error out.
4517 	 * Exception: Instruction PDIR is only available on the fixed counter 0.
4518 	 */
4519 	if ((event->attr.precise_ip == 3) &&
4520 	    !constraint_match(&fixed0_constraint, event->hw.config)) {
4521 		if (c->idxmsk64 & BIT_ULL(0))
4522 			return &counter0_constraint;
4523 
4524 		return &emptyconstraint;
4525 	}
4526 
4527 	return c;
4528 }
4529 
4530 static struct event_constraint *
glp_get_event_constraints(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)4531 glp_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4532 			  struct perf_event *event)
4533 {
4534 	struct event_constraint *c;
4535 
4536 	/* :ppp means to do reduced skid PEBS which is PMC0 only. */
4537 	if (event->attr.precise_ip == 3)
4538 		return &counter0_constraint;
4539 
4540 	c = intel_get_event_constraints(cpuc, idx, event);
4541 
4542 	return c;
4543 }
4544 
4545 static struct event_constraint *
tnt_get_event_constraints(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)4546 tnt_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4547 			  struct perf_event *event)
4548 {
4549 	struct event_constraint *c;
4550 
4551 	c = intel_get_event_constraints(cpuc, idx, event);
4552 
4553 	/*
4554 	 * :ppp means to do reduced skid PEBS,
4555 	 * which is available on PMC0 and fixed counter 0.
4556 	 */
4557 	if (event->attr.precise_ip == 3) {
4558 		/* Force instruction:ppp on PMC0 and Fixed counter 0 */
4559 		if (constraint_match(&fixed0_constraint, event->hw.config))
4560 			return &fixed0_counter0_constraint;
4561 
4562 		return &counter0_constraint;
4563 	}
4564 
4565 	return c;
4566 }
4567 
4568 static bool allow_tsx_force_abort = true;
4569 
4570 static struct event_constraint *
tfa_get_event_constraints(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)4571 tfa_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4572 			  struct perf_event *event)
4573 {
4574 	struct event_constraint *c = hsw_get_event_constraints(cpuc, idx, event);
4575 
4576 	/*
4577 	 * Without TFA we must not use PMC3.
4578 	 */
4579 	if (!allow_tsx_force_abort && test_bit(3, c->idxmsk)) {
4580 		c = dyn_constraint(cpuc, c, idx);
4581 		c->idxmsk64 &= ~(1ULL << 3);
4582 		c->weight--;
4583 	}
4584 
4585 	return c;
4586 }
4587 
4588 static struct event_constraint *
adl_get_event_constraints(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)4589 adl_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4590 			  struct perf_event *event)
4591 {
4592 	struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu);
4593 
4594 	if (pmu->pmu_type == hybrid_big)
4595 		return glc_get_event_constraints(cpuc, idx, event);
4596 	else if (pmu->pmu_type == hybrid_small)
4597 		return tnt_get_event_constraints(cpuc, idx, event);
4598 
4599 	WARN_ON(1);
4600 	return &emptyconstraint;
4601 }
4602 
4603 static struct event_constraint *
cmt_get_event_constraints(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)4604 cmt_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4605 			  struct perf_event *event)
4606 {
4607 	struct event_constraint *c;
4608 
4609 	c = intel_get_event_constraints(cpuc, idx, event);
4610 
4611 	/*
4612 	 * The :ppp indicates the Precise Distribution (PDist) facility, which
4613 	 * is only supported on the GP counter 0 & 1 and Fixed counter 0.
4614 	 * If a :ppp event which is not available on the above eligible counters,
4615 	 * error out.
4616 	 */
4617 	if (event->attr.precise_ip == 3) {
4618 		/* Force instruction:ppp on PMC0, 1 and Fixed counter 0 */
4619 		if (constraint_match(&fixed0_constraint, event->hw.config)) {
4620 			/* The fixed counter 0 doesn't support LBR event logging. */
4621 			if (branch_sample_counters(event))
4622 				return &counter0_1_constraint;
4623 			else
4624 				return &fixed0_counter0_1_constraint;
4625 		}
4626 
4627 		switch (c->idxmsk64 & 0x3ull) {
4628 		case 0x1:
4629 			return &counter0_constraint;
4630 		case 0x2:
4631 			return &counter1_constraint;
4632 		case 0x3:
4633 			return &counter0_1_constraint;
4634 		}
4635 		return &emptyconstraint;
4636 	}
4637 
4638 	return c;
4639 }
4640 
4641 static struct event_constraint *
rwc_get_event_constraints(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)4642 rwc_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4643 			  struct perf_event *event)
4644 {
4645 	struct event_constraint *c;
4646 
4647 	c = glc_get_event_constraints(cpuc, idx, event);
4648 
4649 	/* The Retire Latency is not supported by the fixed counter 0. */
4650 	if (event->attr.precise_ip &&
4651 	    (event->attr.sample_type & PERF_SAMPLE_WEIGHT_TYPE) &&
4652 	    constraint_match(&fixed0_constraint, event->hw.config)) {
4653 		/*
4654 		 * The Instruction PDIR is only available
4655 		 * on the fixed counter 0. Error out for this case.
4656 		 */
4657 		if (event->attr.precise_ip == 3)
4658 			return &emptyconstraint;
4659 		return &counters_1_7_constraint;
4660 	}
4661 
4662 	return c;
4663 }
4664 
4665 static struct event_constraint *
mtl_get_event_constraints(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)4666 mtl_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4667 			  struct perf_event *event)
4668 {
4669 	struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu);
4670 
4671 	if (pmu->pmu_type == hybrid_big)
4672 		return rwc_get_event_constraints(cpuc, idx, event);
4673 	if (pmu->pmu_type == hybrid_small)
4674 		return cmt_get_event_constraints(cpuc, idx, event);
4675 
4676 	WARN_ON(1);
4677 	return &emptyconstraint;
4678 }
4679 
adl_hw_config(struct perf_event * event)4680 static int adl_hw_config(struct perf_event *event)
4681 {
4682 	struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu);
4683 
4684 	if (pmu->pmu_type == hybrid_big)
4685 		return hsw_hw_config(event);
4686 	else if (pmu->pmu_type == hybrid_small)
4687 		return intel_pmu_hw_config(event);
4688 
4689 	WARN_ON(1);
4690 	return -EOPNOTSUPP;
4691 }
4692 
adl_get_hybrid_cpu_type(void)4693 static enum hybrid_cpu_type adl_get_hybrid_cpu_type(void)
4694 {
4695 	return HYBRID_INTEL_CORE;
4696 }
4697 
erratum_hsw11(struct perf_event * event)4698 static inline bool erratum_hsw11(struct perf_event *event)
4699 {
4700 	return (event->hw.config & INTEL_ARCH_EVENT_MASK) ==
4701 		X86_CONFIG(.event=0xc0, .umask=0x01);
4702 }
4703 
4704 static struct event_constraint *
arl_h_get_event_constraints(struct cpu_hw_events * cpuc,int idx,struct perf_event * event)4705 arl_h_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
4706 			  struct perf_event *event)
4707 {
4708 	struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu);
4709 
4710 	if (pmu->pmu_type == hybrid_tiny)
4711 		return cmt_get_event_constraints(cpuc, idx, event);
4712 
4713 	return mtl_get_event_constraints(cpuc, idx, event);
4714 }
4715 
arl_h_hw_config(struct perf_event * event)4716 static int arl_h_hw_config(struct perf_event *event)
4717 {
4718 	struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu);
4719 
4720 	if (pmu->pmu_type == hybrid_tiny)
4721 		return intel_pmu_hw_config(event);
4722 
4723 	return adl_hw_config(event);
4724 }
4725 
4726 /*
4727  * The HSW11 requires a period larger than 100 which is the same as the BDM11.
4728  * A minimum period of 128 is enforced as well for the INST_RETIRED.ALL.
4729  *
4730  * The message 'interrupt took too long' can be observed on any counter which
4731  * was armed with a period < 32 and two events expired in the same NMI.
4732  * A minimum period of 32 is enforced for the rest of the events.
4733  */
hsw_limit_period(struct perf_event * event,s64 * left)4734 static void hsw_limit_period(struct perf_event *event, s64 *left)
4735 {
4736 	*left = max(*left, erratum_hsw11(event) ? 128 : 32);
4737 }
4738 
4739 /*
4740  * Broadwell:
4741  *
4742  * The INST_RETIRED.ALL period always needs to have lowest 6 bits cleared
4743  * (BDM55) and it must not use a period smaller than 100 (BDM11). We combine
4744  * the two to enforce a minimum period of 128 (the smallest value that has bits
4745  * 0-5 cleared and >= 100).
4746  *
4747  * Because of how the code in x86_perf_event_set_period() works, the truncation
4748  * of the lower 6 bits is 'harmless' as we'll occasionally add a longer period
4749  * to make up for the 'lost' events due to carrying the 'error' in period_left.
4750  *
4751  * Therefore the effective (average) period matches the requested period,
4752  * despite coarser hardware granularity.
4753  */
bdw_limit_period(struct perf_event * event,s64 * left)4754 static void bdw_limit_period(struct perf_event *event, s64 *left)
4755 {
4756 	if (erratum_hsw11(event)) {
4757 		if (*left < 128)
4758 			*left = 128;
4759 		*left &= ~0x3fULL;
4760 	}
4761 }
4762 
nhm_limit_period(struct perf_event * event,s64 * left)4763 static void nhm_limit_period(struct perf_event *event, s64 *left)
4764 {
4765 	*left = max(*left, 32LL);
4766 }
4767 
glc_limit_period(struct perf_event * event,s64 * left)4768 static void glc_limit_period(struct perf_event *event, s64 *left)
4769 {
4770 	if (event->attr.precise_ip == 3)
4771 		*left = max(*left, 128LL);
4772 }
4773 
4774 PMU_FORMAT_ATTR(event,	"config:0-7"	);
4775 PMU_FORMAT_ATTR(umask,	"config:8-15"	);
4776 PMU_FORMAT_ATTR(edge,	"config:18"	);
4777 PMU_FORMAT_ATTR(pc,	"config:19"	);
4778 PMU_FORMAT_ATTR(any,	"config:21"	); /* v3 + */
4779 PMU_FORMAT_ATTR(inv,	"config:23"	);
4780 PMU_FORMAT_ATTR(cmask,	"config:24-31"	);
4781 PMU_FORMAT_ATTR(in_tx,  "config:32"	);
4782 PMU_FORMAT_ATTR(in_tx_cp, "config:33"	);
4783 PMU_FORMAT_ATTR(eq,	"config:36"	); /* v6 + */
4784 
4785 PMU_FORMAT_ATTR(metrics_clear,	"config1:0"); /* PERF_CAPABILITIES.RDPMC_METRICS_CLEAR */
4786 
umask2_show(struct device * dev,struct device_attribute * attr,char * page)4787 static ssize_t umask2_show(struct device *dev,
4788 			   struct device_attribute *attr,
4789 			   char *page)
4790 {
4791 	u64 mask = hybrid(dev_get_drvdata(dev), config_mask) & ARCH_PERFMON_EVENTSEL_UMASK2;
4792 
4793 	if (mask == ARCH_PERFMON_EVENTSEL_UMASK2)
4794 		return sprintf(page, "config:8-15,40-47\n");
4795 
4796 	/* Roll back to the old format if umask2 is not supported. */
4797 	return sprintf(page, "config:8-15\n");
4798 }
4799 
4800 static struct device_attribute format_attr_umask2  =
4801 		__ATTR(umask, 0444, umask2_show, NULL);
4802 
4803 static struct attribute *format_evtsel_ext_attrs[] = {
4804 	&format_attr_umask2.attr,
4805 	&format_attr_eq.attr,
4806 	&format_attr_metrics_clear.attr,
4807 	NULL
4808 };
4809 
4810 static umode_t
evtsel_ext_is_visible(struct kobject * kobj,struct attribute * attr,int i)4811 evtsel_ext_is_visible(struct kobject *kobj, struct attribute *attr, int i)
4812 {
4813 	struct device *dev = kobj_to_dev(kobj);
4814 	u64 mask;
4815 
4816 	/*
4817 	 * The umask and umask2 have different formats but share the
4818 	 * same attr name. In update mode, the previous value of the
4819 	 * umask is unconditionally removed before is_visible. If
4820 	 * umask2 format is not enumerated, it's impossible to roll
4821 	 * back to the old format.
4822 	 * Does the check in umask2_show rather than is_visible.
4823 	 */
4824 	if (i == 0)
4825 		return attr->mode;
4826 
4827 	mask = hybrid(dev_get_drvdata(dev), config_mask);
4828 	if (i == 1)
4829 		return (mask & ARCH_PERFMON_EVENTSEL_EQ) ? attr->mode : 0;
4830 
4831 	/* PERF_CAPABILITIES.RDPMC_METRICS_CLEAR */
4832 	if (i == 2) {
4833 		union perf_capabilities intel_cap = hybrid(dev_get_drvdata(dev), intel_cap);
4834 
4835 		return intel_cap.rdpmc_metrics_clear ? attr->mode : 0;
4836 	}
4837 
4838 	return 0;
4839 }
4840 
4841 static struct attribute *intel_arch_formats_attr[] = {
4842 	&format_attr_event.attr,
4843 	&format_attr_umask.attr,
4844 	&format_attr_edge.attr,
4845 	&format_attr_pc.attr,
4846 	&format_attr_inv.attr,
4847 	&format_attr_cmask.attr,
4848 	NULL,
4849 };
4850 
intel_event_sysfs_show(char * page,u64 config)4851 ssize_t intel_event_sysfs_show(char *page, u64 config)
4852 {
4853 	u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT);
4854 
4855 	return x86_event_sysfs_show(page, config, event);
4856 }
4857 
allocate_shared_regs(int cpu)4858 static struct intel_shared_regs *allocate_shared_regs(int cpu)
4859 {
4860 	struct intel_shared_regs *regs;
4861 	int i;
4862 
4863 	regs = kzalloc_node(sizeof(struct intel_shared_regs),
4864 			    GFP_KERNEL, cpu_to_node(cpu));
4865 	if (regs) {
4866 		/*
4867 		 * initialize the locks to keep lockdep happy
4868 		 */
4869 		for (i = 0; i < EXTRA_REG_MAX; i++)
4870 			raw_spin_lock_init(&regs->regs[i].lock);
4871 
4872 		regs->core_id = -1;
4873 	}
4874 	return regs;
4875 }
4876 
allocate_excl_cntrs(int cpu)4877 static struct intel_excl_cntrs *allocate_excl_cntrs(int cpu)
4878 {
4879 	struct intel_excl_cntrs *c;
4880 
4881 	c = kzalloc_node(sizeof(struct intel_excl_cntrs),
4882 			 GFP_KERNEL, cpu_to_node(cpu));
4883 	if (c) {
4884 		raw_spin_lock_init(&c->lock);
4885 		c->core_id = -1;
4886 	}
4887 	return c;
4888 }
4889 
4890 
intel_cpuc_prepare(struct cpu_hw_events * cpuc,int cpu)4891 int intel_cpuc_prepare(struct cpu_hw_events *cpuc, int cpu)
4892 {
4893 	cpuc->pebs_record_size = x86_pmu.pebs_record_size;
4894 
4895 	if (is_hybrid() || x86_pmu.extra_regs || x86_pmu.lbr_sel_map) {
4896 		cpuc->shared_regs = allocate_shared_regs(cpu);
4897 		if (!cpuc->shared_regs)
4898 			goto err;
4899 	}
4900 
4901 	if (x86_pmu.flags & (PMU_FL_EXCL_CNTRS | PMU_FL_TFA | PMU_FL_BR_CNTR)) {
4902 		size_t sz = X86_PMC_IDX_MAX * sizeof(struct event_constraint);
4903 
4904 		cpuc->constraint_list = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu));
4905 		if (!cpuc->constraint_list)
4906 			goto err_shared_regs;
4907 	}
4908 
4909 	if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) {
4910 		cpuc->excl_cntrs = allocate_excl_cntrs(cpu);
4911 		if (!cpuc->excl_cntrs)
4912 			goto err_constraint_list;
4913 
4914 		cpuc->excl_thread_id = 0;
4915 	}
4916 
4917 	return 0;
4918 
4919 err_constraint_list:
4920 	kfree(cpuc->constraint_list);
4921 	cpuc->constraint_list = NULL;
4922 
4923 err_shared_regs:
4924 	kfree(cpuc->shared_regs);
4925 	cpuc->shared_regs = NULL;
4926 
4927 err:
4928 	return -ENOMEM;
4929 }
4930 
intel_pmu_cpu_prepare(int cpu)4931 static int intel_pmu_cpu_prepare(int cpu)
4932 {
4933 	return intel_cpuc_prepare(&per_cpu(cpu_hw_events, cpu), cpu);
4934 }
4935 
flip_smm_bit(void * data)4936 static void flip_smm_bit(void *data)
4937 {
4938 	unsigned long set = *(unsigned long *)data;
4939 
4940 	if (set > 0) {
4941 		msr_set_bit(MSR_IA32_DEBUGCTLMSR,
4942 			    DEBUGCTLMSR_FREEZE_IN_SMM_BIT);
4943 	} else {
4944 		msr_clear_bit(MSR_IA32_DEBUGCTLMSR,
4945 			      DEBUGCTLMSR_FREEZE_IN_SMM_BIT);
4946 	}
4947 }
4948 
intel_pmu_check_counters_mask(u64 * cntr_mask,u64 * fixed_cntr_mask,u64 * intel_ctrl)4949 static void intel_pmu_check_counters_mask(u64 *cntr_mask,
4950 					  u64 *fixed_cntr_mask,
4951 					  u64 *intel_ctrl)
4952 {
4953 	unsigned int bit;
4954 
4955 	bit = fls64(*cntr_mask);
4956 	if (bit > INTEL_PMC_MAX_GENERIC) {
4957 		WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
4958 		     bit, INTEL_PMC_MAX_GENERIC);
4959 		*cntr_mask &= GENMASK_ULL(INTEL_PMC_MAX_GENERIC - 1, 0);
4960 	}
4961 	*intel_ctrl = *cntr_mask;
4962 
4963 	bit = fls64(*fixed_cntr_mask);
4964 	if (bit > INTEL_PMC_MAX_FIXED) {
4965 		WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
4966 		     bit, INTEL_PMC_MAX_FIXED);
4967 		*fixed_cntr_mask &= GENMASK_ULL(INTEL_PMC_MAX_FIXED - 1, 0);
4968 	}
4969 
4970 	*intel_ctrl |= *fixed_cntr_mask << INTEL_PMC_IDX_FIXED;
4971 }
4972 
4973 static void intel_pmu_check_event_constraints(struct event_constraint *event_constraints,
4974 					      u64 cntr_mask,
4975 					      u64 fixed_cntr_mask,
4976 					      u64 intel_ctrl);
4977 
4978 static void intel_pmu_check_extra_regs(struct extra_reg *extra_regs);
4979 
intel_pmu_broken_perf_cap(void)4980 static inline bool intel_pmu_broken_perf_cap(void)
4981 {
4982 	/* The Perf Metric (Bit 15) is always cleared */
4983 	if (boot_cpu_data.x86_vfm == INTEL_METEORLAKE ||
4984 	    boot_cpu_data.x86_vfm == INTEL_METEORLAKE_L)
4985 		return true;
4986 
4987 	return false;
4988 }
4989 
update_pmu_cap(struct x86_hybrid_pmu * pmu)4990 static void update_pmu_cap(struct x86_hybrid_pmu *pmu)
4991 {
4992 	unsigned int cntr, fixed_cntr, ecx, edx;
4993 	union cpuid35_eax eax;
4994 	union cpuid35_ebx ebx;
4995 
4996 	cpuid(ARCH_PERFMON_EXT_LEAF, &eax.full, &ebx.full, &ecx, &edx);
4997 
4998 	if (ebx.split.umask2)
4999 		pmu->config_mask |= ARCH_PERFMON_EVENTSEL_UMASK2;
5000 	if (ebx.split.eq)
5001 		pmu->config_mask |= ARCH_PERFMON_EVENTSEL_EQ;
5002 
5003 	if (eax.split.cntr_subleaf) {
5004 		cpuid_count(ARCH_PERFMON_EXT_LEAF, ARCH_PERFMON_NUM_COUNTER_LEAF,
5005 			    &cntr, &fixed_cntr, &ecx, &edx);
5006 		pmu->cntr_mask64 = cntr;
5007 		pmu->fixed_cntr_mask64 = fixed_cntr;
5008 	}
5009 
5010 	if (!intel_pmu_broken_perf_cap()) {
5011 		/* Perf Metric (Bit 15) and PEBS via PT (Bit 16) are hybrid enumeration */
5012 		rdmsrl(MSR_IA32_PERF_CAPABILITIES, pmu->intel_cap.capabilities);
5013 	}
5014 }
5015 
intel_pmu_check_hybrid_pmus(struct x86_hybrid_pmu * pmu)5016 static void intel_pmu_check_hybrid_pmus(struct x86_hybrid_pmu *pmu)
5017 {
5018 	intel_pmu_check_counters_mask(&pmu->cntr_mask64, &pmu->fixed_cntr_mask64,
5019 				      &pmu->intel_ctrl);
5020 	pmu->pebs_events_mask = intel_pmu_pebs_mask(pmu->cntr_mask64);
5021 	pmu->unconstrained = (struct event_constraint)
5022 			     __EVENT_CONSTRAINT(0, pmu->cntr_mask64,
5023 						0, x86_pmu_num_counters(&pmu->pmu), 0, 0);
5024 
5025 	if (pmu->intel_cap.perf_metrics)
5026 		pmu->intel_ctrl |= 1ULL << GLOBAL_CTRL_EN_PERF_METRICS;
5027 	else
5028 		pmu->intel_ctrl &= ~(1ULL << GLOBAL_CTRL_EN_PERF_METRICS);
5029 
5030 	intel_pmu_check_event_constraints(pmu->event_constraints,
5031 					  pmu->cntr_mask64,
5032 					  pmu->fixed_cntr_mask64,
5033 					  pmu->intel_ctrl);
5034 
5035 	intel_pmu_check_extra_regs(pmu->extra_regs);
5036 }
5037 
find_hybrid_pmu_for_cpu(void)5038 static struct x86_hybrid_pmu *find_hybrid_pmu_for_cpu(void)
5039 {
5040 	u8 cpu_type = get_this_hybrid_cpu_type();
5041 	int i;
5042 
5043 	/*
5044 	 * This is running on a CPU model that is known to have hybrid
5045 	 * configurations. But the CPU told us it is not hybrid, shame
5046 	 * on it. There should be a fixup function provided for these
5047 	 * troublesome CPUs (->get_hybrid_cpu_type).
5048 	 */
5049 	if (cpu_type == HYBRID_INTEL_NONE) {
5050 		if (x86_pmu.get_hybrid_cpu_type)
5051 			cpu_type = x86_pmu.get_hybrid_cpu_type();
5052 		else
5053 			return NULL;
5054 	}
5055 
5056 	/*
5057 	 * This essentially just maps between the 'hybrid_cpu_type'
5058 	 * and 'hybrid_pmu_type' enums except for ARL-H processor
5059 	 * which needs to compare atom uarch native id since ARL-H
5060 	 * contains two different atom uarchs.
5061 	 */
5062 	for (i = 0; i < x86_pmu.num_hybrid_pmus; i++) {
5063 		enum hybrid_pmu_type pmu_type = x86_pmu.hybrid_pmu[i].pmu_type;
5064 		u32 native_id;
5065 
5066 		if (cpu_type == HYBRID_INTEL_CORE && pmu_type == hybrid_big)
5067 			return &x86_pmu.hybrid_pmu[i];
5068 		if (cpu_type == HYBRID_INTEL_ATOM) {
5069 			if (x86_pmu.num_hybrid_pmus == 2 && pmu_type == hybrid_small)
5070 				return &x86_pmu.hybrid_pmu[i];
5071 
5072 			native_id = get_this_hybrid_cpu_native_id();
5073 			if (native_id == skt_native_id && pmu_type == hybrid_small)
5074 				return &x86_pmu.hybrid_pmu[i];
5075 			if (native_id == cmt_native_id && pmu_type == hybrid_tiny)
5076 				return &x86_pmu.hybrid_pmu[i];
5077 		}
5078 	}
5079 
5080 	return NULL;
5081 }
5082 
init_hybrid_pmu(int cpu)5083 static bool init_hybrid_pmu(int cpu)
5084 {
5085 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
5086 	struct x86_hybrid_pmu *pmu = find_hybrid_pmu_for_cpu();
5087 
5088 	if (WARN_ON_ONCE(!pmu || (pmu->pmu.type == -1))) {
5089 		cpuc->pmu = NULL;
5090 		return false;
5091 	}
5092 
5093 	/* Only check and dump the PMU information for the first CPU */
5094 	if (!cpumask_empty(&pmu->supported_cpus))
5095 		goto end;
5096 
5097 	if (this_cpu_has(X86_FEATURE_ARCH_PERFMON_EXT))
5098 		update_pmu_cap(pmu);
5099 
5100 	intel_pmu_check_hybrid_pmus(pmu);
5101 
5102 	if (!check_hw_exists(&pmu->pmu, pmu->cntr_mask, pmu->fixed_cntr_mask))
5103 		return false;
5104 
5105 	pr_info("%s PMU driver: ", pmu->name);
5106 
5107 	pr_cont("\n");
5108 
5109 	x86_pmu_show_pmu_cap(&pmu->pmu);
5110 
5111 end:
5112 	cpumask_set_cpu(cpu, &pmu->supported_cpus);
5113 	cpuc->pmu = &pmu->pmu;
5114 
5115 	return true;
5116 }
5117 
intel_pmu_cpu_starting(int cpu)5118 static void intel_pmu_cpu_starting(int cpu)
5119 {
5120 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
5121 	int core_id = topology_core_id(cpu);
5122 	int i;
5123 
5124 	if (is_hybrid() && !init_hybrid_pmu(cpu))
5125 		return;
5126 
5127 	init_debug_store_on_cpu(cpu);
5128 	/*
5129 	 * Deal with CPUs that don't clear their LBRs on power-up, and that may
5130 	 * even boot with LBRs enabled.
5131 	 */
5132 	if (!static_cpu_has(X86_FEATURE_ARCH_LBR) && x86_pmu.lbr_nr)
5133 		msr_clear_bit(MSR_IA32_DEBUGCTLMSR, DEBUGCTLMSR_LBR_BIT);
5134 	intel_pmu_lbr_reset();
5135 
5136 	cpuc->lbr_sel = NULL;
5137 
5138 	if (x86_pmu.flags & PMU_FL_TFA) {
5139 		WARN_ON_ONCE(cpuc->tfa_shadow);
5140 		cpuc->tfa_shadow = ~0ULL;
5141 		intel_set_tfa(cpuc, false);
5142 	}
5143 
5144 	if (x86_pmu.version > 1)
5145 		flip_smm_bit(&x86_pmu.attr_freeze_on_smi);
5146 
5147 	/*
5148 	 * Disable perf metrics if any added CPU doesn't support it.
5149 	 *
5150 	 * Turn off the check for a hybrid architecture, because the
5151 	 * architecture MSR, MSR_IA32_PERF_CAPABILITIES, only indicate
5152 	 * the architecture features. The perf metrics is a model-specific
5153 	 * feature for now. The corresponding bit should always be 0 on
5154 	 * a hybrid platform, e.g., Alder Lake.
5155 	 */
5156 	if (!is_hybrid() && x86_pmu.intel_cap.perf_metrics) {
5157 		union perf_capabilities perf_cap;
5158 
5159 		rdmsrl(MSR_IA32_PERF_CAPABILITIES, perf_cap.capabilities);
5160 		if (!perf_cap.perf_metrics) {
5161 			x86_pmu.intel_cap.perf_metrics = 0;
5162 			x86_pmu.intel_ctrl &= ~(1ULL << GLOBAL_CTRL_EN_PERF_METRICS);
5163 		}
5164 	}
5165 
5166 	if (!cpuc->shared_regs)
5167 		return;
5168 
5169 	if (!(x86_pmu.flags & PMU_FL_NO_HT_SHARING)) {
5170 		for_each_cpu(i, topology_sibling_cpumask(cpu)) {
5171 			struct intel_shared_regs *pc;
5172 
5173 			pc = per_cpu(cpu_hw_events, i).shared_regs;
5174 			if (pc && pc->core_id == core_id) {
5175 				cpuc->kfree_on_online[0] = cpuc->shared_regs;
5176 				cpuc->shared_regs = pc;
5177 				break;
5178 			}
5179 		}
5180 		cpuc->shared_regs->core_id = core_id;
5181 		cpuc->shared_regs->refcnt++;
5182 	}
5183 
5184 	if (x86_pmu.lbr_sel_map)
5185 		cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR];
5186 
5187 	if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) {
5188 		for_each_cpu(i, topology_sibling_cpumask(cpu)) {
5189 			struct cpu_hw_events *sibling;
5190 			struct intel_excl_cntrs *c;
5191 
5192 			sibling = &per_cpu(cpu_hw_events, i);
5193 			c = sibling->excl_cntrs;
5194 			if (c && c->core_id == core_id) {
5195 				cpuc->kfree_on_online[1] = cpuc->excl_cntrs;
5196 				cpuc->excl_cntrs = c;
5197 				if (!sibling->excl_thread_id)
5198 					cpuc->excl_thread_id = 1;
5199 				break;
5200 			}
5201 		}
5202 		cpuc->excl_cntrs->core_id = core_id;
5203 		cpuc->excl_cntrs->refcnt++;
5204 	}
5205 }
5206 
free_excl_cntrs(struct cpu_hw_events * cpuc)5207 static void free_excl_cntrs(struct cpu_hw_events *cpuc)
5208 {
5209 	struct intel_excl_cntrs *c;
5210 
5211 	c = cpuc->excl_cntrs;
5212 	if (c) {
5213 		if (c->core_id == -1 || --c->refcnt == 0)
5214 			kfree(c);
5215 		cpuc->excl_cntrs = NULL;
5216 	}
5217 
5218 	kfree(cpuc->constraint_list);
5219 	cpuc->constraint_list = NULL;
5220 }
5221 
intel_pmu_cpu_dying(int cpu)5222 static void intel_pmu_cpu_dying(int cpu)
5223 {
5224 	fini_debug_store_on_cpu(cpu);
5225 }
5226 
intel_cpuc_finish(struct cpu_hw_events * cpuc)5227 void intel_cpuc_finish(struct cpu_hw_events *cpuc)
5228 {
5229 	struct intel_shared_regs *pc;
5230 
5231 	pc = cpuc->shared_regs;
5232 	if (pc) {
5233 		if (pc->core_id == -1 || --pc->refcnt == 0)
5234 			kfree(pc);
5235 		cpuc->shared_regs = NULL;
5236 	}
5237 
5238 	free_excl_cntrs(cpuc);
5239 }
5240 
intel_pmu_cpu_dead(int cpu)5241 static void intel_pmu_cpu_dead(int cpu)
5242 {
5243 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
5244 
5245 	intel_cpuc_finish(cpuc);
5246 
5247 	if (is_hybrid() && cpuc->pmu)
5248 		cpumask_clear_cpu(cpu, &hybrid_pmu(cpuc->pmu)->supported_cpus);
5249 }
5250 
intel_pmu_sched_task(struct perf_event_pmu_context * pmu_ctx,struct task_struct * task,bool sched_in)5251 static void intel_pmu_sched_task(struct perf_event_pmu_context *pmu_ctx,
5252 				 struct task_struct *task, bool sched_in)
5253 {
5254 	intel_pmu_pebs_sched_task(pmu_ctx, sched_in);
5255 	intel_pmu_lbr_sched_task(pmu_ctx, task, sched_in);
5256 }
5257 
intel_pmu_swap_task_ctx(struct perf_event_pmu_context * prev_epc,struct perf_event_pmu_context * next_epc)5258 static void intel_pmu_swap_task_ctx(struct perf_event_pmu_context *prev_epc,
5259 				    struct perf_event_pmu_context *next_epc)
5260 {
5261 	intel_pmu_lbr_swap_task_ctx(prev_epc, next_epc);
5262 }
5263 
intel_pmu_check_period(struct perf_event * event,u64 value)5264 static int intel_pmu_check_period(struct perf_event *event, u64 value)
5265 {
5266 	return intel_pmu_has_bts_period(event, value) ? -EINVAL : 0;
5267 }
5268 
intel_aux_output_init(void)5269 static void intel_aux_output_init(void)
5270 {
5271 	/* Refer also intel_pmu_aux_output_match() */
5272 	if (x86_pmu.intel_cap.pebs_output_pt_available)
5273 		x86_pmu.assign = intel_pmu_assign_event;
5274 }
5275 
intel_pmu_aux_output_match(struct perf_event * event)5276 static int intel_pmu_aux_output_match(struct perf_event *event)
5277 {
5278 	/* intel_pmu_assign_event() is needed, refer intel_aux_output_init() */
5279 	if (!x86_pmu.intel_cap.pebs_output_pt_available)
5280 		return 0;
5281 
5282 	return is_intel_pt_event(event);
5283 }
5284 
intel_pmu_filter(struct pmu * pmu,int cpu,bool * ret)5285 static void intel_pmu_filter(struct pmu *pmu, int cpu, bool *ret)
5286 {
5287 	struct x86_hybrid_pmu *hpmu = hybrid_pmu(pmu);
5288 
5289 	*ret = !cpumask_test_cpu(cpu, &hpmu->supported_cpus);
5290 }
5291 
5292 PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63");
5293 
5294 PMU_FORMAT_ATTR(ldlat, "config1:0-15");
5295 
5296 PMU_FORMAT_ATTR(frontend, "config1:0-23");
5297 
5298 PMU_FORMAT_ATTR(snoop_rsp, "config1:0-63");
5299 
5300 static struct attribute *intel_arch3_formats_attr[] = {
5301 	&format_attr_event.attr,
5302 	&format_attr_umask.attr,
5303 	&format_attr_edge.attr,
5304 	&format_attr_pc.attr,
5305 	&format_attr_any.attr,
5306 	&format_attr_inv.attr,
5307 	&format_attr_cmask.attr,
5308 	NULL,
5309 };
5310 
5311 static struct attribute *hsw_format_attr[] = {
5312 	&format_attr_in_tx.attr,
5313 	&format_attr_in_tx_cp.attr,
5314 	&format_attr_offcore_rsp.attr,
5315 	&format_attr_ldlat.attr,
5316 	NULL
5317 };
5318 
5319 static struct attribute *nhm_format_attr[] = {
5320 	&format_attr_offcore_rsp.attr,
5321 	&format_attr_ldlat.attr,
5322 	NULL
5323 };
5324 
5325 static struct attribute *slm_format_attr[] = {
5326 	&format_attr_offcore_rsp.attr,
5327 	NULL
5328 };
5329 
5330 static struct attribute *cmt_format_attr[] = {
5331 	&format_attr_offcore_rsp.attr,
5332 	&format_attr_ldlat.attr,
5333 	&format_attr_snoop_rsp.attr,
5334 	NULL
5335 };
5336 
5337 static struct attribute *skl_format_attr[] = {
5338 	&format_attr_frontend.attr,
5339 	NULL,
5340 };
5341 
5342 static __initconst const struct x86_pmu core_pmu = {
5343 	.name			= "core",
5344 	.handle_irq		= x86_pmu_handle_irq,
5345 	.disable_all		= x86_pmu_disable_all,
5346 	.enable_all		= core_pmu_enable_all,
5347 	.enable			= core_pmu_enable_event,
5348 	.disable		= x86_pmu_disable_event,
5349 	.hw_config		= core_pmu_hw_config,
5350 	.schedule_events	= x86_schedule_events,
5351 	.eventsel		= MSR_ARCH_PERFMON_EVENTSEL0,
5352 	.perfctr		= MSR_ARCH_PERFMON_PERFCTR0,
5353 	.fixedctr		= MSR_ARCH_PERFMON_FIXED_CTR0,
5354 	.event_map		= intel_pmu_event_map,
5355 	.max_events		= ARRAY_SIZE(intel_perfmon_event_map),
5356 	.apic			= 1,
5357 	.large_pebs_flags	= LARGE_PEBS_FLAGS,
5358 
5359 	/*
5360 	 * Intel PMCs cannot be accessed sanely above 32-bit width,
5361 	 * so we install an artificial 1<<31 period regardless of
5362 	 * the generic event period:
5363 	 */
5364 	.max_period		= (1ULL<<31) - 1,
5365 	.get_event_constraints	= intel_get_event_constraints,
5366 	.put_event_constraints	= intel_put_event_constraints,
5367 	.event_constraints	= intel_core_event_constraints,
5368 	.guest_get_msrs		= core_guest_get_msrs,
5369 	.format_attrs		= intel_arch_formats_attr,
5370 	.events_sysfs_show	= intel_event_sysfs_show,
5371 
5372 	/*
5373 	 * Virtual (or funny metal) CPU can define x86_pmu.extra_regs
5374 	 * together with PMU version 1 and thus be using core_pmu with
5375 	 * shared_regs. We need following callbacks here to allocate
5376 	 * it properly.
5377 	 */
5378 	.cpu_prepare		= intel_pmu_cpu_prepare,
5379 	.cpu_starting		= intel_pmu_cpu_starting,
5380 	.cpu_dying		= intel_pmu_cpu_dying,
5381 	.cpu_dead		= intel_pmu_cpu_dead,
5382 
5383 	.check_period		= intel_pmu_check_period,
5384 
5385 	.lbr_reset		= intel_pmu_lbr_reset_64,
5386 	.lbr_read		= intel_pmu_lbr_read_64,
5387 	.lbr_save		= intel_pmu_lbr_save,
5388 	.lbr_restore		= intel_pmu_lbr_restore,
5389 };
5390 
5391 static __initconst const struct x86_pmu intel_pmu = {
5392 	.name			= "Intel",
5393 	.handle_irq		= intel_pmu_handle_irq,
5394 	.disable_all		= intel_pmu_disable_all,
5395 	.enable_all		= intel_pmu_enable_all,
5396 	.enable			= intel_pmu_enable_event,
5397 	.disable		= intel_pmu_disable_event,
5398 	.add			= intel_pmu_add_event,
5399 	.del			= intel_pmu_del_event,
5400 	.read			= intel_pmu_read_event,
5401 	.set_period		= intel_pmu_set_period,
5402 	.update			= intel_pmu_update,
5403 	.hw_config		= intel_pmu_hw_config,
5404 	.schedule_events	= x86_schedule_events,
5405 	.eventsel		= MSR_ARCH_PERFMON_EVENTSEL0,
5406 	.perfctr		= MSR_ARCH_PERFMON_PERFCTR0,
5407 	.fixedctr		= MSR_ARCH_PERFMON_FIXED_CTR0,
5408 	.event_map		= intel_pmu_event_map,
5409 	.max_events		= ARRAY_SIZE(intel_perfmon_event_map),
5410 	.apic			= 1,
5411 	.large_pebs_flags	= LARGE_PEBS_FLAGS,
5412 	/*
5413 	 * Intel PMCs cannot be accessed sanely above 32 bit width,
5414 	 * so we install an artificial 1<<31 period regardless of
5415 	 * the generic event period:
5416 	 */
5417 	.max_period		= (1ULL << 31) - 1,
5418 	.get_event_constraints	= intel_get_event_constraints,
5419 	.put_event_constraints	= intel_put_event_constraints,
5420 	.pebs_aliases		= intel_pebs_aliases_core2,
5421 
5422 	.format_attrs		= intel_arch3_formats_attr,
5423 	.events_sysfs_show	= intel_event_sysfs_show,
5424 
5425 	.cpu_prepare		= intel_pmu_cpu_prepare,
5426 	.cpu_starting		= intel_pmu_cpu_starting,
5427 	.cpu_dying		= intel_pmu_cpu_dying,
5428 	.cpu_dead		= intel_pmu_cpu_dead,
5429 
5430 	.guest_get_msrs		= intel_guest_get_msrs,
5431 	.sched_task		= intel_pmu_sched_task,
5432 	.swap_task_ctx		= intel_pmu_swap_task_ctx,
5433 
5434 	.check_period		= intel_pmu_check_period,
5435 
5436 	.aux_output_match	= intel_pmu_aux_output_match,
5437 
5438 	.lbr_reset		= intel_pmu_lbr_reset_64,
5439 	.lbr_read		= intel_pmu_lbr_read_64,
5440 	.lbr_save		= intel_pmu_lbr_save,
5441 	.lbr_restore		= intel_pmu_lbr_restore,
5442 
5443 	/*
5444 	 * SMM has access to all 4 rings and while traditionally SMM code only
5445 	 * ran in CPL0, 2021-era firmware is starting to make use of CPL3 in SMM.
5446 	 *
5447 	 * Since the EVENTSEL.{USR,OS} CPL filtering makes no distinction
5448 	 * between SMM or not, this results in what should be pure userspace
5449 	 * counters including SMM data.
5450 	 *
5451 	 * This is a clear privilege issue, therefore globally disable
5452 	 * counting SMM by default.
5453 	 */
5454 	.attr_freeze_on_smi	= 1,
5455 };
5456 
intel_clovertown_quirk(void)5457 static __init void intel_clovertown_quirk(void)
5458 {
5459 	/*
5460 	 * PEBS is unreliable due to:
5461 	 *
5462 	 *   AJ67  - PEBS may experience CPL leaks
5463 	 *   AJ68  - PEBS PMI may be delayed by one event
5464 	 *   AJ69  - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12]
5465 	 *   AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS
5466 	 *
5467 	 * AJ67 could be worked around by restricting the OS/USR flags.
5468 	 * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI.
5469 	 *
5470 	 * AJ106 could possibly be worked around by not allowing LBR
5471 	 *       usage from PEBS, including the fixup.
5472 	 * AJ68  could possibly be worked around by always programming
5473 	 *	 a pebs_event_reset[0] value and coping with the lost events.
5474 	 *
5475 	 * But taken together it might just make sense to not enable PEBS on
5476 	 * these chips.
5477 	 */
5478 	pr_warn("PEBS disabled due to CPU errata\n");
5479 	x86_pmu.pebs = 0;
5480 	x86_pmu.pebs_constraints = NULL;
5481 }
5482 
5483 static const struct x86_cpu_id isolation_ucodes[] = {
5484 	X86_MATCH_VFM_STEPS(INTEL_HASWELL,	 3,  3, 0x0000001f),
5485 	X86_MATCH_VFM_STEPS(INTEL_HASWELL_L,	 1,  1, 0x0000001e),
5486 	X86_MATCH_VFM_STEPS(INTEL_HASWELL_G,	 1,  1, 0x00000015),
5487 	X86_MATCH_VFM_STEPS(INTEL_HASWELL_X,	 2,  2, 0x00000037),
5488 	X86_MATCH_VFM_STEPS(INTEL_HASWELL_X,	 4,  4, 0x0000000a),
5489 	X86_MATCH_VFM_STEPS(INTEL_BROADWELL,	 4,  4, 0x00000023),
5490 	X86_MATCH_VFM_STEPS(INTEL_BROADWELL_G,	 1,  1, 0x00000014),
5491 	X86_MATCH_VFM_STEPS(INTEL_BROADWELL_D,	 2,  2, 0x00000010),
5492 	X86_MATCH_VFM_STEPS(INTEL_BROADWELL_D,	 3,  3, 0x07000009),
5493 	X86_MATCH_VFM_STEPS(INTEL_BROADWELL_D,	 4,  4, 0x0f000009),
5494 	X86_MATCH_VFM_STEPS(INTEL_BROADWELL_D,	 5,  5, 0x0e000002),
5495 	X86_MATCH_VFM_STEPS(INTEL_BROADWELL_X,	 1,  1, 0x0b000014),
5496 	X86_MATCH_VFM_STEPS(INTEL_SKYLAKE_X,	 3,  3, 0x00000021),
5497 	X86_MATCH_VFM_STEPS(INTEL_SKYLAKE_X,	 4,  7, 0x00000000),
5498 	X86_MATCH_VFM_STEPS(INTEL_SKYLAKE_X,	11, 11, 0x00000000),
5499 	X86_MATCH_VFM_STEPS(INTEL_SKYLAKE_L,	 3,  3, 0x0000007c),
5500 	X86_MATCH_VFM_STEPS(INTEL_SKYLAKE,	 3,  3, 0x0000007c),
5501 	X86_MATCH_VFM_STEPS(INTEL_KABYLAKE,	 9, 13, 0x0000004e),
5502 	X86_MATCH_VFM_STEPS(INTEL_KABYLAKE_L,	 9, 12, 0x0000004e),
5503 	{}
5504 };
5505 
intel_check_pebs_isolation(void)5506 static void intel_check_pebs_isolation(void)
5507 {
5508 	x86_pmu.pebs_no_isolation = !x86_match_min_microcode_rev(isolation_ucodes);
5509 }
5510 
intel_pebs_isolation_quirk(void)5511 static __init void intel_pebs_isolation_quirk(void)
5512 {
5513 	WARN_ON_ONCE(x86_pmu.check_microcode);
5514 	x86_pmu.check_microcode = intel_check_pebs_isolation;
5515 	intel_check_pebs_isolation();
5516 }
5517 
5518 static const struct x86_cpu_id pebs_ucodes[] = {
5519 	X86_MATCH_VFM_STEPS(INTEL_SANDYBRIDGE,	7, 7, 0x00000028),
5520 	X86_MATCH_VFM_STEPS(INTEL_SANDYBRIDGE_X,	6, 6, 0x00000618),
5521 	X86_MATCH_VFM_STEPS(INTEL_SANDYBRIDGE_X,	7, 7, 0x0000070c),
5522 	{}
5523 };
5524 
intel_snb_pebs_broken(void)5525 static bool intel_snb_pebs_broken(void)
5526 {
5527 	return !x86_match_min_microcode_rev(pebs_ucodes);
5528 }
5529 
intel_snb_check_microcode(void)5530 static void intel_snb_check_microcode(void)
5531 {
5532 	if (intel_snb_pebs_broken() == x86_pmu.pebs_broken)
5533 		return;
5534 
5535 	/*
5536 	 * Serialized by the microcode lock..
5537 	 */
5538 	if (x86_pmu.pebs_broken) {
5539 		pr_info("PEBS enabled due to microcode update\n");
5540 		x86_pmu.pebs_broken = 0;
5541 	} else {
5542 		pr_info("PEBS disabled due to CPU errata, please upgrade microcode\n");
5543 		x86_pmu.pebs_broken = 1;
5544 	}
5545 }
5546 
is_lbr_from(unsigned long msr)5547 static bool is_lbr_from(unsigned long msr)
5548 {
5549 	unsigned long lbr_from_nr = x86_pmu.lbr_from + x86_pmu.lbr_nr;
5550 
5551 	return x86_pmu.lbr_from <= msr && msr < lbr_from_nr;
5552 }
5553 
5554 /*
5555  * Under certain circumstances, access certain MSR may cause #GP.
5556  * The function tests if the input MSR can be safely accessed.
5557  */
check_msr(unsigned long msr,u64 mask)5558 static bool check_msr(unsigned long msr, u64 mask)
5559 {
5560 	u64 val_old, val_new, val_tmp;
5561 
5562 	/*
5563 	 * Disable the check for real HW, so we don't
5564 	 * mess with potentially enabled registers:
5565 	 */
5566 	if (!boot_cpu_has(X86_FEATURE_HYPERVISOR))
5567 		return true;
5568 
5569 	/*
5570 	 * Read the current value, change it and read it back to see if it
5571 	 * matches, this is needed to detect certain hardware emulators
5572 	 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
5573 	 */
5574 	if (rdmsrl_safe(msr, &val_old))
5575 		return false;
5576 
5577 	/*
5578 	 * Only change the bits which can be updated by wrmsrl.
5579 	 */
5580 	val_tmp = val_old ^ mask;
5581 
5582 	if (is_lbr_from(msr))
5583 		val_tmp = lbr_from_signext_quirk_wr(val_tmp);
5584 
5585 	if (wrmsrl_safe(msr, val_tmp) ||
5586 	    rdmsrl_safe(msr, &val_new))
5587 		return false;
5588 
5589 	/*
5590 	 * Quirk only affects validation in wrmsr(), so wrmsrl()'s value
5591 	 * should equal rdmsrl()'s even with the quirk.
5592 	 */
5593 	if (val_new != val_tmp)
5594 		return false;
5595 
5596 	if (is_lbr_from(msr))
5597 		val_old = lbr_from_signext_quirk_wr(val_old);
5598 
5599 	/* Here it's sure that the MSR can be safely accessed.
5600 	 * Restore the old value and return.
5601 	 */
5602 	wrmsrl(msr, val_old);
5603 
5604 	return true;
5605 }
5606 
intel_sandybridge_quirk(void)5607 static __init void intel_sandybridge_quirk(void)
5608 {
5609 	x86_pmu.check_microcode = intel_snb_check_microcode;
5610 	cpus_read_lock();
5611 	intel_snb_check_microcode();
5612 	cpus_read_unlock();
5613 }
5614 
5615 static const struct { int id; char *name; } intel_arch_events_map[] __initconst = {
5616 	{ PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" },
5617 	{ PERF_COUNT_HW_INSTRUCTIONS, "instructions" },
5618 	{ PERF_COUNT_HW_BUS_CYCLES, "bus cycles" },
5619 	{ PERF_COUNT_HW_CACHE_REFERENCES, "cache references" },
5620 	{ PERF_COUNT_HW_CACHE_MISSES, "cache misses" },
5621 	{ PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" },
5622 	{ PERF_COUNT_HW_BRANCH_MISSES, "branch misses" },
5623 };
5624 
intel_arch_events_quirk(void)5625 static __init void intel_arch_events_quirk(void)
5626 {
5627 	int bit;
5628 
5629 	/* disable event that reported as not present by cpuid */
5630 	for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_arch_events_map)) {
5631 		intel_perfmon_event_map[intel_arch_events_map[bit].id] = 0;
5632 		pr_warn("CPUID marked event: \'%s\' unavailable\n",
5633 			intel_arch_events_map[bit].name);
5634 	}
5635 }
5636 
intel_nehalem_quirk(void)5637 static __init void intel_nehalem_quirk(void)
5638 {
5639 	union cpuid10_ebx ebx;
5640 
5641 	ebx.full = x86_pmu.events_maskl;
5642 	if (ebx.split.no_branch_misses_retired) {
5643 		/*
5644 		 * Erratum AAJ80 detected, we work it around by using
5645 		 * the BR_MISP_EXEC.ANY event. This will over-count
5646 		 * branch-misses, but it's still much better than the
5647 		 * architectural event which is often completely bogus:
5648 		 */
5649 		intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89;
5650 		ebx.split.no_branch_misses_retired = 0;
5651 		x86_pmu.events_maskl = ebx.full;
5652 		pr_info("CPU erratum AAJ80 worked around\n");
5653 	}
5654 }
5655 
5656 /*
5657  * enable software workaround for errata:
5658  * SNB: BJ122
5659  * IVB: BV98
5660  * HSW: HSD29
5661  *
5662  * Only needed when HT is enabled. However detecting
5663  * if HT is enabled is difficult (model specific). So instead,
5664  * we enable the workaround in the early boot, and verify if
5665  * it is needed in a later initcall phase once we have valid
5666  * topology information to check if HT is actually enabled
5667  */
intel_ht_bug(void)5668 static __init void intel_ht_bug(void)
5669 {
5670 	x86_pmu.flags |= PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED;
5671 
5672 	x86_pmu.start_scheduling = intel_start_scheduling;
5673 	x86_pmu.commit_scheduling = intel_commit_scheduling;
5674 	x86_pmu.stop_scheduling = intel_stop_scheduling;
5675 }
5676 
5677 EVENT_ATTR_STR(mem-loads,	mem_ld_hsw,	"event=0xcd,umask=0x1,ldlat=3");
5678 EVENT_ATTR_STR(mem-stores,	mem_st_hsw,	"event=0xd0,umask=0x82")
5679 
5680 /* Haswell special events */
5681 EVENT_ATTR_STR(tx-start,	tx_start,	"event=0xc9,umask=0x1");
5682 EVENT_ATTR_STR(tx-commit,	tx_commit,	"event=0xc9,umask=0x2");
5683 EVENT_ATTR_STR(tx-abort,	tx_abort,	"event=0xc9,umask=0x4");
5684 EVENT_ATTR_STR(tx-capacity,	tx_capacity,	"event=0x54,umask=0x2");
5685 EVENT_ATTR_STR(tx-conflict,	tx_conflict,	"event=0x54,umask=0x1");
5686 EVENT_ATTR_STR(el-start,	el_start,	"event=0xc8,umask=0x1");
5687 EVENT_ATTR_STR(el-commit,	el_commit,	"event=0xc8,umask=0x2");
5688 EVENT_ATTR_STR(el-abort,	el_abort,	"event=0xc8,umask=0x4");
5689 EVENT_ATTR_STR(el-capacity,	el_capacity,	"event=0x54,umask=0x2");
5690 EVENT_ATTR_STR(el-conflict,	el_conflict,	"event=0x54,umask=0x1");
5691 EVENT_ATTR_STR(cycles-t,	cycles_t,	"event=0x3c,in_tx=1");
5692 EVENT_ATTR_STR(cycles-ct,	cycles_ct,	"event=0x3c,in_tx=1,in_tx_cp=1");
5693 
5694 static struct attribute *hsw_events_attrs[] = {
5695 	EVENT_PTR(td_slots_issued),
5696 	EVENT_PTR(td_slots_retired),
5697 	EVENT_PTR(td_fetch_bubbles),
5698 	EVENT_PTR(td_total_slots),
5699 	EVENT_PTR(td_total_slots_scale),
5700 	EVENT_PTR(td_recovery_bubbles),
5701 	EVENT_PTR(td_recovery_bubbles_scale),
5702 	NULL
5703 };
5704 
5705 static struct attribute *hsw_mem_events_attrs[] = {
5706 	EVENT_PTR(mem_ld_hsw),
5707 	EVENT_PTR(mem_st_hsw),
5708 	NULL,
5709 };
5710 
5711 static struct attribute *hsw_tsx_events_attrs[] = {
5712 	EVENT_PTR(tx_start),
5713 	EVENT_PTR(tx_commit),
5714 	EVENT_PTR(tx_abort),
5715 	EVENT_PTR(tx_capacity),
5716 	EVENT_PTR(tx_conflict),
5717 	EVENT_PTR(el_start),
5718 	EVENT_PTR(el_commit),
5719 	EVENT_PTR(el_abort),
5720 	EVENT_PTR(el_capacity),
5721 	EVENT_PTR(el_conflict),
5722 	EVENT_PTR(cycles_t),
5723 	EVENT_PTR(cycles_ct),
5724 	NULL
5725 };
5726 
5727 EVENT_ATTR_STR(tx-capacity-read,  tx_capacity_read,  "event=0x54,umask=0x80");
5728 EVENT_ATTR_STR(tx-capacity-write, tx_capacity_write, "event=0x54,umask=0x2");
5729 EVENT_ATTR_STR(el-capacity-read,  el_capacity_read,  "event=0x54,umask=0x80");
5730 EVENT_ATTR_STR(el-capacity-write, el_capacity_write, "event=0x54,umask=0x2");
5731 
5732 static struct attribute *icl_events_attrs[] = {
5733 	EVENT_PTR(mem_ld_hsw),
5734 	EVENT_PTR(mem_st_hsw),
5735 	NULL,
5736 };
5737 
5738 static struct attribute *icl_td_events_attrs[] = {
5739 	EVENT_PTR(slots),
5740 	EVENT_PTR(td_retiring),
5741 	EVENT_PTR(td_bad_spec),
5742 	EVENT_PTR(td_fe_bound),
5743 	EVENT_PTR(td_be_bound),
5744 	NULL,
5745 };
5746 
5747 static struct attribute *icl_tsx_events_attrs[] = {
5748 	EVENT_PTR(tx_start),
5749 	EVENT_PTR(tx_abort),
5750 	EVENT_PTR(tx_commit),
5751 	EVENT_PTR(tx_capacity_read),
5752 	EVENT_PTR(tx_capacity_write),
5753 	EVENT_PTR(tx_conflict),
5754 	EVENT_PTR(el_start),
5755 	EVENT_PTR(el_abort),
5756 	EVENT_PTR(el_commit),
5757 	EVENT_PTR(el_capacity_read),
5758 	EVENT_PTR(el_capacity_write),
5759 	EVENT_PTR(el_conflict),
5760 	EVENT_PTR(cycles_t),
5761 	EVENT_PTR(cycles_ct),
5762 	NULL,
5763 };
5764 
5765 
5766 EVENT_ATTR_STR(mem-stores,	mem_st_spr,	"event=0xcd,umask=0x2");
5767 EVENT_ATTR_STR(mem-loads-aux,	mem_ld_aux,	"event=0x03,umask=0x82");
5768 
5769 static struct attribute *glc_events_attrs[] = {
5770 	EVENT_PTR(mem_ld_hsw),
5771 	EVENT_PTR(mem_st_spr),
5772 	EVENT_PTR(mem_ld_aux),
5773 	NULL,
5774 };
5775 
5776 static struct attribute *glc_td_events_attrs[] = {
5777 	EVENT_PTR(slots),
5778 	EVENT_PTR(td_retiring),
5779 	EVENT_PTR(td_bad_spec),
5780 	EVENT_PTR(td_fe_bound),
5781 	EVENT_PTR(td_be_bound),
5782 	EVENT_PTR(td_heavy_ops),
5783 	EVENT_PTR(td_br_mispredict),
5784 	EVENT_PTR(td_fetch_lat),
5785 	EVENT_PTR(td_mem_bound),
5786 	NULL,
5787 };
5788 
5789 static struct attribute *glc_tsx_events_attrs[] = {
5790 	EVENT_PTR(tx_start),
5791 	EVENT_PTR(tx_abort),
5792 	EVENT_PTR(tx_commit),
5793 	EVENT_PTR(tx_capacity_read),
5794 	EVENT_PTR(tx_capacity_write),
5795 	EVENT_PTR(tx_conflict),
5796 	EVENT_PTR(cycles_t),
5797 	EVENT_PTR(cycles_ct),
5798 	NULL,
5799 };
5800 
freeze_on_smi_show(struct device * cdev,struct device_attribute * attr,char * buf)5801 static ssize_t freeze_on_smi_show(struct device *cdev,
5802 				  struct device_attribute *attr,
5803 				  char *buf)
5804 {
5805 	return sprintf(buf, "%lu\n", x86_pmu.attr_freeze_on_smi);
5806 }
5807 
5808 static DEFINE_MUTEX(freeze_on_smi_mutex);
5809 
freeze_on_smi_store(struct device * cdev,struct device_attribute * attr,const char * buf,size_t count)5810 static ssize_t freeze_on_smi_store(struct device *cdev,
5811 				   struct device_attribute *attr,
5812 				   const char *buf, size_t count)
5813 {
5814 	unsigned long val;
5815 	ssize_t ret;
5816 
5817 	ret = kstrtoul(buf, 0, &val);
5818 	if (ret)
5819 		return ret;
5820 
5821 	if (val > 1)
5822 		return -EINVAL;
5823 
5824 	mutex_lock(&freeze_on_smi_mutex);
5825 
5826 	if (x86_pmu.attr_freeze_on_smi == val)
5827 		goto done;
5828 
5829 	x86_pmu.attr_freeze_on_smi = val;
5830 
5831 	cpus_read_lock();
5832 	on_each_cpu(flip_smm_bit, &val, 1);
5833 	cpus_read_unlock();
5834 done:
5835 	mutex_unlock(&freeze_on_smi_mutex);
5836 
5837 	return count;
5838 }
5839 
update_tfa_sched(void * ignored)5840 static void update_tfa_sched(void *ignored)
5841 {
5842 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
5843 
5844 	/*
5845 	 * check if PMC3 is used
5846 	 * and if so force schedule out for all event types all contexts
5847 	 */
5848 	if (test_bit(3, cpuc->active_mask))
5849 		perf_pmu_resched(x86_get_pmu(smp_processor_id()));
5850 }
5851 
show_sysctl_tfa(struct device * cdev,struct device_attribute * attr,char * buf)5852 static ssize_t show_sysctl_tfa(struct device *cdev,
5853 			      struct device_attribute *attr,
5854 			      char *buf)
5855 {
5856 	return snprintf(buf, 40, "%d\n", allow_tsx_force_abort);
5857 }
5858 
set_sysctl_tfa(struct device * cdev,struct device_attribute * attr,const char * buf,size_t count)5859 static ssize_t set_sysctl_tfa(struct device *cdev,
5860 			      struct device_attribute *attr,
5861 			      const char *buf, size_t count)
5862 {
5863 	bool val;
5864 	ssize_t ret;
5865 
5866 	ret = kstrtobool(buf, &val);
5867 	if (ret)
5868 		return ret;
5869 
5870 	/* no change */
5871 	if (val == allow_tsx_force_abort)
5872 		return count;
5873 
5874 	allow_tsx_force_abort = val;
5875 
5876 	cpus_read_lock();
5877 	on_each_cpu(update_tfa_sched, NULL, 1);
5878 	cpus_read_unlock();
5879 
5880 	return count;
5881 }
5882 
5883 
5884 static DEVICE_ATTR_RW(freeze_on_smi);
5885 
branches_show(struct device * cdev,struct device_attribute * attr,char * buf)5886 static ssize_t branches_show(struct device *cdev,
5887 			     struct device_attribute *attr,
5888 			     char *buf)
5889 {
5890 	return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu.lbr_nr);
5891 }
5892 
5893 static DEVICE_ATTR_RO(branches);
5894 
branch_counter_nr_show(struct device * cdev,struct device_attribute * attr,char * buf)5895 static ssize_t branch_counter_nr_show(struct device *cdev,
5896 				      struct device_attribute *attr,
5897 				      char *buf)
5898 {
5899 	return snprintf(buf, PAGE_SIZE, "%d\n", fls(x86_pmu.lbr_counters));
5900 }
5901 
5902 static DEVICE_ATTR_RO(branch_counter_nr);
5903 
branch_counter_width_show(struct device * cdev,struct device_attribute * attr,char * buf)5904 static ssize_t branch_counter_width_show(struct device *cdev,
5905 					 struct device_attribute *attr,
5906 					 char *buf)
5907 {
5908 	return snprintf(buf, PAGE_SIZE, "%d\n", LBR_INFO_BR_CNTR_BITS);
5909 }
5910 
5911 static DEVICE_ATTR_RO(branch_counter_width);
5912 
5913 static struct attribute *lbr_attrs[] = {
5914 	&dev_attr_branches.attr,
5915 	&dev_attr_branch_counter_nr.attr,
5916 	&dev_attr_branch_counter_width.attr,
5917 	NULL
5918 };
5919 
5920 static umode_t
lbr_is_visible(struct kobject * kobj,struct attribute * attr,int i)5921 lbr_is_visible(struct kobject *kobj, struct attribute *attr, int i)
5922 {
5923 	/* branches */
5924 	if (i == 0)
5925 		return x86_pmu.lbr_nr ? attr->mode : 0;
5926 
5927 	return (x86_pmu.flags & PMU_FL_BR_CNTR) ? attr->mode : 0;
5928 }
5929 
5930 static char pmu_name_str[30];
5931 
5932 static DEVICE_STRING_ATTR_RO(pmu_name, 0444, pmu_name_str);
5933 
5934 static struct attribute *intel_pmu_caps_attrs[] = {
5935 	&dev_attr_pmu_name.attr.attr,
5936 	NULL
5937 };
5938 
5939 static DEVICE_ATTR(allow_tsx_force_abort, 0644,
5940 		   show_sysctl_tfa,
5941 		   set_sysctl_tfa);
5942 
5943 static struct attribute *intel_pmu_attrs[] = {
5944 	&dev_attr_freeze_on_smi.attr,
5945 	&dev_attr_allow_tsx_force_abort.attr,
5946 	NULL,
5947 };
5948 
5949 static umode_t
default_is_visible(struct kobject * kobj,struct attribute * attr,int i)5950 default_is_visible(struct kobject *kobj, struct attribute *attr, int i)
5951 {
5952 	if (attr == &dev_attr_allow_tsx_force_abort.attr)
5953 		return x86_pmu.flags & PMU_FL_TFA ? attr->mode : 0;
5954 
5955 	return attr->mode;
5956 }
5957 
5958 static umode_t
tsx_is_visible(struct kobject * kobj,struct attribute * attr,int i)5959 tsx_is_visible(struct kobject *kobj, struct attribute *attr, int i)
5960 {
5961 	return boot_cpu_has(X86_FEATURE_RTM) ? attr->mode : 0;
5962 }
5963 
5964 static umode_t
pebs_is_visible(struct kobject * kobj,struct attribute * attr,int i)5965 pebs_is_visible(struct kobject *kobj, struct attribute *attr, int i)
5966 {
5967 	return x86_pmu.pebs ? attr->mode : 0;
5968 }
5969 
5970 static umode_t
mem_is_visible(struct kobject * kobj,struct attribute * attr,int i)5971 mem_is_visible(struct kobject *kobj, struct attribute *attr, int i)
5972 {
5973 	if (attr == &event_attr_mem_ld_aux.attr.attr)
5974 		return x86_pmu.flags & PMU_FL_MEM_LOADS_AUX ? attr->mode : 0;
5975 
5976 	return pebs_is_visible(kobj, attr, i);
5977 }
5978 
5979 static umode_t
exra_is_visible(struct kobject * kobj,struct attribute * attr,int i)5980 exra_is_visible(struct kobject *kobj, struct attribute *attr, int i)
5981 {
5982 	return x86_pmu.version >= 2 ? attr->mode : 0;
5983 }
5984 
5985 static umode_t
td_is_visible(struct kobject * kobj,struct attribute * attr,int i)5986 td_is_visible(struct kobject *kobj, struct attribute *attr, int i)
5987 {
5988 	/*
5989 	 * Hide the perf metrics topdown events
5990 	 * if the feature is not enumerated.
5991 	 */
5992 	if (x86_pmu.num_topdown_events)
5993 		return x86_pmu.intel_cap.perf_metrics ? attr->mode : 0;
5994 
5995 	return attr->mode;
5996 }
5997 
5998 static struct attribute_group group_events_td  = {
5999 	.name = "events",
6000 	.is_visible = td_is_visible,
6001 };
6002 
6003 static struct attribute_group group_events_mem = {
6004 	.name       = "events",
6005 	.is_visible = mem_is_visible,
6006 };
6007 
6008 static struct attribute_group group_events_tsx = {
6009 	.name       = "events",
6010 	.is_visible = tsx_is_visible,
6011 };
6012 
6013 static struct attribute_group group_caps_gen = {
6014 	.name  = "caps",
6015 	.attrs = intel_pmu_caps_attrs,
6016 };
6017 
6018 static struct attribute_group group_caps_lbr = {
6019 	.name       = "caps",
6020 	.attrs	    = lbr_attrs,
6021 	.is_visible = lbr_is_visible,
6022 };
6023 
6024 static struct attribute_group group_format_extra = {
6025 	.name       = "format",
6026 	.is_visible = exra_is_visible,
6027 };
6028 
6029 static struct attribute_group group_format_extra_skl = {
6030 	.name       = "format",
6031 	.is_visible = exra_is_visible,
6032 };
6033 
6034 static struct attribute_group group_format_evtsel_ext = {
6035 	.name       = "format",
6036 	.attrs      = format_evtsel_ext_attrs,
6037 	.is_visible = evtsel_ext_is_visible,
6038 };
6039 
6040 static struct attribute_group group_default = {
6041 	.attrs      = intel_pmu_attrs,
6042 	.is_visible = default_is_visible,
6043 };
6044 
6045 static const struct attribute_group *attr_update[] = {
6046 	&group_events_td,
6047 	&group_events_mem,
6048 	&group_events_tsx,
6049 	&group_caps_gen,
6050 	&group_caps_lbr,
6051 	&group_format_extra,
6052 	&group_format_extra_skl,
6053 	&group_format_evtsel_ext,
6054 	&group_default,
6055 	NULL,
6056 };
6057 
6058 EVENT_ATTR_STR_HYBRID(slots,                 slots_adl,        "event=0x00,umask=0x4",                       hybrid_big);
6059 EVENT_ATTR_STR_HYBRID(topdown-retiring,      td_retiring_adl,  "event=0xc2,umask=0x0;event=0x00,umask=0x80", hybrid_big_small);
6060 EVENT_ATTR_STR_HYBRID(topdown-bad-spec,      td_bad_spec_adl,  "event=0x73,umask=0x0;event=0x00,umask=0x81", hybrid_big_small);
6061 EVENT_ATTR_STR_HYBRID(topdown-fe-bound,      td_fe_bound_adl,  "event=0x71,umask=0x0;event=0x00,umask=0x82", hybrid_big_small);
6062 EVENT_ATTR_STR_HYBRID(topdown-be-bound,      td_be_bound_adl,  "event=0x74,umask=0x0;event=0x00,umask=0x83", hybrid_big_small);
6063 EVENT_ATTR_STR_HYBRID(topdown-heavy-ops,     td_heavy_ops_adl, "event=0x00,umask=0x84",                      hybrid_big);
6064 EVENT_ATTR_STR_HYBRID(topdown-br-mispredict, td_br_mis_adl,    "event=0x00,umask=0x85",                      hybrid_big);
6065 EVENT_ATTR_STR_HYBRID(topdown-fetch-lat,     td_fetch_lat_adl, "event=0x00,umask=0x86",                      hybrid_big);
6066 EVENT_ATTR_STR_HYBRID(topdown-mem-bound,     td_mem_bound_adl, "event=0x00,umask=0x87",                      hybrid_big);
6067 
6068 static struct attribute *adl_hybrid_events_attrs[] = {
6069 	EVENT_PTR(slots_adl),
6070 	EVENT_PTR(td_retiring_adl),
6071 	EVENT_PTR(td_bad_spec_adl),
6072 	EVENT_PTR(td_fe_bound_adl),
6073 	EVENT_PTR(td_be_bound_adl),
6074 	EVENT_PTR(td_heavy_ops_adl),
6075 	EVENT_PTR(td_br_mis_adl),
6076 	EVENT_PTR(td_fetch_lat_adl),
6077 	EVENT_PTR(td_mem_bound_adl),
6078 	NULL,
6079 };
6080 
6081 EVENT_ATTR_STR_HYBRID(topdown-retiring,      td_retiring_lnl,  "event=0xc2,umask=0x02;event=0x00,umask=0x80", hybrid_big_small);
6082 EVENT_ATTR_STR_HYBRID(topdown-fe-bound,      td_fe_bound_lnl,  "event=0x9c,umask=0x01;event=0x00,umask=0x82", hybrid_big_small);
6083 EVENT_ATTR_STR_HYBRID(topdown-be-bound,      td_be_bound_lnl,  "event=0xa4,umask=0x02;event=0x00,umask=0x83", hybrid_big_small);
6084 
6085 static struct attribute *lnl_hybrid_events_attrs[] = {
6086 	EVENT_PTR(slots_adl),
6087 	EVENT_PTR(td_retiring_lnl),
6088 	EVENT_PTR(td_bad_spec_adl),
6089 	EVENT_PTR(td_fe_bound_lnl),
6090 	EVENT_PTR(td_be_bound_lnl),
6091 	EVENT_PTR(td_heavy_ops_adl),
6092 	EVENT_PTR(td_br_mis_adl),
6093 	EVENT_PTR(td_fetch_lat_adl),
6094 	EVENT_PTR(td_mem_bound_adl),
6095 	NULL
6096 };
6097 
6098 /* The event string must be in PMU IDX order. */
6099 EVENT_ATTR_STR_HYBRID(topdown-retiring,
6100 		      td_retiring_arl_h,
6101 		      "event=0xc2,umask=0x02;event=0x00,umask=0x80;event=0xc2,umask=0x0",
6102 		      hybrid_big_small_tiny);
6103 EVENT_ATTR_STR_HYBRID(topdown-bad-spec,
6104 		      td_bad_spec_arl_h,
6105 		      "event=0x73,umask=0x0;event=0x00,umask=0x81;event=0x73,umask=0x0",
6106 		      hybrid_big_small_tiny);
6107 EVENT_ATTR_STR_HYBRID(topdown-fe-bound,
6108 		      td_fe_bound_arl_h,
6109 		      "event=0x9c,umask=0x01;event=0x00,umask=0x82;event=0x71,umask=0x0",
6110 		      hybrid_big_small_tiny);
6111 EVENT_ATTR_STR_HYBRID(topdown-be-bound,
6112 		      td_be_bound_arl_h,
6113 		      "event=0xa4,umask=0x02;event=0x00,umask=0x83;event=0x74,umask=0x0",
6114 		      hybrid_big_small_tiny);
6115 
6116 static struct attribute *arl_h_hybrid_events_attrs[] = {
6117 	EVENT_PTR(slots_adl),
6118 	EVENT_PTR(td_retiring_arl_h),
6119 	EVENT_PTR(td_bad_spec_arl_h),
6120 	EVENT_PTR(td_fe_bound_arl_h),
6121 	EVENT_PTR(td_be_bound_arl_h),
6122 	EVENT_PTR(td_heavy_ops_adl),
6123 	EVENT_PTR(td_br_mis_adl),
6124 	EVENT_PTR(td_fetch_lat_adl),
6125 	EVENT_PTR(td_mem_bound_adl),
6126 	NULL,
6127 };
6128 
6129 /* Must be in IDX order */
6130 EVENT_ATTR_STR_HYBRID(mem-loads,     mem_ld_adl,     "event=0xd0,umask=0x5,ldlat=3;event=0xcd,umask=0x1,ldlat=3", hybrid_big_small);
6131 EVENT_ATTR_STR_HYBRID(mem-stores,    mem_st_adl,     "event=0xd0,umask=0x6;event=0xcd,umask=0x2",                 hybrid_big_small);
6132 EVENT_ATTR_STR_HYBRID(mem-loads-aux, mem_ld_aux_adl, "event=0x03,umask=0x82",                                     hybrid_big);
6133 
6134 static struct attribute *adl_hybrid_mem_attrs[] = {
6135 	EVENT_PTR(mem_ld_adl),
6136 	EVENT_PTR(mem_st_adl),
6137 	EVENT_PTR(mem_ld_aux_adl),
6138 	NULL,
6139 };
6140 
6141 static struct attribute *mtl_hybrid_mem_attrs[] = {
6142 	EVENT_PTR(mem_ld_adl),
6143 	EVENT_PTR(mem_st_adl),
6144 	NULL
6145 };
6146 
6147 EVENT_ATTR_STR_HYBRID(mem-loads,
6148 		      mem_ld_arl_h,
6149 		      "event=0xd0,umask=0x5,ldlat=3;event=0xcd,umask=0x1,ldlat=3;event=0xd0,umask=0x5,ldlat=3",
6150 		      hybrid_big_small_tiny);
6151 EVENT_ATTR_STR_HYBRID(mem-stores,
6152 		      mem_st_arl_h,
6153 		      "event=0xd0,umask=0x6;event=0xcd,umask=0x2;event=0xd0,umask=0x6",
6154 		      hybrid_big_small_tiny);
6155 
6156 static struct attribute *arl_h_hybrid_mem_attrs[] = {
6157 	EVENT_PTR(mem_ld_arl_h),
6158 	EVENT_PTR(mem_st_arl_h),
6159 	NULL,
6160 };
6161 
6162 EVENT_ATTR_STR_HYBRID(tx-start,          tx_start_adl,          "event=0xc9,umask=0x1",          hybrid_big);
6163 EVENT_ATTR_STR_HYBRID(tx-commit,         tx_commit_adl,         "event=0xc9,umask=0x2",          hybrid_big);
6164 EVENT_ATTR_STR_HYBRID(tx-abort,          tx_abort_adl,          "event=0xc9,umask=0x4",          hybrid_big);
6165 EVENT_ATTR_STR_HYBRID(tx-conflict,       tx_conflict_adl,       "event=0x54,umask=0x1",          hybrid_big);
6166 EVENT_ATTR_STR_HYBRID(cycles-t,          cycles_t_adl,          "event=0x3c,in_tx=1",            hybrid_big);
6167 EVENT_ATTR_STR_HYBRID(cycles-ct,         cycles_ct_adl,         "event=0x3c,in_tx=1,in_tx_cp=1", hybrid_big);
6168 EVENT_ATTR_STR_HYBRID(tx-capacity-read,  tx_capacity_read_adl,  "event=0x54,umask=0x80",         hybrid_big);
6169 EVENT_ATTR_STR_HYBRID(tx-capacity-write, tx_capacity_write_adl, "event=0x54,umask=0x2",          hybrid_big);
6170 
6171 static struct attribute *adl_hybrid_tsx_attrs[] = {
6172 	EVENT_PTR(tx_start_adl),
6173 	EVENT_PTR(tx_abort_adl),
6174 	EVENT_PTR(tx_commit_adl),
6175 	EVENT_PTR(tx_capacity_read_adl),
6176 	EVENT_PTR(tx_capacity_write_adl),
6177 	EVENT_PTR(tx_conflict_adl),
6178 	EVENT_PTR(cycles_t_adl),
6179 	EVENT_PTR(cycles_ct_adl),
6180 	NULL,
6181 };
6182 
6183 FORMAT_ATTR_HYBRID(in_tx,       hybrid_big);
6184 FORMAT_ATTR_HYBRID(in_tx_cp,    hybrid_big);
6185 FORMAT_ATTR_HYBRID(offcore_rsp, hybrid_big_small_tiny);
6186 FORMAT_ATTR_HYBRID(ldlat,       hybrid_big_small_tiny);
6187 FORMAT_ATTR_HYBRID(frontend,    hybrid_big);
6188 
6189 #define ADL_HYBRID_RTM_FORMAT_ATTR	\
6190 	FORMAT_HYBRID_PTR(in_tx),	\
6191 	FORMAT_HYBRID_PTR(in_tx_cp)
6192 
6193 #define ADL_HYBRID_FORMAT_ATTR		\
6194 	FORMAT_HYBRID_PTR(offcore_rsp),	\
6195 	FORMAT_HYBRID_PTR(ldlat),	\
6196 	FORMAT_HYBRID_PTR(frontend)
6197 
6198 static struct attribute *adl_hybrid_extra_attr_rtm[] = {
6199 	ADL_HYBRID_RTM_FORMAT_ATTR,
6200 	ADL_HYBRID_FORMAT_ATTR,
6201 	NULL
6202 };
6203 
6204 static struct attribute *adl_hybrid_extra_attr[] = {
6205 	ADL_HYBRID_FORMAT_ATTR,
6206 	NULL
6207 };
6208 
6209 FORMAT_ATTR_HYBRID(snoop_rsp,	hybrid_small_tiny);
6210 
6211 static struct attribute *mtl_hybrid_extra_attr_rtm[] = {
6212 	ADL_HYBRID_RTM_FORMAT_ATTR,
6213 	ADL_HYBRID_FORMAT_ATTR,
6214 	FORMAT_HYBRID_PTR(snoop_rsp),
6215 	NULL
6216 };
6217 
6218 static struct attribute *mtl_hybrid_extra_attr[] = {
6219 	ADL_HYBRID_FORMAT_ATTR,
6220 	FORMAT_HYBRID_PTR(snoop_rsp),
6221 	NULL
6222 };
6223 
is_attr_for_this_pmu(struct kobject * kobj,struct attribute * attr)6224 static bool is_attr_for_this_pmu(struct kobject *kobj, struct attribute *attr)
6225 {
6226 	struct device *dev = kobj_to_dev(kobj);
6227 	struct x86_hybrid_pmu *pmu =
6228 		container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu);
6229 	struct perf_pmu_events_hybrid_attr *pmu_attr =
6230 		container_of(attr, struct perf_pmu_events_hybrid_attr, attr.attr);
6231 
6232 	return pmu->pmu_type & pmu_attr->pmu_type;
6233 }
6234 
hybrid_events_is_visible(struct kobject * kobj,struct attribute * attr,int i)6235 static umode_t hybrid_events_is_visible(struct kobject *kobj,
6236 					struct attribute *attr, int i)
6237 {
6238 	return is_attr_for_this_pmu(kobj, attr) ? attr->mode : 0;
6239 }
6240 
hybrid_find_supported_cpu(struct x86_hybrid_pmu * pmu)6241 static inline int hybrid_find_supported_cpu(struct x86_hybrid_pmu *pmu)
6242 {
6243 	int cpu = cpumask_first(&pmu->supported_cpus);
6244 
6245 	return (cpu >= nr_cpu_ids) ? -1 : cpu;
6246 }
6247 
hybrid_tsx_is_visible(struct kobject * kobj,struct attribute * attr,int i)6248 static umode_t hybrid_tsx_is_visible(struct kobject *kobj,
6249 				     struct attribute *attr, int i)
6250 {
6251 	struct device *dev = kobj_to_dev(kobj);
6252 	struct x86_hybrid_pmu *pmu =
6253 		 container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu);
6254 	int cpu = hybrid_find_supported_cpu(pmu);
6255 
6256 	return (cpu >= 0) && is_attr_for_this_pmu(kobj, attr) && cpu_has(&cpu_data(cpu), X86_FEATURE_RTM) ? attr->mode : 0;
6257 }
6258 
hybrid_format_is_visible(struct kobject * kobj,struct attribute * attr,int i)6259 static umode_t hybrid_format_is_visible(struct kobject *kobj,
6260 					struct attribute *attr, int i)
6261 {
6262 	struct device *dev = kobj_to_dev(kobj);
6263 	struct x86_hybrid_pmu *pmu =
6264 		container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu);
6265 	struct perf_pmu_format_hybrid_attr *pmu_attr =
6266 		container_of(attr, struct perf_pmu_format_hybrid_attr, attr.attr);
6267 	int cpu = hybrid_find_supported_cpu(pmu);
6268 
6269 	return (cpu >= 0) && (pmu->pmu_type & pmu_attr->pmu_type) ? attr->mode : 0;
6270 }
6271 
hybrid_td_is_visible(struct kobject * kobj,struct attribute * attr,int i)6272 static umode_t hybrid_td_is_visible(struct kobject *kobj,
6273 				    struct attribute *attr, int i)
6274 {
6275 	struct device *dev = kobj_to_dev(kobj);
6276 	struct x86_hybrid_pmu *pmu =
6277 		 container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu);
6278 
6279 	if (!is_attr_for_this_pmu(kobj, attr))
6280 		return 0;
6281 
6282 
6283 	/* Only the big core supports perf metrics */
6284 	if (pmu->pmu_type == hybrid_big)
6285 		return pmu->intel_cap.perf_metrics ? attr->mode : 0;
6286 
6287 	return attr->mode;
6288 }
6289 
6290 static struct attribute_group hybrid_group_events_td  = {
6291 	.name		= "events",
6292 	.is_visible	= hybrid_td_is_visible,
6293 };
6294 
6295 static struct attribute_group hybrid_group_events_mem = {
6296 	.name		= "events",
6297 	.is_visible	= hybrid_events_is_visible,
6298 };
6299 
6300 static struct attribute_group hybrid_group_events_tsx = {
6301 	.name		= "events",
6302 	.is_visible	= hybrid_tsx_is_visible,
6303 };
6304 
6305 static struct attribute_group hybrid_group_format_extra = {
6306 	.name		= "format",
6307 	.is_visible	= hybrid_format_is_visible,
6308 };
6309 
intel_hybrid_get_attr_cpus(struct device * dev,struct device_attribute * attr,char * buf)6310 static ssize_t intel_hybrid_get_attr_cpus(struct device *dev,
6311 					  struct device_attribute *attr,
6312 					  char *buf)
6313 {
6314 	struct x86_hybrid_pmu *pmu =
6315 		container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu);
6316 
6317 	return cpumap_print_to_pagebuf(true, buf, &pmu->supported_cpus);
6318 }
6319 
6320 static DEVICE_ATTR(cpus, S_IRUGO, intel_hybrid_get_attr_cpus, NULL);
6321 static struct attribute *intel_hybrid_cpus_attrs[] = {
6322 	&dev_attr_cpus.attr,
6323 	NULL,
6324 };
6325 
6326 static struct attribute_group hybrid_group_cpus = {
6327 	.attrs		= intel_hybrid_cpus_attrs,
6328 };
6329 
6330 static const struct attribute_group *hybrid_attr_update[] = {
6331 	&hybrid_group_events_td,
6332 	&hybrid_group_events_mem,
6333 	&hybrid_group_events_tsx,
6334 	&group_caps_gen,
6335 	&group_caps_lbr,
6336 	&hybrid_group_format_extra,
6337 	&group_format_evtsel_ext,
6338 	&group_default,
6339 	&hybrid_group_cpus,
6340 	NULL,
6341 };
6342 
6343 static struct attribute *empty_attrs;
6344 
intel_pmu_check_event_constraints(struct event_constraint * event_constraints,u64 cntr_mask,u64 fixed_cntr_mask,u64 intel_ctrl)6345 static void intel_pmu_check_event_constraints(struct event_constraint *event_constraints,
6346 					      u64 cntr_mask,
6347 					      u64 fixed_cntr_mask,
6348 					      u64 intel_ctrl)
6349 {
6350 	struct event_constraint *c;
6351 
6352 	if (!event_constraints)
6353 		return;
6354 
6355 	/*
6356 	 * event on fixed counter2 (REF_CYCLES) only works on this
6357 	 * counter, so do not extend mask to generic counters
6358 	 */
6359 	for_each_event_constraint(c, event_constraints) {
6360 		/*
6361 		 * Don't extend the topdown slots and metrics
6362 		 * events to the generic counters.
6363 		 */
6364 		if (c->idxmsk64 & INTEL_PMC_MSK_TOPDOWN) {
6365 			/*
6366 			 * Disable topdown slots and metrics events,
6367 			 * if slots event is not in CPUID.
6368 			 */
6369 			if (!(INTEL_PMC_MSK_FIXED_SLOTS & intel_ctrl))
6370 				c->idxmsk64 = 0;
6371 			c->weight = hweight64(c->idxmsk64);
6372 			continue;
6373 		}
6374 
6375 		if (c->cmask == FIXED_EVENT_FLAGS) {
6376 			/* Disabled fixed counters which are not in CPUID */
6377 			c->idxmsk64 &= intel_ctrl;
6378 
6379 			/*
6380 			 * Don't extend the pseudo-encoding to the
6381 			 * generic counters
6382 			 */
6383 			if (!use_fixed_pseudo_encoding(c->code))
6384 				c->idxmsk64 |= cntr_mask;
6385 		}
6386 		c->idxmsk64 &= cntr_mask | (fixed_cntr_mask << INTEL_PMC_IDX_FIXED);
6387 		c->weight = hweight64(c->idxmsk64);
6388 	}
6389 }
6390 
intel_pmu_check_extra_regs(struct extra_reg * extra_regs)6391 static void intel_pmu_check_extra_regs(struct extra_reg *extra_regs)
6392 {
6393 	struct extra_reg *er;
6394 
6395 	/*
6396 	 * Access extra MSR may cause #GP under certain circumstances.
6397 	 * E.g. KVM doesn't support offcore event
6398 	 * Check all extra_regs here.
6399 	 */
6400 	if (!extra_regs)
6401 		return;
6402 
6403 	for (er = extra_regs; er->msr; er++) {
6404 		er->extra_msr_access = check_msr(er->msr, 0x11UL);
6405 		/* Disable LBR select mapping */
6406 		if ((er->idx == EXTRA_REG_LBR) && !er->extra_msr_access)
6407 			x86_pmu.lbr_sel_map = NULL;
6408 	}
6409 }
6410 
intel_pmu_v6_addr_offset(int index,bool eventsel)6411 static inline int intel_pmu_v6_addr_offset(int index, bool eventsel)
6412 {
6413 	return MSR_IA32_PMC_V6_STEP * index;
6414 }
6415 
6416 static const struct { enum hybrid_pmu_type id; char *name; } intel_hybrid_pmu_type_map[] __initconst = {
6417 	{ hybrid_small,	"cpu_atom" },
6418 	{ hybrid_big,	"cpu_core" },
6419 	{ hybrid_tiny,	"cpu_lowpower" },
6420 };
6421 
intel_pmu_init_hybrid(enum hybrid_pmu_type pmus)6422 static __always_inline int intel_pmu_init_hybrid(enum hybrid_pmu_type pmus)
6423 {
6424 	unsigned long pmus_mask = pmus;
6425 	struct x86_hybrid_pmu *pmu;
6426 	int idx = 0, bit;
6427 
6428 	x86_pmu.num_hybrid_pmus = hweight_long(pmus_mask);
6429 	x86_pmu.hybrid_pmu = kcalloc(x86_pmu.num_hybrid_pmus,
6430 				     sizeof(struct x86_hybrid_pmu),
6431 				     GFP_KERNEL);
6432 	if (!x86_pmu.hybrid_pmu)
6433 		return -ENOMEM;
6434 
6435 	static_branch_enable(&perf_is_hybrid);
6436 	x86_pmu.filter = intel_pmu_filter;
6437 
6438 	for_each_set_bit(bit, &pmus_mask, ARRAY_SIZE(intel_hybrid_pmu_type_map)) {
6439 		pmu = &x86_pmu.hybrid_pmu[idx++];
6440 		pmu->pmu_type = intel_hybrid_pmu_type_map[bit].id;
6441 		pmu->name = intel_hybrid_pmu_type_map[bit].name;
6442 
6443 		pmu->cntr_mask64 = x86_pmu.cntr_mask64;
6444 		pmu->fixed_cntr_mask64 = x86_pmu.fixed_cntr_mask64;
6445 		pmu->pebs_events_mask = intel_pmu_pebs_mask(pmu->cntr_mask64);
6446 		pmu->config_mask = X86_RAW_EVENT_MASK;
6447 		pmu->unconstrained = (struct event_constraint)
6448 				     __EVENT_CONSTRAINT(0, pmu->cntr_mask64,
6449 							0, x86_pmu_num_counters(&pmu->pmu), 0, 0);
6450 
6451 		pmu->intel_cap.capabilities = x86_pmu.intel_cap.capabilities;
6452 		if (pmu->pmu_type & hybrid_small_tiny) {
6453 			pmu->intel_cap.perf_metrics = 0;
6454 			pmu->mid_ack = true;
6455 		} else if (pmu->pmu_type & hybrid_big) {
6456 			pmu->intel_cap.perf_metrics = 1;
6457 			pmu->late_ack = true;
6458 		}
6459 	}
6460 
6461 	return 0;
6462 }
6463 
intel_pmu_ref_cycles_ext(void)6464 static __always_inline void intel_pmu_ref_cycles_ext(void)
6465 {
6466 	if (!(x86_pmu.events_maskl & (INTEL_PMC_MSK_FIXED_REF_CYCLES >> INTEL_PMC_IDX_FIXED)))
6467 		intel_perfmon_event_map[PERF_COUNT_HW_REF_CPU_CYCLES] = 0x013c;
6468 }
6469 
intel_pmu_init_glc(struct pmu * pmu)6470 static __always_inline void intel_pmu_init_glc(struct pmu *pmu)
6471 {
6472 	x86_pmu.late_ack = true;
6473 	x86_pmu.limit_period = glc_limit_period;
6474 	x86_pmu.pebs_aliases = NULL;
6475 	x86_pmu.pebs_prec_dist = true;
6476 	x86_pmu.pebs_block = true;
6477 	x86_pmu.flags |= PMU_FL_HAS_RSP_1;
6478 	x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
6479 	x86_pmu.flags |= PMU_FL_INSTR_LATENCY;
6480 	x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04);
6481 	x86_pmu.lbr_pt_coexist = true;
6482 	x86_pmu.num_topdown_events = 8;
6483 	static_call_update(intel_pmu_update_topdown_event,
6484 			   &icl_update_topdown_event);
6485 	static_call_update(intel_pmu_set_topdown_event_period,
6486 			   &icl_set_topdown_event_period);
6487 
6488 	memcpy(hybrid_var(pmu, hw_cache_event_ids), glc_hw_cache_event_ids, sizeof(hw_cache_event_ids));
6489 	memcpy(hybrid_var(pmu, hw_cache_extra_regs), glc_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
6490 	hybrid(pmu, event_constraints) = intel_glc_event_constraints;
6491 	hybrid(pmu, pebs_constraints) = intel_glc_pebs_event_constraints;
6492 
6493 	intel_pmu_ref_cycles_ext();
6494 }
6495 
intel_pmu_init_grt(struct pmu * pmu)6496 static __always_inline void intel_pmu_init_grt(struct pmu *pmu)
6497 {
6498 	x86_pmu.mid_ack = true;
6499 	x86_pmu.limit_period = glc_limit_period;
6500 	x86_pmu.pebs_aliases = NULL;
6501 	x86_pmu.pebs_prec_dist = true;
6502 	x86_pmu.pebs_block = true;
6503 	x86_pmu.lbr_pt_coexist = true;
6504 	x86_pmu.flags |= PMU_FL_HAS_RSP_1;
6505 	x86_pmu.flags |= PMU_FL_INSTR_LATENCY;
6506 
6507 	memcpy(hybrid_var(pmu, hw_cache_event_ids), glp_hw_cache_event_ids, sizeof(hw_cache_event_ids));
6508 	memcpy(hybrid_var(pmu, hw_cache_extra_regs), tnt_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
6509 	hybrid_var(pmu, hw_cache_event_ids)[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1;
6510 	hybrid(pmu, event_constraints) = intel_grt_event_constraints;
6511 	hybrid(pmu, pebs_constraints) = intel_grt_pebs_event_constraints;
6512 	hybrid(pmu, extra_regs) = intel_grt_extra_regs;
6513 
6514 	intel_pmu_ref_cycles_ext();
6515 }
6516 
intel_pmu_init_lnc(struct pmu * pmu)6517 static __always_inline void intel_pmu_init_lnc(struct pmu *pmu)
6518 {
6519 	intel_pmu_init_glc(pmu);
6520 	hybrid(pmu, event_constraints) = intel_lnc_event_constraints;
6521 	hybrid(pmu, pebs_constraints) = intel_lnc_pebs_event_constraints;
6522 	hybrid(pmu, extra_regs) = intel_lnc_extra_regs;
6523 }
6524 
intel_pmu_init_skt(struct pmu * pmu)6525 static __always_inline void intel_pmu_init_skt(struct pmu *pmu)
6526 {
6527 	intel_pmu_init_grt(pmu);
6528 	hybrid(pmu, event_constraints) = intel_skt_event_constraints;
6529 	hybrid(pmu, extra_regs) = intel_cmt_extra_regs;
6530 }
6531 
intel_pmu_init(void)6532 __init int intel_pmu_init(void)
6533 {
6534 	struct attribute **extra_skl_attr = &empty_attrs;
6535 	struct attribute **extra_attr = &empty_attrs;
6536 	struct attribute **td_attr    = &empty_attrs;
6537 	struct attribute **mem_attr   = &empty_attrs;
6538 	struct attribute **tsx_attr   = &empty_attrs;
6539 	union cpuid10_edx edx;
6540 	union cpuid10_eax eax;
6541 	union cpuid10_ebx ebx;
6542 	unsigned int fixed_mask;
6543 	bool pmem = false;
6544 	int version, i;
6545 	char *name;
6546 	struct x86_hybrid_pmu *pmu;
6547 
6548 	if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
6549 		switch (boot_cpu_data.x86) {
6550 		case 0x6:
6551 			return p6_pmu_init();
6552 		case 0xb:
6553 			return knc_pmu_init();
6554 		case 0xf:
6555 			return p4_pmu_init();
6556 		}
6557 		return -ENODEV;
6558 	}
6559 
6560 	/*
6561 	 * Check whether the Architectural PerfMon supports
6562 	 * Branch Misses Retired hw_event or not.
6563 	 */
6564 	cpuid(10, &eax.full, &ebx.full, &fixed_mask, &edx.full);
6565 	if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT)
6566 		return -ENODEV;
6567 
6568 	version = eax.split.version_id;
6569 	if (version < 2)
6570 		x86_pmu = core_pmu;
6571 	else
6572 		x86_pmu = intel_pmu;
6573 
6574 	x86_pmu.version			= version;
6575 	x86_pmu.cntr_mask64		= GENMASK_ULL(eax.split.num_counters - 1, 0);
6576 	x86_pmu.cntval_bits		= eax.split.bit_width;
6577 	x86_pmu.cntval_mask		= (1ULL << eax.split.bit_width) - 1;
6578 
6579 	x86_pmu.events_maskl		= ebx.full;
6580 	x86_pmu.events_mask_len		= eax.split.mask_length;
6581 
6582 	x86_pmu.pebs_events_mask	= intel_pmu_pebs_mask(x86_pmu.cntr_mask64);
6583 	x86_pmu.pebs_capable		= PEBS_COUNTER_MASK;
6584 
6585 	/*
6586 	 * Quirk: v2 perfmon does not report fixed-purpose events, so
6587 	 * assume at least 3 events, when not running in a hypervisor:
6588 	 */
6589 	if (version > 1 && version < 5) {
6590 		int assume = 3 * !boot_cpu_has(X86_FEATURE_HYPERVISOR);
6591 
6592 		x86_pmu.fixed_cntr_mask64 =
6593 			GENMASK_ULL(max((int)edx.split.num_counters_fixed, assume) - 1, 0);
6594 	} else if (version >= 5)
6595 		x86_pmu.fixed_cntr_mask64 = fixed_mask;
6596 
6597 	if (boot_cpu_has(X86_FEATURE_PDCM)) {
6598 		u64 capabilities;
6599 
6600 		rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities);
6601 		x86_pmu.intel_cap.capabilities = capabilities;
6602 	}
6603 
6604 	if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_32) {
6605 		x86_pmu.lbr_reset = intel_pmu_lbr_reset_32;
6606 		x86_pmu.lbr_read = intel_pmu_lbr_read_32;
6607 	}
6608 
6609 	if (boot_cpu_has(X86_FEATURE_ARCH_LBR))
6610 		intel_pmu_arch_lbr_init();
6611 
6612 	intel_ds_init();
6613 
6614 	x86_add_quirk(intel_arch_events_quirk); /* Install first, so it runs last */
6615 
6616 	if (version >= 5) {
6617 		x86_pmu.intel_cap.anythread_deprecated = edx.split.anythread_deprecated;
6618 		if (x86_pmu.intel_cap.anythread_deprecated)
6619 			pr_cont(" AnyThread deprecated, ");
6620 	}
6621 
6622 	/*
6623 	 * Install the hw-cache-events table:
6624 	 */
6625 	switch (boot_cpu_data.x86_vfm) {
6626 	case INTEL_CORE_YONAH:
6627 		pr_cont("Core events, ");
6628 		name = "core";
6629 		break;
6630 
6631 	case INTEL_CORE2_MEROM:
6632 		x86_add_quirk(intel_clovertown_quirk);
6633 		fallthrough;
6634 
6635 	case INTEL_CORE2_MEROM_L:
6636 	case INTEL_CORE2_PENRYN:
6637 	case INTEL_CORE2_DUNNINGTON:
6638 		memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
6639 		       sizeof(hw_cache_event_ids));
6640 
6641 		intel_pmu_lbr_init_core();
6642 
6643 		x86_pmu.event_constraints = intel_core2_event_constraints;
6644 		x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints;
6645 		pr_cont("Core2 events, ");
6646 		name = "core2";
6647 		break;
6648 
6649 	case INTEL_NEHALEM:
6650 	case INTEL_NEHALEM_EP:
6651 	case INTEL_NEHALEM_EX:
6652 		memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
6653 		       sizeof(hw_cache_event_ids));
6654 		memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
6655 		       sizeof(hw_cache_extra_regs));
6656 
6657 		intel_pmu_lbr_init_nhm();
6658 
6659 		x86_pmu.event_constraints = intel_nehalem_event_constraints;
6660 		x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints;
6661 		x86_pmu.enable_all = intel_pmu_nhm_enable_all;
6662 		x86_pmu.extra_regs = intel_nehalem_extra_regs;
6663 		x86_pmu.limit_period = nhm_limit_period;
6664 
6665 		mem_attr = nhm_mem_events_attrs;
6666 
6667 		/* UOPS_ISSUED.STALLED_CYCLES */
6668 		intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
6669 			X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
6670 		/* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
6671 		intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
6672 			X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);
6673 
6674 		intel_pmu_pebs_data_source_nhm();
6675 		x86_add_quirk(intel_nehalem_quirk);
6676 		x86_pmu.pebs_no_tlb = 1;
6677 		extra_attr = nhm_format_attr;
6678 
6679 		pr_cont("Nehalem events, ");
6680 		name = "nehalem";
6681 		break;
6682 
6683 	case INTEL_ATOM_BONNELL:
6684 	case INTEL_ATOM_BONNELL_MID:
6685 	case INTEL_ATOM_SALTWELL:
6686 	case INTEL_ATOM_SALTWELL_MID:
6687 	case INTEL_ATOM_SALTWELL_TABLET:
6688 		memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
6689 		       sizeof(hw_cache_event_ids));
6690 
6691 		intel_pmu_lbr_init_atom();
6692 
6693 		x86_pmu.event_constraints = intel_gen_event_constraints;
6694 		x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints;
6695 		x86_pmu.pebs_aliases = intel_pebs_aliases_core2;
6696 		pr_cont("Atom events, ");
6697 		name = "bonnell";
6698 		break;
6699 
6700 	case INTEL_ATOM_SILVERMONT:
6701 	case INTEL_ATOM_SILVERMONT_D:
6702 	case INTEL_ATOM_SILVERMONT_MID:
6703 	case INTEL_ATOM_AIRMONT:
6704 	case INTEL_ATOM_AIRMONT_MID:
6705 		memcpy(hw_cache_event_ids, slm_hw_cache_event_ids,
6706 			sizeof(hw_cache_event_ids));
6707 		memcpy(hw_cache_extra_regs, slm_hw_cache_extra_regs,
6708 		       sizeof(hw_cache_extra_regs));
6709 
6710 		intel_pmu_lbr_init_slm();
6711 
6712 		x86_pmu.event_constraints = intel_slm_event_constraints;
6713 		x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints;
6714 		x86_pmu.extra_regs = intel_slm_extra_regs;
6715 		x86_pmu.flags |= PMU_FL_HAS_RSP_1;
6716 		td_attr = slm_events_attrs;
6717 		extra_attr = slm_format_attr;
6718 		pr_cont("Silvermont events, ");
6719 		name = "silvermont";
6720 		break;
6721 
6722 	case INTEL_ATOM_GOLDMONT:
6723 	case INTEL_ATOM_GOLDMONT_D:
6724 		memcpy(hw_cache_event_ids, glm_hw_cache_event_ids,
6725 		       sizeof(hw_cache_event_ids));
6726 		memcpy(hw_cache_extra_regs, glm_hw_cache_extra_regs,
6727 		       sizeof(hw_cache_extra_regs));
6728 
6729 		intel_pmu_lbr_init_skl();
6730 
6731 		x86_pmu.event_constraints = intel_slm_event_constraints;
6732 		x86_pmu.pebs_constraints = intel_glm_pebs_event_constraints;
6733 		x86_pmu.extra_regs = intel_glm_extra_regs;
6734 		/*
6735 		 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS
6736 		 * for precise cycles.
6737 		 * :pp is identical to :ppp
6738 		 */
6739 		x86_pmu.pebs_aliases = NULL;
6740 		x86_pmu.pebs_prec_dist = true;
6741 		x86_pmu.lbr_pt_coexist = true;
6742 		x86_pmu.flags |= PMU_FL_HAS_RSP_1;
6743 		td_attr = glm_events_attrs;
6744 		extra_attr = slm_format_attr;
6745 		pr_cont("Goldmont events, ");
6746 		name = "goldmont";
6747 		break;
6748 
6749 	case INTEL_ATOM_GOLDMONT_PLUS:
6750 		memcpy(hw_cache_event_ids, glp_hw_cache_event_ids,
6751 		       sizeof(hw_cache_event_ids));
6752 		memcpy(hw_cache_extra_regs, glp_hw_cache_extra_regs,
6753 		       sizeof(hw_cache_extra_regs));
6754 
6755 		intel_pmu_lbr_init_skl();
6756 
6757 		x86_pmu.event_constraints = intel_slm_event_constraints;
6758 		x86_pmu.extra_regs = intel_glm_extra_regs;
6759 		/*
6760 		 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS
6761 		 * for precise cycles.
6762 		 */
6763 		x86_pmu.pebs_aliases = NULL;
6764 		x86_pmu.pebs_prec_dist = true;
6765 		x86_pmu.lbr_pt_coexist = true;
6766 		x86_pmu.pebs_capable = ~0ULL;
6767 		x86_pmu.flags |= PMU_FL_HAS_RSP_1;
6768 		x86_pmu.flags |= PMU_FL_PEBS_ALL;
6769 		x86_pmu.get_event_constraints = glp_get_event_constraints;
6770 		td_attr = glm_events_attrs;
6771 		/* Goldmont Plus has 4-wide pipeline */
6772 		event_attr_td_total_slots_scale_glm.event_str = "4";
6773 		extra_attr = slm_format_attr;
6774 		pr_cont("Goldmont plus events, ");
6775 		name = "goldmont_plus";
6776 		break;
6777 
6778 	case INTEL_ATOM_TREMONT_D:
6779 	case INTEL_ATOM_TREMONT:
6780 	case INTEL_ATOM_TREMONT_L:
6781 		x86_pmu.late_ack = true;
6782 		memcpy(hw_cache_event_ids, glp_hw_cache_event_ids,
6783 		       sizeof(hw_cache_event_ids));
6784 		memcpy(hw_cache_extra_regs, tnt_hw_cache_extra_regs,
6785 		       sizeof(hw_cache_extra_regs));
6786 		hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1;
6787 
6788 		intel_pmu_lbr_init_skl();
6789 
6790 		x86_pmu.event_constraints = intel_slm_event_constraints;
6791 		x86_pmu.extra_regs = intel_tnt_extra_regs;
6792 		/*
6793 		 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS
6794 		 * for precise cycles.
6795 		 */
6796 		x86_pmu.pebs_aliases = NULL;
6797 		x86_pmu.pebs_prec_dist = true;
6798 		x86_pmu.lbr_pt_coexist = true;
6799 		x86_pmu.flags |= PMU_FL_HAS_RSP_1;
6800 		x86_pmu.get_event_constraints = tnt_get_event_constraints;
6801 		td_attr = tnt_events_attrs;
6802 		extra_attr = slm_format_attr;
6803 		pr_cont("Tremont events, ");
6804 		name = "Tremont";
6805 		break;
6806 
6807 	case INTEL_ATOM_GRACEMONT:
6808 		intel_pmu_init_grt(NULL);
6809 		intel_pmu_pebs_data_source_grt();
6810 		x86_pmu.pebs_latency_data = grt_latency_data;
6811 		x86_pmu.get_event_constraints = tnt_get_event_constraints;
6812 		td_attr = tnt_events_attrs;
6813 		mem_attr = grt_mem_attrs;
6814 		extra_attr = nhm_format_attr;
6815 		pr_cont("Gracemont events, ");
6816 		name = "gracemont";
6817 		break;
6818 
6819 	case INTEL_ATOM_CRESTMONT:
6820 	case INTEL_ATOM_CRESTMONT_X:
6821 		intel_pmu_init_grt(NULL);
6822 		x86_pmu.extra_regs = intel_cmt_extra_regs;
6823 		intel_pmu_pebs_data_source_cmt();
6824 		x86_pmu.pebs_latency_data = cmt_latency_data;
6825 		x86_pmu.get_event_constraints = cmt_get_event_constraints;
6826 		td_attr = cmt_events_attrs;
6827 		mem_attr = grt_mem_attrs;
6828 		extra_attr = cmt_format_attr;
6829 		pr_cont("Crestmont events, ");
6830 		name = "crestmont";
6831 		break;
6832 
6833 	case INTEL_WESTMERE:
6834 	case INTEL_WESTMERE_EP:
6835 	case INTEL_WESTMERE_EX:
6836 		memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids,
6837 		       sizeof(hw_cache_event_ids));
6838 		memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
6839 		       sizeof(hw_cache_extra_regs));
6840 
6841 		intel_pmu_lbr_init_nhm();
6842 
6843 		x86_pmu.event_constraints = intel_westmere_event_constraints;
6844 		x86_pmu.enable_all = intel_pmu_nhm_enable_all;
6845 		x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints;
6846 		x86_pmu.extra_regs = intel_westmere_extra_regs;
6847 		x86_pmu.flags |= PMU_FL_HAS_RSP_1;
6848 
6849 		mem_attr = nhm_mem_events_attrs;
6850 
6851 		/* UOPS_ISSUED.STALLED_CYCLES */
6852 		intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
6853 			X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
6854 		/* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
6855 		intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
6856 			X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);
6857 
6858 		intel_pmu_pebs_data_source_nhm();
6859 		extra_attr = nhm_format_attr;
6860 		pr_cont("Westmere events, ");
6861 		name = "westmere";
6862 		break;
6863 
6864 	case INTEL_SANDYBRIDGE:
6865 	case INTEL_SANDYBRIDGE_X:
6866 		x86_add_quirk(intel_sandybridge_quirk);
6867 		x86_add_quirk(intel_ht_bug);
6868 		memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
6869 		       sizeof(hw_cache_event_ids));
6870 		memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs,
6871 		       sizeof(hw_cache_extra_regs));
6872 
6873 		intel_pmu_lbr_init_snb();
6874 
6875 		x86_pmu.event_constraints = intel_snb_event_constraints;
6876 		x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints;
6877 		x86_pmu.pebs_aliases = intel_pebs_aliases_snb;
6878 		if (boot_cpu_data.x86_vfm == INTEL_SANDYBRIDGE_X)
6879 			x86_pmu.extra_regs = intel_snbep_extra_regs;
6880 		else
6881 			x86_pmu.extra_regs = intel_snb_extra_regs;
6882 
6883 
6884 		/* all extra regs are per-cpu when HT is on */
6885 		x86_pmu.flags |= PMU_FL_HAS_RSP_1;
6886 		x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
6887 
6888 		td_attr  = snb_events_attrs;
6889 		mem_attr = snb_mem_events_attrs;
6890 
6891 		/* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
6892 		intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
6893 			X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
6894 		/* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/
6895 		intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
6896 			X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1);
6897 
6898 		extra_attr = nhm_format_attr;
6899 
6900 		pr_cont("SandyBridge events, ");
6901 		name = "sandybridge";
6902 		break;
6903 
6904 	case INTEL_IVYBRIDGE:
6905 	case INTEL_IVYBRIDGE_X:
6906 		x86_add_quirk(intel_ht_bug);
6907 		memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
6908 		       sizeof(hw_cache_event_ids));
6909 		/* dTLB-load-misses on IVB is different than SNB */
6910 		hw_cache_event_ids[C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = 0x8108; /* DTLB_LOAD_MISSES.DEMAND_LD_MISS_CAUSES_A_WALK */
6911 
6912 		memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs,
6913 		       sizeof(hw_cache_extra_regs));
6914 
6915 		intel_pmu_lbr_init_snb();
6916 
6917 		x86_pmu.event_constraints = intel_ivb_event_constraints;
6918 		x86_pmu.pebs_constraints = intel_ivb_pebs_event_constraints;
6919 		x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
6920 		x86_pmu.pebs_prec_dist = true;
6921 		if (boot_cpu_data.x86_vfm == INTEL_IVYBRIDGE_X)
6922 			x86_pmu.extra_regs = intel_snbep_extra_regs;
6923 		else
6924 			x86_pmu.extra_regs = intel_snb_extra_regs;
6925 		/* all extra regs are per-cpu when HT is on */
6926 		x86_pmu.flags |= PMU_FL_HAS_RSP_1;
6927 		x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
6928 
6929 		td_attr  = snb_events_attrs;
6930 		mem_attr = snb_mem_events_attrs;
6931 
6932 		/* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
6933 		intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
6934 			X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
6935 
6936 		extra_attr = nhm_format_attr;
6937 
6938 		pr_cont("IvyBridge events, ");
6939 		name = "ivybridge";
6940 		break;
6941 
6942 
6943 	case INTEL_HASWELL:
6944 	case INTEL_HASWELL_X:
6945 	case INTEL_HASWELL_L:
6946 	case INTEL_HASWELL_G:
6947 		x86_add_quirk(intel_ht_bug);
6948 		x86_add_quirk(intel_pebs_isolation_quirk);
6949 		x86_pmu.late_ack = true;
6950 		memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids));
6951 		memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
6952 
6953 		intel_pmu_lbr_init_hsw();
6954 
6955 		x86_pmu.event_constraints = intel_hsw_event_constraints;
6956 		x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints;
6957 		x86_pmu.extra_regs = intel_snbep_extra_regs;
6958 		x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
6959 		x86_pmu.pebs_prec_dist = true;
6960 		/* all extra regs are per-cpu when HT is on */
6961 		x86_pmu.flags |= PMU_FL_HAS_RSP_1;
6962 		x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
6963 
6964 		x86_pmu.hw_config = hsw_hw_config;
6965 		x86_pmu.get_event_constraints = hsw_get_event_constraints;
6966 		x86_pmu.limit_period = hsw_limit_period;
6967 		x86_pmu.lbr_double_abort = true;
6968 		extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
6969 			hsw_format_attr : nhm_format_attr;
6970 		td_attr  = hsw_events_attrs;
6971 		mem_attr = hsw_mem_events_attrs;
6972 		tsx_attr = hsw_tsx_events_attrs;
6973 		pr_cont("Haswell events, ");
6974 		name = "haswell";
6975 		break;
6976 
6977 	case INTEL_BROADWELL:
6978 	case INTEL_BROADWELL_D:
6979 	case INTEL_BROADWELL_G:
6980 	case INTEL_BROADWELL_X:
6981 		x86_add_quirk(intel_pebs_isolation_quirk);
6982 		x86_pmu.late_ack = true;
6983 		memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids));
6984 		memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
6985 
6986 		/* L3_MISS_LOCAL_DRAM is BIT(26) in Broadwell */
6987 		hw_cache_extra_regs[C(LL)][C(OP_READ)][C(RESULT_MISS)] = HSW_DEMAND_READ |
6988 									 BDW_L3_MISS|HSW_SNOOP_DRAM;
6989 		hw_cache_extra_regs[C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = HSW_DEMAND_WRITE|BDW_L3_MISS|
6990 									  HSW_SNOOP_DRAM;
6991 		hw_cache_extra_regs[C(NODE)][C(OP_READ)][C(RESULT_ACCESS)] = HSW_DEMAND_READ|
6992 									     BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM;
6993 		hw_cache_extra_regs[C(NODE)][C(OP_WRITE)][C(RESULT_ACCESS)] = HSW_DEMAND_WRITE|
6994 									      BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM;
6995 
6996 		intel_pmu_lbr_init_hsw();
6997 
6998 		x86_pmu.event_constraints = intel_bdw_event_constraints;
6999 		x86_pmu.pebs_constraints = intel_bdw_pebs_event_constraints;
7000 		x86_pmu.extra_regs = intel_snbep_extra_regs;
7001 		x86_pmu.pebs_aliases = intel_pebs_aliases_ivb;
7002 		x86_pmu.pebs_prec_dist = true;
7003 		/* all extra regs are per-cpu when HT is on */
7004 		x86_pmu.flags |= PMU_FL_HAS_RSP_1;
7005 		x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
7006 
7007 		x86_pmu.hw_config = hsw_hw_config;
7008 		x86_pmu.get_event_constraints = hsw_get_event_constraints;
7009 		x86_pmu.limit_period = bdw_limit_period;
7010 		extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
7011 			hsw_format_attr : nhm_format_attr;
7012 		td_attr  = hsw_events_attrs;
7013 		mem_attr = hsw_mem_events_attrs;
7014 		tsx_attr = hsw_tsx_events_attrs;
7015 		pr_cont("Broadwell events, ");
7016 		name = "broadwell";
7017 		break;
7018 
7019 	case INTEL_XEON_PHI_KNL:
7020 	case INTEL_XEON_PHI_KNM:
7021 		memcpy(hw_cache_event_ids,
7022 		       slm_hw_cache_event_ids, sizeof(hw_cache_event_ids));
7023 		memcpy(hw_cache_extra_regs,
7024 		       knl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
7025 		intel_pmu_lbr_init_knl();
7026 
7027 		x86_pmu.event_constraints = intel_slm_event_constraints;
7028 		x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints;
7029 		x86_pmu.extra_regs = intel_knl_extra_regs;
7030 
7031 		/* all extra regs are per-cpu when HT is on */
7032 		x86_pmu.flags |= PMU_FL_HAS_RSP_1;
7033 		x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
7034 		extra_attr = slm_format_attr;
7035 		pr_cont("Knights Landing/Mill events, ");
7036 		name = "knights-landing";
7037 		break;
7038 
7039 	case INTEL_SKYLAKE_X:
7040 		pmem = true;
7041 		fallthrough;
7042 	case INTEL_SKYLAKE_L:
7043 	case INTEL_SKYLAKE:
7044 	case INTEL_KABYLAKE_L:
7045 	case INTEL_KABYLAKE:
7046 	case INTEL_COMETLAKE_L:
7047 	case INTEL_COMETLAKE:
7048 		x86_add_quirk(intel_pebs_isolation_quirk);
7049 		x86_pmu.late_ack = true;
7050 		memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids));
7051 		memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
7052 		intel_pmu_lbr_init_skl();
7053 
7054 		/* INT_MISC.RECOVERY_CYCLES has umask 1 in Skylake */
7055 		event_attr_td_recovery_bubbles.event_str_noht =
7056 			"event=0xd,umask=0x1,cmask=1";
7057 		event_attr_td_recovery_bubbles.event_str_ht =
7058 			"event=0xd,umask=0x1,cmask=1,any=1";
7059 
7060 		x86_pmu.event_constraints = intel_skl_event_constraints;
7061 		x86_pmu.pebs_constraints = intel_skl_pebs_event_constraints;
7062 		x86_pmu.extra_regs = intel_skl_extra_regs;
7063 		x86_pmu.pebs_aliases = intel_pebs_aliases_skl;
7064 		x86_pmu.pebs_prec_dist = true;
7065 		/* all extra regs are per-cpu when HT is on */
7066 		x86_pmu.flags |= PMU_FL_HAS_RSP_1;
7067 		x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
7068 
7069 		x86_pmu.hw_config = hsw_hw_config;
7070 		x86_pmu.get_event_constraints = hsw_get_event_constraints;
7071 		extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
7072 			hsw_format_attr : nhm_format_attr;
7073 		extra_skl_attr = skl_format_attr;
7074 		td_attr  = hsw_events_attrs;
7075 		mem_attr = hsw_mem_events_attrs;
7076 		tsx_attr = hsw_tsx_events_attrs;
7077 		intel_pmu_pebs_data_source_skl(pmem);
7078 
7079 		/*
7080 		 * Processors with CPUID.RTM_ALWAYS_ABORT have TSX deprecated by default.
7081 		 * TSX force abort hooks are not required on these systems. Only deploy
7082 		 * workaround when microcode has not enabled X86_FEATURE_RTM_ALWAYS_ABORT.
7083 		 */
7084 		if (boot_cpu_has(X86_FEATURE_TSX_FORCE_ABORT) &&
7085 		   !boot_cpu_has(X86_FEATURE_RTM_ALWAYS_ABORT)) {
7086 			x86_pmu.flags |= PMU_FL_TFA;
7087 			x86_pmu.get_event_constraints = tfa_get_event_constraints;
7088 			x86_pmu.enable_all = intel_tfa_pmu_enable_all;
7089 			x86_pmu.commit_scheduling = intel_tfa_commit_scheduling;
7090 		}
7091 
7092 		pr_cont("Skylake events, ");
7093 		name = "skylake";
7094 		break;
7095 
7096 	case INTEL_ICELAKE_X:
7097 	case INTEL_ICELAKE_D:
7098 		x86_pmu.pebs_ept = 1;
7099 		pmem = true;
7100 		fallthrough;
7101 	case INTEL_ICELAKE_L:
7102 	case INTEL_ICELAKE:
7103 	case INTEL_TIGERLAKE_L:
7104 	case INTEL_TIGERLAKE:
7105 	case INTEL_ROCKETLAKE:
7106 		x86_pmu.late_ack = true;
7107 		memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids));
7108 		memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs));
7109 		hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1;
7110 		intel_pmu_lbr_init_skl();
7111 
7112 		x86_pmu.event_constraints = intel_icl_event_constraints;
7113 		x86_pmu.pebs_constraints = intel_icl_pebs_event_constraints;
7114 		x86_pmu.extra_regs = intel_icl_extra_regs;
7115 		x86_pmu.pebs_aliases = NULL;
7116 		x86_pmu.pebs_prec_dist = true;
7117 		x86_pmu.flags |= PMU_FL_HAS_RSP_1;
7118 		x86_pmu.flags |= PMU_FL_NO_HT_SHARING;
7119 
7120 		x86_pmu.hw_config = hsw_hw_config;
7121 		x86_pmu.get_event_constraints = icl_get_event_constraints;
7122 		extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
7123 			hsw_format_attr : nhm_format_attr;
7124 		extra_skl_attr = skl_format_attr;
7125 		mem_attr = icl_events_attrs;
7126 		td_attr = icl_td_events_attrs;
7127 		tsx_attr = icl_tsx_events_attrs;
7128 		x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04);
7129 		x86_pmu.lbr_pt_coexist = true;
7130 		intel_pmu_pebs_data_source_skl(pmem);
7131 		x86_pmu.num_topdown_events = 4;
7132 		static_call_update(intel_pmu_update_topdown_event,
7133 				   &icl_update_topdown_event);
7134 		static_call_update(intel_pmu_set_topdown_event_period,
7135 				   &icl_set_topdown_event_period);
7136 		pr_cont("Icelake events, ");
7137 		name = "icelake";
7138 		break;
7139 
7140 	case INTEL_SAPPHIRERAPIDS_X:
7141 	case INTEL_EMERALDRAPIDS_X:
7142 		x86_pmu.flags |= PMU_FL_MEM_LOADS_AUX;
7143 		x86_pmu.extra_regs = intel_glc_extra_regs;
7144 		pr_cont("Sapphire Rapids events, ");
7145 		name = "sapphire_rapids";
7146 		goto glc_common;
7147 
7148 	case INTEL_GRANITERAPIDS_X:
7149 	case INTEL_GRANITERAPIDS_D:
7150 		x86_pmu.extra_regs = intel_rwc_extra_regs;
7151 		pr_cont("Granite Rapids events, ");
7152 		name = "granite_rapids";
7153 
7154 	glc_common:
7155 		intel_pmu_init_glc(NULL);
7156 		x86_pmu.pebs_ept = 1;
7157 		x86_pmu.hw_config = hsw_hw_config;
7158 		x86_pmu.get_event_constraints = glc_get_event_constraints;
7159 		extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
7160 			hsw_format_attr : nhm_format_attr;
7161 		extra_skl_attr = skl_format_attr;
7162 		mem_attr = glc_events_attrs;
7163 		td_attr = glc_td_events_attrs;
7164 		tsx_attr = glc_tsx_events_attrs;
7165 		intel_pmu_pebs_data_source_skl(true);
7166 		break;
7167 
7168 	case INTEL_ALDERLAKE:
7169 	case INTEL_ALDERLAKE_L:
7170 	case INTEL_RAPTORLAKE:
7171 	case INTEL_RAPTORLAKE_P:
7172 	case INTEL_RAPTORLAKE_S:
7173 		/*
7174 		 * Alder Lake has 2 types of CPU, core and atom.
7175 		 *
7176 		 * Initialize the common PerfMon capabilities here.
7177 		 */
7178 		intel_pmu_init_hybrid(hybrid_big_small);
7179 
7180 		x86_pmu.pebs_latency_data = grt_latency_data;
7181 		x86_pmu.get_event_constraints = adl_get_event_constraints;
7182 		x86_pmu.hw_config = adl_hw_config;
7183 		x86_pmu.get_hybrid_cpu_type = adl_get_hybrid_cpu_type;
7184 
7185 		td_attr = adl_hybrid_events_attrs;
7186 		mem_attr = adl_hybrid_mem_attrs;
7187 		tsx_attr = adl_hybrid_tsx_attrs;
7188 		extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
7189 			adl_hybrid_extra_attr_rtm : adl_hybrid_extra_attr;
7190 
7191 		/* Initialize big core specific PerfMon capabilities.*/
7192 		pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX];
7193 		intel_pmu_init_glc(&pmu->pmu);
7194 		if (cpu_feature_enabled(X86_FEATURE_HYBRID_CPU)) {
7195 			pmu->cntr_mask64 <<= 2;
7196 			pmu->cntr_mask64 |= 0x3;
7197 			pmu->fixed_cntr_mask64 <<= 1;
7198 			pmu->fixed_cntr_mask64 |= 0x1;
7199 		} else {
7200 			pmu->cntr_mask64 = x86_pmu.cntr_mask64;
7201 			pmu->fixed_cntr_mask64 = x86_pmu.fixed_cntr_mask64;
7202 		}
7203 
7204 		/*
7205 		 * Quirk: For some Alder Lake machine, when all E-cores are disabled in
7206 		 * a BIOS, the leaf 0xA will enumerate all counters of P-cores. However,
7207 		 * the X86_FEATURE_HYBRID_CPU is still set. The above codes will
7208 		 * mistakenly add extra counters for P-cores. Correct the number of
7209 		 * counters here.
7210 		 */
7211 		if ((x86_pmu_num_counters(&pmu->pmu) > 8) || (x86_pmu_num_counters_fixed(&pmu->pmu) > 4)) {
7212 			pmu->cntr_mask64 = x86_pmu.cntr_mask64;
7213 			pmu->fixed_cntr_mask64 = x86_pmu.fixed_cntr_mask64;
7214 		}
7215 
7216 		pmu->pebs_events_mask = intel_pmu_pebs_mask(pmu->cntr_mask64);
7217 		pmu->unconstrained = (struct event_constraint)
7218 				     __EVENT_CONSTRAINT(0, pmu->cntr_mask64,
7219 				     0, x86_pmu_num_counters(&pmu->pmu), 0, 0);
7220 
7221 		pmu->extra_regs = intel_glc_extra_regs;
7222 
7223 		/* Initialize Atom core specific PerfMon capabilities.*/
7224 		pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX];
7225 		intel_pmu_init_grt(&pmu->pmu);
7226 
7227 		x86_pmu.flags |= PMU_FL_MEM_LOADS_AUX;
7228 		intel_pmu_pebs_data_source_adl();
7229 		pr_cont("Alderlake Hybrid events, ");
7230 		name = "alderlake_hybrid";
7231 		break;
7232 
7233 	case INTEL_METEORLAKE:
7234 	case INTEL_METEORLAKE_L:
7235 	case INTEL_ARROWLAKE_U:
7236 		intel_pmu_init_hybrid(hybrid_big_small);
7237 
7238 		x86_pmu.pebs_latency_data = cmt_latency_data;
7239 		x86_pmu.get_event_constraints = mtl_get_event_constraints;
7240 		x86_pmu.hw_config = adl_hw_config;
7241 
7242 		td_attr = adl_hybrid_events_attrs;
7243 		mem_attr = mtl_hybrid_mem_attrs;
7244 		tsx_attr = adl_hybrid_tsx_attrs;
7245 		extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
7246 			mtl_hybrid_extra_attr_rtm : mtl_hybrid_extra_attr;
7247 
7248 		/* Initialize big core specific PerfMon capabilities.*/
7249 		pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX];
7250 		intel_pmu_init_glc(&pmu->pmu);
7251 		pmu->extra_regs = intel_rwc_extra_regs;
7252 
7253 		/* Initialize Atom core specific PerfMon capabilities.*/
7254 		pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX];
7255 		intel_pmu_init_grt(&pmu->pmu);
7256 		pmu->extra_regs = intel_cmt_extra_regs;
7257 
7258 		intel_pmu_pebs_data_source_mtl();
7259 		pr_cont("Meteorlake Hybrid events, ");
7260 		name = "meteorlake_hybrid";
7261 		break;
7262 
7263 	case INTEL_LUNARLAKE_M:
7264 	case INTEL_ARROWLAKE:
7265 		intel_pmu_init_hybrid(hybrid_big_small);
7266 
7267 		x86_pmu.pebs_latency_data = lnl_latency_data;
7268 		x86_pmu.get_event_constraints = mtl_get_event_constraints;
7269 		x86_pmu.hw_config = adl_hw_config;
7270 
7271 		td_attr = lnl_hybrid_events_attrs;
7272 		mem_attr = mtl_hybrid_mem_attrs;
7273 		tsx_attr = adl_hybrid_tsx_attrs;
7274 		extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
7275 			mtl_hybrid_extra_attr_rtm : mtl_hybrid_extra_attr;
7276 
7277 		/* Initialize big core specific PerfMon capabilities.*/
7278 		pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX];
7279 		intel_pmu_init_lnc(&pmu->pmu);
7280 
7281 		/* Initialize Atom core specific PerfMon capabilities.*/
7282 		pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX];
7283 		intel_pmu_init_skt(&pmu->pmu);
7284 
7285 		intel_pmu_pebs_data_source_lnl();
7286 		pr_cont("Lunarlake Hybrid events, ");
7287 		name = "lunarlake_hybrid";
7288 		break;
7289 
7290 	case INTEL_ARROWLAKE_H:
7291 		intel_pmu_init_hybrid(hybrid_big_small_tiny);
7292 
7293 		x86_pmu.pebs_latency_data = arl_h_latency_data;
7294 		x86_pmu.get_event_constraints = arl_h_get_event_constraints;
7295 		x86_pmu.hw_config = arl_h_hw_config;
7296 
7297 		td_attr = arl_h_hybrid_events_attrs;
7298 		mem_attr = arl_h_hybrid_mem_attrs;
7299 		tsx_attr = adl_hybrid_tsx_attrs;
7300 		extra_attr = boot_cpu_has(X86_FEATURE_RTM) ?
7301 			mtl_hybrid_extra_attr_rtm : mtl_hybrid_extra_attr;
7302 
7303 		/* Initialize big core specific PerfMon capabilities. */
7304 		pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX];
7305 		intel_pmu_init_lnc(&pmu->pmu);
7306 
7307 		/* Initialize Atom core specific PerfMon capabilities. */
7308 		pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX];
7309 		intel_pmu_init_skt(&pmu->pmu);
7310 
7311 		/* Initialize Lower Power Atom specific PerfMon capabilities. */
7312 		pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_TINY_IDX];
7313 		intel_pmu_init_grt(&pmu->pmu);
7314 		pmu->extra_regs = intel_cmt_extra_regs;
7315 
7316 		intel_pmu_pebs_data_source_arl_h();
7317 		pr_cont("ArrowLake-H Hybrid events, ");
7318 		name = "arrowlake_h_hybrid";
7319 		break;
7320 
7321 	default:
7322 		switch (x86_pmu.version) {
7323 		case 1:
7324 			x86_pmu.event_constraints = intel_v1_event_constraints;
7325 			pr_cont("generic architected perfmon v1, ");
7326 			name = "generic_arch_v1";
7327 			break;
7328 		case 2:
7329 		case 3:
7330 		case 4:
7331 			/*
7332 			 * default constraints for v2 and up
7333 			 */
7334 			x86_pmu.event_constraints = intel_gen_event_constraints;
7335 			pr_cont("generic architected perfmon, ");
7336 			name = "generic_arch_v2+";
7337 			break;
7338 		default:
7339 			/*
7340 			 * The default constraints for v5 and up can support up to
7341 			 * 16 fixed counters. For the fixed counters 4 and later,
7342 			 * the pseudo-encoding is applied.
7343 			 * The constraints may be cut according to the CPUID enumeration
7344 			 * by inserting the EVENT_CONSTRAINT_END.
7345 			 */
7346 			if (fls64(x86_pmu.fixed_cntr_mask64) > INTEL_PMC_MAX_FIXED)
7347 				x86_pmu.fixed_cntr_mask64 &= GENMASK_ULL(INTEL_PMC_MAX_FIXED - 1, 0);
7348 			intel_v5_gen_event_constraints[fls64(x86_pmu.fixed_cntr_mask64)].weight = -1;
7349 			x86_pmu.event_constraints = intel_v5_gen_event_constraints;
7350 			pr_cont("generic architected perfmon, ");
7351 			name = "generic_arch_v5+";
7352 			break;
7353 		}
7354 	}
7355 
7356 	snprintf(pmu_name_str, sizeof(pmu_name_str), "%s", name);
7357 
7358 	if (!is_hybrid()) {
7359 		group_events_td.attrs  = td_attr;
7360 		group_events_mem.attrs = mem_attr;
7361 		group_events_tsx.attrs = tsx_attr;
7362 		group_format_extra.attrs = extra_attr;
7363 		group_format_extra_skl.attrs = extra_skl_attr;
7364 
7365 		x86_pmu.attr_update = attr_update;
7366 	} else {
7367 		hybrid_group_events_td.attrs  = td_attr;
7368 		hybrid_group_events_mem.attrs = mem_attr;
7369 		hybrid_group_events_tsx.attrs = tsx_attr;
7370 		hybrid_group_format_extra.attrs = extra_attr;
7371 
7372 		x86_pmu.attr_update = hybrid_attr_update;
7373 	}
7374 
7375 	intel_pmu_check_counters_mask(&x86_pmu.cntr_mask64,
7376 				      &x86_pmu.fixed_cntr_mask64,
7377 				      &x86_pmu.intel_ctrl);
7378 
7379 	/* AnyThread may be deprecated on arch perfmon v5 or later */
7380 	if (x86_pmu.intel_cap.anythread_deprecated)
7381 		x86_pmu.format_attrs = intel_arch_formats_attr;
7382 
7383 	intel_pmu_check_event_constraints(x86_pmu.event_constraints,
7384 					  x86_pmu.cntr_mask64,
7385 					  x86_pmu.fixed_cntr_mask64,
7386 					  x86_pmu.intel_ctrl);
7387 	/*
7388 	 * Access LBR MSR may cause #GP under certain circumstances.
7389 	 * Check all LBR MSR here.
7390 	 * Disable LBR access if any LBR MSRs can not be accessed.
7391 	 */
7392 	if (x86_pmu.lbr_tos && !check_msr(x86_pmu.lbr_tos, 0x3UL))
7393 		x86_pmu.lbr_nr = 0;
7394 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
7395 		if (!(check_msr(x86_pmu.lbr_from + i, 0xffffUL) &&
7396 		      check_msr(x86_pmu.lbr_to + i, 0xffffUL)))
7397 			x86_pmu.lbr_nr = 0;
7398 	}
7399 
7400 	if (x86_pmu.lbr_nr) {
7401 		intel_pmu_lbr_init();
7402 
7403 		pr_cont("%d-deep LBR, ", x86_pmu.lbr_nr);
7404 
7405 		/* only support branch_stack snapshot for perfmon >= v2 */
7406 		if (x86_pmu.disable_all == intel_pmu_disable_all) {
7407 			if (boot_cpu_has(X86_FEATURE_ARCH_LBR)) {
7408 				static_call_update(perf_snapshot_branch_stack,
7409 						   intel_pmu_snapshot_arch_branch_stack);
7410 			} else {
7411 				static_call_update(perf_snapshot_branch_stack,
7412 						   intel_pmu_snapshot_branch_stack);
7413 			}
7414 		}
7415 	}
7416 
7417 	intel_pmu_check_extra_regs(x86_pmu.extra_regs);
7418 
7419 	/* Support full width counters using alternative MSR range */
7420 	if (x86_pmu.intel_cap.full_width_write) {
7421 		x86_pmu.max_period = x86_pmu.cntval_mask >> 1;
7422 		x86_pmu.perfctr = MSR_IA32_PMC0;
7423 		pr_cont("full-width counters, ");
7424 	}
7425 
7426 	/* Support V6+ MSR Aliasing */
7427 	if (x86_pmu.version >= 6) {
7428 		x86_pmu.perfctr = MSR_IA32_PMC_V6_GP0_CTR;
7429 		x86_pmu.eventsel = MSR_IA32_PMC_V6_GP0_CFG_A;
7430 		x86_pmu.fixedctr = MSR_IA32_PMC_V6_FX0_CTR;
7431 		x86_pmu.addr_offset = intel_pmu_v6_addr_offset;
7432 	}
7433 
7434 	if (!is_hybrid() && x86_pmu.intel_cap.perf_metrics)
7435 		x86_pmu.intel_ctrl |= 1ULL << GLOBAL_CTRL_EN_PERF_METRICS;
7436 
7437 	if (x86_pmu.intel_cap.pebs_timing_info)
7438 		x86_pmu.flags |= PMU_FL_RETIRE_LATENCY;
7439 
7440 	intel_aux_output_init();
7441 
7442 	return 0;
7443 }
7444 
7445 /*
7446  * HT bug: phase 2 init
7447  * Called once we have valid topology information to check
7448  * whether or not HT is enabled
7449  * If HT is off, then we disable the workaround
7450  */
fixup_ht_bug(void)7451 static __init int fixup_ht_bug(void)
7452 {
7453 	int c;
7454 	/*
7455 	 * problem not present on this CPU model, nothing to do
7456 	 */
7457 	if (!(x86_pmu.flags & PMU_FL_EXCL_ENABLED))
7458 		return 0;
7459 
7460 	if (topology_max_smt_threads() > 1) {
7461 		pr_info("PMU erratum BJ122, BV98, HSD29 worked around, HT is on\n");
7462 		return 0;
7463 	}
7464 
7465 	cpus_read_lock();
7466 
7467 	hardlockup_detector_perf_stop();
7468 
7469 	x86_pmu.flags &= ~(PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED);
7470 
7471 	x86_pmu.start_scheduling = NULL;
7472 	x86_pmu.commit_scheduling = NULL;
7473 	x86_pmu.stop_scheduling = NULL;
7474 
7475 	hardlockup_detector_perf_restart();
7476 
7477 	for_each_online_cpu(c)
7478 		free_excl_cntrs(&per_cpu(cpu_hw_events, c));
7479 
7480 	cpus_read_unlock();
7481 	pr_info("PMU erratum BJ122, BV98, HSD29 workaround disabled, HT off\n");
7482 	return 0;
7483 }
7484 subsys_initcall(fixup_ht_bug)
7485