?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
x86, xsave: Use xsaveopt in context-switch path when supported
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / arch / x86 / kernel / cpu / perf_event_p4.c
blobae85d69644d182f31c9711eba6c5ddb21862bcec
1 /*
2 * Netburst Perfomance Events (P4, old Xeon)
3 *
4 * Copyright (C) 2010 Parallels, Inc., Cyrill Gorcunov <gorcunov@openvz.org>
5 * Copyright (C) 2010 Intel Corporation, Lin Ming <ming.m.lin@intel.com>
6 *
7 * For licencing details see kernel-base/COPYING
8 */
9
10 #ifdef CONFIG_CPU_SUP_INTEL
11
12 #include <asm/perf_event_p4.h>
13
14 #define P4_CNTR_LIMIT 3
15 /*
16 * array indices: 0,1 - HT threads, used with HT enabled cpu
17 */
18 struct p4_event_bind {
19 unsigned int opcode; /* Event code and ESCR selector */
20 unsigned int escr_msr[2]; /* ESCR MSR for this event */
21 char cntr[2][P4_CNTR_LIMIT]; /* counter index (offset), -1 on abscence */
22 };
23
24 struct p4_cache_event_bind {
25 unsigned int metric_pebs;
26 unsigned int metric_vert;
27 };
28
29 #define P4_GEN_CACHE_EVENT_BIND(name) \
30 [P4_CACHE__##name] = { \
31 .metric_pebs = P4_PEBS__##name, \
32 .metric_vert = P4_VERT__##name, \
33 }
34
35 static struct p4_cache_event_bind p4_cache_event_bind_map[] = {
36 P4_GEN_CACHE_EVENT_BIND(1stl_cache_load_miss_retired),
37 P4_GEN_CACHE_EVENT_BIND(2ndl_cache_load_miss_retired),
38 P4_GEN_CACHE_EVENT_BIND(dtlb_load_miss_retired),
39 P4_GEN_CACHE_EVENT_BIND(dtlb_store_miss_retired),
40 };
41
42 /*
43 * Note that we don't use CCCR1 here, there is an
44 * exception for P4_BSQ_ALLOCATION but we just have
45 * no workaround
46 *
47 * consider this binding as resources which particular
48 * event may borrow, it doesn't contain EventMask,
49 * Tags and friends -- they are left to a caller
50 */
51 static struct p4_event_bind p4_event_bind_map[] = {
52 [P4_EVENT_TC_DELIVER_MODE] = {
53 .opcode = P4_OPCODE(P4_EVENT_TC_DELIVER_MODE),
54 .escr_msr = { MSR_P4_TC_ESCR0, MSR_P4_TC_ESCR1 },
55 .cntr = { {4, 5, -1}, {6, 7, -1} },
56 },
57 [P4_EVENT_BPU_FETCH_REQUEST] = {
58 .opcode = P4_OPCODE(P4_EVENT_BPU_FETCH_REQUEST),
59 .escr_msr = { MSR_P4_BPU_ESCR0, MSR_P4_BPU_ESCR1 },
60 .cntr = { {0, -1, -1}, {2, -1, -1} },
61 },
62 [P4_EVENT_ITLB_REFERENCE] = {
63 .opcode = P4_OPCODE(P4_EVENT_ITLB_REFERENCE),
64 .escr_msr = { MSR_P4_ITLB_ESCR0, MSR_P4_ITLB_ESCR1 },
65 .cntr = { {0, -1, -1}, {2, -1, -1} },
66 },
67 [P4_EVENT_MEMORY_CANCEL] = {
68 .opcode = P4_OPCODE(P4_EVENT_MEMORY_CANCEL),
69 .escr_msr = { MSR_P4_DAC_ESCR0, MSR_P4_DAC_ESCR1 },
70 .cntr = { {8, 9, -1}, {10, 11, -1} },
71 },
72 [P4_EVENT_MEMORY_COMPLETE] = {
73 .opcode = P4_OPCODE(P4_EVENT_MEMORY_COMPLETE),
74 .escr_msr = { MSR_P4_SAAT_ESCR0 , MSR_P4_SAAT_ESCR1 },
75 .cntr = { {8, 9, -1}, {10, 11, -1} },
76 },
77 [P4_EVENT_LOAD_PORT_REPLAY] = {
78 .opcode = P4_OPCODE(P4_EVENT_LOAD_PORT_REPLAY),
79 .escr_msr = { MSR_P4_SAAT_ESCR0, MSR_P4_SAAT_ESCR1 },
80 .cntr = { {8, 9, -1}, {10, 11, -1} },
81 },
82 [P4_EVENT_STORE_PORT_REPLAY] = {
83 .opcode = P4_OPCODE(P4_EVENT_STORE_PORT_REPLAY),
84 .escr_msr = { MSR_P4_SAAT_ESCR0 , MSR_P4_SAAT_ESCR1 },
85 .cntr = { {8, 9, -1}, {10, 11, -1} },
86 },
87 [P4_EVENT_MOB_LOAD_REPLAY] = {
88 .opcode = P4_OPCODE(P4_EVENT_MOB_LOAD_REPLAY),
89 .escr_msr = { MSR_P4_MOB_ESCR0, MSR_P4_MOB_ESCR1 },
90 .cntr = { {0, -1, -1}, {2, -1, -1} },
91 },
92 [P4_EVENT_PAGE_WALK_TYPE] = {
93 .opcode = P4_OPCODE(P4_EVENT_PAGE_WALK_TYPE),
94 .escr_msr = { MSR_P4_PMH_ESCR0, MSR_P4_PMH_ESCR1 },
95 .cntr = { {0, -1, -1}, {2, -1, -1} },
96 },
97 [P4_EVENT_BSQ_CACHE_REFERENCE] = {
98 .opcode = P4_OPCODE(P4_EVENT_BSQ_CACHE_REFERENCE),
99 .escr_msr = { MSR_P4_BSU_ESCR0, MSR_P4_BSU_ESCR1 },
100 .cntr = { {0, -1, -1}, {2, -1, -1} },
101 },
102 [P4_EVENT_IOQ_ALLOCATION] = {
103 .opcode = P4_OPCODE(P4_EVENT_IOQ_ALLOCATION),
104 .escr_msr = { MSR_P4_FSB_ESCR0, MSR_P4_FSB_ESCR1 },
105 .cntr = { {0, -1, -1}, {2, -1, -1} },
106 },
107 [P4_EVENT_IOQ_ACTIVE_ENTRIES] = { /* shared ESCR */
108 .opcode = P4_OPCODE(P4_EVENT_IOQ_ACTIVE_ENTRIES),
109 .escr_msr = { MSR_P4_FSB_ESCR1, MSR_P4_FSB_ESCR1 },
110 .cntr = { {2, -1, -1}, {3, -1, -1} },
111 },
112 [P4_EVENT_FSB_DATA_ACTIVITY] = {
113 .opcode = P4_OPCODE(P4_EVENT_FSB_DATA_ACTIVITY),
114 .escr_msr = { MSR_P4_FSB_ESCR0, MSR_P4_FSB_ESCR1 },
115 .cntr = { {0, -1, -1}, {2, -1, -1} },
116 },
117 [P4_EVENT_BSQ_ALLOCATION] = { /* shared ESCR, broken CCCR1 */
118 .opcode = P4_OPCODE(P4_EVENT_BSQ_ALLOCATION),
119 .escr_msr = { MSR_P4_BSU_ESCR0, MSR_P4_BSU_ESCR0 },
120 .cntr = { {0, -1, -1}, {1, -1, -1} },
121 },
122 [P4_EVENT_BSQ_ACTIVE_ENTRIES] = { /* shared ESCR */
123 .opcode = P4_OPCODE(P4_EVENT_BSQ_ACTIVE_ENTRIES),
124 .escr_msr = { MSR_P4_BSU_ESCR1 , MSR_P4_BSU_ESCR1 },
125 .cntr = { {2, -1, -1}, {3, -1, -1} },
126 },
127 [P4_EVENT_SSE_INPUT_ASSIST] = {
128 .opcode = P4_OPCODE(P4_EVENT_SSE_INPUT_ASSIST),
129 .escr_msr = { MSR_P4_FIRM_ESCR0, MSR_P4_FIRM_ESCR1 },
130 .cntr = { {8, 9, -1}, {10, 11, -1} },
131 },
132 [P4_EVENT_PACKED_SP_UOP] = {
133 .opcode = P4_OPCODE(P4_EVENT_PACKED_SP_UOP),
134 .escr_msr = { MSR_P4_FIRM_ESCR0, MSR_P4_FIRM_ESCR1 },
135 .cntr = { {8, 9, -1}, {10, 11, -1} },
136 },
137 [P4_EVENT_PACKED_DP_UOP] = {
138 .opcode = P4_OPCODE(P4_EVENT_PACKED_DP_UOP),
139 .escr_msr = { MSR_P4_FIRM_ESCR0, MSR_P4_FIRM_ESCR1 },
140 .cntr = { {8, 9, -1}, {10, 11, -1} },
141 },
142 [P4_EVENT_SCALAR_SP_UOP] = {
143 .opcode = P4_OPCODE(P4_EVENT_SCALAR_SP_UOP),
144 .escr_msr = { MSR_P4_FIRM_ESCR0, MSR_P4_FIRM_ESCR1 },
145 .cntr = { {8, 9, -1}, {10, 11, -1} },
146 },
147 [P4_EVENT_SCALAR_DP_UOP] = {
148 .opcode = P4_OPCODE(P4_EVENT_SCALAR_DP_UOP),
149 .escr_msr = { MSR_P4_FIRM_ESCR0, MSR_P4_FIRM_ESCR1 },
150 .cntr = { {8, 9, -1}, {10, 11, -1} },
151 },
152 [P4_EVENT_64BIT_MMX_UOP] = {
153 .opcode = P4_OPCODE(P4_EVENT_64BIT_MMX_UOP),
154 .escr_msr = { MSR_P4_FIRM_ESCR0, MSR_P4_FIRM_ESCR1 },
155 .cntr = { {8, 9, -1}, {10, 11, -1} },
156 },
157 [P4_EVENT_128BIT_MMX_UOP] = {
158 .opcode = P4_OPCODE(P4_EVENT_128BIT_MMX_UOP),
159 .escr_msr = { MSR_P4_FIRM_ESCR0, MSR_P4_FIRM_ESCR1 },
160 .cntr = { {8, 9, -1}, {10, 11, -1} },
161 },
162 [P4_EVENT_X87_FP_UOP] = {
163 .opcode = P4_OPCODE(P4_EVENT_X87_FP_UOP),
164 .escr_msr = { MSR_P4_FIRM_ESCR0, MSR_P4_FIRM_ESCR1 },
165 .cntr = { {8, 9, -1}, {10, 11, -1} },
166 },
167 [P4_EVENT_TC_MISC] = {
168 .opcode = P4_OPCODE(P4_EVENT_TC_MISC),
169 .escr_msr = { MSR_P4_TC_ESCR0, MSR_P4_TC_ESCR1 },
170 .cntr = { {4, 5, -1}, {6, 7, -1} },
171 },
172 [P4_EVENT_GLOBAL_POWER_EVENTS] = {
173 .opcode = P4_OPCODE(P4_EVENT_GLOBAL_POWER_EVENTS),
174 .escr_msr = { MSR_P4_FSB_ESCR0, MSR_P4_FSB_ESCR1 },
175 .cntr = { {0, -1, -1}, {2, -1, -1} },
176 },
177 [P4_EVENT_TC_MS_XFER] = {
178 .opcode = P4_OPCODE(P4_EVENT_TC_MS_XFER),
179 .escr_msr = { MSR_P4_MS_ESCR0, MSR_P4_MS_ESCR1 },
180 .cntr = { {4, 5, -1}, {6, 7, -1} },
181 },
182 [P4_EVENT_UOP_QUEUE_WRITES] = {
183 .opcode = P4_OPCODE(P4_EVENT_UOP_QUEUE_WRITES),
184 .escr_msr = { MSR_P4_MS_ESCR0, MSR_P4_MS_ESCR1 },
185 .cntr = { {4, 5, -1}, {6, 7, -1} },
186 },
187 [P4_EVENT_RETIRED_MISPRED_BRANCH_TYPE] = {
188 .opcode = P4_OPCODE(P4_EVENT_RETIRED_MISPRED_BRANCH_TYPE),
189 .escr_msr = { MSR_P4_TBPU_ESCR0 , MSR_P4_TBPU_ESCR0 },
190 .cntr = { {4, 5, -1}, {6, 7, -1} },
191 },
192 [P4_EVENT_RETIRED_BRANCH_TYPE] = {
193 .opcode = P4_OPCODE(P4_EVENT_RETIRED_BRANCH_TYPE),
194 .escr_msr = { MSR_P4_TBPU_ESCR0 , MSR_P4_TBPU_ESCR1 },
195 .cntr = { {4, 5, -1}, {6, 7, -1} },
196 },
197 [P4_EVENT_RESOURCE_STALL] = {
198 .opcode = P4_OPCODE(P4_EVENT_RESOURCE_STALL),
199 .escr_msr = { MSR_P4_ALF_ESCR0, MSR_P4_ALF_ESCR1 },
200 .cntr = { {12, 13, 16}, {14, 15, 17} },
201 },
202 [P4_EVENT_WC_BUFFER] = {
203 .opcode = P4_OPCODE(P4_EVENT_WC_BUFFER),
204 .escr_msr = { MSR_P4_DAC_ESCR0, MSR_P4_DAC_ESCR1 },
205 .cntr = { {8, 9, -1}, {10, 11, -1} },
206 },
207 [P4_EVENT_B2B_CYCLES] = {
208 .opcode = P4_OPCODE(P4_EVENT_B2B_CYCLES),
209 .escr_msr = { MSR_P4_FSB_ESCR0, MSR_P4_FSB_ESCR1 },
210 .cntr = { {0, -1, -1}, {2, -1, -1} },
211 },
212 [P4_EVENT_BNR] = {
213 .opcode = P4_OPCODE(P4_EVENT_BNR),
214 .escr_msr = { MSR_P4_FSB_ESCR0, MSR_P4_FSB_ESCR1 },
215 .cntr = { {0, -1, -1}, {2, -1, -1} },
216 },
217 [P4_EVENT_SNOOP] = {
218 .opcode = P4_OPCODE(P4_EVENT_SNOOP),
219 .escr_msr = { MSR_P4_FSB_ESCR0, MSR_P4_FSB_ESCR1 },
220 .cntr = { {0, -1, -1}, {2, -1, -1} },
221 },
222 [P4_EVENT_RESPONSE] = {
223 .opcode = P4_OPCODE(P4_EVENT_RESPONSE),
224 .escr_msr = { MSR_P4_FSB_ESCR0, MSR_P4_FSB_ESCR1 },
225 .cntr = { {0, -1, -1}, {2, -1, -1} },
226 },
227 [P4_EVENT_FRONT_END_EVENT] = {
228 .opcode = P4_OPCODE(P4_EVENT_FRONT_END_EVENT),
229 .escr_msr = { MSR_P4_CRU_ESCR2, MSR_P4_CRU_ESCR3 },
230 .cntr = { {12, 13, 16}, {14, 15, 17} },
231 },
232 [P4_EVENT_EXECUTION_EVENT] = {
233 .opcode = P4_OPCODE(P4_EVENT_EXECUTION_EVENT),
234 .escr_msr = { MSR_P4_CRU_ESCR2, MSR_P4_CRU_ESCR3 },
235 .cntr = { {12, 13, 16}, {14, 15, 17} },
236 },
237 [P4_EVENT_REPLAY_EVENT] = {
238 .opcode = P4_OPCODE(P4_EVENT_REPLAY_EVENT),
239 .escr_msr = { MSR_P4_CRU_ESCR2, MSR_P4_CRU_ESCR3 },
240 .cntr = { {12, 13, 16}, {14, 15, 17} },
241 },
242 [P4_EVENT_INSTR_RETIRED] = {
243 .opcode = P4_OPCODE(P4_EVENT_INSTR_RETIRED),
244 .escr_msr = { MSR_P4_CRU_ESCR0, MSR_P4_CRU_ESCR1 },
245 .cntr = { {12, 13, 16}, {14, 15, 17} },
246 },
247 [P4_EVENT_UOPS_RETIRED] = {
248 .opcode = P4_OPCODE(P4_EVENT_UOPS_RETIRED),
249 .escr_msr = { MSR_P4_CRU_ESCR0, MSR_P4_CRU_ESCR1 },
250 .cntr = { {12, 13, 16}, {14, 15, 17} },
251 },
252 [P4_EVENT_UOP_TYPE] = {
253 .opcode = P4_OPCODE(P4_EVENT_UOP_TYPE),
254 .escr_msr = { MSR_P4_RAT_ESCR0, MSR_P4_RAT_ESCR1 },
255 .cntr = { {12, 13, 16}, {14, 15, 17} },
256 },
257 [P4_EVENT_BRANCH_RETIRED] = {
258 .opcode = P4_OPCODE(P4_EVENT_BRANCH_RETIRED),
259 .escr_msr = { MSR_P4_CRU_ESCR2, MSR_P4_CRU_ESCR3 },
260 .cntr = { {12, 13, 16}, {14, 15, 17} },
261 },
262 [P4_EVENT_MISPRED_BRANCH_RETIRED] = {
263 .opcode = P4_OPCODE(P4_EVENT_MISPRED_BRANCH_RETIRED),
264 .escr_msr = { MSR_P4_CRU_ESCR0, MSR_P4_CRU_ESCR1 },
265 .cntr = { {12, 13, 16}, {14, 15, 17} },
266 },
267 [P4_EVENT_X87_ASSIST] = {
268 .opcode = P4_OPCODE(P4_EVENT_X87_ASSIST),
269 .escr_msr = { MSR_P4_CRU_ESCR2, MSR_P4_CRU_ESCR3 },
270 .cntr = { {12, 13, 16}, {14, 15, 17} },
271 },
272 [P4_EVENT_MACHINE_CLEAR] = {
273 .opcode = P4_OPCODE(P4_EVENT_MACHINE_CLEAR),
274 .escr_msr = { MSR_P4_CRU_ESCR2, MSR_P4_CRU_ESCR3 },
275 .cntr = { {12, 13, 16}, {14, 15, 17} },
276 },
277 [P4_EVENT_INSTR_COMPLETED] = {
278 .opcode = P4_OPCODE(P4_EVENT_INSTR_COMPLETED),
279 .escr_msr = { MSR_P4_CRU_ESCR0, MSR_P4_CRU_ESCR1 },
280 .cntr = { {12, 13, 16}, {14, 15, 17} },
281 },
282 };
283
284 #define P4_GEN_CACHE_EVENT(event, bit, cache_event) \
285 p4_config_pack_escr(P4_ESCR_EVENT(event) | \
286 P4_ESCR_EMASK_BIT(event, bit)) | \
287 p4_config_pack_cccr(cache_event | \
288 P4_CCCR_ESEL(P4_OPCODE_ESEL(P4_OPCODE(event))))
289
290 static __initconst const u64 p4_hw_cache_event_ids
291 [PERF_COUNT_HW_CACHE_MAX]
292 [PERF_COUNT_HW_CACHE_OP_MAX]
293 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
294 {
295 [ C(L1D ) ] = {
296 [ C(OP_READ) ] = {
297 [ C(RESULT_ACCESS) ] = 0x0,
298 [ C(RESULT_MISS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_REPLAY_EVENT, NBOGUS,
299 P4_CACHE__1stl_cache_load_miss_retired),
300 },
301 },
302 [ C(LL ) ] = {
303 [ C(OP_READ) ] = {
304 [ C(RESULT_ACCESS) ] = 0x0,
305 [ C(RESULT_MISS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_REPLAY_EVENT, NBOGUS,
306 P4_CACHE__2ndl_cache_load_miss_retired),
307 },
308 },
309 [ C(DTLB) ] = {
310 [ C(OP_READ) ] = {
311 [ C(RESULT_ACCESS) ] = 0x0,
312 [ C(RESULT_MISS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_REPLAY_EVENT, NBOGUS,
313 P4_CACHE__dtlb_load_miss_retired),
314 },
315 [ C(OP_WRITE) ] = {
316 [ C(RESULT_ACCESS) ] = 0x0,
317 [ C(RESULT_MISS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_REPLAY_EVENT, NBOGUS,
318 P4_CACHE__dtlb_store_miss_retired),
319 },
320 },
321 [ C(ITLB) ] = {
322 [ C(OP_READ) ] = {
323 [ C(RESULT_ACCESS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_ITLB_REFERENCE, HIT,
324 P4_CACHE__itlb_reference_hit),
325 [ C(RESULT_MISS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_ITLB_REFERENCE, MISS,
326 P4_CACHE__itlb_reference_miss),
327 },
328 [ C(OP_WRITE) ] = {
329 [ C(RESULT_ACCESS) ] = -1,
330 [ C(RESULT_MISS) ] = -1,
331 },
332 [ C(OP_PREFETCH) ] = {
333 [ C(RESULT_ACCESS) ] = -1,
334 [ C(RESULT_MISS) ] = -1,
335 },
336 },
337 };
338
339 static u64 p4_general_events[PERF_COUNT_HW_MAX] = {
340 /* non-halted CPU clocks */
341 [PERF_COUNT_HW_CPU_CYCLES] =
342 p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_GLOBAL_POWER_EVENTS) |
343 P4_ESCR_EMASK_BIT(P4_EVENT_GLOBAL_POWER_EVENTS, RUNNING)),
344
345 /*
346 * retired instructions
347 * in a sake of simplicity we don't use the FSB tagging
348 */
349 [PERF_COUNT_HW_INSTRUCTIONS] =
350 p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_INSTR_RETIRED) |
351 P4_ESCR_EMASK_BIT(P4_EVENT_INSTR_RETIRED, NBOGUSNTAG) |
352 P4_ESCR_EMASK_BIT(P4_EVENT_INSTR_RETIRED, BOGUSNTAG)),
353
354 /* cache hits */
355 [PERF_COUNT_HW_CACHE_REFERENCES] =
356 p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_BSQ_CACHE_REFERENCE) |
357 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_2ndL_HITS) |
358 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_2ndL_HITE) |
359 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_2ndL_HITM) |
360 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_3rdL_HITS) |
361 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_3rdL_HITE) |
362 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_3rdL_HITM)),
363
364 /* cache misses */
365 [PERF_COUNT_HW_CACHE_MISSES] =
366 p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_BSQ_CACHE_REFERENCE) |
367 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_2ndL_MISS) |
368 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_3rdL_MISS) |
369 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, WR_2ndL_MISS)),
370
371 /* branch instructions retired */
372 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] =
373 p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_RETIRED_BRANCH_TYPE) |
374 P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_BRANCH_TYPE, CONDITIONAL) |
375 P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_BRANCH_TYPE, CALL) |
376 P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_BRANCH_TYPE, RETURN) |
377 P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_BRANCH_TYPE, INDIRECT)),
378
379 /* mispredicted branches retired */
380 [PERF_COUNT_HW_BRANCH_MISSES] =
381 p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_MISPRED_BRANCH_RETIRED) |
382 P4_ESCR_EMASK_BIT(P4_EVENT_MISPRED_BRANCH_RETIRED, NBOGUS)),
383
384 /* bus ready clocks (cpu is driving #DRDY_DRV\#DRDY_OWN): */
385 [PERF_COUNT_HW_BUS_CYCLES] =
386 p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_FSB_DATA_ACTIVITY) |
387 P4_ESCR_EMASK_BIT(P4_EVENT_FSB_DATA_ACTIVITY, DRDY_DRV) |
388 P4_ESCR_EMASK_BIT(P4_EVENT_FSB_DATA_ACTIVITY, DRDY_OWN)) |
389 p4_config_pack_cccr(P4_CCCR_EDGE | P4_CCCR_COMPARE),
390 };
391
392 static struct p4_event_bind *p4_config_get_bind(u64 config)
393 {
394 unsigned int evnt = p4_config_unpack_event(config);
395 struct p4_event_bind *bind = NULL;
396
397 if (evnt < ARRAY_SIZE(p4_event_bind_map))
398 bind = &p4_event_bind_map[evnt];
399
400 return bind;
401 }
402
403 static u64 p4_pmu_event_map(int hw_event)
404 {
405 struct p4_event_bind *bind;
406 unsigned int esel;
407 u64 config;
408
409 config = p4_general_events[hw_event];
410 bind = p4_config_get_bind(config);
411 esel = P4_OPCODE_ESEL(bind->opcode);
412 config |= p4_config_pack_cccr(P4_CCCR_ESEL(esel));
413
414 return config;
415 }
416
417 static int p4_hw_config(struct perf_event *event)
418 {
419 int cpu = get_cpu();
420 int rc = 0;
421 unsigned int evnt;
422 u32 escr, cccr;
423
424 /*
425 * the reason we use cpu that early is that: if we get scheduled
426 * first time on the same cpu -- we will not need swap thread
427 * specific flags in config (and will save some cpu cycles)
428 */
429
430 cccr = p4_default_cccr_conf(cpu);
431 escr = p4_default_escr_conf(cpu, event->attr.exclude_kernel,
432 event->attr.exclude_user);
433 event->hw.config = p4_config_pack_escr(escr) |
434 p4_config_pack_cccr(cccr);
435
436 if (p4_ht_active() && p4_ht_thread(cpu))
437 event->hw.config = p4_set_ht_bit(event->hw.config);
438
439 if (event->attr.type == PERF_TYPE_RAW) {
440
441 /* user data may have out-of-bound event index */
442 evnt = p4_config_unpack_event(event->attr.config);
443 if (evnt >= ARRAY_SIZE(p4_event_bind_map)) {
444 rc = -EINVAL;
445 goto out;
446 }
447
448 /*
449 * We don't control raw events so it's up to the caller
450 * to pass sane values (and we don't count the thread number
451 * on HT machine but allow HT-compatible specifics to be
452 * passed on)
453 *
454 * XXX: HT wide things should check perf_paranoid_cpu() &&
455 * CAP_SYS_ADMIN
456 */
457 event->hw.config |= event->attr.config &
458 (p4_config_pack_escr(P4_ESCR_MASK_HT) |
459 p4_config_pack_cccr(P4_CCCR_MASK_HT));
460 }
461
462 rc = x86_setup_perfctr(event);
463 out:
464 put_cpu();
465 return rc;
466 }
467
468 static inline int p4_pmu_clear_cccr_ovf(struct hw_perf_event *hwc)
469 {
470 int overflow = 0;
471 u32 low, high;
472
473 rdmsr(hwc->config_base + hwc->idx, low, high);
474
475 /* we need to check high bit for unflagged overflows */
476 if ((low & P4_CCCR_OVF) || !(high & (1 << 31))) {
477 overflow = 1;
478 (void)checking_wrmsrl(hwc->config_base + hwc->idx,
479 ((u64)low) & ~P4_CCCR_OVF);
480 }
481
482 return overflow;
483 }
484
485 static inline void p4_pmu_disable_event(struct perf_event *event)
486 {
487 struct hw_perf_event *hwc = &event->hw;
488
489 /*
490 * If event gets disabled while counter is in overflowed
491 * state we need to clear P4_CCCR_OVF, otherwise interrupt get
492 * asserted again and again
493 */
494 (void)checking_wrmsrl(hwc->config_base + hwc->idx,
495 (u64)(p4_config_unpack_cccr(hwc->config)) &
496 ~P4_CCCR_ENABLE & ~P4_CCCR_OVF & ~P4_CCCR_RESERVED);
497 }
498
499 static void p4_pmu_disable_all(void)
500 {
501 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
502 int idx;
503
504 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
505 struct perf_event *event = cpuc->events[idx];
506 if (!test_bit(idx, cpuc->active_mask))
507 continue;
508 p4_pmu_disable_event(event);
509 }
510 }
511
512 static void p4_pmu_enable_event(struct perf_event *event)
513 {
514 struct hw_perf_event *hwc = &event->hw;
515 int thread = p4_ht_config_thread(hwc->config);
516 u64 escr_conf = p4_config_unpack_escr(p4_clear_ht_bit(hwc->config));
517 unsigned int idx = p4_config_unpack_event(hwc->config);
518 unsigned int idx_cache = p4_config_unpack_cache_event(hwc->config);
519 struct p4_event_bind *bind;
520 struct p4_cache_event_bind *bind_cache;
521 u64 escr_addr, cccr;
522
523 bind = &p4_event_bind_map[idx];
524 escr_addr = (u64)bind->escr_msr[thread];
525
526 /*
527 * - we dont support cascaded counters yet
528 * - and counter 1 is broken (erratum)
529 */
530 WARN_ON_ONCE(p4_is_event_cascaded(hwc->config));
531 WARN_ON_ONCE(hwc->idx == 1);
532
533 /* we need a real Event value */
534 escr_conf &= ~P4_ESCR_EVENT_MASK;
535 escr_conf |= P4_ESCR_EVENT(P4_OPCODE_EVNT(bind->opcode));
536
537 cccr = p4_config_unpack_cccr(hwc->config);
538
539 /*
540 * it could be Cache event so that we need to
541 * set metrics into additional MSRs
542 */
543 BUILD_BUG_ON(P4_CACHE__MAX > P4_CCCR_CACHE_OPS_MASK);
544 if (idx_cache > P4_CACHE__NONE &&
545 idx_cache < ARRAY_SIZE(p4_cache_event_bind_map)) {
546 bind_cache = &p4_cache_event_bind_map[idx_cache];
547 (void)checking_wrmsrl(MSR_IA32_PEBS_ENABLE, (u64)bind_cache->metric_pebs);
548 (void)checking_wrmsrl(MSR_P4_PEBS_MATRIX_VERT, (u64)bind_cache->metric_vert);
549 }
550
551 (void)checking_wrmsrl(escr_addr, escr_conf);
552 (void)checking_wrmsrl(hwc->config_base + hwc->idx,
553 (cccr & ~P4_CCCR_RESERVED) | P4_CCCR_ENABLE);
554 }
555
556 static void p4_pmu_enable_all(int added)
557 {
558 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
559 int idx;
560
561 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
562 struct perf_event *event = cpuc->events[idx];
563 if (!test_bit(idx, cpuc->active_mask))
564 continue;
565 p4_pmu_enable_event(event);
566 }
567 }
568
569 static int p4_pmu_handle_irq(struct pt_regs *regs)
570 {
571 struct perf_sample_data data;
572 struct cpu_hw_events *cpuc;
573 struct perf_event *event;
574 struct hw_perf_event *hwc;
575 int idx, handled = 0;
576 u64 val;
577
578 data.addr = 0;
579 data.raw = NULL;
580
581 cpuc = &__get_cpu_var(cpu_hw_events);
582
583 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
584
585 if (!test_bit(idx, cpuc->active_mask))
586 continue;
587
588 event = cpuc->events[idx];
589 hwc = &event->hw;
590
591 WARN_ON_ONCE(hwc->idx != idx);
592
593 /* it might be unflagged overflow */
594 handled = p4_pmu_clear_cccr_ovf(hwc);
595
596 val = x86_perf_event_update(event);
597 if (!handled && (val & (1ULL << (x86_pmu.cntval_bits - 1))))
598 continue;
599
600 /* event overflow for sure */
601 data.period = event->hw.last_period;
602
603 if (!x86_perf_event_set_period(event))
604 continue;
605 if (perf_event_overflow(event, 1, &data, regs))
606 p4_pmu_disable_event(event);
607 }
608
609 if (handled) {
610 /* p4 quirk: unmask it again */
611 apic_write(APIC_LVTPC, apic_read(APIC_LVTPC) & ~APIC_LVT_MASKED);
612 inc_irq_stat(apic_perf_irqs);
613 }
614
615 return handled;
616 }
617
618 /*
619 * swap thread specific fields according to a thread
620 * we are going to run on
621 */
622 static void p4_pmu_swap_config_ts(struct hw_perf_event *hwc, int cpu)
623 {
624 u32 escr, cccr;
625
626 /*
627 * we either lucky and continue on same cpu or no HT support
628 */
629 if (!p4_should_swap_ts(hwc->config, cpu))
630 return;
631
632 /*
633 * the event is migrated from an another logical
634 * cpu, so we need to swap thread specific flags
635 */
636
637 escr = p4_config_unpack_escr(hwc->config);
638 cccr = p4_config_unpack_cccr(hwc->config);
639
640 if (p4_ht_thread(cpu)) {
641 cccr &= ~P4_CCCR_OVF_PMI_T0;
642 cccr |= P4_CCCR_OVF_PMI_T1;
643 if (escr & P4_ESCR_T0_OS) {
644 escr &= ~P4_ESCR_T0_OS;
645 escr |= P4_ESCR_T1_OS;
646 }
647 if (escr & P4_ESCR_T0_USR) {
648 escr &= ~P4_ESCR_T0_USR;
649 escr |= P4_ESCR_T1_USR;
650 }
651 hwc->config = p4_config_pack_escr(escr);
652 hwc->config |= p4_config_pack_cccr(cccr);
653 hwc->config |= P4_CONFIG_HT;
654 } else {
655 cccr &= ~P4_CCCR_OVF_PMI_T1;
656 cccr |= P4_CCCR_OVF_PMI_T0;
657 if (escr & P4_ESCR_T1_OS) {
658 escr &= ~P4_ESCR_T1_OS;
659 escr |= P4_ESCR_T0_OS;
660 }
661 if (escr & P4_ESCR_T1_USR) {
662 escr &= ~P4_ESCR_T1_USR;
663 escr |= P4_ESCR_T0_USR;
664 }
665 hwc->config = p4_config_pack_escr(escr);
666 hwc->config |= p4_config_pack_cccr(cccr);
667 hwc->config &= ~P4_CONFIG_HT;
668 }
669 }
670
671 /*
672 * ESCR address hashing is tricky, ESCRs are not sequential
673 * in memory but all starts from MSR_P4_BSU_ESCR0 (0x03a0) and
674 * the metric between any ESCRs is laid in range [0xa0,0xe1]
675 *
676 * so we make ~70% filled hashtable
677 */
678
679 #define P4_ESCR_MSR_BASE 0x000003a0
680 #define P4_ESCR_MSR_MAX 0x000003e1
681 #define P4_ESCR_MSR_TABLE_SIZE (P4_ESCR_MSR_MAX - P4_ESCR_MSR_BASE + 1)
682 #define P4_ESCR_MSR_IDX(msr) (msr - P4_ESCR_MSR_BASE)
683 #define P4_ESCR_MSR_TABLE_ENTRY(msr) [P4_ESCR_MSR_IDX(msr)] = msr
684
685 static const unsigned int p4_escr_table[P4_ESCR_MSR_TABLE_SIZE] = {
686 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_ALF_ESCR0),
687 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_ALF_ESCR1),
688 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_BPU_ESCR0),
689 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_BPU_ESCR1),
690 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_BSU_ESCR0),
691 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_BSU_ESCR1),
692 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR0),
693 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR1),
694 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR2),
695 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR3),
696 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR4),
697 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR5),
698 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_DAC_ESCR0),
699 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_DAC_ESCR1),
700 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FIRM_ESCR0),
701 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FIRM_ESCR1),
702 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FLAME_ESCR0),
703 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FLAME_ESCR1),
704 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FSB_ESCR0),
705 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FSB_ESCR1),
706 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IQ_ESCR0),
707 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IQ_ESCR1),
708 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IS_ESCR0),
709 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IS_ESCR1),
710 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_ITLB_ESCR0),
711 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_ITLB_ESCR1),
712 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IX_ESCR0),
713 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IX_ESCR1),
714 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_MOB_ESCR0),
715 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_MOB_ESCR1),
716 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_MS_ESCR0),
717 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_MS_ESCR1),
718 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_PMH_ESCR0),
719 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_PMH_ESCR1),
720 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_RAT_ESCR0),
721 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_RAT_ESCR1),
722 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_SAAT_ESCR0),
723 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_SAAT_ESCR1),
724 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_SSU_ESCR0),
725 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_SSU_ESCR1),
726 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_TBPU_ESCR0),
727 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_TBPU_ESCR1),
728 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_TC_ESCR0),
729 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_TC_ESCR1),
730 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_U2L_ESCR0),
731 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_U2L_ESCR1),
732 };
733
734 static int p4_get_escr_idx(unsigned int addr)
735 {
736 unsigned int idx = P4_ESCR_MSR_IDX(addr);
737
738 if (unlikely(idx >= P4_ESCR_MSR_TABLE_SIZE ||
739 !p4_escr_table[idx] ||
740 p4_escr_table[idx] != addr)) {
741 WARN_ONCE(1, "P4 PMU: Wrong address passed: %x\n", addr);
742 return -1;
743 }
744
745 return idx;
746 }
747
748 static int p4_next_cntr(int thread, unsigned long *used_mask,
749 struct p4_event_bind *bind)
750 {
751 int i, j;
752
753 for (i = 0; i < P4_CNTR_LIMIT; i++) {
754 j = bind->cntr[thread][i];
755 if (j != -1 && !test_bit(j, used_mask))
756 return j;
757 }
758
759 return -1;
760 }
761
762 static int p4_pmu_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
763 {
764 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
765 unsigned long escr_mask[BITS_TO_LONGS(P4_ESCR_MSR_TABLE_SIZE)];
766 int cpu = smp_processor_id();
767 struct hw_perf_event *hwc;
768 struct p4_event_bind *bind;
769 unsigned int i, thread, num;
770 int cntr_idx, escr_idx;
771
772 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
773 bitmap_zero(escr_mask, P4_ESCR_MSR_TABLE_SIZE);
774
775 for (i = 0, num = n; i < n; i++, num--) {
776
777 hwc = &cpuc->event_list[i]->hw;
778 thread = p4_ht_thread(cpu);
779 bind = p4_config_get_bind(hwc->config);
780 escr_idx = p4_get_escr_idx(bind->escr_msr[thread]);
781 if (unlikely(escr_idx == -1))
782 goto done;
783
784 if (hwc->idx != -1 && !p4_should_swap_ts(hwc->config, cpu)) {
785 cntr_idx = hwc->idx;
786 if (assign)
787 assign[i] = hwc->idx;
788 goto reserve;
789 }
790
791 cntr_idx = p4_next_cntr(thread, used_mask, bind);
792 if (cntr_idx == -1 || test_bit(escr_idx, escr_mask))
793 goto done;
794
795 p4_pmu_swap_config_ts(hwc, cpu);
796 if (assign)
797 assign[i] = cntr_idx;
798 reserve:
799 set_bit(cntr_idx, used_mask);
800 set_bit(escr_idx, escr_mask);
801 }
802
803 done:
804 return num ? -ENOSPC : 0;
805 }
806
807 static __initconst const struct x86_pmu p4_pmu = {
808 .name = "Netburst P4/Xeon",
809 .handle_irq = p4_pmu_handle_irq,
810 .disable_all = p4_pmu_disable_all,
811 .enable_all = p4_pmu_enable_all,
812 .enable = p4_pmu_enable_event,
813 .disable = p4_pmu_disable_event,
814 .eventsel = MSR_P4_BPU_CCCR0,
815 .perfctr = MSR_P4_BPU_PERFCTR0,
816 .event_map = p4_pmu_event_map,
817 .max_events = ARRAY_SIZE(p4_general_events),
818 .get_event_constraints = x86_get_event_constraints,
819 /*
820 * IF HT disabled we may need to use all
821 * ARCH_P4_MAX_CCCR counters simulaneously
822 * though leave it restricted at moment assuming
823 * HT is on
824 */
825 .num_counters = ARCH_P4_MAX_CCCR,
826 .apic = 1,
827 .cntval_bits = 40,
828 .cntval_mask = (1ULL << 40) - 1,
829 .max_period = (1ULL << 39) - 1,
830 .hw_config = p4_hw_config,
831 .schedule_events = p4_pmu_schedule_events,
832 };
833
834 static __init int p4_pmu_init(void)
835 {
836 unsigned int low, high;
837
838 /* If we get stripped -- indexig fails */
839 BUILD_BUG_ON(ARCH_P4_MAX_CCCR > X86_PMC_MAX_GENERIC);
840
841 rdmsr(MSR_IA32_MISC_ENABLE, low, high);
842 if (!(low & (1 << 7))) {
843 pr_cont("unsupported Netburst CPU model %d ",
844 boot_cpu_data.x86_model);
845 return -ENODEV;
846 }
847
848 memcpy(hw_cache_event_ids, p4_hw_cache_event_ids,
849 sizeof(hw_cache_event_ids));
850
851 pr_cont("Netburst events, ");
852
853 x86_pmu = p4_pmu;
854
855 return 0;
856 }
857
858 #endif /* CONFIG_CPU_SUP_INTEL */
Linux 2.6 ibm-acpi/thinkpad-acpi driver git tree