Merge branch 'master' of /home/davem/src/GIT/linux-2.6/
[sfrench/cifs-2.6.git] / arch / x86 / kernel / cpu / perf_event_intel.c
1 #ifdef CONFIG_CPU_SUP_INTEL
2
3 /*
4  * Intel PerfMon, used on Core and later.
5  */
6 static const u64 intel_perfmon_event_map[] =
7 {
8   [PERF_COUNT_HW_CPU_CYCLES]            = 0x003c,
9   [PERF_COUNT_HW_INSTRUCTIONS]          = 0x00c0,
10   [PERF_COUNT_HW_CACHE_REFERENCES]      = 0x4f2e,
11   [PERF_COUNT_HW_CACHE_MISSES]          = 0x412e,
12   [PERF_COUNT_HW_BRANCH_INSTRUCTIONS]   = 0x00c4,
13   [PERF_COUNT_HW_BRANCH_MISSES]         = 0x00c5,
14   [PERF_COUNT_HW_BUS_CYCLES]            = 0x013c,
15 };
16
17 static struct event_constraint intel_core_event_constraints[] =
18 {
19         INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
20         INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
21         INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
22         INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
23         INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
24         INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */
25         EVENT_CONSTRAINT_END
26 };
27
28 static struct event_constraint intel_core2_event_constraints[] =
29 {
30         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
31         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
32         /*
33          * Core2 has Fixed Counter 2 listed as CPU_CLK_UNHALTED.REF and event
34          * 0x013c as CPU_CLK_UNHALTED.BUS and specifies there is a fixed
35          * ratio between these counters.
36          */
37         /* FIXED_EVENT_CONSTRAINT(0x013c, 2),  CPU_CLK_UNHALTED.REF */
38         INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
39         INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
40         INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
41         INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
42         INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
43         INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */
44         INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
45         INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */
46         INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */
47         INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */
48         EVENT_CONSTRAINT_END
49 };
50
51 static struct event_constraint intel_nehalem_event_constraints[] =
52 {
53         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
54         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
55         /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
56         INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */
57         INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */
58         INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */
59         INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */
60         INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */
61         INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */
62         INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
63         INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
64         EVENT_CONSTRAINT_END
65 };
66
67 static struct event_constraint intel_westmere_event_constraints[] =
68 {
69         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
70         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
71         /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
72         INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
73         INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */
74         INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
75         EVENT_CONSTRAINT_END
76 };
77
78 static struct event_constraint intel_gen_event_constraints[] =
79 {
80         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
81         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
82         /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
83         EVENT_CONSTRAINT_END
84 };
85
86 static u64 intel_pmu_event_map(int hw_event)
87 {
88         return intel_perfmon_event_map[hw_event];
89 }
90
91 static __initconst u64 westmere_hw_cache_event_ids
92                                 [PERF_COUNT_HW_CACHE_MAX]
93                                 [PERF_COUNT_HW_CACHE_OP_MAX]
94                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
95 {
96  [ C(L1D) ] = {
97         [ C(OP_READ) ] = {
98                 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS       */
99                 [ C(RESULT_MISS)   ] = 0x0151, /* L1D.REPL                     */
100         },
101         [ C(OP_WRITE) ] = {
102                 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES      */
103                 [ C(RESULT_MISS)   ] = 0x0251, /* L1D.M_REPL                   */
104         },
105         [ C(OP_PREFETCH) ] = {
106                 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS        */
107                 [ C(RESULT_MISS)   ] = 0x024e, /* L1D_PREFETCH.MISS            */
108         },
109  },
110  [ C(L1I ) ] = {
111         [ C(OP_READ) ] = {
112                 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS                    */
113                 [ C(RESULT_MISS)   ] = 0x0280, /* L1I.MISSES                   */
114         },
115         [ C(OP_WRITE) ] = {
116                 [ C(RESULT_ACCESS) ] = -1,
117                 [ C(RESULT_MISS)   ] = -1,
118         },
119         [ C(OP_PREFETCH) ] = {
120                 [ C(RESULT_ACCESS) ] = 0x0,
121                 [ C(RESULT_MISS)   ] = 0x0,
122         },
123  },
124  [ C(LL  ) ] = {
125         [ C(OP_READ) ] = {
126                 [ C(RESULT_ACCESS) ] = 0x0324, /* L2_RQSTS.LOADS               */
127                 [ C(RESULT_MISS)   ] = 0x0224, /* L2_RQSTS.LD_MISS             */
128         },
129         [ C(OP_WRITE) ] = {
130                 [ C(RESULT_ACCESS) ] = 0x0c24, /* L2_RQSTS.RFOS                */
131                 [ C(RESULT_MISS)   ] = 0x0824, /* L2_RQSTS.RFO_MISS            */
132         },
133         [ C(OP_PREFETCH) ] = {
134                 [ C(RESULT_ACCESS) ] = 0x4f2e, /* LLC Reference                */
135                 [ C(RESULT_MISS)   ] = 0x412e, /* LLC Misses                   */
136         },
137  },
138  [ C(DTLB) ] = {
139         [ C(OP_READ) ] = {
140                 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS       */
141                 [ C(RESULT_MISS)   ] = 0x0108, /* DTLB_LOAD_MISSES.ANY         */
142         },
143         [ C(OP_WRITE) ] = {
144                 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES      */
145                 [ C(RESULT_MISS)   ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS  */
146         },
147         [ C(OP_PREFETCH) ] = {
148                 [ C(RESULT_ACCESS) ] = 0x0,
149                 [ C(RESULT_MISS)   ] = 0x0,
150         },
151  },
152  [ C(ITLB) ] = {
153         [ C(OP_READ) ] = {
154                 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P           */
155                 [ C(RESULT_MISS)   ] = 0x0185, /* ITLB_MISSES.ANY              */
156         },
157         [ C(OP_WRITE) ] = {
158                 [ C(RESULT_ACCESS) ] = -1,
159                 [ C(RESULT_MISS)   ] = -1,
160         },
161         [ C(OP_PREFETCH) ] = {
162                 [ C(RESULT_ACCESS) ] = -1,
163                 [ C(RESULT_MISS)   ] = -1,
164         },
165  },
166  [ C(BPU ) ] = {
167         [ C(OP_READ) ] = {
168                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
169                 [ C(RESULT_MISS)   ] = 0x03e8, /* BPU_CLEARS.ANY               */
170         },
171         [ C(OP_WRITE) ] = {
172                 [ C(RESULT_ACCESS) ] = -1,
173                 [ C(RESULT_MISS)   ] = -1,
174         },
175         [ C(OP_PREFETCH) ] = {
176                 [ C(RESULT_ACCESS) ] = -1,
177                 [ C(RESULT_MISS)   ] = -1,
178         },
179  },
180 };
181
182 static __initconst u64 nehalem_hw_cache_event_ids
183                                 [PERF_COUNT_HW_CACHE_MAX]
184                                 [PERF_COUNT_HW_CACHE_OP_MAX]
185                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
186 {
187  [ C(L1D) ] = {
188         [ C(OP_READ) ] = {
189                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI            */
190                 [ C(RESULT_MISS)   ] = 0x0140, /* L1D_CACHE_LD.I_STATE         */
191         },
192         [ C(OP_WRITE) ] = {
193                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI            */
194                 [ C(RESULT_MISS)   ] = 0x0141, /* L1D_CACHE_ST.I_STATE         */
195         },
196         [ C(OP_PREFETCH) ] = {
197                 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS        */
198                 [ C(RESULT_MISS)   ] = 0x024e, /* L1D_PREFETCH.MISS            */
199         },
200  },
201  [ C(L1I ) ] = {
202         [ C(OP_READ) ] = {
203                 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS                    */
204                 [ C(RESULT_MISS)   ] = 0x0280, /* L1I.MISSES                   */
205         },
206         [ C(OP_WRITE) ] = {
207                 [ C(RESULT_ACCESS) ] = -1,
208                 [ C(RESULT_MISS)   ] = -1,
209         },
210         [ C(OP_PREFETCH) ] = {
211                 [ C(RESULT_ACCESS) ] = 0x0,
212                 [ C(RESULT_MISS)   ] = 0x0,
213         },
214  },
215  [ C(LL  ) ] = {
216         [ C(OP_READ) ] = {
217                 [ C(RESULT_ACCESS) ] = 0x0324, /* L2_RQSTS.LOADS               */
218                 [ C(RESULT_MISS)   ] = 0x0224, /* L2_RQSTS.LD_MISS             */
219         },
220         [ C(OP_WRITE) ] = {
221                 [ C(RESULT_ACCESS) ] = 0x0c24, /* L2_RQSTS.RFOS                */
222                 [ C(RESULT_MISS)   ] = 0x0824, /* L2_RQSTS.RFO_MISS            */
223         },
224         [ C(OP_PREFETCH) ] = {
225                 [ C(RESULT_ACCESS) ] = 0x4f2e, /* LLC Reference                */
226                 [ C(RESULT_MISS)   ] = 0x412e, /* LLC Misses                   */
227         },
228  },
229  [ C(DTLB) ] = {
230         [ C(OP_READ) ] = {
231                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI   (alias)  */
232                 [ C(RESULT_MISS)   ] = 0x0108, /* DTLB_LOAD_MISSES.ANY         */
233         },
234         [ C(OP_WRITE) ] = {
235                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI   (alias)  */
236                 [ C(RESULT_MISS)   ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS  */
237         },
238         [ C(OP_PREFETCH) ] = {
239                 [ C(RESULT_ACCESS) ] = 0x0,
240                 [ C(RESULT_MISS)   ] = 0x0,
241         },
242  },
243  [ C(ITLB) ] = {
244         [ C(OP_READ) ] = {
245                 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P           */
246                 [ C(RESULT_MISS)   ] = 0x20c8, /* ITLB_MISS_RETIRED            */
247         },
248         [ C(OP_WRITE) ] = {
249                 [ C(RESULT_ACCESS) ] = -1,
250                 [ C(RESULT_MISS)   ] = -1,
251         },
252         [ C(OP_PREFETCH) ] = {
253                 [ C(RESULT_ACCESS) ] = -1,
254                 [ C(RESULT_MISS)   ] = -1,
255         },
256  },
257  [ C(BPU ) ] = {
258         [ C(OP_READ) ] = {
259                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
260                 [ C(RESULT_MISS)   ] = 0x03e8, /* BPU_CLEARS.ANY               */
261         },
262         [ C(OP_WRITE) ] = {
263                 [ C(RESULT_ACCESS) ] = -1,
264                 [ C(RESULT_MISS)   ] = -1,
265         },
266         [ C(OP_PREFETCH) ] = {
267                 [ C(RESULT_ACCESS) ] = -1,
268                 [ C(RESULT_MISS)   ] = -1,
269         },
270  },
271 };
272
273 static __initconst u64 core2_hw_cache_event_ids
274                                 [PERF_COUNT_HW_CACHE_MAX]
275                                 [PERF_COUNT_HW_CACHE_OP_MAX]
276                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
277 {
278  [ C(L1D) ] = {
279         [ C(OP_READ) ] = {
280                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI          */
281                 [ C(RESULT_MISS)   ] = 0x0140, /* L1D_CACHE_LD.I_STATE       */
282         },
283         [ C(OP_WRITE) ] = {
284                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI          */
285                 [ C(RESULT_MISS)   ] = 0x0141, /* L1D_CACHE_ST.I_STATE       */
286         },
287         [ C(OP_PREFETCH) ] = {
288                 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS      */
289                 [ C(RESULT_MISS)   ] = 0,
290         },
291  },
292  [ C(L1I ) ] = {
293         [ C(OP_READ) ] = {
294                 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS                  */
295                 [ C(RESULT_MISS)   ] = 0x0081, /* L1I.MISSES                 */
296         },
297         [ C(OP_WRITE) ] = {
298                 [ C(RESULT_ACCESS) ] = -1,
299                 [ C(RESULT_MISS)   ] = -1,
300         },
301         [ C(OP_PREFETCH) ] = {
302                 [ C(RESULT_ACCESS) ] = 0,
303                 [ C(RESULT_MISS)   ] = 0,
304         },
305  },
306  [ C(LL  ) ] = {
307         [ C(OP_READ) ] = {
308                 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI                 */
309                 [ C(RESULT_MISS)   ] = 0x4129, /* L2_LD.ISTATE               */
310         },
311         [ C(OP_WRITE) ] = {
312                 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI                 */
313                 [ C(RESULT_MISS)   ] = 0x412A, /* L2_ST.ISTATE               */
314         },
315         [ C(OP_PREFETCH) ] = {
316                 [ C(RESULT_ACCESS) ] = 0,
317                 [ C(RESULT_MISS)   ] = 0,
318         },
319  },
320  [ C(DTLB) ] = {
321         [ C(OP_READ) ] = {
322                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI  (alias) */
323                 [ C(RESULT_MISS)   ] = 0x0208, /* DTLB_MISSES.MISS_LD        */
324         },
325         [ C(OP_WRITE) ] = {
326                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI  (alias) */
327                 [ C(RESULT_MISS)   ] = 0x0808, /* DTLB_MISSES.MISS_ST        */
328         },
329         [ C(OP_PREFETCH) ] = {
330                 [ C(RESULT_ACCESS) ] = 0,
331                 [ C(RESULT_MISS)   ] = 0,
332         },
333  },
334  [ C(ITLB) ] = {
335         [ C(OP_READ) ] = {
336                 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P         */
337                 [ C(RESULT_MISS)   ] = 0x1282, /* ITLBMISSES                 */
338         },
339         [ C(OP_WRITE) ] = {
340                 [ C(RESULT_ACCESS) ] = -1,
341                 [ C(RESULT_MISS)   ] = -1,
342         },
343         [ C(OP_PREFETCH) ] = {
344                 [ C(RESULT_ACCESS) ] = -1,
345                 [ C(RESULT_MISS)   ] = -1,
346         },
347  },
348  [ C(BPU ) ] = {
349         [ C(OP_READ) ] = {
350                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY        */
351                 [ C(RESULT_MISS)   ] = 0x00c5, /* BP_INST_RETIRED.MISPRED    */
352         },
353         [ C(OP_WRITE) ] = {
354                 [ C(RESULT_ACCESS) ] = -1,
355                 [ C(RESULT_MISS)   ] = -1,
356         },
357         [ C(OP_PREFETCH) ] = {
358                 [ C(RESULT_ACCESS) ] = -1,
359                 [ C(RESULT_MISS)   ] = -1,
360         },
361  },
362 };
363
364 static __initconst u64 atom_hw_cache_event_ids
365                                 [PERF_COUNT_HW_CACHE_MAX]
366                                 [PERF_COUNT_HW_CACHE_OP_MAX]
367                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
368 {
369  [ C(L1D) ] = {
370         [ C(OP_READ) ] = {
371                 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD               */
372                 [ C(RESULT_MISS)   ] = 0,
373         },
374         [ C(OP_WRITE) ] = {
375                 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST               */
376                 [ C(RESULT_MISS)   ] = 0,
377         },
378         [ C(OP_PREFETCH) ] = {
379                 [ C(RESULT_ACCESS) ] = 0x0,
380                 [ C(RESULT_MISS)   ] = 0,
381         },
382  },
383  [ C(L1I ) ] = {
384         [ C(OP_READ) ] = {
385                 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS                  */
386                 [ C(RESULT_MISS)   ] = 0x0280, /* L1I.MISSES                 */
387         },
388         [ C(OP_WRITE) ] = {
389                 [ C(RESULT_ACCESS) ] = -1,
390                 [ C(RESULT_MISS)   ] = -1,
391         },
392         [ C(OP_PREFETCH) ] = {
393                 [ C(RESULT_ACCESS) ] = 0,
394                 [ C(RESULT_MISS)   ] = 0,
395         },
396  },
397  [ C(LL  ) ] = {
398         [ C(OP_READ) ] = {
399                 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI                 */
400                 [ C(RESULT_MISS)   ] = 0x4129, /* L2_LD.ISTATE               */
401         },
402         [ C(OP_WRITE) ] = {
403                 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI                 */
404                 [ C(RESULT_MISS)   ] = 0x412A, /* L2_ST.ISTATE               */
405         },
406         [ C(OP_PREFETCH) ] = {
407                 [ C(RESULT_ACCESS) ] = 0,
408                 [ C(RESULT_MISS)   ] = 0,
409         },
410  },
411  [ C(DTLB) ] = {
412         [ C(OP_READ) ] = {
413                 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI  (alias) */
414                 [ C(RESULT_MISS)   ] = 0x0508, /* DTLB_MISSES.MISS_LD        */
415         },
416         [ C(OP_WRITE) ] = {
417                 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI  (alias) */
418                 [ C(RESULT_MISS)   ] = 0x0608, /* DTLB_MISSES.MISS_ST        */
419         },
420         [ C(OP_PREFETCH) ] = {
421                 [ C(RESULT_ACCESS) ] = 0,
422                 [ C(RESULT_MISS)   ] = 0,
423         },
424  },
425  [ C(ITLB) ] = {
426         [ C(OP_READ) ] = {
427                 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P         */
428                 [ C(RESULT_MISS)   ] = 0x0282, /* ITLB.MISSES                */
429         },
430         [ C(OP_WRITE) ] = {
431                 [ C(RESULT_ACCESS) ] = -1,
432                 [ C(RESULT_MISS)   ] = -1,
433         },
434         [ C(OP_PREFETCH) ] = {
435                 [ C(RESULT_ACCESS) ] = -1,
436                 [ C(RESULT_MISS)   ] = -1,
437         },
438  },
439  [ C(BPU ) ] = {
440         [ C(OP_READ) ] = {
441                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY        */
442                 [ C(RESULT_MISS)   ] = 0x00c5, /* BP_INST_RETIRED.MISPRED    */
443         },
444         [ C(OP_WRITE) ] = {
445                 [ C(RESULT_ACCESS) ] = -1,
446                 [ C(RESULT_MISS)   ] = -1,
447         },
448         [ C(OP_PREFETCH) ] = {
449                 [ C(RESULT_ACCESS) ] = -1,
450                 [ C(RESULT_MISS)   ] = -1,
451         },
452  },
453 };
454
455 static u64 intel_pmu_raw_event(u64 hw_event)
456 {
457 #define CORE_EVNTSEL_EVENT_MASK         0x000000FFULL
458 #define CORE_EVNTSEL_UNIT_MASK          0x0000FF00ULL
459 #define CORE_EVNTSEL_EDGE_MASK          0x00040000ULL
460 #define CORE_EVNTSEL_INV_MASK           0x00800000ULL
461 #define CORE_EVNTSEL_REG_MASK           0xFF000000ULL
462
463 #define CORE_EVNTSEL_MASK               \
464         (INTEL_ARCH_EVTSEL_MASK |       \
465          INTEL_ARCH_UNIT_MASK   |       \
466          INTEL_ARCH_EDGE_MASK   |       \
467          INTEL_ARCH_INV_MASK    |       \
468          INTEL_ARCH_CNT_MASK)
469
470         return hw_event & CORE_EVNTSEL_MASK;
471 }
472
473 static void intel_pmu_enable_bts(u64 config)
474 {
475         unsigned long debugctlmsr;
476
477         debugctlmsr = get_debugctlmsr();
478
479         debugctlmsr |= X86_DEBUGCTL_TR;
480         debugctlmsr |= X86_DEBUGCTL_BTS;
481         debugctlmsr |= X86_DEBUGCTL_BTINT;
482
483         if (!(config & ARCH_PERFMON_EVENTSEL_OS))
484                 debugctlmsr |= X86_DEBUGCTL_BTS_OFF_OS;
485
486         if (!(config & ARCH_PERFMON_EVENTSEL_USR))
487                 debugctlmsr |= X86_DEBUGCTL_BTS_OFF_USR;
488
489         update_debugctlmsr(debugctlmsr);
490 }
491
492 static void intel_pmu_disable_bts(void)
493 {
494         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
495         unsigned long debugctlmsr;
496
497         if (!cpuc->ds)
498                 return;
499
500         debugctlmsr = get_debugctlmsr();
501
502         debugctlmsr &=
503                 ~(X86_DEBUGCTL_TR | X86_DEBUGCTL_BTS | X86_DEBUGCTL_BTINT |
504                   X86_DEBUGCTL_BTS_OFF_OS | X86_DEBUGCTL_BTS_OFF_USR);
505
506         update_debugctlmsr(debugctlmsr);
507 }
508
509 static void intel_pmu_disable_all(void)
510 {
511         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
512
513         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
514
515         if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask))
516                 intel_pmu_disable_bts();
517 }
518
519 static void intel_pmu_enable_all(void)
520 {
521         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
522
523         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, x86_pmu.intel_ctrl);
524
525         if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
526                 struct perf_event *event =
527                         cpuc->events[X86_PMC_IDX_FIXED_BTS];
528
529                 if (WARN_ON_ONCE(!event))
530                         return;
531
532                 intel_pmu_enable_bts(event->hw.config);
533         }
534 }
535
536 static inline u64 intel_pmu_get_status(void)
537 {
538         u64 status;
539
540         rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
541
542         return status;
543 }
544
545 static inline void intel_pmu_ack_status(u64 ack)
546 {
547         wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
548 }
549
550 static inline void
551 intel_pmu_disable_fixed(struct hw_perf_event *hwc, int __idx)
552 {
553         int idx = __idx - X86_PMC_IDX_FIXED;
554         u64 ctrl_val, mask;
555
556         mask = 0xfULL << (idx * 4);
557
558         rdmsrl(hwc->config_base, ctrl_val);
559         ctrl_val &= ~mask;
560         (void)checking_wrmsrl(hwc->config_base, ctrl_val);
561 }
562
563 static void intel_pmu_drain_bts_buffer(void)
564 {
565         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
566         struct debug_store *ds = cpuc->ds;
567         struct bts_record {
568                 u64     from;
569                 u64     to;
570                 u64     flags;
571         };
572         struct perf_event *event = cpuc->events[X86_PMC_IDX_FIXED_BTS];
573         struct bts_record *at, *top;
574         struct perf_output_handle handle;
575         struct perf_event_header header;
576         struct perf_sample_data data;
577         struct pt_regs regs;
578
579         if (!event)
580                 return;
581
582         if (!ds)
583                 return;
584
585         at  = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
586         top = (struct bts_record *)(unsigned long)ds->bts_index;
587
588         if (top <= at)
589                 return;
590
591         ds->bts_index = ds->bts_buffer_base;
592
593         perf_sample_data_init(&data, 0);
594
595         data.period     = event->hw.last_period;
596         regs.ip         = 0;
597
598         /*
599          * Prepare a generic sample, i.e. fill in the invariant fields.
600          * We will overwrite the from and to address before we output
601          * the sample.
602          */
603         perf_prepare_sample(&header, &data, event, &regs);
604
605         if (perf_output_begin(&handle, event,
606                               header.size * (top - at), 1, 1))
607                 return;
608
609         for (; at < top; at++) {
610                 data.ip         = at->from;
611                 data.addr       = at->to;
612
613                 perf_output_sample(&handle, &header, &data, event);
614         }
615
616         perf_output_end(&handle);
617
618         /* There's new data available. */
619         event->hw.interrupts++;
620         event->pending_kill = POLL_IN;
621 }
622
623 static inline void
624 intel_pmu_disable_event(struct hw_perf_event *hwc, int idx)
625 {
626         if (unlikely(idx == X86_PMC_IDX_FIXED_BTS)) {
627                 intel_pmu_disable_bts();
628                 intel_pmu_drain_bts_buffer();
629                 return;
630         }
631
632         if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
633                 intel_pmu_disable_fixed(hwc, idx);
634                 return;
635         }
636
637         x86_pmu_disable_event(hwc, idx);
638 }
639
640 static inline void
641 intel_pmu_enable_fixed(struct hw_perf_event *hwc, int __idx)
642 {
643         int idx = __idx - X86_PMC_IDX_FIXED;
644         u64 ctrl_val, bits, mask;
645         int err;
646
647         /*
648          * Enable IRQ generation (0x8),
649          * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
650          * if requested:
651          */
652         bits = 0x8ULL;
653         if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
654                 bits |= 0x2;
655         if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
656                 bits |= 0x1;
657
658         /*
659          * ANY bit is supported in v3 and up
660          */
661         if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY)
662                 bits |= 0x4;
663
664         bits <<= (idx * 4);
665         mask = 0xfULL << (idx * 4);
666
667         rdmsrl(hwc->config_base, ctrl_val);
668         ctrl_val &= ~mask;
669         ctrl_val |= bits;
670         err = checking_wrmsrl(hwc->config_base, ctrl_val);
671 }
672
673 static void intel_pmu_enable_event(struct hw_perf_event *hwc, int idx)
674 {
675         if (unlikely(idx == X86_PMC_IDX_FIXED_BTS)) {
676                 if (!__get_cpu_var(cpu_hw_events).enabled)
677                         return;
678
679                 intel_pmu_enable_bts(hwc->config);
680                 return;
681         }
682
683         if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
684                 intel_pmu_enable_fixed(hwc, idx);
685                 return;
686         }
687
688         __x86_pmu_enable_event(hwc, idx);
689 }
690
691 /*
692  * Save and restart an expired event. Called by NMI contexts,
693  * so it has to be careful about preempting normal event ops:
694  */
695 static int intel_pmu_save_and_restart(struct perf_event *event)
696 {
697         struct hw_perf_event *hwc = &event->hw;
698         int idx = hwc->idx;
699         int ret;
700
701         x86_perf_event_update(event, hwc, idx);
702         ret = x86_perf_event_set_period(event, hwc, idx);
703
704         return ret;
705 }
706
707 static void intel_pmu_reset(void)
708 {
709         struct debug_store *ds = __get_cpu_var(cpu_hw_events).ds;
710         unsigned long flags;
711         int idx;
712
713         if (!x86_pmu.num_events)
714                 return;
715
716         local_irq_save(flags);
717
718         printk("clearing PMU state on CPU#%d\n", smp_processor_id());
719
720         for (idx = 0; idx < x86_pmu.num_events; idx++) {
721                 checking_wrmsrl(x86_pmu.eventsel + idx, 0ull);
722                 checking_wrmsrl(x86_pmu.perfctr  + idx, 0ull);
723         }
724         for (idx = 0; idx < x86_pmu.num_events_fixed; idx++) {
725                 checking_wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
726         }
727         if (ds)
728                 ds->bts_index = ds->bts_buffer_base;
729
730         local_irq_restore(flags);
731 }
732
733 /*
734  * This handler is triggered by the local APIC, so the APIC IRQ handling
735  * rules apply:
736  */
737 static int intel_pmu_handle_irq(struct pt_regs *regs)
738 {
739         struct perf_sample_data data;
740         struct cpu_hw_events *cpuc;
741         int bit, loops;
742         u64 ack, status;
743
744         perf_sample_data_init(&data, 0);
745
746         cpuc = &__get_cpu_var(cpu_hw_events);
747
748         perf_disable();
749         intel_pmu_drain_bts_buffer();
750         status = intel_pmu_get_status();
751         if (!status) {
752                 perf_enable();
753                 return 0;
754         }
755
756         loops = 0;
757 again:
758         if (++loops > 100) {
759                 WARN_ONCE(1, "perfevents: irq loop stuck!\n");
760                 perf_event_print_debug();
761                 intel_pmu_reset();
762                 perf_enable();
763                 return 1;
764         }
765
766         inc_irq_stat(apic_perf_irqs);
767         ack = status;
768         for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
769                 struct perf_event *event = cpuc->events[bit];
770
771                 clear_bit(bit, (unsigned long *) &status);
772                 if (!test_bit(bit, cpuc->active_mask))
773                         continue;
774
775                 if (!intel_pmu_save_and_restart(event))
776                         continue;
777
778                 data.period = event->hw.last_period;
779
780                 if (perf_event_overflow(event, 1, &data, regs))
781                         intel_pmu_disable_event(&event->hw, bit);
782         }
783
784         intel_pmu_ack_status(ack);
785
786         /*
787          * Repeat if there is more work to be done:
788          */
789         status = intel_pmu_get_status();
790         if (status)
791                 goto again;
792
793         perf_enable();
794
795         return 1;
796 }
797
798 static struct event_constraint bts_constraint =
799         EVENT_CONSTRAINT(0, 1ULL << X86_PMC_IDX_FIXED_BTS, 0);
800
801 static struct event_constraint *
802 intel_special_constraints(struct perf_event *event)
803 {
804         unsigned int hw_event;
805
806         hw_event = event->hw.config & INTEL_ARCH_EVENT_MASK;
807
808         if (unlikely((hw_event ==
809                       x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS)) &&
810                      (event->hw.sample_period == 1))) {
811
812                 return &bts_constraint;
813         }
814         return NULL;
815 }
816
817 static struct event_constraint *
818 intel_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
819 {
820         struct event_constraint *c;
821
822         c = intel_special_constraints(event);
823         if (c)
824                 return c;
825
826         return x86_get_event_constraints(cpuc, event);
827 }
828
829 static __initconst struct x86_pmu core_pmu = {
830         .name                   = "core",
831         .handle_irq             = x86_pmu_handle_irq,
832         .disable_all            = x86_pmu_disable_all,
833         .enable_all             = x86_pmu_enable_all,
834         .enable                 = x86_pmu_enable_event,
835         .disable                = x86_pmu_disable_event,
836         .eventsel               = MSR_ARCH_PERFMON_EVENTSEL0,
837         .perfctr                = MSR_ARCH_PERFMON_PERFCTR0,
838         .event_map              = intel_pmu_event_map,
839         .raw_event              = intel_pmu_raw_event,
840         .max_events             = ARRAY_SIZE(intel_perfmon_event_map),
841         .apic                   = 1,
842         /*
843          * Intel PMCs cannot be accessed sanely above 32 bit width,
844          * so we install an artificial 1<<31 period regardless of
845          * the generic event period:
846          */
847         .max_period             = (1ULL << 31) - 1,
848         .get_event_constraints  = intel_get_event_constraints,
849         .event_constraints      = intel_core_event_constraints,
850 };
851
852 static __initconst struct x86_pmu intel_pmu = {
853         .name                   = "Intel",
854         .handle_irq             = intel_pmu_handle_irq,
855         .disable_all            = intel_pmu_disable_all,
856         .enable_all             = intel_pmu_enable_all,
857         .enable                 = intel_pmu_enable_event,
858         .disable                = intel_pmu_disable_event,
859         .eventsel               = MSR_ARCH_PERFMON_EVENTSEL0,
860         .perfctr                = MSR_ARCH_PERFMON_PERFCTR0,
861         .event_map              = intel_pmu_event_map,
862         .raw_event              = intel_pmu_raw_event,
863         .max_events             = ARRAY_SIZE(intel_perfmon_event_map),
864         .apic                   = 1,
865         /*
866          * Intel PMCs cannot be accessed sanely above 32 bit width,
867          * so we install an artificial 1<<31 period regardless of
868          * the generic event period:
869          */
870         .max_period             = (1ULL << 31) - 1,
871         .enable_bts             = intel_pmu_enable_bts,
872         .disable_bts            = intel_pmu_disable_bts,
873         .get_event_constraints  = intel_get_event_constraints
874 };
875
876 static __init int intel_pmu_init(void)
877 {
878         union cpuid10_edx edx;
879         union cpuid10_eax eax;
880         unsigned int unused;
881         unsigned int ebx;
882         int version;
883
884         if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
885                 /* check for P6 processor family */
886            if (boot_cpu_data.x86 == 6) {
887                 return p6_pmu_init();
888            } else {
889                 return -ENODEV;
890            }
891         }
892
893         /*
894          * Check whether the Architectural PerfMon supports
895          * Branch Misses Retired hw_event or not.
896          */
897         cpuid(10, &eax.full, &ebx, &unused, &edx.full);
898         if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED)
899                 return -ENODEV;
900
901         version = eax.split.version_id;
902         if (version < 2)
903                 x86_pmu = core_pmu;
904         else
905                 x86_pmu = intel_pmu;
906
907         x86_pmu.version                 = version;
908         x86_pmu.num_events              = eax.split.num_events;
909         x86_pmu.event_bits              = eax.split.bit_width;
910         x86_pmu.event_mask              = (1ULL << eax.split.bit_width) - 1;
911
912         /*
913          * Quirk: v2 perfmon does not report fixed-purpose events, so
914          * assume at least 3 events:
915          */
916         if (version > 1)
917                 x86_pmu.num_events_fixed = max((int)edx.split.num_events_fixed, 3);
918
919         /*
920          * Install the hw-cache-events table:
921          */
922         switch (boot_cpu_data.x86_model) {
923         case 14: /* 65 nm core solo/duo, "Yonah" */
924                 pr_cont("Core events, ");
925                 break;
926
927         case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */
928         case 22: /* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */
929         case 23: /* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */
930         case 29: /* six-core 45 nm xeon "Dunnington" */
931                 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
932                        sizeof(hw_cache_event_ids));
933
934                 x86_pmu.event_constraints = intel_core2_event_constraints;
935                 pr_cont("Core2 events, ");
936                 break;
937
938         case 26: /* 45 nm nehalem, "Bloomfield" */
939         case 30: /* 45 nm nehalem, "Lynnfield" */
940                 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
941                        sizeof(hw_cache_event_ids));
942
943                 x86_pmu.event_constraints = intel_nehalem_event_constraints;
944                 pr_cont("Nehalem/Corei7 events, ");
945                 break;
946         case 28: /* Atom */
947                 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
948                        sizeof(hw_cache_event_ids));
949
950                 x86_pmu.event_constraints = intel_gen_event_constraints;
951                 pr_cont("Atom events, ");
952                 break;
953
954         case 37: /* 32 nm nehalem, "Clarkdale" */
955         case 44: /* 32 nm nehalem, "Gulftown" */
956                 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids,
957                        sizeof(hw_cache_event_ids));
958
959                 x86_pmu.event_constraints = intel_westmere_event_constraints;
960                 pr_cont("Westmere events, ");
961                 break;
962
963         default:
964                 /*
965                  * default constraints for v2 and up
966                  */
967                 x86_pmu.event_constraints = intel_gen_event_constraints;
968                 pr_cont("generic architected perfmon, ");
969         }
970         return 0;
971 }
972
973 #else /* CONFIG_CPU_SUP_INTEL */
974
975 static int intel_pmu_init(void)
976 {
977         return 0;
978 }
979
980 #endif /* CONFIG_CPU_SUP_INTEL */