]> bbs.cooldavid.org Git - net-next-2.6.git/blob - arch/x86/kernel/cpu/perf_event_intel.c
perf, x86: Add Nehelem PMU programming errata workaround
[net-next-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_disable_all(void)
474 {
475         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
476
477         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
478
479         if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask))
480                 intel_pmu_disable_bts();
481
482         intel_pmu_pebs_disable_all();
483         intel_pmu_lbr_disable_all();
484 }
485
486 static void intel_pmu_enable_all(int added)
487 {
488         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
489
490         intel_pmu_pebs_enable_all();
491         intel_pmu_lbr_enable_all();
492         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, x86_pmu.intel_ctrl);
493
494         if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
495                 struct perf_event *event =
496                         cpuc->events[X86_PMC_IDX_FIXED_BTS];
497
498                 if (WARN_ON_ONCE(!event))
499                         return;
500
501                 intel_pmu_enable_bts(event->hw.config);
502         }
503 }
504
505 /*
506  * Workaround for:
507  *   Intel Errata AAK100 (model 26)
508  *   Intel Errata AAP53  (model 30)
509  *
510  * These chips need to be 'reset' when adding counters by programming
511  * the magic three (non counting) events 0x4300D2, 0x4300B1 and 0x4300B5
512  * either in sequence on the same PMC or on different PMCs.
513  */
514 static void intel_pmu_nhm_enable_all(int added)
515 {
516         if (added) {
517                 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
518                 int i;
519
520                 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + 0, 0x4300D2);
521                 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + 1, 0x4300B1);
522                 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + 2, 0x4300B5);
523
524                 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x3);
525                 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0);
526
527                 for (i = 0; i < 3; i++) {
528                         struct perf_event *event = cpuc->events[i];
529
530                         if (!event)
531                                 continue;
532
533                         __x86_pmu_enable_event(&event->hw);
534                 }
535         }
536         intel_pmu_enable_all(added);
537 }
538
539 static inline u64 intel_pmu_get_status(void)
540 {
541         u64 status;
542
543         rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
544
545         return status;
546 }
547
548 static inline void intel_pmu_ack_status(u64 ack)
549 {
550         wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
551 }
552
553 static void intel_pmu_disable_fixed(struct hw_perf_event *hwc)
554 {
555         int idx = hwc->idx - X86_PMC_IDX_FIXED;
556         u64 ctrl_val, mask;
557
558         mask = 0xfULL << (idx * 4);
559
560         rdmsrl(hwc->config_base, ctrl_val);
561         ctrl_val &= ~mask;
562         wrmsrl(hwc->config_base, ctrl_val);
563 }
564
565 static void intel_pmu_disable_event(struct perf_event *event)
566 {
567         struct hw_perf_event *hwc = &event->hw;
568
569         if (unlikely(hwc->idx == X86_PMC_IDX_FIXED_BTS)) {
570                 intel_pmu_disable_bts();
571                 intel_pmu_drain_bts_buffer();
572                 return;
573         }
574
575         if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
576                 intel_pmu_disable_fixed(hwc);
577                 return;
578         }
579
580         x86_pmu_disable_event(event);
581
582         if (unlikely(event->attr.precise))
583                 intel_pmu_pebs_disable(event);
584 }
585
586 static void intel_pmu_enable_fixed(struct hw_perf_event *hwc)
587 {
588         int idx = hwc->idx - X86_PMC_IDX_FIXED;
589         u64 ctrl_val, bits, mask;
590
591         /*
592          * Enable IRQ generation (0x8),
593          * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
594          * if requested:
595          */
596         bits = 0x8ULL;
597         if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
598                 bits |= 0x2;
599         if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
600                 bits |= 0x1;
601
602         /*
603          * ANY bit is supported in v3 and up
604          */
605         if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY)
606                 bits |= 0x4;
607
608         bits <<= (idx * 4);
609         mask = 0xfULL << (idx * 4);
610
611         rdmsrl(hwc->config_base, ctrl_val);
612         ctrl_val &= ~mask;
613         ctrl_val |= bits;
614         wrmsrl(hwc->config_base, ctrl_val);
615 }
616
617 static void intel_pmu_enable_event(struct perf_event *event)
618 {
619         struct hw_perf_event *hwc = &event->hw;
620
621         if (unlikely(hwc->idx == X86_PMC_IDX_FIXED_BTS)) {
622                 if (!__get_cpu_var(cpu_hw_events).enabled)
623                         return;
624
625                 intel_pmu_enable_bts(hwc->config);
626                 return;
627         }
628
629         if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
630                 intel_pmu_enable_fixed(hwc);
631                 return;
632         }
633
634         if (unlikely(event->attr.precise))
635                 intel_pmu_pebs_enable(event);
636
637         __x86_pmu_enable_event(hwc);
638 }
639
640 /*
641  * Save and restart an expired event. Called by NMI contexts,
642  * so it has to be careful about preempting normal event ops:
643  */
644 static int intel_pmu_save_and_restart(struct perf_event *event)
645 {
646         x86_perf_event_update(event);
647         return x86_perf_event_set_period(event);
648 }
649
650 static void intel_pmu_reset(void)
651 {
652         struct debug_store *ds = __get_cpu_var(cpu_hw_events).ds;
653         unsigned long flags;
654         int idx;
655
656         if (!x86_pmu.num_events)
657                 return;
658
659         local_irq_save(flags);
660
661         printk("clearing PMU state on CPU#%d\n", smp_processor_id());
662
663         for (idx = 0; idx < x86_pmu.num_events; idx++) {
664                 checking_wrmsrl(x86_pmu.eventsel + idx, 0ull);
665                 checking_wrmsrl(x86_pmu.perfctr  + idx, 0ull);
666         }
667         for (idx = 0; idx < x86_pmu.num_events_fixed; idx++) {
668                 checking_wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
669         }
670         if (ds)
671                 ds->bts_index = ds->bts_buffer_base;
672
673         local_irq_restore(flags);
674 }
675
676 /*
677  * This handler is triggered by the local APIC, so the APIC IRQ handling
678  * rules apply:
679  */
680 static int intel_pmu_handle_irq(struct pt_regs *regs)
681 {
682         struct perf_sample_data data;
683         struct cpu_hw_events *cpuc;
684         int bit, loops;
685         u64 ack, status;
686
687         perf_sample_data_init(&data, 0);
688
689         cpuc = &__get_cpu_var(cpu_hw_events);
690
691         intel_pmu_disable_all();
692         intel_pmu_drain_bts_buffer();
693         status = intel_pmu_get_status();
694         if (!status) {
695                 intel_pmu_enable_all(0);
696                 return 0;
697         }
698
699         loops = 0;
700 again:
701         if (++loops > 100) {
702                 WARN_ONCE(1, "perfevents: irq loop stuck!\n");
703                 perf_event_print_debug();
704                 intel_pmu_reset();
705                 goto done;
706         }
707
708         inc_irq_stat(apic_perf_irqs);
709         ack = status;
710
711         intel_pmu_lbr_read();
712
713         /*
714          * PEBS overflow sets bit 62 in the global status register
715          */
716         if (__test_and_clear_bit(62, (unsigned long *)&status))
717                 x86_pmu.drain_pebs(regs);
718
719         for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
720                 struct perf_event *event = cpuc->events[bit];
721
722                 if (!test_bit(bit, cpuc->active_mask))
723                         continue;
724
725                 if (!intel_pmu_save_and_restart(event))
726                         continue;
727
728                 data.period = event->hw.last_period;
729
730                 if (perf_event_overflow(event, 1, &data, regs))
731                         x86_pmu_stop(event);
732         }
733
734         intel_pmu_ack_status(ack);
735
736         /*
737          * Repeat if there is more work to be done:
738          */
739         status = intel_pmu_get_status();
740         if (status)
741                 goto again;
742
743 done:
744         intel_pmu_enable_all(0);
745         return 1;
746 }
747
748 static struct event_constraint *
749 intel_bts_constraints(struct perf_event *event)
750 {
751         struct hw_perf_event *hwc = &event->hw;
752         unsigned int hw_event, bts_event;
753
754         hw_event = hwc->config & INTEL_ARCH_EVENT_MASK;
755         bts_event = x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS);
756
757         if (unlikely(hw_event == bts_event && hwc->sample_period == 1))
758                 return &bts_constraint;
759
760         return NULL;
761 }
762
763 static struct event_constraint *
764 intel_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
765 {
766         struct event_constraint *c;
767
768         c = intel_bts_constraints(event);
769         if (c)
770                 return c;
771
772         c = intel_pebs_constraints(event);
773         if (c)
774                 return c;
775
776         return x86_get_event_constraints(cpuc, event);
777 }
778
779 static __initconst struct x86_pmu core_pmu = {
780         .name                   = "core",
781         .handle_irq             = x86_pmu_handle_irq,
782         .disable_all            = x86_pmu_disable_all,
783         .enable_all             = x86_pmu_enable_all,
784         .enable                 = x86_pmu_enable_event,
785         .disable                = x86_pmu_disable_event,
786         .hw_config              = x86_hw_config,
787         .schedule_events        = x86_schedule_events,
788         .eventsel               = MSR_ARCH_PERFMON_EVENTSEL0,
789         .perfctr                = MSR_ARCH_PERFMON_PERFCTR0,
790         .event_map              = intel_pmu_event_map,
791         .raw_event              = intel_pmu_raw_event,
792         .max_events             = ARRAY_SIZE(intel_perfmon_event_map),
793         .apic                   = 1,
794         /*
795          * Intel PMCs cannot be accessed sanely above 32 bit width,
796          * so we install an artificial 1<<31 period regardless of
797          * the generic event period:
798          */
799         .max_period             = (1ULL << 31) - 1,
800         .get_event_constraints  = intel_get_event_constraints,
801         .event_constraints      = intel_core_event_constraints,
802 };
803
804 static void intel_pmu_cpu_starting(int cpu)
805 {
806         init_debug_store_on_cpu(cpu);
807         /*
808          * Deal with CPUs that don't clear their LBRs on power-up.
809          */
810         intel_pmu_lbr_reset();
811 }
812
813 static void intel_pmu_cpu_dying(int cpu)
814 {
815         fini_debug_store_on_cpu(cpu);
816 }
817
818 static __initconst struct x86_pmu intel_pmu = {
819         .name                   = "Intel",
820         .handle_irq             = intel_pmu_handle_irq,
821         .disable_all            = intel_pmu_disable_all,
822         .enable_all             = intel_pmu_enable_all,
823         .enable                 = intel_pmu_enable_event,
824         .disable                = intel_pmu_disable_event,
825         .hw_config              = x86_hw_config,
826         .schedule_events        = x86_schedule_events,
827         .eventsel               = MSR_ARCH_PERFMON_EVENTSEL0,
828         .perfctr                = MSR_ARCH_PERFMON_PERFCTR0,
829         .event_map              = intel_pmu_event_map,
830         .raw_event              = intel_pmu_raw_event,
831         .max_events             = ARRAY_SIZE(intel_perfmon_event_map),
832         .apic                   = 1,
833         /*
834          * Intel PMCs cannot be accessed sanely above 32 bit width,
835          * so we install an artificial 1<<31 period regardless of
836          * the generic event period:
837          */
838         .max_period             = (1ULL << 31) - 1,
839         .get_event_constraints  = intel_get_event_constraints,
840
841         .cpu_starting           = intel_pmu_cpu_starting,
842         .cpu_dying              = intel_pmu_cpu_dying,
843 };
844
845 static void intel_clovertown_quirks(void)
846 {
847         /*
848          * PEBS is unreliable due to:
849          *
850          *   AJ67  - PEBS may experience CPL leaks
851          *   AJ68  - PEBS PMI may be delayed by one event
852          *   AJ69  - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12]
853          *   AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS
854          *
855          * AJ67 could be worked around by restricting the OS/USR flags.
856          * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI.
857          *
858          * AJ106 could possibly be worked around by not allowing LBR
859          *       usage from PEBS, including the fixup.
860          * AJ68  could possibly be worked around by always programming
861          *       a pebs_event_reset[0] value and coping with the lost events.
862          *
863          * But taken together it might just make sense to not enable PEBS on
864          * these chips.
865          */
866         printk(KERN_WARNING "PEBS disabled due to CPU errata.\n");
867         x86_pmu.pebs = 0;
868         x86_pmu.pebs_constraints = NULL;
869 }
870
871 static __init int intel_pmu_init(void)
872 {
873         union cpuid10_edx edx;
874         union cpuid10_eax eax;
875         unsigned int unused;
876         unsigned int ebx;
877         int version;
878
879         if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
880                 switch (boot_cpu_data.x86) {
881                 case 0x6:
882                         return p6_pmu_init();
883                 case 0xf:
884                         return p4_pmu_init();
885                 }
886                 return -ENODEV;
887         }
888
889         /*
890          * Check whether the Architectural PerfMon supports
891          * Branch Misses Retired hw_event or not.
892          */
893         cpuid(10, &eax.full, &ebx, &unused, &edx.full);
894         if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED)
895                 return -ENODEV;
896
897         version = eax.split.version_id;
898         if (version < 2)
899                 x86_pmu = core_pmu;
900         else
901                 x86_pmu = intel_pmu;
902
903         x86_pmu.version                 = version;
904         x86_pmu.num_events              = eax.split.num_events;
905         x86_pmu.event_bits              = eax.split.bit_width;
906         x86_pmu.event_mask              = (1ULL << eax.split.bit_width) - 1;
907
908         /*
909          * Quirk: v2 perfmon does not report fixed-purpose events, so
910          * assume at least 3 events:
911          */
912         if (version > 1)
913                 x86_pmu.num_events_fixed = max((int)edx.split.num_events_fixed, 3);
914
915         /*
916          * v2 and above have a perf capabilities MSR
917          */
918         if (version > 1) {
919                 u64 capabilities;
920
921                 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities);
922                 x86_pmu.intel_cap.capabilities = capabilities;
923         }
924
925         intel_ds_init();
926
927         /*
928          * Install the hw-cache-events table:
929          */
930         switch (boot_cpu_data.x86_model) {
931         case 14: /* 65 nm core solo/duo, "Yonah" */
932                 pr_cont("Core events, ");
933                 break;
934
935         case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */
936                 x86_pmu.quirks = intel_clovertown_quirks;
937         case 22: /* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */
938         case 23: /* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */
939         case 29: /* six-core 45 nm xeon "Dunnington" */
940                 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
941                        sizeof(hw_cache_event_ids));
942
943                 intel_pmu_lbr_init_core();
944
945                 x86_pmu.event_constraints = intel_core2_event_constraints;
946                 pr_cont("Core2 events, ");
947                 break;
948
949         case 26: /* 45 nm nehalem, "Bloomfield" */
950         case 30: /* 45 nm nehalem, "Lynnfield" */
951                 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
952                        sizeof(hw_cache_event_ids));
953
954                 intel_pmu_lbr_init_nhm();
955
956                 x86_pmu.event_constraints = intel_nehalem_event_constraints;
957                 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
958                 pr_cont("Nehalem events, ");
959                 break;
960
961         case 28: /* Atom */
962                 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
963                        sizeof(hw_cache_event_ids));
964
965                 intel_pmu_lbr_init_atom();
966
967                 x86_pmu.event_constraints = intel_gen_event_constraints;
968                 pr_cont("Atom events, ");
969                 break;
970
971         case 37: /* 32 nm nehalem, "Clarkdale" */
972         case 44: /* 32 nm nehalem, "Gulftown" */
973                 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids,
974                        sizeof(hw_cache_event_ids));
975
976                 intel_pmu_lbr_init_nhm();
977
978                 x86_pmu.event_constraints = intel_westmere_event_constraints;
979                 pr_cont("Westmere events, ");
980                 break;
981
982         default:
983                 /*
984                  * default constraints for v2 and up
985                  */
986                 x86_pmu.event_constraints = intel_gen_event_constraints;
987                 pr_cont("generic architected perfmon, ");
988         }
989         return 0;
990 }
991
992 #else /* CONFIG_CPU_SUP_INTEL */
993
994 static int intel_pmu_init(void)
995 {
996         return 0;
997 }
998
999 #endif /* CONFIG_CPU_SUP_INTEL */