Merge git://git.kernel.org/pub/scm/linux/kernel/git/herbert/crypto-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         INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */
76         EVENT_CONSTRAINT_END
77 };
78
79 static struct event_constraint intel_gen_event_constraints[] =
80 {
81         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
82         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
83         /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
84         EVENT_CONSTRAINT_END
85 };
86
87 static u64 intel_pmu_event_map(int hw_event)
88 {
89         return intel_perfmon_event_map[hw_event];
90 }
91
92 static __initconst const u64 westmere_hw_cache_event_ids
93                                 [PERF_COUNT_HW_CACHE_MAX]
94                                 [PERF_COUNT_HW_CACHE_OP_MAX]
95                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
96 {
97  [ C(L1D) ] = {
98         [ C(OP_READ) ] = {
99                 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS       */
100                 [ C(RESULT_MISS)   ] = 0x0151, /* L1D.REPL                     */
101         },
102         [ C(OP_WRITE) ] = {
103                 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES      */
104                 [ C(RESULT_MISS)   ] = 0x0251, /* L1D.M_REPL                   */
105         },
106         [ C(OP_PREFETCH) ] = {
107                 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS        */
108                 [ C(RESULT_MISS)   ] = 0x024e, /* L1D_PREFETCH.MISS            */
109         },
110  },
111  [ C(L1I ) ] = {
112         [ C(OP_READ) ] = {
113                 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS                    */
114                 [ C(RESULT_MISS)   ] = 0x0280, /* L1I.MISSES                   */
115         },
116         [ C(OP_WRITE) ] = {
117                 [ C(RESULT_ACCESS) ] = -1,
118                 [ C(RESULT_MISS)   ] = -1,
119         },
120         [ C(OP_PREFETCH) ] = {
121                 [ C(RESULT_ACCESS) ] = 0x0,
122                 [ C(RESULT_MISS)   ] = 0x0,
123         },
124  },
125  [ C(LL  ) ] = {
126         [ C(OP_READ) ] = {
127                 [ C(RESULT_ACCESS) ] = 0x0324, /* L2_RQSTS.LOADS               */
128                 [ C(RESULT_MISS)   ] = 0x0224, /* L2_RQSTS.LD_MISS             */
129         },
130         [ C(OP_WRITE) ] = {
131                 [ C(RESULT_ACCESS) ] = 0x0c24, /* L2_RQSTS.RFOS                */
132                 [ C(RESULT_MISS)   ] = 0x0824, /* L2_RQSTS.RFO_MISS            */
133         },
134         [ C(OP_PREFETCH) ] = {
135                 [ C(RESULT_ACCESS) ] = 0x4f2e, /* LLC Reference                */
136                 [ C(RESULT_MISS)   ] = 0x412e, /* LLC Misses                   */
137         },
138  },
139  [ C(DTLB) ] = {
140         [ C(OP_READ) ] = {
141                 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS       */
142                 [ C(RESULT_MISS)   ] = 0x0108, /* DTLB_LOAD_MISSES.ANY         */
143         },
144         [ C(OP_WRITE) ] = {
145                 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES      */
146                 [ C(RESULT_MISS)   ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS  */
147         },
148         [ C(OP_PREFETCH) ] = {
149                 [ C(RESULT_ACCESS) ] = 0x0,
150                 [ C(RESULT_MISS)   ] = 0x0,
151         },
152  },
153  [ C(ITLB) ] = {
154         [ C(OP_READ) ] = {
155                 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P           */
156                 [ C(RESULT_MISS)   ] = 0x0185, /* ITLB_MISSES.ANY              */
157         },
158         [ C(OP_WRITE) ] = {
159                 [ C(RESULT_ACCESS) ] = -1,
160                 [ C(RESULT_MISS)   ] = -1,
161         },
162         [ C(OP_PREFETCH) ] = {
163                 [ C(RESULT_ACCESS) ] = -1,
164                 [ C(RESULT_MISS)   ] = -1,
165         },
166  },
167  [ C(BPU ) ] = {
168         [ C(OP_READ) ] = {
169                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
170                 [ C(RESULT_MISS)   ] = 0x03e8, /* BPU_CLEARS.ANY               */
171         },
172         [ C(OP_WRITE) ] = {
173                 [ C(RESULT_ACCESS) ] = -1,
174                 [ C(RESULT_MISS)   ] = -1,
175         },
176         [ C(OP_PREFETCH) ] = {
177                 [ C(RESULT_ACCESS) ] = -1,
178                 [ C(RESULT_MISS)   ] = -1,
179         },
180  },
181 };
182
183 static __initconst const u64 nehalem_hw_cache_event_ids
184                                 [PERF_COUNT_HW_CACHE_MAX]
185                                 [PERF_COUNT_HW_CACHE_OP_MAX]
186                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
187 {
188  [ C(L1D) ] = {
189         [ C(OP_READ) ] = {
190                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI            */
191                 [ C(RESULT_MISS)   ] = 0x0140, /* L1D_CACHE_LD.I_STATE         */
192         },
193         [ C(OP_WRITE) ] = {
194                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI            */
195                 [ C(RESULT_MISS)   ] = 0x0141, /* L1D_CACHE_ST.I_STATE         */
196         },
197         [ C(OP_PREFETCH) ] = {
198                 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS        */
199                 [ C(RESULT_MISS)   ] = 0x024e, /* L1D_PREFETCH.MISS            */
200         },
201  },
202  [ C(L1I ) ] = {
203         [ C(OP_READ) ] = {
204                 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS                    */
205                 [ C(RESULT_MISS)   ] = 0x0280, /* L1I.MISSES                   */
206         },
207         [ C(OP_WRITE) ] = {
208                 [ C(RESULT_ACCESS) ] = -1,
209                 [ C(RESULT_MISS)   ] = -1,
210         },
211         [ C(OP_PREFETCH) ] = {
212                 [ C(RESULT_ACCESS) ] = 0x0,
213                 [ C(RESULT_MISS)   ] = 0x0,
214         },
215  },
216  [ C(LL  ) ] = {
217         [ C(OP_READ) ] = {
218                 [ C(RESULT_ACCESS) ] = 0x0324, /* L2_RQSTS.LOADS               */
219                 [ C(RESULT_MISS)   ] = 0x0224, /* L2_RQSTS.LD_MISS             */
220         },
221         [ C(OP_WRITE) ] = {
222                 [ C(RESULT_ACCESS) ] = 0x0c24, /* L2_RQSTS.RFOS                */
223                 [ C(RESULT_MISS)   ] = 0x0824, /* L2_RQSTS.RFO_MISS            */
224         },
225         [ C(OP_PREFETCH) ] = {
226                 [ C(RESULT_ACCESS) ] = 0x4f2e, /* LLC Reference                */
227                 [ C(RESULT_MISS)   ] = 0x412e, /* LLC Misses                   */
228         },
229  },
230  [ C(DTLB) ] = {
231         [ C(OP_READ) ] = {
232                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI   (alias)  */
233                 [ C(RESULT_MISS)   ] = 0x0108, /* DTLB_LOAD_MISSES.ANY         */
234         },
235         [ C(OP_WRITE) ] = {
236                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI   (alias)  */
237                 [ C(RESULT_MISS)   ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS  */
238         },
239         [ C(OP_PREFETCH) ] = {
240                 [ C(RESULT_ACCESS) ] = 0x0,
241                 [ C(RESULT_MISS)   ] = 0x0,
242         },
243  },
244  [ C(ITLB) ] = {
245         [ C(OP_READ) ] = {
246                 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P           */
247                 [ C(RESULT_MISS)   ] = 0x20c8, /* ITLB_MISS_RETIRED            */
248         },
249         [ C(OP_WRITE) ] = {
250                 [ C(RESULT_ACCESS) ] = -1,
251                 [ C(RESULT_MISS)   ] = -1,
252         },
253         [ C(OP_PREFETCH) ] = {
254                 [ C(RESULT_ACCESS) ] = -1,
255                 [ C(RESULT_MISS)   ] = -1,
256         },
257  },
258  [ C(BPU ) ] = {
259         [ C(OP_READ) ] = {
260                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
261                 [ C(RESULT_MISS)   ] = 0x03e8, /* BPU_CLEARS.ANY               */
262         },
263         [ C(OP_WRITE) ] = {
264                 [ C(RESULT_ACCESS) ] = -1,
265                 [ C(RESULT_MISS)   ] = -1,
266         },
267         [ C(OP_PREFETCH) ] = {
268                 [ C(RESULT_ACCESS) ] = -1,
269                 [ C(RESULT_MISS)   ] = -1,
270         },
271  },
272 };
273
274 static __initconst const u64 core2_hw_cache_event_ids
275                                 [PERF_COUNT_HW_CACHE_MAX]
276                                 [PERF_COUNT_HW_CACHE_OP_MAX]
277                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
278 {
279  [ C(L1D) ] = {
280         [ C(OP_READ) ] = {
281                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI          */
282                 [ C(RESULT_MISS)   ] = 0x0140, /* L1D_CACHE_LD.I_STATE       */
283         },
284         [ C(OP_WRITE) ] = {
285                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI          */
286                 [ C(RESULT_MISS)   ] = 0x0141, /* L1D_CACHE_ST.I_STATE       */
287         },
288         [ C(OP_PREFETCH) ] = {
289                 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS      */
290                 [ C(RESULT_MISS)   ] = 0,
291         },
292  },
293  [ C(L1I ) ] = {
294         [ C(OP_READ) ] = {
295                 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS                  */
296                 [ C(RESULT_MISS)   ] = 0x0081, /* L1I.MISSES                 */
297         },
298         [ C(OP_WRITE) ] = {
299                 [ C(RESULT_ACCESS) ] = -1,
300                 [ C(RESULT_MISS)   ] = -1,
301         },
302         [ C(OP_PREFETCH) ] = {
303                 [ C(RESULT_ACCESS) ] = 0,
304                 [ C(RESULT_MISS)   ] = 0,
305         },
306  },
307  [ C(LL  ) ] = {
308         [ C(OP_READ) ] = {
309                 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI                 */
310                 [ C(RESULT_MISS)   ] = 0x4129, /* L2_LD.ISTATE               */
311         },
312         [ C(OP_WRITE) ] = {
313                 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI                 */
314                 [ C(RESULT_MISS)   ] = 0x412A, /* L2_ST.ISTATE               */
315         },
316         [ C(OP_PREFETCH) ] = {
317                 [ C(RESULT_ACCESS) ] = 0,
318                 [ C(RESULT_MISS)   ] = 0,
319         },
320  },
321  [ C(DTLB) ] = {
322         [ C(OP_READ) ] = {
323                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI  (alias) */
324                 [ C(RESULT_MISS)   ] = 0x0208, /* DTLB_MISSES.MISS_LD        */
325         },
326         [ C(OP_WRITE) ] = {
327                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI  (alias) */
328                 [ C(RESULT_MISS)   ] = 0x0808, /* DTLB_MISSES.MISS_ST        */
329         },
330         [ C(OP_PREFETCH) ] = {
331                 [ C(RESULT_ACCESS) ] = 0,
332                 [ C(RESULT_MISS)   ] = 0,
333         },
334  },
335  [ C(ITLB) ] = {
336         [ C(OP_READ) ] = {
337                 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P         */
338                 [ C(RESULT_MISS)   ] = 0x1282, /* ITLBMISSES                 */
339         },
340         [ C(OP_WRITE) ] = {
341                 [ C(RESULT_ACCESS) ] = -1,
342                 [ C(RESULT_MISS)   ] = -1,
343         },
344         [ C(OP_PREFETCH) ] = {
345                 [ C(RESULT_ACCESS) ] = -1,
346                 [ C(RESULT_MISS)   ] = -1,
347         },
348  },
349  [ C(BPU ) ] = {
350         [ C(OP_READ) ] = {
351                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY        */
352                 [ C(RESULT_MISS)   ] = 0x00c5, /* BP_INST_RETIRED.MISPRED    */
353         },
354         [ C(OP_WRITE) ] = {
355                 [ C(RESULT_ACCESS) ] = -1,
356                 [ C(RESULT_MISS)   ] = -1,
357         },
358         [ C(OP_PREFETCH) ] = {
359                 [ C(RESULT_ACCESS) ] = -1,
360                 [ C(RESULT_MISS)   ] = -1,
361         },
362  },
363 };
364
365 static __initconst const u64 atom_hw_cache_event_ids
366                                 [PERF_COUNT_HW_CACHE_MAX]
367                                 [PERF_COUNT_HW_CACHE_OP_MAX]
368                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
369 {
370  [ C(L1D) ] = {
371         [ C(OP_READ) ] = {
372                 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD               */
373                 [ C(RESULT_MISS)   ] = 0,
374         },
375         [ C(OP_WRITE) ] = {
376                 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST               */
377                 [ C(RESULT_MISS)   ] = 0,
378         },
379         [ C(OP_PREFETCH) ] = {
380                 [ C(RESULT_ACCESS) ] = 0x0,
381                 [ C(RESULT_MISS)   ] = 0,
382         },
383  },
384  [ C(L1I ) ] = {
385         [ C(OP_READ) ] = {
386                 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS                  */
387                 [ C(RESULT_MISS)   ] = 0x0280, /* L1I.MISSES                 */
388         },
389         [ C(OP_WRITE) ] = {
390                 [ C(RESULT_ACCESS) ] = -1,
391                 [ C(RESULT_MISS)   ] = -1,
392         },
393         [ C(OP_PREFETCH) ] = {
394                 [ C(RESULT_ACCESS) ] = 0,
395                 [ C(RESULT_MISS)   ] = 0,
396         },
397  },
398  [ C(LL  ) ] = {
399         [ C(OP_READ) ] = {
400                 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI                 */
401                 [ C(RESULT_MISS)   ] = 0x4129, /* L2_LD.ISTATE               */
402         },
403         [ C(OP_WRITE) ] = {
404                 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI                 */
405                 [ C(RESULT_MISS)   ] = 0x412A, /* L2_ST.ISTATE               */
406         },
407         [ C(OP_PREFETCH) ] = {
408                 [ C(RESULT_ACCESS) ] = 0,
409                 [ C(RESULT_MISS)   ] = 0,
410         },
411  },
412  [ C(DTLB) ] = {
413         [ C(OP_READ) ] = {
414                 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI  (alias) */
415                 [ C(RESULT_MISS)   ] = 0x0508, /* DTLB_MISSES.MISS_LD        */
416         },
417         [ C(OP_WRITE) ] = {
418                 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI  (alias) */
419                 [ C(RESULT_MISS)   ] = 0x0608, /* DTLB_MISSES.MISS_ST        */
420         },
421         [ C(OP_PREFETCH) ] = {
422                 [ C(RESULT_ACCESS) ] = 0,
423                 [ C(RESULT_MISS)   ] = 0,
424         },
425  },
426  [ C(ITLB) ] = {
427         [ C(OP_READ) ] = {
428                 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P         */
429                 [ C(RESULT_MISS)   ] = 0x0282, /* ITLB.MISSES                */
430         },
431         [ C(OP_WRITE) ] = {
432                 [ C(RESULT_ACCESS) ] = -1,
433                 [ C(RESULT_MISS)   ] = -1,
434         },
435         [ C(OP_PREFETCH) ] = {
436                 [ C(RESULT_ACCESS) ] = -1,
437                 [ C(RESULT_MISS)   ] = -1,
438         },
439  },
440  [ C(BPU ) ] = {
441         [ C(OP_READ) ] = {
442                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY        */
443                 [ C(RESULT_MISS)   ] = 0x00c5, /* BP_INST_RETIRED.MISPRED    */
444         },
445         [ C(OP_WRITE) ] = {
446                 [ C(RESULT_ACCESS) ] = -1,
447                 [ C(RESULT_MISS)   ] = -1,
448         },
449         [ C(OP_PREFETCH) ] = {
450                 [ C(RESULT_ACCESS) ] = -1,
451                 [ C(RESULT_MISS)   ] = -1,
452         },
453  },
454 };
455
456 static void intel_pmu_disable_all(void)
457 {
458         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
459
460         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
461
462         if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask))
463                 intel_pmu_disable_bts();
464
465         intel_pmu_pebs_disable_all();
466         intel_pmu_lbr_disable_all();
467 }
468
469 static void intel_pmu_enable_all(int added)
470 {
471         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
472
473         intel_pmu_pebs_enable_all();
474         intel_pmu_lbr_enable_all();
475         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, x86_pmu.intel_ctrl);
476
477         if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
478                 struct perf_event *event =
479                         cpuc->events[X86_PMC_IDX_FIXED_BTS];
480
481                 if (WARN_ON_ONCE(!event))
482                         return;
483
484                 intel_pmu_enable_bts(event->hw.config);
485         }
486 }
487
488 /*
489  * Workaround for:
490  *   Intel Errata AAK100 (model 26)
491  *   Intel Errata AAP53  (model 30)
492  *   Intel Errata BD53   (model 44)
493  *
494  * These chips need to be 'reset' when adding counters by programming
495  * the magic three (non counting) events 0x4300D2, 0x4300B1 and 0x4300B5
496  * either in sequence on the same PMC or on different PMCs.
497  */
498 static void intel_pmu_nhm_enable_all(int added)
499 {
500         if (added) {
501                 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
502                 int i;
503
504                 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + 0, 0x4300D2);
505                 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + 1, 0x4300B1);
506                 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + 2, 0x4300B5);
507
508                 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x3);
509                 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0);
510
511                 for (i = 0; i < 3; i++) {
512                         struct perf_event *event = cpuc->events[i];
513
514                         if (!event)
515                                 continue;
516
517                         __x86_pmu_enable_event(&event->hw,
518                                                ARCH_PERFMON_EVENTSEL_ENABLE);
519                 }
520         }
521         intel_pmu_enable_all(added);
522 }
523
524 static inline u64 intel_pmu_get_status(void)
525 {
526         u64 status;
527
528         rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
529
530         return status;
531 }
532
533 static inline void intel_pmu_ack_status(u64 ack)
534 {
535         wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
536 }
537
538 static void intel_pmu_disable_fixed(struct hw_perf_event *hwc)
539 {
540         int idx = hwc->idx - X86_PMC_IDX_FIXED;
541         u64 ctrl_val, mask;
542
543         mask = 0xfULL << (idx * 4);
544
545         rdmsrl(hwc->config_base, ctrl_val);
546         ctrl_val &= ~mask;
547         wrmsrl(hwc->config_base, ctrl_val);
548 }
549
550 static void intel_pmu_disable_event(struct perf_event *event)
551 {
552         struct hw_perf_event *hwc = &event->hw;
553
554         if (unlikely(hwc->idx == X86_PMC_IDX_FIXED_BTS)) {
555                 intel_pmu_disable_bts();
556                 intel_pmu_drain_bts_buffer();
557                 return;
558         }
559
560         if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
561                 intel_pmu_disable_fixed(hwc);
562                 return;
563         }
564
565         x86_pmu_disable_event(event);
566
567         if (unlikely(event->attr.precise_ip))
568                 intel_pmu_pebs_disable(event);
569 }
570
571 static void intel_pmu_enable_fixed(struct hw_perf_event *hwc)
572 {
573         int idx = hwc->idx - X86_PMC_IDX_FIXED;
574         u64 ctrl_val, bits, mask;
575
576         /*
577          * Enable IRQ generation (0x8),
578          * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
579          * if requested:
580          */
581         bits = 0x8ULL;
582         if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
583                 bits |= 0x2;
584         if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
585                 bits |= 0x1;
586
587         /*
588          * ANY bit is supported in v3 and up
589          */
590         if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY)
591                 bits |= 0x4;
592
593         bits <<= (idx * 4);
594         mask = 0xfULL << (idx * 4);
595
596         rdmsrl(hwc->config_base, ctrl_val);
597         ctrl_val &= ~mask;
598         ctrl_val |= bits;
599         wrmsrl(hwc->config_base, ctrl_val);
600 }
601
602 static void intel_pmu_enable_event(struct perf_event *event)
603 {
604         struct hw_perf_event *hwc = &event->hw;
605
606         if (unlikely(hwc->idx == X86_PMC_IDX_FIXED_BTS)) {
607                 if (!__get_cpu_var(cpu_hw_events).enabled)
608                         return;
609
610                 intel_pmu_enable_bts(hwc->config);
611                 return;
612         }
613
614         if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
615                 intel_pmu_enable_fixed(hwc);
616                 return;
617         }
618
619         if (unlikely(event->attr.precise_ip))
620                 intel_pmu_pebs_enable(event);
621
622         __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
623 }
624
625 /*
626  * Save and restart an expired event. Called by NMI contexts,
627  * so it has to be careful about preempting normal event ops:
628  */
629 static int intel_pmu_save_and_restart(struct perf_event *event)
630 {
631         x86_perf_event_update(event);
632         return x86_perf_event_set_period(event);
633 }
634
635 static void intel_pmu_reset(void)
636 {
637         struct debug_store *ds = __get_cpu_var(cpu_hw_events).ds;
638         unsigned long flags;
639         int idx;
640
641         if (!x86_pmu.num_counters)
642                 return;
643
644         local_irq_save(flags);
645
646         printk("clearing PMU state on CPU#%d\n", smp_processor_id());
647
648         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
649                 checking_wrmsrl(x86_pmu.eventsel + idx, 0ull);
650                 checking_wrmsrl(x86_pmu.perfctr  + idx, 0ull);
651         }
652         for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++)
653                 checking_wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
654
655         if (ds)
656                 ds->bts_index = ds->bts_buffer_base;
657
658         local_irq_restore(flags);
659 }
660
661 /*
662  * This handler is triggered by the local APIC, so the APIC IRQ handling
663  * rules apply:
664  */
665 static int intel_pmu_handle_irq(struct pt_regs *regs)
666 {
667         struct perf_sample_data data;
668         struct cpu_hw_events *cpuc;
669         int bit, loops;
670         u64 ack, status;
671
672         perf_sample_data_init(&data, 0);
673
674         cpuc = &__get_cpu_var(cpu_hw_events);
675
676         intel_pmu_disable_all();
677         intel_pmu_drain_bts_buffer();
678         status = intel_pmu_get_status();
679         if (!status) {
680                 intel_pmu_enable_all(0);
681                 return 0;
682         }
683
684         loops = 0;
685 again:
686         if (++loops > 100) {
687                 WARN_ONCE(1, "perfevents: irq loop stuck!\n");
688                 perf_event_print_debug();
689                 intel_pmu_reset();
690                 goto done;
691         }
692
693         inc_irq_stat(apic_perf_irqs);
694         ack = status;
695
696         intel_pmu_lbr_read();
697
698         /*
699          * PEBS overflow sets bit 62 in the global status register
700          */
701         if (__test_and_clear_bit(62, (unsigned long *)&status))
702                 x86_pmu.drain_pebs(regs);
703
704         for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
705                 struct perf_event *event = cpuc->events[bit];
706
707                 if (!test_bit(bit, cpuc->active_mask))
708                         continue;
709
710                 if (!intel_pmu_save_and_restart(event))
711                         continue;
712
713                 data.period = event->hw.last_period;
714
715                 if (perf_event_overflow(event, 1, &data, regs))
716                         x86_pmu_stop(event);
717         }
718
719         intel_pmu_ack_status(ack);
720
721         /*
722          * Repeat if there is more work to be done:
723          */
724         status = intel_pmu_get_status();
725         if (status)
726                 goto again;
727
728 done:
729         intel_pmu_enable_all(0);
730         return 1;
731 }
732
733 static struct event_constraint *
734 intel_bts_constraints(struct perf_event *event)
735 {
736         struct hw_perf_event *hwc = &event->hw;
737         unsigned int hw_event, bts_event;
738
739         hw_event = hwc->config & INTEL_ARCH_EVENT_MASK;
740         bts_event = x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS);
741
742         if (unlikely(hw_event == bts_event && hwc->sample_period == 1))
743                 return &bts_constraint;
744
745         return NULL;
746 }
747
748 static struct event_constraint *
749 intel_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
750 {
751         struct event_constraint *c;
752
753         c = intel_bts_constraints(event);
754         if (c)
755                 return c;
756
757         c = intel_pebs_constraints(event);
758         if (c)
759                 return c;
760
761         return x86_get_event_constraints(cpuc, event);
762 }
763
764 static int intel_pmu_hw_config(struct perf_event *event)
765 {
766         int ret = x86_pmu_hw_config(event);
767
768         if (ret)
769                 return ret;
770
771         if (event->attr.type != PERF_TYPE_RAW)
772                 return 0;
773
774         if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY))
775                 return 0;
776
777         if (x86_pmu.version < 3)
778                 return -EINVAL;
779
780         if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
781                 return -EACCES;
782
783         event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY;
784
785         return 0;
786 }
787
788 static __initconst const struct x86_pmu core_pmu = {
789         .name                   = "core",
790         .handle_irq             = x86_pmu_handle_irq,
791         .disable_all            = x86_pmu_disable_all,
792         .enable_all             = x86_pmu_enable_all,
793         .enable                 = x86_pmu_enable_event,
794         .disable                = x86_pmu_disable_event,
795         .hw_config              = x86_pmu_hw_config,
796         .schedule_events        = x86_schedule_events,
797         .eventsel               = MSR_ARCH_PERFMON_EVENTSEL0,
798         .perfctr                = MSR_ARCH_PERFMON_PERFCTR0,
799         .event_map              = intel_pmu_event_map,
800         .max_events             = ARRAY_SIZE(intel_perfmon_event_map),
801         .apic                   = 1,
802         /*
803          * Intel PMCs cannot be accessed sanely above 32 bit width,
804          * so we install an artificial 1<<31 period regardless of
805          * the generic event period:
806          */
807         .max_period             = (1ULL << 31) - 1,
808         .get_event_constraints  = intel_get_event_constraints,
809         .event_constraints      = intel_core_event_constraints,
810 };
811
812 static void intel_pmu_cpu_starting(int cpu)
813 {
814         init_debug_store_on_cpu(cpu);
815         /*
816          * Deal with CPUs that don't clear their LBRs on power-up.
817          */
818         intel_pmu_lbr_reset();
819 }
820
821 static void intel_pmu_cpu_dying(int cpu)
822 {
823         fini_debug_store_on_cpu(cpu);
824 }
825
826 static __initconst const struct x86_pmu intel_pmu = {
827         .name                   = "Intel",
828         .handle_irq             = intel_pmu_handle_irq,
829         .disable_all            = intel_pmu_disable_all,
830         .enable_all             = intel_pmu_enable_all,
831         .enable                 = intel_pmu_enable_event,
832         .disable                = intel_pmu_disable_event,
833         .hw_config              = intel_pmu_hw_config,
834         .schedule_events        = x86_schedule_events,
835         .eventsel               = MSR_ARCH_PERFMON_EVENTSEL0,
836         .perfctr                = MSR_ARCH_PERFMON_PERFCTR0,
837         .event_map              = intel_pmu_event_map,
838         .max_events             = ARRAY_SIZE(intel_perfmon_event_map),
839         .apic                   = 1,
840         /*
841          * Intel PMCs cannot be accessed sanely above 32 bit width,
842          * so we install an artificial 1<<31 period regardless of
843          * the generic event period:
844          */
845         .max_period             = (1ULL << 31) - 1,
846         .get_event_constraints  = intel_get_event_constraints,
847
848         .cpu_starting           = intel_pmu_cpu_starting,
849         .cpu_dying              = intel_pmu_cpu_dying,
850 };
851
852 static void intel_clovertown_quirks(void)
853 {
854         /*
855          * PEBS is unreliable due to:
856          *
857          *   AJ67  - PEBS may experience CPL leaks
858          *   AJ68  - PEBS PMI may be delayed by one event
859          *   AJ69  - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12]
860          *   AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS
861          *
862          * AJ67 could be worked around by restricting the OS/USR flags.
863          * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI.
864          *
865          * AJ106 could possibly be worked around by not allowing LBR
866          *       usage from PEBS, including the fixup.
867          * AJ68  could possibly be worked around by always programming
868          *       a pebs_event_reset[0] value and coping with the lost events.
869          *
870          * But taken together it might just make sense to not enable PEBS on
871          * these chips.
872          */
873         printk(KERN_WARNING "PEBS disabled due to CPU errata.\n");
874         x86_pmu.pebs = 0;
875         x86_pmu.pebs_constraints = NULL;
876 }
877
878 static __init int intel_pmu_init(void)
879 {
880         union cpuid10_edx edx;
881         union cpuid10_eax eax;
882         unsigned int unused;
883         unsigned int ebx;
884         int version;
885
886         if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
887                 switch (boot_cpu_data.x86) {
888                 case 0x6:
889                         return p6_pmu_init();
890                 case 0xf:
891                         return p4_pmu_init();
892                 }
893                 return -ENODEV;
894         }
895
896         /*
897          * Check whether the Architectural PerfMon supports
898          * Branch Misses Retired hw_event or not.
899          */
900         cpuid(10, &eax.full, &ebx, &unused, &edx.full);
901         if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED)
902                 return -ENODEV;
903
904         version = eax.split.version_id;
905         if (version < 2)
906                 x86_pmu = core_pmu;
907         else
908                 x86_pmu = intel_pmu;
909
910         x86_pmu.version                 = version;
911         x86_pmu.num_counters            = eax.split.num_counters;
912         x86_pmu.cntval_bits             = eax.split.bit_width;
913         x86_pmu.cntval_mask             = (1ULL << eax.split.bit_width) - 1;
914
915         /*
916          * Quirk: v2 perfmon does not report fixed-purpose events, so
917          * assume at least 3 events:
918          */
919         if (version > 1)
920                 x86_pmu.num_counters_fixed = max((int)edx.split.num_counters_fixed, 3);
921
922         /*
923          * v2 and above have a perf capabilities MSR
924          */
925         if (version > 1) {
926                 u64 capabilities;
927
928                 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities);
929                 x86_pmu.intel_cap.capabilities = capabilities;
930         }
931
932         intel_ds_init();
933
934         /*
935          * Install the hw-cache-events table:
936          */
937         switch (boot_cpu_data.x86_model) {
938         case 14: /* 65 nm core solo/duo, "Yonah" */
939                 pr_cont("Core events, ");
940                 break;
941
942         case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */
943                 x86_pmu.quirks = intel_clovertown_quirks;
944         case 22: /* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */
945         case 23: /* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */
946         case 29: /* six-core 45 nm xeon "Dunnington" */
947                 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
948                        sizeof(hw_cache_event_ids));
949
950                 intel_pmu_lbr_init_core();
951
952                 x86_pmu.event_constraints = intel_core2_event_constraints;
953                 pr_cont("Core2 events, ");
954                 break;
955
956         case 26: /* 45 nm nehalem, "Bloomfield" */
957         case 30: /* 45 nm nehalem, "Lynnfield" */
958         case 46: /* 45 nm nehalem-ex, "Beckton" */
959                 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
960                        sizeof(hw_cache_event_ids));
961
962                 intel_pmu_lbr_init_nhm();
963
964                 x86_pmu.event_constraints = intel_nehalem_event_constraints;
965                 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
966                 pr_cont("Nehalem events, ");
967                 break;
968
969         case 28: /* Atom */
970                 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
971                        sizeof(hw_cache_event_ids));
972
973                 intel_pmu_lbr_init_atom();
974
975                 x86_pmu.event_constraints = intel_gen_event_constraints;
976                 pr_cont("Atom events, ");
977                 break;
978
979         case 37: /* 32 nm nehalem, "Clarkdale" */
980         case 44: /* 32 nm nehalem, "Gulftown" */
981                 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids,
982                        sizeof(hw_cache_event_ids));
983
984                 intel_pmu_lbr_init_nhm();
985
986                 x86_pmu.event_constraints = intel_westmere_event_constraints;
987                 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
988                 pr_cont("Westmere events, ");
989                 break;
990
991         default:
992                 /*
993                  * default constraints for v2 and up
994                  */
995                 x86_pmu.event_constraints = intel_gen_event_constraints;
996                 pr_cont("generic architected perfmon, ");
997         }
998         return 0;
999 }
1000
1001 #else /* CONFIG_CPU_SUP_INTEL */
1002
1003 static int intel_pmu_init(void)
1004 {
1005         return 0;
1006 }
1007
1008 #endif /* CONFIG_CPU_SUP_INTEL */