perf_event: x86: Reduce some overly long lines with some MACROs
[linux-2.6.git] / arch / x86 / kernel / cpu / perf_event.c
1 /*
2  * Performance events x86 architecture code
3  *
4  *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5  *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6  *  Copyright (C) 2009 Jaswinder Singh Rajput
7  *  Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
8  *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
9  *  Copyright (C) 2009 Intel Corporation, <markus.t.metzger@intel.com>
10  *  Copyright (C) 2009 Google, Inc., Stephane Eranian
11  *
12  *  For licencing details see kernel-base/COPYING
13  */
14
15 #include <linux/perf_event.h>
16 #include <linux/capability.h>
17 #include <linux/notifier.h>
18 #include <linux/hardirq.h>
19 #include <linux/kprobes.h>
20 #include <linux/module.h>
21 #include <linux/kdebug.h>
22 #include <linux/sched.h>
23 #include <linux/uaccess.h>
24 #include <linux/highmem.h>
25 #include <linux/cpu.h>
26
27 #include <asm/apic.h>
28 #include <asm/stacktrace.h>
29 #include <asm/nmi.h>
30
31 static u64 perf_event_mask __read_mostly;
32
33 /* The maximal number of PEBS events: */
34 #define MAX_PEBS_EVENTS 4
35
36 /* The size of a BTS record in bytes: */
37 #define BTS_RECORD_SIZE         24
38
39 /* The size of a per-cpu BTS buffer in bytes: */
40 #define BTS_BUFFER_SIZE         (BTS_RECORD_SIZE * 2048)
41
42 /* The BTS overflow threshold in bytes from the end of the buffer: */
43 #define BTS_OVFL_TH             (BTS_RECORD_SIZE * 128)
44
45
46 /*
47  * Bits in the debugctlmsr controlling branch tracing.
48  */
49 #define X86_DEBUGCTL_TR                 (1 << 6)
50 #define X86_DEBUGCTL_BTS                (1 << 7)
51 #define X86_DEBUGCTL_BTINT              (1 << 8)
52 #define X86_DEBUGCTL_BTS_OFF_OS         (1 << 9)
53 #define X86_DEBUGCTL_BTS_OFF_USR        (1 << 10)
54
55 /*
56  * A debug store configuration.
57  *
58  * We only support architectures that use 64bit fields.
59  */
60 struct debug_store {
61         u64     bts_buffer_base;
62         u64     bts_index;
63         u64     bts_absolute_maximum;
64         u64     bts_interrupt_threshold;
65         u64     pebs_buffer_base;
66         u64     pebs_index;
67         u64     pebs_absolute_maximum;
68         u64     pebs_interrupt_threshold;
69         u64     pebs_event_reset[MAX_PEBS_EVENTS];
70 };
71
72 struct event_constraint {
73         union {
74                 unsigned long   idxmsk[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
75                 u64             idxmsk64[1];
76         };
77         int     code;
78         int     cmask;
79 };
80
81 struct cpu_hw_events {
82         struct perf_event       *events[X86_PMC_IDX_MAX]; /* in counter order */
83         unsigned long           active_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
84         unsigned long           interrupts;
85         int                     enabled;
86         struct debug_store      *ds;
87
88         int                     n_events;
89         int                     n_added;
90         int                     assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
91         struct perf_event       *event_list[X86_PMC_IDX_MAX]; /* in enabled order */
92 };
93
94 #define EVENT_CONSTRAINT(c, n, m) {     \
95         { .idxmsk64[0] = (n) },         \
96         .code = (c),                    \
97         .cmask = (m),                   \
98 }
99
100 #define INTEL_EVENT_CONSTRAINT(c, n)    \
101         EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK)
102
103 #define FIXED_EVENT_CONSTRAINT(c, n)    \
104         EVENT_CONSTRAINT(c, n, INTEL_ARCH_FIXED_MASK)
105
106 #define EVENT_CONSTRAINT_END \
107         EVENT_CONSTRAINT(0, 0, 0)
108
109 #define for_each_event_constraint(e, c) \
110         for ((e) = (c); (e)->cmask; (e)++)
111
112 /*
113  * struct x86_pmu - generic x86 pmu
114  */
115 struct x86_pmu {
116         const char      *name;
117         int             version;
118         int             (*handle_irq)(struct pt_regs *);
119         void            (*disable_all)(void);
120         void            (*enable_all)(void);
121         void            (*enable)(struct hw_perf_event *, int);
122         void            (*disable)(struct hw_perf_event *, int);
123         unsigned        eventsel;
124         unsigned        perfctr;
125         u64             (*event_map)(int);
126         u64             (*raw_event)(u64);
127         int             max_events;
128         int             num_events;
129         int             num_events_fixed;
130         int             event_bits;
131         u64             event_mask;
132         int             apic;
133         u64             max_period;
134         u64             intel_ctrl;
135         void            (*enable_bts)(u64 config);
136         void            (*disable_bts)(void);
137         void            (*get_event_constraints)(struct cpu_hw_events *cpuc,
138                                                  struct perf_event *event,
139                                                  unsigned long *idxmsk);
140         void            (*put_event_constraints)(struct cpu_hw_events *cpuc,
141                                                  struct perf_event *event);
142         const struct event_constraint *event_constraints;
143 };
144
145 static struct x86_pmu x86_pmu __read_mostly;
146
147 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
148         .enabled = 1,
149 };
150
151 static int x86_perf_event_set_period(struct perf_event *event,
152                              struct hw_perf_event *hwc, int idx);
153
154 /*
155  * Not sure about some of these
156  */
157 static const u64 p6_perfmon_event_map[] =
158 {
159   [PERF_COUNT_HW_CPU_CYCLES]            = 0x0079,
160   [PERF_COUNT_HW_INSTRUCTIONS]          = 0x00c0,
161   [PERF_COUNT_HW_CACHE_REFERENCES]      = 0x0f2e,
162   [PERF_COUNT_HW_CACHE_MISSES]          = 0x012e,
163   [PERF_COUNT_HW_BRANCH_INSTRUCTIONS]   = 0x00c4,
164   [PERF_COUNT_HW_BRANCH_MISSES]         = 0x00c5,
165   [PERF_COUNT_HW_BUS_CYCLES]            = 0x0062,
166 };
167
168 static u64 p6_pmu_event_map(int hw_event)
169 {
170         return p6_perfmon_event_map[hw_event];
171 }
172
173 /*
174  * Event setting that is specified not to count anything.
175  * We use this to effectively disable a counter.
176  *
177  * L2_RQSTS with 0 MESI unit mask.
178  */
179 #define P6_NOP_EVENT                    0x0000002EULL
180
181 static u64 p6_pmu_raw_event(u64 hw_event)
182 {
183 #define P6_EVNTSEL_EVENT_MASK           0x000000FFULL
184 #define P6_EVNTSEL_UNIT_MASK            0x0000FF00ULL
185 #define P6_EVNTSEL_EDGE_MASK            0x00040000ULL
186 #define P6_EVNTSEL_INV_MASK             0x00800000ULL
187 #define P6_EVNTSEL_REG_MASK             0xFF000000ULL
188
189 #define P6_EVNTSEL_MASK                 \
190         (P6_EVNTSEL_EVENT_MASK |        \
191          P6_EVNTSEL_UNIT_MASK  |        \
192          P6_EVNTSEL_EDGE_MASK  |        \
193          P6_EVNTSEL_INV_MASK   |        \
194          P6_EVNTSEL_REG_MASK)
195
196         return hw_event & P6_EVNTSEL_MASK;
197 }
198
199 static struct event_constraint intel_p6_event_constraints[] =
200 {
201         INTEL_EVENT_CONSTRAINT(0xc1, 0x1),      /* FLOPS */
202         INTEL_EVENT_CONSTRAINT(0x10, 0x1),      /* FP_COMP_OPS_EXE */
203         INTEL_EVENT_CONSTRAINT(0x11, 0x1),      /* FP_ASSIST */
204         INTEL_EVENT_CONSTRAINT(0x12, 0x2),      /* MUL */
205         INTEL_EVENT_CONSTRAINT(0x13, 0x2),      /* DIV */
206         INTEL_EVENT_CONSTRAINT(0x14, 0x1),      /* CYCLES_DIV_BUSY */
207         EVENT_CONSTRAINT_END
208 };
209
210 /*
211  * Intel PerfMon v3. Used on Core2 and later.
212  */
213 static const u64 intel_perfmon_event_map[] =
214 {
215   [PERF_COUNT_HW_CPU_CYCLES]            = 0x003c,
216   [PERF_COUNT_HW_INSTRUCTIONS]          = 0x00c0,
217   [PERF_COUNT_HW_CACHE_REFERENCES]      = 0x4f2e,
218   [PERF_COUNT_HW_CACHE_MISSES]          = 0x412e,
219   [PERF_COUNT_HW_BRANCH_INSTRUCTIONS]   = 0x00c4,
220   [PERF_COUNT_HW_BRANCH_MISSES]         = 0x00c5,
221   [PERF_COUNT_HW_BUS_CYCLES]            = 0x013c,
222 };
223
224 static struct event_constraint intel_core_event_constraints[] =
225 {
226         FIXED_EVENT_CONSTRAINT(0xc0, (0x3|(1ULL<<32))), /* INSTRUCTIONS_RETIRED */
227         FIXED_EVENT_CONSTRAINT(0x3c, (0x3|(1ULL<<33))), /* UNHALTED_CORE_CYCLES */
228         INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
229         INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
230         INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
231         INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
232         INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
233         INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */
234         INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
235         INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */
236         INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */
237         EVENT_CONSTRAINT_END
238 };
239
240 static struct event_constraint intel_nehalem_event_constraints[] =
241 {
242         FIXED_EVENT_CONSTRAINT(0xc0, (0x3|(1ULL<<32))), /* INSTRUCTIONS_RETIRED */
243         FIXED_EVENT_CONSTRAINT(0x3c, (0x3|(1ULL<<33))), /* UNHALTED_CORE_CYCLES */
244         INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */
245         INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */
246         INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */
247         INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */
248         INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */
249         INTEL_EVENT_CONSTRAINT(0x4c, 0x3), /* LOAD_HIT_PRE */
250         INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
251         INTEL_EVENT_CONSTRAINT(0x52, 0x3), /* L1D_CACHE_PREFETCH_LOCK_FB_HIT */
252         INTEL_EVENT_CONSTRAINT(0x53, 0x3), /* L1D_CACHE_LOCK_FB_HIT */
253         INTEL_EVENT_CONSTRAINT(0xc5, 0x3), /* CACHE_LOCK_CYCLES */
254         EVENT_CONSTRAINT_END
255 };
256
257 static struct event_constraint intel_gen_event_constraints[] =
258 {
259         FIXED_EVENT_CONSTRAINT(0xc0, (0x3|(1ULL<<32))), /* INSTRUCTIONS_RETIRED */
260         FIXED_EVENT_CONSTRAINT(0x3c, (0x3|(1ULL<<33))), /* UNHALTED_CORE_CYCLES */
261         EVENT_CONSTRAINT_END
262 };
263
264 static u64 intel_pmu_event_map(int hw_event)
265 {
266         return intel_perfmon_event_map[hw_event];
267 }
268
269 /*
270  * Generalized hw caching related hw_event table, filled
271  * in on a per model basis. A value of 0 means
272  * 'not supported', -1 means 'hw_event makes no sense on
273  * this CPU', any other value means the raw hw_event
274  * ID.
275  */
276
277 #define C(x) PERF_COUNT_HW_CACHE_##x
278
279 static u64 __read_mostly hw_cache_event_ids
280                                 [PERF_COUNT_HW_CACHE_MAX]
281                                 [PERF_COUNT_HW_CACHE_OP_MAX]
282                                 [PERF_COUNT_HW_CACHE_RESULT_MAX];
283
284 static __initconst u64 nehalem_hw_cache_event_ids
285                                 [PERF_COUNT_HW_CACHE_MAX]
286                                 [PERF_COUNT_HW_CACHE_OP_MAX]
287                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
288 {
289  [ C(L1D) ] = {
290         [ C(OP_READ) ] = {
291                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI            */
292                 [ C(RESULT_MISS)   ] = 0x0140, /* L1D_CACHE_LD.I_STATE         */
293         },
294         [ C(OP_WRITE) ] = {
295                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI            */
296                 [ C(RESULT_MISS)   ] = 0x0141, /* L1D_CACHE_ST.I_STATE         */
297         },
298         [ C(OP_PREFETCH) ] = {
299                 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS        */
300                 [ C(RESULT_MISS)   ] = 0x024e, /* L1D_PREFETCH.MISS            */
301         },
302  },
303  [ C(L1I ) ] = {
304         [ C(OP_READ) ] = {
305                 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS                    */
306                 [ C(RESULT_MISS)   ] = 0x0280, /* L1I.MISSES                   */
307         },
308         [ C(OP_WRITE) ] = {
309                 [ C(RESULT_ACCESS) ] = -1,
310                 [ C(RESULT_MISS)   ] = -1,
311         },
312         [ C(OP_PREFETCH) ] = {
313                 [ C(RESULT_ACCESS) ] = 0x0,
314                 [ C(RESULT_MISS)   ] = 0x0,
315         },
316  },
317  [ C(LL  ) ] = {
318         [ C(OP_READ) ] = {
319                 [ C(RESULT_ACCESS) ] = 0x0324, /* L2_RQSTS.LOADS               */
320                 [ C(RESULT_MISS)   ] = 0x0224, /* L2_RQSTS.LD_MISS             */
321         },
322         [ C(OP_WRITE) ] = {
323                 [ C(RESULT_ACCESS) ] = 0x0c24, /* L2_RQSTS.RFOS                */
324                 [ C(RESULT_MISS)   ] = 0x0824, /* L2_RQSTS.RFO_MISS            */
325         },
326         [ C(OP_PREFETCH) ] = {
327                 [ C(RESULT_ACCESS) ] = 0x4f2e, /* LLC Reference                */
328                 [ C(RESULT_MISS)   ] = 0x412e, /* LLC Misses                   */
329         },
330  },
331  [ C(DTLB) ] = {
332         [ C(OP_READ) ] = {
333                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI   (alias)  */
334                 [ C(RESULT_MISS)   ] = 0x0108, /* DTLB_LOAD_MISSES.ANY         */
335         },
336         [ C(OP_WRITE) ] = {
337                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI   (alias)  */
338                 [ C(RESULT_MISS)   ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS  */
339         },
340         [ C(OP_PREFETCH) ] = {
341                 [ C(RESULT_ACCESS) ] = 0x0,
342                 [ C(RESULT_MISS)   ] = 0x0,
343         },
344  },
345  [ C(ITLB) ] = {
346         [ C(OP_READ) ] = {
347                 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P           */
348                 [ C(RESULT_MISS)   ] = 0x20c8, /* ITLB_MISS_RETIRED            */
349         },
350         [ C(OP_WRITE) ] = {
351                 [ C(RESULT_ACCESS) ] = -1,
352                 [ C(RESULT_MISS)   ] = -1,
353         },
354         [ C(OP_PREFETCH) ] = {
355                 [ C(RESULT_ACCESS) ] = -1,
356                 [ C(RESULT_MISS)   ] = -1,
357         },
358  },
359  [ C(BPU ) ] = {
360         [ C(OP_READ) ] = {
361                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
362                 [ C(RESULT_MISS)   ] = 0x03e8, /* BPU_CLEARS.ANY               */
363         },
364         [ C(OP_WRITE) ] = {
365                 [ C(RESULT_ACCESS) ] = -1,
366                 [ C(RESULT_MISS)   ] = -1,
367         },
368         [ C(OP_PREFETCH) ] = {
369                 [ C(RESULT_ACCESS) ] = -1,
370                 [ C(RESULT_MISS)   ] = -1,
371         },
372  },
373 };
374
375 static __initconst u64 core2_hw_cache_event_ids
376                                 [PERF_COUNT_HW_CACHE_MAX]
377                                 [PERF_COUNT_HW_CACHE_OP_MAX]
378                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
379 {
380  [ C(L1D) ] = {
381         [ C(OP_READ) ] = {
382                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI          */
383                 [ C(RESULT_MISS)   ] = 0x0140, /* L1D_CACHE_LD.I_STATE       */
384         },
385         [ C(OP_WRITE) ] = {
386                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI          */
387                 [ C(RESULT_MISS)   ] = 0x0141, /* L1D_CACHE_ST.I_STATE       */
388         },
389         [ C(OP_PREFETCH) ] = {
390                 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS      */
391                 [ C(RESULT_MISS)   ] = 0,
392         },
393  },
394  [ C(L1I ) ] = {
395         [ C(OP_READ) ] = {
396                 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS                  */
397                 [ C(RESULT_MISS)   ] = 0x0081, /* L1I.MISSES                 */
398         },
399         [ C(OP_WRITE) ] = {
400                 [ C(RESULT_ACCESS) ] = -1,
401                 [ C(RESULT_MISS)   ] = -1,
402         },
403         [ C(OP_PREFETCH) ] = {
404                 [ C(RESULT_ACCESS) ] = 0,
405                 [ C(RESULT_MISS)   ] = 0,
406         },
407  },
408  [ C(LL  ) ] = {
409         [ C(OP_READ) ] = {
410                 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI                 */
411                 [ C(RESULT_MISS)   ] = 0x4129, /* L2_LD.ISTATE               */
412         },
413         [ C(OP_WRITE) ] = {
414                 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI                 */
415                 [ C(RESULT_MISS)   ] = 0x412A, /* L2_ST.ISTATE               */
416         },
417         [ C(OP_PREFETCH) ] = {
418                 [ C(RESULT_ACCESS) ] = 0,
419                 [ C(RESULT_MISS)   ] = 0,
420         },
421  },
422  [ C(DTLB) ] = {
423         [ C(OP_READ) ] = {
424                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI  (alias) */
425                 [ C(RESULT_MISS)   ] = 0x0208, /* DTLB_MISSES.MISS_LD        */
426         },
427         [ C(OP_WRITE) ] = {
428                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI  (alias) */
429                 [ C(RESULT_MISS)   ] = 0x0808, /* DTLB_MISSES.MISS_ST        */
430         },
431         [ C(OP_PREFETCH) ] = {
432                 [ C(RESULT_ACCESS) ] = 0,
433                 [ C(RESULT_MISS)   ] = 0,
434         },
435  },
436  [ C(ITLB) ] = {
437         [ C(OP_READ) ] = {
438                 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P         */
439                 [ C(RESULT_MISS)   ] = 0x1282, /* ITLBMISSES                 */
440         },
441         [ C(OP_WRITE) ] = {
442                 [ C(RESULT_ACCESS) ] = -1,
443                 [ C(RESULT_MISS)   ] = -1,
444         },
445         [ C(OP_PREFETCH) ] = {
446                 [ C(RESULT_ACCESS) ] = -1,
447                 [ C(RESULT_MISS)   ] = -1,
448         },
449  },
450  [ C(BPU ) ] = {
451         [ C(OP_READ) ] = {
452                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY        */
453                 [ C(RESULT_MISS)   ] = 0x00c5, /* BP_INST_RETIRED.MISPRED    */
454         },
455         [ C(OP_WRITE) ] = {
456                 [ C(RESULT_ACCESS) ] = -1,
457                 [ C(RESULT_MISS)   ] = -1,
458         },
459         [ C(OP_PREFETCH) ] = {
460                 [ C(RESULT_ACCESS) ] = -1,
461                 [ C(RESULT_MISS)   ] = -1,
462         },
463  },
464 };
465
466 static __initconst u64 atom_hw_cache_event_ids
467                                 [PERF_COUNT_HW_CACHE_MAX]
468                                 [PERF_COUNT_HW_CACHE_OP_MAX]
469                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
470 {
471  [ C(L1D) ] = {
472         [ C(OP_READ) ] = {
473                 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD               */
474                 [ C(RESULT_MISS)   ] = 0,
475         },
476         [ C(OP_WRITE) ] = {
477                 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST               */
478                 [ C(RESULT_MISS)   ] = 0,
479         },
480         [ C(OP_PREFETCH) ] = {
481                 [ C(RESULT_ACCESS) ] = 0x0,
482                 [ C(RESULT_MISS)   ] = 0,
483         },
484  },
485  [ C(L1I ) ] = {
486         [ C(OP_READ) ] = {
487                 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS                  */
488                 [ C(RESULT_MISS)   ] = 0x0280, /* L1I.MISSES                 */
489         },
490         [ C(OP_WRITE) ] = {
491                 [ C(RESULT_ACCESS) ] = -1,
492                 [ C(RESULT_MISS)   ] = -1,
493         },
494         [ C(OP_PREFETCH) ] = {
495                 [ C(RESULT_ACCESS) ] = 0,
496                 [ C(RESULT_MISS)   ] = 0,
497         },
498  },
499  [ C(LL  ) ] = {
500         [ C(OP_READ) ] = {
501                 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI                 */
502                 [ C(RESULT_MISS)   ] = 0x4129, /* L2_LD.ISTATE               */
503         },
504         [ C(OP_WRITE) ] = {
505                 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI                 */
506                 [ C(RESULT_MISS)   ] = 0x412A, /* L2_ST.ISTATE               */
507         },
508         [ C(OP_PREFETCH) ] = {
509                 [ C(RESULT_ACCESS) ] = 0,
510                 [ C(RESULT_MISS)   ] = 0,
511         },
512  },
513  [ C(DTLB) ] = {
514         [ C(OP_READ) ] = {
515                 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI  (alias) */
516                 [ C(RESULT_MISS)   ] = 0x0508, /* DTLB_MISSES.MISS_LD        */
517         },
518         [ C(OP_WRITE) ] = {
519                 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI  (alias) */
520                 [ C(RESULT_MISS)   ] = 0x0608, /* DTLB_MISSES.MISS_ST        */
521         },
522         [ C(OP_PREFETCH) ] = {
523                 [ C(RESULT_ACCESS) ] = 0,
524                 [ C(RESULT_MISS)   ] = 0,
525         },
526  },
527  [ C(ITLB) ] = {
528         [ C(OP_READ) ] = {
529                 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P         */
530                 [ C(RESULT_MISS)   ] = 0x0282, /* ITLB.MISSES                */
531         },
532         [ C(OP_WRITE) ] = {
533                 [ C(RESULT_ACCESS) ] = -1,
534                 [ C(RESULT_MISS)   ] = -1,
535         },
536         [ C(OP_PREFETCH) ] = {
537                 [ C(RESULT_ACCESS) ] = -1,
538                 [ C(RESULT_MISS)   ] = -1,
539         },
540  },
541  [ C(BPU ) ] = {
542         [ C(OP_READ) ] = {
543                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY        */
544                 [ C(RESULT_MISS)   ] = 0x00c5, /* BP_INST_RETIRED.MISPRED    */
545         },
546         [ C(OP_WRITE) ] = {
547                 [ C(RESULT_ACCESS) ] = -1,
548                 [ C(RESULT_MISS)   ] = -1,
549         },
550         [ C(OP_PREFETCH) ] = {
551                 [ C(RESULT_ACCESS) ] = -1,
552                 [ C(RESULT_MISS)   ] = -1,
553         },
554  },
555 };
556
557 static u64 intel_pmu_raw_event(u64 hw_event)
558 {
559 #define CORE_EVNTSEL_EVENT_MASK         0x000000FFULL
560 #define CORE_EVNTSEL_UNIT_MASK          0x0000FF00ULL
561 #define CORE_EVNTSEL_EDGE_MASK          0x00040000ULL
562 #define CORE_EVNTSEL_INV_MASK           0x00800000ULL
563 #define CORE_EVNTSEL_REG_MASK           0xFF000000ULL
564
565 #define CORE_EVNTSEL_MASK               \
566         (INTEL_ARCH_EVTSEL_MASK |       \
567          INTEL_ARCH_UNIT_MASK   |       \
568          INTEL_ARCH_EDGE_MASK   |       \
569          INTEL_ARCH_INV_MASK    |       \
570          INTEL_ARCH_CNT_MASK)
571
572         return hw_event & CORE_EVNTSEL_MASK;
573 }
574
575 static __initconst u64 amd_hw_cache_event_ids
576                                 [PERF_COUNT_HW_CACHE_MAX]
577                                 [PERF_COUNT_HW_CACHE_OP_MAX]
578                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
579 {
580  [ C(L1D) ] = {
581         [ C(OP_READ) ] = {
582                 [ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses        */
583                 [ C(RESULT_MISS)   ] = 0x0041, /* Data Cache Misses          */
584         },
585         [ C(OP_WRITE) ] = {
586                 [ C(RESULT_ACCESS) ] = 0x0142, /* Data Cache Refills :system */
587                 [ C(RESULT_MISS)   ] = 0,
588         },
589         [ C(OP_PREFETCH) ] = {
590                 [ C(RESULT_ACCESS) ] = 0x0267, /* Data Prefetcher :attempts  */
591                 [ C(RESULT_MISS)   ] = 0x0167, /* Data Prefetcher :cancelled */
592         },
593  },
594  [ C(L1I ) ] = {
595         [ C(OP_READ) ] = {
596                 [ C(RESULT_ACCESS) ] = 0x0080, /* Instruction cache fetches  */
597                 [ C(RESULT_MISS)   ] = 0x0081, /* Instruction cache misses   */
598         },
599         [ C(OP_WRITE) ] = {
600                 [ C(RESULT_ACCESS) ] = -1,
601                 [ C(RESULT_MISS)   ] = -1,
602         },
603         [ C(OP_PREFETCH) ] = {
604                 [ C(RESULT_ACCESS) ] = 0x014B, /* Prefetch Instructions :Load */
605                 [ C(RESULT_MISS)   ] = 0,
606         },
607  },
608  [ C(LL  ) ] = {
609         [ C(OP_READ) ] = {
610                 [ C(RESULT_ACCESS) ] = 0x037D, /* Requests to L2 Cache :IC+DC */
611                 [ C(RESULT_MISS)   ] = 0x037E, /* L2 Cache Misses : IC+DC     */
612         },
613         [ C(OP_WRITE) ] = {
614                 [ C(RESULT_ACCESS) ] = 0x017F, /* L2 Fill/Writeback           */
615                 [ C(RESULT_MISS)   ] = 0,
616         },
617         [ C(OP_PREFETCH) ] = {
618                 [ C(RESULT_ACCESS) ] = 0,
619                 [ C(RESULT_MISS)   ] = 0,
620         },
621  },
622  [ C(DTLB) ] = {
623         [ C(OP_READ) ] = {
624                 [ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses        */
625                 [ C(RESULT_MISS)   ] = 0x0046, /* L1 DTLB and L2 DLTB Miss   */
626         },
627         [ C(OP_WRITE) ] = {
628                 [ C(RESULT_ACCESS) ] = 0,
629                 [ C(RESULT_MISS)   ] = 0,
630         },
631         [ C(OP_PREFETCH) ] = {
632                 [ C(RESULT_ACCESS) ] = 0,
633                 [ C(RESULT_MISS)   ] = 0,
634         },
635  },
636  [ C(ITLB) ] = {
637         [ C(OP_READ) ] = {
638                 [ C(RESULT_ACCESS) ] = 0x0080, /* Instruction fecthes        */
639                 [ C(RESULT_MISS)   ] = 0x0085, /* Instr. fetch ITLB misses   */
640         },
641         [ C(OP_WRITE) ] = {
642                 [ C(RESULT_ACCESS) ] = -1,
643                 [ C(RESULT_MISS)   ] = -1,
644         },
645         [ C(OP_PREFETCH) ] = {
646                 [ C(RESULT_ACCESS) ] = -1,
647                 [ C(RESULT_MISS)   ] = -1,
648         },
649  },
650  [ C(BPU ) ] = {
651         [ C(OP_READ) ] = {
652                 [ C(RESULT_ACCESS) ] = 0x00c2, /* Retired Branch Instr.      */
653                 [ C(RESULT_MISS)   ] = 0x00c3, /* Retired Mispredicted BI    */
654         },
655         [ C(OP_WRITE) ] = {
656                 [ C(RESULT_ACCESS) ] = -1,
657                 [ C(RESULT_MISS)   ] = -1,
658         },
659         [ C(OP_PREFETCH) ] = {
660                 [ C(RESULT_ACCESS) ] = -1,
661                 [ C(RESULT_MISS)   ] = -1,
662         },
663  },
664 };
665
666 /*
667  * AMD Performance Monitor K7 and later.
668  */
669 static const u64 amd_perfmon_event_map[] =
670 {
671   [PERF_COUNT_HW_CPU_CYCLES]            = 0x0076,
672   [PERF_COUNT_HW_INSTRUCTIONS]          = 0x00c0,
673   [PERF_COUNT_HW_CACHE_REFERENCES]      = 0x0080,
674   [PERF_COUNT_HW_CACHE_MISSES]          = 0x0081,
675   [PERF_COUNT_HW_BRANCH_INSTRUCTIONS]   = 0x00c4,
676   [PERF_COUNT_HW_BRANCH_MISSES]         = 0x00c5,
677 };
678
679 static u64 amd_pmu_event_map(int hw_event)
680 {
681         return amd_perfmon_event_map[hw_event];
682 }
683
684 static u64 amd_pmu_raw_event(u64 hw_event)
685 {
686 #define K7_EVNTSEL_EVENT_MASK   0x7000000FFULL
687 #define K7_EVNTSEL_UNIT_MASK    0x00000FF00ULL
688 #define K7_EVNTSEL_EDGE_MASK    0x000040000ULL
689 #define K7_EVNTSEL_INV_MASK     0x000800000ULL
690 #define K7_EVNTSEL_REG_MASK     0x0FF000000ULL
691
692 #define K7_EVNTSEL_MASK                 \
693         (K7_EVNTSEL_EVENT_MASK |        \
694          K7_EVNTSEL_UNIT_MASK  |        \
695          K7_EVNTSEL_EDGE_MASK  |        \
696          K7_EVNTSEL_INV_MASK   |        \
697          K7_EVNTSEL_REG_MASK)
698
699         return hw_event & K7_EVNTSEL_MASK;
700 }
701
702 /*
703  * Propagate event elapsed time into the generic event.
704  * Can only be executed on the CPU where the event is active.
705  * Returns the delta events processed.
706  */
707 static u64
708 x86_perf_event_update(struct perf_event *event,
709                         struct hw_perf_event *hwc, int idx)
710 {
711         int shift = 64 - x86_pmu.event_bits;
712         u64 prev_raw_count, new_raw_count;
713         s64 delta;
714
715         if (idx == X86_PMC_IDX_FIXED_BTS)
716                 return 0;
717
718         /*
719          * Careful: an NMI might modify the previous event value.
720          *
721          * Our tactic to handle this is to first atomically read and
722          * exchange a new raw count - then add that new-prev delta
723          * count to the generic event atomically:
724          */
725 again:
726         prev_raw_count = atomic64_read(&hwc->prev_count);
727         rdmsrl(hwc->event_base + idx, new_raw_count);
728
729         if (atomic64_cmpxchg(&hwc->prev_count, prev_raw_count,
730                                         new_raw_count) != prev_raw_count)
731                 goto again;
732
733         /*
734          * Now we have the new raw value and have updated the prev
735          * timestamp already. We can now calculate the elapsed delta
736          * (event-)time and add that to the generic event.
737          *
738          * Careful, not all hw sign-extends above the physical width
739          * of the count.
740          */
741         delta = (new_raw_count << shift) - (prev_raw_count << shift);
742         delta >>= shift;
743
744         atomic64_add(delta, &event->count);
745         atomic64_sub(delta, &hwc->period_left);
746
747         return new_raw_count;
748 }
749
750 static atomic_t active_events;
751 static DEFINE_MUTEX(pmc_reserve_mutex);
752
753 static bool reserve_pmc_hardware(void)
754 {
755 #ifdef CONFIG_X86_LOCAL_APIC
756         int i;
757
758         if (nmi_watchdog == NMI_LOCAL_APIC)
759                 disable_lapic_nmi_watchdog();
760
761         for (i = 0; i < x86_pmu.num_events; i++) {
762                 if (!reserve_perfctr_nmi(x86_pmu.perfctr + i))
763                         goto perfctr_fail;
764         }
765
766         for (i = 0; i < x86_pmu.num_events; i++) {
767                 if (!reserve_evntsel_nmi(x86_pmu.eventsel + i))
768                         goto eventsel_fail;
769         }
770 #endif
771
772         return true;
773
774 #ifdef CONFIG_X86_LOCAL_APIC
775 eventsel_fail:
776         for (i--; i >= 0; i--)
777                 release_evntsel_nmi(x86_pmu.eventsel + i);
778
779         i = x86_pmu.num_events;
780
781 perfctr_fail:
782         for (i--; i >= 0; i--)
783                 release_perfctr_nmi(x86_pmu.perfctr + i);
784
785         if (nmi_watchdog == NMI_LOCAL_APIC)
786                 enable_lapic_nmi_watchdog();
787
788         return false;
789 #endif
790 }
791
792 static void release_pmc_hardware(void)
793 {
794 #ifdef CONFIG_X86_LOCAL_APIC
795         int i;
796
797         for (i = 0; i < x86_pmu.num_events; i++) {
798                 release_perfctr_nmi(x86_pmu.perfctr + i);
799                 release_evntsel_nmi(x86_pmu.eventsel + i);
800         }
801
802         if (nmi_watchdog == NMI_LOCAL_APIC)
803                 enable_lapic_nmi_watchdog();
804 #endif
805 }
806
807 static inline bool bts_available(void)
808 {
809         return x86_pmu.enable_bts != NULL;
810 }
811
812 static inline void init_debug_store_on_cpu(int cpu)
813 {
814         struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
815
816         if (!ds)
817                 return;
818
819         wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
820                      (u32)((u64)(unsigned long)ds),
821                      (u32)((u64)(unsigned long)ds >> 32));
822 }
823
824 static inline void fini_debug_store_on_cpu(int cpu)
825 {
826         if (!per_cpu(cpu_hw_events, cpu).ds)
827                 return;
828
829         wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
830 }
831
832 static void release_bts_hardware(void)
833 {
834         int cpu;
835
836         if (!bts_available())
837                 return;
838
839         get_online_cpus();
840
841         for_each_online_cpu(cpu)
842                 fini_debug_store_on_cpu(cpu);
843
844         for_each_possible_cpu(cpu) {
845                 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
846
847                 if (!ds)
848                         continue;
849
850                 per_cpu(cpu_hw_events, cpu).ds = NULL;
851
852                 kfree((void *)(unsigned long)ds->bts_buffer_base);
853                 kfree(ds);
854         }
855
856         put_online_cpus();
857 }
858
859 static int reserve_bts_hardware(void)
860 {
861         int cpu, err = 0;
862
863         if (!bts_available())
864                 return 0;
865
866         get_online_cpus();
867
868         for_each_possible_cpu(cpu) {
869                 struct debug_store *ds;
870                 void *buffer;
871
872                 err = -ENOMEM;
873                 buffer = kzalloc(BTS_BUFFER_SIZE, GFP_KERNEL);
874                 if (unlikely(!buffer))
875                         break;
876
877                 ds = kzalloc(sizeof(*ds), GFP_KERNEL);
878                 if (unlikely(!ds)) {
879                         kfree(buffer);
880                         break;
881                 }
882
883                 ds->bts_buffer_base = (u64)(unsigned long)buffer;
884                 ds->bts_index = ds->bts_buffer_base;
885                 ds->bts_absolute_maximum =
886                         ds->bts_buffer_base + BTS_BUFFER_SIZE;
887                 ds->bts_interrupt_threshold =
888                         ds->bts_absolute_maximum - BTS_OVFL_TH;
889
890                 per_cpu(cpu_hw_events, cpu).ds = ds;
891                 err = 0;
892         }
893
894         if (err)
895                 release_bts_hardware();
896         else {
897                 for_each_online_cpu(cpu)
898                         init_debug_store_on_cpu(cpu);
899         }
900
901         put_online_cpus();
902
903         return err;
904 }
905
906 static void hw_perf_event_destroy(struct perf_event *event)
907 {
908         if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) {
909                 release_pmc_hardware();
910                 release_bts_hardware();
911                 mutex_unlock(&pmc_reserve_mutex);
912         }
913 }
914
915 static inline int x86_pmu_initialized(void)
916 {
917         return x86_pmu.handle_irq != NULL;
918 }
919
920 static inline int
921 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event_attr *attr)
922 {
923         unsigned int cache_type, cache_op, cache_result;
924         u64 config, val;
925
926         config = attr->config;
927
928         cache_type = (config >>  0) & 0xff;
929         if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
930                 return -EINVAL;
931
932         cache_op = (config >>  8) & 0xff;
933         if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
934                 return -EINVAL;
935
936         cache_result = (config >> 16) & 0xff;
937         if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
938                 return -EINVAL;
939
940         val = hw_cache_event_ids[cache_type][cache_op][cache_result];
941
942         if (val == 0)
943                 return -ENOENT;
944
945         if (val == -1)
946                 return -EINVAL;
947
948         hwc->config |= val;
949
950         return 0;
951 }
952
953 static void intel_pmu_enable_bts(u64 config)
954 {
955         unsigned long debugctlmsr;
956
957         debugctlmsr = get_debugctlmsr();
958
959         debugctlmsr |= X86_DEBUGCTL_TR;
960         debugctlmsr |= X86_DEBUGCTL_BTS;
961         debugctlmsr |= X86_DEBUGCTL_BTINT;
962
963         if (!(config & ARCH_PERFMON_EVENTSEL_OS))
964                 debugctlmsr |= X86_DEBUGCTL_BTS_OFF_OS;
965
966         if (!(config & ARCH_PERFMON_EVENTSEL_USR))
967                 debugctlmsr |= X86_DEBUGCTL_BTS_OFF_USR;
968
969         update_debugctlmsr(debugctlmsr);
970 }
971
972 static void intel_pmu_disable_bts(void)
973 {
974         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
975         unsigned long debugctlmsr;
976
977         if (!cpuc->ds)
978                 return;
979
980         debugctlmsr = get_debugctlmsr();
981
982         debugctlmsr &=
983                 ~(X86_DEBUGCTL_TR | X86_DEBUGCTL_BTS | X86_DEBUGCTL_BTINT |
984                   X86_DEBUGCTL_BTS_OFF_OS | X86_DEBUGCTL_BTS_OFF_USR);
985
986         update_debugctlmsr(debugctlmsr);
987 }
988
989 /*
990  * Setup the hardware configuration for a given attr_type
991  */
992 static int __hw_perf_event_init(struct perf_event *event)
993 {
994         struct perf_event_attr *attr = &event->attr;
995         struct hw_perf_event *hwc = &event->hw;
996         u64 config;
997         int err;
998
999         if (!x86_pmu_initialized())
1000                 return -ENODEV;
1001
1002         err = 0;
1003         if (!atomic_inc_not_zero(&active_events)) {
1004                 mutex_lock(&pmc_reserve_mutex);
1005                 if (atomic_read(&active_events) == 0) {
1006                         if (!reserve_pmc_hardware())
1007                                 err = -EBUSY;
1008                         else
1009                                 err = reserve_bts_hardware();
1010                 }
1011                 if (!err)
1012                         atomic_inc(&active_events);
1013                 mutex_unlock(&pmc_reserve_mutex);
1014         }
1015         if (err)
1016                 return err;
1017
1018         event->destroy = hw_perf_event_destroy;
1019
1020         /*
1021          * Generate PMC IRQs:
1022          * (keep 'enabled' bit clear for now)
1023          */
1024         hwc->config = ARCH_PERFMON_EVENTSEL_INT;
1025
1026         hwc->idx = -1;
1027
1028         /*
1029          * Count user and OS events unless requested not to.
1030          */
1031         if (!attr->exclude_user)
1032                 hwc->config |= ARCH_PERFMON_EVENTSEL_USR;
1033         if (!attr->exclude_kernel)
1034                 hwc->config |= ARCH_PERFMON_EVENTSEL_OS;
1035
1036         if (!hwc->sample_period) {
1037                 hwc->sample_period = x86_pmu.max_period;
1038                 hwc->last_period = hwc->sample_period;
1039                 atomic64_set(&hwc->period_left, hwc->sample_period);
1040         } else {
1041                 /*
1042                  * If we have a PMU initialized but no APIC
1043                  * interrupts, we cannot sample hardware
1044                  * events (user-space has to fall back and
1045                  * sample via a hrtimer based software event):
1046                  */
1047                 if (!x86_pmu.apic)
1048                         return -EOPNOTSUPP;
1049         }
1050
1051         /*
1052          * Raw hw_event type provide the config in the hw_event structure
1053          */
1054         if (attr->type == PERF_TYPE_RAW) {
1055                 hwc->config |= x86_pmu.raw_event(attr->config);
1056                 return 0;
1057         }
1058
1059         if (attr->type == PERF_TYPE_HW_CACHE)
1060                 return set_ext_hw_attr(hwc, attr);
1061
1062         if (attr->config >= x86_pmu.max_events)
1063                 return -EINVAL;
1064
1065         /*
1066          * The generic map:
1067          */
1068         config = x86_pmu.event_map(attr->config);
1069
1070         if (config == 0)
1071                 return -ENOENT;
1072
1073         if (config == -1LL)
1074                 return -EINVAL;
1075
1076         /*
1077          * Branch tracing:
1078          */
1079         if ((attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS) &&
1080             (hwc->sample_period == 1)) {
1081                 /* BTS is not supported by this architecture. */
1082                 if (!bts_available())
1083                         return -EOPNOTSUPP;
1084
1085                 /* BTS is currently only allowed for user-mode. */
1086                 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
1087                         return -EOPNOTSUPP;
1088         }
1089
1090         hwc->config |= config;
1091
1092         return 0;
1093 }
1094
1095 static void p6_pmu_disable_all(void)
1096 {
1097         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1098         u64 val;
1099
1100         if (!cpuc->enabled)
1101                 return;
1102
1103         cpuc->enabled = 0;
1104         barrier();
1105
1106         /* p6 only has one enable register */
1107         rdmsrl(MSR_P6_EVNTSEL0, val);
1108         val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
1109         wrmsrl(MSR_P6_EVNTSEL0, val);
1110 }
1111
1112 static void intel_pmu_disable_all(void)
1113 {
1114         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1115
1116         if (!cpuc->enabled)
1117                 return;
1118
1119         cpuc->enabled = 0;
1120         barrier();
1121
1122         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
1123
1124         if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask))
1125                 intel_pmu_disable_bts();
1126 }
1127
1128 static void amd_pmu_disable_all(void)
1129 {
1130         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1131         int idx;
1132
1133         if (!cpuc->enabled)
1134                 return;
1135
1136         cpuc->enabled = 0;
1137         /*
1138          * ensure we write the disable before we start disabling the
1139          * events proper, so that amd_pmu_enable_event() does the
1140          * right thing.
1141          */
1142         barrier();
1143
1144         for (idx = 0; idx < x86_pmu.num_events; idx++) {
1145                 u64 val;
1146
1147                 if (!test_bit(idx, cpuc->active_mask))
1148                         continue;
1149                 rdmsrl(MSR_K7_EVNTSEL0 + idx, val);
1150                 if (!(val & ARCH_PERFMON_EVENTSEL0_ENABLE))
1151                         continue;
1152                 val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
1153                 wrmsrl(MSR_K7_EVNTSEL0 + idx, val);
1154         }
1155 }
1156
1157 void hw_perf_disable(void)
1158 {
1159         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1160
1161         if (!x86_pmu_initialized())
1162                 return;
1163
1164         if (cpuc->enabled)
1165                 cpuc->n_added = 0;
1166
1167         x86_pmu.disable_all();
1168 }
1169
1170 static void p6_pmu_enable_all(void)
1171 {
1172         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1173         unsigned long val;
1174
1175         if (cpuc->enabled)
1176                 return;
1177
1178         cpuc->enabled = 1;
1179         barrier();
1180
1181         /* p6 only has one enable register */
1182         rdmsrl(MSR_P6_EVNTSEL0, val);
1183         val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
1184         wrmsrl(MSR_P6_EVNTSEL0, val);
1185 }
1186
1187 static void intel_pmu_enable_all(void)
1188 {
1189         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1190
1191         if (cpuc->enabled)
1192                 return;
1193
1194         cpuc->enabled = 1;
1195         barrier();
1196
1197         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, x86_pmu.intel_ctrl);
1198
1199         if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
1200                 struct perf_event *event =
1201                         cpuc->events[X86_PMC_IDX_FIXED_BTS];
1202
1203                 if (WARN_ON_ONCE(!event))
1204                         return;
1205
1206                 intel_pmu_enable_bts(event->hw.config);
1207         }
1208 }
1209
1210 static void amd_pmu_enable_all(void)
1211 {
1212         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1213         int idx;
1214
1215         if (cpuc->enabled)
1216                 return;
1217
1218         cpuc->enabled = 1;
1219         barrier();
1220
1221         for (idx = 0; idx < x86_pmu.num_events; idx++) {
1222                 struct perf_event *event = cpuc->events[idx];
1223                 u64 val;
1224
1225                 if (!test_bit(idx, cpuc->active_mask))
1226                         continue;
1227
1228                 val = event->hw.config;
1229                 val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
1230                 wrmsrl(MSR_K7_EVNTSEL0 + idx, val);
1231         }
1232 }
1233
1234 static const struct pmu pmu;
1235
1236 static inline int is_x86_event(struct perf_event *event)
1237 {
1238         return event->pmu == &pmu;
1239 }
1240
1241 static int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
1242 {
1243         int i, j , w, num;
1244         int weight, wmax;
1245         unsigned long *c;
1246         unsigned long constraints[X86_PMC_IDX_MAX][BITS_TO_LONGS(X86_PMC_IDX_MAX)];
1247         unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
1248         struct hw_perf_event *hwc;
1249
1250         bitmap_zero(used_mask, X86_PMC_IDX_MAX);
1251
1252         for (i = 0; i < n; i++) {
1253                 x86_pmu.get_event_constraints(cpuc,
1254                                               cpuc->event_list[i],
1255                                               constraints[i]);
1256         }
1257
1258         /*
1259          * fastpath, try to reuse previous register
1260          */
1261         for (i = 0, num = n; i < n; i++, num--) {
1262                 hwc = &cpuc->event_list[i]->hw;
1263                 c = constraints[i];
1264
1265                 /* never assigned */
1266                 if (hwc->idx == -1)
1267                         break;
1268
1269                 /* constraint still honored */
1270                 if (!test_bit(hwc->idx, c))
1271                         break;
1272
1273                 /* not already used */
1274                 if (test_bit(hwc->idx, used_mask))
1275                         break;
1276
1277 #if 0
1278                 pr_debug("CPU%d fast config=0x%llx idx=%d assign=%c\n",
1279                          smp_processor_id(),
1280                          hwc->config,
1281                          hwc->idx,
1282                          assign ? 'y' : 'n');
1283 #endif
1284
1285                 set_bit(hwc->idx, used_mask);
1286                 if (assign)
1287                         assign[i] = hwc->idx;
1288         }
1289         if (!num)
1290                 goto done;
1291
1292         /*
1293          * begin slow path
1294          */
1295
1296         bitmap_zero(used_mask, X86_PMC_IDX_MAX);
1297
1298         /*
1299          * weight = number of possible counters
1300          *
1301          * 1    = most constrained, only works on one counter
1302          * wmax = least constrained, works on any counter
1303          *
1304          * assign events to counters starting with most
1305          * constrained events.
1306          */
1307         wmax = x86_pmu.num_events;
1308
1309         /*
1310          * when fixed event counters are present,
1311          * wmax is incremented by 1 to account
1312          * for one more choice
1313          */
1314         if (x86_pmu.num_events_fixed)
1315                 wmax++;
1316
1317         for (w = 1, num = n; num && w <= wmax; w++) {
1318                 /* for each event */
1319                 for (i = 0; num && i < n; i++) {
1320                         c = constraints[i];
1321                         hwc = &cpuc->event_list[i]->hw;
1322
1323                         weight = bitmap_weight(c, X86_PMC_IDX_MAX);
1324                         if (weight != w)
1325                                 continue;
1326
1327                         for_each_bit(j, c, X86_PMC_IDX_MAX) {
1328                                 if (!test_bit(j, used_mask))
1329                                         break;
1330                         }
1331
1332                         if (j == X86_PMC_IDX_MAX)
1333                                 break;
1334
1335 #if 0
1336                         pr_debug("CPU%d slow config=0x%llx idx=%d assign=%c\n",
1337                                 smp_processor_id(),
1338                                 hwc->config,
1339                                 j,
1340                                 assign ? 'y' : 'n');
1341 #endif
1342
1343                         set_bit(j, used_mask);
1344
1345                         if (assign)
1346                                 assign[i] = j;
1347                         num--;
1348                 }
1349         }
1350 done:
1351         /*
1352          * scheduling failed or is just a simulation,
1353          * free resources if necessary
1354          */
1355         if (!assign || num) {
1356                 for (i = 0; i < n; i++) {
1357                         if (x86_pmu.put_event_constraints)
1358                                 x86_pmu.put_event_constraints(cpuc, cpuc->event_list[i]);
1359                 }
1360         }
1361         return num ? -ENOSPC : 0;
1362 }
1363
1364 /*
1365  * dogrp: true if must collect siblings events (group)
1366  * returns total number of events and error code
1367  */
1368 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
1369 {
1370         struct perf_event *event;
1371         int n, max_count;
1372
1373         max_count = x86_pmu.num_events + x86_pmu.num_events_fixed;
1374
1375         /* current number of events already accepted */
1376         n = cpuc->n_events;
1377
1378         if (is_x86_event(leader)) {
1379                 if (n >= max_count)
1380                         return -ENOSPC;
1381                 cpuc->event_list[n] = leader;
1382                 n++;
1383         }
1384         if (!dogrp)
1385                 return n;
1386
1387         list_for_each_entry(event, &leader->sibling_list, group_entry) {
1388                 if (!is_x86_event(event) ||
1389                     event->state <= PERF_EVENT_STATE_OFF)
1390                         continue;
1391
1392                 if (n >= max_count)
1393                         return -ENOSPC;
1394
1395                 cpuc->event_list[n] = event;
1396                 n++;
1397         }
1398         return n;
1399 }
1400
1401
1402 static inline void x86_assign_hw_event(struct perf_event *event,
1403                                 struct hw_perf_event *hwc, int idx)
1404 {
1405         hwc->idx = idx;
1406
1407         if (hwc->idx == X86_PMC_IDX_FIXED_BTS) {
1408                 hwc->config_base = 0;
1409                 hwc->event_base = 0;
1410         } else if (hwc->idx >= X86_PMC_IDX_FIXED) {
1411                 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
1412                 /*
1413                  * We set it so that event_base + idx in wrmsr/rdmsr maps to
1414                  * MSR_ARCH_PERFMON_FIXED_CTR0 ... CTR2:
1415                  */
1416                 hwc->event_base =
1417                         MSR_ARCH_PERFMON_FIXED_CTR0 - X86_PMC_IDX_FIXED;
1418         } else {
1419                 hwc->config_base = x86_pmu.eventsel;
1420                 hwc->event_base  = x86_pmu.perfctr;
1421         }
1422 }
1423
1424 void hw_perf_enable(void)
1425 {
1426         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1427         struct perf_event *event;
1428         struct hw_perf_event *hwc;
1429         int i;
1430
1431         if (!x86_pmu_initialized())
1432                 return;
1433         if (cpuc->n_added) {
1434                 /*
1435                  * apply assignment obtained either from
1436                  * hw_perf_group_sched_in() or x86_pmu_enable()
1437                  *
1438                  * step1: save events moving to new counters
1439                  * step2: reprogram moved events into new counters
1440                  */
1441                 for (i = 0; i < cpuc->n_events; i++) {
1442
1443                         event = cpuc->event_list[i];
1444                         hwc = &event->hw;
1445
1446                         if (hwc->idx == -1 || hwc->idx == cpuc->assign[i])
1447                                 continue;
1448
1449                         x86_pmu.disable(hwc, hwc->idx);
1450
1451                         clear_bit(hwc->idx, cpuc->active_mask);
1452                         barrier();
1453                         cpuc->events[hwc->idx] = NULL;
1454
1455                         x86_perf_event_update(event, hwc, hwc->idx);
1456
1457                         hwc->idx = -1;
1458                 }
1459
1460                 for (i = 0; i < cpuc->n_events; i++) {
1461
1462                         event = cpuc->event_list[i];
1463                         hwc = &event->hw;
1464
1465                         if (hwc->idx == -1) {
1466                                 x86_assign_hw_event(event, hwc, cpuc->assign[i]);
1467                                 x86_perf_event_set_period(event, hwc, hwc->idx);
1468                         }
1469                         /*
1470                          * need to mark as active because x86_pmu_disable()
1471                          * clear active_mask and eventsp[] yet it preserves
1472                          * idx
1473                          */
1474                         set_bit(hwc->idx, cpuc->active_mask);
1475                         cpuc->events[hwc->idx] = event;
1476
1477                         x86_pmu.enable(hwc, hwc->idx);
1478                         perf_event_update_userpage(event);
1479                 }
1480                 cpuc->n_added = 0;
1481                 perf_events_lapic_init();
1482         }
1483         x86_pmu.enable_all();
1484 }
1485
1486 static inline u64 intel_pmu_get_status(void)
1487 {
1488         u64 status;
1489
1490         rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1491
1492         return status;
1493 }
1494
1495 static inline void intel_pmu_ack_status(u64 ack)
1496 {
1497         wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
1498 }
1499
1500 static inline void x86_pmu_enable_event(struct hw_perf_event *hwc, int idx)
1501 {
1502         (void)checking_wrmsrl(hwc->config_base + idx,
1503                               hwc->config | ARCH_PERFMON_EVENTSEL0_ENABLE);
1504 }
1505
1506 static inline void x86_pmu_disable_event(struct hw_perf_event *hwc, int idx)
1507 {
1508         (void)checking_wrmsrl(hwc->config_base + idx, hwc->config);
1509 }
1510
1511 static inline void
1512 intel_pmu_disable_fixed(struct hw_perf_event *hwc, int __idx)
1513 {
1514         int idx = __idx - X86_PMC_IDX_FIXED;
1515         u64 ctrl_val, mask;
1516
1517         mask = 0xfULL << (idx * 4);
1518
1519         rdmsrl(hwc->config_base, ctrl_val);
1520         ctrl_val &= ~mask;
1521         (void)checking_wrmsrl(hwc->config_base, ctrl_val);
1522 }
1523
1524 static inline void
1525 p6_pmu_disable_event(struct hw_perf_event *hwc, int idx)
1526 {
1527         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1528         u64 val = P6_NOP_EVENT;
1529
1530         if (cpuc->enabled)
1531                 val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
1532
1533         (void)checking_wrmsrl(hwc->config_base + idx, val);
1534 }
1535
1536 static inline void
1537 intel_pmu_disable_event(struct hw_perf_event *hwc, int idx)
1538 {
1539         if (unlikely(idx == X86_PMC_IDX_FIXED_BTS)) {
1540                 intel_pmu_disable_bts();
1541                 return;
1542         }
1543
1544         if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
1545                 intel_pmu_disable_fixed(hwc, idx);
1546                 return;
1547         }
1548
1549         x86_pmu_disable_event(hwc, idx);
1550 }
1551
1552 static inline void
1553 amd_pmu_disable_event(struct hw_perf_event *hwc, int idx)
1554 {
1555         x86_pmu_disable_event(hwc, idx);
1556 }
1557
1558 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
1559
1560 /*
1561  * Set the next IRQ period, based on the hwc->period_left value.
1562  * To be called with the event disabled in hw:
1563  */
1564 static int
1565 x86_perf_event_set_period(struct perf_event *event,
1566                              struct hw_perf_event *hwc, int idx)
1567 {
1568         s64 left = atomic64_read(&hwc->period_left);
1569         s64 period = hwc->sample_period;
1570         int err, ret = 0;
1571
1572         if (idx == X86_PMC_IDX_FIXED_BTS)
1573                 return 0;
1574
1575         /*
1576          * If we are way outside a reasonable range then just skip forward:
1577          */
1578         if (unlikely(left <= -period)) {
1579                 left = period;
1580                 atomic64_set(&hwc->period_left, left);
1581                 hwc->last_period = period;
1582                 ret = 1;
1583         }
1584
1585         if (unlikely(left <= 0)) {
1586                 left += period;
1587                 atomic64_set(&hwc->period_left, left);
1588                 hwc->last_period = period;
1589                 ret = 1;
1590         }
1591         /*
1592          * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1593          */
1594         if (unlikely(left < 2))
1595                 left = 2;
1596
1597         if (left > x86_pmu.max_period)
1598                 left = x86_pmu.max_period;
1599
1600         per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
1601
1602         /*
1603          * The hw event starts counting from this event offset,
1604          * mark it to be able to extra future deltas:
1605          */
1606         atomic64_set(&hwc->prev_count, (u64)-left);
1607
1608         err = checking_wrmsrl(hwc->event_base + idx,
1609                              (u64)(-left) & x86_pmu.event_mask);
1610
1611         perf_event_update_userpage(event);
1612
1613         return ret;
1614 }
1615
1616 static inline void
1617 intel_pmu_enable_fixed(struct hw_perf_event *hwc, int __idx)
1618 {
1619         int idx = __idx - X86_PMC_IDX_FIXED;
1620         u64 ctrl_val, bits, mask;
1621         int err;
1622
1623         /*
1624          * Enable IRQ generation (0x8),
1625          * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
1626          * if requested:
1627          */
1628         bits = 0x8ULL;
1629         if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
1630                 bits |= 0x2;
1631         if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
1632                 bits |= 0x1;
1633         bits <<= (idx * 4);
1634         mask = 0xfULL << (idx * 4);
1635
1636         rdmsrl(hwc->config_base, ctrl_val);
1637         ctrl_val &= ~mask;
1638         ctrl_val |= bits;
1639         err = checking_wrmsrl(hwc->config_base, ctrl_val);
1640 }
1641
1642 static void p6_pmu_enable_event(struct hw_perf_event *hwc, int idx)
1643 {
1644         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1645         u64 val;
1646
1647         val = hwc->config;
1648         if (cpuc->enabled)
1649                 val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
1650
1651         (void)checking_wrmsrl(hwc->config_base + idx, val);
1652 }
1653
1654
1655 static void intel_pmu_enable_event(struct hw_perf_event *hwc, int idx)
1656 {
1657         if (unlikely(idx == X86_PMC_IDX_FIXED_BTS)) {
1658                 if (!__get_cpu_var(cpu_hw_events).enabled)
1659                         return;
1660
1661                 intel_pmu_enable_bts(hwc->config);
1662                 return;
1663         }
1664
1665         if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
1666                 intel_pmu_enable_fixed(hwc, idx);
1667                 return;
1668         }
1669
1670         x86_pmu_enable_event(hwc, idx);
1671 }
1672
1673 static void amd_pmu_enable_event(struct hw_perf_event *hwc, int idx)
1674 {
1675         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1676
1677         if (cpuc->enabled)
1678                 x86_pmu_enable_event(hwc, idx);
1679 }
1680
1681 /*
1682  * activate a single event
1683  *
1684  * The event is added to the group of enabled events
1685  * but only if it can be scehduled with existing events.
1686  *
1687  * Called with PMU disabled. If successful and return value 1,
1688  * then guaranteed to call perf_enable() and hw_perf_enable()
1689  */
1690 static int x86_pmu_enable(struct perf_event *event)
1691 {
1692         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1693         struct hw_perf_event *hwc;
1694         int assign[X86_PMC_IDX_MAX];
1695         int n, n0, ret;
1696
1697         hwc = &event->hw;
1698
1699         n0 = cpuc->n_events;
1700         n = collect_events(cpuc, event, false);
1701         if (n < 0)
1702                 return n;
1703
1704         ret = x86_schedule_events(cpuc, n, assign);
1705         if (ret)
1706                 return ret;
1707         /*
1708          * copy new assignment, now we know it is possible
1709          * will be used by hw_perf_enable()
1710          */
1711         memcpy(cpuc->assign, assign, n*sizeof(int));
1712
1713         cpuc->n_events = n;
1714         cpuc->n_added  = n - n0;
1715
1716         if (hwc->idx != -1)
1717                 x86_perf_event_set_period(event, hwc, hwc->idx);
1718
1719         return 0;
1720 }
1721
1722 static void x86_pmu_unthrottle(struct perf_event *event)
1723 {
1724         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1725         struct hw_perf_event *hwc = &event->hw;
1726
1727         if (WARN_ON_ONCE(hwc->idx >= X86_PMC_IDX_MAX ||
1728                                 cpuc->events[hwc->idx] != event))
1729                 return;
1730
1731         x86_pmu.enable(hwc, hwc->idx);
1732 }
1733
1734 void perf_event_print_debug(void)
1735 {
1736         u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1737         struct cpu_hw_events *cpuc;
1738         unsigned long flags;
1739         int cpu, idx;
1740
1741         if (!x86_pmu.num_events)
1742                 return;
1743
1744         local_irq_save(flags);
1745
1746         cpu = smp_processor_id();
1747         cpuc = &per_cpu(cpu_hw_events, cpu);
1748
1749         if (x86_pmu.version >= 2) {
1750                 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1751                 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1752                 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1753                 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1754
1755                 pr_info("\n");
1756                 pr_info("CPU#%d: ctrl:       %016llx\n", cpu, ctrl);
1757                 pr_info("CPU#%d: status:     %016llx\n", cpu, status);
1758                 pr_info("CPU#%d: overflow:   %016llx\n", cpu, overflow);
1759                 pr_info("CPU#%d: fixed:      %016llx\n", cpu, fixed);
1760         }
1761         pr_info("CPU#%d: active:       %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1762
1763         for (idx = 0; idx < x86_pmu.num_events; idx++) {
1764                 rdmsrl(x86_pmu.eventsel + idx, pmc_ctrl);
1765                 rdmsrl(x86_pmu.perfctr  + idx, pmc_count);
1766
1767                 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1768
1769                 pr_info("CPU#%d:   gen-PMC%d ctrl:  %016llx\n",
1770                         cpu, idx, pmc_ctrl);
1771                 pr_info("CPU#%d:   gen-PMC%d count: %016llx\n",
1772                         cpu, idx, pmc_count);
1773                 pr_info("CPU#%d:   gen-PMC%d left:  %016llx\n",
1774                         cpu, idx, prev_left);
1775         }
1776         for (idx = 0; idx < x86_pmu.num_events_fixed; idx++) {
1777                 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1778
1779                 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1780                         cpu, idx, pmc_count);
1781         }
1782         local_irq_restore(flags);
1783 }
1784
1785 static void intel_pmu_drain_bts_buffer(struct cpu_hw_events *cpuc)
1786 {
1787         struct debug_store *ds = cpuc->ds;
1788         struct bts_record {
1789                 u64     from;
1790                 u64     to;
1791                 u64     flags;
1792         };
1793         struct perf_event *event = cpuc->events[X86_PMC_IDX_FIXED_BTS];
1794         struct bts_record *at, *top;
1795         struct perf_output_handle handle;
1796         struct perf_event_header header;
1797         struct perf_sample_data data;
1798         struct pt_regs regs;
1799
1800         if (!event)
1801                 return;
1802
1803         if (!ds)
1804                 return;
1805
1806         at  = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
1807         top = (struct bts_record *)(unsigned long)ds->bts_index;
1808
1809         if (top <= at)
1810                 return;
1811
1812         ds->bts_index = ds->bts_buffer_base;
1813
1814
1815         data.period     = event->hw.last_period;
1816         data.addr       = 0;
1817         data.raw        = NULL;
1818         regs.ip         = 0;
1819
1820         /*
1821          * Prepare a generic sample, i.e. fill in the invariant fields.
1822          * We will overwrite the from and to address before we output
1823          * the sample.
1824          */
1825         perf_prepare_sample(&header, &data, event, &regs);
1826
1827         if (perf_output_begin(&handle, event,
1828                               header.size * (top - at), 1, 1))
1829                 return;
1830
1831         for (; at < top; at++) {
1832                 data.ip         = at->from;
1833                 data.addr       = at->to;
1834
1835                 perf_output_sample(&handle, &header, &data, event);
1836         }
1837
1838         perf_output_end(&handle);
1839
1840         /* There's new data available. */
1841         event->hw.interrupts++;
1842         event->pending_kill = POLL_IN;
1843 }
1844
1845 static void x86_pmu_disable(struct perf_event *event)
1846 {
1847         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1848         struct hw_perf_event *hwc = &event->hw;
1849         int i, idx = hwc->idx;
1850
1851         /*
1852          * Must be done before we disable, otherwise the nmi handler
1853          * could reenable again:
1854          */
1855         clear_bit(idx, cpuc->active_mask);
1856         x86_pmu.disable(hwc, idx);
1857
1858         /*
1859          * Make sure the cleared pointer becomes visible before we
1860          * (potentially) free the event:
1861          */
1862         barrier();
1863
1864         /*
1865          * Drain the remaining delta count out of a event
1866          * that we are disabling:
1867          */
1868         x86_perf_event_update(event, hwc, idx);
1869
1870         /* Drain the remaining BTS records. */
1871         if (unlikely(idx == X86_PMC_IDX_FIXED_BTS))
1872                 intel_pmu_drain_bts_buffer(cpuc);
1873
1874         cpuc->events[idx] = NULL;
1875
1876         for (i = 0; i < cpuc->n_events; i++) {
1877                 if (event == cpuc->event_list[i]) {
1878
1879                         if (x86_pmu.put_event_constraints)
1880                                 x86_pmu.put_event_constraints(cpuc, event);
1881
1882                         while (++i < cpuc->n_events)
1883                                 cpuc->event_list[i-1] = cpuc->event_list[i];
1884
1885                         --cpuc->n_events;
1886                 }
1887         }
1888         perf_event_update_userpage(event);
1889 }
1890
1891 /*
1892  * Save and restart an expired event. Called by NMI contexts,
1893  * so it has to be careful about preempting normal event ops:
1894  */
1895 static int intel_pmu_save_and_restart(struct perf_event *event)
1896 {
1897         struct hw_perf_event *hwc = &event->hw;
1898         int idx = hwc->idx;
1899         int ret;
1900
1901         x86_perf_event_update(event, hwc, idx);
1902         ret = x86_perf_event_set_period(event, hwc, idx);
1903
1904         if (event->state == PERF_EVENT_STATE_ACTIVE)
1905                 intel_pmu_enable_event(hwc, idx);
1906
1907         return ret;
1908 }
1909
1910 static void intel_pmu_reset(void)
1911 {
1912         struct debug_store *ds = __get_cpu_var(cpu_hw_events).ds;
1913         unsigned long flags;
1914         int idx;
1915
1916         if (!x86_pmu.num_events)
1917                 return;
1918
1919         local_irq_save(flags);
1920
1921         printk("clearing PMU state on CPU#%d\n", smp_processor_id());
1922
1923         for (idx = 0; idx < x86_pmu.num_events; idx++) {
1924                 checking_wrmsrl(x86_pmu.eventsel + idx, 0ull);
1925                 checking_wrmsrl(x86_pmu.perfctr  + idx, 0ull);
1926         }
1927         for (idx = 0; idx < x86_pmu.num_events_fixed; idx++) {
1928                 checking_wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
1929         }
1930         if (ds)
1931                 ds->bts_index = ds->bts_buffer_base;
1932
1933         local_irq_restore(flags);
1934 }
1935
1936 static int p6_pmu_handle_irq(struct pt_regs *regs)
1937 {
1938         struct perf_sample_data data;
1939         struct cpu_hw_events *cpuc;
1940         struct perf_event *event;
1941         struct hw_perf_event *hwc;
1942         int idx, handled = 0;
1943         u64 val;
1944
1945         data.addr = 0;
1946         data.raw = NULL;
1947
1948         cpuc = &__get_cpu_var(cpu_hw_events);
1949
1950         for (idx = 0; idx < x86_pmu.num_events; idx++) {
1951                 if (!test_bit(idx, cpuc->active_mask))
1952                         continue;
1953
1954                 event = cpuc->events[idx];
1955                 hwc = &event->hw;
1956
1957                 val = x86_perf_event_update(event, hwc, idx);
1958                 if (val & (1ULL << (x86_pmu.event_bits - 1)))
1959                         continue;
1960
1961                 /*
1962                  * event overflow
1963                  */
1964                 handled         = 1;
1965                 data.period     = event->hw.last_period;
1966
1967                 if (!x86_perf_event_set_period(event, hwc, idx))
1968                         continue;
1969
1970                 if (perf_event_overflow(event, 1, &data, regs))
1971                         p6_pmu_disable_event(hwc, idx);
1972         }
1973
1974         if (handled)
1975                 inc_irq_stat(apic_perf_irqs);
1976
1977         return handled;
1978 }
1979
1980 /*
1981  * This handler is triggered by the local APIC, so the APIC IRQ handling
1982  * rules apply:
1983  */
1984 static int intel_pmu_handle_irq(struct pt_regs *regs)
1985 {
1986         struct perf_sample_data data;
1987         struct cpu_hw_events *cpuc;
1988         int bit, loops;
1989         u64 ack, status;
1990
1991         data.addr = 0;
1992         data.raw = NULL;
1993
1994         cpuc = &__get_cpu_var(cpu_hw_events);
1995
1996         perf_disable();
1997         intel_pmu_drain_bts_buffer(cpuc);
1998         status = intel_pmu_get_status();
1999         if (!status) {
2000                 perf_enable();
2001                 return 0;
2002         }
2003
2004         loops = 0;
2005 again:
2006         if (++loops > 100) {
2007                 WARN_ONCE(1, "perfevents: irq loop stuck!\n");
2008                 perf_event_print_debug();
2009                 intel_pmu_reset();
2010                 perf_enable();
2011                 return 1;
2012         }
2013
2014         inc_irq_stat(apic_perf_irqs);
2015         ack = status;
2016         for_each_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
2017                 struct perf_event *event = cpuc->events[bit];
2018
2019                 clear_bit(bit, (unsigned long *) &status);
2020                 if (!test_bit(bit, cpuc->active_mask))
2021                         continue;
2022
2023                 if (!intel_pmu_save_and_restart(event))
2024                         continue;
2025
2026                 data.period = event->hw.last_period;
2027
2028                 if (perf_event_overflow(event, 1, &data, regs))
2029                         intel_pmu_disable_event(&event->hw, bit);
2030         }
2031
2032         intel_pmu_ack_status(ack);
2033
2034         /*
2035          * Repeat if there is more work to be done:
2036          */
2037         status = intel_pmu_get_status();
2038         if (status)
2039                 goto again;
2040
2041         perf_enable();
2042
2043         return 1;
2044 }
2045
2046 static int amd_pmu_handle_irq(struct pt_regs *regs)
2047 {
2048         struct perf_sample_data data;
2049         struct cpu_hw_events *cpuc;
2050         struct perf_event *event;
2051         struct hw_perf_event *hwc;
2052         int idx, handled = 0;
2053         u64 val;
2054
2055         data.addr = 0;
2056         data.raw = NULL;
2057
2058         cpuc = &__get_cpu_var(cpu_hw_events);
2059
2060         for (idx = 0; idx < x86_pmu.num_events; idx++) {
2061                 if (!test_bit(idx, cpuc->active_mask))
2062                         continue;
2063
2064                 event = cpuc->events[idx];
2065                 hwc = &event->hw;
2066
2067                 val = x86_perf_event_update(event, hwc, idx);
2068                 if (val & (1ULL << (x86_pmu.event_bits - 1)))
2069                         continue;
2070
2071                 /*
2072                  * event overflow
2073                  */
2074                 handled         = 1;
2075                 data.period     = event->hw.last_period;
2076
2077                 if (!x86_perf_event_set_period(event, hwc, idx))
2078                         continue;
2079
2080                 if (perf_event_overflow(event, 1, &data, regs))
2081                         amd_pmu_disable_event(hwc, idx);
2082         }
2083
2084         if (handled)
2085                 inc_irq_stat(apic_perf_irqs);
2086
2087         return handled;
2088 }
2089
2090 void smp_perf_pending_interrupt(struct pt_regs *regs)
2091 {
2092         irq_enter();
2093         ack_APIC_irq();
2094         inc_irq_stat(apic_pending_irqs);
2095         perf_event_do_pending();
2096         irq_exit();
2097 }
2098
2099 void set_perf_event_pending(void)
2100 {
2101 #ifdef CONFIG_X86_LOCAL_APIC
2102         if (!x86_pmu.apic || !x86_pmu_initialized())
2103                 return;
2104
2105         apic->send_IPI_self(LOCAL_PENDING_VECTOR);
2106 #endif
2107 }
2108
2109 void perf_events_lapic_init(void)
2110 {
2111 #ifdef CONFIG_X86_LOCAL_APIC
2112         if (!x86_pmu.apic || !x86_pmu_initialized())
2113                 return;
2114
2115         /*
2116          * Always use NMI for PMU
2117          */
2118         apic_write(APIC_LVTPC, APIC_DM_NMI);
2119 #endif
2120 }
2121
2122 static int __kprobes
2123 perf_event_nmi_handler(struct notifier_block *self,
2124                          unsigned long cmd, void *__args)
2125 {
2126         struct die_args *args = __args;
2127         struct pt_regs *regs;
2128
2129         if (!atomic_read(&active_events))
2130                 return NOTIFY_DONE;
2131
2132         switch (cmd) {
2133         case DIE_NMI:
2134         case DIE_NMI_IPI:
2135                 break;
2136
2137         default:
2138                 return NOTIFY_DONE;
2139         }
2140
2141         regs = args->regs;
2142
2143 #ifdef CONFIG_X86_LOCAL_APIC
2144         apic_write(APIC_LVTPC, APIC_DM_NMI);
2145 #endif
2146         /*
2147          * Can't rely on the handled return value to say it was our NMI, two
2148          * events could trigger 'simultaneously' raising two back-to-back NMIs.
2149          *
2150          * If the first NMI handles both, the latter will be empty and daze
2151          * the CPU.
2152          */
2153         x86_pmu.handle_irq(regs);
2154
2155         return NOTIFY_STOP;
2156 }
2157
2158 static struct event_constraint bts_constraint =
2159         EVENT_CONSTRAINT(0, 1ULL << X86_PMC_IDX_FIXED_BTS, 0);
2160
2161 static int intel_special_constraints(struct perf_event *event,
2162                                      unsigned long *idxmsk)
2163 {
2164         unsigned int hw_event;
2165
2166         hw_event = event->hw.config & INTEL_ARCH_EVENT_MASK;
2167
2168         if (unlikely((hw_event ==
2169                       x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS)) &&
2170                      (event->hw.sample_period == 1))) {
2171
2172                 bitmap_copy((unsigned long *)idxmsk,
2173                             (unsigned long *)bts_constraint.idxmsk,
2174                             X86_PMC_IDX_MAX);
2175                 return 1;
2176         }
2177         return 0;
2178 }
2179
2180 static void intel_get_event_constraints(struct cpu_hw_events *cpuc,
2181                                         struct perf_event *event,
2182                                         unsigned long *idxmsk)
2183 {
2184         const struct event_constraint *c;
2185
2186         /*
2187          * cleanup bitmask
2188          */
2189         bitmap_zero(idxmsk, X86_PMC_IDX_MAX);
2190
2191         if (intel_special_constraints(event, idxmsk))
2192                 return;
2193
2194         if (x86_pmu.event_constraints) {
2195                 for_each_event_constraint(c, x86_pmu.event_constraints) {
2196                         if ((event->hw.config & c->cmask) == c->code) {
2197                                 bitmap_copy(idxmsk, c->idxmsk, X86_PMC_IDX_MAX);
2198                                 return;
2199                         }
2200                 }
2201         }
2202         /* no constraints, means supports all generic counters */
2203         bitmap_fill((unsigned long *)idxmsk, x86_pmu.num_events);
2204 }
2205
2206 static void amd_get_event_constraints(struct cpu_hw_events *cpuc,
2207                                       struct perf_event *event,
2208                                       unsigned long *idxmsk)
2209 {
2210         /* no constraints, means supports all generic counters */
2211         bitmap_fill(idxmsk, x86_pmu.num_events);
2212 }
2213
2214 static int x86_event_sched_in(struct perf_event *event,
2215                           struct perf_cpu_context *cpuctx, int cpu)
2216 {
2217         int ret = 0;
2218
2219         event->state = PERF_EVENT_STATE_ACTIVE;
2220         event->oncpu = cpu;
2221         event->tstamp_running += event->ctx->time - event->tstamp_stopped;
2222
2223         if (!is_x86_event(event))
2224                 ret = event->pmu->enable(event);
2225
2226         if (!ret && !is_software_event(event))
2227                 cpuctx->active_oncpu++;
2228
2229         if (!ret && event->attr.exclusive)
2230                 cpuctx->exclusive = 1;
2231
2232         return ret;
2233 }
2234
2235 static void x86_event_sched_out(struct perf_event *event,
2236                             struct perf_cpu_context *cpuctx, int cpu)
2237 {
2238         event->state = PERF_EVENT_STATE_INACTIVE;
2239         event->oncpu = -1;
2240
2241         if (!is_x86_event(event))
2242                 event->pmu->disable(event);
2243
2244         event->tstamp_running -= event->ctx->time - event->tstamp_stopped;
2245
2246         if (!is_software_event(event))
2247                 cpuctx->active_oncpu--;
2248
2249         if (event->attr.exclusive || !cpuctx->active_oncpu)
2250                 cpuctx->exclusive = 0;
2251 }
2252
2253 /*
2254  * Called to enable a whole group of events.
2255  * Returns 1 if the group was enabled, or -EAGAIN if it could not be.
2256  * Assumes the caller has disabled interrupts and has
2257  * frozen the PMU with hw_perf_save_disable.
2258  *
2259  * called with PMU disabled. If successful and return value 1,
2260  * then guaranteed to call perf_enable() and hw_perf_enable()
2261  */
2262 int hw_perf_group_sched_in(struct perf_event *leader,
2263                struct perf_cpu_context *cpuctx,
2264                struct perf_event_context *ctx, int cpu)
2265 {
2266         struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
2267         struct perf_event *sub;
2268         int assign[X86_PMC_IDX_MAX];
2269         int n0, n1, ret;
2270
2271         /* n0 = total number of events */
2272         n0 = collect_events(cpuc, leader, true);
2273         if (n0 < 0)
2274                 return n0;
2275
2276         ret = x86_schedule_events(cpuc, n0, assign);
2277         if (ret)
2278                 return ret;
2279
2280         ret = x86_event_sched_in(leader, cpuctx, cpu);
2281         if (ret)
2282                 return ret;
2283
2284         n1 = 1;
2285         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2286                 if (sub->state > PERF_EVENT_STATE_OFF) {
2287                         ret = x86_event_sched_in(sub, cpuctx, cpu);
2288                         if (ret)
2289                                 goto undo;
2290                         ++n1;
2291                 }
2292         }
2293         /*
2294          * copy new assignment, now we know it is possible
2295          * will be used by hw_perf_enable()
2296          */
2297         memcpy(cpuc->assign, assign, n0*sizeof(int));
2298
2299         cpuc->n_events  = n0;
2300         cpuc->n_added   = n1;
2301         ctx->nr_active += n1;
2302
2303         /*
2304          * 1 means successful and events are active
2305          * This is not quite true because we defer
2306          * actual activation until hw_perf_enable() but
2307          * this way we* ensure caller won't try to enable
2308          * individual events
2309          */
2310         return 1;
2311 undo:
2312         x86_event_sched_out(leader, cpuctx, cpu);
2313         n0  = 1;
2314         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2315                 if (sub->state == PERF_EVENT_STATE_ACTIVE) {
2316                         x86_event_sched_out(sub, cpuctx, cpu);
2317                         if (++n0 == n1)
2318                                 break;
2319                 }
2320         }
2321         return ret;
2322 }
2323
2324 static __read_mostly struct notifier_block perf_event_nmi_notifier = {
2325         .notifier_call          = perf_event_nmi_handler,
2326         .next                   = NULL,
2327         .priority               = 1
2328 };
2329
2330 static __initconst struct x86_pmu p6_pmu = {
2331         .name                   = "p6",
2332         .handle_irq             = p6_pmu_handle_irq,
2333         .disable_all            = p6_pmu_disable_all,
2334         .enable_all             = p6_pmu_enable_all,
2335         .enable                 = p6_pmu_enable_event,
2336         .disable                = p6_pmu_disable_event,
2337         .eventsel               = MSR_P6_EVNTSEL0,
2338         .perfctr                = MSR_P6_PERFCTR0,
2339         .event_map              = p6_pmu_event_map,
2340         .raw_event              = p6_pmu_raw_event,
2341         .max_events             = ARRAY_SIZE(p6_perfmon_event_map),
2342         .apic                   = 1,
2343         .max_period             = (1ULL << 31) - 1,
2344         .version                = 0,
2345         .num_events             = 2,
2346         /*
2347          * Events have 40 bits implemented. However they are designed such
2348          * that bits [32-39] are sign extensions of bit 31. As such the
2349          * effective width of a event for P6-like PMU is 32 bits only.
2350          *
2351          * See IA-32 Intel Architecture Software developer manual Vol 3B
2352          */
2353         .event_bits             = 32,
2354         .event_mask             = (1ULL << 32) - 1,
2355         .get_event_constraints  = intel_get_event_constraints,
2356         .event_constraints      = intel_p6_event_constraints
2357 };
2358
2359 static __initconst struct x86_pmu intel_pmu = {
2360         .name                   = "Intel",
2361         .handle_irq             = intel_pmu_handle_irq,
2362         .disable_all            = intel_pmu_disable_all,
2363         .enable_all             = intel_pmu_enable_all,
2364         .enable                 = intel_pmu_enable_event,
2365         .disable                = intel_pmu_disable_event,
2366         .eventsel               = MSR_ARCH_PERFMON_EVENTSEL0,
2367         .perfctr                = MSR_ARCH_PERFMON_PERFCTR0,
2368         .event_map              = intel_pmu_event_map,
2369         .raw_event              = intel_pmu_raw_event,
2370         .max_events             = ARRAY_SIZE(intel_perfmon_event_map),
2371         .apic                   = 1,
2372         /*
2373          * Intel PMCs cannot be accessed sanely above 32 bit width,
2374          * so we install an artificial 1<<31 period regardless of
2375          * the generic event period:
2376          */
2377         .max_period             = (1ULL << 31) - 1,
2378         .enable_bts             = intel_pmu_enable_bts,
2379         .disable_bts            = intel_pmu_disable_bts,
2380         .get_event_constraints  = intel_get_event_constraints
2381 };
2382
2383 static __initconst struct x86_pmu amd_pmu = {
2384         .name                   = "AMD",
2385         .handle_irq             = amd_pmu_handle_irq,
2386         .disable_all            = amd_pmu_disable_all,
2387         .enable_all             = amd_pmu_enable_all,
2388         .enable                 = amd_pmu_enable_event,
2389         .disable                = amd_pmu_disable_event,
2390         .eventsel               = MSR_K7_EVNTSEL0,
2391         .perfctr                = MSR_K7_PERFCTR0,
2392         .event_map              = amd_pmu_event_map,
2393         .raw_event              = amd_pmu_raw_event,
2394         .max_events             = ARRAY_SIZE(amd_perfmon_event_map),
2395         .num_events             = 4,
2396         .event_bits             = 48,
2397         .event_mask             = (1ULL << 48) - 1,
2398         .apic                   = 1,
2399         /* use highest bit to detect overflow */
2400         .max_period             = (1ULL << 47) - 1,
2401         .get_event_constraints  = amd_get_event_constraints
2402 };
2403
2404 static __init int p6_pmu_init(void)
2405 {
2406         switch (boot_cpu_data.x86_model) {
2407         case 1:
2408         case 3:  /* Pentium Pro */
2409         case 5:
2410         case 6:  /* Pentium II */
2411         case 7:
2412         case 8:
2413         case 11: /* Pentium III */
2414         case 9:
2415         case 13:
2416                 /* Pentium M */
2417                 break;
2418         default:
2419                 pr_cont("unsupported p6 CPU model %d ",
2420                         boot_cpu_data.x86_model);
2421                 return -ENODEV;
2422         }
2423
2424         x86_pmu = p6_pmu;
2425
2426         return 0;
2427 }
2428
2429 static __init int intel_pmu_init(void)
2430 {
2431         union cpuid10_edx edx;
2432         union cpuid10_eax eax;
2433         unsigned int unused;
2434         unsigned int ebx;
2435         int version;
2436
2437         if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
2438                 /* check for P6 processor family */
2439            if (boot_cpu_data.x86 == 6) {
2440                 return p6_pmu_init();
2441            } else {
2442                 return -ENODEV;
2443            }
2444         }
2445
2446         /*
2447          * Check whether the Architectural PerfMon supports
2448          * Branch Misses Retired hw_event or not.
2449          */
2450         cpuid(10, &eax.full, &ebx, &unused, &edx.full);
2451         if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED)
2452                 return -ENODEV;
2453
2454         version = eax.split.version_id;
2455         if (version < 2)
2456                 return -ENODEV;
2457
2458         x86_pmu                         = intel_pmu;
2459         x86_pmu.version                 = version;
2460         x86_pmu.num_events              = eax.split.num_events;
2461         x86_pmu.event_bits              = eax.split.bit_width;
2462         x86_pmu.event_mask              = (1ULL << eax.split.bit_width) - 1;
2463
2464         /*
2465          * Quirk: v2 perfmon does not report fixed-purpose events, so
2466          * assume at least 3 events:
2467          */
2468         x86_pmu.num_events_fixed        = max((int)edx.split.num_events_fixed, 3);
2469
2470         /*
2471          * Install the hw-cache-events table:
2472          */
2473         switch (boot_cpu_data.x86_model) {
2474         case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */
2475         case 22: /* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */
2476         case 23: /* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */
2477         case 29: /* six-core 45 nm xeon "Dunnington" */
2478                 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
2479                        sizeof(hw_cache_event_ids));
2480
2481                 x86_pmu.event_constraints = intel_core_event_constraints;
2482                 pr_cont("Core2 events, ");
2483                 break;
2484         case 26:
2485                 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
2486                        sizeof(hw_cache_event_ids));
2487
2488                 x86_pmu.event_constraints = intel_nehalem_event_constraints;
2489                 pr_cont("Nehalem/Corei7 events, ");
2490                 break;
2491         case 28:
2492                 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
2493                        sizeof(hw_cache_event_ids));
2494
2495                 x86_pmu.event_constraints = intel_gen_event_constraints;
2496                 pr_cont("Atom events, ");
2497                 break;
2498         default:
2499                 /*
2500                  * default constraints for v2 and up
2501                  */
2502                 x86_pmu.event_constraints = intel_gen_event_constraints;
2503                 pr_cont("generic architected perfmon, ");
2504         }
2505         return 0;
2506 }
2507
2508 static __init int amd_pmu_init(void)
2509 {
2510         /* Performance-monitoring supported from K7 and later: */
2511         if (boot_cpu_data.x86 < 6)
2512                 return -ENODEV;
2513
2514         x86_pmu = amd_pmu;
2515
2516         /* Events are common for all AMDs */
2517         memcpy(hw_cache_event_ids, amd_hw_cache_event_ids,
2518                sizeof(hw_cache_event_ids));
2519
2520         return 0;
2521 }
2522
2523 static void __init pmu_check_apic(void)
2524 {
2525         if (cpu_has_apic)
2526                 return;
2527
2528         x86_pmu.apic = 0;
2529         pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
2530         pr_info("no hardware sampling interrupt available.\n");
2531 }
2532
2533 void __init init_hw_perf_events(void)
2534 {
2535         int err;
2536
2537         pr_info("Performance Events: ");
2538
2539         switch (boot_cpu_data.x86_vendor) {
2540         case X86_VENDOR_INTEL:
2541                 err = intel_pmu_init();
2542                 break;
2543         case X86_VENDOR_AMD:
2544                 err = amd_pmu_init();
2545                 break;
2546         default:
2547                 return;
2548         }
2549         if (err != 0) {
2550                 pr_cont("no PMU driver, software events only.\n");
2551                 return;
2552         }
2553
2554         pmu_check_apic();
2555
2556         pr_cont("%s PMU driver.\n", x86_pmu.name);
2557
2558         if (x86_pmu.num_events > X86_PMC_MAX_GENERIC) {
2559                 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
2560                      x86_pmu.num_events, X86_PMC_MAX_GENERIC);
2561                 x86_pmu.num_events = X86_PMC_MAX_GENERIC;
2562         }
2563         perf_event_mask = (1 << x86_pmu.num_events) - 1;
2564         perf_max_events = x86_pmu.num_events;
2565
2566         if (x86_pmu.num_events_fixed > X86_PMC_MAX_FIXED) {
2567                 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
2568                      x86_pmu.num_events_fixed, X86_PMC_MAX_FIXED);
2569                 x86_pmu.num_events_fixed = X86_PMC_MAX_FIXED;
2570         }
2571
2572         perf_event_mask |=
2573                 ((1LL << x86_pmu.num_events_fixed)-1) << X86_PMC_IDX_FIXED;
2574         x86_pmu.intel_ctrl = perf_event_mask;
2575
2576         perf_events_lapic_init();
2577         register_die_notifier(&perf_event_nmi_notifier);
2578
2579         pr_info("... version:                %d\n",     x86_pmu.version);
2580         pr_info("... bit width:              %d\n",     x86_pmu.event_bits);
2581         pr_info("... generic registers:      %d\n",     x86_pmu.num_events);
2582         pr_info("... value mask:             %016Lx\n", x86_pmu.event_mask);
2583         pr_info("... max period:             %016Lx\n", x86_pmu.max_period);
2584         pr_info("... fixed-purpose events:   %d\n",     x86_pmu.num_events_fixed);
2585         pr_info("... event mask:             %016Lx\n", perf_event_mask);
2586 }
2587
2588 static inline void x86_pmu_read(struct perf_event *event)
2589 {
2590         x86_perf_event_update(event, &event->hw, event->hw.idx);
2591 }
2592
2593 static const struct pmu pmu = {
2594         .enable         = x86_pmu_enable,
2595         .disable        = x86_pmu_disable,
2596         .read           = x86_pmu_read,
2597         .unthrottle     = x86_pmu_unthrottle,
2598 };
2599
2600 /*
2601  * validate a single event group
2602  *
2603  * validation include:
2604  *      - check events are compatible which each other
2605  *      - events do not compete for the same counter
2606  *      - number of events <= number of counters
2607  *
2608  * validation ensures the group can be loaded onto the
2609  * PMU if it was the only group available.
2610  */
2611 static int validate_group(struct perf_event *event)
2612 {
2613         struct perf_event *leader = event->group_leader;
2614         struct cpu_hw_events *fake_cpuc;
2615         int ret, n;
2616
2617         ret = -ENOMEM;
2618         fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO);
2619         if (!fake_cpuc)
2620                 goto out;
2621
2622         /*
2623          * the event is not yet connected with its
2624          * siblings therefore we must first collect
2625          * existing siblings, then add the new event
2626          * before we can simulate the scheduling
2627          */
2628         ret = -ENOSPC;
2629         n = collect_events(fake_cpuc, leader, true);
2630         if (n < 0)
2631                 goto out_free;
2632
2633         fake_cpuc->n_events = n;
2634         n = collect_events(fake_cpuc, event, false);
2635         if (n < 0)
2636                 goto out_free;
2637
2638         fake_cpuc->n_events = n;
2639
2640         ret = x86_schedule_events(fake_cpuc, n, NULL);
2641
2642 out_free:
2643         kfree(fake_cpuc);
2644 out:
2645         return ret;
2646 }
2647
2648 const struct pmu *hw_perf_event_init(struct perf_event *event)
2649 {
2650         const struct pmu *tmp;
2651         int err;
2652
2653         err = __hw_perf_event_init(event);
2654         if (!err) {
2655                 /*
2656                  * we temporarily connect event to its pmu
2657                  * such that validate_group() can classify
2658                  * it as an x86 event using is_x86_event()
2659                  */
2660                 tmp = event->pmu;
2661                 event->pmu = &pmu;
2662
2663                 if (event->group_leader != event)
2664                         err = validate_group(event);
2665
2666                 event->pmu = tmp;
2667         }
2668         if (err) {
2669                 if (event->destroy)
2670                         event->destroy(event);
2671                 return ERR_PTR(err);
2672         }
2673
2674         return &pmu;
2675 }
2676
2677 /*
2678  * callchain support
2679  */
2680
2681 static inline
2682 void callchain_store(struct perf_callchain_entry *entry, u64 ip)
2683 {
2684         if (entry->nr < PERF_MAX_STACK_DEPTH)
2685                 entry->ip[entry->nr++] = ip;
2686 }
2687
2688 static DEFINE_PER_CPU(struct perf_callchain_entry, pmc_irq_entry);
2689 static DEFINE_PER_CPU(struct perf_callchain_entry, pmc_nmi_entry);
2690
2691
2692 static void
2693 backtrace_warning_symbol(void *data, char *msg, unsigned long symbol)
2694 {
2695         /* Ignore warnings */
2696 }
2697
2698 static void backtrace_warning(void *data, char *msg)
2699 {
2700         /* Ignore warnings */
2701 }
2702
2703 static int backtrace_stack(void *data, char *name)
2704 {
2705         return 0;
2706 }
2707
2708 static void backtrace_address(void *data, unsigned long addr, int reliable)
2709 {
2710         struct perf_callchain_entry *entry = data;
2711
2712         if (reliable)
2713                 callchain_store(entry, addr);
2714 }
2715
2716 static const struct stacktrace_ops backtrace_ops = {
2717         .warning                = backtrace_warning,
2718         .warning_symbol         = backtrace_warning_symbol,
2719         .stack                  = backtrace_stack,
2720         .address                = backtrace_address,
2721         .walk_stack             = print_context_stack_bp,
2722 };
2723
2724 #include "../dumpstack.h"
2725
2726 static void
2727 perf_callchain_kernel(struct pt_regs *regs, struct perf_callchain_entry *entry)
2728 {
2729         callchain_store(entry, PERF_CONTEXT_KERNEL);
2730         callchain_store(entry, regs->ip);
2731
2732         dump_trace(NULL, regs, NULL, regs->bp, &backtrace_ops, entry);
2733 }
2734
2735 /*
2736  * best effort, GUP based copy_from_user() that assumes IRQ or NMI context
2737  */
2738 static unsigned long
2739 copy_from_user_nmi(void *to, const void __user *from, unsigned long n)
2740 {
2741         unsigned long offset, addr = (unsigned long)from;
2742         int type = in_nmi() ? KM_NMI : KM_IRQ0;
2743         unsigned long size, len = 0;
2744         struct page *page;
2745         void *map;
2746         int ret;
2747
2748         do {
2749                 ret = __get_user_pages_fast(addr, 1, 0, &page);
2750                 if (!ret)
2751                         break;
2752
2753                 offset = addr & (PAGE_SIZE - 1);
2754                 size = min(PAGE_SIZE - offset, n - len);
2755
2756                 map = kmap_atomic(page, type);
2757                 memcpy(to, map+offset, size);
2758                 kunmap_atomic(map, type);
2759                 put_page(page);
2760
2761                 len  += size;
2762                 to   += size;
2763                 addr += size;
2764
2765         } while (len < n);
2766
2767         return len;
2768 }
2769
2770 static int copy_stack_frame(const void __user *fp, struct stack_frame *frame)
2771 {
2772         unsigned long bytes;
2773
2774         bytes = copy_from_user_nmi(frame, fp, sizeof(*frame));
2775
2776         return bytes == sizeof(*frame);
2777 }
2778
2779 static void
2780 perf_callchain_user(struct pt_regs *regs, struct perf_callchain_entry *entry)
2781 {
2782         struct stack_frame frame;
2783         const void __user *fp;
2784
2785         if (!user_mode(regs))
2786                 regs = task_pt_regs(current);
2787
2788         fp = (void __user *)regs->bp;
2789
2790         callchain_store(entry, PERF_CONTEXT_USER);
2791         callchain_store(entry, regs->ip);
2792
2793         while (entry->nr < PERF_MAX_STACK_DEPTH) {
2794                 frame.next_frame             = NULL;
2795                 frame.return_address = 0;
2796
2797                 if (!copy_stack_frame(fp, &frame))
2798                         break;
2799
2800                 if ((unsigned long)fp < regs->sp)
2801                         break;
2802
2803                 callchain_store(entry, frame.return_address);
2804                 fp = frame.next_frame;
2805         }
2806 }
2807
2808 static void
2809 perf_do_callchain(struct pt_regs *regs, struct perf_callchain_entry *entry)
2810 {
2811         int is_user;
2812
2813         if (!regs)
2814                 return;
2815
2816         is_user = user_mode(regs);
2817
2818         if (is_user && current->state != TASK_RUNNING)
2819                 return;
2820
2821         if (!is_user)
2822                 perf_callchain_kernel(regs, entry);
2823
2824         if (current->mm)
2825                 perf_callchain_user(regs, entry);
2826 }
2827
2828 struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2829 {
2830         struct perf_callchain_entry *entry;
2831
2832         if (in_nmi())
2833                 entry = &__get_cpu_var(pmc_nmi_entry);
2834         else
2835                 entry = &__get_cpu_var(pmc_irq_entry);
2836
2837         entry->nr = 0;
2838
2839         perf_do_callchain(regs, entry);
2840
2841         return entry;
2842 }
2843
2844 void hw_perf_event_setup_online(int cpu)
2845 {
2846         init_debug_store_on_cpu(cpu);
2847 }