?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge branch 'for-linus' of git://oss.sgi.com/xfs/xfs
[linux/fpc-iii.git] / arch / sparc / kernel / perf_event.c
blob0ec92c8861dd3a68fe0465b71ccf8007ad6e40a7
1 /* Performance event support for sparc64.
2 *
3 * Copyright (C) 2009, 2010 David S. Miller <davem@davemloft.net>
4 *
5 * This code is based almost entirely upon the x86 perf event
6 * code, which is:
7 *
8 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
9 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
10 * Copyright (C) 2009 Jaswinder Singh Rajput
11 * Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
12 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
13 */
14
15 #include <linux/perf_event.h>
16 #include <linux/kprobes.h>
17 #include <linux/ftrace.h>
18 #include <linux/kernel.h>
19 #include <linux/kdebug.h>
20 #include <linux/mutex.h>
21
22 #include <asm/stacktrace.h>
23 #include <asm/cpudata.h>
24 #include <asm/uaccess.h>
25 #include <asm/atomic.h>
26 #include <asm/nmi.h>
27 #include <asm/pcr.h>
28
29 #include "kstack.h"
30
31 /* Sparc64 chips have two performance counters, 32-bits each, with
32 * overflow interrupts generated on transition from 0xffffffff to 0.
33 * The counters are accessed in one go using a 64-bit register.
34 *
35 * Both counters are controlled using a single control register. The
36 * only way to stop all sampling is to clear all of the context (user,
37 * supervisor, hypervisor) sampling enable bits. But these bits apply
38 * to both counters, thus the two counters can't be enabled/disabled
39 * individually.
40 *
41 * The control register has two event fields, one for each of the two
42 * counters. It's thus nearly impossible to have one counter going
43 * while keeping the other one stopped. Therefore it is possible to
44 * get overflow interrupts for counters not currently "in use" and
45 * that condition must be checked in the overflow interrupt handler.
46 *
47 * So we use a hack, in that we program inactive counters with the
48 * "sw_count0" and "sw_count1" events. These count how many times
49 * the instruction "sethi %hi(0xfc000), %g0" is executed. It's an
50 * unusual way to encode a NOP and therefore will not trigger in
51 * normal code.
52 */
53
54 #define MAX_HWEVENTS 2
55 #define MAX_PERIOD ((1UL << 32) - 1)
56
57 #define PIC_UPPER_INDEX 0
58 #define PIC_LOWER_INDEX 1
59 #define PIC_NO_INDEX -1
60
61 struct cpu_hw_events {
62 /* Number of events currently scheduled onto this cpu.
63 * This tells how many entries in the arrays below
64 * are valid.
65 */
66 int n_events;
67
68 /* Number of new events added since the last hw_perf_disable().
69 * This works because the perf event layer always adds new
70 * events inside of a perf_{disable,enable}() sequence.
71 */
72 int n_added;
73
74 /* Array of events current scheduled on this cpu. */
75 struct perf_event *event[MAX_HWEVENTS];
76
77 /* Array of encoded longs, specifying the %pcr register
78 * encoding and the mask of PIC counters this even can
79 * be scheduled on. See perf_event_encode() et al.
80 */
81 unsigned long events[MAX_HWEVENTS];
82
83 /* The current counter index assigned to an event. When the
84 * event hasn't been programmed into the cpu yet, this will
85 * hold PIC_NO_INDEX. The event->hw.idx value tells us where
86 * we ought to schedule the event.
87 */
88 int current_idx[MAX_HWEVENTS];
89
90 /* Software copy of %pcr register on this cpu. */
91 u64 pcr;
92
93 /* Enabled/disable state. */
94 int enabled;
95
96 unsigned int group_flag;
97 };
98 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = { .enabled = 1, };
99
100 /* An event map describes the characteristics of a performance
101 * counter event. In particular it gives the encoding as well as
102 * a mask telling which counters the event can be measured on.
103 */
104 struct perf_event_map {
105 u16 encoding;
106 u8 pic_mask;
107 #define PIC_NONE 0x00
108 #define PIC_UPPER 0x01
109 #define PIC_LOWER 0x02
110 };
111
112 /* Encode a perf_event_map entry into a long. */
113 static unsigned long perf_event_encode(const struct perf_event_map *pmap)
114 {
115 return ((unsigned long) pmap->encoding << 16) | pmap->pic_mask;
116 }
117
118 static u8 perf_event_get_msk(unsigned long val)
119 {
120 return val & 0xff;
121 }
122
123 static u64 perf_event_get_enc(unsigned long val)
124 {
125 return val >> 16;
126 }
127
128 #define C(x) PERF_COUNT_HW_CACHE_##x
129
130 #define CACHE_OP_UNSUPPORTED 0xfffe
131 #define CACHE_OP_NONSENSE 0xffff
132
133 typedef struct perf_event_map cache_map_t
134 [PERF_COUNT_HW_CACHE_MAX]
135 [PERF_COUNT_HW_CACHE_OP_MAX]
136 [PERF_COUNT_HW_CACHE_RESULT_MAX];
137
138 struct sparc_pmu {
139 const struct perf_event_map *(*event_map)(int);
140 const cache_map_t *cache_map;
141 int max_events;
142 int upper_shift;
143 int lower_shift;
144 int event_mask;
145 int hv_bit;
146 int irq_bit;
147 int upper_nop;
148 int lower_nop;
149 };
150
151 static const struct perf_event_map ultra3_perfmon_event_map[] = {
152 [PERF_COUNT_HW_CPU_CYCLES] = { 0x0000, PIC_UPPER | PIC_LOWER },
153 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x0001, PIC_UPPER | PIC_LOWER },
154 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0009, PIC_LOWER },
155 [PERF_COUNT_HW_CACHE_MISSES] = { 0x0009, PIC_UPPER },
156 };
157
158 static const struct perf_event_map *ultra3_event_map(int event_id)
159 {
160 return &ultra3_perfmon_event_map[event_id];
161 }
162
163 static const cache_map_t ultra3_cache_map = {
164 [C(L1D)] = {
165 [C(OP_READ)] = {
166 [C(RESULT_ACCESS)] = { 0x09, PIC_LOWER, },
167 [C(RESULT_MISS)] = { 0x09, PIC_UPPER, },
168 },
169 [C(OP_WRITE)] = {
170 [C(RESULT_ACCESS)] = { 0x0a, PIC_LOWER },
171 [C(RESULT_MISS)] = { 0x0a, PIC_UPPER },
172 },
173 [C(OP_PREFETCH)] = {
174 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
175 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
176 },
177 },
178 [C(L1I)] = {
179 [C(OP_READ)] = {
180 [C(RESULT_ACCESS)] = { 0x09, PIC_LOWER, },
181 [C(RESULT_MISS)] = { 0x09, PIC_UPPER, },
182 },
183 [ C(OP_WRITE) ] = {
184 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
185 [ C(RESULT_MISS) ] = { CACHE_OP_NONSENSE },
186 },
187 [ C(OP_PREFETCH) ] = {
188 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
189 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
190 },
191 },
192 [C(LL)] = {
193 [C(OP_READ)] = {
194 [C(RESULT_ACCESS)] = { 0x0c, PIC_LOWER, },
195 [C(RESULT_MISS)] = { 0x0c, PIC_UPPER, },
196 },
197 [C(OP_WRITE)] = {
198 [C(RESULT_ACCESS)] = { 0x0c, PIC_LOWER },
199 [C(RESULT_MISS)] = { 0x0c, PIC_UPPER },
200 },
201 [C(OP_PREFETCH)] = {
202 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
203 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
204 },
205 },
206 [C(DTLB)] = {
207 [C(OP_READ)] = {
208 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
209 [C(RESULT_MISS)] = { 0x12, PIC_UPPER, },
210 },
211 [ C(OP_WRITE) ] = {
212 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
213 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
214 },
215 [ C(OP_PREFETCH) ] = {
216 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
217 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
218 },
219 },
220 [C(ITLB)] = {
221 [C(OP_READ)] = {
222 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
223 [C(RESULT_MISS)] = { 0x11, PIC_UPPER, },
224 },
225 [ C(OP_WRITE) ] = {
226 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
227 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
228 },
229 [ C(OP_PREFETCH) ] = {
230 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
231 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
232 },
233 },
234 [C(BPU)] = {
235 [C(OP_READ)] = {
236 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
237 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
238 },
239 [ C(OP_WRITE) ] = {
240 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
241 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
242 },
243 [ C(OP_PREFETCH) ] = {
244 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
245 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
246 },
247 },
248 };
249
250 static const struct sparc_pmu ultra3_pmu = {
251 .event_map = ultra3_event_map,
252 .cache_map = &ultra3_cache_map,
253 .max_events = ARRAY_SIZE(ultra3_perfmon_event_map),
254 .upper_shift = 11,
255 .lower_shift = 4,
256 .event_mask = 0x3f,
257 .upper_nop = 0x1c,
258 .lower_nop = 0x14,
259 };
260
261 /* Niagara1 is very limited. The upper PIC is hard-locked to count
262 * only instructions, so it is free running which creates all kinds of
263 * problems. Some hardware designs make one wonder if the creator
264 * even looked at how this stuff gets used by software.
265 */
266 static const struct perf_event_map niagara1_perfmon_event_map[] = {
267 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, PIC_UPPER },
268 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x00, PIC_UPPER },
269 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0, PIC_NONE },
270 [PERF_COUNT_HW_CACHE_MISSES] = { 0x03, PIC_LOWER },
271 };
272
273 static const struct perf_event_map *niagara1_event_map(int event_id)
274 {
275 return &niagara1_perfmon_event_map[event_id];
276 }
277
278 static const cache_map_t niagara1_cache_map = {
279 [C(L1D)] = {
280 [C(OP_READ)] = {
281 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
282 [C(RESULT_MISS)] = { 0x03, PIC_LOWER, },
283 },
284 [C(OP_WRITE)] = {
285 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
286 [C(RESULT_MISS)] = { 0x03, PIC_LOWER, },
287 },
288 [C(OP_PREFETCH)] = {
289 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
290 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
291 },
292 },
293 [C(L1I)] = {
294 [C(OP_READ)] = {
295 [C(RESULT_ACCESS)] = { 0x00, PIC_UPPER },
296 [C(RESULT_MISS)] = { 0x02, PIC_LOWER, },
297 },
298 [ C(OP_WRITE) ] = {
299 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
300 [ C(RESULT_MISS) ] = { CACHE_OP_NONSENSE },
301 },
302 [ C(OP_PREFETCH) ] = {
303 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
304 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
305 },
306 },
307 [C(LL)] = {
308 [C(OP_READ)] = {
309 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
310 [C(RESULT_MISS)] = { 0x07, PIC_LOWER, },
311 },
312 [C(OP_WRITE)] = {
313 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
314 [C(RESULT_MISS)] = { 0x07, PIC_LOWER, },
315 },
316 [C(OP_PREFETCH)] = {
317 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
318 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
319 },
320 },
321 [C(DTLB)] = {
322 [C(OP_READ)] = {
323 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
324 [C(RESULT_MISS)] = { 0x05, PIC_LOWER, },
325 },
326 [ C(OP_WRITE) ] = {
327 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
328 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
329 },
330 [ C(OP_PREFETCH) ] = {
331 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
332 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
333 },
334 },
335 [C(ITLB)] = {
336 [C(OP_READ)] = {
337 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
338 [C(RESULT_MISS)] = { 0x04, PIC_LOWER, },
339 },
340 [ C(OP_WRITE) ] = {
341 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
342 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
343 },
344 [ C(OP_PREFETCH) ] = {
345 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
346 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
347 },
348 },
349 [C(BPU)] = {
350 [C(OP_READ)] = {
351 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
352 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
353 },
354 [ C(OP_WRITE) ] = {
355 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
356 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
357 },
358 [ C(OP_PREFETCH) ] = {
359 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
360 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
361 },
362 },
363 };
364
365 static const struct sparc_pmu niagara1_pmu = {
366 .event_map = niagara1_event_map,
367 .cache_map = &niagara1_cache_map,
368 .max_events = ARRAY_SIZE(niagara1_perfmon_event_map),
369 .upper_shift = 0,
370 .lower_shift = 4,
371 .event_mask = 0x7,
372 .upper_nop = 0x0,
373 .lower_nop = 0x0,
374 };
375
376 static const struct perf_event_map niagara2_perfmon_event_map[] = {
377 [PERF_COUNT_HW_CPU_CYCLES] = { 0x02ff, PIC_UPPER | PIC_LOWER },
378 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x02ff, PIC_UPPER | PIC_LOWER },
379 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0208, PIC_UPPER | PIC_LOWER },
380 [PERF_COUNT_HW_CACHE_MISSES] = { 0x0302, PIC_UPPER | PIC_LOWER },
381 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x0201, PIC_UPPER | PIC_LOWER },
382 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x0202, PIC_UPPER | PIC_LOWER },
383 };
384
385 static const struct perf_event_map *niagara2_event_map(int event_id)
386 {
387 return &niagara2_perfmon_event_map[event_id];
388 }
389
390 static const cache_map_t niagara2_cache_map = {
391 [C(L1D)] = {
392 [C(OP_READ)] = {
393 [C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, },
394 [C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, },
395 },
396 [C(OP_WRITE)] = {
397 [C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, },
398 [C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, },
399 },
400 [C(OP_PREFETCH)] = {
401 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
402 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
403 },
404 },
405 [C(L1I)] = {
406 [C(OP_READ)] = {
407 [C(RESULT_ACCESS)] = { 0x02ff, PIC_UPPER | PIC_LOWER, },
408 [C(RESULT_MISS)] = { 0x0301, PIC_UPPER | PIC_LOWER, },
409 },
410 [ C(OP_WRITE) ] = {
411 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
412 [ C(RESULT_MISS) ] = { CACHE_OP_NONSENSE },
413 },
414 [ C(OP_PREFETCH) ] = {
415 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
416 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
417 },
418 },
419 [C(LL)] = {
420 [C(OP_READ)] = {
421 [C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, },
422 [C(RESULT_MISS)] = { 0x0330, PIC_UPPER | PIC_LOWER, },
423 },
424 [C(OP_WRITE)] = {
425 [C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, },
426 [C(RESULT_MISS)] = { 0x0320, PIC_UPPER | PIC_LOWER, },
427 },
428 [C(OP_PREFETCH)] = {
429 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
430 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
431 },
432 },
433 [C(DTLB)] = {
434 [C(OP_READ)] = {
435 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
436 [C(RESULT_MISS)] = { 0x0b08, PIC_UPPER | PIC_LOWER, },
437 },
438 [ C(OP_WRITE) ] = {
439 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
440 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
441 },
442 [ C(OP_PREFETCH) ] = {
443 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
444 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
445 },
446 },
447 [C(ITLB)] = {
448 [C(OP_READ)] = {
449 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
450 [C(RESULT_MISS)] = { 0xb04, PIC_UPPER | PIC_LOWER, },
451 },
452 [ C(OP_WRITE) ] = {
453 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
454 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
455 },
456 [ C(OP_PREFETCH) ] = {
457 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
458 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
459 },
460 },
461 [C(BPU)] = {
462 [C(OP_READ)] = {
463 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
464 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
465 },
466 [ C(OP_WRITE) ] = {
467 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
468 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
469 },
470 [ C(OP_PREFETCH) ] = {
471 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
472 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
473 },
474 },
475 };
476
477 static const struct sparc_pmu niagara2_pmu = {
478 .event_map = niagara2_event_map,
479 .cache_map = &niagara2_cache_map,
480 .max_events = ARRAY_SIZE(niagara2_perfmon_event_map),
481 .upper_shift = 19,
482 .lower_shift = 6,
483 .event_mask = 0xfff,
484 .hv_bit = 0x8,
485 .irq_bit = 0x30,
486 .upper_nop = 0x220,
487 .lower_nop = 0x220,
488 };
489
490 static const struct sparc_pmu *sparc_pmu __read_mostly;
491
492 static u64 event_encoding(u64 event_id, int idx)
493 {
494 if (idx == PIC_UPPER_INDEX)
495 event_id <<= sparc_pmu->upper_shift;
496 else
497 event_id <<= sparc_pmu->lower_shift;
498 return event_id;
499 }
500
501 static u64 mask_for_index(int idx)
502 {
503 return event_encoding(sparc_pmu->event_mask, idx);
504 }
505
506 static u64 nop_for_index(int idx)
507 {
508 return event_encoding(idx == PIC_UPPER_INDEX ?
509 sparc_pmu->upper_nop :
510 sparc_pmu->lower_nop, idx);
511 }
512
513 static inline void sparc_pmu_enable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx)
514 {
515 u64 val, mask = mask_for_index(idx);
516
517 val = cpuc->pcr;
518 val &= ~mask;
519 val |= hwc->config;
520 cpuc->pcr = val;
521
522 pcr_ops->write(cpuc->pcr);
523 }
524
525 static inline void sparc_pmu_disable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx)
526 {
527 u64 mask = mask_for_index(idx);
528 u64 nop = nop_for_index(idx);
529 u64 val;
530
531 val = cpuc->pcr;
532 val &= ~mask;
533 val |= nop;
534 cpuc->pcr = val;
535
536 pcr_ops->write(cpuc->pcr);
537 }
538
539 static u32 read_pmc(int idx)
540 {
541 u64 val;
542
543 read_pic(val);
544 if (idx == PIC_UPPER_INDEX)
545 val >>= 32;
546
547 return val & 0xffffffff;
548 }
549
550 static void write_pmc(int idx, u64 val)
551 {
552 u64 shift, mask, pic;
553
554 shift = 0;
555 if (idx == PIC_UPPER_INDEX)
556 shift = 32;
557
558 mask = ((u64) 0xffffffff) << shift;
559 val <<= shift;
560
561 read_pic(pic);
562 pic &= ~mask;
563 pic |= val;
564 write_pic(pic);
565 }
566
567 static u64 sparc_perf_event_update(struct perf_event *event,
568 struct hw_perf_event *hwc, int idx)
569 {
570 int shift = 64 - 32;
571 u64 prev_raw_count, new_raw_count;
572 s64 delta;
573
574 again:
575 prev_raw_count = atomic64_read(&hwc->prev_count);
576 new_raw_count = read_pmc(idx);
577
578 if (atomic64_cmpxchg(&hwc->prev_count, prev_raw_count,
579 new_raw_count) != prev_raw_count)
580 goto again;
581
582 delta = (new_raw_count << shift) - (prev_raw_count << shift);
583 delta >>= shift;
584
585 atomic64_add(delta, &event->count);
586 atomic64_sub(delta, &hwc->period_left);
587
588 return new_raw_count;
589 }
590
591 static int sparc_perf_event_set_period(struct perf_event *event,
592 struct hw_perf_event *hwc, int idx)
593 {
594 s64 left = atomic64_read(&hwc->period_left);
595 s64 period = hwc->sample_period;
596 int ret = 0;
597
598 if (unlikely(left <= -period)) {
599 left = period;
600 atomic64_set(&hwc->period_left, left);
601 hwc->last_period = period;
602 ret = 1;
603 }
604
605 if (unlikely(left <= 0)) {
606 left += period;
607 atomic64_set(&hwc->period_left, left);
608 hwc->last_period = period;
609 ret = 1;
610 }
611 if (left > MAX_PERIOD)
612 left = MAX_PERIOD;
613
614 atomic64_set(&hwc->prev_count, (u64)-left);
615
616 write_pmc(idx, (u64)(-left) & 0xffffffff);
617
618 perf_event_update_userpage(event);
619
620 return ret;
621 }
622
623 /* If performance event entries have been added, move existing
624 * events around (if necessary) and then assign new entries to
625 * counters.
626 */
627 static u64 maybe_change_configuration(struct cpu_hw_events *cpuc, u64 pcr)
628 {
629 int i;
630
631 if (!cpuc->n_added)
632 goto out;
633
634 /* Read in the counters which are moving. */
635 for (i = 0; i < cpuc->n_events; i++) {
636 struct perf_event *cp = cpuc->event[i];
637
638 if (cpuc->current_idx[i] != PIC_NO_INDEX &&
639 cpuc->current_idx[i] != cp->hw.idx) {
640 sparc_perf_event_update(cp, &cp->hw,
641 cpuc->current_idx[i]);
642 cpuc->current_idx[i] = PIC_NO_INDEX;
643 }
644 }
645
646 /* Assign to counters all unassigned events. */
647 for (i = 0; i < cpuc->n_events; i++) {
648 struct perf_event *cp = cpuc->event[i];
649 struct hw_perf_event *hwc = &cp->hw;
650 int idx = hwc->idx;
651 u64 enc;
652
653 if (cpuc->current_idx[i] != PIC_NO_INDEX)
654 continue;
655
656 sparc_perf_event_set_period(cp, hwc, idx);
657 cpuc->current_idx[i] = idx;
658
659 enc = perf_event_get_enc(cpuc->events[i]);
660 pcr |= event_encoding(enc, idx);
661 }
662 out:
663 return pcr;
664 }
665
666 void hw_perf_enable(void)
667 {
668 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
669 u64 pcr;
670
671 if (cpuc->enabled)
672 return;
673
674 cpuc->enabled = 1;
675 barrier();
676
677 pcr = cpuc->pcr;
678 if (!cpuc->n_events) {
679 pcr = 0;
680 } else {
681 pcr = maybe_change_configuration(cpuc, pcr);
682
683 /* We require that all of the events have the same
684 * configuration, so just fetch the settings from the
685 * first entry.
686 */
687 cpuc->pcr = pcr | cpuc->event[0]->hw.config_base;
688 }
689
690 pcr_ops->write(cpuc->pcr);
691 }
692
693 void hw_perf_disable(void)
694 {
695 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
696 u64 val;
697
698 if (!cpuc->enabled)
699 return;
700
701 cpuc->enabled = 0;
702 cpuc->n_added = 0;
703
704 val = cpuc->pcr;
705 val &= ~(PCR_UTRACE | PCR_STRACE |
706 sparc_pmu->hv_bit | sparc_pmu->irq_bit);
707 cpuc->pcr = val;
708
709 pcr_ops->write(cpuc->pcr);
710 }
711
712 static void sparc_pmu_disable(struct perf_event *event)
713 {
714 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
715 struct hw_perf_event *hwc = &event->hw;
716 unsigned long flags;
717 int i;
718
719 local_irq_save(flags);
720 perf_disable();
721
722 for (i = 0; i < cpuc->n_events; i++) {
723 if (event == cpuc->event[i]) {
724 int idx = cpuc->current_idx[i];
725
726 /* Shift remaining entries down into
727 * the existing slot.
728 */
729 while (++i < cpuc->n_events) {
730 cpuc->event[i - 1] = cpuc->event[i];
731 cpuc->events[i - 1] = cpuc->events[i];
732 cpuc->current_idx[i - 1] =
733 cpuc->current_idx[i];
734 }
735
736 /* Absorb the final count and turn off the
737 * event.
738 */
739 sparc_pmu_disable_event(cpuc, hwc, idx);
740 barrier();
741 sparc_perf_event_update(event, hwc, idx);
742
743 perf_event_update_userpage(event);
744
745 cpuc->n_events--;
746 break;
747 }
748 }
749
750 perf_enable();
751 local_irq_restore(flags);
752 }
753
754 static int active_event_index(struct cpu_hw_events *cpuc,
755 struct perf_event *event)
756 {
757 int i;
758
759 for (i = 0; i < cpuc->n_events; i++) {
760 if (cpuc->event[i] == event)
761 break;
762 }
763 BUG_ON(i == cpuc->n_events);
764 return cpuc->current_idx[i];
765 }
766
767 static void sparc_pmu_read(struct perf_event *event)
768 {
769 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
770 int idx = active_event_index(cpuc, event);
771 struct hw_perf_event *hwc = &event->hw;
772
773 sparc_perf_event_update(event, hwc, idx);
774 }
775
776 static void sparc_pmu_unthrottle(struct perf_event *event)
777 {
778 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
779 int idx = active_event_index(cpuc, event);
780 struct hw_perf_event *hwc = &event->hw;
781
782 sparc_pmu_enable_event(cpuc, hwc, idx);
783 }
784
785 static atomic_t active_events = ATOMIC_INIT(0);
786 static DEFINE_MUTEX(pmc_grab_mutex);
787
788 static void perf_stop_nmi_watchdog(void *unused)
789 {
790 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
791
792 stop_nmi_watchdog(NULL);
793 cpuc->pcr = pcr_ops->read();
794 }
795
796 void perf_event_grab_pmc(void)
797 {
798 if (atomic_inc_not_zero(&active_events))
799 return;
800
801 mutex_lock(&pmc_grab_mutex);
802 if (atomic_read(&active_events) == 0) {
803 if (atomic_read(&nmi_active) > 0) {
804 on_each_cpu(perf_stop_nmi_watchdog, NULL, 1);
805 BUG_ON(atomic_read(&nmi_active) != 0);
806 }
807 atomic_inc(&active_events);
808 }
809 mutex_unlock(&pmc_grab_mutex);
810 }
811
812 void perf_event_release_pmc(void)
813 {
814 if (atomic_dec_and_mutex_lock(&active_events, &pmc_grab_mutex)) {
815 if (atomic_read(&nmi_active) == 0)
816 on_each_cpu(start_nmi_watchdog, NULL, 1);
817 mutex_unlock(&pmc_grab_mutex);
818 }
819 }
820
821 static const struct perf_event_map *sparc_map_cache_event(u64 config)
822 {
823 unsigned int cache_type, cache_op, cache_result;
824 const struct perf_event_map *pmap;
825
826 if (!sparc_pmu->cache_map)
827 return ERR_PTR(-ENOENT);
828
829 cache_type = (config >> 0) & 0xff;
830 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
831 return ERR_PTR(-EINVAL);
832
833 cache_op = (config >> 8) & 0xff;
834 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
835 return ERR_PTR(-EINVAL);
836
837 cache_result = (config >> 16) & 0xff;
838 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
839 return ERR_PTR(-EINVAL);
840
841 pmap = &((*sparc_pmu->cache_map)[cache_type][cache_op][cache_result]);
842
843 if (pmap->encoding == CACHE_OP_UNSUPPORTED)
844 return ERR_PTR(-ENOENT);
845
846 if (pmap->encoding == CACHE_OP_NONSENSE)
847 return ERR_PTR(-EINVAL);
848
849 return pmap;
850 }
851
852 static void hw_perf_event_destroy(struct perf_event *event)
853 {
854 perf_event_release_pmc();
855 }
856
857 /* Make sure all events can be scheduled into the hardware at
858 * the same time. This is simplified by the fact that we only
859 * need to support 2 simultaneous HW events.
860 *
861 * As a side effect, the evts[]->hw.idx values will be assigned
862 * on success. These are pending indexes. When the events are
863 * actually programmed into the chip, these values will propagate
864 * to the per-cpu cpuc->current_idx[] slots, see the code in
865 * maybe_change_configuration() for details.
866 */
867 static int sparc_check_constraints(struct perf_event **evts,
868 unsigned long *events, int n_ev)
869 {
870 u8 msk0 = 0, msk1 = 0;
871 int idx0 = 0;
872
873 /* This case is possible when we are invoked from
874 * hw_perf_group_sched_in().
875 */
876 if (!n_ev)
877 return 0;
878
879 if (n_ev > perf_max_events)
880 return -1;
881
882 msk0 = perf_event_get_msk(events[0]);
883 if (n_ev == 1) {
884 if (msk0 & PIC_LOWER)
885 idx0 = 1;
886 goto success;
887 }
888 BUG_ON(n_ev != 2);
889 msk1 = perf_event_get_msk(events[1]);
890
891 /* If both events can go on any counter, OK. */
892 if (msk0 == (PIC_UPPER | PIC_LOWER) &&
893 msk1 == (PIC_UPPER | PIC_LOWER))
894 goto success;
895
896 /* If one event is limited to a specific counter,
897 * and the other can go on both, OK.
898 */
899 if ((msk0 == PIC_UPPER || msk0 == PIC_LOWER) &&
900 msk1 == (PIC_UPPER | PIC_LOWER)) {
901 if (msk0 & PIC_LOWER)
902 idx0 = 1;
903 goto success;
904 }
905
906 if ((msk1 == PIC_UPPER || msk1 == PIC_LOWER) &&
907 msk0 == (PIC_UPPER | PIC_LOWER)) {
908 if (msk1 & PIC_UPPER)
909 idx0 = 1;
910 goto success;
911 }
912
913 /* If the events are fixed to different counters, OK. */
914 if ((msk0 == PIC_UPPER && msk1 == PIC_LOWER) ||
915 (msk0 == PIC_LOWER && msk1 == PIC_UPPER)) {
916 if (msk0 & PIC_LOWER)
917 idx0 = 1;
918 goto success;
919 }
920
921 /* Otherwise, there is a conflict. */
922 return -1;
923
924 success:
925 evts[0]->hw.idx = idx0;
926 if (n_ev == 2)
927 evts[1]->hw.idx = idx0 ^ 1;
928 return 0;
929 }
930
931 static int check_excludes(struct perf_event **evts, int n_prev, int n_new)
932 {
933 int eu = 0, ek = 0, eh = 0;
934 struct perf_event *event;
935 int i, n, first;
936
937 n = n_prev + n_new;
938 if (n <= 1)
939 return 0;
940
941 first = 1;
942 for (i = 0; i < n; i++) {
943 event = evts[i];
944 if (first) {
945 eu = event->attr.exclude_user;
946 ek = event->attr.exclude_kernel;
947 eh = event->attr.exclude_hv;
948 first = 0;
949 } else if (event->attr.exclude_user != eu ||
950 event->attr.exclude_kernel != ek ||
951 event->attr.exclude_hv != eh) {
952 return -EAGAIN;
953 }
954 }
955
956 return 0;
957 }
958
959 static int collect_events(struct perf_event *group, int max_count,
960 struct perf_event *evts[], unsigned long *events,
961 int *current_idx)
962 {
963 struct perf_event *event;
964 int n = 0;
965
966 if (!is_software_event(group)) {
967 if (n >= max_count)
968 return -1;
969 evts[n] = group;
970 events[n] = group->hw.event_base;
971 current_idx[n++] = PIC_NO_INDEX;
972 }
973 list_for_each_entry(event, &group->sibling_list, group_entry) {
974 if (!is_software_event(event) &&
975 event->state != PERF_EVENT_STATE_OFF) {
976 if (n >= max_count)
977 return -1;
978 evts[n] = event;
979 events[n] = event->hw.event_base;
980 current_idx[n++] = PIC_NO_INDEX;
981 }
982 }
983 return n;
984 }
985
986 static int sparc_pmu_enable(struct perf_event *event)
987 {
988 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
989 int n0, ret = -EAGAIN;
990 unsigned long flags;
991
992 local_irq_save(flags);
993 perf_disable();
994
995 n0 = cpuc->n_events;
996 if (n0 >= perf_max_events)
997 goto out;
998
999 cpuc->event[n0] = event;
1000 cpuc->events[n0] = event->hw.event_base;
1001 cpuc->current_idx[n0] = PIC_NO_INDEX;
1002
1003 /*
1004 * If group events scheduling transaction was started,
1005 * skip the schedulability test here, it will be peformed
1006 * at commit time(->commit_txn) as a whole
1007 */
1008 if (cpuc->group_flag & PERF_EVENT_TXN_STARTED)
1009 goto nocheck;
1010
1011 if (check_excludes(cpuc->event, n0, 1))
1012 goto out;
1013 if (sparc_check_constraints(cpuc->event, cpuc->events, n0 + 1))
1014 goto out;
1015
1016 nocheck:
1017 cpuc->n_events++;
1018 cpuc->n_added++;
1019
1020 ret = 0;
1021 out:
1022 perf_enable();
1023 local_irq_restore(flags);
1024 return ret;
1025 }
1026
1027 static int __hw_perf_event_init(struct perf_event *event)
1028 {
1029 struct perf_event_attr *attr = &event->attr;
1030 struct perf_event *evts[MAX_HWEVENTS];
1031 struct hw_perf_event *hwc = &event->hw;
1032 unsigned long events[MAX_HWEVENTS];
1033 int current_idx_dmy[MAX_HWEVENTS];
1034 const struct perf_event_map *pmap;
1035 int n;
1036
1037 if (atomic_read(&nmi_active) < 0)
1038 return -ENODEV;
1039
1040 if (attr->type == PERF_TYPE_HARDWARE) {
1041 if (attr->config >= sparc_pmu->max_events)
1042 return -EINVAL;
1043 pmap = sparc_pmu->event_map(attr->config);
1044 } else if (attr->type == PERF_TYPE_HW_CACHE) {
1045 pmap = sparc_map_cache_event(attr->config);
1046 if (IS_ERR(pmap))
1047 return PTR_ERR(pmap);
1048 } else
1049 return -EOPNOTSUPP;
1050
1051 /* We save the enable bits in the config_base. */
1052 hwc->config_base = sparc_pmu->irq_bit;
1053 if (!attr->exclude_user)
1054 hwc->config_base |= PCR_UTRACE;
1055 if (!attr->exclude_kernel)
1056 hwc->config_base |= PCR_STRACE;
1057 if (!attr->exclude_hv)
1058 hwc->config_base |= sparc_pmu->hv_bit;
1059
1060 hwc->event_base = perf_event_encode(pmap);
1061
1062 n = 0;
1063 if (event->group_leader != event) {
1064 n = collect_events(event->group_leader,
1065 perf_max_events - 1,
1066 evts, events, current_idx_dmy);
1067 if (n < 0)
1068 return -EINVAL;
1069 }
1070 events[n] = hwc->event_base;
1071 evts[n] = event;
1072
1073 if (check_excludes(evts, n, 1))
1074 return -EINVAL;
1075
1076 if (sparc_check_constraints(evts, events, n + 1))
1077 return -EINVAL;
1078
1079 hwc->idx = PIC_NO_INDEX;
1080
1081 /* Try to do all error checking before this point, as unwinding
1082 * state after grabbing the PMC is difficult.
1083 */
1084 perf_event_grab_pmc();
1085 event->destroy = hw_perf_event_destroy;
1086
1087 if (!hwc->sample_period) {
1088 hwc->sample_period = MAX_PERIOD;
1089 hwc->last_period = hwc->sample_period;
1090 atomic64_set(&hwc->period_left, hwc->sample_period);
1091 }
1092
1093 return 0;
1094 }
1095
1096 /*
1097 * Start group events scheduling transaction
1098 * Set the flag to make pmu::enable() not perform the
1099 * schedulability test, it will be performed at commit time
1100 */
1101 static void sparc_pmu_start_txn(const struct pmu *pmu)
1102 {
1103 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1104
1105 cpuhw->group_flag |= PERF_EVENT_TXN_STARTED;
1106 }
1107
1108 /*
1109 * Stop group events scheduling transaction
1110 * Clear the flag and pmu::enable() will perform the
1111 * schedulability test.
1112 */
1113 static void sparc_pmu_cancel_txn(const struct pmu *pmu)
1114 {
1115 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1116
1117 cpuhw->group_flag &= ~PERF_EVENT_TXN_STARTED;
1118 }
1119
1120 /*
1121 * Commit group events scheduling transaction
1122 * Perform the group schedulability test as a whole
1123 * Return 0 if success
1124 */
1125 static int sparc_pmu_commit_txn(const struct pmu *pmu)
1126 {
1127 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1128 int n;
1129
1130 if (!sparc_pmu)
1131 return -EINVAL;
1132
1133 cpuc = &__get_cpu_var(cpu_hw_events);
1134 n = cpuc->n_events;
1135 if (check_excludes(cpuc->event, 0, n))
1136 return -EINVAL;
1137 if (sparc_check_constraints(cpuc->event, cpuc->events, n))
1138 return -EAGAIN;
1139
1140 return 0;
1141 }
1142
1143 static const struct pmu pmu = {
1144 .enable = sparc_pmu_enable,
1145 .disable = sparc_pmu_disable,
1146 .read = sparc_pmu_read,
1147 .unthrottle = sparc_pmu_unthrottle,
1148 .start_txn = sparc_pmu_start_txn,
1149 .cancel_txn = sparc_pmu_cancel_txn,
1150 .commit_txn = sparc_pmu_commit_txn,
1151 };
1152
1153 const struct pmu *hw_perf_event_init(struct perf_event *event)
1154 {
1155 int err = __hw_perf_event_init(event);
1156
1157 if (err)
1158 return ERR_PTR(err);
1159 return &pmu;
1160 }
1161
1162 void perf_event_print_debug(void)
1163 {
1164 unsigned long flags;
1165 u64 pcr, pic;
1166 int cpu;
1167
1168 if (!sparc_pmu)
1169 return;
1170
1171 local_irq_save(flags);
1172
1173 cpu = smp_processor_id();
1174
1175 pcr = pcr_ops->read();
1176 read_pic(pic);
1177
1178 pr_info("\n");
1179 pr_info("CPU#%d: PCR[%016llx] PIC[%016llx]\n",
1180 cpu, pcr, pic);
1181
1182 local_irq_restore(flags);
1183 }
1184
1185 static int __kprobes perf_event_nmi_handler(struct notifier_block *self,
1186 unsigned long cmd, void *__args)
1187 {
1188 struct die_args *args = __args;
1189 struct perf_sample_data data;
1190 struct cpu_hw_events *cpuc;
1191 struct pt_regs *regs;
1192 int i;
1193
1194 if (!atomic_read(&active_events))
1195 return NOTIFY_DONE;
1196
1197 switch (cmd) {
1198 case DIE_NMI:
1199 break;
1200
1201 default:
1202 return NOTIFY_DONE;
1203 }
1204
1205 regs = args->regs;
1206
1207 perf_sample_data_init(&data, 0);
1208
1209 cpuc = &__get_cpu_var(cpu_hw_events);
1210
1211 /* If the PMU has the TOE IRQ enable bits, we need to do a
1212 * dummy write to the %pcr to clear the overflow bits and thus
1213 * the interrupt.
1214 *
1215 * Do this before we peek at the counters to determine
1216 * overflow so we don't lose any events.
1217 */
1218 if (sparc_pmu->irq_bit)
1219 pcr_ops->write(cpuc->pcr);
1220
1221 for (i = 0; i < cpuc->n_events; i++) {
1222 struct perf_event *event = cpuc->event[i];
1223 int idx = cpuc->current_idx[i];
1224 struct hw_perf_event *hwc;
1225 u64 val;
1226
1227 hwc = &event->hw;
1228 val = sparc_perf_event_update(event, hwc, idx);
1229 if (val & (1ULL << 31))
1230 continue;
1231
1232 data.period = event->hw.last_period;
1233 if (!sparc_perf_event_set_period(event, hwc, idx))
1234 continue;
1235
1236 if (perf_event_overflow(event, 1, &data, regs))
1237 sparc_pmu_disable_event(cpuc, hwc, idx);
1238 }
1239
1240 return NOTIFY_STOP;
1241 }
1242
1243 static __read_mostly struct notifier_block perf_event_nmi_notifier = {
1244 .notifier_call = perf_event_nmi_handler,
1245 };
1246
1247 static bool __init supported_pmu(void)
1248 {
1249 if (!strcmp(sparc_pmu_type, "ultra3") ||
1250 !strcmp(sparc_pmu_type, "ultra3+") ||
1251 !strcmp(sparc_pmu_type, "ultra3i") ||
1252 !strcmp(sparc_pmu_type, "ultra4+")) {
1253 sparc_pmu = &ultra3_pmu;
1254 return true;
1255 }
1256 if (!strcmp(sparc_pmu_type, "niagara")) {
1257 sparc_pmu = &niagara1_pmu;
1258 return true;
1259 }
1260 if (!strcmp(sparc_pmu_type, "niagara2")) {
1261 sparc_pmu = &niagara2_pmu;
1262 return true;
1263 }
1264 return false;
1265 }
1266
1267 void __init init_hw_perf_events(void)
1268 {
1269 pr_info("Performance events: ");
1270
1271 if (!supported_pmu()) {
1272 pr_cont("No support for PMU type '%s'\n", sparc_pmu_type);
1273 return;
1274 }
1275
1276 pr_cont("Supported PMU type is '%s'\n", sparc_pmu_type);
1277
1278 /* All sparc64 PMUs currently have 2 events. */
1279 perf_max_events = 2;
1280
1281 register_die_notifier(&perf_event_nmi_notifier);
1282 }
1283
1284 static inline void callchain_store(struct perf_callchain_entry *entry, u64 ip)
1285 {
1286 if (entry->nr < PERF_MAX_STACK_DEPTH)
1287 entry->ip[entry->nr++] = ip;
1288 }
1289
1290 static void perf_callchain_kernel(struct pt_regs *regs,
1291 struct perf_callchain_entry *entry)
1292 {
1293 unsigned long ksp, fp;
1294 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
1295 int graph = 0;
1296 #endif
1297
1298 callchain_store(entry, PERF_CONTEXT_KERNEL);
1299 callchain_store(entry, regs->tpc);
1300
1301 ksp = regs->u_regs[UREG_I6];
1302 fp = ksp + STACK_BIAS;
1303 do {
1304 struct sparc_stackf *sf;
1305 struct pt_regs *regs;
1306 unsigned long pc;
1307
1308 if (!kstack_valid(current_thread_info(), fp))
1309 break;
1310
1311 sf = (struct sparc_stackf *) fp;
1312 regs = (struct pt_regs *) (sf + 1);
1313
1314 if (kstack_is_trap_frame(current_thread_info(), regs)) {
1315 if (user_mode(regs))
1316 break;
1317 pc = regs->tpc;
1318 fp = regs->u_regs[UREG_I6] + STACK_BIAS;
1319 } else {
1320 pc = sf->callers_pc;
1321 fp = (unsigned long)sf->fp + STACK_BIAS;
1322 }
1323 callchain_store(entry, pc);
1324 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
1325 if ((pc + 8UL) == (unsigned long) &return_to_handler) {
1326 int index = current->curr_ret_stack;
1327 if (current->ret_stack && index >= graph) {
1328 pc = current->ret_stack[index - graph].ret;
1329 callchain_store(entry, pc);
1330 graph++;
1331 }
1332 }
1333 #endif
1334 } while (entry->nr < PERF_MAX_STACK_DEPTH);
1335 }
1336
1337 static void perf_callchain_user_64(struct pt_regs *regs,
1338 struct perf_callchain_entry *entry)
1339 {
1340 unsigned long ufp;
1341
1342 callchain_store(entry, PERF_CONTEXT_USER);
1343 callchain_store(entry, regs->tpc);
1344
1345 ufp = regs->u_regs[UREG_I6] + STACK_BIAS;
1346 do {
1347 struct sparc_stackf *usf, sf;
1348 unsigned long pc;
1349
1350 usf = (struct sparc_stackf *) ufp;
1351 if (__copy_from_user_inatomic(&sf, usf, sizeof(sf)))
1352 break;
1353
1354 pc = sf.callers_pc;
1355 ufp = (unsigned long)sf.fp + STACK_BIAS;
1356 callchain_store(entry, pc);
1357 } while (entry->nr < PERF_MAX_STACK_DEPTH);
1358 }
1359
1360 static void perf_callchain_user_32(struct pt_regs *regs,
1361 struct perf_callchain_entry *entry)
1362 {
1363 unsigned long ufp;
1364
1365 callchain_store(entry, PERF_CONTEXT_USER);
1366 callchain_store(entry, regs->tpc);
1367
1368 ufp = regs->u_regs[UREG_I6] & 0xffffffffUL;
1369 do {
1370 struct sparc_stackf32 *usf, sf;
1371 unsigned long pc;
1372
1373 usf = (struct sparc_stackf32 *) ufp;
1374 if (__copy_from_user_inatomic(&sf, usf, sizeof(sf)))
1375 break;
1376
1377 pc = sf.callers_pc;
1378 ufp = (unsigned long)sf.fp;
1379 callchain_store(entry, pc);
1380 } while (entry->nr < PERF_MAX_STACK_DEPTH);
1381 }
1382
1383 /* Like powerpc we can't get PMU interrupts within the PMU handler,
1384 * so no need for separate NMI and IRQ chains as on x86.
1385 */
1386 static DEFINE_PER_CPU(struct perf_callchain_entry, callchain);
1387
1388 struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1389 {
1390 struct perf_callchain_entry *entry = &__get_cpu_var(callchain);
1391
1392 entry->nr = 0;
1393 if (!user_mode(regs)) {
1394 stack_trace_flush();
1395 perf_callchain_kernel(regs, entry);
1396 if (current->mm)
1397 regs = task_pt_regs(current);
1398 else
1399 regs = NULL;
1400 }
1401 if (regs) {
1402 flushw_user();
1403 if (test_thread_flag(TIF_32BIT))
1404 perf_callchain_user_32(regs, entry);
1405 else
1406 perf_callchain_user_64(regs, entry);
1407 }
1408 return entry;
1409 }
Linux with added FPC-III support