Loading...
1// SPDX-License-Identifier: GPL-2.0
2/* Performance event support for sparc64.
3 *
4 * Copyright (C) 2009, 2010 David S. Miller <davem@davemloft.net>
5 *
6 * This code is based almost entirely upon the x86 perf event
7 * code, which is:
8 *
9 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
10 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
11 * Copyright (C) 2009 Jaswinder Singh Rajput
12 * Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
13 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra
14 */
15
16#include <linux/perf_event.h>
17#include <linux/kprobes.h>
18#include <linux/ftrace.h>
19#include <linux/kernel.h>
20#include <linux/kdebug.h>
21#include <linux/mutex.h>
22
23#include <asm/stacktrace.h>
24#include <asm/cpudata.h>
25#include <linux/uaccess.h>
26#include <linux/atomic.h>
27#include <asm/nmi.h>
28#include <asm/pcr.h>
29#include <asm/cacheflush.h>
30
31#include "kernel.h"
32#include "kstack.h"
33
34/* Two classes of sparc64 chips currently exist. All of which have
35 * 32-bit counters which can generate overflow interrupts on the
36 * transition from 0xffffffff to 0.
37 *
38 * All chips upto and including SPARC-T3 have two performance
39 * counters. The two 32-bit counters are accessed in one go using a
40 * single 64-bit register.
41 *
42 * On these older chips both counters are controlled using a single
43 * control register. The only way to stop all sampling is to clear
44 * all of the context (user, supervisor, hypervisor) sampling enable
45 * bits. But these bits apply to both counters, thus the two counters
46 * can't be enabled/disabled individually.
47 *
48 * Furthermore, the control register on these older chips have two
49 * event fields, one for each of the two counters. It's thus nearly
50 * impossible to have one counter going while keeping the other one
51 * stopped. Therefore it is possible to get overflow interrupts for
52 * counters not currently "in use" and that condition must be checked
53 * in the overflow interrupt handler.
54 *
55 * So we use a hack, in that we program inactive counters with the
56 * "sw_count0" and "sw_count1" events. These count how many times
57 * the instruction "sethi %hi(0xfc000), %g0" is executed. It's an
58 * unusual way to encode a NOP and therefore will not trigger in
59 * normal code.
60 *
61 * Starting with SPARC-T4 we have one control register per counter.
62 * And the counters are stored in individual registers. The registers
63 * for the counters are 64-bit but only a 32-bit counter is
64 * implemented. The event selections on SPARC-T4 lack any
65 * restrictions, therefore we can elide all of the complicated
66 * conflict resolution code we have for SPARC-T3 and earlier chips.
67 */
68
69#define MAX_HWEVENTS 4
70#define MAX_PCRS 4
71#define MAX_PERIOD ((1UL << 32) - 1)
72
73#define PIC_UPPER_INDEX 0
74#define PIC_LOWER_INDEX 1
75#define PIC_NO_INDEX -1
76
77struct cpu_hw_events {
78 /* Number of events currently scheduled onto this cpu.
79 * This tells how many entries in the arrays below
80 * are valid.
81 */
82 int n_events;
83
84 /* Number of new events added since the last hw_perf_disable().
85 * This works because the perf event layer always adds new
86 * events inside of a perf_{disable,enable}() sequence.
87 */
88 int n_added;
89
90 /* Array of events current scheduled on this cpu. */
91 struct perf_event *event[MAX_HWEVENTS];
92
93 /* Array of encoded longs, specifying the %pcr register
94 * encoding and the mask of PIC counters this even can
95 * be scheduled on. See perf_event_encode() et al.
96 */
97 unsigned long events[MAX_HWEVENTS];
98
99 /* The current counter index assigned to an event. When the
100 * event hasn't been programmed into the cpu yet, this will
101 * hold PIC_NO_INDEX. The event->hw.idx value tells us where
102 * we ought to schedule the event.
103 */
104 int current_idx[MAX_HWEVENTS];
105
106 /* Software copy of %pcr register(s) on this cpu. */
107 u64 pcr[MAX_HWEVENTS];
108
109 /* Enabled/disable state. */
110 int enabled;
111
112 unsigned int txn_flags;
113};
114static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = { .enabled = 1, };
115
116/* An event map describes the characteristics of a performance
117 * counter event. In particular it gives the encoding as well as
118 * a mask telling which counters the event can be measured on.
119 *
120 * The mask is unused on SPARC-T4 and later.
121 */
122struct perf_event_map {
123 u16 encoding;
124 u8 pic_mask;
125#define PIC_NONE 0x00
126#define PIC_UPPER 0x01
127#define PIC_LOWER 0x02
128};
129
130/* Encode a perf_event_map entry into a long. */
131static unsigned long perf_event_encode(const struct perf_event_map *pmap)
132{
133 return ((unsigned long) pmap->encoding << 16) | pmap->pic_mask;
134}
135
136static u8 perf_event_get_msk(unsigned long val)
137{
138 return val & 0xff;
139}
140
141static u64 perf_event_get_enc(unsigned long val)
142{
143 return val >> 16;
144}
145
146#define C(x) PERF_COUNT_HW_CACHE_##x
147
148#define CACHE_OP_UNSUPPORTED 0xfffe
149#define CACHE_OP_NONSENSE 0xffff
150
151typedef struct perf_event_map cache_map_t
152 [PERF_COUNT_HW_CACHE_MAX]
153 [PERF_COUNT_HW_CACHE_OP_MAX]
154 [PERF_COUNT_HW_CACHE_RESULT_MAX];
155
156struct sparc_pmu {
157 const struct perf_event_map *(*event_map)(int);
158 const cache_map_t *cache_map;
159 int max_events;
160 u32 (*read_pmc)(int);
161 void (*write_pmc)(int, u64);
162 int upper_shift;
163 int lower_shift;
164 int event_mask;
165 int user_bit;
166 int priv_bit;
167 int hv_bit;
168 int irq_bit;
169 int upper_nop;
170 int lower_nop;
171 unsigned int flags;
172#define SPARC_PMU_ALL_EXCLUDES_SAME 0x00000001
173#define SPARC_PMU_HAS_CONFLICTS 0x00000002
174 int max_hw_events;
175 int num_pcrs;
176 int num_pic_regs;
177};
178
179static u32 sparc_default_read_pmc(int idx)
180{
181 u64 val;
182
183 val = pcr_ops->read_pic(0);
184 if (idx == PIC_UPPER_INDEX)
185 val >>= 32;
186
187 return val & 0xffffffff;
188}
189
190static void sparc_default_write_pmc(int idx, u64 val)
191{
192 u64 shift, mask, pic;
193
194 shift = 0;
195 if (idx == PIC_UPPER_INDEX)
196 shift = 32;
197
198 mask = ((u64) 0xffffffff) << shift;
199 val <<= shift;
200
201 pic = pcr_ops->read_pic(0);
202 pic &= ~mask;
203 pic |= val;
204 pcr_ops->write_pic(0, pic);
205}
206
207static const struct perf_event_map ultra3_perfmon_event_map[] = {
208 [PERF_COUNT_HW_CPU_CYCLES] = { 0x0000, PIC_UPPER | PIC_LOWER },
209 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x0001, PIC_UPPER | PIC_LOWER },
210 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0009, PIC_LOWER },
211 [PERF_COUNT_HW_CACHE_MISSES] = { 0x0009, PIC_UPPER },
212};
213
214static const struct perf_event_map *ultra3_event_map(int event_id)
215{
216 return &ultra3_perfmon_event_map[event_id];
217}
218
219static const cache_map_t ultra3_cache_map = {
220[C(L1D)] = {
221 [C(OP_READ)] = {
222 [C(RESULT_ACCESS)] = { 0x09, PIC_LOWER, },
223 [C(RESULT_MISS)] = { 0x09, PIC_UPPER, },
224 },
225 [C(OP_WRITE)] = {
226 [C(RESULT_ACCESS)] = { 0x0a, PIC_LOWER },
227 [C(RESULT_MISS)] = { 0x0a, PIC_UPPER },
228 },
229 [C(OP_PREFETCH)] = {
230 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
231 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
232 },
233},
234[C(L1I)] = {
235 [C(OP_READ)] = {
236 [C(RESULT_ACCESS)] = { 0x09, PIC_LOWER, },
237 [C(RESULT_MISS)] = { 0x09, PIC_UPPER, },
238 },
239 [ C(OP_WRITE) ] = {
240 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
241 [ C(RESULT_MISS) ] = { CACHE_OP_NONSENSE },
242 },
243 [ C(OP_PREFETCH) ] = {
244 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
245 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
246 },
247},
248[C(LL)] = {
249 [C(OP_READ)] = {
250 [C(RESULT_ACCESS)] = { 0x0c, PIC_LOWER, },
251 [C(RESULT_MISS)] = { 0x0c, PIC_UPPER, },
252 },
253 [C(OP_WRITE)] = {
254 [C(RESULT_ACCESS)] = { 0x0c, PIC_LOWER },
255 [C(RESULT_MISS)] = { 0x0c, PIC_UPPER },
256 },
257 [C(OP_PREFETCH)] = {
258 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
259 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
260 },
261},
262[C(DTLB)] = {
263 [C(OP_READ)] = {
264 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
265 [C(RESULT_MISS)] = { 0x12, PIC_UPPER, },
266 },
267 [ C(OP_WRITE) ] = {
268 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
269 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
270 },
271 [ C(OP_PREFETCH) ] = {
272 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
273 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
274 },
275},
276[C(ITLB)] = {
277 [C(OP_READ)] = {
278 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
279 [C(RESULT_MISS)] = { 0x11, PIC_UPPER, },
280 },
281 [ C(OP_WRITE) ] = {
282 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
283 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
284 },
285 [ C(OP_PREFETCH) ] = {
286 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
287 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
288 },
289},
290[C(BPU)] = {
291 [C(OP_READ)] = {
292 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
293 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
294 },
295 [ C(OP_WRITE) ] = {
296 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
297 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
298 },
299 [ C(OP_PREFETCH) ] = {
300 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
301 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
302 },
303},
304[C(NODE)] = {
305 [C(OP_READ)] = {
306 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
307 [C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
308 },
309 [ C(OP_WRITE) ] = {
310 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
311 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
312 },
313 [ C(OP_PREFETCH) ] = {
314 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
315 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
316 },
317},
318};
319
320static const struct sparc_pmu ultra3_pmu = {
321 .event_map = ultra3_event_map,
322 .cache_map = &ultra3_cache_map,
323 .max_events = ARRAY_SIZE(ultra3_perfmon_event_map),
324 .read_pmc = sparc_default_read_pmc,
325 .write_pmc = sparc_default_write_pmc,
326 .upper_shift = 11,
327 .lower_shift = 4,
328 .event_mask = 0x3f,
329 .user_bit = PCR_UTRACE,
330 .priv_bit = PCR_STRACE,
331 .upper_nop = 0x1c,
332 .lower_nop = 0x14,
333 .flags = (SPARC_PMU_ALL_EXCLUDES_SAME |
334 SPARC_PMU_HAS_CONFLICTS),
335 .max_hw_events = 2,
336 .num_pcrs = 1,
337 .num_pic_regs = 1,
338};
339
340/* Niagara1 is very limited. The upper PIC is hard-locked to count
341 * only instructions, so it is free running which creates all kinds of
342 * problems. Some hardware designs make one wonder if the creator
343 * even looked at how this stuff gets used by software.
344 */
345static const struct perf_event_map niagara1_perfmon_event_map[] = {
346 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, PIC_UPPER },
347 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x00, PIC_UPPER },
348 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0, PIC_NONE },
349 [PERF_COUNT_HW_CACHE_MISSES] = { 0x03, PIC_LOWER },
350};
351
352static const struct perf_event_map *niagara1_event_map(int event_id)
353{
354 return &niagara1_perfmon_event_map[event_id];
355}
356
357static const cache_map_t niagara1_cache_map = {
358[C(L1D)] = {
359 [C(OP_READ)] = {
360 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
361 [C(RESULT_MISS)] = { 0x03, PIC_LOWER, },
362 },
363 [C(OP_WRITE)] = {
364 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
365 [C(RESULT_MISS)] = { 0x03, PIC_LOWER, },
366 },
367 [C(OP_PREFETCH)] = {
368 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
369 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
370 },
371},
372[C(L1I)] = {
373 [C(OP_READ)] = {
374 [C(RESULT_ACCESS)] = { 0x00, PIC_UPPER },
375 [C(RESULT_MISS)] = { 0x02, PIC_LOWER, },
376 },
377 [ C(OP_WRITE) ] = {
378 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
379 [ C(RESULT_MISS) ] = { CACHE_OP_NONSENSE },
380 },
381 [ C(OP_PREFETCH) ] = {
382 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
383 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
384 },
385},
386[C(LL)] = {
387 [C(OP_READ)] = {
388 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
389 [C(RESULT_MISS)] = { 0x07, PIC_LOWER, },
390 },
391 [C(OP_WRITE)] = {
392 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
393 [C(RESULT_MISS)] = { 0x07, PIC_LOWER, },
394 },
395 [C(OP_PREFETCH)] = {
396 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
397 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
398 },
399},
400[C(DTLB)] = {
401 [C(OP_READ)] = {
402 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
403 [C(RESULT_MISS)] = { 0x05, PIC_LOWER, },
404 },
405 [ C(OP_WRITE) ] = {
406 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
407 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
408 },
409 [ C(OP_PREFETCH) ] = {
410 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
411 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
412 },
413},
414[C(ITLB)] = {
415 [C(OP_READ)] = {
416 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
417 [C(RESULT_MISS)] = { 0x04, PIC_LOWER, },
418 },
419 [ C(OP_WRITE) ] = {
420 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
421 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
422 },
423 [ C(OP_PREFETCH) ] = {
424 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
425 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
426 },
427},
428[C(BPU)] = {
429 [C(OP_READ)] = {
430 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
431 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
432 },
433 [ C(OP_WRITE) ] = {
434 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
435 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
436 },
437 [ C(OP_PREFETCH) ] = {
438 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
439 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
440 },
441},
442[C(NODE)] = {
443 [C(OP_READ)] = {
444 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
445 [C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
446 },
447 [ C(OP_WRITE) ] = {
448 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
449 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
450 },
451 [ C(OP_PREFETCH) ] = {
452 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
453 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
454 },
455},
456};
457
458static const struct sparc_pmu niagara1_pmu = {
459 .event_map = niagara1_event_map,
460 .cache_map = &niagara1_cache_map,
461 .max_events = ARRAY_SIZE(niagara1_perfmon_event_map),
462 .read_pmc = sparc_default_read_pmc,
463 .write_pmc = sparc_default_write_pmc,
464 .upper_shift = 0,
465 .lower_shift = 4,
466 .event_mask = 0x7,
467 .user_bit = PCR_UTRACE,
468 .priv_bit = PCR_STRACE,
469 .upper_nop = 0x0,
470 .lower_nop = 0x0,
471 .flags = (SPARC_PMU_ALL_EXCLUDES_SAME |
472 SPARC_PMU_HAS_CONFLICTS),
473 .max_hw_events = 2,
474 .num_pcrs = 1,
475 .num_pic_regs = 1,
476};
477
478static const struct perf_event_map niagara2_perfmon_event_map[] = {
479 [PERF_COUNT_HW_CPU_CYCLES] = { 0x02ff, PIC_UPPER | PIC_LOWER },
480 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x02ff, PIC_UPPER | PIC_LOWER },
481 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0208, PIC_UPPER | PIC_LOWER },
482 [PERF_COUNT_HW_CACHE_MISSES] = { 0x0302, PIC_UPPER | PIC_LOWER },
483 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x0201, PIC_UPPER | PIC_LOWER },
484 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x0202, PIC_UPPER | PIC_LOWER },
485};
486
487static const struct perf_event_map *niagara2_event_map(int event_id)
488{
489 return &niagara2_perfmon_event_map[event_id];
490}
491
492static const cache_map_t niagara2_cache_map = {
493[C(L1D)] = {
494 [C(OP_READ)] = {
495 [C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, },
496 [C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, },
497 },
498 [C(OP_WRITE)] = {
499 [C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, },
500 [C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, },
501 },
502 [C(OP_PREFETCH)] = {
503 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
504 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
505 },
506},
507[C(L1I)] = {
508 [C(OP_READ)] = {
509 [C(RESULT_ACCESS)] = { 0x02ff, PIC_UPPER | PIC_LOWER, },
510 [C(RESULT_MISS)] = { 0x0301, PIC_UPPER | PIC_LOWER, },
511 },
512 [ C(OP_WRITE) ] = {
513 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
514 [ C(RESULT_MISS) ] = { CACHE_OP_NONSENSE },
515 },
516 [ C(OP_PREFETCH) ] = {
517 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
518 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
519 },
520},
521[C(LL)] = {
522 [C(OP_READ)] = {
523 [C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, },
524 [C(RESULT_MISS)] = { 0x0330, PIC_UPPER | PIC_LOWER, },
525 },
526 [C(OP_WRITE)] = {
527 [C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, },
528 [C(RESULT_MISS)] = { 0x0320, PIC_UPPER | PIC_LOWER, },
529 },
530 [C(OP_PREFETCH)] = {
531 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
532 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
533 },
534},
535[C(DTLB)] = {
536 [C(OP_READ)] = {
537 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
538 [C(RESULT_MISS)] = { 0x0b08, PIC_UPPER | PIC_LOWER, },
539 },
540 [ C(OP_WRITE) ] = {
541 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
542 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
543 },
544 [ C(OP_PREFETCH) ] = {
545 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
546 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
547 },
548},
549[C(ITLB)] = {
550 [C(OP_READ)] = {
551 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
552 [C(RESULT_MISS)] = { 0xb04, PIC_UPPER | PIC_LOWER, },
553 },
554 [ C(OP_WRITE) ] = {
555 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
556 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
557 },
558 [ C(OP_PREFETCH) ] = {
559 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
560 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
561 },
562},
563[C(BPU)] = {
564 [C(OP_READ)] = {
565 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
566 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
567 },
568 [ C(OP_WRITE) ] = {
569 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
570 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
571 },
572 [ C(OP_PREFETCH) ] = {
573 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
574 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
575 },
576},
577[C(NODE)] = {
578 [C(OP_READ)] = {
579 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
580 [C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
581 },
582 [ C(OP_WRITE) ] = {
583 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
584 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
585 },
586 [ C(OP_PREFETCH) ] = {
587 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
588 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
589 },
590},
591};
592
593static const struct sparc_pmu niagara2_pmu = {
594 .event_map = niagara2_event_map,
595 .cache_map = &niagara2_cache_map,
596 .max_events = ARRAY_SIZE(niagara2_perfmon_event_map),
597 .read_pmc = sparc_default_read_pmc,
598 .write_pmc = sparc_default_write_pmc,
599 .upper_shift = 19,
600 .lower_shift = 6,
601 .event_mask = 0xfff,
602 .user_bit = PCR_UTRACE,
603 .priv_bit = PCR_STRACE,
604 .hv_bit = PCR_N2_HTRACE,
605 .irq_bit = 0x30,
606 .upper_nop = 0x220,
607 .lower_nop = 0x220,
608 .flags = (SPARC_PMU_ALL_EXCLUDES_SAME |
609 SPARC_PMU_HAS_CONFLICTS),
610 .max_hw_events = 2,
611 .num_pcrs = 1,
612 .num_pic_regs = 1,
613};
614
615static const struct perf_event_map niagara4_perfmon_event_map[] = {
616 [PERF_COUNT_HW_CPU_CYCLES] = { (26 << 6) },
617 [PERF_COUNT_HW_INSTRUCTIONS] = { (3 << 6) | 0x3f },
618 [PERF_COUNT_HW_CACHE_REFERENCES] = { (3 << 6) | 0x04 },
619 [PERF_COUNT_HW_CACHE_MISSES] = { (16 << 6) | 0x07 },
620 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { (4 << 6) | 0x01 },
621 [PERF_COUNT_HW_BRANCH_MISSES] = { (25 << 6) | 0x0f },
622};
623
624static const struct perf_event_map *niagara4_event_map(int event_id)
625{
626 return &niagara4_perfmon_event_map[event_id];
627}
628
629static const cache_map_t niagara4_cache_map = {
630[C(L1D)] = {
631 [C(OP_READ)] = {
632 [C(RESULT_ACCESS)] = { (3 << 6) | 0x04 },
633 [C(RESULT_MISS)] = { (16 << 6) | 0x07 },
634 },
635 [C(OP_WRITE)] = {
636 [C(RESULT_ACCESS)] = { (3 << 6) | 0x08 },
637 [C(RESULT_MISS)] = { (16 << 6) | 0x07 },
638 },
639 [C(OP_PREFETCH)] = {
640 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
641 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
642 },
643},
644[C(L1I)] = {
645 [C(OP_READ)] = {
646 [C(RESULT_ACCESS)] = { (3 << 6) | 0x3f },
647 [C(RESULT_MISS)] = { (11 << 6) | 0x03 },
648 },
649 [ C(OP_WRITE) ] = {
650 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
651 [ C(RESULT_MISS) ] = { CACHE_OP_NONSENSE },
652 },
653 [ C(OP_PREFETCH) ] = {
654 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
655 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
656 },
657},
658[C(LL)] = {
659 [C(OP_READ)] = {
660 [C(RESULT_ACCESS)] = { (3 << 6) | 0x04 },
661 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
662 },
663 [C(OP_WRITE)] = {
664 [C(RESULT_ACCESS)] = { (3 << 6) | 0x08 },
665 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
666 },
667 [C(OP_PREFETCH)] = {
668 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
669 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
670 },
671},
672[C(DTLB)] = {
673 [C(OP_READ)] = {
674 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
675 [C(RESULT_MISS)] = { (17 << 6) | 0x3f },
676 },
677 [ C(OP_WRITE) ] = {
678 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
679 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
680 },
681 [ C(OP_PREFETCH) ] = {
682 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
683 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
684 },
685},
686[C(ITLB)] = {
687 [C(OP_READ)] = {
688 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
689 [C(RESULT_MISS)] = { (6 << 6) | 0x3f },
690 },
691 [ C(OP_WRITE) ] = {
692 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
693 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
694 },
695 [ C(OP_PREFETCH) ] = {
696 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
697 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
698 },
699},
700[C(BPU)] = {
701 [C(OP_READ)] = {
702 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
703 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
704 },
705 [ C(OP_WRITE) ] = {
706 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
707 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
708 },
709 [ C(OP_PREFETCH) ] = {
710 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
711 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
712 },
713},
714[C(NODE)] = {
715 [C(OP_READ)] = {
716 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
717 [C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
718 },
719 [ C(OP_WRITE) ] = {
720 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
721 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
722 },
723 [ C(OP_PREFETCH) ] = {
724 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
725 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
726 },
727},
728};
729
730static u32 sparc_vt_read_pmc(int idx)
731{
732 u64 val = pcr_ops->read_pic(idx);
733
734 return val & 0xffffffff;
735}
736
737static void sparc_vt_write_pmc(int idx, u64 val)
738{
739 u64 pcr;
740
741 pcr = pcr_ops->read_pcr(idx);
742 /* ensure ov and ntc are reset */
743 pcr &= ~(PCR_N4_OV | PCR_N4_NTC);
744
745 pcr_ops->write_pic(idx, val & 0xffffffff);
746
747 pcr_ops->write_pcr(idx, pcr);
748}
749
750static const struct sparc_pmu niagara4_pmu = {
751 .event_map = niagara4_event_map,
752 .cache_map = &niagara4_cache_map,
753 .max_events = ARRAY_SIZE(niagara4_perfmon_event_map),
754 .read_pmc = sparc_vt_read_pmc,
755 .write_pmc = sparc_vt_write_pmc,
756 .upper_shift = 5,
757 .lower_shift = 5,
758 .event_mask = 0x7ff,
759 .user_bit = PCR_N4_UTRACE,
760 .priv_bit = PCR_N4_STRACE,
761
762 /* We explicitly don't support hypervisor tracing. The T4
763 * generates the overflow event for precise events via a trap
764 * which will not be generated (ie. it's completely lost) if
765 * we happen to be in the hypervisor when the event triggers.
766 * Essentially, the overflow event reporting is completely
767 * unusable when you have hypervisor mode tracing enabled.
768 */
769 .hv_bit = 0,
770
771 .irq_bit = PCR_N4_TOE,
772 .upper_nop = 0,
773 .lower_nop = 0,
774 .flags = 0,
775 .max_hw_events = 4,
776 .num_pcrs = 4,
777 .num_pic_regs = 4,
778};
779
780static const struct sparc_pmu sparc_m7_pmu = {
781 .event_map = niagara4_event_map,
782 .cache_map = &niagara4_cache_map,
783 .max_events = ARRAY_SIZE(niagara4_perfmon_event_map),
784 .read_pmc = sparc_vt_read_pmc,
785 .write_pmc = sparc_vt_write_pmc,
786 .upper_shift = 5,
787 .lower_shift = 5,
788 .event_mask = 0x7ff,
789 .user_bit = PCR_N4_UTRACE,
790 .priv_bit = PCR_N4_STRACE,
791
792 /* We explicitly don't support hypervisor tracing. */
793 .hv_bit = 0,
794
795 .irq_bit = PCR_N4_TOE,
796 .upper_nop = 0,
797 .lower_nop = 0,
798 .flags = 0,
799 .max_hw_events = 4,
800 .num_pcrs = 4,
801 .num_pic_regs = 4,
802};
803static const struct sparc_pmu *sparc_pmu __read_mostly;
804
805static u64 event_encoding(u64 event_id, int idx)
806{
807 if (idx == PIC_UPPER_INDEX)
808 event_id <<= sparc_pmu->upper_shift;
809 else
810 event_id <<= sparc_pmu->lower_shift;
811 return event_id;
812}
813
814static u64 mask_for_index(int idx)
815{
816 return event_encoding(sparc_pmu->event_mask, idx);
817}
818
819static u64 nop_for_index(int idx)
820{
821 return event_encoding(idx == PIC_UPPER_INDEX ?
822 sparc_pmu->upper_nop :
823 sparc_pmu->lower_nop, idx);
824}
825
826static inline void sparc_pmu_enable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx)
827{
828 u64 enc, val, mask = mask_for_index(idx);
829 int pcr_index = 0;
830
831 if (sparc_pmu->num_pcrs > 1)
832 pcr_index = idx;
833
834 enc = perf_event_get_enc(cpuc->events[idx]);
835
836 val = cpuc->pcr[pcr_index];
837 val &= ~mask;
838 val |= event_encoding(enc, idx);
839 cpuc->pcr[pcr_index] = val;
840
841 pcr_ops->write_pcr(pcr_index, cpuc->pcr[pcr_index]);
842}
843
844static inline void sparc_pmu_disable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx)
845{
846 u64 mask = mask_for_index(idx);
847 u64 nop = nop_for_index(idx);
848 int pcr_index = 0;
849 u64 val;
850
851 if (sparc_pmu->num_pcrs > 1)
852 pcr_index = idx;
853
854 val = cpuc->pcr[pcr_index];
855 val &= ~mask;
856 val |= nop;
857 cpuc->pcr[pcr_index] = val;
858
859 pcr_ops->write_pcr(pcr_index, cpuc->pcr[pcr_index]);
860}
861
862static u64 sparc_perf_event_update(struct perf_event *event,
863 struct hw_perf_event *hwc, int idx)
864{
865 int shift = 64 - 32;
866 u64 prev_raw_count, new_raw_count;
867 s64 delta;
868
869again:
870 prev_raw_count = local64_read(&hwc->prev_count);
871 new_raw_count = sparc_pmu->read_pmc(idx);
872
873 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
874 new_raw_count) != prev_raw_count)
875 goto again;
876
877 delta = (new_raw_count << shift) - (prev_raw_count << shift);
878 delta >>= shift;
879
880 local64_add(delta, &event->count);
881 local64_sub(delta, &hwc->period_left);
882
883 return new_raw_count;
884}
885
886static int sparc_perf_event_set_period(struct perf_event *event,
887 struct hw_perf_event *hwc, int idx)
888{
889 s64 left = local64_read(&hwc->period_left);
890 s64 period = hwc->sample_period;
891 int ret = 0;
892
893 if (unlikely(left <= -period)) {
894 left = period;
895 local64_set(&hwc->period_left, left);
896 hwc->last_period = period;
897 ret = 1;
898 }
899
900 if (unlikely(left <= 0)) {
901 left += period;
902 local64_set(&hwc->period_left, left);
903 hwc->last_period = period;
904 ret = 1;
905 }
906 if (left > MAX_PERIOD)
907 left = MAX_PERIOD;
908
909 local64_set(&hwc->prev_count, (u64)-left);
910
911 sparc_pmu->write_pmc(idx, (u64)(-left) & 0xffffffff);
912
913 perf_event_update_userpage(event);
914
915 return ret;
916}
917
918static void read_in_all_counters(struct cpu_hw_events *cpuc)
919{
920 int i;
921
922 for (i = 0; i < cpuc->n_events; i++) {
923 struct perf_event *cp = cpuc->event[i];
924
925 if (cpuc->current_idx[i] != PIC_NO_INDEX &&
926 cpuc->current_idx[i] != cp->hw.idx) {
927 sparc_perf_event_update(cp, &cp->hw,
928 cpuc->current_idx[i]);
929 cpuc->current_idx[i] = PIC_NO_INDEX;
930 }
931 }
932}
933
934/* On this PMU all PICs are programmed using a single PCR. Calculate
935 * the combined control register value.
936 *
937 * For such chips we require that all of the events have the same
938 * configuration, so just fetch the settings from the first entry.
939 */
940static void calculate_single_pcr(struct cpu_hw_events *cpuc)
941{
942 int i;
943
944 if (!cpuc->n_added)
945 goto out;
946
947 /* Assign to counters all unassigned events. */
948 for (i = 0; i < cpuc->n_events; i++) {
949 struct perf_event *cp = cpuc->event[i];
950 struct hw_perf_event *hwc = &cp->hw;
951 int idx = hwc->idx;
952 u64 enc;
953
954 if (cpuc->current_idx[i] != PIC_NO_INDEX)
955 continue;
956
957 sparc_perf_event_set_period(cp, hwc, idx);
958 cpuc->current_idx[i] = idx;
959
960 enc = perf_event_get_enc(cpuc->events[i]);
961 cpuc->pcr[0] &= ~mask_for_index(idx);
962 if (hwc->state & PERF_HES_STOPPED)
963 cpuc->pcr[0] |= nop_for_index(idx);
964 else
965 cpuc->pcr[0] |= event_encoding(enc, idx);
966 }
967out:
968 cpuc->pcr[0] |= cpuc->event[0]->hw.config_base;
969}
970
971static void sparc_pmu_start(struct perf_event *event, int flags);
972
973/* On this PMU each PIC has it's own PCR control register. */
974static void calculate_multiple_pcrs(struct cpu_hw_events *cpuc)
975{
976 int i;
977
978 if (!cpuc->n_added)
979 goto out;
980
981 for (i = 0; i < cpuc->n_events; i++) {
982 struct perf_event *cp = cpuc->event[i];
983 struct hw_perf_event *hwc = &cp->hw;
984 int idx = hwc->idx;
985
986 if (cpuc->current_idx[i] != PIC_NO_INDEX)
987 continue;
988
989 cpuc->current_idx[i] = idx;
990
991 sparc_pmu_start(cp, PERF_EF_RELOAD);
992 }
993out:
994 for (i = 0; i < cpuc->n_events; i++) {
995 struct perf_event *cp = cpuc->event[i];
996 int idx = cp->hw.idx;
997
998 cpuc->pcr[idx] |= cp->hw.config_base;
999 }
1000}
1001
1002/* If performance event entries have been added, move existing events
1003 * around (if necessary) and then assign new entries to counters.
1004 */
1005static void update_pcrs_for_enable(struct cpu_hw_events *cpuc)
1006{
1007 if (cpuc->n_added)
1008 read_in_all_counters(cpuc);
1009
1010 if (sparc_pmu->num_pcrs == 1) {
1011 calculate_single_pcr(cpuc);
1012 } else {
1013 calculate_multiple_pcrs(cpuc);
1014 }
1015}
1016
1017static void sparc_pmu_enable(struct pmu *pmu)
1018{
1019 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1020 int i;
1021
1022 if (cpuc->enabled)
1023 return;
1024
1025 cpuc->enabled = 1;
1026 barrier();
1027
1028 if (cpuc->n_events)
1029 update_pcrs_for_enable(cpuc);
1030
1031 for (i = 0; i < sparc_pmu->num_pcrs; i++)
1032 pcr_ops->write_pcr(i, cpuc->pcr[i]);
1033}
1034
1035static void sparc_pmu_disable(struct pmu *pmu)
1036{
1037 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1038 int i;
1039
1040 if (!cpuc->enabled)
1041 return;
1042
1043 cpuc->enabled = 0;
1044 cpuc->n_added = 0;
1045
1046 for (i = 0; i < sparc_pmu->num_pcrs; i++) {
1047 u64 val = cpuc->pcr[i];
1048
1049 val &= ~(sparc_pmu->user_bit | sparc_pmu->priv_bit |
1050 sparc_pmu->hv_bit | sparc_pmu->irq_bit);
1051 cpuc->pcr[i] = val;
1052 pcr_ops->write_pcr(i, cpuc->pcr[i]);
1053 }
1054}
1055
1056static int active_event_index(struct cpu_hw_events *cpuc,
1057 struct perf_event *event)
1058{
1059 int i;
1060
1061 for (i = 0; i < cpuc->n_events; i++) {
1062 if (cpuc->event[i] == event)
1063 break;
1064 }
1065 BUG_ON(i == cpuc->n_events);
1066 return cpuc->current_idx[i];
1067}
1068
1069static void sparc_pmu_start(struct perf_event *event, int flags)
1070{
1071 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1072 int idx = active_event_index(cpuc, event);
1073
1074 if (flags & PERF_EF_RELOAD) {
1075 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1076 sparc_perf_event_set_period(event, &event->hw, idx);
1077 }
1078
1079 event->hw.state = 0;
1080
1081 sparc_pmu_enable_event(cpuc, &event->hw, idx);
1082}
1083
1084static void sparc_pmu_stop(struct perf_event *event, int flags)
1085{
1086 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1087 int idx = active_event_index(cpuc, event);
1088
1089 if (!(event->hw.state & PERF_HES_STOPPED)) {
1090 sparc_pmu_disable_event(cpuc, &event->hw, idx);
1091 event->hw.state |= PERF_HES_STOPPED;
1092 }
1093
1094 if (!(event->hw.state & PERF_HES_UPTODATE) && (flags & PERF_EF_UPDATE)) {
1095 sparc_perf_event_update(event, &event->hw, idx);
1096 event->hw.state |= PERF_HES_UPTODATE;
1097 }
1098}
1099
1100static void sparc_pmu_del(struct perf_event *event, int _flags)
1101{
1102 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1103 unsigned long flags;
1104 int i;
1105
1106 local_irq_save(flags);
1107
1108 for (i = 0; i < cpuc->n_events; i++) {
1109 if (event == cpuc->event[i]) {
1110 /* Absorb the final count and turn off the
1111 * event.
1112 */
1113 sparc_pmu_stop(event, PERF_EF_UPDATE);
1114
1115 /* Shift remaining entries down into
1116 * the existing slot.
1117 */
1118 while (++i < cpuc->n_events) {
1119 cpuc->event[i - 1] = cpuc->event[i];
1120 cpuc->events[i - 1] = cpuc->events[i];
1121 cpuc->current_idx[i - 1] =
1122 cpuc->current_idx[i];
1123 }
1124
1125 perf_event_update_userpage(event);
1126
1127 cpuc->n_events--;
1128 break;
1129 }
1130 }
1131
1132 local_irq_restore(flags);
1133}
1134
1135static void sparc_pmu_read(struct perf_event *event)
1136{
1137 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1138 int idx = active_event_index(cpuc, event);
1139 struct hw_perf_event *hwc = &event->hw;
1140
1141 sparc_perf_event_update(event, hwc, idx);
1142}
1143
1144static atomic_t active_events = ATOMIC_INIT(0);
1145static DEFINE_MUTEX(pmc_grab_mutex);
1146
1147static void perf_stop_nmi_watchdog(void *unused)
1148{
1149 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1150 int i;
1151
1152 stop_nmi_watchdog(NULL);
1153 for (i = 0; i < sparc_pmu->num_pcrs; i++)
1154 cpuc->pcr[i] = pcr_ops->read_pcr(i);
1155}
1156
1157static void perf_event_grab_pmc(void)
1158{
1159 if (atomic_inc_not_zero(&active_events))
1160 return;
1161
1162 mutex_lock(&pmc_grab_mutex);
1163 if (atomic_read(&active_events) == 0) {
1164 if (atomic_read(&nmi_active) > 0) {
1165 on_each_cpu(perf_stop_nmi_watchdog, NULL, 1);
1166 BUG_ON(atomic_read(&nmi_active) != 0);
1167 }
1168 atomic_inc(&active_events);
1169 }
1170 mutex_unlock(&pmc_grab_mutex);
1171}
1172
1173static void perf_event_release_pmc(void)
1174{
1175 if (atomic_dec_and_mutex_lock(&active_events, &pmc_grab_mutex)) {
1176 if (atomic_read(&nmi_active) == 0)
1177 on_each_cpu(start_nmi_watchdog, NULL, 1);
1178 mutex_unlock(&pmc_grab_mutex);
1179 }
1180}
1181
1182static const struct perf_event_map *sparc_map_cache_event(u64 config)
1183{
1184 unsigned int cache_type, cache_op, cache_result;
1185 const struct perf_event_map *pmap;
1186
1187 if (!sparc_pmu->cache_map)
1188 return ERR_PTR(-ENOENT);
1189
1190 cache_type = (config >> 0) & 0xff;
1191 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
1192 return ERR_PTR(-EINVAL);
1193
1194 cache_op = (config >> 8) & 0xff;
1195 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
1196 return ERR_PTR(-EINVAL);
1197
1198 cache_result = (config >> 16) & 0xff;
1199 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
1200 return ERR_PTR(-EINVAL);
1201
1202 pmap = &((*sparc_pmu->cache_map)[cache_type][cache_op][cache_result]);
1203
1204 if (pmap->encoding == CACHE_OP_UNSUPPORTED)
1205 return ERR_PTR(-ENOENT);
1206
1207 if (pmap->encoding == CACHE_OP_NONSENSE)
1208 return ERR_PTR(-EINVAL);
1209
1210 return pmap;
1211}
1212
1213static void hw_perf_event_destroy(struct perf_event *event)
1214{
1215 perf_event_release_pmc();
1216}
1217
1218/* Make sure all events can be scheduled into the hardware at
1219 * the same time. This is simplified by the fact that we only
1220 * need to support 2 simultaneous HW events.
1221 *
1222 * As a side effect, the evts[]->hw.idx values will be assigned
1223 * on success. These are pending indexes. When the events are
1224 * actually programmed into the chip, these values will propagate
1225 * to the per-cpu cpuc->current_idx[] slots, see the code in
1226 * maybe_change_configuration() for details.
1227 */
1228static int sparc_check_constraints(struct perf_event **evts,
1229 unsigned long *events, int n_ev)
1230{
1231 u8 msk0 = 0, msk1 = 0;
1232 int idx0 = 0;
1233
1234 /* This case is possible when we are invoked from
1235 * hw_perf_group_sched_in().
1236 */
1237 if (!n_ev)
1238 return 0;
1239
1240 if (n_ev > sparc_pmu->max_hw_events)
1241 return -1;
1242
1243 if (!(sparc_pmu->flags & SPARC_PMU_HAS_CONFLICTS)) {
1244 int i;
1245
1246 for (i = 0; i < n_ev; i++)
1247 evts[i]->hw.idx = i;
1248 return 0;
1249 }
1250
1251 msk0 = perf_event_get_msk(events[0]);
1252 if (n_ev == 1) {
1253 if (msk0 & PIC_LOWER)
1254 idx0 = 1;
1255 goto success;
1256 }
1257 BUG_ON(n_ev != 2);
1258 msk1 = perf_event_get_msk(events[1]);
1259
1260 /* If both events can go on any counter, OK. */
1261 if (msk0 == (PIC_UPPER | PIC_LOWER) &&
1262 msk1 == (PIC_UPPER | PIC_LOWER))
1263 goto success;
1264
1265 /* If one event is limited to a specific counter,
1266 * and the other can go on both, OK.
1267 */
1268 if ((msk0 == PIC_UPPER || msk0 == PIC_LOWER) &&
1269 msk1 == (PIC_UPPER | PIC_LOWER)) {
1270 if (msk0 & PIC_LOWER)
1271 idx0 = 1;
1272 goto success;
1273 }
1274
1275 if ((msk1 == PIC_UPPER || msk1 == PIC_LOWER) &&
1276 msk0 == (PIC_UPPER | PIC_LOWER)) {
1277 if (msk1 & PIC_UPPER)
1278 idx0 = 1;
1279 goto success;
1280 }
1281
1282 /* If the events are fixed to different counters, OK. */
1283 if ((msk0 == PIC_UPPER && msk1 == PIC_LOWER) ||
1284 (msk0 == PIC_LOWER && msk1 == PIC_UPPER)) {
1285 if (msk0 & PIC_LOWER)
1286 idx0 = 1;
1287 goto success;
1288 }
1289
1290 /* Otherwise, there is a conflict. */
1291 return -1;
1292
1293success:
1294 evts[0]->hw.idx = idx0;
1295 if (n_ev == 2)
1296 evts[1]->hw.idx = idx0 ^ 1;
1297 return 0;
1298}
1299
1300static int check_excludes(struct perf_event **evts, int n_prev, int n_new)
1301{
1302 int eu = 0, ek = 0, eh = 0;
1303 struct perf_event *event;
1304 int i, n, first;
1305
1306 if (!(sparc_pmu->flags & SPARC_PMU_ALL_EXCLUDES_SAME))
1307 return 0;
1308
1309 n = n_prev + n_new;
1310 if (n <= 1)
1311 return 0;
1312
1313 first = 1;
1314 for (i = 0; i < n; i++) {
1315 event = evts[i];
1316 if (first) {
1317 eu = event->attr.exclude_user;
1318 ek = event->attr.exclude_kernel;
1319 eh = event->attr.exclude_hv;
1320 first = 0;
1321 } else if (event->attr.exclude_user != eu ||
1322 event->attr.exclude_kernel != ek ||
1323 event->attr.exclude_hv != eh) {
1324 return -EAGAIN;
1325 }
1326 }
1327
1328 return 0;
1329}
1330
1331static int collect_events(struct perf_event *group, int max_count,
1332 struct perf_event *evts[], unsigned long *events,
1333 int *current_idx)
1334{
1335 struct perf_event *event;
1336 int n = 0;
1337
1338 if (!is_software_event(group)) {
1339 if (n >= max_count)
1340 return -1;
1341 evts[n] = group;
1342 events[n] = group->hw.event_base;
1343 current_idx[n++] = PIC_NO_INDEX;
1344 }
1345 for_each_sibling_event(event, group) {
1346 if (!is_software_event(event) &&
1347 event->state != PERF_EVENT_STATE_OFF) {
1348 if (n >= max_count)
1349 return -1;
1350 evts[n] = event;
1351 events[n] = event->hw.event_base;
1352 current_idx[n++] = PIC_NO_INDEX;
1353 }
1354 }
1355 return n;
1356}
1357
1358static int sparc_pmu_add(struct perf_event *event, int ef_flags)
1359{
1360 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1361 int n0, ret = -EAGAIN;
1362 unsigned long flags;
1363
1364 local_irq_save(flags);
1365
1366 n0 = cpuc->n_events;
1367 if (n0 >= sparc_pmu->max_hw_events)
1368 goto out;
1369
1370 cpuc->event[n0] = event;
1371 cpuc->events[n0] = event->hw.event_base;
1372 cpuc->current_idx[n0] = PIC_NO_INDEX;
1373
1374 event->hw.state = PERF_HES_UPTODATE;
1375 if (!(ef_flags & PERF_EF_START))
1376 event->hw.state |= PERF_HES_STOPPED;
1377
1378 /*
1379 * If group events scheduling transaction was started,
1380 * skip the schedulability test here, it will be performed
1381 * at commit time(->commit_txn) as a whole
1382 */
1383 if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1384 goto nocheck;
1385
1386 if (check_excludes(cpuc->event, n0, 1))
1387 goto out;
1388 if (sparc_check_constraints(cpuc->event, cpuc->events, n0 + 1))
1389 goto out;
1390
1391nocheck:
1392 cpuc->n_events++;
1393 cpuc->n_added++;
1394
1395 ret = 0;
1396out:
1397 local_irq_restore(flags);
1398 return ret;
1399}
1400
1401static int sparc_pmu_event_init(struct perf_event *event)
1402{
1403 struct perf_event_attr *attr = &event->attr;
1404 struct perf_event *evts[MAX_HWEVENTS];
1405 struct hw_perf_event *hwc = &event->hw;
1406 unsigned long events[MAX_HWEVENTS];
1407 int current_idx_dmy[MAX_HWEVENTS];
1408 const struct perf_event_map *pmap;
1409 int n;
1410
1411 if (atomic_read(&nmi_active) < 0)
1412 return -ENODEV;
1413
1414 /* does not support taken branch sampling */
1415 if (has_branch_stack(event))
1416 return -EOPNOTSUPP;
1417
1418 switch (attr->type) {
1419 case PERF_TYPE_HARDWARE:
1420 if (attr->config >= sparc_pmu->max_events)
1421 return -EINVAL;
1422 pmap = sparc_pmu->event_map(attr->config);
1423 break;
1424
1425 case PERF_TYPE_HW_CACHE:
1426 pmap = sparc_map_cache_event(attr->config);
1427 if (IS_ERR(pmap))
1428 return PTR_ERR(pmap);
1429 break;
1430
1431 case PERF_TYPE_RAW:
1432 pmap = NULL;
1433 break;
1434
1435 default:
1436 return -ENOENT;
1437
1438 }
1439
1440 if (pmap) {
1441 hwc->event_base = perf_event_encode(pmap);
1442 } else {
1443 /*
1444 * User gives us "(encoding << 16) | pic_mask" for
1445 * PERF_TYPE_RAW events.
1446 */
1447 hwc->event_base = attr->config;
1448 }
1449
1450 /* We save the enable bits in the config_base. */
1451 hwc->config_base = sparc_pmu->irq_bit;
1452 if (!attr->exclude_user)
1453 hwc->config_base |= sparc_pmu->user_bit;
1454 if (!attr->exclude_kernel)
1455 hwc->config_base |= sparc_pmu->priv_bit;
1456 if (!attr->exclude_hv)
1457 hwc->config_base |= sparc_pmu->hv_bit;
1458
1459 n = 0;
1460 if (event->group_leader != event) {
1461 n = collect_events(event->group_leader,
1462 sparc_pmu->max_hw_events - 1,
1463 evts, events, current_idx_dmy);
1464 if (n < 0)
1465 return -EINVAL;
1466 }
1467 events[n] = hwc->event_base;
1468 evts[n] = event;
1469
1470 if (check_excludes(evts, n, 1))
1471 return -EINVAL;
1472
1473 if (sparc_check_constraints(evts, events, n + 1))
1474 return -EINVAL;
1475
1476 hwc->idx = PIC_NO_INDEX;
1477
1478 /* Try to do all error checking before this point, as unwinding
1479 * state after grabbing the PMC is difficult.
1480 */
1481 perf_event_grab_pmc();
1482 event->destroy = hw_perf_event_destroy;
1483
1484 if (!hwc->sample_period) {
1485 hwc->sample_period = MAX_PERIOD;
1486 hwc->last_period = hwc->sample_period;
1487 local64_set(&hwc->period_left, hwc->sample_period);
1488 }
1489
1490 return 0;
1491}
1492
1493/*
1494 * Start group events scheduling transaction
1495 * Set the flag to make pmu::enable() not perform the
1496 * schedulability test, it will be performed at commit time
1497 */
1498static void sparc_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
1499{
1500 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1501
1502 WARN_ON_ONCE(cpuhw->txn_flags); /* txn already in flight */
1503
1504 cpuhw->txn_flags = txn_flags;
1505 if (txn_flags & ~PERF_PMU_TXN_ADD)
1506 return;
1507
1508 perf_pmu_disable(pmu);
1509}
1510
1511/*
1512 * Stop group events scheduling transaction
1513 * Clear the flag and pmu::enable() will perform the
1514 * schedulability test.
1515 */
1516static void sparc_pmu_cancel_txn(struct pmu *pmu)
1517{
1518 struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1519 unsigned int txn_flags;
1520
1521 WARN_ON_ONCE(!cpuhw->txn_flags); /* no txn in flight */
1522
1523 txn_flags = cpuhw->txn_flags;
1524 cpuhw->txn_flags = 0;
1525 if (txn_flags & ~PERF_PMU_TXN_ADD)
1526 return;
1527
1528 perf_pmu_enable(pmu);
1529}
1530
1531/*
1532 * Commit group events scheduling transaction
1533 * Perform the group schedulability test as a whole
1534 * Return 0 if success
1535 */
1536static int sparc_pmu_commit_txn(struct pmu *pmu)
1537{
1538 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1539 int n;
1540
1541 if (!sparc_pmu)
1542 return -EINVAL;
1543
1544 WARN_ON_ONCE(!cpuc->txn_flags); /* no txn in flight */
1545
1546 if (cpuc->txn_flags & ~PERF_PMU_TXN_ADD) {
1547 cpuc->txn_flags = 0;
1548 return 0;
1549 }
1550
1551 n = cpuc->n_events;
1552 if (check_excludes(cpuc->event, 0, n))
1553 return -EINVAL;
1554 if (sparc_check_constraints(cpuc->event, cpuc->events, n))
1555 return -EAGAIN;
1556
1557 cpuc->txn_flags = 0;
1558 perf_pmu_enable(pmu);
1559 return 0;
1560}
1561
1562static struct pmu pmu = {
1563 .pmu_enable = sparc_pmu_enable,
1564 .pmu_disable = sparc_pmu_disable,
1565 .event_init = sparc_pmu_event_init,
1566 .add = sparc_pmu_add,
1567 .del = sparc_pmu_del,
1568 .start = sparc_pmu_start,
1569 .stop = sparc_pmu_stop,
1570 .read = sparc_pmu_read,
1571 .start_txn = sparc_pmu_start_txn,
1572 .cancel_txn = sparc_pmu_cancel_txn,
1573 .commit_txn = sparc_pmu_commit_txn,
1574};
1575
1576void perf_event_print_debug(void)
1577{
1578 unsigned long flags;
1579 int cpu, i;
1580
1581 if (!sparc_pmu)
1582 return;
1583
1584 local_irq_save(flags);
1585
1586 cpu = smp_processor_id();
1587
1588 pr_info("\n");
1589 for (i = 0; i < sparc_pmu->num_pcrs; i++)
1590 pr_info("CPU#%d: PCR%d[%016llx]\n",
1591 cpu, i, pcr_ops->read_pcr(i));
1592 for (i = 0; i < sparc_pmu->num_pic_regs; i++)
1593 pr_info("CPU#%d: PIC%d[%016llx]\n",
1594 cpu, i, pcr_ops->read_pic(i));
1595
1596 local_irq_restore(flags);
1597}
1598
1599static int __kprobes perf_event_nmi_handler(struct notifier_block *self,
1600 unsigned long cmd, void *__args)
1601{
1602 struct die_args *args = __args;
1603 struct perf_sample_data data;
1604 struct cpu_hw_events *cpuc;
1605 struct pt_regs *regs;
1606 int i;
1607
1608 if (!atomic_read(&active_events))
1609 return NOTIFY_DONE;
1610
1611 switch (cmd) {
1612 case DIE_NMI:
1613 break;
1614
1615 default:
1616 return NOTIFY_DONE;
1617 }
1618
1619 regs = args->regs;
1620
1621 cpuc = this_cpu_ptr(&cpu_hw_events);
1622
1623 /* If the PMU has the TOE IRQ enable bits, we need to do a
1624 * dummy write to the %pcr to clear the overflow bits and thus
1625 * the interrupt.
1626 *
1627 * Do this before we peek at the counters to determine
1628 * overflow so we don't lose any events.
1629 */
1630 if (sparc_pmu->irq_bit &&
1631 sparc_pmu->num_pcrs == 1)
1632 pcr_ops->write_pcr(0, cpuc->pcr[0]);
1633
1634 for (i = 0; i < cpuc->n_events; i++) {
1635 struct perf_event *event = cpuc->event[i];
1636 int idx = cpuc->current_idx[i];
1637 struct hw_perf_event *hwc;
1638 u64 val;
1639
1640 if (sparc_pmu->irq_bit &&
1641 sparc_pmu->num_pcrs > 1)
1642 pcr_ops->write_pcr(idx, cpuc->pcr[idx]);
1643
1644 hwc = &event->hw;
1645 val = sparc_perf_event_update(event, hwc, idx);
1646 if (val & (1ULL << 31))
1647 continue;
1648
1649 perf_sample_data_init(&data, 0, hwc->last_period);
1650 if (!sparc_perf_event_set_period(event, hwc, idx))
1651 continue;
1652
1653 if (perf_event_overflow(event, &data, regs))
1654 sparc_pmu_stop(event, 0);
1655 }
1656
1657 return NOTIFY_STOP;
1658}
1659
1660static __read_mostly struct notifier_block perf_event_nmi_notifier = {
1661 .notifier_call = perf_event_nmi_handler,
1662};
1663
1664static bool __init supported_pmu(void)
1665{
1666 if (!strcmp(sparc_pmu_type, "ultra3") ||
1667 !strcmp(sparc_pmu_type, "ultra3+") ||
1668 !strcmp(sparc_pmu_type, "ultra3i") ||
1669 !strcmp(sparc_pmu_type, "ultra4+")) {
1670 sparc_pmu = &ultra3_pmu;
1671 return true;
1672 }
1673 if (!strcmp(sparc_pmu_type, "niagara")) {
1674 sparc_pmu = &niagara1_pmu;
1675 return true;
1676 }
1677 if (!strcmp(sparc_pmu_type, "niagara2") ||
1678 !strcmp(sparc_pmu_type, "niagara3")) {
1679 sparc_pmu = &niagara2_pmu;
1680 return true;
1681 }
1682 if (!strcmp(sparc_pmu_type, "niagara4") ||
1683 !strcmp(sparc_pmu_type, "niagara5")) {
1684 sparc_pmu = &niagara4_pmu;
1685 return true;
1686 }
1687 if (!strcmp(sparc_pmu_type, "sparc-m7")) {
1688 sparc_pmu = &sparc_m7_pmu;
1689 return true;
1690 }
1691 return false;
1692}
1693
1694static int __init init_hw_perf_events(void)
1695{
1696 int err;
1697
1698 pr_info("Performance events: ");
1699
1700 err = pcr_arch_init();
1701 if (err || !supported_pmu()) {
1702 pr_cont("No support for PMU type '%s'\n", sparc_pmu_type);
1703 return 0;
1704 }
1705
1706 pr_cont("Supported PMU type is '%s'\n", sparc_pmu_type);
1707
1708 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1709 register_die_notifier(&perf_event_nmi_notifier);
1710
1711 return 0;
1712}
1713pure_initcall(init_hw_perf_events);
1714
1715void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry,
1716 struct pt_regs *regs)
1717{
1718 unsigned long ksp, fp;
1719#ifdef CONFIG_FUNCTION_GRAPH_TRACER
1720 int graph = 0;
1721#endif
1722
1723 stack_trace_flush();
1724
1725 perf_callchain_store(entry, regs->tpc);
1726
1727 ksp = regs->u_regs[UREG_I6];
1728 fp = ksp + STACK_BIAS;
1729 do {
1730 struct sparc_stackf *sf;
1731 struct pt_regs *regs;
1732 unsigned long pc;
1733
1734 if (!kstack_valid(current_thread_info(), fp))
1735 break;
1736
1737 sf = (struct sparc_stackf *) fp;
1738 regs = (struct pt_regs *) (sf + 1);
1739
1740 if (kstack_is_trap_frame(current_thread_info(), regs)) {
1741 if (user_mode(regs))
1742 break;
1743 pc = regs->tpc;
1744 fp = regs->u_regs[UREG_I6] + STACK_BIAS;
1745 } else {
1746 pc = sf->callers_pc;
1747 fp = (unsigned long)sf->fp + STACK_BIAS;
1748 }
1749 perf_callchain_store(entry, pc);
1750#ifdef CONFIG_FUNCTION_GRAPH_TRACER
1751 if ((pc + 8UL) == (unsigned long) &return_to_handler) {
1752 int index = current->curr_ret_stack;
1753 if (current->ret_stack && index >= graph) {
1754 pc = current->ret_stack[index - graph].ret;
1755 perf_callchain_store(entry, pc);
1756 graph++;
1757 }
1758 }
1759#endif
1760 } while (entry->nr < entry->max_stack);
1761}
1762
1763static inline int
1764valid_user_frame(const void __user *fp, unsigned long size)
1765{
1766 /* addresses should be at least 4-byte aligned */
1767 if (((unsigned long) fp) & 3)
1768 return 0;
1769
1770 return (__range_not_ok(fp, size, TASK_SIZE) == 0);
1771}
1772
1773static void perf_callchain_user_64(struct perf_callchain_entry_ctx *entry,
1774 struct pt_regs *regs)
1775{
1776 unsigned long ufp;
1777
1778 ufp = regs->u_regs[UREG_FP] + STACK_BIAS;
1779 do {
1780 struct sparc_stackf __user *usf;
1781 struct sparc_stackf sf;
1782 unsigned long pc;
1783
1784 usf = (struct sparc_stackf __user *)ufp;
1785 if (!valid_user_frame(usf, sizeof(sf)))
1786 break;
1787
1788 if (__copy_from_user_inatomic(&sf, usf, sizeof(sf)))
1789 break;
1790
1791 pc = sf.callers_pc;
1792 ufp = (unsigned long)sf.fp + STACK_BIAS;
1793 perf_callchain_store(entry, pc);
1794 } while (entry->nr < entry->max_stack);
1795}
1796
1797static void perf_callchain_user_32(struct perf_callchain_entry_ctx *entry,
1798 struct pt_regs *regs)
1799{
1800 unsigned long ufp;
1801
1802 ufp = regs->u_regs[UREG_FP] & 0xffffffffUL;
1803 do {
1804 unsigned long pc;
1805
1806 if (thread32_stack_is_64bit(ufp)) {
1807 struct sparc_stackf __user *usf;
1808 struct sparc_stackf sf;
1809
1810 ufp += STACK_BIAS;
1811 usf = (struct sparc_stackf __user *)ufp;
1812 if (__copy_from_user_inatomic(&sf, usf, sizeof(sf)))
1813 break;
1814 pc = sf.callers_pc & 0xffffffff;
1815 ufp = ((unsigned long) sf.fp) & 0xffffffff;
1816 } else {
1817 struct sparc_stackf32 __user *usf;
1818 struct sparc_stackf32 sf;
1819 usf = (struct sparc_stackf32 __user *)ufp;
1820 if (__copy_from_user_inatomic(&sf, usf, sizeof(sf)))
1821 break;
1822 pc = sf.callers_pc;
1823 ufp = (unsigned long)sf.fp;
1824 }
1825 perf_callchain_store(entry, pc);
1826 } while (entry->nr < entry->max_stack);
1827}
1828
1829void
1830perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
1831{
1832 u64 saved_fault_address = current_thread_info()->fault_address;
1833 u8 saved_fault_code = get_thread_fault_code();
1834 mm_segment_t old_fs;
1835
1836 perf_callchain_store(entry, regs->tpc);
1837
1838 if (!current->mm)
1839 return;
1840
1841 old_fs = get_fs();
1842 set_fs(USER_DS);
1843
1844 flushw_user();
1845
1846 pagefault_disable();
1847
1848 if (test_thread_flag(TIF_32BIT))
1849 perf_callchain_user_32(entry, regs);
1850 else
1851 perf_callchain_user_64(entry, regs);
1852
1853 pagefault_enable();
1854
1855 set_fs(old_fs);
1856 set_thread_fault_code(saved_fault_code);
1857 current_thread_info()->fault_address = saved_fault_address;
1858}
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 <linux/atomic.h>
26#include <asm/nmi.h>
27#include <asm/pcr.h>
28
29#include "kernel.h"
30#include "kstack.h"
31
32/* Sparc64 chips have two performance counters, 32-bits each, with
33 * overflow interrupts generated on transition from 0xffffffff to 0.
34 * The counters are accessed in one go using a 64-bit register.
35 *
36 * Both counters are controlled using a single control register. The
37 * only way to stop all sampling is to clear all of the context (user,
38 * supervisor, hypervisor) sampling enable bits. But these bits apply
39 * to both counters, thus the two counters can't be enabled/disabled
40 * individually.
41 *
42 * The control register has two event fields, one for each of the two
43 * counters. It's thus nearly impossible to have one counter going
44 * while keeping the other one stopped. Therefore it is possible to
45 * get overflow interrupts for counters not currently "in use" and
46 * that condition must be checked in the overflow interrupt handler.
47 *
48 * So we use a hack, in that we program inactive counters with the
49 * "sw_count0" and "sw_count1" events. These count how many times
50 * the instruction "sethi %hi(0xfc000), %g0" is executed. It's an
51 * unusual way to encode a NOP and therefore will not trigger in
52 * normal code.
53 */
54
55#define MAX_HWEVENTS 2
56#define MAX_PERIOD ((1UL << 32) - 1)
57
58#define PIC_UPPER_INDEX 0
59#define PIC_LOWER_INDEX 1
60#define PIC_NO_INDEX -1
61
62struct cpu_hw_events {
63 /* Number of events currently scheduled onto this cpu.
64 * This tells how many entries in the arrays below
65 * are valid.
66 */
67 int n_events;
68
69 /* Number of new events added since the last hw_perf_disable().
70 * This works because the perf event layer always adds new
71 * events inside of a perf_{disable,enable}() sequence.
72 */
73 int n_added;
74
75 /* Array of events current scheduled on this cpu. */
76 struct perf_event *event[MAX_HWEVENTS];
77
78 /* Array of encoded longs, specifying the %pcr register
79 * encoding and the mask of PIC counters this even can
80 * be scheduled on. See perf_event_encode() et al.
81 */
82 unsigned long events[MAX_HWEVENTS];
83
84 /* The current counter index assigned to an event. When the
85 * event hasn't been programmed into the cpu yet, this will
86 * hold PIC_NO_INDEX. The event->hw.idx value tells us where
87 * we ought to schedule the event.
88 */
89 int current_idx[MAX_HWEVENTS];
90
91 /* Software copy of %pcr register on this cpu. */
92 u64 pcr;
93
94 /* Enabled/disable state. */
95 int enabled;
96
97 unsigned int group_flag;
98};
99DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = { .enabled = 1, };
100
101/* An event map describes the characteristics of a performance
102 * counter event. In particular it gives the encoding as well as
103 * a mask telling which counters the event can be measured on.
104 */
105struct perf_event_map {
106 u16 encoding;
107 u8 pic_mask;
108#define PIC_NONE 0x00
109#define PIC_UPPER 0x01
110#define PIC_LOWER 0x02
111};
112
113/* Encode a perf_event_map entry into a long. */
114static unsigned long perf_event_encode(const struct perf_event_map *pmap)
115{
116 return ((unsigned long) pmap->encoding << 16) | pmap->pic_mask;
117}
118
119static u8 perf_event_get_msk(unsigned long val)
120{
121 return val & 0xff;
122}
123
124static u64 perf_event_get_enc(unsigned long val)
125{
126 return val >> 16;
127}
128
129#define C(x) PERF_COUNT_HW_CACHE_##x
130
131#define CACHE_OP_UNSUPPORTED 0xfffe
132#define CACHE_OP_NONSENSE 0xffff
133
134typedef struct perf_event_map cache_map_t
135 [PERF_COUNT_HW_CACHE_MAX]
136 [PERF_COUNT_HW_CACHE_OP_MAX]
137 [PERF_COUNT_HW_CACHE_RESULT_MAX];
138
139struct sparc_pmu {
140 const struct perf_event_map *(*event_map)(int);
141 const cache_map_t *cache_map;
142 int max_events;
143 int upper_shift;
144 int lower_shift;
145 int event_mask;
146 int hv_bit;
147 int irq_bit;
148 int upper_nop;
149 int lower_nop;
150};
151
152static const struct perf_event_map ultra3_perfmon_event_map[] = {
153 [PERF_COUNT_HW_CPU_CYCLES] = { 0x0000, PIC_UPPER | PIC_LOWER },
154 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x0001, PIC_UPPER | PIC_LOWER },
155 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0009, PIC_LOWER },
156 [PERF_COUNT_HW_CACHE_MISSES] = { 0x0009, PIC_UPPER },
157};
158
159static const struct perf_event_map *ultra3_event_map(int event_id)
160{
161 return &ultra3_perfmon_event_map[event_id];
162}
163
164static const cache_map_t ultra3_cache_map = {
165[C(L1D)] = {
166 [C(OP_READ)] = {
167 [C(RESULT_ACCESS)] = { 0x09, PIC_LOWER, },
168 [C(RESULT_MISS)] = { 0x09, PIC_UPPER, },
169 },
170 [C(OP_WRITE)] = {
171 [C(RESULT_ACCESS)] = { 0x0a, PIC_LOWER },
172 [C(RESULT_MISS)] = { 0x0a, PIC_UPPER },
173 },
174 [C(OP_PREFETCH)] = {
175 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
176 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
177 },
178},
179[C(L1I)] = {
180 [C(OP_READ)] = {
181 [C(RESULT_ACCESS)] = { 0x09, PIC_LOWER, },
182 [C(RESULT_MISS)] = { 0x09, PIC_UPPER, },
183 },
184 [ C(OP_WRITE) ] = {
185 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
186 [ C(RESULT_MISS) ] = { CACHE_OP_NONSENSE },
187 },
188 [ C(OP_PREFETCH) ] = {
189 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
190 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
191 },
192},
193[C(LL)] = {
194 [C(OP_READ)] = {
195 [C(RESULT_ACCESS)] = { 0x0c, PIC_LOWER, },
196 [C(RESULT_MISS)] = { 0x0c, PIC_UPPER, },
197 },
198 [C(OP_WRITE)] = {
199 [C(RESULT_ACCESS)] = { 0x0c, PIC_LOWER },
200 [C(RESULT_MISS)] = { 0x0c, PIC_UPPER },
201 },
202 [C(OP_PREFETCH)] = {
203 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
204 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
205 },
206},
207[C(DTLB)] = {
208 [C(OP_READ)] = {
209 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
210 [C(RESULT_MISS)] = { 0x12, PIC_UPPER, },
211 },
212 [ C(OP_WRITE) ] = {
213 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
214 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
215 },
216 [ C(OP_PREFETCH) ] = {
217 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
218 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
219 },
220},
221[C(ITLB)] = {
222 [C(OP_READ)] = {
223 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
224 [C(RESULT_MISS)] = { 0x11, PIC_UPPER, },
225 },
226 [ C(OP_WRITE) ] = {
227 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
228 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
229 },
230 [ C(OP_PREFETCH) ] = {
231 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
232 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
233 },
234},
235[C(BPU)] = {
236 [C(OP_READ)] = {
237 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
238 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
239 },
240 [ C(OP_WRITE) ] = {
241 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
242 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
243 },
244 [ C(OP_PREFETCH) ] = {
245 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
246 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
247 },
248},
249[C(NODE)] = {
250 [C(OP_READ)] = {
251 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
252 [C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
253 },
254 [ C(OP_WRITE) ] = {
255 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
256 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
257 },
258 [ C(OP_PREFETCH) ] = {
259 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
260 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
261 },
262},
263};
264
265static const struct sparc_pmu ultra3_pmu = {
266 .event_map = ultra3_event_map,
267 .cache_map = &ultra3_cache_map,
268 .max_events = ARRAY_SIZE(ultra3_perfmon_event_map),
269 .upper_shift = 11,
270 .lower_shift = 4,
271 .event_mask = 0x3f,
272 .upper_nop = 0x1c,
273 .lower_nop = 0x14,
274};
275
276/* Niagara1 is very limited. The upper PIC is hard-locked to count
277 * only instructions, so it is free running which creates all kinds of
278 * problems. Some hardware designs make one wonder if the creator
279 * even looked at how this stuff gets used by software.
280 */
281static const struct perf_event_map niagara1_perfmon_event_map[] = {
282 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, PIC_UPPER },
283 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x00, PIC_UPPER },
284 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0, PIC_NONE },
285 [PERF_COUNT_HW_CACHE_MISSES] = { 0x03, PIC_LOWER },
286};
287
288static const struct perf_event_map *niagara1_event_map(int event_id)
289{
290 return &niagara1_perfmon_event_map[event_id];
291}
292
293static const cache_map_t niagara1_cache_map = {
294[C(L1D)] = {
295 [C(OP_READ)] = {
296 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
297 [C(RESULT_MISS)] = { 0x03, PIC_LOWER, },
298 },
299 [C(OP_WRITE)] = {
300 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
301 [C(RESULT_MISS)] = { 0x03, PIC_LOWER, },
302 },
303 [C(OP_PREFETCH)] = {
304 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
305 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
306 },
307},
308[C(L1I)] = {
309 [C(OP_READ)] = {
310 [C(RESULT_ACCESS)] = { 0x00, PIC_UPPER },
311 [C(RESULT_MISS)] = { 0x02, PIC_LOWER, },
312 },
313 [ C(OP_WRITE) ] = {
314 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
315 [ C(RESULT_MISS) ] = { CACHE_OP_NONSENSE },
316 },
317 [ C(OP_PREFETCH) ] = {
318 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
319 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
320 },
321},
322[C(LL)] = {
323 [C(OP_READ)] = {
324 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
325 [C(RESULT_MISS)] = { 0x07, PIC_LOWER, },
326 },
327 [C(OP_WRITE)] = {
328 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
329 [C(RESULT_MISS)] = { 0x07, PIC_LOWER, },
330 },
331 [C(OP_PREFETCH)] = {
332 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
333 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
334 },
335},
336[C(DTLB)] = {
337 [C(OP_READ)] = {
338 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
339 [C(RESULT_MISS)] = { 0x05, PIC_LOWER, },
340 },
341 [ C(OP_WRITE) ] = {
342 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
343 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
344 },
345 [ C(OP_PREFETCH) ] = {
346 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
347 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
348 },
349},
350[C(ITLB)] = {
351 [C(OP_READ)] = {
352 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
353 [C(RESULT_MISS)] = { 0x04, PIC_LOWER, },
354 },
355 [ C(OP_WRITE) ] = {
356 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
357 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
358 },
359 [ C(OP_PREFETCH) ] = {
360 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
361 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
362 },
363},
364[C(BPU)] = {
365 [C(OP_READ)] = {
366 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
367 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
368 },
369 [ C(OP_WRITE) ] = {
370 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
371 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
372 },
373 [ C(OP_PREFETCH) ] = {
374 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
375 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
376 },
377},
378[C(NODE)] = {
379 [C(OP_READ)] = {
380 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
381 [C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
382 },
383 [ C(OP_WRITE) ] = {
384 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
385 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
386 },
387 [ C(OP_PREFETCH) ] = {
388 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
389 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
390 },
391},
392};
393
394static const struct sparc_pmu niagara1_pmu = {
395 .event_map = niagara1_event_map,
396 .cache_map = &niagara1_cache_map,
397 .max_events = ARRAY_SIZE(niagara1_perfmon_event_map),
398 .upper_shift = 0,
399 .lower_shift = 4,
400 .event_mask = 0x7,
401 .upper_nop = 0x0,
402 .lower_nop = 0x0,
403};
404
405static const struct perf_event_map niagara2_perfmon_event_map[] = {
406 [PERF_COUNT_HW_CPU_CYCLES] = { 0x02ff, PIC_UPPER | PIC_LOWER },
407 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x02ff, PIC_UPPER | PIC_LOWER },
408 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0208, PIC_UPPER | PIC_LOWER },
409 [PERF_COUNT_HW_CACHE_MISSES] = { 0x0302, PIC_UPPER | PIC_LOWER },
410 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x0201, PIC_UPPER | PIC_LOWER },
411 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x0202, PIC_UPPER | PIC_LOWER },
412};
413
414static const struct perf_event_map *niagara2_event_map(int event_id)
415{
416 return &niagara2_perfmon_event_map[event_id];
417}
418
419static const cache_map_t niagara2_cache_map = {
420[C(L1D)] = {
421 [C(OP_READ)] = {
422 [C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, },
423 [C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, },
424 },
425 [C(OP_WRITE)] = {
426 [C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, },
427 [C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, },
428 },
429 [C(OP_PREFETCH)] = {
430 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
431 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
432 },
433},
434[C(L1I)] = {
435 [C(OP_READ)] = {
436 [C(RESULT_ACCESS)] = { 0x02ff, PIC_UPPER | PIC_LOWER, },
437 [C(RESULT_MISS)] = { 0x0301, PIC_UPPER | PIC_LOWER, },
438 },
439 [ C(OP_WRITE) ] = {
440 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
441 [ C(RESULT_MISS) ] = { CACHE_OP_NONSENSE },
442 },
443 [ C(OP_PREFETCH) ] = {
444 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
445 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
446 },
447},
448[C(LL)] = {
449 [C(OP_READ)] = {
450 [C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, },
451 [C(RESULT_MISS)] = { 0x0330, PIC_UPPER | PIC_LOWER, },
452 },
453 [C(OP_WRITE)] = {
454 [C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, },
455 [C(RESULT_MISS)] = { 0x0320, PIC_UPPER | PIC_LOWER, },
456 },
457 [C(OP_PREFETCH)] = {
458 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
459 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
460 },
461},
462[C(DTLB)] = {
463 [C(OP_READ)] = {
464 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
465 [C(RESULT_MISS)] = { 0x0b08, PIC_UPPER | PIC_LOWER, },
466 },
467 [ C(OP_WRITE) ] = {
468 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
469 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
470 },
471 [ C(OP_PREFETCH) ] = {
472 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
473 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
474 },
475},
476[C(ITLB)] = {
477 [C(OP_READ)] = {
478 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
479 [C(RESULT_MISS)] = { 0xb04, PIC_UPPER | PIC_LOWER, },
480 },
481 [ C(OP_WRITE) ] = {
482 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
483 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
484 },
485 [ C(OP_PREFETCH) ] = {
486 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
487 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
488 },
489},
490[C(BPU)] = {
491 [C(OP_READ)] = {
492 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
493 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
494 },
495 [ C(OP_WRITE) ] = {
496 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
497 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
498 },
499 [ C(OP_PREFETCH) ] = {
500 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
501 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
502 },
503},
504[C(NODE)] = {
505 [C(OP_READ)] = {
506 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
507 [C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
508 },
509 [ C(OP_WRITE) ] = {
510 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
511 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
512 },
513 [ C(OP_PREFETCH) ] = {
514 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
515 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED },
516 },
517},
518};
519
520static const struct sparc_pmu niagara2_pmu = {
521 .event_map = niagara2_event_map,
522 .cache_map = &niagara2_cache_map,
523 .max_events = ARRAY_SIZE(niagara2_perfmon_event_map),
524 .upper_shift = 19,
525 .lower_shift = 6,
526 .event_mask = 0xfff,
527 .hv_bit = 0x8,
528 .irq_bit = 0x30,
529 .upper_nop = 0x220,
530 .lower_nop = 0x220,
531};
532
533static const struct sparc_pmu *sparc_pmu __read_mostly;
534
535static u64 event_encoding(u64 event_id, int idx)
536{
537 if (idx == PIC_UPPER_INDEX)
538 event_id <<= sparc_pmu->upper_shift;
539 else
540 event_id <<= sparc_pmu->lower_shift;
541 return event_id;
542}
543
544static u64 mask_for_index(int idx)
545{
546 return event_encoding(sparc_pmu->event_mask, idx);
547}
548
549static u64 nop_for_index(int idx)
550{
551 return event_encoding(idx == PIC_UPPER_INDEX ?
552 sparc_pmu->upper_nop :
553 sparc_pmu->lower_nop, idx);
554}
555
556static inline void sparc_pmu_enable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx)
557{
558 u64 val, mask = mask_for_index(idx);
559
560 val = cpuc->pcr;
561 val &= ~mask;
562 val |= hwc->config;
563 cpuc->pcr = val;
564
565 pcr_ops->write(cpuc->pcr);
566}
567
568static inline void sparc_pmu_disable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx)
569{
570 u64 mask = mask_for_index(idx);
571 u64 nop = nop_for_index(idx);
572 u64 val;
573
574 val = cpuc->pcr;
575 val &= ~mask;
576 val |= nop;
577 cpuc->pcr = val;
578
579 pcr_ops->write(cpuc->pcr);
580}
581
582static u32 read_pmc(int idx)
583{
584 u64 val;
585
586 read_pic(val);
587 if (idx == PIC_UPPER_INDEX)
588 val >>= 32;
589
590 return val & 0xffffffff;
591}
592
593static void write_pmc(int idx, u64 val)
594{
595 u64 shift, mask, pic;
596
597 shift = 0;
598 if (idx == PIC_UPPER_INDEX)
599 shift = 32;
600
601 mask = ((u64) 0xffffffff) << shift;
602 val <<= shift;
603
604 read_pic(pic);
605 pic &= ~mask;
606 pic |= val;
607 write_pic(pic);
608}
609
610static u64 sparc_perf_event_update(struct perf_event *event,
611 struct hw_perf_event *hwc, int idx)
612{
613 int shift = 64 - 32;
614 u64 prev_raw_count, new_raw_count;
615 s64 delta;
616
617again:
618 prev_raw_count = local64_read(&hwc->prev_count);
619 new_raw_count = read_pmc(idx);
620
621 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
622 new_raw_count) != prev_raw_count)
623 goto again;
624
625 delta = (new_raw_count << shift) - (prev_raw_count << shift);
626 delta >>= shift;
627
628 local64_add(delta, &event->count);
629 local64_sub(delta, &hwc->period_left);
630
631 return new_raw_count;
632}
633
634static int sparc_perf_event_set_period(struct perf_event *event,
635 struct hw_perf_event *hwc, int idx)
636{
637 s64 left = local64_read(&hwc->period_left);
638 s64 period = hwc->sample_period;
639 int ret = 0;
640
641 if (unlikely(left <= -period)) {
642 left = period;
643 local64_set(&hwc->period_left, left);
644 hwc->last_period = period;
645 ret = 1;
646 }
647
648 if (unlikely(left <= 0)) {
649 left += period;
650 local64_set(&hwc->period_left, left);
651 hwc->last_period = period;
652 ret = 1;
653 }
654 if (left > MAX_PERIOD)
655 left = MAX_PERIOD;
656
657 local64_set(&hwc->prev_count, (u64)-left);
658
659 write_pmc(idx, (u64)(-left) & 0xffffffff);
660
661 perf_event_update_userpage(event);
662
663 return ret;
664}
665
666/* If performance event entries have been added, move existing
667 * events around (if necessary) and then assign new entries to
668 * counters.
669 */
670static u64 maybe_change_configuration(struct cpu_hw_events *cpuc, u64 pcr)
671{
672 int i;
673
674 if (!cpuc->n_added)
675 goto out;
676
677 /* Read in the counters which are moving. */
678 for (i = 0; i < cpuc->n_events; i++) {
679 struct perf_event *cp = cpuc->event[i];
680
681 if (cpuc->current_idx[i] != PIC_NO_INDEX &&
682 cpuc->current_idx[i] != cp->hw.idx) {
683 sparc_perf_event_update(cp, &cp->hw,
684 cpuc->current_idx[i]);
685 cpuc->current_idx[i] = PIC_NO_INDEX;
686 }
687 }
688
689 /* Assign to counters all unassigned events. */
690 for (i = 0; i < cpuc->n_events; i++) {
691 struct perf_event *cp = cpuc->event[i];
692 struct hw_perf_event *hwc = &cp->hw;
693 int idx = hwc->idx;
694 u64 enc;
695
696 if (cpuc->current_idx[i] != PIC_NO_INDEX)
697 continue;
698
699 sparc_perf_event_set_period(cp, hwc, idx);
700 cpuc->current_idx[i] = idx;
701
702 enc = perf_event_get_enc(cpuc->events[i]);
703 pcr &= ~mask_for_index(idx);
704 if (hwc->state & PERF_HES_STOPPED)
705 pcr |= nop_for_index(idx);
706 else
707 pcr |= event_encoding(enc, idx);
708 }
709out:
710 return pcr;
711}
712
713static void sparc_pmu_enable(struct pmu *pmu)
714{
715 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
716 u64 pcr;
717
718 if (cpuc->enabled)
719 return;
720
721 cpuc->enabled = 1;
722 barrier();
723
724 pcr = cpuc->pcr;
725 if (!cpuc->n_events) {
726 pcr = 0;
727 } else {
728 pcr = maybe_change_configuration(cpuc, pcr);
729
730 /* We require that all of the events have the same
731 * configuration, so just fetch the settings from the
732 * first entry.
733 */
734 cpuc->pcr = pcr | cpuc->event[0]->hw.config_base;
735 }
736
737 pcr_ops->write(cpuc->pcr);
738}
739
740static void sparc_pmu_disable(struct pmu *pmu)
741{
742 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
743 u64 val;
744
745 if (!cpuc->enabled)
746 return;
747
748 cpuc->enabled = 0;
749 cpuc->n_added = 0;
750
751 val = cpuc->pcr;
752 val &= ~(PCR_UTRACE | PCR_STRACE |
753 sparc_pmu->hv_bit | sparc_pmu->irq_bit);
754 cpuc->pcr = val;
755
756 pcr_ops->write(cpuc->pcr);
757}
758
759static int active_event_index(struct cpu_hw_events *cpuc,
760 struct perf_event *event)
761{
762 int i;
763
764 for (i = 0; i < cpuc->n_events; i++) {
765 if (cpuc->event[i] == event)
766 break;
767 }
768 BUG_ON(i == cpuc->n_events);
769 return cpuc->current_idx[i];
770}
771
772static void sparc_pmu_start(struct perf_event *event, int flags)
773{
774 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
775 int idx = active_event_index(cpuc, event);
776
777 if (flags & PERF_EF_RELOAD) {
778 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
779 sparc_perf_event_set_period(event, &event->hw, idx);
780 }
781
782 event->hw.state = 0;
783
784 sparc_pmu_enable_event(cpuc, &event->hw, idx);
785}
786
787static void sparc_pmu_stop(struct perf_event *event, int flags)
788{
789 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
790 int idx = active_event_index(cpuc, event);
791
792 if (!(event->hw.state & PERF_HES_STOPPED)) {
793 sparc_pmu_disable_event(cpuc, &event->hw, idx);
794 event->hw.state |= PERF_HES_STOPPED;
795 }
796
797 if (!(event->hw.state & PERF_HES_UPTODATE) && (flags & PERF_EF_UPDATE)) {
798 sparc_perf_event_update(event, &event->hw, idx);
799 event->hw.state |= PERF_HES_UPTODATE;
800 }
801}
802
803static void sparc_pmu_del(struct perf_event *event, int _flags)
804{
805 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
806 unsigned long flags;
807 int i;
808
809 local_irq_save(flags);
810 perf_pmu_disable(event->pmu);
811
812 for (i = 0; i < cpuc->n_events; i++) {
813 if (event == cpuc->event[i]) {
814 /* Absorb the final count and turn off the
815 * event.
816 */
817 sparc_pmu_stop(event, PERF_EF_UPDATE);
818
819 /* Shift remaining entries down into
820 * the existing slot.
821 */
822 while (++i < cpuc->n_events) {
823 cpuc->event[i - 1] = cpuc->event[i];
824 cpuc->events[i - 1] = cpuc->events[i];
825 cpuc->current_idx[i - 1] =
826 cpuc->current_idx[i];
827 }
828
829 perf_event_update_userpage(event);
830
831 cpuc->n_events--;
832 break;
833 }
834 }
835
836 perf_pmu_enable(event->pmu);
837 local_irq_restore(flags);
838}
839
840static void sparc_pmu_read(struct perf_event *event)
841{
842 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
843 int idx = active_event_index(cpuc, event);
844 struct hw_perf_event *hwc = &event->hw;
845
846 sparc_perf_event_update(event, hwc, idx);
847}
848
849static atomic_t active_events = ATOMIC_INIT(0);
850static DEFINE_MUTEX(pmc_grab_mutex);
851
852static void perf_stop_nmi_watchdog(void *unused)
853{
854 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
855
856 stop_nmi_watchdog(NULL);
857 cpuc->pcr = pcr_ops->read();
858}
859
860void perf_event_grab_pmc(void)
861{
862 if (atomic_inc_not_zero(&active_events))
863 return;
864
865 mutex_lock(&pmc_grab_mutex);
866 if (atomic_read(&active_events) == 0) {
867 if (atomic_read(&nmi_active) > 0) {
868 on_each_cpu(perf_stop_nmi_watchdog, NULL, 1);
869 BUG_ON(atomic_read(&nmi_active) != 0);
870 }
871 atomic_inc(&active_events);
872 }
873 mutex_unlock(&pmc_grab_mutex);
874}
875
876void perf_event_release_pmc(void)
877{
878 if (atomic_dec_and_mutex_lock(&active_events, &pmc_grab_mutex)) {
879 if (atomic_read(&nmi_active) == 0)
880 on_each_cpu(start_nmi_watchdog, NULL, 1);
881 mutex_unlock(&pmc_grab_mutex);
882 }
883}
884
885static const struct perf_event_map *sparc_map_cache_event(u64 config)
886{
887 unsigned int cache_type, cache_op, cache_result;
888 const struct perf_event_map *pmap;
889
890 if (!sparc_pmu->cache_map)
891 return ERR_PTR(-ENOENT);
892
893 cache_type = (config >> 0) & 0xff;
894 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
895 return ERR_PTR(-EINVAL);
896
897 cache_op = (config >> 8) & 0xff;
898 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
899 return ERR_PTR(-EINVAL);
900
901 cache_result = (config >> 16) & 0xff;
902 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
903 return ERR_PTR(-EINVAL);
904
905 pmap = &((*sparc_pmu->cache_map)[cache_type][cache_op][cache_result]);
906
907 if (pmap->encoding == CACHE_OP_UNSUPPORTED)
908 return ERR_PTR(-ENOENT);
909
910 if (pmap->encoding == CACHE_OP_NONSENSE)
911 return ERR_PTR(-EINVAL);
912
913 return pmap;
914}
915
916static void hw_perf_event_destroy(struct perf_event *event)
917{
918 perf_event_release_pmc();
919}
920
921/* Make sure all events can be scheduled into the hardware at
922 * the same time. This is simplified by the fact that we only
923 * need to support 2 simultaneous HW events.
924 *
925 * As a side effect, the evts[]->hw.idx values will be assigned
926 * on success. These are pending indexes. When the events are
927 * actually programmed into the chip, these values will propagate
928 * to the per-cpu cpuc->current_idx[] slots, see the code in
929 * maybe_change_configuration() for details.
930 */
931static int sparc_check_constraints(struct perf_event **evts,
932 unsigned long *events, int n_ev)
933{
934 u8 msk0 = 0, msk1 = 0;
935 int idx0 = 0;
936
937 /* This case is possible when we are invoked from
938 * hw_perf_group_sched_in().
939 */
940 if (!n_ev)
941 return 0;
942
943 if (n_ev > MAX_HWEVENTS)
944 return -1;
945
946 msk0 = perf_event_get_msk(events[0]);
947 if (n_ev == 1) {
948 if (msk0 & PIC_LOWER)
949 idx0 = 1;
950 goto success;
951 }
952 BUG_ON(n_ev != 2);
953 msk1 = perf_event_get_msk(events[1]);
954
955 /* If both events can go on any counter, OK. */
956 if (msk0 == (PIC_UPPER | PIC_LOWER) &&
957 msk1 == (PIC_UPPER | PIC_LOWER))
958 goto success;
959
960 /* If one event is limited to a specific counter,
961 * and the other can go on both, OK.
962 */
963 if ((msk0 == PIC_UPPER || msk0 == PIC_LOWER) &&
964 msk1 == (PIC_UPPER | PIC_LOWER)) {
965 if (msk0 & PIC_LOWER)
966 idx0 = 1;
967 goto success;
968 }
969
970 if ((msk1 == PIC_UPPER || msk1 == PIC_LOWER) &&
971 msk0 == (PIC_UPPER | PIC_LOWER)) {
972 if (msk1 & PIC_UPPER)
973 idx0 = 1;
974 goto success;
975 }
976
977 /* If the events are fixed to different counters, OK. */
978 if ((msk0 == PIC_UPPER && msk1 == PIC_LOWER) ||
979 (msk0 == PIC_LOWER && msk1 == PIC_UPPER)) {
980 if (msk0 & PIC_LOWER)
981 idx0 = 1;
982 goto success;
983 }
984
985 /* Otherwise, there is a conflict. */
986 return -1;
987
988success:
989 evts[0]->hw.idx = idx0;
990 if (n_ev == 2)
991 evts[1]->hw.idx = idx0 ^ 1;
992 return 0;
993}
994
995static int check_excludes(struct perf_event **evts, int n_prev, int n_new)
996{
997 int eu = 0, ek = 0, eh = 0;
998 struct perf_event *event;
999 int i, n, first;
1000
1001 n = n_prev + n_new;
1002 if (n <= 1)
1003 return 0;
1004
1005 first = 1;
1006 for (i = 0; i < n; i++) {
1007 event = evts[i];
1008 if (first) {
1009 eu = event->attr.exclude_user;
1010 ek = event->attr.exclude_kernel;
1011 eh = event->attr.exclude_hv;
1012 first = 0;
1013 } else if (event->attr.exclude_user != eu ||
1014 event->attr.exclude_kernel != ek ||
1015 event->attr.exclude_hv != eh) {
1016 return -EAGAIN;
1017 }
1018 }
1019
1020 return 0;
1021}
1022
1023static int collect_events(struct perf_event *group, int max_count,
1024 struct perf_event *evts[], unsigned long *events,
1025 int *current_idx)
1026{
1027 struct perf_event *event;
1028 int n = 0;
1029
1030 if (!is_software_event(group)) {
1031 if (n >= max_count)
1032 return -1;
1033 evts[n] = group;
1034 events[n] = group->hw.event_base;
1035 current_idx[n++] = PIC_NO_INDEX;
1036 }
1037 list_for_each_entry(event, &group->sibling_list, group_entry) {
1038 if (!is_software_event(event) &&
1039 event->state != PERF_EVENT_STATE_OFF) {
1040 if (n >= max_count)
1041 return -1;
1042 evts[n] = event;
1043 events[n] = event->hw.event_base;
1044 current_idx[n++] = PIC_NO_INDEX;
1045 }
1046 }
1047 return n;
1048}
1049
1050static int sparc_pmu_add(struct perf_event *event, int ef_flags)
1051{
1052 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1053 int n0, ret = -EAGAIN;
1054 unsigned long flags;
1055
1056 local_irq_save(flags);
1057 perf_pmu_disable(event->pmu);
1058
1059 n0 = cpuc->n_events;
1060 if (n0 >= MAX_HWEVENTS)
1061 goto out;
1062
1063 cpuc->event[n0] = event;
1064 cpuc->events[n0] = event->hw.event_base;
1065 cpuc->current_idx[n0] = PIC_NO_INDEX;
1066
1067 event->hw.state = PERF_HES_UPTODATE;
1068 if (!(ef_flags & PERF_EF_START))
1069 event->hw.state |= PERF_HES_STOPPED;
1070
1071 /*
1072 * If group events scheduling transaction was started,
1073 * skip the schedulability test here, it will be performed
1074 * at commit time(->commit_txn) as a whole
1075 */
1076 if (cpuc->group_flag & PERF_EVENT_TXN)
1077 goto nocheck;
1078
1079 if (check_excludes(cpuc->event, n0, 1))
1080 goto out;
1081 if (sparc_check_constraints(cpuc->event, cpuc->events, n0 + 1))
1082 goto out;
1083
1084nocheck:
1085 cpuc->n_events++;
1086 cpuc->n_added++;
1087
1088 ret = 0;
1089out:
1090 perf_pmu_enable(event->pmu);
1091 local_irq_restore(flags);
1092 return ret;
1093}
1094
1095static int sparc_pmu_event_init(struct perf_event *event)
1096{
1097 struct perf_event_attr *attr = &event->attr;
1098 struct perf_event *evts[MAX_HWEVENTS];
1099 struct hw_perf_event *hwc = &event->hw;
1100 unsigned long events[MAX_HWEVENTS];
1101 int current_idx_dmy[MAX_HWEVENTS];
1102 const struct perf_event_map *pmap;
1103 int n;
1104
1105 if (atomic_read(&nmi_active) < 0)
1106 return -ENODEV;
1107
1108 switch (attr->type) {
1109 case PERF_TYPE_HARDWARE:
1110 if (attr->config >= sparc_pmu->max_events)
1111 return -EINVAL;
1112 pmap = sparc_pmu->event_map(attr->config);
1113 break;
1114
1115 case PERF_TYPE_HW_CACHE:
1116 pmap = sparc_map_cache_event(attr->config);
1117 if (IS_ERR(pmap))
1118 return PTR_ERR(pmap);
1119 break;
1120
1121 case PERF_TYPE_RAW:
1122 pmap = NULL;
1123 break;
1124
1125 default:
1126 return -ENOENT;
1127
1128 }
1129
1130 if (pmap) {
1131 hwc->event_base = perf_event_encode(pmap);
1132 } else {
1133 /*
1134 * User gives us "(encoding << 16) | pic_mask" for
1135 * PERF_TYPE_RAW events.
1136 */
1137 hwc->event_base = attr->config;
1138 }
1139
1140 /* We save the enable bits in the config_base. */
1141 hwc->config_base = sparc_pmu->irq_bit;
1142 if (!attr->exclude_user)
1143 hwc->config_base |= PCR_UTRACE;
1144 if (!attr->exclude_kernel)
1145 hwc->config_base |= PCR_STRACE;
1146 if (!attr->exclude_hv)
1147 hwc->config_base |= sparc_pmu->hv_bit;
1148
1149 n = 0;
1150 if (event->group_leader != event) {
1151 n = collect_events(event->group_leader,
1152 MAX_HWEVENTS - 1,
1153 evts, events, current_idx_dmy);
1154 if (n < 0)
1155 return -EINVAL;
1156 }
1157 events[n] = hwc->event_base;
1158 evts[n] = event;
1159
1160 if (check_excludes(evts, n, 1))
1161 return -EINVAL;
1162
1163 if (sparc_check_constraints(evts, events, n + 1))
1164 return -EINVAL;
1165
1166 hwc->idx = PIC_NO_INDEX;
1167
1168 /* Try to do all error checking before this point, as unwinding
1169 * state after grabbing the PMC is difficult.
1170 */
1171 perf_event_grab_pmc();
1172 event->destroy = hw_perf_event_destroy;
1173
1174 if (!hwc->sample_period) {
1175 hwc->sample_period = MAX_PERIOD;
1176 hwc->last_period = hwc->sample_period;
1177 local64_set(&hwc->period_left, hwc->sample_period);
1178 }
1179
1180 return 0;
1181}
1182
1183/*
1184 * Start group events scheduling transaction
1185 * Set the flag to make pmu::enable() not perform the
1186 * schedulability test, it will be performed at commit time
1187 */
1188static void sparc_pmu_start_txn(struct pmu *pmu)
1189{
1190 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1191
1192 perf_pmu_disable(pmu);
1193 cpuhw->group_flag |= PERF_EVENT_TXN;
1194}
1195
1196/*
1197 * Stop group events scheduling transaction
1198 * Clear the flag and pmu::enable() will perform the
1199 * schedulability test.
1200 */
1201static void sparc_pmu_cancel_txn(struct pmu *pmu)
1202{
1203 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1204
1205 cpuhw->group_flag &= ~PERF_EVENT_TXN;
1206 perf_pmu_enable(pmu);
1207}
1208
1209/*
1210 * Commit group events scheduling transaction
1211 * Perform the group schedulability test as a whole
1212 * Return 0 if success
1213 */
1214static int sparc_pmu_commit_txn(struct pmu *pmu)
1215{
1216 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1217 int n;
1218
1219 if (!sparc_pmu)
1220 return -EINVAL;
1221
1222 cpuc = &__get_cpu_var(cpu_hw_events);
1223 n = cpuc->n_events;
1224 if (check_excludes(cpuc->event, 0, n))
1225 return -EINVAL;
1226 if (sparc_check_constraints(cpuc->event, cpuc->events, n))
1227 return -EAGAIN;
1228
1229 cpuc->group_flag &= ~PERF_EVENT_TXN;
1230 perf_pmu_enable(pmu);
1231 return 0;
1232}
1233
1234static struct pmu pmu = {
1235 .pmu_enable = sparc_pmu_enable,
1236 .pmu_disable = sparc_pmu_disable,
1237 .event_init = sparc_pmu_event_init,
1238 .add = sparc_pmu_add,
1239 .del = sparc_pmu_del,
1240 .start = sparc_pmu_start,
1241 .stop = sparc_pmu_stop,
1242 .read = sparc_pmu_read,
1243 .start_txn = sparc_pmu_start_txn,
1244 .cancel_txn = sparc_pmu_cancel_txn,
1245 .commit_txn = sparc_pmu_commit_txn,
1246};
1247
1248void perf_event_print_debug(void)
1249{
1250 unsigned long flags;
1251 u64 pcr, pic;
1252 int cpu;
1253
1254 if (!sparc_pmu)
1255 return;
1256
1257 local_irq_save(flags);
1258
1259 cpu = smp_processor_id();
1260
1261 pcr = pcr_ops->read();
1262 read_pic(pic);
1263
1264 pr_info("\n");
1265 pr_info("CPU#%d: PCR[%016llx] PIC[%016llx]\n",
1266 cpu, pcr, pic);
1267
1268 local_irq_restore(flags);
1269}
1270
1271static int __kprobes perf_event_nmi_handler(struct notifier_block *self,
1272 unsigned long cmd, void *__args)
1273{
1274 struct die_args *args = __args;
1275 struct perf_sample_data data;
1276 struct cpu_hw_events *cpuc;
1277 struct pt_regs *regs;
1278 int i;
1279
1280 if (!atomic_read(&active_events))
1281 return NOTIFY_DONE;
1282
1283 switch (cmd) {
1284 case DIE_NMI:
1285 break;
1286
1287 default:
1288 return NOTIFY_DONE;
1289 }
1290
1291 regs = args->regs;
1292
1293 perf_sample_data_init(&data, 0);
1294
1295 cpuc = &__get_cpu_var(cpu_hw_events);
1296
1297 /* If the PMU has the TOE IRQ enable bits, we need to do a
1298 * dummy write to the %pcr to clear the overflow bits and thus
1299 * the interrupt.
1300 *
1301 * Do this before we peek at the counters to determine
1302 * overflow so we don't lose any events.
1303 */
1304 if (sparc_pmu->irq_bit)
1305 pcr_ops->write(cpuc->pcr);
1306
1307 for (i = 0; i < cpuc->n_events; i++) {
1308 struct perf_event *event = cpuc->event[i];
1309 int idx = cpuc->current_idx[i];
1310 struct hw_perf_event *hwc;
1311 u64 val;
1312
1313 hwc = &event->hw;
1314 val = sparc_perf_event_update(event, hwc, idx);
1315 if (val & (1ULL << 31))
1316 continue;
1317
1318 data.period = event->hw.last_period;
1319 if (!sparc_perf_event_set_period(event, hwc, idx))
1320 continue;
1321
1322 if (perf_event_overflow(event, &data, regs))
1323 sparc_pmu_stop(event, 0);
1324 }
1325
1326 return NOTIFY_STOP;
1327}
1328
1329static __read_mostly struct notifier_block perf_event_nmi_notifier = {
1330 .notifier_call = perf_event_nmi_handler,
1331};
1332
1333static bool __init supported_pmu(void)
1334{
1335 if (!strcmp(sparc_pmu_type, "ultra3") ||
1336 !strcmp(sparc_pmu_type, "ultra3+") ||
1337 !strcmp(sparc_pmu_type, "ultra3i") ||
1338 !strcmp(sparc_pmu_type, "ultra4+")) {
1339 sparc_pmu = &ultra3_pmu;
1340 return true;
1341 }
1342 if (!strcmp(sparc_pmu_type, "niagara")) {
1343 sparc_pmu = &niagara1_pmu;
1344 return true;
1345 }
1346 if (!strcmp(sparc_pmu_type, "niagara2") ||
1347 !strcmp(sparc_pmu_type, "niagara3")) {
1348 sparc_pmu = &niagara2_pmu;
1349 return true;
1350 }
1351 return false;
1352}
1353
1354int __init init_hw_perf_events(void)
1355{
1356 pr_info("Performance events: ");
1357
1358 if (!supported_pmu()) {
1359 pr_cont("No support for PMU type '%s'\n", sparc_pmu_type);
1360 return 0;
1361 }
1362
1363 pr_cont("Supported PMU type is '%s'\n", sparc_pmu_type);
1364
1365 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1366 register_die_notifier(&perf_event_nmi_notifier);
1367
1368 return 0;
1369}
1370early_initcall(init_hw_perf_events);
1371
1372void perf_callchain_kernel(struct perf_callchain_entry *entry,
1373 struct pt_regs *regs)
1374{
1375 unsigned long ksp, fp;
1376#ifdef CONFIG_FUNCTION_GRAPH_TRACER
1377 int graph = 0;
1378#endif
1379
1380 stack_trace_flush();
1381
1382 perf_callchain_store(entry, regs->tpc);
1383
1384 ksp = regs->u_regs[UREG_I6];
1385 fp = ksp + STACK_BIAS;
1386 do {
1387 struct sparc_stackf *sf;
1388 struct pt_regs *regs;
1389 unsigned long pc;
1390
1391 if (!kstack_valid(current_thread_info(), fp))
1392 break;
1393
1394 sf = (struct sparc_stackf *) fp;
1395 regs = (struct pt_regs *) (sf + 1);
1396
1397 if (kstack_is_trap_frame(current_thread_info(), regs)) {
1398 if (user_mode(regs))
1399 break;
1400 pc = regs->tpc;
1401 fp = regs->u_regs[UREG_I6] + STACK_BIAS;
1402 } else {
1403 pc = sf->callers_pc;
1404 fp = (unsigned long)sf->fp + STACK_BIAS;
1405 }
1406 perf_callchain_store(entry, pc);
1407#ifdef CONFIG_FUNCTION_GRAPH_TRACER
1408 if ((pc + 8UL) == (unsigned long) &return_to_handler) {
1409 int index = current->curr_ret_stack;
1410 if (current->ret_stack && index >= graph) {
1411 pc = current->ret_stack[index - graph].ret;
1412 perf_callchain_store(entry, pc);
1413 graph++;
1414 }
1415 }
1416#endif
1417 } while (entry->nr < PERF_MAX_STACK_DEPTH);
1418}
1419
1420static void perf_callchain_user_64(struct perf_callchain_entry *entry,
1421 struct pt_regs *regs)
1422{
1423 unsigned long ufp;
1424
1425 perf_callchain_store(entry, regs->tpc);
1426
1427 ufp = regs->u_regs[UREG_I6] + STACK_BIAS;
1428 do {
1429 struct sparc_stackf *usf, sf;
1430 unsigned long pc;
1431
1432 usf = (struct sparc_stackf *) ufp;
1433 if (__copy_from_user_inatomic(&sf, usf, sizeof(sf)))
1434 break;
1435
1436 pc = sf.callers_pc;
1437 ufp = (unsigned long)sf.fp + STACK_BIAS;
1438 perf_callchain_store(entry, pc);
1439 } while (entry->nr < PERF_MAX_STACK_DEPTH);
1440}
1441
1442static void perf_callchain_user_32(struct perf_callchain_entry *entry,
1443 struct pt_regs *regs)
1444{
1445 unsigned long ufp;
1446
1447 perf_callchain_store(entry, regs->tpc);
1448
1449 ufp = regs->u_regs[UREG_I6] & 0xffffffffUL;
1450 do {
1451 struct sparc_stackf32 *usf, sf;
1452 unsigned long pc;
1453
1454 usf = (struct sparc_stackf32 *) ufp;
1455 if (__copy_from_user_inatomic(&sf, usf, sizeof(sf)))
1456 break;
1457
1458 pc = sf.callers_pc;
1459 ufp = (unsigned long)sf.fp;
1460 perf_callchain_store(entry, pc);
1461 } while (entry->nr < PERF_MAX_STACK_DEPTH);
1462}
1463
1464void
1465perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
1466{
1467 flushw_user();
1468 if (test_thread_flag(TIF_32BIT))
1469 perf_callchain_user_32(entry, regs);
1470 else
1471 perf_callchain_user_64(entry, regs);
1472}