xref: /openbmc/linux/arch/sparc/kernel/perf_event.c (revision 54525552)
1 /* Performance event support for sparc64.
2  *
3  * Copyright (C) 2009, 2010 David S. Miller <davem@davemloft.net>
4  *
5  * This code is based almost entirely upon the x86 perf event
6  * code, which is:
7  *
8  *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
9  *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
10  *  Copyright (C) 2009 Jaswinder Singh Rajput
11  *  Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
12  *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
13  */
14 
15 #include <linux/perf_event.h>
16 #include <linux/kprobes.h>
17 #include <linux/ftrace.h>
18 #include <linux/kernel.h>
19 #include <linux/kdebug.h>
20 #include <linux/mutex.h>
21 
22 #include <asm/stacktrace.h>
23 #include <asm/cpudata.h>
24 #include <asm/uaccess.h>
25 #include <asm/atomic.h>
26 #include <asm/nmi.h>
27 #include <asm/pcr.h>
28 
29 #include "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 
62 struct 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 };
99 DEFINE_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  */
105 struct 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.  */
114 static unsigned long perf_event_encode(const struct perf_event_map *pmap)
115 {
116 	return ((unsigned long) pmap->encoding << 16) | pmap->pic_mask;
117 }
118 
119 static u8 perf_event_get_msk(unsigned long val)
120 {
121 	return val & 0xff;
122 }
123 
124 static 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 
134 typedef 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 
139 struct 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 
152 static 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 
159 static const struct perf_event_map *ultra3_event_map(int event_id)
160 {
161 	return &ultra3_perfmon_event_map[event_id];
162 }
163 
164 static 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 };
250 
251 static const struct sparc_pmu ultra3_pmu = {
252 	.event_map	= ultra3_event_map,
253 	.cache_map	= &ultra3_cache_map,
254 	.max_events	= ARRAY_SIZE(ultra3_perfmon_event_map),
255 	.upper_shift	= 11,
256 	.lower_shift	= 4,
257 	.event_mask	= 0x3f,
258 	.upper_nop	= 0x1c,
259 	.lower_nop	= 0x14,
260 };
261 
262 /* Niagara1 is very limited.  The upper PIC is hard-locked to count
263  * only instructions, so it is free running which creates all kinds of
264  * problems.  Some hardware designs make one wonder if the creator
265  * even looked at how this stuff gets used by software.
266  */
267 static const struct perf_event_map niagara1_perfmon_event_map[] = {
268 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00, PIC_UPPER },
269 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x00, PIC_UPPER },
270 	[PERF_COUNT_HW_CACHE_REFERENCES] = { 0, PIC_NONE },
271 	[PERF_COUNT_HW_CACHE_MISSES] = { 0x03, PIC_LOWER },
272 };
273 
274 static const struct perf_event_map *niagara1_event_map(int event_id)
275 {
276 	return &niagara1_perfmon_event_map[event_id];
277 }
278 
279 static const cache_map_t niagara1_cache_map = {
280 [C(L1D)] = {
281 	[C(OP_READ)] = {
282 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
283 		[C(RESULT_MISS)] = { 0x03, PIC_LOWER, },
284 	},
285 	[C(OP_WRITE)] = {
286 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
287 		[C(RESULT_MISS)] = { 0x03, PIC_LOWER, },
288 	},
289 	[C(OP_PREFETCH)] = {
290 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
291 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
292 	},
293 },
294 [C(L1I)] = {
295 	[C(OP_READ)] = {
296 		[C(RESULT_ACCESS)] = { 0x00, PIC_UPPER },
297 		[C(RESULT_MISS)] = { 0x02, PIC_LOWER, },
298 	},
299 	[ C(OP_WRITE) ] = {
300 		[ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
301 		[ C(RESULT_MISS)   ] = { CACHE_OP_NONSENSE },
302 	},
303 	[ C(OP_PREFETCH) ] = {
304 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
305 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
306 	},
307 },
308 [C(LL)] = {
309 	[C(OP_READ)] = {
310 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
311 		[C(RESULT_MISS)] = { 0x07, PIC_LOWER, },
312 	},
313 	[C(OP_WRITE)] = {
314 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
315 		[C(RESULT_MISS)] = { 0x07, PIC_LOWER, },
316 	},
317 	[C(OP_PREFETCH)] = {
318 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
319 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
320 	},
321 },
322 [C(DTLB)] = {
323 	[C(OP_READ)] = {
324 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
325 		[C(RESULT_MISS)] = { 0x05, PIC_LOWER, },
326 	},
327 	[ C(OP_WRITE) ] = {
328 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
329 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
330 	},
331 	[ C(OP_PREFETCH) ] = {
332 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
333 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
334 	},
335 },
336 [C(ITLB)] = {
337 	[C(OP_READ)] = {
338 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
339 		[C(RESULT_MISS)] = { 0x04, 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(BPU)] = {
351 	[C(OP_READ)] = {
352 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
353 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
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 };
365 
366 static const struct sparc_pmu niagara1_pmu = {
367 	.event_map	= niagara1_event_map,
368 	.cache_map	= &niagara1_cache_map,
369 	.max_events	= ARRAY_SIZE(niagara1_perfmon_event_map),
370 	.upper_shift	= 0,
371 	.lower_shift	= 4,
372 	.event_mask	= 0x7,
373 	.upper_nop	= 0x0,
374 	.lower_nop	= 0x0,
375 };
376 
377 static const struct perf_event_map niagara2_perfmon_event_map[] = {
378 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x02ff, PIC_UPPER | PIC_LOWER },
379 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x02ff, PIC_UPPER | PIC_LOWER },
380 	[PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0208, PIC_UPPER | PIC_LOWER },
381 	[PERF_COUNT_HW_CACHE_MISSES] = { 0x0302, PIC_UPPER | PIC_LOWER },
382 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x0201, PIC_UPPER | PIC_LOWER },
383 	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x0202, PIC_UPPER | PIC_LOWER },
384 };
385 
386 static const struct perf_event_map *niagara2_event_map(int event_id)
387 {
388 	return &niagara2_perfmon_event_map[event_id];
389 }
390 
391 static const cache_map_t niagara2_cache_map = {
392 [C(L1D)] = {
393 	[C(OP_READ)] = {
394 		[C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, },
395 		[C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, },
396 	},
397 	[C(OP_WRITE)] = {
398 		[C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, },
399 		[C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, },
400 	},
401 	[C(OP_PREFETCH)] = {
402 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
403 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
404 	},
405 },
406 [C(L1I)] = {
407 	[C(OP_READ)] = {
408 		[C(RESULT_ACCESS)] = { 0x02ff, PIC_UPPER | PIC_LOWER, },
409 		[C(RESULT_MISS)] = { 0x0301, PIC_UPPER | PIC_LOWER, },
410 	},
411 	[ C(OP_WRITE) ] = {
412 		[ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
413 		[ C(RESULT_MISS)   ] = { CACHE_OP_NONSENSE },
414 	},
415 	[ C(OP_PREFETCH) ] = {
416 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
417 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
418 	},
419 },
420 [C(LL)] = {
421 	[C(OP_READ)] = {
422 		[C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, },
423 		[C(RESULT_MISS)] = { 0x0330, PIC_UPPER | PIC_LOWER, },
424 	},
425 	[C(OP_WRITE)] = {
426 		[C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, },
427 		[C(RESULT_MISS)] = { 0x0320, 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(DTLB)] = {
435 	[C(OP_READ)] = {
436 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
437 		[C(RESULT_MISS)] = { 0x0b08, PIC_UPPER | PIC_LOWER, },
438 	},
439 	[ C(OP_WRITE) ] = {
440 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
441 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
442 	},
443 	[ C(OP_PREFETCH) ] = {
444 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
445 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
446 	},
447 },
448 [C(ITLB)] = {
449 	[C(OP_READ)] = {
450 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
451 		[C(RESULT_MISS)] = { 0xb04, PIC_UPPER | PIC_LOWER, },
452 	},
453 	[ C(OP_WRITE) ] = {
454 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
455 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
456 	},
457 	[ C(OP_PREFETCH) ] = {
458 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
459 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
460 	},
461 },
462 [C(BPU)] = {
463 	[C(OP_READ)] = {
464 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
465 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
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 };
477 
478 static const struct sparc_pmu niagara2_pmu = {
479 	.event_map	= niagara2_event_map,
480 	.cache_map	= &niagara2_cache_map,
481 	.max_events	= ARRAY_SIZE(niagara2_perfmon_event_map),
482 	.upper_shift	= 19,
483 	.lower_shift	= 6,
484 	.event_mask	= 0xfff,
485 	.hv_bit		= 0x8,
486 	.irq_bit	= 0x30,
487 	.upper_nop	= 0x220,
488 	.lower_nop	= 0x220,
489 };
490 
491 static const struct sparc_pmu *sparc_pmu __read_mostly;
492 
493 static u64 event_encoding(u64 event_id, int idx)
494 {
495 	if (idx == PIC_UPPER_INDEX)
496 		event_id <<= sparc_pmu->upper_shift;
497 	else
498 		event_id <<= sparc_pmu->lower_shift;
499 	return event_id;
500 }
501 
502 static u64 mask_for_index(int idx)
503 {
504 	return event_encoding(sparc_pmu->event_mask, idx);
505 }
506 
507 static u64 nop_for_index(int idx)
508 {
509 	return event_encoding(idx == PIC_UPPER_INDEX ?
510 			      sparc_pmu->upper_nop :
511 			      sparc_pmu->lower_nop, idx);
512 }
513 
514 static inline void sparc_pmu_enable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx)
515 {
516 	u64 val, mask = mask_for_index(idx);
517 
518 	val = cpuc->pcr;
519 	val &= ~mask;
520 	val |= hwc->config;
521 	cpuc->pcr = val;
522 
523 	pcr_ops->write(cpuc->pcr);
524 }
525 
526 static inline void sparc_pmu_disable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx)
527 {
528 	u64 mask = mask_for_index(idx);
529 	u64 nop = nop_for_index(idx);
530 	u64 val;
531 
532 	val = cpuc->pcr;
533 	val &= ~mask;
534 	val |= nop;
535 	cpuc->pcr = val;
536 
537 	pcr_ops->write(cpuc->pcr);
538 }
539 
540 static u32 read_pmc(int idx)
541 {
542 	u64 val;
543 
544 	read_pic(val);
545 	if (idx == PIC_UPPER_INDEX)
546 		val >>= 32;
547 
548 	return val & 0xffffffff;
549 }
550 
551 static void write_pmc(int idx, u64 val)
552 {
553 	u64 shift, mask, pic;
554 
555 	shift = 0;
556 	if (idx == PIC_UPPER_INDEX)
557 		shift = 32;
558 
559 	mask = ((u64) 0xffffffff) << shift;
560 	val <<= shift;
561 
562 	read_pic(pic);
563 	pic &= ~mask;
564 	pic |= val;
565 	write_pic(pic);
566 }
567 
568 static u64 sparc_perf_event_update(struct perf_event *event,
569 				   struct hw_perf_event *hwc, int idx)
570 {
571 	int shift = 64 - 32;
572 	u64 prev_raw_count, new_raw_count;
573 	s64 delta;
574 
575 again:
576 	prev_raw_count = local64_read(&hwc->prev_count);
577 	new_raw_count = read_pmc(idx);
578 
579 	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
580 			     new_raw_count) != prev_raw_count)
581 		goto again;
582 
583 	delta = (new_raw_count << shift) - (prev_raw_count << shift);
584 	delta >>= shift;
585 
586 	local64_add(delta, &event->count);
587 	local64_sub(delta, &hwc->period_left);
588 
589 	return new_raw_count;
590 }
591 
592 static int sparc_perf_event_set_period(struct perf_event *event,
593 				       struct hw_perf_event *hwc, int idx)
594 {
595 	s64 left = local64_read(&hwc->period_left);
596 	s64 period = hwc->sample_period;
597 	int ret = 0;
598 
599 	if (unlikely(left <= -period)) {
600 		left = period;
601 		local64_set(&hwc->period_left, left);
602 		hwc->last_period = period;
603 		ret = 1;
604 	}
605 
606 	if (unlikely(left <= 0)) {
607 		left += period;
608 		local64_set(&hwc->period_left, left);
609 		hwc->last_period = period;
610 		ret = 1;
611 	}
612 	if (left > MAX_PERIOD)
613 		left = MAX_PERIOD;
614 
615 	local64_set(&hwc->prev_count, (u64)-left);
616 
617 	write_pmc(idx, (u64)(-left) & 0xffffffff);
618 
619 	perf_event_update_userpage(event);
620 
621 	return ret;
622 }
623 
624 /* If performance event entries have been added, move existing
625  * events around (if necessary) and then assign new entries to
626  * counters.
627  */
628 static u64 maybe_change_configuration(struct cpu_hw_events *cpuc, u64 pcr)
629 {
630 	int i;
631 
632 	if (!cpuc->n_added)
633 		goto out;
634 
635 	/* Read in the counters which are moving.  */
636 	for (i = 0; i < cpuc->n_events; i++) {
637 		struct perf_event *cp = cpuc->event[i];
638 
639 		if (cpuc->current_idx[i] != PIC_NO_INDEX &&
640 		    cpuc->current_idx[i] != cp->hw.idx) {
641 			sparc_perf_event_update(cp, &cp->hw,
642 						cpuc->current_idx[i]);
643 			cpuc->current_idx[i] = PIC_NO_INDEX;
644 		}
645 	}
646 
647 	/* Assign to counters all unassigned events.  */
648 	for (i = 0; i < cpuc->n_events; i++) {
649 		struct perf_event *cp = cpuc->event[i];
650 		struct hw_perf_event *hwc = &cp->hw;
651 		int idx = hwc->idx;
652 		u64 enc;
653 
654 		if (cpuc->current_idx[i] != PIC_NO_INDEX)
655 			continue;
656 
657 		sparc_perf_event_set_period(cp, hwc, idx);
658 		cpuc->current_idx[i] = idx;
659 
660 		enc = perf_event_get_enc(cpuc->events[i]);
661 		pcr &= ~mask_for_index(idx);
662 		if (hwc->state & PERF_HES_STOPPED)
663 			pcr |= nop_for_index(idx);
664 		else
665 			pcr |= event_encoding(enc, idx);
666 	}
667 out:
668 	return pcr;
669 }
670 
671 static void sparc_pmu_enable(struct pmu *pmu)
672 {
673 	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
674 	u64 pcr;
675 
676 	if (cpuc->enabled)
677 		return;
678 
679 	cpuc->enabled = 1;
680 	barrier();
681 
682 	pcr = cpuc->pcr;
683 	if (!cpuc->n_events) {
684 		pcr = 0;
685 	} else {
686 		pcr = maybe_change_configuration(cpuc, pcr);
687 
688 		/* We require that all of the events have the same
689 		 * configuration, so just fetch the settings from the
690 		 * first entry.
691 		 */
692 		cpuc->pcr = pcr | cpuc->event[0]->hw.config_base;
693 	}
694 
695 	pcr_ops->write(cpuc->pcr);
696 }
697 
698 static void sparc_pmu_disable(struct pmu *pmu)
699 {
700 	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
701 	u64 val;
702 
703 	if (!cpuc->enabled)
704 		return;
705 
706 	cpuc->enabled = 0;
707 	cpuc->n_added = 0;
708 
709 	val = cpuc->pcr;
710 	val &= ~(PCR_UTRACE | PCR_STRACE |
711 		 sparc_pmu->hv_bit | sparc_pmu->irq_bit);
712 	cpuc->pcr = val;
713 
714 	pcr_ops->write(cpuc->pcr);
715 }
716 
717 static int active_event_index(struct cpu_hw_events *cpuc,
718 			      struct perf_event *event)
719 {
720 	int i;
721 
722 	for (i = 0; i < cpuc->n_events; i++) {
723 		if (cpuc->event[i] == event)
724 			break;
725 	}
726 	BUG_ON(i == cpuc->n_events);
727 	return cpuc->current_idx[i];
728 }
729 
730 static void sparc_pmu_start(struct perf_event *event, int flags)
731 {
732 	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
733 	int idx = active_event_index(cpuc, event);
734 
735 	if (flags & PERF_EF_RELOAD) {
736 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
737 		sparc_perf_event_set_period(event, &event->hw, idx);
738 	}
739 
740 	event->hw.state = 0;
741 
742 	sparc_pmu_enable_event(cpuc, &event->hw, idx);
743 }
744 
745 static void sparc_pmu_stop(struct perf_event *event, int flags)
746 {
747 	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
748 	int idx = active_event_index(cpuc, event);
749 
750 	if (!(event->hw.state & PERF_HES_STOPPED)) {
751 		sparc_pmu_disable_event(cpuc, &event->hw, idx);
752 		event->hw.state |= PERF_HES_STOPPED;
753 	}
754 
755 	if (!(event->hw.state & PERF_HES_UPTODATE) && (flags & PERF_EF_UPDATE)) {
756 		sparc_perf_event_update(event, &event->hw, idx);
757 		event->hw.state |= PERF_HES_UPTODATE;
758 	}
759 }
760 
761 static void sparc_pmu_del(struct perf_event *event, int _flags)
762 {
763 	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
764 	unsigned long flags;
765 	int i;
766 
767 	local_irq_save(flags);
768 	perf_pmu_disable(event->pmu);
769 
770 	for (i = 0; i < cpuc->n_events; i++) {
771 		if (event == cpuc->event[i]) {
772 			/* Absorb the final count and turn off the
773 			 * event.
774 			 */
775 			sparc_pmu_stop(event, PERF_EF_UPDATE);
776 
777 			/* Shift remaining entries down into
778 			 * the existing slot.
779 			 */
780 			while (++i < cpuc->n_events) {
781 				cpuc->event[i - 1] = cpuc->event[i];
782 				cpuc->events[i - 1] = cpuc->events[i];
783 				cpuc->current_idx[i - 1] =
784 					cpuc->current_idx[i];
785 			}
786 
787 			perf_event_update_userpage(event);
788 
789 			cpuc->n_events--;
790 			break;
791 		}
792 	}
793 
794 	perf_pmu_enable(event->pmu);
795 	local_irq_restore(flags);
796 }
797 
798 static void sparc_pmu_read(struct perf_event *event)
799 {
800 	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
801 	int idx = active_event_index(cpuc, event);
802 	struct hw_perf_event *hwc = &event->hw;
803 
804 	sparc_perf_event_update(event, hwc, idx);
805 }
806 
807 static atomic_t active_events = ATOMIC_INIT(0);
808 static DEFINE_MUTEX(pmc_grab_mutex);
809 
810 static void perf_stop_nmi_watchdog(void *unused)
811 {
812 	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
813 
814 	stop_nmi_watchdog(NULL);
815 	cpuc->pcr = pcr_ops->read();
816 }
817 
818 void perf_event_grab_pmc(void)
819 {
820 	if (atomic_inc_not_zero(&active_events))
821 		return;
822 
823 	mutex_lock(&pmc_grab_mutex);
824 	if (atomic_read(&active_events) == 0) {
825 		if (atomic_read(&nmi_active) > 0) {
826 			on_each_cpu(perf_stop_nmi_watchdog, NULL, 1);
827 			BUG_ON(atomic_read(&nmi_active) != 0);
828 		}
829 		atomic_inc(&active_events);
830 	}
831 	mutex_unlock(&pmc_grab_mutex);
832 }
833 
834 void perf_event_release_pmc(void)
835 {
836 	if (atomic_dec_and_mutex_lock(&active_events, &pmc_grab_mutex)) {
837 		if (atomic_read(&nmi_active) == 0)
838 			on_each_cpu(start_nmi_watchdog, NULL, 1);
839 		mutex_unlock(&pmc_grab_mutex);
840 	}
841 }
842 
843 static const struct perf_event_map *sparc_map_cache_event(u64 config)
844 {
845 	unsigned int cache_type, cache_op, cache_result;
846 	const struct perf_event_map *pmap;
847 
848 	if (!sparc_pmu->cache_map)
849 		return ERR_PTR(-ENOENT);
850 
851 	cache_type = (config >>  0) & 0xff;
852 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
853 		return ERR_PTR(-EINVAL);
854 
855 	cache_op = (config >>  8) & 0xff;
856 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
857 		return ERR_PTR(-EINVAL);
858 
859 	cache_result = (config >> 16) & 0xff;
860 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
861 		return ERR_PTR(-EINVAL);
862 
863 	pmap = &((*sparc_pmu->cache_map)[cache_type][cache_op][cache_result]);
864 
865 	if (pmap->encoding == CACHE_OP_UNSUPPORTED)
866 		return ERR_PTR(-ENOENT);
867 
868 	if (pmap->encoding == CACHE_OP_NONSENSE)
869 		return ERR_PTR(-EINVAL);
870 
871 	return pmap;
872 }
873 
874 static void hw_perf_event_destroy(struct perf_event *event)
875 {
876 	perf_event_release_pmc();
877 }
878 
879 /* Make sure all events can be scheduled into the hardware at
880  * the same time.  This is simplified by the fact that we only
881  * need to support 2 simultaneous HW events.
882  *
883  * As a side effect, the evts[]->hw.idx values will be assigned
884  * on success.  These are pending indexes.  When the events are
885  * actually programmed into the chip, these values will propagate
886  * to the per-cpu cpuc->current_idx[] slots, see the code in
887  * maybe_change_configuration() for details.
888  */
889 static int sparc_check_constraints(struct perf_event **evts,
890 				   unsigned long *events, int n_ev)
891 {
892 	u8 msk0 = 0, msk1 = 0;
893 	int idx0 = 0;
894 
895 	/* This case is possible when we are invoked from
896 	 * hw_perf_group_sched_in().
897 	 */
898 	if (!n_ev)
899 		return 0;
900 
901 	if (n_ev > MAX_HWEVENTS)
902 		return -1;
903 
904 	msk0 = perf_event_get_msk(events[0]);
905 	if (n_ev == 1) {
906 		if (msk0 & PIC_LOWER)
907 			idx0 = 1;
908 		goto success;
909 	}
910 	BUG_ON(n_ev != 2);
911 	msk1 = perf_event_get_msk(events[1]);
912 
913 	/* If both events can go on any counter, OK.  */
914 	if (msk0 == (PIC_UPPER | PIC_LOWER) &&
915 	    msk1 == (PIC_UPPER | PIC_LOWER))
916 		goto success;
917 
918 	/* If one event is limited to a specific counter,
919 	 * and the other can go on both, OK.
920 	 */
921 	if ((msk0 == PIC_UPPER || msk0 == PIC_LOWER) &&
922 	    msk1 == (PIC_UPPER | PIC_LOWER)) {
923 		if (msk0 & PIC_LOWER)
924 			idx0 = 1;
925 		goto success;
926 	}
927 
928 	if ((msk1 == PIC_UPPER || msk1 == PIC_LOWER) &&
929 	    msk0 == (PIC_UPPER | PIC_LOWER)) {
930 		if (msk1 & PIC_UPPER)
931 			idx0 = 1;
932 		goto success;
933 	}
934 
935 	/* If the events are fixed to different counters, OK.  */
936 	if ((msk0 == PIC_UPPER && msk1 == PIC_LOWER) ||
937 	    (msk0 == PIC_LOWER && msk1 == PIC_UPPER)) {
938 		if (msk0 & PIC_LOWER)
939 			idx0 = 1;
940 		goto success;
941 	}
942 
943 	/* Otherwise, there is a conflict.  */
944 	return -1;
945 
946 success:
947 	evts[0]->hw.idx = idx0;
948 	if (n_ev == 2)
949 		evts[1]->hw.idx = idx0 ^ 1;
950 	return 0;
951 }
952 
953 static int check_excludes(struct perf_event **evts, int n_prev, int n_new)
954 {
955 	int eu = 0, ek = 0, eh = 0;
956 	struct perf_event *event;
957 	int i, n, first;
958 
959 	n = n_prev + n_new;
960 	if (n <= 1)
961 		return 0;
962 
963 	first = 1;
964 	for (i = 0; i < n; i++) {
965 		event = evts[i];
966 		if (first) {
967 			eu = event->attr.exclude_user;
968 			ek = event->attr.exclude_kernel;
969 			eh = event->attr.exclude_hv;
970 			first = 0;
971 		} else if (event->attr.exclude_user != eu ||
972 			   event->attr.exclude_kernel != ek ||
973 			   event->attr.exclude_hv != eh) {
974 			return -EAGAIN;
975 		}
976 	}
977 
978 	return 0;
979 }
980 
981 static int collect_events(struct perf_event *group, int max_count,
982 			  struct perf_event *evts[], unsigned long *events,
983 			  int *current_idx)
984 {
985 	struct perf_event *event;
986 	int n = 0;
987 
988 	if (!is_software_event(group)) {
989 		if (n >= max_count)
990 			return -1;
991 		evts[n] = group;
992 		events[n] = group->hw.event_base;
993 		current_idx[n++] = PIC_NO_INDEX;
994 	}
995 	list_for_each_entry(event, &group->sibling_list, group_entry) {
996 		if (!is_software_event(event) &&
997 		    event->state != PERF_EVENT_STATE_OFF) {
998 			if (n >= max_count)
999 				return -1;
1000 			evts[n] = event;
1001 			events[n] = event->hw.event_base;
1002 			current_idx[n++] = PIC_NO_INDEX;
1003 		}
1004 	}
1005 	return n;
1006 }
1007 
1008 static int sparc_pmu_add(struct perf_event *event, int ef_flags)
1009 {
1010 	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1011 	int n0, ret = -EAGAIN;
1012 	unsigned long flags;
1013 
1014 	local_irq_save(flags);
1015 	perf_pmu_disable(event->pmu);
1016 
1017 	n0 = cpuc->n_events;
1018 	if (n0 >= MAX_HWEVENTS)
1019 		goto out;
1020 
1021 	cpuc->event[n0] = event;
1022 	cpuc->events[n0] = event->hw.event_base;
1023 	cpuc->current_idx[n0] = PIC_NO_INDEX;
1024 
1025 	event->hw.state = PERF_HES_UPTODATE;
1026 	if (!(ef_flags & PERF_EF_START))
1027 		event->hw.state |= PERF_HES_STOPPED;
1028 
1029 	/*
1030 	 * If group events scheduling transaction was started,
1031 	 * skip the schedulability test here, it will be performed
1032 	 * at commit time(->commit_txn) as a whole
1033 	 */
1034 	if (cpuc->group_flag & PERF_EVENT_TXN)
1035 		goto nocheck;
1036 
1037 	if (check_excludes(cpuc->event, n0, 1))
1038 		goto out;
1039 	if (sparc_check_constraints(cpuc->event, cpuc->events, n0 + 1))
1040 		goto out;
1041 
1042 nocheck:
1043 	cpuc->n_events++;
1044 	cpuc->n_added++;
1045 
1046 	ret = 0;
1047 out:
1048 	perf_pmu_enable(event->pmu);
1049 	local_irq_restore(flags);
1050 	return ret;
1051 }
1052 
1053 static int sparc_pmu_event_init(struct perf_event *event)
1054 {
1055 	struct perf_event_attr *attr = &event->attr;
1056 	struct perf_event *evts[MAX_HWEVENTS];
1057 	struct hw_perf_event *hwc = &event->hw;
1058 	unsigned long events[MAX_HWEVENTS];
1059 	int current_idx_dmy[MAX_HWEVENTS];
1060 	const struct perf_event_map *pmap;
1061 	int n;
1062 
1063 	if (atomic_read(&nmi_active) < 0)
1064 		return -ENODEV;
1065 
1066 	switch (attr->type) {
1067 	case PERF_TYPE_HARDWARE:
1068 		if (attr->config >= sparc_pmu->max_events)
1069 			return -EINVAL;
1070 		pmap = sparc_pmu->event_map(attr->config);
1071 		break;
1072 
1073 	case PERF_TYPE_HW_CACHE:
1074 		pmap = sparc_map_cache_event(attr->config);
1075 		if (IS_ERR(pmap))
1076 			return PTR_ERR(pmap);
1077 		break;
1078 
1079 	case PERF_TYPE_RAW:
1080 		pmap = NULL;
1081 		break;
1082 
1083 	default:
1084 		return -ENOENT;
1085 
1086 	}
1087 
1088 	if (pmap) {
1089 		hwc->event_base = perf_event_encode(pmap);
1090 	} else {
1091 		/*
1092 		 * User gives us "(encoding << 16) | pic_mask" for
1093 		 * PERF_TYPE_RAW events.
1094 		 */
1095 		hwc->event_base = attr->config;
1096 	}
1097 
1098 	/* We save the enable bits in the config_base.  */
1099 	hwc->config_base = sparc_pmu->irq_bit;
1100 	if (!attr->exclude_user)
1101 		hwc->config_base |= PCR_UTRACE;
1102 	if (!attr->exclude_kernel)
1103 		hwc->config_base |= PCR_STRACE;
1104 	if (!attr->exclude_hv)
1105 		hwc->config_base |= sparc_pmu->hv_bit;
1106 
1107 	n = 0;
1108 	if (event->group_leader != event) {
1109 		n = collect_events(event->group_leader,
1110 				   MAX_HWEVENTS - 1,
1111 				   evts, events, current_idx_dmy);
1112 		if (n < 0)
1113 			return -EINVAL;
1114 	}
1115 	events[n] = hwc->event_base;
1116 	evts[n] = event;
1117 
1118 	if (check_excludes(evts, n, 1))
1119 		return -EINVAL;
1120 
1121 	if (sparc_check_constraints(evts, events, n + 1))
1122 		return -EINVAL;
1123 
1124 	hwc->idx = PIC_NO_INDEX;
1125 
1126 	/* Try to do all error checking before this point, as unwinding
1127 	 * state after grabbing the PMC is difficult.
1128 	 */
1129 	perf_event_grab_pmc();
1130 	event->destroy = hw_perf_event_destroy;
1131 
1132 	if (!hwc->sample_period) {
1133 		hwc->sample_period = MAX_PERIOD;
1134 		hwc->last_period = hwc->sample_period;
1135 		local64_set(&hwc->period_left, hwc->sample_period);
1136 	}
1137 
1138 	return 0;
1139 }
1140 
1141 /*
1142  * Start group events scheduling transaction
1143  * Set the flag to make pmu::enable() not perform the
1144  * schedulability test, it will be performed at commit time
1145  */
1146 static void sparc_pmu_start_txn(struct pmu *pmu)
1147 {
1148 	struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1149 
1150 	perf_pmu_disable(pmu);
1151 	cpuhw->group_flag |= PERF_EVENT_TXN;
1152 }
1153 
1154 /*
1155  * Stop group events scheduling transaction
1156  * Clear the flag and pmu::enable() will perform the
1157  * schedulability test.
1158  */
1159 static void sparc_pmu_cancel_txn(struct pmu *pmu)
1160 {
1161 	struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1162 
1163 	cpuhw->group_flag &= ~PERF_EVENT_TXN;
1164 	perf_pmu_enable(pmu);
1165 }
1166 
1167 /*
1168  * Commit group events scheduling transaction
1169  * Perform the group schedulability test as a whole
1170  * Return 0 if success
1171  */
1172 static int sparc_pmu_commit_txn(struct pmu *pmu)
1173 {
1174 	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1175 	int n;
1176 
1177 	if (!sparc_pmu)
1178 		return -EINVAL;
1179 
1180 	cpuc = &__get_cpu_var(cpu_hw_events);
1181 	n = cpuc->n_events;
1182 	if (check_excludes(cpuc->event, 0, n))
1183 		return -EINVAL;
1184 	if (sparc_check_constraints(cpuc->event, cpuc->events, n))
1185 		return -EAGAIN;
1186 
1187 	cpuc->group_flag &= ~PERF_EVENT_TXN;
1188 	perf_pmu_enable(pmu);
1189 	return 0;
1190 }
1191 
1192 static struct pmu pmu = {
1193 	.pmu_enable	= sparc_pmu_enable,
1194 	.pmu_disable	= sparc_pmu_disable,
1195 	.event_init	= sparc_pmu_event_init,
1196 	.add		= sparc_pmu_add,
1197 	.del		= sparc_pmu_del,
1198 	.start		= sparc_pmu_start,
1199 	.stop		= sparc_pmu_stop,
1200 	.read		= sparc_pmu_read,
1201 	.start_txn	= sparc_pmu_start_txn,
1202 	.cancel_txn	= sparc_pmu_cancel_txn,
1203 	.commit_txn	= sparc_pmu_commit_txn,
1204 };
1205 
1206 void perf_event_print_debug(void)
1207 {
1208 	unsigned long flags;
1209 	u64 pcr, pic;
1210 	int cpu;
1211 
1212 	if (!sparc_pmu)
1213 		return;
1214 
1215 	local_irq_save(flags);
1216 
1217 	cpu = smp_processor_id();
1218 
1219 	pcr = pcr_ops->read();
1220 	read_pic(pic);
1221 
1222 	pr_info("\n");
1223 	pr_info("CPU#%d: PCR[%016llx] PIC[%016llx]\n",
1224 		cpu, pcr, pic);
1225 
1226 	local_irq_restore(flags);
1227 }
1228 
1229 static int __kprobes perf_event_nmi_handler(struct notifier_block *self,
1230 					    unsigned long cmd, void *__args)
1231 {
1232 	struct die_args *args = __args;
1233 	struct perf_sample_data data;
1234 	struct cpu_hw_events *cpuc;
1235 	struct pt_regs *regs;
1236 	int i;
1237 
1238 	if (!atomic_read(&active_events))
1239 		return NOTIFY_DONE;
1240 
1241 	switch (cmd) {
1242 	case DIE_NMI:
1243 		break;
1244 
1245 	default:
1246 		return NOTIFY_DONE;
1247 	}
1248 
1249 	regs = args->regs;
1250 
1251 	perf_sample_data_init(&data, 0);
1252 
1253 	cpuc = &__get_cpu_var(cpu_hw_events);
1254 
1255 	/* If the PMU has the TOE IRQ enable bits, we need to do a
1256 	 * dummy write to the %pcr to clear the overflow bits and thus
1257 	 * the interrupt.
1258 	 *
1259 	 * Do this before we peek at the counters to determine
1260 	 * overflow so we don't lose any events.
1261 	 */
1262 	if (sparc_pmu->irq_bit)
1263 		pcr_ops->write(cpuc->pcr);
1264 
1265 	for (i = 0; i < cpuc->n_events; i++) {
1266 		struct perf_event *event = cpuc->event[i];
1267 		int idx = cpuc->current_idx[i];
1268 		struct hw_perf_event *hwc;
1269 		u64 val;
1270 
1271 		hwc = &event->hw;
1272 		val = sparc_perf_event_update(event, hwc, idx);
1273 		if (val & (1ULL << 31))
1274 			continue;
1275 
1276 		data.period = event->hw.last_period;
1277 		if (!sparc_perf_event_set_period(event, hwc, idx))
1278 			continue;
1279 
1280 		if (perf_event_overflow(event, 1, &data, regs))
1281 			sparc_pmu_stop(event, 0);
1282 	}
1283 
1284 	return NOTIFY_STOP;
1285 }
1286 
1287 static __read_mostly struct notifier_block perf_event_nmi_notifier = {
1288 	.notifier_call		= perf_event_nmi_handler,
1289 };
1290 
1291 static bool __init supported_pmu(void)
1292 {
1293 	if (!strcmp(sparc_pmu_type, "ultra3") ||
1294 	    !strcmp(sparc_pmu_type, "ultra3+") ||
1295 	    !strcmp(sparc_pmu_type, "ultra3i") ||
1296 	    !strcmp(sparc_pmu_type, "ultra4+")) {
1297 		sparc_pmu = &ultra3_pmu;
1298 		return true;
1299 	}
1300 	if (!strcmp(sparc_pmu_type, "niagara")) {
1301 		sparc_pmu = &niagara1_pmu;
1302 		return true;
1303 	}
1304 	if (!strcmp(sparc_pmu_type, "niagara2")) {
1305 		sparc_pmu = &niagara2_pmu;
1306 		return true;
1307 	}
1308 	return false;
1309 }
1310 
1311 int __init init_hw_perf_events(void)
1312 {
1313 	pr_info("Performance events: ");
1314 
1315 	if (!supported_pmu()) {
1316 		pr_cont("No support for PMU type '%s'\n", sparc_pmu_type);
1317 		return 0;
1318 	}
1319 
1320 	pr_cont("Supported PMU type is '%s'\n", sparc_pmu_type);
1321 
1322 	perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1323 	register_die_notifier(&perf_event_nmi_notifier);
1324 
1325 	return 0;
1326 }
1327 early_initcall(init_hw_perf_events);
1328 
1329 void perf_callchain_kernel(struct perf_callchain_entry *entry,
1330 			   struct pt_regs *regs)
1331 {
1332 	unsigned long ksp, fp;
1333 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
1334 	int graph = 0;
1335 #endif
1336 
1337 	stack_trace_flush();
1338 
1339 	perf_callchain_store(entry, regs->tpc);
1340 
1341 	ksp = regs->u_regs[UREG_I6];
1342 	fp = ksp + STACK_BIAS;
1343 	do {
1344 		struct sparc_stackf *sf;
1345 		struct pt_regs *regs;
1346 		unsigned long pc;
1347 
1348 		if (!kstack_valid(current_thread_info(), fp))
1349 			break;
1350 
1351 		sf = (struct sparc_stackf *) fp;
1352 		regs = (struct pt_regs *) (sf + 1);
1353 
1354 		if (kstack_is_trap_frame(current_thread_info(), regs)) {
1355 			if (user_mode(regs))
1356 				break;
1357 			pc = regs->tpc;
1358 			fp = regs->u_regs[UREG_I6] + STACK_BIAS;
1359 		} else {
1360 			pc = sf->callers_pc;
1361 			fp = (unsigned long)sf->fp + STACK_BIAS;
1362 		}
1363 		perf_callchain_store(entry, pc);
1364 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
1365 		if ((pc + 8UL) == (unsigned long) &return_to_handler) {
1366 			int index = current->curr_ret_stack;
1367 			if (current->ret_stack && index >= graph) {
1368 				pc = current->ret_stack[index - graph].ret;
1369 				perf_callchain_store(entry, pc);
1370 				graph++;
1371 			}
1372 		}
1373 #endif
1374 	} while (entry->nr < PERF_MAX_STACK_DEPTH);
1375 }
1376 
1377 static void perf_callchain_user_64(struct perf_callchain_entry *entry,
1378 				   struct pt_regs *regs)
1379 {
1380 	unsigned long ufp;
1381 
1382 	perf_callchain_store(entry, regs->tpc);
1383 
1384 	ufp = regs->u_regs[UREG_I6] + STACK_BIAS;
1385 	do {
1386 		struct sparc_stackf *usf, sf;
1387 		unsigned long pc;
1388 
1389 		usf = (struct sparc_stackf *) ufp;
1390 		if (__copy_from_user_inatomic(&sf, usf, sizeof(sf)))
1391 			break;
1392 
1393 		pc = sf.callers_pc;
1394 		ufp = (unsigned long)sf.fp + STACK_BIAS;
1395 		perf_callchain_store(entry, pc);
1396 	} while (entry->nr < PERF_MAX_STACK_DEPTH);
1397 }
1398 
1399 static void perf_callchain_user_32(struct perf_callchain_entry *entry,
1400 				   struct pt_regs *regs)
1401 {
1402 	unsigned long ufp;
1403 
1404 	perf_callchain_store(entry, regs->tpc);
1405 
1406 	ufp = regs->u_regs[UREG_I6] & 0xffffffffUL;
1407 	do {
1408 		struct sparc_stackf32 *usf, sf;
1409 		unsigned long pc;
1410 
1411 		usf = (struct sparc_stackf32 *) ufp;
1412 		if (__copy_from_user_inatomic(&sf, usf, sizeof(sf)))
1413 			break;
1414 
1415 		pc = sf.callers_pc;
1416 		ufp = (unsigned long)sf.fp;
1417 		perf_callchain_store(entry, pc);
1418 	} while (entry->nr < PERF_MAX_STACK_DEPTH);
1419 }
1420 
1421 void
1422 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
1423 {
1424 	flushw_user();
1425 	if (test_thread_flag(TIF_32BIT))
1426 		perf_callchain_user_32(entry, regs);
1427 	else
1428 		perf_callchain_user_64(entry, regs);
1429 }
1430