xref: /openbmc/linux/arch/sparc/kernel/perf_event.c (revision a2818ee4)
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 <linux/sched/clock.h>
28 #include <asm/nmi.h>
29 #include <asm/pcr.h>
30 #include <asm/cacheflush.h>
31 
32 #include "kernel.h"
33 #include "kstack.h"
34 
35 /* Two classes of sparc64 chips currently exist.  All of which have
36  * 32-bit counters which can generate overflow interrupts on the
37  * transition from 0xffffffff to 0.
38  *
39  * All chips upto and including SPARC-T3 have two performance
40  * counters.  The two 32-bit counters are accessed in one go using a
41  * single 64-bit register.
42  *
43  * On these older chips both counters are controlled using a single
44  * control register.  The only way to stop all sampling is to clear
45  * all of the context (user, supervisor, hypervisor) sampling enable
46  * bits.  But these bits apply to both counters, thus the two counters
47  * can't be enabled/disabled individually.
48  *
49  * Furthermore, the control register on these older chips have two
50  * event fields, one for each of the two counters.  It's thus nearly
51  * impossible to have one counter going while keeping the other one
52  * stopped.  Therefore it is possible to get overflow interrupts for
53  * counters not currently "in use" and that condition must be checked
54  * in the overflow interrupt handler.
55  *
56  * So we use a hack, in that we program inactive counters with the
57  * "sw_count0" and "sw_count1" events.  These count how many times
58  * the instruction "sethi %hi(0xfc000), %g0" is executed.  It's an
59  * unusual way to encode a NOP and therefore will not trigger in
60  * normal code.
61  *
62  * Starting with SPARC-T4 we have one control register per counter.
63  * And the counters are stored in individual registers.  The registers
64  * for the counters are 64-bit but only a 32-bit counter is
65  * implemented.  The event selections on SPARC-T4 lack any
66  * restrictions, therefore we can elide all of the complicated
67  * conflict resolution code we have for SPARC-T3 and earlier chips.
68  */
69 
70 #define MAX_HWEVENTS			4
71 #define MAX_PCRS			4
72 #define MAX_PERIOD			((1UL << 32) - 1)
73 
74 #define PIC_UPPER_INDEX			0
75 #define PIC_LOWER_INDEX			1
76 #define PIC_NO_INDEX			-1
77 
78 struct cpu_hw_events {
79 	/* Number of events currently scheduled onto this cpu.
80 	 * This tells how many entries in the arrays below
81 	 * are valid.
82 	 */
83 	int			n_events;
84 
85 	/* Number of new events added since the last hw_perf_disable().
86 	 * This works because the perf event layer always adds new
87 	 * events inside of a perf_{disable,enable}() sequence.
88 	 */
89 	int			n_added;
90 
91 	/* Array of events current scheduled on this cpu.  */
92 	struct perf_event	*event[MAX_HWEVENTS];
93 
94 	/* Array of encoded longs, specifying the %pcr register
95 	 * encoding and the mask of PIC counters this even can
96 	 * be scheduled on.  See perf_event_encode() et al.
97 	 */
98 	unsigned long		events[MAX_HWEVENTS];
99 
100 	/* The current counter index assigned to an event.  When the
101 	 * event hasn't been programmed into the cpu yet, this will
102 	 * hold PIC_NO_INDEX.  The event->hw.idx value tells us where
103 	 * we ought to schedule the event.
104 	 */
105 	int			current_idx[MAX_HWEVENTS];
106 
107 	/* Software copy of %pcr register(s) on this cpu.  */
108 	u64			pcr[MAX_HWEVENTS];
109 
110 	/* Enabled/disable state.  */
111 	int			enabled;
112 
113 	unsigned int		txn_flags;
114 };
115 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = { .enabled = 1, };
116 
117 /* An event map describes the characteristics of a performance
118  * counter event.  In particular it gives the encoding as well as
119  * a mask telling which counters the event can be measured on.
120  *
121  * The mask is unused on SPARC-T4 and later.
122  */
123 struct perf_event_map {
124 	u16	encoding;
125 	u8	pic_mask;
126 #define PIC_NONE	0x00
127 #define PIC_UPPER	0x01
128 #define PIC_LOWER	0x02
129 };
130 
131 /* Encode a perf_event_map entry into a long.  */
132 static unsigned long perf_event_encode(const struct perf_event_map *pmap)
133 {
134 	return ((unsigned long) pmap->encoding << 16) | pmap->pic_mask;
135 }
136 
137 static u8 perf_event_get_msk(unsigned long val)
138 {
139 	return val & 0xff;
140 }
141 
142 static u64 perf_event_get_enc(unsigned long val)
143 {
144 	return val >> 16;
145 }
146 
147 #define C(x) PERF_COUNT_HW_CACHE_##x
148 
149 #define CACHE_OP_UNSUPPORTED	0xfffe
150 #define CACHE_OP_NONSENSE	0xffff
151 
152 typedef struct perf_event_map cache_map_t
153 				[PERF_COUNT_HW_CACHE_MAX]
154 				[PERF_COUNT_HW_CACHE_OP_MAX]
155 				[PERF_COUNT_HW_CACHE_RESULT_MAX];
156 
157 struct sparc_pmu {
158 	const struct perf_event_map	*(*event_map)(int);
159 	const cache_map_t		*cache_map;
160 	int				max_events;
161 	u32				(*read_pmc)(int);
162 	void				(*write_pmc)(int, u64);
163 	int				upper_shift;
164 	int				lower_shift;
165 	int				event_mask;
166 	int				user_bit;
167 	int				priv_bit;
168 	int				hv_bit;
169 	int				irq_bit;
170 	int				upper_nop;
171 	int				lower_nop;
172 	unsigned int			flags;
173 #define SPARC_PMU_ALL_EXCLUDES_SAME	0x00000001
174 #define SPARC_PMU_HAS_CONFLICTS		0x00000002
175 	int				max_hw_events;
176 	int				num_pcrs;
177 	int				num_pic_regs;
178 };
179 
180 static u32 sparc_default_read_pmc(int idx)
181 {
182 	u64 val;
183 
184 	val = pcr_ops->read_pic(0);
185 	if (idx == PIC_UPPER_INDEX)
186 		val >>= 32;
187 
188 	return val & 0xffffffff;
189 }
190 
191 static void sparc_default_write_pmc(int idx, u64 val)
192 {
193 	u64 shift, mask, pic;
194 
195 	shift = 0;
196 	if (idx == PIC_UPPER_INDEX)
197 		shift = 32;
198 
199 	mask = ((u64) 0xffffffff) << shift;
200 	val <<= shift;
201 
202 	pic = pcr_ops->read_pic(0);
203 	pic &= ~mask;
204 	pic |= val;
205 	pcr_ops->write_pic(0, pic);
206 }
207 
208 static const struct perf_event_map ultra3_perfmon_event_map[] = {
209 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x0000, PIC_UPPER | PIC_LOWER },
210 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x0001, PIC_UPPER | PIC_LOWER },
211 	[PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0009, PIC_LOWER },
212 	[PERF_COUNT_HW_CACHE_MISSES] = { 0x0009, PIC_UPPER },
213 };
214 
215 static const struct perf_event_map *ultra3_event_map(int event_id)
216 {
217 	return &ultra3_perfmon_event_map[event_id];
218 }
219 
220 static const cache_map_t ultra3_cache_map = {
221 [C(L1D)] = {
222 	[C(OP_READ)] = {
223 		[C(RESULT_ACCESS)] = { 0x09, PIC_LOWER, },
224 		[C(RESULT_MISS)] = { 0x09, PIC_UPPER, },
225 	},
226 	[C(OP_WRITE)] = {
227 		[C(RESULT_ACCESS)] = { 0x0a, PIC_LOWER },
228 		[C(RESULT_MISS)] = { 0x0a, PIC_UPPER },
229 	},
230 	[C(OP_PREFETCH)] = {
231 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
232 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
233 	},
234 },
235 [C(L1I)] = {
236 	[C(OP_READ)] = {
237 		[C(RESULT_ACCESS)] = { 0x09, PIC_LOWER, },
238 		[C(RESULT_MISS)] = { 0x09, PIC_UPPER, },
239 	},
240 	[ C(OP_WRITE) ] = {
241 		[ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
242 		[ C(RESULT_MISS)   ] = { CACHE_OP_NONSENSE },
243 	},
244 	[ C(OP_PREFETCH) ] = {
245 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
246 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
247 	},
248 },
249 [C(LL)] = {
250 	[C(OP_READ)] = {
251 		[C(RESULT_ACCESS)] = { 0x0c, PIC_LOWER, },
252 		[C(RESULT_MISS)] = { 0x0c, PIC_UPPER, },
253 	},
254 	[C(OP_WRITE)] = {
255 		[C(RESULT_ACCESS)] = { 0x0c, PIC_LOWER },
256 		[C(RESULT_MISS)] = { 0x0c, PIC_UPPER },
257 	},
258 	[C(OP_PREFETCH)] = {
259 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
260 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
261 	},
262 },
263 [C(DTLB)] = {
264 	[C(OP_READ)] = {
265 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
266 		[C(RESULT_MISS)] = { 0x12, PIC_UPPER, },
267 	},
268 	[ C(OP_WRITE) ] = {
269 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
270 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
271 	},
272 	[ C(OP_PREFETCH) ] = {
273 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
274 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
275 	},
276 },
277 [C(ITLB)] = {
278 	[C(OP_READ)] = {
279 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
280 		[C(RESULT_MISS)] = { 0x11, PIC_UPPER, },
281 	},
282 	[ C(OP_WRITE) ] = {
283 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
284 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
285 	},
286 	[ C(OP_PREFETCH) ] = {
287 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
288 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
289 	},
290 },
291 [C(BPU)] = {
292 	[C(OP_READ)] = {
293 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
294 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
295 	},
296 	[ C(OP_WRITE) ] = {
297 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
298 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
299 	},
300 	[ C(OP_PREFETCH) ] = {
301 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
302 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
303 	},
304 },
305 [C(NODE)] = {
306 	[C(OP_READ)] = {
307 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
308 		[C(RESULT_MISS)  ] = { CACHE_OP_UNSUPPORTED },
309 	},
310 	[ C(OP_WRITE) ] = {
311 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
312 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
313 	},
314 	[ C(OP_PREFETCH) ] = {
315 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
316 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
317 	},
318 },
319 };
320 
321 static const struct sparc_pmu ultra3_pmu = {
322 	.event_map	= ultra3_event_map,
323 	.cache_map	= &ultra3_cache_map,
324 	.max_events	= ARRAY_SIZE(ultra3_perfmon_event_map),
325 	.read_pmc	= sparc_default_read_pmc,
326 	.write_pmc	= sparc_default_write_pmc,
327 	.upper_shift	= 11,
328 	.lower_shift	= 4,
329 	.event_mask	= 0x3f,
330 	.user_bit	= PCR_UTRACE,
331 	.priv_bit	= PCR_STRACE,
332 	.upper_nop	= 0x1c,
333 	.lower_nop	= 0x14,
334 	.flags		= (SPARC_PMU_ALL_EXCLUDES_SAME |
335 			   SPARC_PMU_HAS_CONFLICTS),
336 	.max_hw_events	= 2,
337 	.num_pcrs	= 1,
338 	.num_pic_regs	= 1,
339 };
340 
341 /* Niagara1 is very limited.  The upper PIC is hard-locked to count
342  * only instructions, so it is free running which creates all kinds of
343  * problems.  Some hardware designs make one wonder if the creator
344  * even looked at how this stuff gets used by software.
345  */
346 static const struct perf_event_map niagara1_perfmon_event_map[] = {
347 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00, PIC_UPPER },
348 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x00, PIC_UPPER },
349 	[PERF_COUNT_HW_CACHE_REFERENCES] = { 0, PIC_NONE },
350 	[PERF_COUNT_HW_CACHE_MISSES] = { 0x03, PIC_LOWER },
351 };
352 
353 static const struct perf_event_map *niagara1_event_map(int event_id)
354 {
355 	return &niagara1_perfmon_event_map[event_id];
356 }
357 
358 static const cache_map_t niagara1_cache_map = {
359 [C(L1D)] = {
360 	[C(OP_READ)] = {
361 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
362 		[C(RESULT_MISS)] = { 0x03, PIC_LOWER, },
363 	},
364 	[C(OP_WRITE)] = {
365 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
366 		[C(RESULT_MISS)] = { 0x03, PIC_LOWER, },
367 	},
368 	[C(OP_PREFETCH)] = {
369 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
370 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
371 	},
372 },
373 [C(L1I)] = {
374 	[C(OP_READ)] = {
375 		[C(RESULT_ACCESS)] = { 0x00, PIC_UPPER },
376 		[C(RESULT_MISS)] = { 0x02, PIC_LOWER, },
377 	},
378 	[ C(OP_WRITE) ] = {
379 		[ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
380 		[ C(RESULT_MISS)   ] = { CACHE_OP_NONSENSE },
381 	},
382 	[ C(OP_PREFETCH) ] = {
383 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
384 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
385 	},
386 },
387 [C(LL)] = {
388 	[C(OP_READ)] = {
389 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
390 		[C(RESULT_MISS)] = { 0x07, PIC_LOWER, },
391 	},
392 	[C(OP_WRITE)] = {
393 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
394 		[C(RESULT_MISS)] = { 0x07, PIC_LOWER, },
395 	},
396 	[C(OP_PREFETCH)] = {
397 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
398 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
399 	},
400 },
401 [C(DTLB)] = {
402 	[C(OP_READ)] = {
403 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
404 		[C(RESULT_MISS)] = { 0x05, PIC_LOWER, },
405 	},
406 	[ C(OP_WRITE) ] = {
407 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
408 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
409 	},
410 	[ C(OP_PREFETCH) ] = {
411 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
412 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
413 	},
414 },
415 [C(ITLB)] = {
416 	[C(OP_READ)] = {
417 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
418 		[C(RESULT_MISS)] = { 0x04, PIC_LOWER, },
419 	},
420 	[ C(OP_WRITE) ] = {
421 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
422 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
423 	},
424 	[ C(OP_PREFETCH) ] = {
425 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
426 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
427 	},
428 },
429 [C(BPU)] = {
430 	[C(OP_READ)] = {
431 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
432 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
433 	},
434 	[ C(OP_WRITE) ] = {
435 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
436 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
437 	},
438 	[ C(OP_PREFETCH) ] = {
439 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
440 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
441 	},
442 },
443 [C(NODE)] = {
444 	[C(OP_READ)] = {
445 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
446 		[C(RESULT_MISS)  ] = { CACHE_OP_UNSUPPORTED },
447 	},
448 	[ C(OP_WRITE) ] = {
449 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
450 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
451 	},
452 	[ C(OP_PREFETCH) ] = {
453 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
454 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
455 	},
456 },
457 };
458 
459 static const struct sparc_pmu niagara1_pmu = {
460 	.event_map	= niagara1_event_map,
461 	.cache_map	= &niagara1_cache_map,
462 	.max_events	= ARRAY_SIZE(niagara1_perfmon_event_map),
463 	.read_pmc	= sparc_default_read_pmc,
464 	.write_pmc	= sparc_default_write_pmc,
465 	.upper_shift	= 0,
466 	.lower_shift	= 4,
467 	.event_mask	= 0x7,
468 	.user_bit	= PCR_UTRACE,
469 	.priv_bit	= PCR_STRACE,
470 	.upper_nop	= 0x0,
471 	.lower_nop	= 0x0,
472 	.flags		= (SPARC_PMU_ALL_EXCLUDES_SAME |
473 			   SPARC_PMU_HAS_CONFLICTS),
474 	.max_hw_events	= 2,
475 	.num_pcrs	= 1,
476 	.num_pic_regs	= 1,
477 };
478 
479 static const struct perf_event_map niagara2_perfmon_event_map[] = {
480 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x02ff, PIC_UPPER | PIC_LOWER },
481 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x02ff, PIC_UPPER | PIC_LOWER },
482 	[PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0208, PIC_UPPER | PIC_LOWER },
483 	[PERF_COUNT_HW_CACHE_MISSES] = { 0x0302, PIC_UPPER | PIC_LOWER },
484 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x0201, PIC_UPPER | PIC_LOWER },
485 	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x0202, PIC_UPPER | PIC_LOWER },
486 };
487 
488 static const struct perf_event_map *niagara2_event_map(int event_id)
489 {
490 	return &niagara2_perfmon_event_map[event_id];
491 }
492 
493 static const cache_map_t niagara2_cache_map = {
494 [C(L1D)] = {
495 	[C(OP_READ)] = {
496 		[C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, },
497 		[C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, },
498 	},
499 	[C(OP_WRITE)] = {
500 		[C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, },
501 		[C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, },
502 	},
503 	[C(OP_PREFETCH)] = {
504 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
505 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
506 	},
507 },
508 [C(L1I)] = {
509 	[C(OP_READ)] = {
510 		[C(RESULT_ACCESS)] = { 0x02ff, PIC_UPPER | PIC_LOWER, },
511 		[C(RESULT_MISS)] = { 0x0301, PIC_UPPER | PIC_LOWER, },
512 	},
513 	[ C(OP_WRITE) ] = {
514 		[ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
515 		[ C(RESULT_MISS)   ] = { CACHE_OP_NONSENSE },
516 	},
517 	[ C(OP_PREFETCH) ] = {
518 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
519 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
520 	},
521 },
522 [C(LL)] = {
523 	[C(OP_READ)] = {
524 		[C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, },
525 		[C(RESULT_MISS)] = { 0x0330, PIC_UPPER | PIC_LOWER, },
526 	},
527 	[C(OP_WRITE)] = {
528 		[C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, },
529 		[C(RESULT_MISS)] = { 0x0320, PIC_UPPER | PIC_LOWER, },
530 	},
531 	[C(OP_PREFETCH)] = {
532 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
533 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
534 	},
535 },
536 [C(DTLB)] = {
537 	[C(OP_READ)] = {
538 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
539 		[C(RESULT_MISS)] = { 0x0b08, PIC_UPPER | PIC_LOWER, },
540 	},
541 	[ C(OP_WRITE) ] = {
542 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
543 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
544 	},
545 	[ C(OP_PREFETCH) ] = {
546 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
547 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
548 	},
549 },
550 [C(ITLB)] = {
551 	[C(OP_READ)] = {
552 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
553 		[C(RESULT_MISS)] = { 0xb04, PIC_UPPER | PIC_LOWER, },
554 	},
555 	[ C(OP_WRITE) ] = {
556 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
557 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
558 	},
559 	[ C(OP_PREFETCH) ] = {
560 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
561 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
562 	},
563 },
564 [C(BPU)] = {
565 	[C(OP_READ)] = {
566 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
567 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
568 	},
569 	[ C(OP_WRITE) ] = {
570 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
571 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
572 	},
573 	[ C(OP_PREFETCH) ] = {
574 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
575 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
576 	},
577 },
578 [C(NODE)] = {
579 	[C(OP_READ)] = {
580 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
581 		[C(RESULT_MISS)  ] = { CACHE_OP_UNSUPPORTED },
582 	},
583 	[ C(OP_WRITE) ] = {
584 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
585 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
586 	},
587 	[ C(OP_PREFETCH) ] = {
588 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
589 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
590 	},
591 },
592 };
593 
594 static const struct sparc_pmu niagara2_pmu = {
595 	.event_map	= niagara2_event_map,
596 	.cache_map	= &niagara2_cache_map,
597 	.max_events	= ARRAY_SIZE(niagara2_perfmon_event_map),
598 	.read_pmc	= sparc_default_read_pmc,
599 	.write_pmc	= sparc_default_write_pmc,
600 	.upper_shift	= 19,
601 	.lower_shift	= 6,
602 	.event_mask	= 0xfff,
603 	.user_bit	= PCR_UTRACE,
604 	.priv_bit	= PCR_STRACE,
605 	.hv_bit		= PCR_N2_HTRACE,
606 	.irq_bit	= 0x30,
607 	.upper_nop	= 0x220,
608 	.lower_nop	= 0x220,
609 	.flags		= (SPARC_PMU_ALL_EXCLUDES_SAME |
610 			   SPARC_PMU_HAS_CONFLICTS),
611 	.max_hw_events	= 2,
612 	.num_pcrs	= 1,
613 	.num_pic_regs	= 1,
614 };
615 
616 static const struct perf_event_map niagara4_perfmon_event_map[] = {
617 	[PERF_COUNT_HW_CPU_CYCLES] = { (26 << 6) },
618 	[PERF_COUNT_HW_INSTRUCTIONS] = { (3 << 6) | 0x3f },
619 	[PERF_COUNT_HW_CACHE_REFERENCES] = { (3 << 6) | 0x04 },
620 	[PERF_COUNT_HW_CACHE_MISSES] = { (16 << 6) | 0x07 },
621 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { (4 << 6) | 0x01 },
622 	[PERF_COUNT_HW_BRANCH_MISSES] = { (25 << 6) | 0x0f },
623 };
624 
625 static const struct perf_event_map *niagara4_event_map(int event_id)
626 {
627 	return &niagara4_perfmon_event_map[event_id];
628 }
629 
630 static const cache_map_t niagara4_cache_map = {
631 [C(L1D)] = {
632 	[C(OP_READ)] = {
633 		[C(RESULT_ACCESS)] = { (3 << 6) | 0x04 },
634 		[C(RESULT_MISS)] = { (16 << 6) | 0x07 },
635 	},
636 	[C(OP_WRITE)] = {
637 		[C(RESULT_ACCESS)] = { (3 << 6) | 0x08 },
638 		[C(RESULT_MISS)] = { (16 << 6) | 0x07 },
639 	},
640 	[C(OP_PREFETCH)] = {
641 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
642 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
643 	},
644 },
645 [C(L1I)] = {
646 	[C(OP_READ)] = {
647 		[C(RESULT_ACCESS)] = { (3 << 6) | 0x3f },
648 		[C(RESULT_MISS)] = { (11 << 6) | 0x03 },
649 	},
650 	[ C(OP_WRITE) ] = {
651 		[ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE },
652 		[ C(RESULT_MISS)   ] = { CACHE_OP_NONSENSE },
653 	},
654 	[ C(OP_PREFETCH) ] = {
655 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
656 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
657 	},
658 },
659 [C(LL)] = {
660 	[C(OP_READ)] = {
661 		[C(RESULT_ACCESS)] = { (3 << 6) | 0x04 },
662 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
663 	},
664 	[C(OP_WRITE)] = {
665 		[C(RESULT_ACCESS)] = { (3 << 6) | 0x08 },
666 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
667 	},
668 	[C(OP_PREFETCH)] = {
669 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
670 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
671 	},
672 },
673 [C(DTLB)] = {
674 	[C(OP_READ)] = {
675 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
676 		[C(RESULT_MISS)] = { (17 << 6) | 0x3f },
677 	},
678 	[ C(OP_WRITE) ] = {
679 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
680 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
681 	},
682 	[ C(OP_PREFETCH) ] = {
683 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
684 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
685 	},
686 },
687 [C(ITLB)] = {
688 	[C(OP_READ)] = {
689 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
690 		[C(RESULT_MISS)] = { (6 << 6) | 0x3f },
691 	},
692 	[ C(OP_WRITE) ] = {
693 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
694 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
695 	},
696 	[ C(OP_PREFETCH) ] = {
697 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
698 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
699 	},
700 },
701 [C(BPU)] = {
702 	[C(OP_READ)] = {
703 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
704 		[C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED },
705 	},
706 	[ C(OP_WRITE) ] = {
707 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
708 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
709 	},
710 	[ C(OP_PREFETCH) ] = {
711 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
712 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
713 	},
714 },
715 [C(NODE)] = {
716 	[C(OP_READ)] = {
717 		[C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED },
718 		[C(RESULT_MISS)  ] = { CACHE_OP_UNSUPPORTED },
719 	},
720 	[ C(OP_WRITE) ] = {
721 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
722 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
723 	},
724 	[ C(OP_PREFETCH) ] = {
725 		[ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED },
726 		[ C(RESULT_MISS)   ] = { CACHE_OP_UNSUPPORTED },
727 	},
728 },
729 };
730 
731 static u32 sparc_vt_read_pmc(int idx)
732 {
733 	u64 val = pcr_ops->read_pic(idx);
734 
735 	return val & 0xffffffff;
736 }
737 
738 static void sparc_vt_write_pmc(int idx, u64 val)
739 {
740 	u64 pcr;
741 
742 	pcr = pcr_ops->read_pcr(idx);
743 	/* ensure ov and ntc are reset */
744 	pcr &= ~(PCR_N4_OV | PCR_N4_NTC);
745 
746 	pcr_ops->write_pic(idx, val & 0xffffffff);
747 
748 	pcr_ops->write_pcr(idx, pcr);
749 }
750 
751 static const struct sparc_pmu niagara4_pmu = {
752 	.event_map	= niagara4_event_map,
753 	.cache_map	= &niagara4_cache_map,
754 	.max_events	= ARRAY_SIZE(niagara4_perfmon_event_map),
755 	.read_pmc	= sparc_vt_read_pmc,
756 	.write_pmc	= sparc_vt_write_pmc,
757 	.upper_shift	= 5,
758 	.lower_shift	= 5,
759 	.event_mask	= 0x7ff,
760 	.user_bit	= PCR_N4_UTRACE,
761 	.priv_bit	= PCR_N4_STRACE,
762 
763 	/* We explicitly don't support hypervisor tracing.  The T4
764 	 * generates the overflow event for precise events via a trap
765 	 * which will not be generated (ie. it's completely lost) if
766 	 * we happen to be in the hypervisor when the event triggers.
767 	 * Essentially, the overflow event reporting is completely
768 	 * unusable when you have hypervisor mode tracing enabled.
769 	 */
770 	.hv_bit		= 0,
771 
772 	.irq_bit	= PCR_N4_TOE,
773 	.upper_nop	= 0,
774 	.lower_nop	= 0,
775 	.flags		= 0,
776 	.max_hw_events	= 4,
777 	.num_pcrs	= 4,
778 	.num_pic_regs	= 4,
779 };
780 
781 static const struct sparc_pmu sparc_m7_pmu = {
782 	.event_map	= niagara4_event_map,
783 	.cache_map	= &niagara4_cache_map,
784 	.max_events	= ARRAY_SIZE(niagara4_perfmon_event_map),
785 	.read_pmc	= sparc_vt_read_pmc,
786 	.write_pmc	= sparc_vt_write_pmc,
787 	.upper_shift	= 5,
788 	.lower_shift	= 5,
789 	.event_mask	= 0x7ff,
790 	.user_bit	= PCR_N4_UTRACE,
791 	.priv_bit	= PCR_N4_STRACE,
792 
793 	/* We explicitly don't support hypervisor tracing. */
794 	.hv_bit		= 0,
795 
796 	.irq_bit	= PCR_N4_TOE,
797 	.upper_nop	= 0,
798 	.lower_nop	= 0,
799 	.flags		= 0,
800 	.max_hw_events	= 4,
801 	.num_pcrs	= 4,
802 	.num_pic_regs	= 4,
803 };
804 static const struct sparc_pmu *sparc_pmu __read_mostly;
805 
806 static u64 event_encoding(u64 event_id, int idx)
807 {
808 	if (idx == PIC_UPPER_INDEX)
809 		event_id <<= sparc_pmu->upper_shift;
810 	else
811 		event_id <<= sparc_pmu->lower_shift;
812 	return event_id;
813 }
814 
815 static u64 mask_for_index(int idx)
816 {
817 	return event_encoding(sparc_pmu->event_mask, idx);
818 }
819 
820 static u64 nop_for_index(int idx)
821 {
822 	return event_encoding(idx == PIC_UPPER_INDEX ?
823 			      sparc_pmu->upper_nop :
824 			      sparc_pmu->lower_nop, idx);
825 }
826 
827 static inline void sparc_pmu_enable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx)
828 {
829 	u64 enc, val, mask = mask_for_index(idx);
830 	int pcr_index = 0;
831 
832 	if (sparc_pmu->num_pcrs > 1)
833 		pcr_index = idx;
834 
835 	enc = perf_event_get_enc(cpuc->events[idx]);
836 
837 	val = cpuc->pcr[pcr_index];
838 	val &= ~mask;
839 	val |= event_encoding(enc, idx);
840 	cpuc->pcr[pcr_index] = val;
841 
842 	pcr_ops->write_pcr(pcr_index, cpuc->pcr[pcr_index]);
843 }
844 
845 static inline void sparc_pmu_disable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx)
846 {
847 	u64 mask = mask_for_index(idx);
848 	u64 nop = nop_for_index(idx);
849 	int pcr_index = 0;
850 	u64 val;
851 
852 	if (sparc_pmu->num_pcrs > 1)
853 		pcr_index = idx;
854 
855 	val = cpuc->pcr[pcr_index];
856 	val &= ~mask;
857 	val |= nop;
858 	cpuc->pcr[pcr_index] = val;
859 
860 	pcr_ops->write_pcr(pcr_index, cpuc->pcr[pcr_index]);
861 }
862 
863 static u64 sparc_perf_event_update(struct perf_event *event,
864 				   struct hw_perf_event *hwc, int idx)
865 {
866 	int shift = 64 - 32;
867 	u64 prev_raw_count, new_raw_count;
868 	s64 delta;
869 
870 again:
871 	prev_raw_count = local64_read(&hwc->prev_count);
872 	new_raw_count = sparc_pmu->read_pmc(idx);
873 
874 	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
875 			     new_raw_count) != prev_raw_count)
876 		goto again;
877 
878 	delta = (new_raw_count << shift) - (prev_raw_count << shift);
879 	delta >>= shift;
880 
881 	local64_add(delta, &event->count);
882 	local64_sub(delta, &hwc->period_left);
883 
884 	return new_raw_count;
885 }
886 
887 static int sparc_perf_event_set_period(struct perf_event *event,
888 				       struct hw_perf_event *hwc, int idx)
889 {
890 	s64 left = local64_read(&hwc->period_left);
891 	s64 period = hwc->sample_period;
892 	int ret = 0;
893 
894 	if (unlikely(left <= -period)) {
895 		left = period;
896 		local64_set(&hwc->period_left, left);
897 		hwc->last_period = period;
898 		ret = 1;
899 	}
900 
901 	if (unlikely(left <= 0)) {
902 		left += period;
903 		local64_set(&hwc->period_left, left);
904 		hwc->last_period = period;
905 		ret = 1;
906 	}
907 	if (left > MAX_PERIOD)
908 		left = MAX_PERIOD;
909 
910 	local64_set(&hwc->prev_count, (u64)-left);
911 
912 	sparc_pmu->write_pmc(idx, (u64)(-left) & 0xffffffff);
913 
914 	perf_event_update_userpage(event);
915 
916 	return ret;
917 }
918 
919 static void read_in_all_counters(struct cpu_hw_events *cpuc)
920 {
921 	int i;
922 
923 	for (i = 0; i < cpuc->n_events; i++) {
924 		struct perf_event *cp = cpuc->event[i];
925 
926 		if (cpuc->current_idx[i] != PIC_NO_INDEX &&
927 		    cpuc->current_idx[i] != cp->hw.idx) {
928 			sparc_perf_event_update(cp, &cp->hw,
929 						cpuc->current_idx[i]);
930 			cpuc->current_idx[i] = PIC_NO_INDEX;
931 			if (cp->hw.state & PERF_HES_STOPPED)
932 				cp->hw.state |= PERF_HES_ARCH;
933 		}
934 	}
935 }
936 
937 /* On this PMU all PICs are programmed using a single PCR.  Calculate
938  * the combined control register value.
939  *
940  * For such chips we require that all of the events have the same
941  * configuration, so just fetch the settings from the first entry.
942  */
943 static void calculate_single_pcr(struct cpu_hw_events *cpuc)
944 {
945 	int i;
946 
947 	if (!cpuc->n_added)
948 		goto out;
949 
950 	/* Assign to counters all unassigned events.  */
951 	for (i = 0; i < cpuc->n_events; i++) {
952 		struct perf_event *cp = cpuc->event[i];
953 		struct hw_perf_event *hwc = &cp->hw;
954 		int idx = hwc->idx;
955 		u64 enc;
956 
957 		if (cpuc->current_idx[i] != PIC_NO_INDEX)
958 			continue;
959 
960 		sparc_perf_event_set_period(cp, hwc, idx);
961 		cpuc->current_idx[i] = idx;
962 
963 		enc = perf_event_get_enc(cpuc->events[i]);
964 		cpuc->pcr[0] &= ~mask_for_index(idx);
965 		if (hwc->state & PERF_HES_ARCH) {
966 			cpuc->pcr[0] |= nop_for_index(idx);
967 		} else {
968 			cpuc->pcr[0] |= event_encoding(enc, idx);
969 			hwc->state = 0;
970 		}
971 	}
972 out:
973 	cpuc->pcr[0] |= cpuc->event[0]->hw.config_base;
974 }
975 
976 static void sparc_pmu_start(struct perf_event *event, int flags);
977 
978 /* On this PMU each PIC has it's own PCR control register.  */
979 static void calculate_multiple_pcrs(struct cpu_hw_events *cpuc)
980 {
981 	int i;
982 
983 	if (!cpuc->n_added)
984 		goto out;
985 
986 	for (i = 0; i < cpuc->n_events; i++) {
987 		struct perf_event *cp = cpuc->event[i];
988 		struct hw_perf_event *hwc = &cp->hw;
989 		int idx = hwc->idx;
990 
991 		if (cpuc->current_idx[i] != PIC_NO_INDEX)
992 			continue;
993 
994 		cpuc->current_idx[i] = idx;
995 
996 		if (cp->hw.state & PERF_HES_ARCH)
997 			continue;
998 
999 		sparc_pmu_start(cp, PERF_EF_RELOAD);
1000 	}
1001 out:
1002 	for (i = 0; i < cpuc->n_events; i++) {
1003 		struct perf_event *cp = cpuc->event[i];
1004 		int idx = cp->hw.idx;
1005 
1006 		cpuc->pcr[idx] |= cp->hw.config_base;
1007 	}
1008 }
1009 
1010 /* If performance event entries have been added, move existing events
1011  * around (if necessary) and then assign new entries to counters.
1012  */
1013 static void update_pcrs_for_enable(struct cpu_hw_events *cpuc)
1014 {
1015 	if (cpuc->n_added)
1016 		read_in_all_counters(cpuc);
1017 
1018 	if (sparc_pmu->num_pcrs == 1) {
1019 		calculate_single_pcr(cpuc);
1020 	} else {
1021 		calculate_multiple_pcrs(cpuc);
1022 	}
1023 }
1024 
1025 static void sparc_pmu_enable(struct pmu *pmu)
1026 {
1027 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1028 	int i;
1029 
1030 	if (cpuc->enabled)
1031 		return;
1032 
1033 	cpuc->enabled = 1;
1034 	barrier();
1035 
1036 	if (cpuc->n_events)
1037 		update_pcrs_for_enable(cpuc);
1038 
1039 	for (i = 0; i < sparc_pmu->num_pcrs; i++)
1040 		pcr_ops->write_pcr(i, cpuc->pcr[i]);
1041 }
1042 
1043 static void sparc_pmu_disable(struct pmu *pmu)
1044 {
1045 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1046 	int i;
1047 
1048 	if (!cpuc->enabled)
1049 		return;
1050 
1051 	cpuc->enabled = 0;
1052 	cpuc->n_added = 0;
1053 
1054 	for (i = 0; i < sparc_pmu->num_pcrs; i++) {
1055 		u64 val = cpuc->pcr[i];
1056 
1057 		val &= ~(sparc_pmu->user_bit | sparc_pmu->priv_bit |
1058 			 sparc_pmu->hv_bit | sparc_pmu->irq_bit);
1059 		cpuc->pcr[i] = val;
1060 		pcr_ops->write_pcr(i, cpuc->pcr[i]);
1061 	}
1062 }
1063 
1064 static int active_event_index(struct cpu_hw_events *cpuc,
1065 			      struct perf_event *event)
1066 {
1067 	int i;
1068 
1069 	for (i = 0; i < cpuc->n_events; i++) {
1070 		if (cpuc->event[i] == event)
1071 			break;
1072 	}
1073 	BUG_ON(i == cpuc->n_events);
1074 	return cpuc->current_idx[i];
1075 }
1076 
1077 static void sparc_pmu_start(struct perf_event *event, int flags)
1078 {
1079 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1080 	int idx = active_event_index(cpuc, event);
1081 
1082 	if (flags & PERF_EF_RELOAD) {
1083 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1084 		sparc_perf_event_set_period(event, &event->hw, idx);
1085 	}
1086 
1087 	event->hw.state = 0;
1088 
1089 	sparc_pmu_enable_event(cpuc, &event->hw, idx);
1090 
1091 	perf_event_update_userpage(event);
1092 }
1093 
1094 static void sparc_pmu_stop(struct perf_event *event, int flags)
1095 {
1096 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1097 	int idx = active_event_index(cpuc, event);
1098 
1099 	if (!(event->hw.state & PERF_HES_STOPPED)) {
1100 		sparc_pmu_disable_event(cpuc, &event->hw, idx);
1101 		event->hw.state |= PERF_HES_STOPPED;
1102 	}
1103 
1104 	if (!(event->hw.state & PERF_HES_UPTODATE) && (flags & PERF_EF_UPDATE)) {
1105 		sparc_perf_event_update(event, &event->hw, idx);
1106 		event->hw.state |= PERF_HES_UPTODATE;
1107 	}
1108 }
1109 
1110 static void sparc_pmu_del(struct perf_event *event, int _flags)
1111 {
1112 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1113 	unsigned long flags;
1114 	int i;
1115 
1116 	local_irq_save(flags);
1117 
1118 	for (i = 0; i < cpuc->n_events; i++) {
1119 		if (event == cpuc->event[i]) {
1120 			/* Absorb the final count and turn off the
1121 			 * event.
1122 			 */
1123 			sparc_pmu_stop(event, PERF_EF_UPDATE);
1124 
1125 			/* Shift remaining entries down into
1126 			 * the existing slot.
1127 			 */
1128 			while (++i < cpuc->n_events) {
1129 				cpuc->event[i - 1] = cpuc->event[i];
1130 				cpuc->events[i - 1] = cpuc->events[i];
1131 				cpuc->current_idx[i - 1] =
1132 					cpuc->current_idx[i];
1133 			}
1134 
1135 			perf_event_update_userpage(event);
1136 
1137 			cpuc->n_events--;
1138 			break;
1139 		}
1140 	}
1141 
1142 	local_irq_restore(flags);
1143 }
1144 
1145 static void sparc_pmu_read(struct perf_event *event)
1146 {
1147 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1148 	int idx = active_event_index(cpuc, event);
1149 	struct hw_perf_event *hwc = &event->hw;
1150 
1151 	sparc_perf_event_update(event, hwc, idx);
1152 }
1153 
1154 static atomic_t active_events = ATOMIC_INIT(0);
1155 static DEFINE_MUTEX(pmc_grab_mutex);
1156 
1157 static void perf_stop_nmi_watchdog(void *unused)
1158 {
1159 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1160 	int i;
1161 
1162 	stop_nmi_watchdog(NULL);
1163 	for (i = 0; i < sparc_pmu->num_pcrs; i++)
1164 		cpuc->pcr[i] = pcr_ops->read_pcr(i);
1165 }
1166 
1167 static void perf_event_grab_pmc(void)
1168 {
1169 	if (atomic_inc_not_zero(&active_events))
1170 		return;
1171 
1172 	mutex_lock(&pmc_grab_mutex);
1173 	if (atomic_read(&active_events) == 0) {
1174 		if (atomic_read(&nmi_active) > 0) {
1175 			on_each_cpu(perf_stop_nmi_watchdog, NULL, 1);
1176 			BUG_ON(atomic_read(&nmi_active) != 0);
1177 		}
1178 		atomic_inc(&active_events);
1179 	}
1180 	mutex_unlock(&pmc_grab_mutex);
1181 }
1182 
1183 static void perf_event_release_pmc(void)
1184 {
1185 	if (atomic_dec_and_mutex_lock(&active_events, &pmc_grab_mutex)) {
1186 		if (atomic_read(&nmi_active) == 0)
1187 			on_each_cpu(start_nmi_watchdog, NULL, 1);
1188 		mutex_unlock(&pmc_grab_mutex);
1189 	}
1190 }
1191 
1192 static const struct perf_event_map *sparc_map_cache_event(u64 config)
1193 {
1194 	unsigned int cache_type, cache_op, cache_result;
1195 	const struct perf_event_map *pmap;
1196 
1197 	if (!sparc_pmu->cache_map)
1198 		return ERR_PTR(-ENOENT);
1199 
1200 	cache_type = (config >>  0) & 0xff;
1201 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
1202 		return ERR_PTR(-EINVAL);
1203 
1204 	cache_op = (config >>  8) & 0xff;
1205 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
1206 		return ERR_PTR(-EINVAL);
1207 
1208 	cache_result = (config >> 16) & 0xff;
1209 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
1210 		return ERR_PTR(-EINVAL);
1211 
1212 	pmap = &((*sparc_pmu->cache_map)[cache_type][cache_op][cache_result]);
1213 
1214 	if (pmap->encoding == CACHE_OP_UNSUPPORTED)
1215 		return ERR_PTR(-ENOENT);
1216 
1217 	if (pmap->encoding == CACHE_OP_NONSENSE)
1218 		return ERR_PTR(-EINVAL);
1219 
1220 	return pmap;
1221 }
1222 
1223 static void hw_perf_event_destroy(struct perf_event *event)
1224 {
1225 	perf_event_release_pmc();
1226 }
1227 
1228 /* Make sure all events can be scheduled into the hardware at
1229  * the same time.  This is simplified by the fact that we only
1230  * need to support 2 simultaneous HW events.
1231  *
1232  * As a side effect, the evts[]->hw.idx values will be assigned
1233  * on success.  These are pending indexes.  When the events are
1234  * actually programmed into the chip, these values will propagate
1235  * to the per-cpu cpuc->current_idx[] slots, see the code in
1236  * maybe_change_configuration() for details.
1237  */
1238 static int sparc_check_constraints(struct perf_event **evts,
1239 				   unsigned long *events, int n_ev)
1240 {
1241 	u8 msk0 = 0, msk1 = 0;
1242 	int idx0 = 0;
1243 
1244 	/* This case is possible when we are invoked from
1245 	 * hw_perf_group_sched_in().
1246 	 */
1247 	if (!n_ev)
1248 		return 0;
1249 
1250 	if (n_ev > sparc_pmu->max_hw_events)
1251 		return -1;
1252 
1253 	if (!(sparc_pmu->flags & SPARC_PMU_HAS_CONFLICTS)) {
1254 		int i;
1255 
1256 		for (i = 0; i < n_ev; i++)
1257 			evts[i]->hw.idx = i;
1258 		return 0;
1259 	}
1260 
1261 	msk0 = perf_event_get_msk(events[0]);
1262 	if (n_ev == 1) {
1263 		if (msk0 & PIC_LOWER)
1264 			idx0 = 1;
1265 		goto success;
1266 	}
1267 	BUG_ON(n_ev != 2);
1268 	msk1 = perf_event_get_msk(events[1]);
1269 
1270 	/* If both events can go on any counter, OK.  */
1271 	if (msk0 == (PIC_UPPER | PIC_LOWER) &&
1272 	    msk1 == (PIC_UPPER | PIC_LOWER))
1273 		goto success;
1274 
1275 	/* If one event is limited to a specific counter,
1276 	 * and the other can go on both, OK.
1277 	 */
1278 	if ((msk0 == PIC_UPPER || msk0 == PIC_LOWER) &&
1279 	    msk1 == (PIC_UPPER | PIC_LOWER)) {
1280 		if (msk0 & PIC_LOWER)
1281 			idx0 = 1;
1282 		goto success;
1283 	}
1284 
1285 	if ((msk1 == PIC_UPPER || msk1 == PIC_LOWER) &&
1286 	    msk0 == (PIC_UPPER | PIC_LOWER)) {
1287 		if (msk1 & PIC_UPPER)
1288 			idx0 = 1;
1289 		goto success;
1290 	}
1291 
1292 	/* If the events are fixed to different counters, OK.  */
1293 	if ((msk0 == PIC_UPPER && msk1 == PIC_LOWER) ||
1294 	    (msk0 == PIC_LOWER && msk1 == PIC_UPPER)) {
1295 		if (msk0 & PIC_LOWER)
1296 			idx0 = 1;
1297 		goto success;
1298 	}
1299 
1300 	/* Otherwise, there is a conflict.  */
1301 	return -1;
1302 
1303 success:
1304 	evts[0]->hw.idx = idx0;
1305 	if (n_ev == 2)
1306 		evts[1]->hw.idx = idx0 ^ 1;
1307 	return 0;
1308 }
1309 
1310 static int check_excludes(struct perf_event **evts, int n_prev, int n_new)
1311 {
1312 	int eu = 0, ek = 0, eh = 0;
1313 	struct perf_event *event;
1314 	int i, n, first;
1315 
1316 	if (!(sparc_pmu->flags & SPARC_PMU_ALL_EXCLUDES_SAME))
1317 		return 0;
1318 
1319 	n = n_prev + n_new;
1320 	if (n <= 1)
1321 		return 0;
1322 
1323 	first = 1;
1324 	for (i = 0; i < n; i++) {
1325 		event = evts[i];
1326 		if (first) {
1327 			eu = event->attr.exclude_user;
1328 			ek = event->attr.exclude_kernel;
1329 			eh = event->attr.exclude_hv;
1330 			first = 0;
1331 		} else if (event->attr.exclude_user != eu ||
1332 			   event->attr.exclude_kernel != ek ||
1333 			   event->attr.exclude_hv != eh) {
1334 			return -EAGAIN;
1335 		}
1336 	}
1337 
1338 	return 0;
1339 }
1340 
1341 static int collect_events(struct perf_event *group, int max_count,
1342 			  struct perf_event *evts[], unsigned long *events,
1343 			  int *current_idx)
1344 {
1345 	struct perf_event *event;
1346 	int n = 0;
1347 
1348 	if (!is_software_event(group)) {
1349 		if (n >= max_count)
1350 			return -1;
1351 		evts[n] = group;
1352 		events[n] = group->hw.event_base;
1353 		current_idx[n++] = PIC_NO_INDEX;
1354 	}
1355 	for_each_sibling_event(event, group) {
1356 		if (!is_software_event(event) &&
1357 		    event->state != PERF_EVENT_STATE_OFF) {
1358 			if (n >= max_count)
1359 				return -1;
1360 			evts[n] = event;
1361 			events[n] = event->hw.event_base;
1362 			current_idx[n++] = PIC_NO_INDEX;
1363 		}
1364 	}
1365 	return n;
1366 }
1367 
1368 static int sparc_pmu_add(struct perf_event *event, int ef_flags)
1369 {
1370 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1371 	int n0, ret = -EAGAIN;
1372 	unsigned long flags;
1373 
1374 	local_irq_save(flags);
1375 
1376 	n0 = cpuc->n_events;
1377 	if (n0 >= sparc_pmu->max_hw_events)
1378 		goto out;
1379 
1380 	cpuc->event[n0] = event;
1381 	cpuc->events[n0] = event->hw.event_base;
1382 	cpuc->current_idx[n0] = PIC_NO_INDEX;
1383 
1384 	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1385 	if (!(ef_flags & PERF_EF_START))
1386 		event->hw.state |= PERF_HES_ARCH;
1387 
1388 	/*
1389 	 * If group events scheduling transaction was started,
1390 	 * skip the schedulability test here, it will be performed
1391 	 * at commit time(->commit_txn) as a whole
1392 	 */
1393 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1394 		goto nocheck;
1395 
1396 	if (check_excludes(cpuc->event, n0, 1))
1397 		goto out;
1398 	if (sparc_check_constraints(cpuc->event, cpuc->events, n0 + 1))
1399 		goto out;
1400 
1401 nocheck:
1402 	cpuc->n_events++;
1403 	cpuc->n_added++;
1404 
1405 	ret = 0;
1406 out:
1407 	local_irq_restore(flags);
1408 	return ret;
1409 }
1410 
1411 static int sparc_pmu_event_init(struct perf_event *event)
1412 {
1413 	struct perf_event_attr *attr = &event->attr;
1414 	struct perf_event *evts[MAX_HWEVENTS];
1415 	struct hw_perf_event *hwc = &event->hw;
1416 	unsigned long events[MAX_HWEVENTS];
1417 	int current_idx_dmy[MAX_HWEVENTS];
1418 	const struct perf_event_map *pmap;
1419 	int n;
1420 
1421 	if (atomic_read(&nmi_active) < 0)
1422 		return -ENODEV;
1423 
1424 	/* does not support taken branch sampling */
1425 	if (has_branch_stack(event))
1426 		return -EOPNOTSUPP;
1427 
1428 	switch (attr->type) {
1429 	case PERF_TYPE_HARDWARE:
1430 		if (attr->config >= sparc_pmu->max_events)
1431 			return -EINVAL;
1432 		pmap = sparc_pmu->event_map(attr->config);
1433 		break;
1434 
1435 	case PERF_TYPE_HW_CACHE:
1436 		pmap = sparc_map_cache_event(attr->config);
1437 		if (IS_ERR(pmap))
1438 			return PTR_ERR(pmap);
1439 		break;
1440 
1441 	case PERF_TYPE_RAW:
1442 		pmap = NULL;
1443 		break;
1444 
1445 	default:
1446 		return -ENOENT;
1447 
1448 	}
1449 
1450 	if (pmap) {
1451 		hwc->event_base = perf_event_encode(pmap);
1452 	} else {
1453 		/*
1454 		 * User gives us "(encoding << 16) | pic_mask" for
1455 		 * PERF_TYPE_RAW events.
1456 		 */
1457 		hwc->event_base = attr->config;
1458 	}
1459 
1460 	/* We save the enable bits in the config_base.  */
1461 	hwc->config_base = sparc_pmu->irq_bit;
1462 	if (!attr->exclude_user)
1463 		hwc->config_base |= sparc_pmu->user_bit;
1464 	if (!attr->exclude_kernel)
1465 		hwc->config_base |= sparc_pmu->priv_bit;
1466 	if (!attr->exclude_hv)
1467 		hwc->config_base |= sparc_pmu->hv_bit;
1468 
1469 	n = 0;
1470 	if (event->group_leader != event) {
1471 		n = collect_events(event->group_leader,
1472 				   sparc_pmu->max_hw_events - 1,
1473 				   evts, events, current_idx_dmy);
1474 		if (n < 0)
1475 			return -EINVAL;
1476 	}
1477 	events[n] = hwc->event_base;
1478 	evts[n] = event;
1479 
1480 	if (check_excludes(evts, n, 1))
1481 		return -EINVAL;
1482 
1483 	if (sparc_check_constraints(evts, events, n + 1))
1484 		return -EINVAL;
1485 
1486 	hwc->idx = PIC_NO_INDEX;
1487 
1488 	/* Try to do all error checking before this point, as unwinding
1489 	 * state after grabbing the PMC is difficult.
1490 	 */
1491 	perf_event_grab_pmc();
1492 	event->destroy = hw_perf_event_destroy;
1493 
1494 	if (!hwc->sample_period) {
1495 		hwc->sample_period = MAX_PERIOD;
1496 		hwc->last_period = hwc->sample_period;
1497 		local64_set(&hwc->period_left, hwc->sample_period);
1498 	}
1499 
1500 	return 0;
1501 }
1502 
1503 /*
1504  * Start group events scheduling transaction
1505  * Set the flag to make pmu::enable() not perform the
1506  * schedulability test, it will be performed at commit time
1507  */
1508 static void sparc_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
1509 {
1510 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1511 
1512 	WARN_ON_ONCE(cpuhw->txn_flags);		/* txn already in flight */
1513 
1514 	cpuhw->txn_flags = txn_flags;
1515 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1516 		return;
1517 
1518 	perf_pmu_disable(pmu);
1519 }
1520 
1521 /*
1522  * Stop group events scheduling transaction
1523  * Clear the flag and pmu::enable() will perform the
1524  * schedulability test.
1525  */
1526 static void sparc_pmu_cancel_txn(struct pmu *pmu)
1527 {
1528 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1529 	unsigned int txn_flags;
1530 
1531 	WARN_ON_ONCE(!cpuhw->txn_flags);	/* no txn in flight */
1532 
1533 	txn_flags = cpuhw->txn_flags;
1534 	cpuhw->txn_flags = 0;
1535 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1536 		return;
1537 
1538 	perf_pmu_enable(pmu);
1539 }
1540 
1541 /*
1542  * Commit group events scheduling transaction
1543  * Perform the group schedulability test as a whole
1544  * Return 0 if success
1545  */
1546 static int sparc_pmu_commit_txn(struct pmu *pmu)
1547 {
1548 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1549 	int n;
1550 
1551 	if (!sparc_pmu)
1552 		return -EINVAL;
1553 
1554 	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
1555 
1556 	if (cpuc->txn_flags & ~PERF_PMU_TXN_ADD) {
1557 		cpuc->txn_flags = 0;
1558 		return 0;
1559 	}
1560 
1561 	n = cpuc->n_events;
1562 	if (check_excludes(cpuc->event, 0, n))
1563 		return -EINVAL;
1564 	if (sparc_check_constraints(cpuc->event, cpuc->events, n))
1565 		return -EAGAIN;
1566 
1567 	cpuc->txn_flags = 0;
1568 	perf_pmu_enable(pmu);
1569 	return 0;
1570 }
1571 
1572 static struct pmu pmu = {
1573 	.pmu_enable	= sparc_pmu_enable,
1574 	.pmu_disable	= sparc_pmu_disable,
1575 	.event_init	= sparc_pmu_event_init,
1576 	.add		= sparc_pmu_add,
1577 	.del		= sparc_pmu_del,
1578 	.start		= sparc_pmu_start,
1579 	.stop		= sparc_pmu_stop,
1580 	.read		= sparc_pmu_read,
1581 	.start_txn	= sparc_pmu_start_txn,
1582 	.cancel_txn	= sparc_pmu_cancel_txn,
1583 	.commit_txn	= sparc_pmu_commit_txn,
1584 };
1585 
1586 void perf_event_print_debug(void)
1587 {
1588 	unsigned long flags;
1589 	int cpu, i;
1590 
1591 	if (!sparc_pmu)
1592 		return;
1593 
1594 	local_irq_save(flags);
1595 
1596 	cpu = smp_processor_id();
1597 
1598 	pr_info("\n");
1599 	for (i = 0; i < sparc_pmu->num_pcrs; i++)
1600 		pr_info("CPU#%d: PCR%d[%016llx]\n",
1601 			cpu, i, pcr_ops->read_pcr(i));
1602 	for (i = 0; i < sparc_pmu->num_pic_regs; i++)
1603 		pr_info("CPU#%d: PIC%d[%016llx]\n",
1604 			cpu, i, pcr_ops->read_pic(i));
1605 
1606 	local_irq_restore(flags);
1607 }
1608 
1609 static int __kprobes perf_event_nmi_handler(struct notifier_block *self,
1610 					    unsigned long cmd, void *__args)
1611 {
1612 	struct die_args *args = __args;
1613 	struct perf_sample_data data;
1614 	struct cpu_hw_events *cpuc;
1615 	struct pt_regs *regs;
1616 	u64 finish_clock;
1617 	u64 start_clock;
1618 	int i;
1619 
1620 	if (!atomic_read(&active_events))
1621 		return NOTIFY_DONE;
1622 
1623 	switch (cmd) {
1624 	case DIE_NMI:
1625 		break;
1626 
1627 	default:
1628 		return NOTIFY_DONE;
1629 	}
1630 
1631 	start_clock = sched_clock();
1632 
1633 	regs = args->regs;
1634 
1635 	cpuc = this_cpu_ptr(&cpu_hw_events);
1636 
1637 	/* If the PMU has the TOE IRQ enable bits, we need to do a
1638 	 * dummy write to the %pcr to clear the overflow bits and thus
1639 	 * the interrupt.
1640 	 *
1641 	 * Do this before we peek at the counters to determine
1642 	 * overflow so we don't lose any events.
1643 	 */
1644 	if (sparc_pmu->irq_bit &&
1645 	    sparc_pmu->num_pcrs == 1)
1646 		pcr_ops->write_pcr(0, cpuc->pcr[0]);
1647 
1648 	for (i = 0; i < cpuc->n_events; i++) {
1649 		struct perf_event *event = cpuc->event[i];
1650 		int idx = cpuc->current_idx[i];
1651 		struct hw_perf_event *hwc;
1652 		u64 val;
1653 
1654 		if (sparc_pmu->irq_bit &&
1655 		    sparc_pmu->num_pcrs > 1)
1656 			pcr_ops->write_pcr(idx, cpuc->pcr[idx]);
1657 
1658 		hwc = &event->hw;
1659 		val = sparc_perf_event_update(event, hwc, idx);
1660 		if (val & (1ULL << 31))
1661 			continue;
1662 
1663 		perf_sample_data_init(&data, 0, hwc->last_period);
1664 		if (!sparc_perf_event_set_period(event, hwc, idx))
1665 			continue;
1666 
1667 		if (perf_event_overflow(event, &data, regs))
1668 			sparc_pmu_stop(event, 0);
1669 	}
1670 
1671 	finish_clock = sched_clock();
1672 
1673 	perf_sample_event_took(finish_clock - start_clock);
1674 
1675 	return NOTIFY_STOP;
1676 }
1677 
1678 static __read_mostly struct notifier_block perf_event_nmi_notifier = {
1679 	.notifier_call		= perf_event_nmi_handler,
1680 };
1681 
1682 static bool __init supported_pmu(void)
1683 {
1684 	if (!strcmp(sparc_pmu_type, "ultra3") ||
1685 	    !strcmp(sparc_pmu_type, "ultra3+") ||
1686 	    !strcmp(sparc_pmu_type, "ultra3i") ||
1687 	    !strcmp(sparc_pmu_type, "ultra4+")) {
1688 		sparc_pmu = &ultra3_pmu;
1689 		return true;
1690 	}
1691 	if (!strcmp(sparc_pmu_type, "niagara")) {
1692 		sparc_pmu = &niagara1_pmu;
1693 		return true;
1694 	}
1695 	if (!strcmp(sparc_pmu_type, "niagara2") ||
1696 	    !strcmp(sparc_pmu_type, "niagara3")) {
1697 		sparc_pmu = &niagara2_pmu;
1698 		return true;
1699 	}
1700 	if (!strcmp(sparc_pmu_type, "niagara4") ||
1701 	    !strcmp(sparc_pmu_type, "niagara5")) {
1702 		sparc_pmu = &niagara4_pmu;
1703 		return true;
1704 	}
1705 	if (!strcmp(sparc_pmu_type, "sparc-m7")) {
1706 		sparc_pmu = &sparc_m7_pmu;
1707 		return true;
1708 	}
1709 	return false;
1710 }
1711 
1712 static int __init init_hw_perf_events(void)
1713 {
1714 	int err;
1715 
1716 	pr_info("Performance events: ");
1717 
1718 	err = pcr_arch_init();
1719 	if (err || !supported_pmu()) {
1720 		pr_cont("No support for PMU type '%s'\n", sparc_pmu_type);
1721 		return 0;
1722 	}
1723 
1724 	pr_cont("Supported PMU type is '%s'\n", sparc_pmu_type);
1725 
1726 	perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1727 	register_die_notifier(&perf_event_nmi_notifier);
1728 
1729 	return 0;
1730 }
1731 pure_initcall(init_hw_perf_events);
1732 
1733 void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry,
1734 			   struct pt_regs *regs)
1735 {
1736 	unsigned long ksp, fp;
1737 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
1738 	int graph = 0;
1739 #endif
1740 
1741 	stack_trace_flush();
1742 
1743 	perf_callchain_store(entry, regs->tpc);
1744 
1745 	ksp = regs->u_regs[UREG_I6];
1746 	fp = ksp + STACK_BIAS;
1747 	do {
1748 		struct sparc_stackf *sf;
1749 		struct pt_regs *regs;
1750 		unsigned long pc;
1751 
1752 		if (!kstack_valid(current_thread_info(), fp))
1753 			break;
1754 
1755 		sf = (struct sparc_stackf *) fp;
1756 		regs = (struct pt_regs *) (sf + 1);
1757 
1758 		if (kstack_is_trap_frame(current_thread_info(), regs)) {
1759 			if (user_mode(regs))
1760 				break;
1761 			pc = regs->tpc;
1762 			fp = regs->u_regs[UREG_I6] + STACK_BIAS;
1763 		} else {
1764 			pc = sf->callers_pc;
1765 			fp = (unsigned long)sf->fp + STACK_BIAS;
1766 		}
1767 		perf_callchain_store(entry, pc);
1768 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
1769 		if ((pc + 8UL) == (unsigned long) &return_to_handler) {
1770 			struct ftrace_ret_stack *ret_stack;
1771 			ret_stack = ftrace_graph_get_ret_stack(current,
1772 							       graph);
1773 			if (ret_stack) {
1774 				pc = ret_stack->ret;
1775 				perf_callchain_store(entry, pc);
1776 				graph++;
1777 			}
1778 		}
1779 #endif
1780 	} while (entry->nr < entry->max_stack);
1781 }
1782 
1783 static inline int
1784 valid_user_frame(const void __user *fp, unsigned long size)
1785 {
1786 	/* addresses should be at least 4-byte aligned */
1787 	if (((unsigned long) fp) & 3)
1788 		return 0;
1789 
1790 	return (__range_not_ok(fp, size, TASK_SIZE) == 0);
1791 }
1792 
1793 static void perf_callchain_user_64(struct perf_callchain_entry_ctx *entry,
1794 				   struct pt_regs *regs)
1795 {
1796 	unsigned long ufp;
1797 
1798 	ufp = regs->u_regs[UREG_FP] + STACK_BIAS;
1799 	do {
1800 		struct sparc_stackf __user *usf;
1801 		struct sparc_stackf sf;
1802 		unsigned long pc;
1803 
1804 		usf = (struct sparc_stackf __user *)ufp;
1805 		if (!valid_user_frame(usf, sizeof(sf)))
1806 			break;
1807 
1808 		if (__copy_from_user_inatomic(&sf, usf, sizeof(sf)))
1809 			break;
1810 
1811 		pc = sf.callers_pc;
1812 		ufp = (unsigned long)sf.fp + STACK_BIAS;
1813 		perf_callchain_store(entry, pc);
1814 	} while (entry->nr < entry->max_stack);
1815 }
1816 
1817 static void perf_callchain_user_32(struct perf_callchain_entry_ctx *entry,
1818 				   struct pt_regs *regs)
1819 {
1820 	unsigned long ufp;
1821 
1822 	ufp = regs->u_regs[UREG_FP] & 0xffffffffUL;
1823 	do {
1824 		unsigned long pc;
1825 
1826 		if (thread32_stack_is_64bit(ufp)) {
1827 			struct sparc_stackf __user *usf;
1828 			struct sparc_stackf sf;
1829 
1830 			ufp += STACK_BIAS;
1831 			usf = (struct sparc_stackf __user *)ufp;
1832 			if (__copy_from_user_inatomic(&sf, usf, sizeof(sf)))
1833 				break;
1834 			pc = sf.callers_pc & 0xffffffff;
1835 			ufp = ((unsigned long) sf.fp) & 0xffffffff;
1836 		} else {
1837 			struct sparc_stackf32 __user *usf;
1838 			struct sparc_stackf32 sf;
1839 			usf = (struct sparc_stackf32 __user *)ufp;
1840 			if (__copy_from_user_inatomic(&sf, usf, sizeof(sf)))
1841 				break;
1842 			pc = sf.callers_pc;
1843 			ufp = (unsigned long)sf.fp;
1844 		}
1845 		perf_callchain_store(entry, pc);
1846 	} while (entry->nr < entry->max_stack);
1847 }
1848 
1849 void
1850 perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
1851 {
1852 	u64 saved_fault_address = current_thread_info()->fault_address;
1853 	u8 saved_fault_code = get_thread_fault_code();
1854 
1855 	perf_callchain_store(entry, regs->tpc);
1856 
1857 	if (!current->mm)
1858 		return;
1859 
1860 	flushw_user();
1861 
1862 	pagefault_disable();
1863 
1864 	if (test_thread_flag(TIF_32BIT))
1865 		perf_callchain_user_32(entry, regs);
1866 	else
1867 		perf_callchain_user_64(entry, regs);
1868 
1869 	pagefault_enable();
1870 
1871 	set_thread_fault_code(saved_fault_code);
1872 	current_thread_info()->fault_address = saved_fault_address;
1873 }
1874