xref: /openbmc/linux/arch/powerpc/perf/power8-pmu.c (revision 80ecbd24)
1 /*
2  * Performance counter support for POWER8 processors.
3  *
4  * Copyright 2009 Paul Mackerras, IBM Corporation.
5  * Copyright 2013 Michael Ellerman, IBM Corporation.
6  *
7  * This program is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU General Public License
9  * as published by the Free Software Foundation; either version
10  * 2 of the License, or (at your option) any later version.
11  */
12 
13 #include <linux/kernel.h>
14 #include <linux/perf_event.h>
15 #include <asm/firmware.h>
16 
17 
18 /*
19  * Some power8 event codes.
20  */
21 #define PM_CYC				0x0001e
22 #define PM_GCT_NOSLOT_CYC		0x100f8
23 #define PM_CMPLU_STALL			0x4000a
24 #define PM_INST_CMPL			0x00002
25 #define PM_BRU_FIN			0x10068
26 #define PM_BR_MPRED_CMPL		0x400f6
27 
28 
29 /*
30  * Raw event encoding for POWER8:
31  *
32  *        60        56        52        48        44        40        36        32
33  * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
34  *   |                                 [      thresh_cmp     ]   [  thresh_ctl   ]
35  *   |                                                                   |
36  *   *- EBB (Linux)                      thresh start/stop OR FAB match -*
37  *
38  *        28        24        20        16        12         8         4         0
39  * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
40  *   [   ] [  sample ]   [cache]   [ pmc ]   [unit ]   c     m   [    pmcxsel    ]
41  *     |        |           |                          |     |
42  *     |        |           |                          |     *- mark
43  *     |        |           *- L1/L2/L3 cache_sel      |
44  *     |        |                                      |
45  *     |        *- sampling mode for marked events     *- combine
46  *     |
47  *     *- thresh_sel
48  *
49  * Below uses IBM bit numbering.
50  *
51  * MMCR1[x:y] = unit    (PMCxUNIT)
52  * MMCR1[x]   = combine (PMCxCOMB)
53  *
54  * if pmc == 3 and unit == 0 and pmcxsel[0:6] == 0b0101011
55  *	# PM_MRK_FAB_RSP_MATCH
56  *	MMCR1[20:27] = thresh_ctl   (FAB_CRESP_MATCH / FAB_TYPE_MATCH)
57  * else if pmc == 4 and unit == 0xf and pmcxsel[0:6] == 0b0101001
58  *	# PM_MRK_FAB_RSP_MATCH_CYC
59  *	MMCR1[20:27] = thresh_ctl   (FAB_CRESP_MATCH / FAB_TYPE_MATCH)
60  * else
61  *	MMCRA[48:55] = thresh_ctl   (THRESH START/END)
62  *
63  * if thresh_sel:
64  *	MMCRA[45:47] = thresh_sel
65  *
66  * if thresh_cmp:
67  *	MMCRA[22:24] = thresh_cmp[0:2]
68  *	MMCRA[25:31] = thresh_cmp[3:9]
69  *
70  * if unit == 6 or unit == 7
71  *	MMCRC[53:55] = cache_sel[1:3]      (L2EVENT_SEL)
72  * else if unit == 8 or unit == 9:
73  *	if cache_sel[0] == 0: # L3 bank
74  *		MMCRC[47:49] = cache_sel[1:3]  (L3EVENT_SEL0)
75  *	else if cache_sel[0] == 1:
76  *		MMCRC[50:51] = cache_sel[2:3]  (L3EVENT_SEL1)
77  * else if cache_sel[1]: # L1 event
78  *	MMCR1[16] = cache_sel[2]
79  *	MMCR1[17] = cache_sel[3]
80  *
81  * if mark:
82  *	MMCRA[63]    = 1		(SAMPLE_ENABLE)
83  *	MMCRA[57:59] = sample[0:2]	(RAND_SAMP_ELIG)
84  *	MMCRA[61:62] = sample[3:4]	(RAND_SAMP_MODE)
85  *
86  */
87 
88 #define EVENT_EBB_MASK		1ull
89 #define EVENT_THR_CMP_SHIFT	40	/* Threshold CMP value */
90 #define EVENT_THR_CMP_MASK	0x3ff
91 #define EVENT_THR_CTL_SHIFT	32	/* Threshold control value (start/stop) */
92 #define EVENT_THR_CTL_MASK	0xffull
93 #define EVENT_THR_SEL_SHIFT	29	/* Threshold select value */
94 #define EVENT_THR_SEL_MASK	0x7
95 #define EVENT_THRESH_SHIFT	29	/* All threshold bits */
96 #define EVENT_THRESH_MASK	0x1fffffull
97 #define EVENT_SAMPLE_SHIFT	24	/* Sampling mode & eligibility */
98 #define EVENT_SAMPLE_MASK	0x1f
99 #define EVENT_CACHE_SEL_SHIFT	20	/* L2/L3 cache select */
100 #define EVENT_CACHE_SEL_MASK	0xf
101 #define EVENT_IS_L1		(4 << EVENT_CACHE_SEL_SHIFT)
102 #define EVENT_PMC_SHIFT		16	/* PMC number (1-based) */
103 #define EVENT_PMC_MASK		0xf
104 #define EVENT_UNIT_SHIFT	12	/* Unit */
105 #define EVENT_UNIT_MASK		0xf
106 #define EVENT_COMBINE_SHIFT	11	/* Combine bit */
107 #define EVENT_COMBINE_MASK	0x1
108 #define EVENT_MARKED_SHIFT	8	/* Marked bit */
109 #define EVENT_MARKED_MASK	0x1
110 #define EVENT_IS_MARKED		(EVENT_MARKED_MASK << EVENT_MARKED_SHIFT)
111 #define EVENT_PSEL_MASK		0xff	/* PMCxSEL value */
112 
113 #define EVENT_VALID_MASK	\
114 	((EVENT_THRESH_MASK    << EVENT_THRESH_SHIFT)		|	\
115 	 (EVENT_SAMPLE_MASK    << EVENT_SAMPLE_SHIFT)		|	\
116 	 (EVENT_CACHE_SEL_MASK << EVENT_CACHE_SEL_SHIFT)	|	\
117 	 (EVENT_PMC_MASK       << EVENT_PMC_SHIFT)		|	\
118 	 (EVENT_UNIT_MASK      << EVENT_UNIT_SHIFT)		|	\
119 	 (EVENT_COMBINE_MASK   << EVENT_COMBINE_SHIFT)		|	\
120 	 (EVENT_MARKED_MASK    << EVENT_MARKED_SHIFT)		|	\
121 	 (EVENT_EBB_MASK       << PERF_EVENT_CONFIG_EBB_SHIFT)	|	\
122 	  EVENT_PSEL_MASK)
123 
124 /* MMCRA IFM bits - POWER8 */
125 #define	POWER8_MMCRA_IFM1		0x0000000040000000UL
126 #define	POWER8_MMCRA_IFM2		0x0000000080000000UL
127 #define	POWER8_MMCRA_IFM3		0x00000000C0000000UL
128 
129 #define ONLY_PLM \
130 	(PERF_SAMPLE_BRANCH_USER        |\
131 	 PERF_SAMPLE_BRANCH_KERNEL      |\
132 	 PERF_SAMPLE_BRANCH_HV)
133 
134 /*
135  * Layout of constraint bits:
136  *
137  *        60        56        52        48        44        40        36        32
138  * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
139  *   [   fab_match   ]         [       thresh_cmp      ] [   thresh_ctl    ] [   ]
140  *                                                                             |
141  *                                                                 thresh_sel -*
142  *
143  *        28        24        20        16        12         8         4         0
144  * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
145  *                   |   [ ]   [  sample ]   [     ]   [6] [5]   [4] [3]   [2] [1]
146  *              EBB -*    |                     |
147  *                        |                     |      Count of events for each PMC.
148  *      L1 I/D qualifier -*                     |        p1, p2, p3, p4, p5, p6.
149  *                     nc - number of counters -*
150  *
151  * The PMC fields P1..P6, and NC, are adder fields. As we accumulate constraints
152  * we want the low bit of each field to be added to any existing value.
153  *
154  * Everything else is a value field.
155  */
156 
157 #define CNST_FAB_MATCH_VAL(v)	(((v) & EVENT_THR_CTL_MASK) << 56)
158 #define CNST_FAB_MATCH_MASK	CNST_FAB_MATCH_VAL(EVENT_THR_CTL_MASK)
159 
160 /* We just throw all the threshold bits into the constraint */
161 #define CNST_THRESH_VAL(v)	(((v) & EVENT_THRESH_MASK) << 32)
162 #define CNST_THRESH_MASK	CNST_THRESH_VAL(EVENT_THRESH_MASK)
163 
164 #define CNST_EBB_VAL(v)		(((v) & EVENT_EBB_MASK) << 24)
165 #define CNST_EBB_MASK		CNST_EBB_VAL(EVENT_EBB_MASK)
166 
167 #define CNST_L1_QUAL_VAL(v)	(((v) & 3) << 22)
168 #define CNST_L1_QUAL_MASK	CNST_L1_QUAL_VAL(3)
169 
170 #define CNST_SAMPLE_VAL(v)	(((v) & EVENT_SAMPLE_MASK) << 16)
171 #define CNST_SAMPLE_MASK	CNST_SAMPLE_VAL(EVENT_SAMPLE_MASK)
172 
173 /*
174  * For NC we are counting up to 4 events. This requires three bits, and we need
175  * the fifth event to overflow and set the 4th bit. To achieve that we bias the
176  * fields by 3 in test_adder.
177  */
178 #define CNST_NC_SHIFT		12
179 #define CNST_NC_VAL		(1 << CNST_NC_SHIFT)
180 #define CNST_NC_MASK		(8 << CNST_NC_SHIFT)
181 #define POWER8_TEST_ADDER	(3 << CNST_NC_SHIFT)
182 
183 /*
184  * For the per-PMC fields we have two bits. The low bit is added, so if two
185  * events ask for the same PMC the sum will overflow, setting the high bit,
186  * indicating an error. So our mask sets the high bit.
187  */
188 #define CNST_PMC_SHIFT(pmc)	((pmc - 1) * 2)
189 #define CNST_PMC_VAL(pmc)	(1 << CNST_PMC_SHIFT(pmc))
190 #define CNST_PMC_MASK(pmc)	(2 << CNST_PMC_SHIFT(pmc))
191 
192 /* Our add_fields is defined as: */
193 #define POWER8_ADD_FIELDS	\
194 	CNST_PMC_VAL(1) | CNST_PMC_VAL(2) | CNST_PMC_VAL(3) | \
195 	CNST_PMC_VAL(4) | CNST_PMC_VAL(5) | CNST_PMC_VAL(6) | CNST_NC_VAL
196 
197 
198 /* Bits in MMCR1 for POWER8 */
199 #define MMCR1_UNIT_SHIFT(pmc)		(60 - (4 * ((pmc) - 1)))
200 #define MMCR1_COMBINE_SHIFT(pmc)	(35 - ((pmc) - 1))
201 #define MMCR1_PMCSEL_SHIFT(pmc)		(24 - (((pmc) - 1)) * 8)
202 #define MMCR1_DC_QUAL_SHIFT		47
203 #define MMCR1_IC_QUAL_SHIFT		46
204 
205 /* Bits in MMCRA for POWER8 */
206 #define MMCRA_SAMP_MODE_SHIFT		1
207 #define MMCRA_SAMP_ELIG_SHIFT		4
208 #define MMCRA_THR_CTL_SHIFT		8
209 #define MMCRA_THR_SEL_SHIFT		16
210 #define MMCRA_THR_CMP_SHIFT		32
211 #define MMCRA_SDAR_MODE_TLB		(1ull << 42)
212 
213 
214 static inline bool event_is_fab_match(u64 event)
215 {
216 	/* Only check pmc, unit and pmcxsel, ignore the edge bit (0) */
217 	event &= 0xff0fe;
218 
219 	/* PM_MRK_FAB_RSP_MATCH & PM_MRK_FAB_RSP_MATCH_CYC */
220 	return (event == 0x30056 || event == 0x4f052);
221 }
222 
223 static int power8_get_constraint(u64 event, unsigned long *maskp, unsigned long *valp)
224 {
225 	unsigned int unit, pmc, cache, ebb;
226 	unsigned long mask, value;
227 
228 	mask = value = 0;
229 
230 	if (event & ~EVENT_VALID_MASK)
231 		return -1;
232 
233 	pmc   = (event >> EVENT_PMC_SHIFT)        & EVENT_PMC_MASK;
234 	unit  = (event >> EVENT_UNIT_SHIFT)       & EVENT_UNIT_MASK;
235 	cache = (event >> EVENT_CACHE_SEL_SHIFT)  & EVENT_CACHE_SEL_MASK;
236 	ebb   = (event >> PERF_EVENT_CONFIG_EBB_SHIFT) & EVENT_EBB_MASK;
237 
238 	/* Clear the EBB bit in the event, so event checks work below */
239 	event &= ~(EVENT_EBB_MASK << PERF_EVENT_CONFIG_EBB_SHIFT);
240 
241 	if (pmc) {
242 		if (pmc > 6)
243 			return -1;
244 
245 		mask  |= CNST_PMC_MASK(pmc);
246 		value |= CNST_PMC_VAL(pmc);
247 
248 		if (pmc >= 5 && event != 0x500fa && event != 0x600f4)
249 			return -1;
250 	}
251 
252 	if (pmc <= 4) {
253 		/*
254 		 * Add to number of counters in use. Note this includes events with
255 		 * a PMC of 0 - they still need a PMC, it's just assigned later.
256 		 * Don't count events on PMC 5 & 6, there is only one valid event
257 		 * on each of those counters, and they are handled above.
258 		 */
259 		mask  |= CNST_NC_MASK;
260 		value |= CNST_NC_VAL;
261 	}
262 
263 	if (unit >= 6 && unit <= 9) {
264 		/*
265 		 * L2/L3 events contain a cache selector field, which is
266 		 * supposed to be programmed into MMCRC. However MMCRC is only
267 		 * HV writable, and there is no API for guest kernels to modify
268 		 * it. The solution is for the hypervisor to initialise the
269 		 * field to zeroes, and for us to only ever allow events that
270 		 * have a cache selector of zero.
271 		 */
272 		if (cache)
273 			return -1;
274 
275 	} else if (event & EVENT_IS_L1) {
276 		mask  |= CNST_L1_QUAL_MASK;
277 		value |= CNST_L1_QUAL_VAL(cache);
278 	}
279 
280 	if (event & EVENT_IS_MARKED) {
281 		mask  |= CNST_SAMPLE_MASK;
282 		value |= CNST_SAMPLE_VAL(event >> EVENT_SAMPLE_SHIFT);
283 	}
284 
285 	/*
286 	 * Special case for PM_MRK_FAB_RSP_MATCH and PM_MRK_FAB_RSP_MATCH_CYC,
287 	 * the threshold control bits are used for the match value.
288 	 */
289 	if (event_is_fab_match(event)) {
290 		mask  |= CNST_FAB_MATCH_MASK;
291 		value |= CNST_FAB_MATCH_VAL(event >> EVENT_THR_CTL_SHIFT);
292 	} else {
293 		/*
294 		 * Check the mantissa upper two bits are not zero, unless the
295 		 * exponent is also zero. See the THRESH_CMP_MANTISSA doc.
296 		 */
297 		unsigned int cmp, exp;
298 
299 		cmp = (event >> EVENT_THR_CMP_SHIFT) & EVENT_THR_CMP_MASK;
300 		exp = cmp >> 7;
301 
302 		if (exp && (cmp & 0x60) == 0)
303 			return -1;
304 
305 		mask  |= CNST_THRESH_MASK;
306 		value |= CNST_THRESH_VAL(event >> EVENT_THRESH_SHIFT);
307 	}
308 
309 	if (!pmc && ebb)
310 		/* EBB events must specify the PMC */
311 		return -1;
312 
313 	/*
314 	 * All events must agree on EBB, either all request it or none.
315 	 * EBB events are pinned & exclusive, so this should never actually
316 	 * hit, but we leave it as a fallback in case.
317 	 */
318 	mask  |= CNST_EBB_VAL(ebb);
319 	value |= CNST_EBB_MASK;
320 
321 	*maskp = mask;
322 	*valp = value;
323 
324 	return 0;
325 }
326 
327 static int power8_compute_mmcr(u64 event[], int n_ev,
328 			       unsigned int hwc[], unsigned long mmcr[])
329 {
330 	unsigned long mmcra, mmcr1, unit, combine, psel, cache, val;
331 	unsigned int pmc, pmc_inuse;
332 	int i;
333 
334 	pmc_inuse = 0;
335 
336 	/* First pass to count resource use */
337 	for (i = 0; i < n_ev; ++i) {
338 		pmc = (event[i] >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK;
339 		if (pmc)
340 			pmc_inuse |= 1 << pmc;
341 	}
342 
343 	/* In continous sampling mode, update SDAR on TLB miss */
344 	mmcra = MMCRA_SDAR_MODE_TLB;
345 	mmcr1 = 0;
346 
347 	/* Second pass: assign PMCs, set all MMCR1 fields */
348 	for (i = 0; i < n_ev; ++i) {
349 		pmc     = (event[i] >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK;
350 		unit    = (event[i] >> EVENT_UNIT_SHIFT) & EVENT_UNIT_MASK;
351 		combine = (event[i] >> EVENT_COMBINE_SHIFT) & EVENT_COMBINE_MASK;
352 		psel    =  event[i] & EVENT_PSEL_MASK;
353 
354 		if (!pmc) {
355 			for (pmc = 1; pmc <= 4; ++pmc) {
356 				if (!(pmc_inuse & (1 << pmc)))
357 					break;
358 			}
359 
360 			pmc_inuse |= 1 << pmc;
361 		}
362 
363 		if (pmc <= 4) {
364 			mmcr1 |= unit << MMCR1_UNIT_SHIFT(pmc);
365 			mmcr1 |= combine << MMCR1_COMBINE_SHIFT(pmc);
366 			mmcr1 |= psel << MMCR1_PMCSEL_SHIFT(pmc);
367 		}
368 
369 		if (event[i] & EVENT_IS_L1) {
370 			cache = event[i] >> EVENT_CACHE_SEL_SHIFT;
371 			mmcr1 |= (cache & 1) << MMCR1_IC_QUAL_SHIFT;
372 			cache >>= 1;
373 			mmcr1 |= (cache & 1) << MMCR1_DC_QUAL_SHIFT;
374 		}
375 
376 		if (event[i] & EVENT_IS_MARKED) {
377 			mmcra |= MMCRA_SAMPLE_ENABLE;
378 
379 			val = (event[i] >> EVENT_SAMPLE_SHIFT) & EVENT_SAMPLE_MASK;
380 			if (val) {
381 				mmcra |= (val &  3) << MMCRA_SAMP_MODE_SHIFT;
382 				mmcra |= (val >> 2) << MMCRA_SAMP_ELIG_SHIFT;
383 			}
384 		}
385 
386 		/*
387 		 * PM_MRK_FAB_RSP_MATCH and PM_MRK_FAB_RSP_MATCH_CYC,
388 		 * the threshold bits are used for the match value.
389 		 */
390 		if (event_is_fab_match(event[i])) {
391 			mmcr1 |= (event[i] >> EVENT_THR_CTL_SHIFT) &
392 				  EVENT_THR_CTL_MASK;
393 		} else {
394 			val = (event[i] >> EVENT_THR_CTL_SHIFT) & EVENT_THR_CTL_MASK;
395 			mmcra |= val << MMCRA_THR_CTL_SHIFT;
396 			val = (event[i] >> EVENT_THR_SEL_SHIFT) & EVENT_THR_SEL_MASK;
397 			mmcra |= val << MMCRA_THR_SEL_SHIFT;
398 			val = (event[i] >> EVENT_THR_CMP_SHIFT) & EVENT_THR_CMP_MASK;
399 			mmcra |= val << MMCRA_THR_CMP_SHIFT;
400 		}
401 
402 		hwc[i] = pmc - 1;
403 	}
404 
405 	/* Return MMCRx values */
406 	mmcr[0] = 0;
407 
408 	/* pmc_inuse is 1-based */
409 	if (pmc_inuse & 2)
410 		mmcr[0] = MMCR0_PMC1CE;
411 
412 	if (pmc_inuse & 0x7c)
413 		mmcr[0] |= MMCR0_PMCjCE;
414 
415 	/* If we're not using PMC 5 or 6, freeze them */
416 	if (!(pmc_inuse & 0x60))
417 		mmcr[0] |= MMCR0_FC56;
418 
419 	mmcr[1] = mmcr1;
420 	mmcr[2] = mmcra;
421 
422 	return 0;
423 }
424 
425 #define MAX_ALT	2
426 
427 /* Table of alternatives, sorted by column 0 */
428 static const unsigned int event_alternatives[][MAX_ALT] = {
429 	{ 0x10134, 0x301e2 },		/* PM_MRK_ST_CMPL */
430 	{ 0x10138, 0x40138 },		/* PM_BR_MRK_2PATH */
431 	{ 0x18082, 0x3e05e },		/* PM_L3_CO_MEPF */
432 	{ 0x1d14e, 0x401e8 },		/* PM_MRK_DATA_FROM_L2MISS */
433 	{ 0x1e054, 0x4000a },		/* PM_CMPLU_STALL */
434 	{ 0x20036, 0x40036 },		/* PM_BR_2PATH */
435 	{ 0x200f2, 0x300f2 },		/* PM_INST_DISP */
436 	{ 0x200f4, 0x600f4 },		/* PM_RUN_CYC */
437 	{ 0x2013c, 0x3012e },		/* PM_MRK_FILT_MATCH */
438 	{ 0x3e054, 0x400f0 },		/* PM_LD_MISS_L1 */
439 	{ 0x400fa, 0x500fa },		/* PM_RUN_INST_CMPL */
440 };
441 
442 /*
443  * Scan the alternatives table for a match and return the
444  * index into the alternatives table if found, else -1.
445  */
446 static int find_alternative(u64 event)
447 {
448 	int i, j;
449 
450 	for (i = 0; i < ARRAY_SIZE(event_alternatives); ++i) {
451 		if (event < event_alternatives[i][0])
452 			break;
453 
454 		for (j = 0; j < MAX_ALT && event_alternatives[i][j]; ++j)
455 			if (event == event_alternatives[i][j])
456 				return i;
457 	}
458 
459 	return -1;
460 }
461 
462 static int power8_get_alternatives(u64 event, unsigned int flags, u64 alt[])
463 {
464 	int i, j, num_alt = 0;
465 	u64 alt_event;
466 
467 	alt[num_alt++] = event;
468 
469 	i = find_alternative(event);
470 	if (i >= 0) {
471 		/* Filter out the original event, it's already in alt[0] */
472 		for (j = 0; j < MAX_ALT; ++j) {
473 			alt_event = event_alternatives[i][j];
474 			if (alt_event && alt_event != event)
475 				alt[num_alt++] = alt_event;
476 		}
477 	}
478 
479 	if (flags & PPMU_ONLY_COUNT_RUN) {
480 		/*
481 		 * We're only counting in RUN state, so PM_CYC is equivalent to
482 		 * PM_RUN_CYC and PM_INST_CMPL === PM_RUN_INST_CMPL.
483 		 */
484 		j = num_alt;
485 		for (i = 0; i < num_alt; ++i) {
486 			switch (alt[i]) {
487 			case 0x1e:	/* PM_CYC */
488 				alt[j++] = 0x600f4;	/* PM_RUN_CYC */
489 				break;
490 			case 0x600f4:	/* PM_RUN_CYC */
491 				alt[j++] = 0x1e;
492 				break;
493 			case 0x2:	/* PM_PPC_CMPL */
494 				alt[j++] = 0x500fa;	/* PM_RUN_INST_CMPL */
495 				break;
496 			case 0x500fa:	/* PM_RUN_INST_CMPL */
497 				alt[j++] = 0x2;	/* PM_PPC_CMPL */
498 				break;
499 			}
500 		}
501 		num_alt = j;
502 	}
503 
504 	return num_alt;
505 }
506 
507 static void power8_disable_pmc(unsigned int pmc, unsigned long mmcr[])
508 {
509 	if (pmc <= 3)
510 		mmcr[1] &= ~(0xffUL << MMCR1_PMCSEL_SHIFT(pmc + 1));
511 }
512 
513 PMU_FORMAT_ATTR(event,		"config:0-49");
514 PMU_FORMAT_ATTR(pmcxsel,	"config:0-7");
515 PMU_FORMAT_ATTR(mark,		"config:8");
516 PMU_FORMAT_ATTR(combine,	"config:11");
517 PMU_FORMAT_ATTR(unit,		"config:12-15");
518 PMU_FORMAT_ATTR(pmc,		"config:16-19");
519 PMU_FORMAT_ATTR(cache_sel,	"config:20-23");
520 PMU_FORMAT_ATTR(sample_mode,	"config:24-28");
521 PMU_FORMAT_ATTR(thresh_sel,	"config:29-31");
522 PMU_FORMAT_ATTR(thresh_stop,	"config:32-35");
523 PMU_FORMAT_ATTR(thresh_start,	"config:36-39");
524 PMU_FORMAT_ATTR(thresh_cmp,	"config:40-49");
525 
526 static struct attribute *power8_pmu_format_attr[] = {
527 	&format_attr_event.attr,
528 	&format_attr_pmcxsel.attr,
529 	&format_attr_mark.attr,
530 	&format_attr_combine.attr,
531 	&format_attr_unit.attr,
532 	&format_attr_pmc.attr,
533 	&format_attr_cache_sel.attr,
534 	&format_attr_sample_mode.attr,
535 	&format_attr_thresh_sel.attr,
536 	&format_attr_thresh_stop.attr,
537 	&format_attr_thresh_start.attr,
538 	&format_attr_thresh_cmp.attr,
539 	NULL,
540 };
541 
542 struct attribute_group power8_pmu_format_group = {
543 	.name = "format",
544 	.attrs = power8_pmu_format_attr,
545 };
546 
547 static const struct attribute_group *power8_pmu_attr_groups[] = {
548 	&power8_pmu_format_group,
549 	NULL,
550 };
551 
552 static int power8_generic_events[] = {
553 	[PERF_COUNT_HW_CPU_CYCLES] =			PM_CYC,
554 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =	PM_GCT_NOSLOT_CYC,
555 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =	PM_CMPLU_STALL,
556 	[PERF_COUNT_HW_INSTRUCTIONS] =			PM_INST_CMPL,
557 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] =		PM_BRU_FIN,
558 	[PERF_COUNT_HW_BRANCH_MISSES] =			PM_BR_MPRED_CMPL,
559 };
560 
561 static u64 power8_bhrb_filter_map(u64 branch_sample_type)
562 {
563 	u64 pmu_bhrb_filter = 0;
564 
565 	/* BHRB and regular PMU events share the same privilege state
566 	 * filter configuration. BHRB is always recorded along with a
567 	 * regular PMU event. As the privilege state filter is handled
568 	 * in the basic PMC configuration of the accompanying regular
569 	 * PMU event, we ignore any separate BHRB specific request.
570 	 */
571 
572 	/* No branch filter requested */
573 	if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY)
574 		return pmu_bhrb_filter;
575 
576 	/* Invalid branch filter options - HW does not support */
577 	if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_RETURN)
578 		return -1;
579 
580 	if (branch_sample_type & PERF_SAMPLE_BRANCH_IND_CALL)
581 		return -1;
582 
583 	if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_CALL) {
584 		pmu_bhrb_filter |= POWER8_MMCRA_IFM1;
585 		return pmu_bhrb_filter;
586 	}
587 
588 	/* Every thing else is unsupported */
589 	return -1;
590 }
591 
592 static void power8_config_bhrb(u64 pmu_bhrb_filter)
593 {
594 	/* Enable BHRB filter in PMU */
595 	mtspr(SPRN_MMCRA, (mfspr(SPRN_MMCRA) | pmu_bhrb_filter));
596 }
597 
598 static struct power_pmu power8_pmu = {
599 	.name			= "POWER8",
600 	.n_counter		= 6,
601 	.max_alternatives	= MAX_ALT + 1,
602 	.add_fields		= POWER8_ADD_FIELDS,
603 	.test_adder		= POWER8_TEST_ADDER,
604 	.compute_mmcr		= power8_compute_mmcr,
605 	.config_bhrb		= power8_config_bhrb,
606 	.bhrb_filter_map	= power8_bhrb_filter_map,
607 	.get_constraint		= power8_get_constraint,
608 	.get_alternatives	= power8_get_alternatives,
609 	.disable_pmc		= power8_disable_pmc,
610 	.flags			= PPMU_HAS_SSLOT | PPMU_HAS_SIER | PPMU_BHRB | PPMU_EBB,
611 	.n_generic		= ARRAY_SIZE(power8_generic_events),
612 	.generic_events		= power8_generic_events,
613 	.attr_groups		= power8_pmu_attr_groups,
614 	.bhrb_nr		= 32,
615 };
616 
617 static int __init init_power8_pmu(void)
618 {
619 	int rc;
620 
621 	if (!cur_cpu_spec->oprofile_cpu_type ||
622 	    strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power8"))
623 		return -ENODEV;
624 
625 	rc = register_power_pmu(&power8_pmu);
626 	if (rc)
627 		return rc;
628 
629 	/* Tell userspace that EBB is supported */
630 	cur_cpu_spec->cpu_user_features2 |= PPC_FEATURE2_EBB;
631 
632 	return 0;
633 }
634 early_initcall(init_power8_pmu);
635