xref: /openbmc/linux/arch/powerpc/perf/power8-pmu.c (revision ee89bd6b)
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  *                                       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_THR_CMP_SHIFT	40	/* Threshold CMP value */
89 #define EVENT_THR_CMP_MASK	0x3ff
90 #define EVENT_THR_CTL_SHIFT	32	/* Threshold control value (start/stop) */
91 #define EVENT_THR_CTL_MASK	0xffull
92 #define EVENT_THR_SEL_SHIFT	29	/* Threshold select value */
93 #define EVENT_THR_SEL_MASK	0x7
94 #define EVENT_THRESH_SHIFT	29	/* All threshold bits */
95 #define EVENT_THRESH_MASK	0x1fffffull
96 #define EVENT_SAMPLE_SHIFT	24	/* Sampling mode & eligibility */
97 #define EVENT_SAMPLE_MASK	0x1f
98 #define EVENT_CACHE_SEL_SHIFT	20	/* L2/L3 cache select */
99 #define EVENT_CACHE_SEL_MASK	0xf
100 #define EVENT_IS_L1		(4 << EVENT_CACHE_SEL_SHIFT)
101 #define EVENT_PMC_SHIFT		16	/* PMC number (1-based) */
102 #define EVENT_PMC_MASK		0xf
103 #define EVENT_UNIT_SHIFT	12	/* Unit */
104 #define EVENT_UNIT_MASK		0xf
105 #define EVENT_COMBINE_SHIFT	11	/* Combine bit */
106 #define EVENT_COMBINE_MASK	0x1
107 #define EVENT_MARKED_SHIFT	8	/* Marked bit */
108 #define EVENT_MARKED_MASK	0x1
109 #define EVENT_IS_MARKED		(EVENT_MARKED_MASK << EVENT_MARKED_SHIFT)
110 #define EVENT_PSEL_MASK		0xff	/* PMCxSEL value */
111 
112 /* MMCRA IFM bits - POWER8 */
113 #define	POWER8_MMCRA_IFM1		0x0000000040000000UL
114 #define	POWER8_MMCRA_IFM2		0x0000000080000000UL
115 #define	POWER8_MMCRA_IFM3		0x00000000C0000000UL
116 
117 #define ONLY_PLM \
118 	(PERF_SAMPLE_BRANCH_USER        |\
119 	 PERF_SAMPLE_BRANCH_KERNEL      |\
120 	 PERF_SAMPLE_BRANCH_HV)
121 
122 /*
123  * Layout of constraint bits:
124  *
125  *        60        56        52        48        44        40        36        32
126  * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
127  *   [   fab_match   ]         [       thresh_cmp      ] [   thresh_ctl    ] [   ]
128  *                                                                             |
129  *                                                                 thresh_sel -*
130  *
131  *        28        24        20        16        12         8         4         0
132  * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
133  *                       [ ]   [  sample ]   [     ]   [6] [5]   [4] [3]   [2] [1]
134  *                        |                     |
135  *      L1 I/D qualifier -*                     |      Count of events for each PMC.
136  *                                              |        p1, p2, p3, p4, p5, p6.
137  *                     nc - number of counters -*
138  *
139  * The PMC fields P1..P6, and NC, are adder fields. As we accumulate constraints
140  * we want the low bit of each field to be added to any existing value.
141  *
142  * Everything else is a value field.
143  */
144 
145 #define CNST_FAB_MATCH_VAL(v)	(((v) & EVENT_THR_CTL_MASK) << 56)
146 #define CNST_FAB_MATCH_MASK	CNST_FAB_MATCH_VAL(EVENT_THR_CTL_MASK)
147 
148 /* We just throw all the threshold bits into the constraint */
149 #define CNST_THRESH_VAL(v)	(((v) & EVENT_THRESH_MASK) << 32)
150 #define CNST_THRESH_MASK	CNST_THRESH_VAL(EVENT_THRESH_MASK)
151 
152 #define CNST_L1_QUAL_VAL(v)	(((v) & 3) << 22)
153 #define CNST_L1_QUAL_MASK	CNST_L1_QUAL_VAL(3)
154 
155 #define CNST_SAMPLE_VAL(v)	(((v) & EVENT_SAMPLE_MASK) << 16)
156 #define CNST_SAMPLE_MASK	CNST_SAMPLE_VAL(EVENT_SAMPLE_MASK)
157 
158 /*
159  * For NC we are counting up to 4 events. This requires three bits, and we need
160  * the fifth event to overflow and set the 4th bit. To achieve that we bias the
161  * fields by 3 in test_adder.
162  */
163 #define CNST_NC_SHIFT		12
164 #define CNST_NC_VAL		(1 << CNST_NC_SHIFT)
165 #define CNST_NC_MASK		(8 << CNST_NC_SHIFT)
166 #define POWER8_TEST_ADDER	(3 << CNST_NC_SHIFT)
167 
168 /*
169  * For the per-PMC fields we have two bits. The low bit is added, so if two
170  * events ask for the same PMC the sum will overflow, setting the high bit,
171  * indicating an error. So our mask sets the high bit.
172  */
173 #define CNST_PMC_SHIFT(pmc)	((pmc - 1) * 2)
174 #define CNST_PMC_VAL(pmc)	(1 << CNST_PMC_SHIFT(pmc))
175 #define CNST_PMC_MASK(pmc)	(2 << CNST_PMC_SHIFT(pmc))
176 
177 /* Our add_fields is defined as: */
178 #define POWER8_ADD_FIELDS	\
179 	CNST_PMC_VAL(1) | CNST_PMC_VAL(2) | CNST_PMC_VAL(3) | \
180 	CNST_PMC_VAL(4) | CNST_PMC_VAL(5) | CNST_PMC_VAL(6) | CNST_NC_VAL
181 
182 
183 /* Bits in MMCR1 for POWER8 */
184 #define MMCR1_UNIT_SHIFT(pmc)		(60 - (4 * ((pmc) - 1)))
185 #define MMCR1_COMBINE_SHIFT(pmc)	(35 - ((pmc) - 1))
186 #define MMCR1_PMCSEL_SHIFT(pmc)		(24 - (((pmc) - 1)) * 8)
187 #define MMCR1_DC_QUAL_SHIFT		47
188 #define MMCR1_IC_QUAL_SHIFT		46
189 
190 /* Bits in MMCRA for POWER8 */
191 #define MMCRA_SAMP_MODE_SHIFT		1
192 #define MMCRA_SAMP_ELIG_SHIFT		4
193 #define MMCRA_THR_CTL_SHIFT		8
194 #define MMCRA_THR_SEL_SHIFT		16
195 #define MMCRA_THR_CMP_SHIFT		32
196 #define MMCRA_SDAR_MODE_TLB		(1ull << 42)
197 
198 
199 static inline bool event_is_fab_match(u64 event)
200 {
201 	/* Only check pmc, unit and pmcxsel, ignore the edge bit (0) */
202 	event &= 0xff0fe;
203 
204 	/* PM_MRK_FAB_RSP_MATCH & PM_MRK_FAB_RSP_MATCH_CYC */
205 	return (event == 0x30056 || event == 0x4f052);
206 }
207 
208 static int power8_get_constraint(u64 event, unsigned long *maskp, unsigned long *valp)
209 {
210 	unsigned int unit, pmc, cache;
211 	unsigned long mask, value;
212 
213 	mask = value = 0;
214 
215 	pmc   = (event >> EVENT_PMC_SHIFT)       & EVENT_PMC_MASK;
216 	unit  = (event >> EVENT_UNIT_SHIFT)      & EVENT_UNIT_MASK;
217 	cache = (event >> EVENT_CACHE_SEL_SHIFT) & EVENT_CACHE_SEL_MASK;
218 
219 	if (pmc) {
220 		if (pmc > 6)
221 			return -1;
222 
223 		mask  |= CNST_PMC_MASK(pmc);
224 		value |= CNST_PMC_VAL(pmc);
225 
226 		if (pmc >= 5 && event != 0x500fa && event != 0x600f4)
227 			return -1;
228 	}
229 
230 	if (pmc <= 4) {
231 		/*
232 		 * Add to number of counters in use. Note this includes events with
233 		 * a PMC of 0 - they still need a PMC, it's just assigned later.
234 		 * Don't count events on PMC 5 & 6, there is only one valid event
235 		 * on each of those counters, and they are handled above.
236 		 */
237 		mask  |= CNST_NC_MASK;
238 		value |= CNST_NC_VAL;
239 	}
240 
241 	if (unit >= 6 && unit <= 9) {
242 		/*
243 		 * L2/L3 events contain a cache selector field, which is
244 		 * supposed to be programmed into MMCRC. However MMCRC is only
245 		 * HV writable, and there is no API for guest kernels to modify
246 		 * it. The solution is for the hypervisor to initialise the
247 		 * field to zeroes, and for us to only ever allow events that
248 		 * have a cache selector of zero.
249 		 */
250 		if (cache)
251 			return -1;
252 
253 	} else if (event & EVENT_IS_L1) {
254 		mask  |= CNST_L1_QUAL_MASK;
255 		value |= CNST_L1_QUAL_VAL(cache);
256 	}
257 
258 	if (event & EVENT_IS_MARKED) {
259 		mask  |= CNST_SAMPLE_MASK;
260 		value |= CNST_SAMPLE_VAL(event >> EVENT_SAMPLE_SHIFT);
261 	}
262 
263 	/*
264 	 * Special case for PM_MRK_FAB_RSP_MATCH and PM_MRK_FAB_RSP_MATCH_CYC,
265 	 * the threshold control bits are used for the match value.
266 	 */
267 	if (event_is_fab_match(event)) {
268 		mask  |= CNST_FAB_MATCH_MASK;
269 		value |= CNST_FAB_MATCH_VAL(event >> EVENT_THR_CTL_SHIFT);
270 	} else {
271 		/*
272 		 * Check the mantissa upper two bits are not zero, unless the
273 		 * exponent is also zero. See the THRESH_CMP_MANTISSA doc.
274 		 */
275 		unsigned int cmp, exp;
276 
277 		cmp = (event >> EVENT_THR_CMP_SHIFT) & EVENT_THR_CMP_MASK;
278 		exp = cmp >> 7;
279 
280 		if (exp && (cmp & 0x60) == 0)
281 			return -1;
282 
283 		mask  |= CNST_THRESH_MASK;
284 		value |= CNST_THRESH_VAL(event >> EVENT_THRESH_SHIFT);
285 	}
286 
287 	*maskp = mask;
288 	*valp = value;
289 
290 	return 0;
291 }
292 
293 static int power8_compute_mmcr(u64 event[], int n_ev,
294 			       unsigned int hwc[], unsigned long mmcr[])
295 {
296 	unsigned long mmcra, mmcr1, unit, combine, psel, cache, val;
297 	unsigned int pmc, pmc_inuse;
298 	int i;
299 
300 	pmc_inuse = 0;
301 
302 	/* First pass to count resource use */
303 	for (i = 0; i < n_ev; ++i) {
304 		pmc = (event[i] >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK;
305 		if (pmc)
306 			pmc_inuse |= 1 << pmc;
307 	}
308 
309 	/* In continous sampling mode, update SDAR on TLB miss */
310 	mmcra = MMCRA_SDAR_MODE_TLB;
311 	mmcr1 = 0;
312 
313 	/* Second pass: assign PMCs, set all MMCR1 fields */
314 	for (i = 0; i < n_ev; ++i) {
315 		pmc     = (event[i] >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK;
316 		unit    = (event[i] >> EVENT_UNIT_SHIFT) & EVENT_UNIT_MASK;
317 		combine = (event[i] >> EVENT_COMBINE_SHIFT) & EVENT_COMBINE_MASK;
318 		psel    =  event[i] & EVENT_PSEL_MASK;
319 
320 		if (!pmc) {
321 			for (pmc = 1; pmc <= 4; ++pmc) {
322 				if (!(pmc_inuse & (1 << pmc)))
323 					break;
324 			}
325 
326 			pmc_inuse |= 1 << pmc;
327 		}
328 
329 		if (pmc <= 4) {
330 			mmcr1 |= unit << MMCR1_UNIT_SHIFT(pmc);
331 			mmcr1 |= combine << MMCR1_COMBINE_SHIFT(pmc);
332 			mmcr1 |= psel << MMCR1_PMCSEL_SHIFT(pmc);
333 		}
334 
335 		if (event[i] & EVENT_IS_L1) {
336 			cache = event[i] >> EVENT_CACHE_SEL_SHIFT;
337 			mmcr1 |= (cache & 1) << MMCR1_IC_QUAL_SHIFT;
338 			cache >>= 1;
339 			mmcr1 |= (cache & 1) << MMCR1_DC_QUAL_SHIFT;
340 		}
341 
342 		if (event[i] & EVENT_IS_MARKED) {
343 			mmcra |= MMCRA_SAMPLE_ENABLE;
344 
345 			val = (event[i] >> EVENT_SAMPLE_SHIFT) & EVENT_SAMPLE_MASK;
346 			if (val) {
347 				mmcra |= (val &  3) << MMCRA_SAMP_MODE_SHIFT;
348 				mmcra |= (val >> 2) << MMCRA_SAMP_ELIG_SHIFT;
349 			}
350 		}
351 
352 		/*
353 		 * PM_MRK_FAB_RSP_MATCH and PM_MRK_FAB_RSP_MATCH_CYC,
354 		 * the threshold bits are used for the match value.
355 		 */
356 		if (event_is_fab_match(event[i])) {
357 			mmcr1 |= (event[i] >> EVENT_THR_CTL_SHIFT) &
358 				  EVENT_THR_CTL_MASK;
359 		} else {
360 			val = (event[i] >> EVENT_THR_CTL_SHIFT) & EVENT_THR_CTL_MASK;
361 			mmcra |= val << MMCRA_THR_CTL_SHIFT;
362 			val = (event[i] >> EVENT_THR_SEL_SHIFT) & EVENT_THR_SEL_MASK;
363 			mmcra |= val << MMCRA_THR_SEL_SHIFT;
364 			val = (event[i] >> EVENT_THR_CMP_SHIFT) & EVENT_THR_CMP_MASK;
365 			mmcra |= val << MMCRA_THR_CMP_SHIFT;
366 		}
367 
368 		hwc[i] = pmc - 1;
369 	}
370 
371 	/* Return MMCRx values */
372 	mmcr[0] = 0;
373 
374 	/* pmc_inuse is 1-based */
375 	if (pmc_inuse & 2)
376 		mmcr[0] = MMCR0_PMC1CE;
377 
378 	if (pmc_inuse & 0x7c)
379 		mmcr[0] |= MMCR0_PMCjCE;
380 
381 	mmcr[1] = mmcr1;
382 	mmcr[2] = mmcra;
383 
384 	return 0;
385 }
386 
387 #define MAX_ALT	2
388 
389 /* Table of alternatives, sorted by column 0 */
390 static const unsigned int event_alternatives[][MAX_ALT] = {
391 	{ 0x10134, 0x301e2 },		/* PM_MRK_ST_CMPL */
392 	{ 0x10138, 0x40138 },		/* PM_BR_MRK_2PATH */
393 	{ 0x18082, 0x3e05e },		/* PM_L3_CO_MEPF */
394 	{ 0x1d14e, 0x401e8 },		/* PM_MRK_DATA_FROM_L2MISS */
395 	{ 0x1e054, 0x4000a },		/* PM_CMPLU_STALL */
396 	{ 0x20036, 0x40036 },		/* PM_BR_2PATH */
397 	{ 0x200f2, 0x300f2 },		/* PM_INST_DISP */
398 	{ 0x200f4, 0x600f4 },		/* PM_RUN_CYC */
399 	{ 0x2013c, 0x3012e },		/* PM_MRK_FILT_MATCH */
400 	{ 0x3e054, 0x400f0 },		/* PM_LD_MISS_L1 */
401 	{ 0x400fa, 0x500fa },		/* PM_RUN_INST_CMPL */
402 };
403 
404 /*
405  * Scan the alternatives table for a match and return the
406  * index into the alternatives table if found, else -1.
407  */
408 static int find_alternative(u64 event)
409 {
410 	int i, j;
411 
412 	for (i = 0; i < ARRAY_SIZE(event_alternatives); ++i) {
413 		if (event < event_alternatives[i][0])
414 			break;
415 
416 		for (j = 0; j < MAX_ALT && event_alternatives[i][j]; ++j)
417 			if (event == event_alternatives[i][j])
418 				return i;
419 	}
420 
421 	return -1;
422 }
423 
424 static int power8_get_alternatives(u64 event, unsigned int flags, u64 alt[])
425 {
426 	int i, j, num_alt = 0;
427 	u64 alt_event;
428 
429 	alt[num_alt++] = event;
430 
431 	i = find_alternative(event);
432 	if (i >= 0) {
433 		/* Filter out the original event, it's already in alt[0] */
434 		for (j = 0; j < MAX_ALT; ++j) {
435 			alt_event = event_alternatives[i][j];
436 			if (alt_event && alt_event != event)
437 				alt[num_alt++] = alt_event;
438 		}
439 	}
440 
441 	if (flags & PPMU_ONLY_COUNT_RUN) {
442 		/*
443 		 * We're only counting in RUN state, so PM_CYC is equivalent to
444 		 * PM_RUN_CYC and PM_INST_CMPL === PM_RUN_INST_CMPL.
445 		 */
446 		j = num_alt;
447 		for (i = 0; i < num_alt; ++i) {
448 			switch (alt[i]) {
449 			case 0x1e:	/* PM_CYC */
450 				alt[j++] = 0x600f4;	/* PM_RUN_CYC */
451 				break;
452 			case 0x600f4:	/* PM_RUN_CYC */
453 				alt[j++] = 0x1e;
454 				break;
455 			case 0x2:	/* PM_PPC_CMPL */
456 				alt[j++] = 0x500fa;	/* PM_RUN_INST_CMPL */
457 				break;
458 			case 0x500fa:	/* PM_RUN_INST_CMPL */
459 				alt[j++] = 0x2;	/* PM_PPC_CMPL */
460 				break;
461 			}
462 		}
463 		num_alt = j;
464 	}
465 
466 	return num_alt;
467 }
468 
469 static void power8_disable_pmc(unsigned int pmc, unsigned long mmcr[])
470 {
471 	if (pmc <= 3)
472 		mmcr[1] &= ~(0xffUL << MMCR1_PMCSEL_SHIFT(pmc + 1));
473 }
474 
475 PMU_FORMAT_ATTR(event,		"config:0-49");
476 PMU_FORMAT_ATTR(pmcxsel,	"config:0-7");
477 PMU_FORMAT_ATTR(mark,		"config:8");
478 PMU_FORMAT_ATTR(combine,	"config:11");
479 PMU_FORMAT_ATTR(unit,		"config:12-15");
480 PMU_FORMAT_ATTR(pmc,		"config:16-19");
481 PMU_FORMAT_ATTR(cache_sel,	"config:20-23");
482 PMU_FORMAT_ATTR(sample_mode,	"config:24-28");
483 PMU_FORMAT_ATTR(thresh_sel,	"config:29-31");
484 PMU_FORMAT_ATTR(thresh_stop,	"config:32-35");
485 PMU_FORMAT_ATTR(thresh_start,	"config:36-39");
486 PMU_FORMAT_ATTR(thresh_cmp,	"config:40-49");
487 
488 static struct attribute *power8_pmu_format_attr[] = {
489 	&format_attr_event.attr,
490 	&format_attr_pmcxsel.attr,
491 	&format_attr_mark.attr,
492 	&format_attr_combine.attr,
493 	&format_attr_unit.attr,
494 	&format_attr_pmc.attr,
495 	&format_attr_cache_sel.attr,
496 	&format_attr_sample_mode.attr,
497 	&format_attr_thresh_sel.attr,
498 	&format_attr_thresh_stop.attr,
499 	&format_attr_thresh_start.attr,
500 	&format_attr_thresh_cmp.attr,
501 	NULL,
502 };
503 
504 struct attribute_group power8_pmu_format_group = {
505 	.name = "format",
506 	.attrs = power8_pmu_format_attr,
507 };
508 
509 static const struct attribute_group *power8_pmu_attr_groups[] = {
510 	&power8_pmu_format_group,
511 	NULL,
512 };
513 
514 static int power8_generic_events[] = {
515 	[PERF_COUNT_HW_CPU_CYCLES] =			PM_CYC,
516 	[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =	PM_GCT_NOSLOT_CYC,
517 	[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =	PM_CMPLU_STALL,
518 	[PERF_COUNT_HW_INSTRUCTIONS] =			PM_INST_CMPL,
519 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] =		PM_BRU_FIN,
520 	[PERF_COUNT_HW_BRANCH_MISSES] =			PM_BR_MPRED_CMPL,
521 };
522 
523 static u64 power8_bhrb_filter_map(u64 branch_sample_type)
524 {
525 	u64 pmu_bhrb_filter = 0;
526 	u64 br_privilege = branch_sample_type & ONLY_PLM;
527 
528 	/* BHRB and regular PMU events share the same prvillege state
529 	 * filter configuration. BHRB is always recorded along with a
530 	 * regular PMU event. So privilege state filter criteria for BHRB
531 	 * and the companion PMU events has to be the same. As a default
532 	 * "perf record" tool sets all privillege bits ON when no filter
533 	 * criteria is provided in the command line. So as along as all
534 	 * privillege bits are ON or they are OFF, we are good to go.
535 	 */
536 	if ((br_privilege != 7) && (br_privilege != 0))
537 		return -1;
538 
539 	/* No branch filter requested */
540 	if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY)
541 		return pmu_bhrb_filter;
542 
543 	/* Invalid branch filter options - HW does not support */
544 	if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_RETURN)
545 		return -1;
546 
547 	if (branch_sample_type & PERF_SAMPLE_BRANCH_IND_CALL)
548 		return -1;
549 
550 	if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_CALL) {
551 		pmu_bhrb_filter |= POWER8_MMCRA_IFM1;
552 		return pmu_bhrb_filter;
553 	}
554 
555 	/* Every thing else is unsupported */
556 	return -1;
557 }
558 
559 static void power8_config_bhrb(u64 pmu_bhrb_filter)
560 {
561 	/* Enable BHRB filter in PMU */
562 	mtspr(SPRN_MMCRA, (mfspr(SPRN_MMCRA) | pmu_bhrb_filter));
563 }
564 
565 static struct power_pmu power8_pmu = {
566 	.name			= "POWER8",
567 	.n_counter		= 6,
568 	.max_alternatives	= MAX_ALT + 1,
569 	.add_fields		= POWER8_ADD_FIELDS,
570 	.test_adder		= POWER8_TEST_ADDER,
571 	.compute_mmcr		= power8_compute_mmcr,
572 	.config_bhrb		= power8_config_bhrb,
573 	.bhrb_filter_map	= power8_bhrb_filter_map,
574 	.get_constraint		= power8_get_constraint,
575 	.get_alternatives	= power8_get_alternatives,
576 	.disable_pmc		= power8_disable_pmc,
577 	.flags			= PPMU_HAS_SSLOT | PPMU_HAS_SIER | PPMU_BHRB,
578 	.n_generic		= ARRAY_SIZE(power8_generic_events),
579 	.generic_events		= power8_generic_events,
580 	.attr_groups		= power8_pmu_attr_groups,
581 	.bhrb_nr		= 32,
582 };
583 
584 static int __init init_power8_pmu(void)
585 {
586 	if (!cur_cpu_spec->oprofile_cpu_type ||
587 	    strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power8"))
588 		return -ENODEV;
589 
590 	return register_power_pmu(&power8_pmu);
591 }
592 early_initcall(init_power8_pmu);
593