1 /*
2  * Common Performance counter support functions for PowerISA v2.07 processors.
3  *
4  * Copyright 2009 Paul Mackerras, IBM Corporation.
5  * Copyright 2013 Michael Ellerman, IBM Corporation.
6  * Copyright 2016 Madhavan Srinivasan, IBM Corporation.
7  *
8  * This program is free software; you can redistribute it and/or
9  * modify it under the terms of the GNU General Public License
10  * as published by the Free Software Foundation; either version
11  * 2 of the License, or (at your option) any later version.
12  */
13 #include "isa207-common.h"
14 
15 PMU_FORMAT_ATTR(event,		"config:0-49");
16 PMU_FORMAT_ATTR(pmcxsel,	"config:0-7");
17 PMU_FORMAT_ATTR(mark,		"config:8");
18 PMU_FORMAT_ATTR(combine,	"config:11");
19 PMU_FORMAT_ATTR(unit,		"config:12-15");
20 PMU_FORMAT_ATTR(pmc,		"config:16-19");
21 PMU_FORMAT_ATTR(cache_sel,	"config:20-23");
22 PMU_FORMAT_ATTR(sample_mode,	"config:24-28");
23 PMU_FORMAT_ATTR(thresh_sel,	"config:29-31");
24 PMU_FORMAT_ATTR(thresh_stop,	"config:32-35");
25 PMU_FORMAT_ATTR(thresh_start,	"config:36-39");
26 PMU_FORMAT_ATTR(thresh_cmp,	"config:40-49");
27 
28 struct attribute *isa207_pmu_format_attr[] = {
29 	&format_attr_event.attr,
30 	&format_attr_pmcxsel.attr,
31 	&format_attr_mark.attr,
32 	&format_attr_combine.attr,
33 	&format_attr_unit.attr,
34 	&format_attr_pmc.attr,
35 	&format_attr_cache_sel.attr,
36 	&format_attr_sample_mode.attr,
37 	&format_attr_thresh_sel.attr,
38 	&format_attr_thresh_stop.attr,
39 	&format_attr_thresh_start.attr,
40 	&format_attr_thresh_cmp.attr,
41 	NULL,
42 };
43 
44 struct attribute_group isa207_pmu_format_group = {
45 	.name = "format",
46 	.attrs = isa207_pmu_format_attr,
47 };
48 
49 static inline bool event_is_fab_match(u64 event)
50 {
51 	/* Only check pmc, unit and pmcxsel, ignore the edge bit (0) */
52 	event &= 0xff0fe;
53 
54 	/* PM_MRK_FAB_RSP_MATCH & PM_MRK_FAB_RSP_MATCH_CYC */
55 	return (event == 0x30056 || event == 0x4f052);
56 }
57 
58 static bool is_event_valid(u64 event)
59 {
60 	u64 valid_mask = EVENT_VALID_MASK;
61 
62 	if (cpu_has_feature(CPU_FTR_ARCH_300))
63 		valid_mask = p9_EVENT_VALID_MASK;
64 
65 	return !(event & ~valid_mask);
66 }
67 
68 static inline bool is_event_marked(u64 event)
69 {
70 	if (event & EVENT_IS_MARKED)
71 		return true;
72 
73 	return false;
74 }
75 
76 static void mmcra_sdar_mode(u64 event, unsigned long *mmcra)
77 {
78 	/*
79 	 * MMCRA[SDAR_MODE] specifices how the SDAR should be updated in
80 	 * continous sampling mode.
81 	 *
82 	 * Incase of Power8:
83 	 * MMCRA[SDAR_MODE] will be programmed as "0b01" for continous sampling
84 	 * mode and will be un-changed when setting MMCRA[63] (Marked events).
85 	 *
86 	 * Incase of Power9:
87 	 * Marked event: MMCRA[SDAR_MODE] will be set to 0b00 ('No Updates'),
88 	 *               or if group already have any marked events.
89 	 * For rest
90 	 *	MMCRA[SDAR_MODE] will be set from event code.
91 	 *      If sdar_mode from event is zero, default to 0b01. Hardware
92 	 *      requires that we set a non-zero value.
93 	 */
94 	if (cpu_has_feature(CPU_FTR_ARCH_300)) {
95 		if (is_event_marked(event) || (*mmcra & MMCRA_SAMPLE_ENABLE))
96 			*mmcra &= MMCRA_SDAR_MODE_NO_UPDATES;
97 		else if (p9_SDAR_MODE(event))
98 			*mmcra |=  p9_SDAR_MODE(event) << MMCRA_SDAR_MODE_SHIFT;
99 		else
100 			*mmcra |= MMCRA_SDAR_MODE_DCACHE;
101 	} else
102 		*mmcra |= MMCRA_SDAR_MODE_TLB;
103 }
104 
105 static u64 thresh_cmp_val(u64 value)
106 {
107 	if (cpu_has_feature(CPU_FTR_ARCH_300))
108 		return value << p9_MMCRA_THR_CMP_SHIFT;
109 
110 	return value << MMCRA_THR_CMP_SHIFT;
111 }
112 
113 static unsigned long combine_from_event(u64 event)
114 {
115 	if (cpu_has_feature(CPU_FTR_ARCH_300))
116 		return p9_EVENT_COMBINE(event);
117 
118 	return EVENT_COMBINE(event);
119 }
120 
121 static unsigned long combine_shift(unsigned long pmc)
122 {
123 	if (cpu_has_feature(CPU_FTR_ARCH_300))
124 		return p9_MMCR1_COMBINE_SHIFT(pmc);
125 
126 	return MMCR1_COMBINE_SHIFT(pmc);
127 }
128 
129 static inline bool event_is_threshold(u64 event)
130 {
131 	return (event >> EVENT_THR_SEL_SHIFT) & EVENT_THR_SEL_MASK;
132 }
133 
134 static bool is_thresh_cmp_valid(u64 event)
135 {
136 	unsigned int cmp, exp;
137 
138 	/*
139 	 * Check the mantissa upper two bits are not zero, unless the
140 	 * exponent is also zero. See the THRESH_CMP_MANTISSA doc.
141 	 */
142 	cmp = (event >> EVENT_THR_CMP_SHIFT) & EVENT_THR_CMP_MASK;
143 	exp = cmp >> 7;
144 
145 	if (exp && (cmp & 0x60) == 0)
146 		return false;
147 
148 	return true;
149 }
150 
151 static unsigned int dc_ic_rld_quad_l1_sel(u64 event)
152 {
153 	unsigned int cache;
154 
155 	cache = (event >> EVENT_CACHE_SEL_SHIFT) & MMCR1_DC_IC_QUAL_MASK;
156 	return cache;
157 }
158 
159 static inline u64 isa207_find_source(u64 idx, u32 sub_idx)
160 {
161 	u64 ret = PERF_MEM_NA;
162 
163 	switch(idx) {
164 	case 0:
165 		/* Nothing to do */
166 		break;
167 	case 1:
168 		ret = PH(LVL, L1);
169 		break;
170 	case 2:
171 		ret = PH(LVL, L2);
172 		break;
173 	case 3:
174 		ret = PH(LVL, L3);
175 		break;
176 	case 4:
177 		if (sub_idx <= 1)
178 			ret = PH(LVL, LOC_RAM);
179 		else if (sub_idx > 1 && sub_idx <= 2)
180 			ret = PH(LVL, REM_RAM1);
181 		else
182 			ret = PH(LVL, REM_RAM2);
183 		ret |= P(SNOOP, HIT);
184 		break;
185 	case 5:
186 		ret = PH(LVL, REM_CCE1);
187 		if ((sub_idx == 0) || (sub_idx == 2) || (sub_idx == 4))
188 			ret |= P(SNOOP, HIT);
189 		else if ((sub_idx == 1) || (sub_idx == 3) || (sub_idx == 5))
190 			ret |= P(SNOOP, HITM);
191 		break;
192 	case 6:
193 		ret = PH(LVL, REM_CCE2);
194 		if ((sub_idx == 0) || (sub_idx == 2))
195 			ret |= P(SNOOP, HIT);
196 		else if ((sub_idx == 1) || (sub_idx == 3))
197 			ret |= P(SNOOP, HITM);
198 		break;
199 	case 7:
200 		ret = PM(LVL, L1);
201 		break;
202 	}
203 
204 	return ret;
205 }
206 
207 void isa207_get_mem_data_src(union perf_mem_data_src *dsrc, u32 flags,
208 							struct pt_regs *regs)
209 {
210 	u64 idx;
211 	u32 sub_idx;
212 	u64 sier;
213 	u64 val;
214 
215 	/* Skip if no SIER support */
216 	if (!(flags & PPMU_HAS_SIER)) {
217 		dsrc->val = 0;
218 		return;
219 	}
220 
221 	sier = mfspr(SPRN_SIER);
222 	val = (sier & ISA207_SIER_TYPE_MASK) >> ISA207_SIER_TYPE_SHIFT;
223 	if (val == 1 || val == 2) {
224 		idx = (sier & ISA207_SIER_LDST_MASK) >> ISA207_SIER_LDST_SHIFT;
225 		sub_idx = (sier & ISA207_SIER_DATA_SRC_MASK) >> ISA207_SIER_DATA_SRC_SHIFT;
226 
227 		dsrc->val = isa207_find_source(idx, sub_idx);
228 		dsrc->val |= (val == 1) ? P(OP, LOAD) : P(OP, STORE);
229 	}
230 }
231 
232 void isa207_get_mem_weight(u64 *weight)
233 {
234 	u64 mmcra = mfspr(SPRN_MMCRA);
235 	u64 exp = MMCRA_THR_CTR_EXP(mmcra);
236 	u64 mantissa = MMCRA_THR_CTR_MANT(mmcra);
237 	u64 sier = mfspr(SPRN_SIER);
238 	u64 val = (sier & ISA207_SIER_TYPE_MASK) >> ISA207_SIER_TYPE_SHIFT;
239 
240 	if (val == 0 || val == 7)
241 		*weight = 0;
242 	else
243 		*weight = mantissa << (2 * exp);
244 }
245 
246 int isa207_get_constraint(u64 event, unsigned long *maskp, unsigned long *valp)
247 {
248 	unsigned int unit, pmc, cache, ebb;
249 	unsigned long mask, value;
250 
251 	mask = value = 0;
252 
253 	if (!is_event_valid(event))
254 		return -1;
255 
256 	pmc   = (event >> EVENT_PMC_SHIFT)        & EVENT_PMC_MASK;
257 	unit  = (event >> EVENT_UNIT_SHIFT)       & EVENT_UNIT_MASK;
258 	cache = (event >> EVENT_CACHE_SEL_SHIFT)  & EVENT_CACHE_SEL_MASK;
259 	ebb   = (event >> EVENT_EBB_SHIFT)        & EVENT_EBB_MASK;
260 
261 	if (pmc) {
262 		u64 base_event;
263 
264 		if (pmc > 6)
265 			return -1;
266 
267 		/* Ignore Linux defined bits when checking event below */
268 		base_event = event & ~EVENT_LINUX_MASK;
269 
270 		if (pmc >= 5 && base_event != 0x500fa &&
271 				base_event != 0x600f4)
272 			return -1;
273 
274 		mask  |= CNST_PMC_MASK(pmc);
275 		value |= CNST_PMC_VAL(pmc);
276 	}
277 
278 	if (pmc <= 4) {
279 		/*
280 		 * Add to number of counters in use. Note this includes events with
281 		 * a PMC of 0 - they still need a PMC, it's just assigned later.
282 		 * Don't count events on PMC 5 & 6, there is only one valid event
283 		 * on each of those counters, and they are handled above.
284 		 */
285 		mask  |= CNST_NC_MASK;
286 		value |= CNST_NC_VAL;
287 	}
288 
289 	if (unit >= 6 && unit <= 9) {
290 		if (cpu_has_feature(CPU_FTR_ARCH_300)) {
291 			mask  |= CNST_CACHE_GROUP_MASK;
292 			value |= CNST_CACHE_GROUP_VAL(event & 0xff);
293 
294 			mask |= CNST_CACHE_PMC4_MASK;
295 			if (pmc == 4)
296 				value |= CNST_CACHE_PMC4_VAL;
297 		} else if (cache & 0x7) {
298 			/*
299 			 * L2/L3 events contain a cache selector field, which is
300 			 * supposed to be programmed into MMCRC. However MMCRC is only
301 			 * HV writable, and there is no API for guest kernels to modify
302 			 * it. The solution is for the hypervisor to initialise the
303 			 * field to zeroes, and for us to only ever allow events that
304 			 * have a cache selector of zero. The bank selector (bit 3) is
305 			 * irrelevant, as long as the rest of the value is 0.
306 			 */
307 			return -1;
308 		}
309 
310 	} else if (cpu_has_feature(CPU_FTR_ARCH_300) || (event & EVENT_IS_L1)) {
311 		mask  |= CNST_L1_QUAL_MASK;
312 		value |= CNST_L1_QUAL_VAL(cache);
313 	}
314 
315 	if (is_event_marked(event)) {
316 		mask  |= CNST_SAMPLE_MASK;
317 		value |= CNST_SAMPLE_VAL(event >> EVENT_SAMPLE_SHIFT);
318 	}
319 
320 	if (cpu_has_feature(CPU_FTR_ARCH_300))  {
321 		if (event_is_threshold(event) && is_thresh_cmp_valid(event)) {
322 			mask  |= CNST_THRESH_MASK;
323 			value |= CNST_THRESH_VAL(event >> EVENT_THRESH_SHIFT);
324 		}
325 	} else {
326 		/*
327 		 * Special case for PM_MRK_FAB_RSP_MATCH and PM_MRK_FAB_RSP_MATCH_CYC,
328 		 * the threshold control bits are used for the match value.
329 		 */
330 		if (event_is_fab_match(event)) {
331 			mask  |= CNST_FAB_MATCH_MASK;
332 			value |= CNST_FAB_MATCH_VAL(event >> EVENT_THR_CTL_SHIFT);
333 		} else {
334 			if (!is_thresh_cmp_valid(event))
335 				return -1;
336 
337 			mask  |= CNST_THRESH_MASK;
338 			value |= CNST_THRESH_VAL(event >> EVENT_THRESH_SHIFT);
339 		}
340 	}
341 
342 	if (!pmc && ebb)
343 		/* EBB events must specify the PMC */
344 		return -1;
345 
346 	if (event & EVENT_WANTS_BHRB) {
347 		if (!ebb)
348 			/* Only EBB events can request BHRB */
349 			return -1;
350 
351 		mask  |= CNST_IFM_MASK;
352 		value |= CNST_IFM_VAL(event >> EVENT_IFM_SHIFT);
353 	}
354 
355 	/*
356 	 * All events must agree on EBB, either all request it or none.
357 	 * EBB events are pinned & exclusive, so this should never actually
358 	 * hit, but we leave it as a fallback in case.
359 	 */
360 	mask  |= CNST_EBB_VAL(ebb);
361 	value |= CNST_EBB_MASK;
362 
363 	*maskp = mask;
364 	*valp = value;
365 
366 	return 0;
367 }
368 
369 int isa207_compute_mmcr(u64 event[], int n_ev,
370 			       unsigned int hwc[], unsigned long mmcr[],
371 			       struct perf_event *pevents[])
372 {
373 	unsigned long mmcra, mmcr1, mmcr2, unit, combine, psel, cache, val;
374 	unsigned int pmc, pmc_inuse;
375 	int i;
376 
377 	pmc_inuse = 0;
378 
379 	/* First pass to count resource use */
380 	for (i = 0; i < n_ev; ++i) {
381 		pmc = (event[i] >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK;
382 		if (pmc)
383 			pmc_inuse |= 1 << pmc;
384 	}
385 
386 	mmcra = mmcr1 = mmcr2 = 0;
387 
388 	/* Second pass: assign PMCs, set all MMCR1 fields */
389 	for (i = 0; i < n_ev; ++i) {
390 		pmc     = (event[i] >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK;
391 		unit    = (event[i] >> EVENT_UNIT_SHIFT) & EVENT_UNIT_MASK;
392 		combine = combine_from_event(event[i]);
393 		psel    =  event[i] & EVENT_PSEL_MASK;
394 
395 		if (!pmc) {
396 			for (pmc = 1; pmc <= 4; ++pmc) {
397 				if (!(pmc_inuse & (1 << pmc)))
398 					break;
399 			}
400 
401 			pmc_inuse |= 1 << pmc;
402 		}
403 
404 		if (pmc <= 4) {
405 			mmcr1 |= unit << MMCR1_UNIT_SHIFT(pmc);
406 			mmcr1 |= combine << combine_shift(pmc);
407 			mmcr1 |= psel << MMCR1_PMCSEL_SHIFT(pmc);
408 		}
409 
410 		/* In continuous sampling mode, update SDAR on TLB miss */
411 		mmcra_sdar_mode(event[i], &mmcra);
412 
413 		if (cpu_has_feature(CPU_FTR_ARCH_300)) {
414 			cache = dc_ic_rld_quad_l1_sel(event[i]);
415 			mmcr1 |= (cache) << MMCR1_DC_IC_QUAL_SHIFT;
416 		} else {
417 			if (event[i] & EVENT_IS_L1) {
418 				cache = dc_ic_rld_quad_l1_sel(event[i]);
419 				mmcr1 |= (cache) << MMCR1_DC_IC_QUAL_SHIFT;
420 			}
421 		}
422 
423 		if (is_event_marked(event[i])) {
424 			mmcra |= MMCRA_SAMPLE_ENABLE;
425 
426 			val = (event[i] >> EVENT_SAMPLE_SHIFT) & EVENT_SAMPLE_MASK;
427 			if (val) {
428 				mmcra |= (val &  3) << MMCRA_SAMP_MODE_SHIFT;
429 				mmcra |= (val >> 2) << MMCRA_SAMP_ELIG_SHIFT;
430 			}
431 		}
432 
433 		/*
434 		 * PM_MRK_FAB_RSP_MATCH and PM_MRK_FAB_RSP_MATCH_CYC,
435 		 * the threshold bits are used for the match value.
436 		 */
437 		if (!cpu_has_feature(CPU_FTR_ARCH_300) && event_is_fab_match(event[i])) {
438 			mmcr1 |= ((event[i] >> EVENT_THR_CTL_SHIFT) &
439 				  EVENT_THR_CTL_MASK) << MMCR1_FAB_SHIFT;
440 		} else {
441 			val = (event[i] >> EVENT_THR_CTL_SHIFT) & EVENT_THR_CTL_MASK;
442 			mmcra |= val << MMCRA_THR_CTL_SHIFT;
443 			val = (event[i] >> EVENT_THR_SEL_SHIFT) & EVENT_THR_SEL_MASK;
444 			mmcra |= val << MMCRA_THR_SEL_SHIFT;
445 			val = (event[i] >> EVENT_THR_CMP_SHIFT) & EVENT_THR_CMP_MASK;
446 			mmcra |= thresh_cmp_val(val);
447 		}
448 
449 		if (event[i] & EVENT_WANTS_BHRB) {
450 			val = (event[i] >> EVENT_IFM_SHIFT) & EVENT_IFM_MASK;
451 			mmcra |= val << MMCRA_IFM_SHIFT;
452 		}
453 
454 		if (pevents[i]->attr.exclude_user)
455 			mmcr2 |= MMCR2_FCP(pmc);
456 
457 		if (pevents[i]->attr.exclude_hv)
458 			mmcr2 |= MMCR2_FCH(pmc);
459 
460 		if (pevents[i]->attr.exclude_kernel) {
461 			if (cpu_has_feature(CPU_FTR_HVMODE))
462 				mmcr2 |= MMCR2_FCH(pmc);
463 			else
464 				mmcr2 |= MMCR2_FCS(pmc);
465 		}
466 
467 		hwc[i] = pmc - 1;
468 	}
469 
470 	/* Return MMCRx values */
471 	mmcr[0] = 0;
472 
473 	/* pmc_inuse is 1-based */
474 	if (pmc_inuse & 2)
475 		mmcr[0] = MMCR0_PMC1CE;
476 
477 	if (pmc_inuse & 0x7c)
478 		mmcr[0] |= MMCR0_PMCjCE;
479 
480 	/* If we're not using PMC 5 or 6, freeze them */
481 	if (!(pmc_inuse & 0x60))
482 		mmcr[0] |= MMCR0_FC56;
483 
484 	mmcr[1] = mmcr1;
485 	mmcr[2] = mmcra;
486 	mmcr[3] = mmcr2;
487 
488 	return 0;
489 }
490 
491 void isa207_disable_pmc(unsigned int pmc, unsigned long mmcr[])
492 {
493 	if (pmc <= 3)
494 		mmcr[1] &= ~(0xffUL << MMCR1_PMCSEL_SHIFT(pmc + 1));
495 }
496 
497 static int find_alternative(u64 event, const unsigned int ev_alt[][MAX_ALT], int size)
498 {
499 	int i, j;
500 
501 	for (i = 0; i < size; ++i) {
502 		if (event < ev_alt[i][0])
503 			break;
504 
505 		for (j = 0; j < MAX_ALT && ev_alt[i][j]; ++j)
506 			if (event == ev_alt[i][j])
507 				return i;
508 	}
509 
510 	return -1;
511 }
512 
513 int isa207_get_alternatives(u64 event, u64 alt[], int size, unsigned int flags,
514 					const unsigned int ev_alt[][MAX_ALT])
515 {
516 	int i, j, num_alt = 0;
517 	u64 alt_event;
518 
519 	alt[num_alt++] = event;
520 	i = find_alternative(event, ev_alt, size);
521 	if (i >= 0) {
522 		/* Filter out the original event, it's already in alt[0] */
523 		for (j = 0; j < MAX_ALT; ++j) {
524 			alt_event = ev_alt[i][j];
525 			if (alt_event && alt_event != event)
526 				alt[num_alt++] = alt_event;
527 		}
528 	}
529 
530 	if (flags & PPMU_ONLY_COUNT_RUN) {
531 		/*
532 		 * We're only counting in RUN state, so PM_CYC is equivalent to
533 		 * PM_RUN_CYC and PM_INST_CMPL === PM_RUN_INST_CMPL.
534 		 */
535 		j = num_alt;
536 		for (i = 0; i < num_alt; ++i) {
537 			switch (alt[i]) {
538 			case 0x1e:			/* PMC_CYC */
539 				alt[j++] = 0x600f4;	/* PM_RUN_CYC */
540 				break;
541 			case 0x600f4:
542 				alt[j++] = 0x1e;
543 				break;
544 			case 0x2:			/* PM_INST_CMPL */
545 				alt[j++] = 0x500fa;	/* PM_RUN_INST_CMPL */
546 				break;
547 			case 0x500fa:
548 				alt[j++] = 0x2;
549 				break;
550 			}
551 		}
552 		num_alt = j;
553 	}
554 
555 	return num_alt;
556 }
557