• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
event_is_fab_match(u64 event)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 
is_event_valid(u64 event)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) && !cpu_has_feature(CPU_FTR_POWER9_DD1))
63 		valid_mask = p9_EVENT_VALID_MASK;
64 
65 	return !(event & ~valid_mask);
66 }
67 
is_event_marked(u64 event)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 
mmcra_sdar_mode(u64 event,unsigned long * mmcra)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 	 * Non-Marked events (for DD1):
90 	 *	MMCRA[SDAR_MODE] will be set to 0b01
91 	 * For rest
92 	 *	MMCRA[SDAR_MODE] will be set from event code.
93 	 *      If sdar_mode from event is zero, default to 0b01. Hardware
94 	 *      requires that we set a non-zero value.
95 	 */
96 	if (cpu_has_feature(CPU_FTR_ARCH_300)) {
97 		if (is_event_marked(event) || (*mmcra & MMCRA_SAMPLE_ENABLE))
98 			*mmcra &= MMCRA_SDAR_MODE_NO_UPDATES;
99 		else if (!cpu_has_feature(CPU_FTR_POWER9_DD1) && p9_SDAR_MODE(event))
100 			*mmcra |=  p9_SDAR_MODE(event) << MMCRA_SDAR_MODE_SHIFT;
101 		else
102 			*mmcra |= MMCRA_SDAR_MODE_DCACHE;
103 	} else
104 		*mmcra |= MMCRA_SDAR_MODE_TLB;
105 }
106 
thresh_cmp_val(u64 value)107 static u64 thresh_cmp_val(u64 value)
108 {
109 	if (cpu_has_feature(CPU_FTR_ARCH_300) && !cpu_has_feature(CPU_FTR_POWER9_DD1))
110 		return value << p9_MMCRA_THR_CMP_SHIFT;
111 
112 	return value << MMCRA_THR_CMP_SHIFT;
113 }
114 
combine_from_event(u64 event)115 static unsigned long combine_from_event(u64 event)
116 {
117 	if (cpu_has_feature(CPU_FTR_ARCH_300) && !cpu_has_feature(CPU_FTR_POWER9_DD1))
118 		return p9_EVENT_COMBINE(event);
119 
120 	return EVENT_COMBINE(event);
121 }
122 
combine_shift(unsigned long pmc)123 static unsigned long combine_shift(unsigned long pmc)
124 {
125 	if (cpu_has_feature(CPU_FTR_ARCH_300) && !cpu_has_feature(CPU_FTR_POWER9_DD1))
126 		return p9_MMCR1_COMBINE_SHIFT(pmc);
127 
128 	return MMCR1_COMBINE_SHIFT(pmc);
129 }
130 
event_is_threshold(u64 event)131 static inline bool event_is_threshold(u64 event)
132 {
133 	return (event >> EVENT_THR_SEL_SHIFT) & EVENT_THR_SEL_MASK;
134 }
135 
is_thresh_cmp_valid(u64 event)136 static bool is_thresh_cmp_valid(u64 event)
137 {
138 	unsigned int cmp, exp;
139 
140 	/*
141 	 * Check the mantissa upper two bits are not zero, unless the
142 	 * exponent is also zero. See the THRESH_CMP_MANTISSA doc.
143 	 */
144 	cmp = (event >> EVENT_THR_CMP_SHIFT) & EVENT_THR_CMP_MASK;
145 	exp = cmp >> 7;
146 
147 	if (exp && (cmp & 0x60) == 0)
148 		return false;
149 
150 	return true;
151 }
152 
dc_ic_rld_quad_l1_sel(u64 event)153 static unsigned int dc_ic_rld_quad_l1_sel(u64 event)
154 {
155 	unsigned int cache;
156 
157 	cache = (event >> EVENT_CACHE_SEL_SHIFT) & MMCR1_DC_IC_QUAL_MASK;
158 	return cache;
159 }
160 
isa207_find_source(u64 idx,u32 sub_idx)161 static inline u64 isa207_find_source(u64 idx, u32 sub_idx)
162 {
163 	u64 ret = PERF_MEM_NA;
164 
165 	switch(idx) {
166 	case 0:
167 		/* Nothing to do */
168 		break;
169 	case 1:
170 		ret = PH(LVL, L1);
171 		break;
172 	case 2:
173 		ret = PH(LVL, L2);
174 		break;
175 	case 3:
176 		ret = PH(LVL, L3);
177 		break;
178 	case 4:
179 		if (sub_idx <= 1)
180 			ret = PH(LVL, LOC_RAM);
181 		else if (sub_idx > 1 && sub_idx <= 2)
182 			ret = PH(LVL, REM_RAM1);
183 		else
184 			ret = PH(LVL, REM_RAM2);
185 		ret |= P(SNOOP, HIT);
186 		break;
187 	case 5:
188 		ret = PH(LVL, REM_CCE1);
189 		if ((sub_idx == 0) || (sub_idx == 2) || (sub_idx == 4))
190 			ret |= P(SNOOP, HIT);
191 		else if ((sub_idx == 1) || (sub_idx == 3) || (sub_idx == 5))
192 			ret |= P(SNOOP, HITM);
193 		break;
194 	case 6:
195 		ret = PH(LVL, REM_CCE2);
196 		if ((sub_idx == 0) || (sub_idx == 2))
197 			ret |= P(SNOOP, HIT);
198 		else if ((sub_idx == 1) || (sub_idx == 3))
199 			ret |= P(SNOOP, HITM);
200 		break;
201 	case 7:
202 		ret = PM(LVL, L1);
203 		break;
204 	}
205 
206 	return ret;
207 }
208 
isa207_get_mem_data_src(union perf_mem_data_src * dsrc,u32 flags,struct pt_regs * regs)209 void isa207_get_mem_data_src(union perf_mem_data_src *dsrc, u32 flags,
210 							struct pt_regs *regs)
211 {
212 	u64 idx;
213 	u32 sub_idx;
214 	u64 sier;
215 	u64 val;
216 
217 	/* Skip if no SIER support */
218 	if (!(flags & PPMU_HAS_SIER)) {
219 		dsrc->val = 0;
220 		return;
221 	}
222 
223 	sier = mfspr(SPRN_SIER);
224 	val = (sier & ISA207_SIER_TYPE_MASK) >> ISA207_SIER_TYPE_SHIFT;
225 	if (val == 1 || val == 2) {
226 		idx = (sier & ISA207_SIER_LDST_MASK) >> ISA207_SIER_LDST_SHIFT;
227 		sub_idx = (sier & ISA207_SIER_DATA_SRC_MASK) >> ISA207_SIER_DATA_SRC_SHIFT;
228 
229 		dsrc->val = isa207_find_source(idx, sub_idx);
230 		dsrc->val |= (val == 1) ? P(OP, LOAD) : P(OP, STORE);
231 	}
232 }
233 
isa207_get_mem_weight(u64 * weight)234 void isa207_get_mem_weight(u64 *weight)
235 {
236 	u64 mmcra = mfspr(SPRN_MMCRA);
237 	u64 exp = MMCRA_THR_CTR_EXP(mmcra);
238 	u64 mantissa = MMCRA_THR_CTR_MANT(mmcra);
239 	u64 sier = mfspr(SPRN_SIER);
240 	u64 val = (sier & ISA207_SIER_TYPE_MASK) >> ISA207_SIER_TYPE_SHIFT;
241 
242 	if (val == 0 || val == 7)
243 		*weight = 0;
244 	else
245 		*weight = mantissa << (2 * exp);
246 }
247 
isa207_get_constraint(u64 event,unsigned long * maskp,unsigned long * valp)248 int isa207_get_constraint(u64 event, unsigned long *maskp, unsigned long *valp)
249 {
250 	unsigned int unit, pmc, cache, ebb;
251 	unsigned long mask, value;
252 
253 	mask = value = 0;
254 
255 	if (!is_event_valid(event))
256 		return -1;
257 
258 	pmc   = (event >> EVENT_PMC_SHIFT)        & EVENT_PMC_MASK;
259 	unit  = (event >> EVENT_UNIT_SHIFT)       & EVENT_UNIT_MASK;
260 	cache = (event >> EVENT_CACHE_SEL_SHIFT)  & EVENT_CACHE_SEL_MASK;
261 	ebb   = (event >> EVENT_EBB_SHIFT)        & EVENT_EBB_MASK;
262 
263 	if (pmc) {
264 		u64 base_event;
265 
266 		if (pmc > 6)
267 			return -1;
268 
269 		/* Ignore Linux defined bits when checking event below */
270 		base_event = event & ~EVENT_LINUX_MASK;
271 
272 		if (pmc >= 5 && base_event != 0x500fa &&
273 				base_event != 0x600f4)
274 			return -1;
275 
276 		mask  |= CNST_PMC_MASK(pmc);
277 		value |= CNST_PMC_VAL(pmc);
278 	}
279 
280 	if (pmc <= 4) {
281 		/*
282 		 * Add to number of counters in use. Note this includes events with
283 		 * a PMC of 0 - they still need a PMC, it's just assigned later.
284 		 * Don't count events on PMC 5 & 6, there is only one valid event
285 		 * on each of those counters, and they are handled above.
286 		 */
287 		mask  |= CNST_NC_MASK;
288 		value |= CNST_NC_VAL;
289 	}
290 
291 	if (unit >= 6 && unit <= 9) {
292 		/*
293 		 * L2/L3 events contain a cache selector field, which is
294 		 * supposed to be programmed into MMCRC. However MMCRC is only
295 		 * HV writable, and there is no API for guest kernels to modify
296 		 * it. The solution is for the hypervisor to initialise the
297 		 * field to zeroes, and for us to only ever allow events that
298 		 * have a cache selector of zero. The bank selector (bit 3) is
299 		 * irrelevant, as long as the rest of the value is 0.
300 		 */
301 		if (!cpu_has_feature(CPU_FTR_ARCH_300) && (cache & 0x7))
302 			return -1;
303 
304 	} else if (cpu_has_feature(CPU_FTR_ARCH_300) || (event & EVENT_IS_L1)) {
305 		mask  |= CNST_L1_QUAL_MASK;
306 		value |= CNST_L1_QUAL_VAL(cache);
307 	}
308 
309 	if (is_event_marked(event)) {
310 		mask  |= CNST_SAMPLE_MASK;
311 		value |= CNST_SAMPLE_VAL(event >> EVENT_SAMPLE_SHIFT);
312 	}
313 
314 	if (cpu_has_feature(CPU_FTR_ARCH_300))  {
315 		if (event_is_threshold(event) && is_thresh_cmp_valid(event)) {
316 			mask  |= CNST_THRESH_MASK;
317 			value |= CNST_THRESH_VAL(event >> EVENT_THRESH_SHIFT);
318 		}
319 	} else {
320 		/*
321 		 * Special case for PM_MRK_FAB_RSP_MATCH and PM_MRK_FAB_RSP_MATCH_CYC,
322 		 * the threshold control bits are used for the match value.
323 		 */
324 		if (event_is_fab_match(event)) {
325 			mask  |= CNST_FAB_MATCH_MASK;
326 			value |= CNST_FAB_MATCH_VAL(event >> EVENT_THR_CTL_SHIFT);
327 		} else {
328 			if (!is_thresh_cmp_valid(event))
329 				return -1;
330 
331 			mask  |= CNST_THRESH_MASK;
332 			value |= CNST_THRESH_VAL(event >> EVENT_THRESH_SHIFT);
333 		}
334 	}
335 
336 	if (!pmc && ebb)
337 		/* EBB events must specify the PMC */
338 		return -1;
339 
340 	if (event & EVENT_WANTS_BHRB) {
341 		if (!ebb)
342 			/* Only EBB events can request BHRB */
343 			return -1;
344 
345 		mask  |= CNST_IFM_MASK;
346 		value |= CNST_IFM_VAL(event >> EVENT_IFM_SHIFT);
347 	}
348 
349 	/*
350 	 * All events must agree on EBB, either all request it or none.
351 	 * EBB events are pinned & exclusive, so this should never actually
352 	 * hit, but we leave it as a fallback in case.
353 	 */
354 	mask  |= CNST_EBB_VAL(ebb);
355 	value |= CNST_EBB_MASK;
356 
357 	*maskp = mask;
358 	*valp = value;
359 
360 	return 0;
361 }
362 
isa207_compute_mmcr(u64 event[],int n_ev,unsigned int hwc[],unsigned long mmcr[],struct perf_event * pevents[])363 int isa207_compute_mmcr(u64 event[], int n_ev,
364 			       unsigned int hwc[], unsigned long mmcr[],
365 			       struct perf_event *pevents[])
366 {
367 	unsigned long mmcra, mmcr1, mmcr2, unit, combine, psel, cache, val;
368 	unsigned int pmc, pmc_inuse;
369 	int i;
370 
371 	pmc_inuse = 0;
372 
373 	/* First pass to count resource use */
374 	for (i = 0; i < n_ev; ++i) {
375 		pmc = (event[i] >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK;
376 		if (pmc)
377 			pmc_inuse |= 1 << pmc;
378 	}
379 
380 	mmcra = mmcr1 = mmcr2 = 0;
381 
382 	/* Second pass: assign PMCs, set all MMCR1 fields */
383 	for (i = 0; i < n_ev; ++i) {
384 		pmc     = (event[i] >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK;
385 		unit    = (event[i] >> EVENT_UNIT_SHIFT) & EVENT_UNIT_MASK;
386 		combine = combine_from_event(event[i]);
387 		psel    =  event[i] & EVENT_PSEL_MASK;
388 
389 		if (!pmc) {
390 			for (pmc = 1; pmc <= 4; ++pmc) {
391 				if (!(pmc_inuse & (1 << pmc)))
392 					break;
393 			}
394 
395 			pmc_inuse |= 1 << pmc;
396 		}
397 
398 		if (pmc <= 4) {
399 			mmcr1 |= unit << MMCR1_UNIT_SHIFT(pmc);
400 			mmcr1 |= combine << combine_shift(pmc);
401 			mmcr1 |= psel << MMCR1_PMCSEL_SHIFT(pmc);
402 		}
403 
404 		/* In continuous sampling mode, update SDAR on TLB miss */
405 		mmcra_sdar_mode(event[i], &mmcra);
406 
407 		if (cpu_has_feature(CPU_FTR_ARCH_300)) {
408 			cache = dc_ic_rld_quad_l1_sel(event[i]);
409 			mmcr1 |= (cache) << MMCR1_DC_IC_QUAL_SHIFT;
410 		} else {
411 			if (event[i] & EVENT_IS_L1) {
412 				cache = dc_ic_rld_quad_l1_sel(event[i]);
413 				mmcr1 |= (cache) << MMCR1_DC_IC_QUAL_SHIFT;
414 			}
415 		}
416 
417 		if (is_event_marked(event[i])) {
418 			mmcra |= MMCRA_SAMPLE_ENABLE;
419 
420 			val = (event[i] >> EVENT_SAMPLE_SHIFT) & EVENT_SAMPLE_MASK;
421 			if (val) {
422 				mmcra |= (val &  3) << MMCRA_SAMP_MODE_SHIFT;
423 				mmcra |= (val >> 2) << MMCRA_SAMP_ELIG_SHIFT;
424 			}
425 		}
426 
427 		/*
428 		 * PM_MRK_FAB_RSP_MATCH and PM_MRK_FAB_RSP_MATCH_CYC,
429 		 * the threshold bits are used for the match value.
430 		 */
431 		if (!cpu_has_feature(CPU_FTR_ARCH_300) && event_is_fab_match(event[i])) {
432 			mmcr1 |= ((event[i] >> EVENT_THR_CTL_SHIFT) &
433 				  EVENT_THR_CTL_MASK) << MMCR1_FAB_SHIFT;
434 		} else {
435 			val = (event[i] >> EVENT_THR_CTL_SHIFT) & EVENT_THR_CTL_MASK;
436 			mmcra |= val << MMCRA_THR_CTL_SHIFT;
437 			val = (event[i] >> EVENT_THR_SEL_SHIFT) & EVENT_THR_SEL_MASK;
438 			mmcra |= val << MMCRA_THR_SEL_SHIFT;
439 			val = (event[i] >> EVENT_THR_CMP_SHIFT) & EVENT_THR_CMP_MASK;
440 			mmcra |= thresh_cmp_val(val);
441 		}
442 
443 		if (event[i] & EVENT_WANTS_BHRB) {
444 			val = (event[i] >> EVENT_IFM_SHIFT) & EVENT_IFM_MASK;
445 			mmcra |= val << MMCRA_IFM_SHIFT;
446 		}
447 
448 		if (pevents[i]->attr.exclude_user)
449 			mmcr2 |= MMCR2_FCP(pmc);
450 
451 		if (pevents[i]->attr.exclude_hv)
452 			mmcr2 |= MMCR2_FCH(pmc);
453 
454 		if (pevents[i]->attr.exclude_kernel) {
455 			if (cpu_has_feature(CPU_FTR_HVMODE))
456 				mmcr2 |= MMCR2_FCH(pmc);
457 			else
458 				mmcr2 |= MMCR2_FCS(pmc);
459 		}
460 
461 		hwc[i] = pmc - 1;
462 	}
463 
464 	/* Return MMCRx values */
465 	mmcr[0] = 0;
466 
467 	/* pmc_inuse is 1-based */
468 	if (pmc_inuse & 2)
469 		mmcr[0] = MMCR0_PMC1CE;
470 
471 	if (pmc_inuse & 0x7c)
472 		mmcr[0] |= MMCR0_PMCjCE;
473 
474 	/* If we're not using PMC 5 or 6, freeze them */
475 	if (!(pmc_inuse & 0x60))
476 		mmcr[0] |= MMCR0_FC56;
477 
478 	mmcr[1] = mmcr1;
479 	mmcr[2] = mmcra;
480 	mmcr[3] = mmcr2;
481 
482 	return 0;
483 }
484 
isa207_disable_pmc(unsigned int pmc,unsigned long mmcr[])485 void isa207_disable_pmc(unsigned int pmc, unsigned long mmcr[])
486 {
487 	if (pmc <= 3)
488 		mmcr[1] &= ~(0xffUL << MMCR1_PMCSEL_SHIFT(pmc + 1));
489 }
490 
find_alternative(u64 event,const unsigned int ev_alt[][MAX_ALT],int size)491 static int find_alternative(u64 event, const unsigned int ev_alt[][MAX_ALT], int size)
492 {
493 	int i, j;
494 
495 	for (i = 0; i < size; ++i) {
496 		if (event < ev_alt[i][0])
497 			break;
498 
499 		for (j = 0; j < MAX_ALT && ev_alt[i][j]; ++j)
500 			if (event == ev_alt[i][j])
501 				return i;
502 	}
503 
504 	return -1;
505 }
506 
isa207_get_alternatives(u64 event,u64 alt[],int size,unsigned int flags,const unsigned int ev_alt[][MAX_ALT])507 int isa207_get_alternatives(u64 event, u64 alt[], int size, unsigned int flags,
508 					const unsigned int ev_alt[][MAX_ALT])
509 {
510 	int i, j, num_alt = 0;
511 	u64 alt_event;
512 
513 	alt[num_alt++] = event;
514 	i = find_alternative(event, ev_alt, size);
515 	if (i >= 0) {
516 		/* Filter out the original event, it's already in alt[0] */
517 		for (j = 0; j < MAX_ALT; ++j) {
518 			alt_event = ev_alt[i][j];
519 			if (alt_event && alt_event != event)
520 				alt[num_alt++] = alt_event;
521 		}
522 	}
523 
524 	if (flags & PPMU_ONLY_COUNT_RUN) {
525 		/*
526 		 * We're only counting in RUN state, so PM_CYC is equivalent to
527 		 * PM_RUN_CYC and PM_INST_CMPL === PM_RUN_INST_CMPL.
528 		 */
529 		j = num_alt;
530 		for (i = 0; i < num_alt; ++i) {
531 			switch (alt[i]) {
532 			case 0x1e:			/* PMC_CYC */
533 				alt[j++] = 0x600f4;	/* PM_RUN_CYC */
534 				break;
535 			case 0x600f4:
536 				alt[j++] = 0x1e;
537 				break;
538 			case 0x2:			/* PM_INST_CMPL */
539 				alt[j++] = 0x500fa;	/* PM_RUN_INST_CMPL */
540 				break;
541 			case 0x500fa:
542 				alt[j++] = 0x2;
543 				break;
544 			}
545 		}
546 		num_alt = j;
547 	}
548 
549 	return num_alt;
550 }
551