2 * Common Performance counter support functions for PowerISA v2.07 processors.
4 * Copyright 2009 Paul Mackerras, IBM Corporation.
5 * Copyright 2013 Michael Ellerman, IBM Corporation.
6 * Copyright 2016 Madhavan Srinivasan, IBM Corporation.
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.
13 #include "isa207-common.h"
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");
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,
44 struct attribute_group isa207_pmu_format_group = {
46 .attrs = isa207_pmu_format_attr,
49 static inline bool event_is_fab_match(u64 event)
51 /* Only check pmc, unit and pmcxsel, ignore the edge bit (0) */
54 /* PM_MRK_FAB_RSP_MATCH & PM_MRK_FAB_RSP_MATCH_CYC */
55 return (event == 0x30056 || event == 0x4f052);
58 static bool is_event_valid(u64 event)
60 u64 valid_mask = EVENT_VALID_MASK;
62 if (cpu_has_feature(CPU_FTR_ARCH_300))
63 valid_mask = p9_EVENT_VALID_MASK;
65 return !(event & ~valid_mask);
68 static inline bool is_event_marked(u64 event)
70 if (event & EVENT_IS_MARKED)
76 static void mmcra_sdar_mode(u64 event, unsigned long *mmcra)
79 * MMCRA[SDAR_MODE] specifices how the SDAR should be updated in
80 * continous sampling mode.
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).
87 * Marked event: MMCRA[SDAR_MODE] will be set to 0b00 ('No Updates'),
88 * or if group already have any marked events.
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.
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;
100 *mmcra |= MMCRA_SDAR_MODE_DCACHE;
102 *mmcra |= MMCRA_SDAR_MODE_TLB;
105 static u64 thresh_cmp_val(u64 value)
107 if (cpu_has_feature(CPU_FTR_ARCH_300))
108 return value << p9_MMCRA_THR_CMP_SHIFT;
110 return value << MMCRA_THR_CMP_SHIFT;
113 static unsigned long combine_from_event(u64 event)
115 if (cpu_has_feature(CPU_FTR_ARCH_300))
116 return p9_EVENT_COMBINE(event);
118 return EVENT_COMBINE(event);
121 static unsigned long combine_shift(unsigned long pmc)
123 if (cpu_has_feature(CPU_FTR_ARCH_300))
124 return p9_MMCR1_COMBINE_SHIFT(pmc);
126 return MMCR1_COMBINE_SHIFT(pmc);
129 static inline bool event_is_threshold(u64 event)
131 return (event >> EVENT_THR_SEL_SHIFT) & EVENT_THR_SEL_MASK;
134 static bool is_thresh_cmp_valid(u64 event)
136 unsigned int cmp, exp;
139 * Check the mantissa upper two bits are not zero, unless the
140 * exponent is also zero. See the THRESH_CMP_MANTISSA doc.
142 cmp = (event >> EVENT_THR_CMP_SHIFT) & EVENT_THR_CMP_MASK;
145 if (exp && (cmp & 0x60) == 0)
151 static unsigned int dc_ic_rld_quad_l1_sel(u64 event)
155 cache = (event >> EVENT_CACHE_SEL_SHIFT) & MMCR1_DC_IC_QUAL_MASK;
159 static inline u64 isa207_find_source(u64 idx, u32 sub_idx)
161 u64 ret = PERF_MEM_NA;
178 ret = PH(LVL, LOC_RAM);
179 else if (sub_idx > 1 && sub_idx <= 2)
180 ret = PH(LVL, REM_RAM1);
182 ret = PH(LVL, REM_RAM2);
183 ret |= P(SNOOP, HIT);
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);
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);
207 void isa207_get_mem_data_src(union perf_mem_data_src *dsrc, u32 flags,
208 struct pt_regs *regs)
215 /* Skip if no SIER support */
216 if (!(flags & PPMU_HAS_SIER)) {
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;
227 dsrc->val = isa207_find_source(idx, sub_idx);
228 dsrc->val |= (val == 1) ? P(OP, LOAD) : P(OP, STORE);
232 void isa207_get_mem_weight(u64 *weight)
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;
240 if (val == 0 || val == 7)
243 *weight = mantissa << (2 * exp);
246 int isa207_get_constraint(u64 event, unsigned long *maskp, unsigned long *valp)
248 unsigned int unit, pmc, cache, ebb;
249 unsigned long mask, value;
253 if (!is_event_valid(event))
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;
267 /* Ignore Linux defined bits when checking event below */
268 base_event = event & ~EVENT_LINUX_MASK;
270 if (pmc >= 5 && base_event != 0x500fa &&
271 base_event != 0x600f4)
274 mask |= CNST_PMC_MASK(pmc);
275 value |= CNST_PMC_VAL(pmc);
278 * PMC5 and PMC6 are used to count cycles and instructions and
279 * they do not support most of the constraint bits. Add a check
280 * to exclude PMC5/6 from most of the constraints except for
289 * Add to number of counters in use. Note this includes events with
290 * a PMC of 0 - they still need a PMC, it's just assigned later.
291 * Don't count events on PMC 5 & 6, there is only one valid event
292 * on each of those counters, and they are handled above.
294 mask |= CNST_NC_MASK;
295 value |= CNST_NC_VAL;
298 if (unit >= 6 && unit <= 9) {
300 * L2/L3 events contain a cache selector field, which is
301 * supposed to be programmed into MMCRC. However MMCRC is only
302 * HV writable, and there is no API for guest kernels to modify
303 * it. The solution is for the hypervisor to initialise the
304 * field to zeroes, and for us to only ever allow events that
305 * have a cache selector of zero. The bank selector (bit 3) is
306 * irrelevant, as long as the rest of the value is 0.
308 if (!cpu_has_feature(CPU_FTR_ARCH_300) && (cache & 0x7))
311 } else if (cpu_has_feature(CPU_FTR_ARCH_300) || (event & EVENT_IS_L1)) {
312 mask |= CNST_L1_QUAL_MASK;
313 value |= CNST_L1_QUAL_VAL(cache);
316 if (is_event_marked(event)) {
317 mask |= CNST_SAMPLE_MASK;
318 value |= CNST_SAMPLE_VAL(event >> EVENT_SAMPLE_SHIFT);
321 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
322 if (event_is_threshold(event) && is_thresh_cmp_valid(event)) {
323 mask |= CNST_THRESH_MASK;
324 value |= CNST_THRESH_VAL(event >> EVENT_THRESH_SHIFT);
325 } else if (event_is_threshold(event))
329 * Special case for PM_MRK_FAB_RSP_MATCH and PM_MRK_FAB_RSP_MATCH_CYC,
330 * the threshold control bits are used for the match value.
332 if (event_is_fab_match(event)) {
333 mask |= CNST_FAB_MATCH_MASK;
334 value |= CNST_FAB_MATCH_VAL(event >> EVENT_THR_CTL_SHIFT);
336 if (!is_thresh_cmp_valid(event))
339 mask |= CNST_THRESH_MASK;
340 value |= CNST_THRESH_VAL(event >> EVENT_THRESH_SHIFT);
346 /* EBB events must specify the PMC */
349 if (event & EVENT_WANTS_BHRB) {
351 /* Only EBB events can request BHRB */
354 mask |= CNST_IFM_MASK;
355 value |= CNST_IFM_VAL(event >> EVENT_IFM_SHIFT);
359 * All events must agree on EBB, either all request it or none.
360 * EBB events are pinned & exclusive, so this should never actually
361 * hit, but we leave it as a fallback in case.
363 mask |= CNST_EBB_MASK;
364 value |= CNST_EBB_VAL(ebb);
372 int isa207_compute_mmcr(u64 event[], int n_ev,
373 unsigned int hwc[], unsigned long mmcr[],
374 struct perf_event *pevents[])
376 unsigned long mmcra, mmcr1, mmcr2, unit, combine, psel, cache, val;
377 unsigned int pmc, pmc_inuse;
382 /* First pass to count resource use */
383 for (i = 0; i < n_ev; ++i) {
384 pmc = (event[i] >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK;
386 pmc_inuse |= 1 << pmc;
389 mmcra = mmcr1 = mmcr2 = 0;
391 /* Second pass: assign PMCs, set all MMCR1 fields */
392 for (i = 0; i < n_ev; ++i) {
393 pmc = (event[i] >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK;
394 unit = (event[i] >> EVENT_UNIT_SHIFT) & EVENT_UNIT_MASK;
395 combine = combine_from_event(event[i]);
396 psel = event[i] & EVENT_PSEL_MASK;
399 for (pmc = 1; pmc <= 4; ++pmc) {
400 if (!(pmc_inuse & (1 << pmc)))
404 pmc_inuse |= 1 << pmc;
408 mmcr1 |= unit << MMCR1_UNIT_SHIFT(pmc);
409 mmcr1 |= combine << combine_shift(pmc);
410 mmcr1 |= psel << MMCR1_PMCSEL_SHIFT(pmc);
413 /* In continuous sampling mode, update SDAR on TLB miss */
414 mmcra_sdar_mode(event[i], &mmcra);
416 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
417 cache = dc_ic_rld_quad_l1_sel(event[i]);
418 mmcr1 |= (cache) << MMCR1_DC_IC_QUAL_SHIFT;
420 if (event[i] & EVENT_IS_L1) {
421 cache = dc_ic_rld_quad_l1_sel(event[i]);
422 mmcr1 |= (cache) << MMCR1_DC_IC_QUAL_SHIFT;
426 if (is_event_marked(event[i])) {
427 mmcra |= MMCRA_SAMPLE_ENABLE;
429 val = (event[i] >> EVENT_SAMPLE_SHIFT) & EVENT_SAMPLE_MASK;
431 mmcra |= (val & 3) << MMCRA_SAMP_MODE_SHIFT;
432 mmcra |= (val >> 2) << MMCRA_SAMP_ELIG_SHIFT;
437 * PM_MRK_FAB_RSP_MATCH and PM_MRK_FAB_RSP_MATCH_CYC,
438 * the threshold bits are used for the match value.
440 if (!cpu_has_feature(CPU_FTR_ARCH_300) && event_is_fab_match(event[i])) {
441 mmcr1 |= ((event[i] >> EVENT_THR_CTL_SHIFT) &
442 EVENT_THR_CTL_MASK) << MMCR1_FAB_SHIFT;
444 val = (event[i] >> EVENT_THR_CTL_SHIFT) & EVENT_THR_CTL_MASK;
445 mmcra |= val << MMCRA_THR_CTL_SHIFT;
446 val = (event[i] >> EVENT_THR_SEL_SHIFT) & EVENT_THR_SEL_MASK;
447 mmcra |= val << MMCRA_THR_SEL_SHIFT;
448 val = (event[i] >> EVENT_THR_CMP_SHIFT) & EVENT_THR_CMP_MASK;
449 mmcra |= thresh_cmp_val(val);
452 if (event[i] & EVENT_WANTS_BHRB) {
453 val = (event[i] >> EVENT_IFM_SHIFT) & EVENT_IFM_MASK;
454 mmcra |= val << MMCRA_IFM_SHIFT;
457 if (pevents[i]->attr.exclude_user)
458 mmcr2 |= MMCR2_FCP(pmc);
460 if (pevents[i]->attr.exclude_hv)
461 mmcr2 |= MMCR2_FCH(pmc);
463 if (pevents[i]->attr.exclude_kernel) {
464 if (cpu_has_feature(CPU_FTR_HVMODE))
465 mmcr2 |= MMCR2_FCH(pmc);
467 mmcr2 |= MMCR2_FCS(pmc);
473 /* Return MMCRx values */
476 /* pmc_inuse is 1-based */
478 mmcr[0] = MMCR0_PMC1CE;
480 if (pmc_inuse & 0x7c)
481 mmcr[0] |= MMCR0_PMCjCE;
483 /* If we're not using PMC 5 or 6, freeze them */
484 if (!(pmc_inuse & 0x60))
485 mmcr[0] |= MMCR0_FC56;
494 void isa207_disable_pmc(unsigned int pmc, unsigned long mmcr[])
497 mmcr[1] &= ~(0xffUL << MMCR1_PMCSEL_SHIFT(pmc + 1));
500 static int find_alternative(u64 event, const unsigned int ev_alt[][MAX_ALT], int size)
504 for (i = 0; i < size; ++i) {
505 if (event < ev_alt[i][0])
508 for (j = 0; j < MAX_ALT && ev_alt[i][j]; ++j)
509 if (event == ev_alt[i][j])
516 int isa207_get_alternatives(u64 event, u64 alt[], int size, unsigned int flags,
517 const unsigned int ev_alt[][MAX_ALT])
519 int i, j, num_alt = 0;
522 alt[num_alt++] = event;
523 i = find_alternative(event, ev_alt, size);
525 /* Filter out the original event, it's already in alt[0] */
526 for (j = 0; j < MAX_ALT; ++j) {
527 alt_event = ev_alt[i][j];
528 if (alt_event && alt_event != event)
529 alt[num_alt++] = alt_event;
533 if (flags & PPMU_ONLY_COUNT_RUN) {
535 * We're only counting in RUN state, so PM_CYC is equivalent to
536 * PM_RUN_CYC and PM_INST_CMPL === PM_RUN_INST_CMPL.
539 for (i = 0; i < num_alt; ++i) {
541 case 0x1e: /* PMC_CYC */
542 alt[j++] = 0x600f4; /* PM_RUN_CYC */
547 case 0x2: /* PM_INST_CMPL */
548 alt[j++] = 0x500fa; /* PM_RUN_INST_CMPL */