1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Performance counter support for POWER9 processors.
5 * Copyright 2009 Paul Mackerras, IBM Corporation.
6 * Copyright 2013 Michael Ellerman, IBM Corporation.
7 * Copyright 2016 Madhavan Srinivasan, IBM Corporation.
10 #define pr_fmt(fmt) "power9-pmu: " fmt
12 #include "isa207-common.h"
15 * Raw event encoding for Power9:
17 * 60 56 52 48 44 40 36 32
18 * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
19 * | | [ ] [ ] [ thresh_cmp ] [ thresh_ctl ]
21 * | | *- IFM (Linux) | thresh start/stop -*
22 * | *- BHRB (Linux) *sm
25 * 28 24 20 16 12 8 4 0
26 * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
27 * [ ] [ sample ] [cache] [ pmc ] [unit ] [] m [ pmcxsel ]
30 * | | *- L1/L2/L3 cache_sel |
32 * | *- sampling mode for marked events *- combine
36 * Below uses IBM bit numbering.
38 * MMCR1[x:y] = unit (PMCxUNIT)
39 * MMCR1[24] = pmc1combine[0]
40 * MMCR1[25] = pmc1combine[1]
41 * MMCR1[26] = pmc2combine[0]
42 * MMCR1[27] = pmc2combine[1]
43 * MMCR1[28] = pmc3combine[0]
44 * MMCR1[29] = pmc3combine[1]
45 * MMCR1[30] = pmc4combine[0]
46 * MMCR1[31] = pmc4combine[1]
48 * if pmc == 3 and unit == 0 and pmcxsel[0:6] == 0b0101011
49 * MMCR1[20:27] = thresh_ctl
50 * else if pmc == 4 and unit == 0xf and pmcxsel[0:6] == 0b0101001
51 * MMCR1[20:27] = thresh_ctl
53 * MMCRA[48:55] = thresh_ctl (THRESH START/END)
56 * MMCRA[45:47] = thresh_sel
59 * MMCRA[9:11] = thresh_cmp[0:2]
60 * MMCRA[12:18] = thresh_cmp[3:9]
62 * MMCR1[16] = cache_sel[2]
63 * MMCR1[17] = cache_sel[3]
66 * MMCRA[63] = 1 (SAMPLE_ENABLE)
67 * MMCRA[57:59] = sample[0:2] (RAND_SAMP_ELIG)
68 * MMCRA[61:62] = sample[3:4] (RAND_SAMP_MODE)
73 * MMCRA[SDAR_MODE] = sm
77 * Some power9 event codes.
79 #define EVENT(_name, _code) _name = _code,
82 #include "power9-events-list.h"
87 /* MMCRA IFM bits - POWER9 */
88 #define POWER9_MMCRA_IFM1 0x0000000040000000UL
89 #define POWER9_MMCRA_IFM2 0x0000000080000000UL
90 #define POWER9_MMCRA_IFM3 0x00000000C0000000UL
91 #define POWER9_MMCRA_BHRB_MASK 0x00000000C0000000UL
93 extern u64 PERF_REG_EXTENDED_MASK;
95 /* Nasty Power9 specific hack */
96 #define PVR_POWER9_CUMULUS 0x00002000
98 /* PowerISA v2.07 format attribute structure*/
99 extern struct attribute_group isa207_pmu_format_group;
101 int p9_dd21_bl_ev[] = {
108 PM_MRK_L2_TM_ST_ABORT_SISTER,
110 PM_RUN_CYC_SMT2_MODE,
111 PM_TM_TX_PASS_RUN_INST,
112 PM_DISP_HELD_SYNC_HOLD,
115 int p9_dd22_bl_ev[] = {
122 PM_MRK_DERAT_MISS_2M,
124 PM_MRK_DTLB_MISS_16G,
126 PM_MRK_DERAT_MISS_1G,
127 PM_MRK_DTLB_MISS_64K,
128 PM_DISP_HELD_SYNC_HOLD,
131 PM_MRK_DTLB_MISS_16M,
134 /* Table of alternatives, sorted by column 0 */
135 static const unsigned int power9_event_alternatives[][MAX_ALT] = {
136 { PM_INST_DISP, PM_INST_DISP_ALT },
137 { PM_RUN_CYC_ALT, PM_RUN_CYC },
138 { PM_RUN_INST_CMPL_ALT, PM_RUN_INST_CMPL },
139 { PM_LD_MISS_L1, PM_LD_MISS_L1_ALT },
140 { PM_BR_2PATH, PM_BR_2PATH_ALT },
143 static int power9_get_alternatives(u64 event, unsigned int flags, u64 alt[])
147 num_alt = isa207_get_alternatives(event, alt,
148 ARRAY_SIZE(power9_event_alternatives), flags,
149 power9_event_alternatives);
154 static int power9_check_attr_config(struct perf_event *ev)
157 u64 event = ev->attr.config;
159 val = (event >> EVENT_SAMPLE_SHIFT) & EVENT_SAMPLE_MASK;
160 if (val == 0xC || isa3XX_check_attr_config(ev))
166 GENERIC_EVENT_ATTR(cpu-cycles, PM_CYC);
167 GENERIC_EVENT_ATTR(stalled-cycles-frontend, PM_ICT_NOSLOT_CYC);
168 GENERIC_EVENT_ATTR(stalled-cycles-backend, PM_CMPLU_STALL);
169 GENERIC_EVENT_ATTR(instructions, PM_INST_CMPL);
170 GENERIC_EVENT_ATTR(branch-instructions, PM_BR_CMPL);
171 GENERIC_EVENT_ATTR(branch-misses, PM_BR_MPRED_CMPL);
172 GENERIC_EVENT_ATTR(cache-references, PM_LD_REF_L1);
173 GENERIC_EVENT_ATTR(cache-misses, PM_LD_MISS_L1_FIN);
174 GENERIC_EVENT_ATTR(mem-loads, MEM_LOADS);
175 GENERIC_EVENT_ATTR(mem-stores, MEM_STORES);
177 CACHE_EVENT_ATTR(L1-dcache-load-misses, PM_LD_MISS_L1_FIN);
178 CACHE_EVENT_ATTR(L1-dcache-loads, PM_LD_REF_L1);
179 CACHE_EVENT_ATTR(L1-dcache-prefetches, PM_L1_PREF);
180 CACHE_EVENT_ATTR(L1-dcache-store-misses, PM_ST_MISS_L1);
181 CACHE_EVENT_ATTR(L1-icache-load-misses, PM_L1_ICACHE_MISS);
182 CACHE_EVENT_ATTR(L1-icache-loads, PM_INST_FROM_L1);
183 CACHE_EVENT_ATTR(L1-icache-prefetches, PM_IC_PREF_WRITE);
184 CACHE_EVENT_ATTR(LLC-load-misses, PM_DATA_FROM_L3MISS);
185 CACHE_EVENT_ATTR(LLC-loads, PM_DATA_FROM_L3);
186 CACHE_EVENT_ATTR(LLC-prefetches, PM_L3_PREF_ALL);
187 CACHE_EVENT_ATTR(branch-load-misses, PM_BR_MPRED_CMPL);
188 CACHE_EVENT_ATTR(branch-loads, PM_BR_CMPL);
189 CACHE_EVENT_ATTR(dTLB-load-misses, PM_DTLB_MISS);
190 CACHE_EVENT_ATTR(iTLB-load-misses, PM_ITLB_MISS);
192 static struct attribute *power9_events_attr[] = {
193 GENERIC_EVENT_PTR(PM_CYC),
194 GENERIC_EVENT_PTR(PM_ICT_NOSLOT_CYC),
195 GENERIC_EVENT_PTR(PM_CMPLU_STALL),
196 GENERIC_EVENT_PTR(PM_INST_CMPL),
197 GENERIC_EVENT_PTR(PM_BR_CMPL),
198 GENERIC_EVENT_PTR(PM_BR_MPRED_CMPL),
199 GENERIC_EVENT_PTR(PM_LD_REF_L1),
200 GENERIC_EVENT_PTR(PM_LD_MISS_L1_FIN),
201 GENERIC_EVENT_PTR(MEM_LOADS),
202 GENERIC_EVENT_PTR(MEM_STORES),
203 CACHE_EVENT_PTR(PM_LD_MISS_L1_FIN),
204 CACHE_EVENT_PTR(PM_LD_REF_L1),
205 CACHE_EVENT_PTR(PM_L1_PREF),
206 CACHE_EVENT_PTR(PM_ST_MISS_L1),
207 CACHE_EVENT_PTR(PM_L1_ICACHE_MISS),
208 CACHE_EVENT_PTR(PM_INST_FROM_L1),
209 CACHE_EVENT_PTR(PM_IC_PREF_WRITE),
210 CACHE_EVENT_PTR(PM_DATA_FROM_L3MISS),
211 CACHE_EVENT_PTR(PM_DATA_FROM_L3),
212 CACHE_EVENT_PTR(PM_L3_PREF_ALL),
213 CACHE_EVENT_PTR(PM_BR_MPRED_CMPL),
214 CACHE_EVENT_PTR(PM_BR_CMPL),
215 CACHE_EVENT_PTR(PM_DTLB_MISS),
216 CACHE_EVENT_PTR(PM_ITLB_MISS),
220 static struct attribute_group power9_pmu_events_group = {
222 .attrs = power9_events_attr,
225 PMU_FORMAT_ATTR(event, "config:0-51");
226 PMU_FORMAT_ATTR(pmcxsel, "config:0-7");
227 PMU_FORMAT_ATTR(mark, "config:8");
228 PMU_FORMAT_ATTR(combine, "config:10-11");
229 PMU_FORMAT_ATTR(unit, "config:12-15");
230 PMU_FORMAT_ATTR(pmc, "config:16-19");
231 PMU_FORMAT_ATTR(cache_sel, "config:20-23");
232 PMU_FORMAT_ATTR(sample_mode, "config:24-28");
233 PMU_FORMAT_ATTR(thresh_sel, "config:29-31");
234 PMU_FORMAT_ATTR(thresh_stop, "config:32-35");
235 PMU_FORMAT_ATTR(thresh_start, "config:36-39");
236 PMU_FORMAT_ATTR(thresh_cmp, "config:40-49");
237 PMU_FORMAT_ATTR(sdar_mode, "config:50-51");
239 static struct attribute *power9_pmu_format_attr[] = {
240 &format_attr_event.attr,
241 &format_attr_pmcxsel.attr,
242 &format_attr_mark.attr,
243 &format_attr_combine.attr,
244 &format_attr_unit.attr,
245 &format_attr_pmc.attr,
246 &format_attr_cache_sel.attr,
247 &format_attr_sample_mode.attr,
248 &format_attr_thresh_sel.attr,
249 &format_attr_thresh_stop.attr,
250 &format_attr_thresh_start.attr,
251 &format_attr_thresh_cmp.attr,
252 &format_attr_sdar_mode.attr,
256 static struct attribute_group power9_pmu_format_group = {
258 .attrs = power9_pmu_format_attr,
261 static const struct attribute_group *power9_pmu_attr_groups[] = {
262 &power9_pmu_format_group,
263 &power9_pmu_events_group,
267 static int power9_generic_events[] = {
268 [PERF_COUNT_HW_CPU_CYCLES] = PM_CYC,
269 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = PM_ICT_NOSLOT_CYC,
270 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = PM_CMPLU_STALL,
271 [PERF_COUNT_HW_INSTRUCTIONS] = PM_INST_CMPL,
272 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = PM_BR_CMPL,
273 [PERF_COUNT_HW_BRANCH_MISSES] = PM_BR_MPRED_CMPL,
274 [PERF_COUNT_HW_CACHE_REFERENCES] = PM_LD_REF_L1,
275 [PERF_COUNT_HW_CACHE_MISSES] = PM_LD_MISS_L1_FIN,
278 static u64 power9_bhrb_filter_map(u64 branch_sample_type)
280 u64 pmu_bhrb_filter = 0;
282 /* BHRB and regular PMU events share the same privilege state
283 * filter configuration. BHRB is always recorded along with a
284 * regular PMU event. As the privilege state filter is handled
285 * in the basic PMC configuration of the accompanying regular
286 * PMU event, we ignore any separate BHRB specific request.
289 /* No branch filter requested */
290 if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY)
291 return pmu_bhrb_filter;
293 /* Invalid branch filter options - HW does not support */
294 if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_RETURN)
297 if (branch_sample_type & PERF_SAMPLE_BRANCH_IND_CALL)
300 if (branch_sample_type & PERF_SAMPLE_BRANCH_CALL)
303 if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_CALL) {
304 pmu_bhrb_filter |= POWER9_MMCRA_IFM1;
305 return pmu_bhrb_filter;
308 /* Every thing else is unsupported */
312 static void power9_config_bhrb(u64 pmu_bhrb_filter)
314 pmu_bhrb_filter &= POWER9_MMCRA_BHRB_MASK;
316 /* Enable BHRB filter in PMU */
317 mtspr(SPRN_MMCRA, (mfspr(SPRN_MMCRA) | pmu_bhrb_filter));
320 #define C(x) PERF_COUNT_HW_CACHE_##x
323 * Table of generalized cache-related events.
324 * 0 means not supported, -1 means nonsensical, other values
327 static u64 power9_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
330 [ C(RESULT_ACCESS) ] = PM_LD_REF_L1,
331 [ C(RESULT_MISS) ] = PM_LD_MISS_L1_FIN,
334 [ C(RESULT_ACCESS) ] = 0,
335 [ C(RESULT_MISS) ] = PM_ST_MISS_L1,
337 [ C(OP_PREFETCH) ] = {
338 [ C(RESULT_ACCESS) ] = PM_L1_PREF,
339 [ C(RESULT_MISS) ] = 0,
344 [ C(RESULT_ACCESS) ] = PM_INST_FROM_L1,
345 [ C(RESULT_MISS) ] = PM_L1_ICACHE_MISS,
348 [ C(RESULT_ACCESS) ] = PM_L1_DEMAND_WRITE,
349 [ C(RESULT_MISS) ] = -1,
351 [ C(OP_PREFETCH) ] = {
352 [ C(RESULT_ACCESS) ] = PM_IC_PREF_WRITE,
353 [ C(RESULT_MISS) ] = 0,
358 [ C(RESULT_ACCESS) ] = PM_DATA_FROM_L3,
359 [ C(RESULT_MISS) ] = PM_DATA_FROM_L3MISS,
362 [ C(RESULT_ACCESS) ] = 0,
363 [ C(RESULT_MISS) ] = 0,
365 [ C(OP_PREFETCH) ] = {
366 [ C(RESULT_ACCESS) ] = PM_L3_PREF_ALL,
367 [ C(RESULT_MISS) ] = 0,
372 [ C(RESULT_ACCESS) ] = 0,
373 [ C(RESULT_MISS) ] = PM_DTLB_MISS,
376 [ C(RESULT_ACCESS) ] = -1,
377 [ C(RESULT_MISS) ] = -1,
379 [ C(OP_PREFETCH) ] = {
380 [ C(RESULT_ACCESS) ] = -1,
381 [ C(RESULT_MISS) ] = -1,
386 [ C(RESULT_ACCESS) ] = 0,
387 [ C(RESULT_MISS) ] = PM_ITLB_MISS,
390 [ C(RESULT_ACCESS) ] = -1,
391 [ C(RESULT_MISS) ] = -1,
393 [ C(OP_PREFETCH) ] = {
394 [ C(RESULT_ACCESS) ] = -1,
395 [ C(RESULT_MISS) ] = -1,
400 [ C(RESULT_ACCESS) ] = PM_BR_CMPL,
401 [ C(RESULT_MISS) ] = PM_BR_MPRED_CMPL,
404 [ C(RESULT_ACCESS) ] = -1,
405 [ C(RESULT_MISS) ] = -1,
407 [ C(OP_PREFETCH) ] = {
408 [ C(RESULT_ACCESS) ] = -1,
409 [ C(RESULT_MISS) ] = -1,
414 [ C(RESULT_ACCESS) ] = -1,
415 [ C(RESULT_MISS) ] = -1,
418 [ C(RESULT_ACCESS) ] = -1,
419 [ C(RESULT_MISS) ] = -1,
421 [ C(OP_PREFETCH) ] = {
422 [ C(RESULT_ACCESS) ] = -1,
423 [ C(RESULT_MISS) ] = -1,
430 static struct power_pmu power9_pmu = {
432 .n_counter = MAX_PMU_COUNTERS,
433 .add_fields = ISA207_ADD_FIELDS,
434 .test_adder = ISA207_TEST_ADDER,
435 .group_constraint_mask = CNST_CACHE_PMC4_MASK,
436 .group_constraint_val = CNST_CACHE_PMC4_VAL,
437 .compute_mmcr = isa207_compute_mmcr,
438 .config_bhrb = power9_config_bhrb,
439 .bhrb_filter_map = power9_bhrb_filter_map,
440 .get_constraint = isa207_get_constraint,
441 .get_alternatives = power9_get_alternatives,
442 .get_mem_data_src = isa207_get_mem_data_src,
443 .get_mem_weight = isa207_get_mem_weight,
444 .disable_pmc = isa207_disable_pmc,
445 .flags = PPMU_HAS_SIER | PPMU_ARCH_207S,
446 .n_generic = ARRAY_SIZE(power9_generic_events),
447 .generic_events = power9_generic_events,
448 .cache_events = &power9_cache_events,
449 .attr_groups = power9_pmu_attr_groups,
451 .capabilities = PERF_PMU_CAP_EXTENDED_REGS,
452 .check_attr_config = power9_check_attr_config,
455 int init_power9_pmu(void)
458 unsigned int pvr = mfspr(SPRN_PVR);
460 /* Comes from cpu_specs[] */
461 if (!cur_cpu_spec->oprofile_cpu_type ||
462 strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power9"))
465 /* Blacklist events */
466 if (!(pvr & PVR_POWER9_CUMULUS)) {
467 if ((PVR_CFG(pvr) == 2) && (PVR_MIN(pvr) == 1)) {
468 power9_pmu.blacklist_ev = p9_dd21_bl_ev;
469 power9_pmu.n_blacklist_ev = ARRAY_SIZE(p9_dd21_bl_ev);
470 } else if ((PVR_CFG(pvr) == 2) && (PVR_MIN(pvr) == 2)) {
471 power9_pmu.blacklist_ev = p9_dd22_bl_ev;
472 power9_pmu.n_blacklist_ev = ARRAY_SIZE(p9_dd22_bl_ev);
476 /* Set the PERF_REG_EXTENDED_MASK here */
477 PERF_REG_EXTENDED_MASK = PERF_REG_PMU_MASK_300;
479 rc = register_power_pmu(&power9_pmu);
483 /* Tell userspace that EBB is supported */
484 cur_cpu_spec->cpu_user_features2 |= PPC_FEATURE2_EBB;