1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * Performance counter support for POWER7 processors.
4 *
5 * Copyright 2009 Paul Mackerras, IBM Corporation.
6 */
7 #include <linux/kernel.h>
8 #include <linux/perf_event.h>
9 #include <linux/string.h>
10 #include <asm/reg.h>
11 #include <asm/cputable.h>
12
13 #include "internal.h"
14
15 /*
16 * Bits in event code for POWER7
17 */
18 #define PM_PMC_SH 16 /* PMC number (1-based) for direct events */
19 #define PM_PMC_MSK 0xf
20 #define PM_PMC_MSKS (PM_PMC_MSK << PM_PMC_SH)
21 #define PM_UNIT_SH 12 /* TTMMUX number and setting - unit select */
22 #define PM_UNIT_MSK 0xf
23 #define PM_COMBINE_SH 11 /* Combined event bit */
24 #define PM_COMBINE_MSK 1
25 #define PM_COMBINE_MSKS 0x800
26 #define PM_L2SEL_SH 8 /* L2 event select */
27 #define PM_L2SEL_MSK 7
28 #define PM_PMCSEL_MSK 0xff
29
30 /*
31 * Bits in MMCR1 for POWER7
32 */
33 #define MMCR1_TTM0SEL_SH 60
34 #define MMCR1_TTM1SEL_SH 56
35 #define MMCR1_TTM2SEL_SH 52
36 #define MMCR1_TTM3SEL_SH 48
37 #define MMCR1_TTMSEL_MSK 0xf
38 #define MMCR1_L2SEL_SH 45
39 #define MMCR1_L2SEL_MSK 7
40 #define MMCR1_PMC1_COMBINE_SH 35
41 #define MMCR1_PMC2_COMBINE_SH 34
42 #define MMCR1_PMC3_COMBINE_SH 33
43 #define MMCR1_PMC4_COMBINE_SH 32
44 #define MMCR1_PMC1SEL_SH 24
45 #define MMCR1_PMC2SEL_SH 16
46 #define MMCR1_PMC3SEL_SH 8
47 #define MMCR1_PMC4SEL_SH 0
48 #define MMCR1_PMCSEL_SH(n) (MMCR1_PMC1SEL_SH - (n) * 8)
49 #define MMCR1_PMCSEL_MSK 0xff
50
51 /*
52 * Power7 event codes.
53 */
54 #define EVENT(_name, _code) \
55 _name = _code,
56
57 enum {
58 #include "power7-events-list.h"
59 };
60 #undef EVENT
61
62 /*
63 * Layout of constraint bits:
64 * 6666555555555544444444443333333333222222222211111111110000000000
65 * 3210987654321098765432109876543210987654321098765432109876543210
66 * < >< ><><><><><><>
67 * L2 NC P6P5P4P3P2P1
68 *
69 * L2 - 16-18 - Required L2SEL value (select field)
70 *
71 * NC - number of counters
72 * 15: NC error 0x8000
73 * 12-14: number of events needing PMC1-4 0x7000
74 *
75 * P6
76 * 11: P6 error 0x800
77 * 10-11: Count of events needing PMC6
78 *
79 * P1..P5
80 * 0-9: Count of events needing PMC1..PMC5
81 */
82
power7_get_constraint(u64 event,unsigned long * maskp,unsigned long * valp,u64 event_config1 __maybe_unused)83 static int power7_get_constraint(u64 event, unsigned long *maskp,
84 unsigned long *valp, u64 event_config1 __maybe_unused)
85 {
86 int pmc, sh, unit;
87 unsigned long mask = 0, value = 0;
88
89 pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
90 if (pmc) {
91 if (pmc > 6)
92 return -1;
93 sh = (pmc - 1) * 2;
94 mask |= 2 << sh;
95 value |= 1 << sh;
96 if (pmc >= 5 && !(event == 0x500fa || event == 0x600f4))
97 return -1;
98 }
99 if (pmc < 5) {
100 /* need a counter from PMC1-4 set */
101 mask |= 0x8000;
102 value |= 0x1000;
103 }
104
105 unit = (event >> PM_UNIT_SH) & PM_UNIT_MSK;
106 if (unit == 6) {
107 /* L2SEL must be identical across events */
108 int l2sel = (event >> PM_L2SEL_SH) & PM_L2SEL_MSK;
109 mask |= 0x7 << 16;
110 value |= l2sel << 16;
111 }
112
113 *maskp = mask;
114 *valp = value;
115 return 0;
116 }
117
118 #define MAX_ALT 2 /* at most 2 alternatives for any event */
119
120 static const unsigned int event_alternatives[][MAX_ALT] = {
121 { 0x200f2, 0x300f2 }, /* PM_INST_DISP */
122 { 0x200f4, 0x600f4 }, /* PM_RUN_CYC */
123 { 0x400fa, 0x500fa }, /* PM_RUN_INST_CMPL */
124 };
125
126 /*
127 * Scan the alternatives table for a match and return the
128 * index into the alternatives table if found, else -1.
129 */
find_alternative(u64 event)130 static int find_alternative(u64 event)
131 {
132 int i, j;
133
134 for (i = 0; i < ARRAY_SIZE(event_alternatives); ++i) {
135 if (event < event_alternatives[i][0])
136 break;
137 for (j = 0; j < MAX_ALT && event_alternatives[i][j]; ++j)
138 if (event == event_alternatives[i][j])
139 return i;
140 }
141 return -1;
142 }
143
find_alternative_decode(u64 event)144 static s64 find_alternative_decode(u64 event)
145 {
146 int pmc, psel;
147
148 /* this only handles the 4x decode events */
149 pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
150 psel = event & PM_PMCSEL_MSK;
151 if ((pmc == 2 || pmc == 4) && (psel & ~7) == 0x40)
152 return event - (1 << PM_PMC_SH) + 8;
153 if ((pmc == 1 || pmc == 3) && (psel & ~7) == 0x48)
154 return event + (1 << PM_PMC_SH) - 8;
155 return -1;
156 }
157
power7_get_alternatives(u64 event,unsigned int flags,u64 alt[])158 static int power7_get_alternatives(u64 event, unsigned int flags, u64 alt[])
159 {
160 int i, j, nalt = 1;
161 s64 ae;
162
163 alt[0] = event;
164 nalt = 1;
165 i = find_alternative(event);
166 if (i >= 0) {
167 for (j = 0; j < MAX_ALT; ++j) {
168 ae = event_alternatives[i][j];
169 if (ae && ae != event)
170 alt[nalt++] = ae;
171 }
172 } else {
173 ae = find_alternative_decode(event);
174 if (ae > 0)
175 alt[nalt++] = ae;
176 }
177
178 if (flags & PPMU_ONLY_COUNT_RUN) {
179 /*
180 * We're only counting in RUN state,
181 * so PM_CYC is equivalent to PM_RUN_CYC
182 * and PM_INST_CMPL === PM_RUN_INST_CMPL.
183 * This doesn't include alternatives that don't provide
184 * any extra flexibility in assigning PMCs.
185 */
186 j = nalt;
187 for (i = 0; i < nalt; ++i) {
188 switch (alt[i]) {
189 case 0x1e: /* PM_CYC */
190 alt[j++] = 0x600f4; /* PM_RUN_CYC */
191 break;
192 case 0x600f4: /* PM_RUN_CYC */
193 alt[j++] = 0x1e;
194 break;
195 case 0x2: /* PM_PPC_CMPL */
196 alt[j++] = 0x500fa; /* PM_RUN_INST_CMPL */
197 break;
198 case 0x500fa: /* PM_RUN_INST_CMPL */
199 alt[j++] = 0x2; /* PM_PPC_CMPL */
200 break;
201 }
202 }
203 nalt = j;
204 }
205
206 return nalt;
207 }
208
209 /*
210 * Returns 1 if event counts things relating to marked instructions
211 * and thus needs the MMCRA_SAMPLE_ENABLE bit set, or 0 if not.
212 */
power7_marked_instr_event(u64 event)213 static int power7_marked_instr_event(u64 event)
214 {
215 int pmc, psel;
216 int unit;
217
218 pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
219 unit = (event >> PM_UNIT_SH) & PM_UNIT_MSK;
220 psel = event & PM_PMCSEL_MSK & ~1; /* trim off edge/level bit */
221 if (pmc >= 5)
222 return 0;
223
224 switch (psel >> 4) {
225 case 2:
226 return pmc == 2 || pmc == 4;
227 case 3:
228 if (psel == 0x3c)
229 return pmc == 1;
230 if (psel == 0x3e)
231 return pmc != 2;
232 return 1;
233 case 4:
234 case 5:
235 return unit == 0xd;
236 case 6:
237 if (psel == 0x64)
238 return pmc >= 3;
239 break;
240 case 8:
241 return unit == 0xd;
242 }
243 return 0;
244 }
245
power7_compute_mmcr(u64 event[],int n_ev,unsigned int hwc[],struct mmcr_regs * mmcr,struct perf_event * pevents[],u32 flags __maybe_unused)246 static int power7_compute_mmcr(u64 event[], int n_ev,
247 unsigned int hwc[], struct mmcr_regs *mmcr,
248 struct perf_event *pevents[],
249 u32 flags __maybe_unused)
250 {
251 unsigned long mmcr1 = 0;
252 unsigned long mmcra = MMCRA_SDAR_DCACHE_MISS | MMCRA_SDAR_ERAT_MISS;
253 unsigned int pmc, unit, combine, l2sel, psel;
254 unsigned int pmc_inuse = 0;
255 int i;
256
257 /* First pass to count resource use */
258 for (i = 0; i < n_ev; ++i) {
259 pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK;
260 if (pmc) {
261 if (pmc > 6)
262 return -1;
263 if (pmc_inuse & (1 << (pmc - 1)))
264 return -1;
265 pmc_inuse |= 1 << (pmc - 1);
266 }
267 }
268
269 /* Second pass: assign PMCs, set all MMCR1 fields */
270 for (i = 0; i < n_ev; ++i) {
271 pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK;
272 unit = (event[i] >> PM_UNIT_SH) & PM_UNIT_MSK;
273 combine = (event[i] >> PM_COMBINE_SH) & PM_COMBINE_MSK;
274 l2sel = (event[i] >> PM_L2SEL_SH) & PM_L2SEL_MSK;
275 psel = event[i] & PM_PMCSEL_MSK;
276 if (!pmc) {
277 /* Bus event or any-PMC direct event */
278 for (pmc = 0; pmc < 4; ++pmc) {
279 if (!(pmc_inuse & (1 << pmc)))
280 break;
281 }
282 if (pmc >= 4)
283 return -1;
284 pmc_inuse |= 1 << pmc;
285 } else {
286 /* Direct or decoded event */
287 --pmc;
288 }
289 if (pmc <= 3) {
290 mmcr1 |= (unsigned long) unit
291 << (MMCR1_TTM0SEL_SH - 4 * pmc);
292 mmcr1 |= (unsigned long) combine
293 << (MMCR1_PMC1_COMBINE_SH - pmc);
294 mmcr1 |= psel << MMCR1_PMCSEL_SH(pmc);
295 if (unit == 6) /* L2 events */
296 mmcr1 |= (unsigned long) l2sel
297 << MMCR1_L2SEL_SH;
298 }
299 if (power7_marked_instr_event(event[i]))
300 mmcra |= MMCRA_SAMPLE_ENABLE;
301 hwc[i] = pmc;
302 }
303
304 /* Return MMCRx values */
305 mmcr->mmcr0 = 0;
306 if (pmc_inuse & 1)
307 mmcr->mmcr0 = MMCR0_PMC1CE;
308 if (pmc_inuse & 0x3e)
309 mmcr->mmcr0 |= MMCR0_PMCjCE;
310 mmcr->mmcr1 = mmcr1;
311 mmcr->mmcra = mmcra;
312 return 0;
313 }
314
power7_disable_pmc(unsigned int pmc,struct mmcr_regs * mmcr)315 static void power7_disable_pmc(unsigned int pmc, struct mmcr_regs *mmcr)
316 {
317 if (pmc <= 3)
318 mmcr->mmcr1 &= ~(0xffUL << MMCR1_PMCSEL_SH(pmc));
319 }
320
321 static int power7_generic_events[] = {
322 [PERF_COUNT_HW_CPU_CYCLES] = PM_CYC,
323 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = PM_GCT_NOSLOT_CYC,
324 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = PM_CMPLU_STALL,
325 [PERF_COUNT_HW_INSTRUCTIONS] = PM_INST_CMPL,
326 [PERF_COUNT_HW_CACHE_REFERENCES] = PM_LD_REF_L1,
327 [PERF_COUNT_HW_CACHE_MISSES] = PM_LD_MISS_L1,
328 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = PM_BRU_FIN,
329 [PERF_COUNT_HW_BRANCH_MISSES] = PM_BR_MPRED,
330 };
331
332 #define C(x) PERF_COUNT_HW_CACHE_##x
333
334 /*
335 * Table of generalized cache-related events.
336 * 0 means not supported, -1 means nonsensical, other values
337 * are event codes.
338 */
339 static u64 power7_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
340 [C(L1D)] = { /* RESULT_ACCESS RESULT_MISS */
341 [C(OP_READ)] = { 0xc880, 0x400f0 },
342 [C(OP_WRITE)] = { 0, 0x300f0 },
343 [C(OP_PREFETCH)] = { 0xd8b8, 0 },
344 },
345 [C(L1I)] = { /* RESULT_ACCESS RESULT_MISS */
346 [C(OP_READ)] = { 0, 0x200fc },
347 [C(OP_WRITE)] = { -1, -1 },
348 [C(OP_PREFETCH)] = { 0x408a, 0 },
349 },
350 [C(LL)] = { /* RESULT_ACCESS RESULT_MISS */
351 [C(OP_READ)] = { 0x16080, 0x26080 },
352 [C(OP_WRITE)] = { 0x16082, 0x26082 },
353 [C(OP_PREFETCH)] = { 0, 0 },
354 },
355 [C(DTLB)] = { /* RESULT_ACCESS RESULT_MISS */
356 [C(OP_READ)] = { 0, 0x300fc },
357 [C(OP_WRITE)] = { -1, -1 },
358 [C(OP_PREFETCH)] = { -1, -1 },
359 },
360 [C(ITLB)] = { /* RESULT_ACCESS RESULT_MISS */
361 [C(OP_READ)] = { 0, 0x400fc },
362 [C(OP_WRITE)] = { -1, -1 },
363 [C(OP_PREFETCH)] = { -1, -1 },
364 },
365 [C(BPU)] = { /* RESULT_ACCESS RESULT_MISS */
366 [C(OP_READ)] = { 0x10068, 0x400f6 },
367 [C(OP_WRITE)] = { -1, -1 },
368 [C(OP_PREFETCH)] = { -1, -1 },
369 },
370 [C(NODE)] = { /* RESULT_ACCESS RESULT_MISS */
371 [C(OP_READ)] = { -1, -1 },
372 [C(OP_WRITE)] = { -1, -1 },
373 [C(OP_PREFETCH)] = { -1, -1 },
374 },
375 };
376
377
378 GENERIC_EVENT_ATTR(cpu-cycles, PM_CYC);
379 GENERIC_EVENT_ATTR(stalled-cycles-frontend, PM_GCT_NOSLOT_CYC);
380 GENERIC_EVENT_ATTR(stalled-cycles-backend, PM_CMPLU_STALL);
381 GENERIC_EVENT_ATTR(instructions, PM_INST_CMPL);
382 GENERIC_EVENT_ATTR(cache-references, PM_LD_REF_L1);
383 GENERIC_EVENT_ATTR(cache-misses, PM_LD_MISS_L1);
384 GENERIC_EVENT_ATTR(branch-instructions, PM_BRU_FIN);
385 GENERIC_EVENT_ATTR(branch-misses, PM_BR_MPRED);
386
387 #define EVENT(_name, _code) POWER_EVENT_ATTR(_name, _name);
388 #include "power7-events-list.h"
389 #undef EVENT
390
391 #define EVENT(_name, _code) POWER_EVENT_PTR(_name),
392
393 static struct attribute *power7_events_attr[] = {
394 GENERIC_EVENT_PTR(PM_CYC),
395 GENERIC_EVENT_PTR(PM_GCT_NOSLOT_CYC),
396 GENERIC_EVENT_PTR(PM_CMPLU_STALL),
397 GENERIC_EVENT_PTR(PM_INST_CMPL),
398 GENERIC_EVENT_PTR(PM_LD_REF_L1),
399 GENERIC_EVENT_PTR(PM_LD_MISS_L1),
400 GENERIC_EVENT_PTR(PM_BRU_FIN),
401 GENERIC_EVENT_PTR(PM_BR_MPRED),
402
403 #include "power7-events-list.h"
404 #undef EVENT
405 NULL
406 };
407
408 static struct attribute_group power7_pmu_events_group = {
409 .name = "events",
410 .attrs = power7_events_attr,
411 };
412
413 PMU_FORMAT_ATTR(event, "config:0-19");
414
415 static struct attribute *power7_pmu_format_attr[] = {
416 &format_attr_event.attr,
417 NULL,
418 };
419
420 static struct attribute_group power7_pmu_format_group = {
421 .name = "format",
422 .attrs = power7_pmu_format_attr,
423 };
424
425 static const struct attribute_group *power7_pmu_attr_groups[] = {
426 &power7_pmu_format_group,
427 &power7_pmu_events_group,
428 NULL,
429 };
430
431 static struct power_pmu power7_pmu = {
432 .name = "POWER7",
433 .n_counter = 6,
434 .max_alternatives = MAX_ALT + 1,
435 .add_fields = 0x1555ul,
436 .test_adder = 0x3000ul,
437 .compute_mmcr = power7_compute_mmcr,
438 .get_constraint = power7_get_constraint,
439 .get_alternatives = power7_get_alternatives,
440 .disable_pmc = power7_disable_pmc,
441 .flags = PPMU_ALT_SIPR,
442 .attr_groups = power7_pmu_attr_groups,
443 .n_generic = ARRAY_SIZE(power7_generic_events),
444 .generic_events = power7_generic_events,
445 .cache_events = &power7_cache_events,
446 };
447
init_power7_pmu(void)448 int init_power7_pmu(void)
449 {
450 if (!cur_cpu_spec->oprofile_cpu_type ||
451 strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power7"))
452 return -ENODEV;
453
454 if (pvr_version_is(PVR_POWER7p))
455 power7_pmu.flags |= PPMU_SIAR_VALID;
456
457 return register_power_pmu(&power7_pmu);
458 }
459