1
2
3
4
5
6
7
8
9
10
11#include <linux/kernel.h>
12#include <linux/perf_event.h>
13#include <linux/string.h>
14#include <asm/reg.h>
15#include <asm/cputable.h>
16
17
18
19
20#define PM_PMC_SH 20
21#define PM_PMC_MSK 0x7
22#define PM_PMC_MSKS (PM_PMC_MSK << PM_PMC_SH)
23#define PM_UNIT_SH 16
24#define PM_UNIT_MSK 0xf
25#define PM_UNIT_MSKS (PM_UNIT_MSK << PM_UNIT_SH)
26#define PM_LLAV 0x8000
27#define PM_LLA 0x4000
28#define PM_BYTE_SH 12
29#define PM_BYTE_MSK 3
30#define PM_SUBUNIT_SH 8
31#define PM_SUBUNIT_MSK 7
32#define PM_SUBUNIT_MSKS (PM_SUBUNIT_MSK << PM_SUBUNIT_SH)
33#define PM_PMCSEL_MSK 0xff
34#define PM_BUSEVENT_MSK 0xf3700
35
36
37
38
39#define MMCR1_TTM0SEL_SH 60
40#define MMCR1_TTMSEL_SH(n) (MMCR1_TTM0SEL_SH - (n) * 4)
41#define MMCR1_TTMSEL_MSK 0xf
42#define MMCR1_TTMSEL(m, n) (((m) >> MMCR1_TTMSEL_SH(n)) & MMCR1_TTMSEL_MSK)
43#define MMCR1_NESTSEL_SH 45
44#define MMCR1_NESTSEL_MSK 0x7
45#define MMCR1_NESTSEL(m) (((m) >> MMCR1_NESTSEL_SH) & MMCR1_NESTSEL_MSK)
46#define MMCR1_PMC1_LLA (1ul << 44)
47#define MMCR1_PMC1_LLA_VALUE (1ul << 39)
48#define MMCR1_PMC1_ADDR_SEL (1ul << 35)
49#define MMCR1_PMC1SEL_SH 24
50#define MMCR1_PMCSEL_SH(n) (MMCR1_PMC1SEL_SH - (n) * 8)
51#define MMCR1_PMCSEL_MSK 0xff
52
53
54
55
56
57
58
59
60
61
62
63
64
65static unsigned char direct_event_is_marked[0x60 >> 1] = {
66 0,
67 0,
68 0,
69 0x07,
70 0x04,
71 0x06,
72 0,
73 0,
74 0x02,
75 0x08,
76 0,
77 0,
78 0x0c,
79 0x0f,
80 0x01,
81 0,
82 0,
83 0,
84 0,
85 0,
86 0x15,
87 0,
88 0,
89 0,
90 0x4f,
91 0x7f,
92 0x4f,
93 0x5f,
94 0x6f,
95 0x4f,
96 0,
97 0x08,
98 0x1f,
99 0x1f,
100 0x1f,
101 0x1f,
102 0x1f,
103 0x1f,
104 0x1f,
105 0x1f,
106 0,
107 0x05,
108 0x1c,
109 0x02,
110 0,
111 0,
112 0,
113 0,
114};
115
116
117
118
119
120static u32 marked_bus_events[16] = {
121 0x01000000,
122 0x00010000,
123 0, 0, 0, 0,
124 0x00000088,
125 0x000000c0,
126 0x04010000,
127 0xff010000u,
128 0,
129 0x00000010,
130 0,
131 0x00000022,
132 0, 0
133};
134
135
136
137
138
139static int power6_marked_instr_event(u64 event)
140{
141 int pmc, psel, ptype;
142 int bit, byte, unit;
143 u32 mask;
144
145 pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
146 psel = (event & PM_PMCSEL_MSK) >> 1;
147 if (pmc >= 5)
148 return 0;
149
150 bit = -1;
151 if (psel < sizeof(direct_event_is_marked)) {
152 ptype = direct_event_is_marked[psel];
153 if (pmc == 0 || !(ptype & (1 << (pmc - 1))))
154 return 0;
155 ptype >>= 4;
156 if (ptype == 0)
157 return 1;
158 if (ptype == 1)
159 bit = 0;
160 else
161 bit = ptype ^ (pmc - 1);
162 } else if ((psel & 0x48) == 0x40)
163 bit = psel & 7;
164
165 if (!(event & PM_BUSEVENT_MSK) || bit == -1)
166 return 0;
167
168 byte = (event >> PM_BYTE_SH) & PM_BYTE_MSK;
169 unit = (event >> PM_UNIT_SH) & PM_UNIT_MSK;
170 mask = marked_bus_events[unit];
171 return (mask >> (byte * 8 + bit)) & 1;
172}
173
174
175
176
177static int p6_compute_mmcr(u64 event[], int n_ev,
178 unsigned int hwc[], unsigned long mmcr[], struct perf_event *pevents[])
179{
180 unsigned long mmcr1 = 0;
181 unsigned long mmcra = MMCRA_SDAR_DCACHE_MISS | MMCRA_SDAR_ERAT_MISS;
182 int i;
183 unsigned int pmc, ev, b, u, s, psel;
184 unsigned int ttmset = 0;
185 unsigned int pmc_inuse = 0;
186
187 if (n_ev > 6)
188 return -1;
189 for (i = 0; i < n_ev; ++i) {
190 pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK;
191 if (pmc) {
192 if (pmc_inuse & (1 << (pmc - 1)))
193 return -1;
194 pmc_inuse |= 1 << (pmc - 1);
195 }
196 }
197 for (i = 0; i < n_ev; ++i) {
198 ev = event[i];
199 pmc = (ev >> PM_PMC_SH) & PM_PMC_MSK;
200 if (pmc) {
201 --pmc;
202 } else {
203
204 for (pmc = 0; pmc < 4; ++pmc)
205 if (!(pmc_inuse & (1 << pmc)))
206 break;
207 if (pmc >= 4)
208 return -1;
209 pmc_inuse |= 1 << pmc;
210 }
211 hwc[i] = pmc;
212 psel = ev & PM_PMCSEL_MSK;
213 if (ev & PM_BUSEVENT_MSK) {
214
215 b = (ev >> PM_BYTE_SH) & PM_BYTE_MSK;
216 u = (ev >> PM_UNIT_SH) & PM_UNIT_MSK;
217
218 if ((ttmset & (1 << b)) && MMCR1_TTMSEL(mmcr1, b) != u)
219 return -1;
220 mmcr1 |= (unsigned long)u << MMCR1_TTMSEL_SH(b);
221 ttmset |= 1 << b;
222 if (u == 5) {
223
224 s = (ev >> PM_SUBUNIT_SH) & PM_SUBUNIT_MSK;
225 if ((ttmset & 0x10) &&
226 MMCR1_NESTSEL(mmcr1) != s)
227 return -1;
228 ttmset |= 0x10;
229 mmcr1 |= (unsigned long)s << MMCR1_NESTSEL_SH;
230 }
231 if (0x30 <= psel && psel <= 0x3d) {
232
233 if (b >= 2)
234 mmcr1 |= MMCR1_PMC1_ADDR_SEL >> pmc;
235 }
236
237 if (pmc >= 2 && (psel & 0x90) == 0x80)
238 psel ^= 0x20;
239 }
240 if (ev & PM_LLA) {
241 mmcr1 |= MMCR1_PMC1_LLA >> pmc;
242 if (ev & PM_LLAV)
243 mmcr1 |= MMCR1_PMC1_LLA_VALUE >> pmc;
244 }
245 if (power6_marked_instr_event(event[i]))
246 mmcra |= MMCRA_SAMPLE_ENABLE;
247 if (pmc < 4)
248 mmcr1 |= (unsigned long)psel << MMCR1_PMCSEL_SH(pmc);
249 }
250 mmcr[0] = 0;
251 if (pmc_inuse & 1)
252 mmcr[0] = MMCR0_PMC1CE;
253 if (pmc_inuse & 0xe)
254 mmcr[0] |= MMCR0_PMCjCE;
255 mmcr[1] = mmcr1;
256 mmcr[2] = mmcra;
257 return 0;
258}
259
260
261
262
263
264
265
266
267
268
269
270static int p6_get_constraint(u64 event, unsigned long *maskp,
271 unsigned long *valp)
272{
273 int pmc, byte, sh, subunit;
274 unsigned long mask = 0, value = 0;
275
276 pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
277 if (pmc) {
278 if (pmc > 4 && !(event == 0x500009 || event == 0x600005))
279 return -1;
280 sh = (pmc - 1) * 2;
281 mask |= 2 << sh;
282 value |= 1 << sh;
283 }
284 if (event & PM_BUSEVENT_MSK) {
285 byte = (event >> PM_BYTE_SH) & PM_BYTE_MSK;
286 sh = byte * 4 + (16 - PM_UNIT_SH);
287 mask |= PM_UNIT_MSKS << sh;
288 value |= (unsigned long)(event & PM_UNIT_MSKS) << sh;
289 if ((event & PM_UNIT_MSKS) == (5 << PM_UNIT_SH)) {
290 subunit = (event >> PM_SUBUNIT_SH) & PM_SUBUNIT_MSK;
291 mask |= (unsigned long)PM_SUBUNIT_MSK << 32;
292 value |= (unsigned long)subunit << 32;
293 }
294 }
295 if (pmc <= 4) {
296 mask |= 0x8000;
297 value |= 0x1000;
298 }
299 *maskp = mask;
300 *valp = value;
301 return 0;
302}
303
304static int p6_limited_pmc_event(u64 event)
305{
306 int pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
307
308 return pmc == 5 || pmc == 6;
309}
310
311#define MAX_ALT 4
312
313static const unsigned int event_alternatives[][MAX_ALT] = {
314 { 0x0130e8, 0x2000f6, 0x3000fc },
315 { 0x080080, 0x10000d, 0x30000c, 0x4000f0 },
316 { 0x080088, 0x200054, 0x3000f0 },
317 { 0x10000a, 0x2000f4, 0x600005 },
318 { 0x10000b, 0x2000f5 },
319 { 0x10000e, 0x400010 },
320 { 0x100010, 0x4000f8 },
321 { 0x10001a, 0x200010 },
322 { 0x100026, 0x3000f8 },
323 { 0x100054, 0x2000f0 },
324 { 0x100056, 0x2000fc },
325 { 0x1000f0, 0x40000a },
326 { 0x1000f8, 0x200008 },
327 { 0x1000fc, 0x400006 },
328 { 0x20000e, 0x400007 },
329 { 0x200012, 0x300012 },
330 { 0x2000f2, 0x3000f2 },
331 { 0x2000f8, 0x300010 },
332 { 0x2000fe, 0x300056 },
333 { 0x2d0030, 0x30001a },
334 { 0x30000a, 0x400018 },
335 { 0x3000f6, 0x40000e },
336 { 0x3000fe, 0x400056 },
337};
338
339
340
341
342
343static int find_alternatives_list(u64 event)
344{
345 int i, j;
346 unsigned int alt;
347
348 for (i = 0; i < ARRAY_SIZE(event_alternatives); ++i) {
349 if (event < event_alternatives[i][0])
350 return -1;
351 for (j = 0; j < MAX_ALT; ++j) {
352 alt = event_alternatives[i][j];
353 if (!alt || event < alt)
354 break;
355 if (event == alt)
356 return i;
357 }
358 }
359 return -1;
360}
361
362static int p6_get_alternatives(u64 event, unsigned int flags, u64 alt[])
363{
364 int i, j, nlim;
365 unsigned int psel, pmc;
366 unsigned int nalt = 1;
367 u64 aevent;
368
369 alt[0] = event;
370 nlim = p6_limited_pmc_event(event);
371
372
373 i = find_alternatives_list(event);
374 if (i >= 0) {
375
376 for (j = 0; j < MAX_ALT; ++j) {
377 aevent = event_alternatives[i][j];
378 if (!aevent)
379 break;
380 if (aevent != event)
381 alt[nalt++] = aevent;
382 nlim += p6_limited_pmc_event(aevent);
383 }
384
385 } else {
386
387
388 psel = event & (PM_PMCSEL_MSK & ~1);
389 pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
390 if (pmc && (psel == 0x32 || psel == 0x34))
391 alt[nalt++] = ((event ^ 0x6) & ~PM_PMC_MSKS) |
392 ((5 - pmc) << PM_PMC_SH);
393
394
395 if (pmc && (psel == 0x38 || psel == 0x3a))
396 alt[nalt++] = ((event ^ 0x2) & ~PM_PMC_MSKS) |
397 ((pmc > 2? pmc - 2: pmc + 2) << PM_PMC_SH);
398 }
399
400 if (flags & PPMU_ONLY_COUNT_RUN) {
401
402
403
404
405
406
407
408
409
410
411 j = nalt;
412 for (i = 0; i < nalt; ++i) {
413 switch (alt[i]) {
414 case 0x1e:
415 alt[j++] = 0x600005;
416 ++nlim;
417 break;
418 case 0x10000a:
419 alt[j++] = 0x1e;
420 break;
421 case 2:
422 alt[j++] = 0x500009;
423 ++nlim;
424 break;
425 case 0x500009:
426 alt[j++] = 2;
427 break;
428 case 0x10000e:
429 alt[j++] = 0x4000f4;
430 break;
431 case 0x4000f4:
432 alt[j++] = 0x10000e;
433 break;
434 }
435 }
436 nalt = j;
437 }
438
439 if (!(flags & PPMU_LIMITED_PMC_OK) && nlim) {
440
441 j = 0;
442 for (i = 0; i < nalt; ++i) {
443 if (!p6_limited_pmc_event(alt[i])) {
444 alt[j] = alt[i];
445 ++j;
446 }
447 }
448 nalt = j;
449 } else if ((flags & PPMU_LIMITED_PMC_REQD) && nlim < nalt) {
450
451 j = 0;
452 for (i = 0; i < nalt; ++i) {
453 if (p6_limited_pmc_event(alt[i])) {
454 alt[j] = alt[i];
455 ++j;
456 }
457 }
458 nalt = j;
459 }
460
461 return nalt;
462}
463
464static void p6_disable_pmc(unsigned int pmc, unsigned long mmcr[])
465{
466
467 if (pmc <= 3)
468 mmcr[1] &= ~(0xffUL << MMCR1_PMCSEL_SH(pmc));
469}
470
471static int power6_generic_events[] = {
472 [PERF_COUNT_HW_CPU_CYCLES] = 0x1e,
473 [PERF_COUNT_HW_INSTRUCTIONS] = 2,
474 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x280030,
475 [PERF_COUNT_HW_CACHE_MISSES] = 0x30000c,
476 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x410a0,
477 [PERF_COUNT_HW_BRANCH_MISSES] = 0x400052,
478};
479
480#define C(x) PERF_COUNT_HW_CACHE_##x
481
482
483
484
485
486
487
488static int power6_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
489 [C(L1D)] = {
490 [C(OP_READ)] = { 0x280030, 0x80080 },
491 [C(OP_WRITE)] = { 0x180032, 0x80088 },
492 [C(OP_PREFETCH)] = { 0x810a4, 0 },
493 },
494 [C(L1I)] = {
495 [C(OP_READ)] = { 0, 0x100056 },
496 [C(OP_WRITE)] = { -1, -1 },
497 [C(OP_PREFETCH)] = { 0x4008c, 0 },
498 },
499 [C(LL)] = {
500 [C(OP_READ)] = { 0x150730, 0x250532 },
501 [C(OP_WRITE)] = { 0x250432, 0x150432 },
502 [C(OP_PREFETCH)] = { 0x810a6, 0 },
503 },
504 [C(DTLB)] = {
505 [C(OP_READ)] = { 0, 0x20000e },
506 [C(OP_WRITE)] = { -1, -1 },
507 [C(OP_PREFETCH)] = { -1, -1 },
508 },
509 [C(ITLB)] = {
510 [C(OP_READ)] = { 0, 0x420ce },
511 [C(OP_WRITE)] = { -1, -1 },
512 [C(OP_PREFETCH)] = { -1, -1 },
513 },
514 [C(BPU)] = {
515 [C(OP_READ)] = { 0x430e6, 0x400052 },
516 [C(OP_WRITE)] = { -1, -1 },
517 [C(OP_PREFETCH)] = { -1, -1 },
518 },
519 [C(NODE)] = {
520 [C(OP_READ)] = { -1, -1 },
521 [C(OP_WRITE)] = { -1, -1 },
522 [C(OP_PREFETCH)] = { -1, -1 },
523 },
524};
525
526static struct power_pmu power6_pmu = {
527 .name = "POWER6",
528 .n_counter = 6,
529 .max_alternatives = MAX_ALT,
530 .add_fields = 0x1555,
531 .test_adder = 0x3000,
532 .compute_mmcr = p6_compute_mmcr,
533 .get_constraint = p6_get_constraint,
534 .get_alternatives = p6_get_alternatives,
535 .disable_pmc = p6_disable_pmc,
536 .limited_pmc_event = p6_limited_pmc_event,
537 .flags = PPMU_LIMITED_PMC5_6 | PPMU_ALT_SIPR,
538 .n_generic = ARRAY_SIZE(power6_generic_events),
539 .generic_events = power6_generic_events,
540 .cache_events = &power6_cache_events,
541};
542
543static int __init init_power6_pmu(void)
544{
545 if (!cur_cpu_spec->oprofile_cpu_type ||
546 strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power6"))
547 return -ENODEV;
548
549 return register_power_pmu(&power6_pmu);
550}
551
552early_initcall(init_power6_pmu);
553