2 * Performance counter support for POWER6 processors.
4 * Copyright 2008-2009 Paul Mackerras, IBM Corporation.
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
11 #include <linux/kernel.h>
12 #include <linux/perf_counter.h>
13 #include <linux/string.h>
15 #include <asm/cputable.h>
18 * Bits in event code for POWER6
20 #define PM_PMC_SH 20 /* PMC number (1-based) for direct events */
21 #define PM_PMC_MSK 0x7
22 #define PM_PMC_MSKS (PM_PMC_MSK << PM_PMC_SH)
23 #define PM_UNIT_SH 16 /* Unit event comes (TTMxSEL encoding) */
24 #define PM_UNIT_MSK 0xf
25 #define PM_UNIT_MSKS (PM_UNIT_MSK << PM_UNIT_SH)
26 #define PM_LLAV 0x8000 /* Load lookahead match value */
27 #define PM_LLA 0x4000 /* Load lookahead match enable */
28 #define PM_BYTE_SH 12 /* Byte of event bus to use */
30 #define PM_SUBUNIT_SH 8 /* Subunit event comes from (NEST_SEL enc.) */
31 #define PM_SUBUNIT_MSK 7
32 #define PM_SUBUNIT_MSKS (PM_SUBUNIT_MSK << PM_SUBUNIT_SH)
33 #define PM_PMCSEL_MSK 0xff /* PMCxSEL value */
34 #define PM_BUSEVENT_MSK 0xf3700
37 * Bits in MMCR1 for POWER6
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
54 * Map of which direct events on which PMCs are marked instruction events.
55 * Indexed by PMCSEL value >> 1.
56 * Bottom 4 bits are a map of which PMCs are interesting,
57 * top 4 bits say what sort of event:
58 * 0 = direct marked event,
59 * 1 = byte decode event,
60 * 4 = add/and event (PMC1 -> bits 0 & 4),
61 * 5 = add/and event (PMC1 -> bits 1 & 5),
62 * 6 = add/and event (PMC1 -> bits 2 & 6),
63 * 7 = add/and event (PMC1 -> bits 3 & 7).
65 static unsigned char direct_event_is_marked[0x60 >> 1] = {
69 0x07, /* 06 PM_MRK_ST_CMPL, PM_MRK_ST_GPS, PM_MRK_ST_CMPL_INT */
70 0x04, /* 08 PM_MRK_DFU_FIN */
71 0x06, /* 0a PM_MRK_IFU_FIN, PM_MRK_INST_FIN */
74 0x02, /* 10 PM_MRK_INST_DISP */
75 0x08, /* 12 PM_MRK_LSU_DERAT_MISS */
78 0x0c, /* 18 PM_THRESH_TIMEO, PM_MRK_INST_FIN */
79 0x0f, /* 1a PM_MRK_INST_DISP, PM_MRK_{FXU,FPU,LSU}_FIN */
80 0x01, /* 1c PM_MRK_INST_ISSUED */
86 0x15, /* 28 PM_MRK_DATA_FROM_L2MISS, PM_MRK_DATA_FROM_L3MISS */
97 0x08, /* 3e PM_MRK_INST_TIMEO */
107 0x05, /* 52 PM_MRK_BR_TAKEN, PM_MRK_BR_MPRED */
108 0x1c, /* 54 PM_MRK_PTEG_FROM_L3MISS, PM_MRK_PTEG_FROM_L2MISS */
109 0x02, /* 56 PM_MRK_LD_MISS_L1 */
117 * Masks showing for each unit which bits are marked events.
118 * These masks are in LE order, i.e. 0x00000001 is byte 0, bit 0.
120 static u32 marked_bus_events[16] = {
121 0x01000000, /* direct events set 1: byte 3 bit 0 */
122 0x00010000, /* direct events set 2: byte 2 bit 0 */
123 0, 0, 0, 0, /* IDU, IFU, nest: nothing */
124 0x00000088, /* VMX set 1: byte 0 bits 3, 7 */
125 0x000000c0, /* VMX set 2: byte 0 bits 4-7 */
126 0x04010000, /* LSU set 1: byte 2 bit 0, byte 3 bit 2 */
127 0xff010000u, /* LSU set 2: byte 2 bit 0, all of byte 3 */
129 0x00000010, /* VMX set 3: byte 0 bit 4 */
131 0x00000022, /* BFP set 2: byte 0 bits 1, 5 */
136 * Returns 1 if event counts things relating to marked instructions
137 * and thus needs the MMCRA_SAMPLE_ENABLE bit set, or 0 if not.
139 static int power6_marked_instr_event(u64 event)
141 int pmc, psel, ptype;
145 pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
146 psel = (event & PM_PMCSEL_MSK) >> 1; /* drop edge/level bit */
151 if (psel < sizeof(direct_event_is_marked)) {
152 ptype = direct_event_is_marked[psel];
153 if (pmc == 0 || !(ptype & (1 << (pmc - 1))))
161 bit = ptype ^ (pmc - 1);
162 } else if ((psel & 0x48) == 0x40)
165 if (!(event & PM_BUSEVENT_MSK) || bit == -1)
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;
175 * Assign PMC numbers and compute MMCR1 value for a set of events
177 static int p6_compute_mmcr(u64 event[], int n_ev,
178 unsigned int hwc[], unsigned long mmcr[])
180 unsigned long mmcr1 = 0;
181 unsigned long mmcra = 0;
183 unsigned int pmc, ev, b, u, s, psel;
184 unsigned int ttmset = 0;
185 unsigned int pmc_inuse = 0;
189 for (i = 0; i < n_ev; ++i) {
190 pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK;
192 if (pmc_inuse & (1 << (pmc - 1)))
193 return -1; /* collision! */
194 pmc_inuse |= 1 << (pmc - 1);
197 for (i = 0; i < n_ev; ++i) {
199 pmc = (ev >> PM_PMC_SH) & PM_PMC_MSK;
203 /* can go on any PMC; find a free one */
204 for (pmc = 0; pmc < 4; ++pmc)
205 if (!(pmc_inuse & (1 << pmc)))
209 pmc_inuse |= 1 << pmc;
212 psel = ev & PM_PMCSEL_MSK;
213 if (ev & PM_BUSEVENT_MSK) {
214 /* this event uses the event bus */
215 b = (ev >> PM_BYTE_SH) & PM_BYTE_MSK;
216 u = (ev >> PM_UNIT_SH) & PM_UNIT_MSK;
217 /* check for conflict on this byte of event bus */
218 if ((ttmset & (1 << b)) && MMCR1_TTMSEL(mmcr1, b) != u)
220 mmcr1 |= (unsigned long)u << MMCR1_TTMSEL_SH(b);
223 /* Nest events have a further mux */
224 s = (ev >> PM_SUBUNIT_SH) & PM_SUBUNIT_MSK;
225 if ((ttmset & 0x10) &&
226 MMCR1_NESTSEL(mmcr1) != s)
229 mmcr1 |= (unsigned long)s << MMCR1_NESTSEL_SH;
231 if (0x30 <= psel && psel <= 0x3d) {
232 /* these need the PMCx_ADDR_SEL bits */
234 mmcr1 |= MMCR1_PMC1_ADDR_SEL >> pmc;
236 /* bus select values are different for PMC3/4 */
237 if (pmc >= 2 && (psel & 0x90) == 0x80)
241 mmcr1 |= MMCR1_PMC1_LLA >> pmc;
243 mmcr1 |= MMCR1_PMC1_LLA_VALUE >> pmc;
245 if (power6_marked_instr_event(event[i]))
246 mmcra |= MMCRA_SAMPLE_ENABLE;
248 mmcr1 |= (unsigned long)psel << MMCR1_PMCSEL_SH(pmc);
252 mmcr[0] = MMCR0_PMC1CE;
254 mmcr[0] |= MMCR0_PMCjCE;
261 * Layout of constraint bits:
263 * 0-1 add field: number of uses of PMC1 (max 1)
264 * 2-3, 4-5, 6-7, 8-9, 10-11: ditto for PMC2, 3, 4, 5, 6
265 * 12-15 add field: number of uses of PMC1-4 (max 4)
266 * 16-19 select field: unit on byte 0 of event bus
267 * 20-23, 24-27, 28-31 ditto for bytes 1, 2, 3
268 * 32-34 select field: nest (subunit) event selector
270 static int p6_get_constraint(u64 event, unsigned long *maskp,
273 int pmc, byte, sh, subunit;
274 unsigned long mask = 0, value = 0;
276 pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
278 if (pmc > 4 && !(event == 0x500009 || event == 0x600005))
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;
296 mask |= 0x8000; /* add field for count of PMC1-4 uses */
304 static int p6_limited_pmc_event(u64 event)
306 int pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
308 return pmc == 5 || pmc == 6;
311 #define MAX_ALT 4 /* at most 4 alternatives for any event */
313 static const unsigned int event_alternatives[][MAX_ALT] = {
314 { 0x0130e8, 0x2000f6, 0x3000fc }, /* PM_PTEG_RELOAD_VALID */
315 { 0x080080, 0x10000d, 0x30000c, 0x4000f0 }, /* PM_LD_MISS_L1 */
316 { 0x080088, 0x200054, 0x3000f0 }, /* PM_ST_MISS_L1 */
317 { 0x10000a, 0x2000f4, 0x600005 }, /* PM_RUN_CYC */
318 { 0x10000b, 0x2000f5 }, /* PM_RUN_COUNT */
319 { 0x10000e, 0x400010 }, /* PM_PURR */
320 { 0x100010, 0x4000f8 }, /* PM_FLUSH */
321 { 0x10001a, 0x200010 }, /* PM_MRK_INST_DISP */
322 { 0x100026, 0x3000f8 }, /* PM_TB_BIT_TRANS */
323 { 0x100054, 0x2000f0 }, /* PM_ST_FIN */
324 { 0x100056, 0x2000fc }, /* PM_L1_ICACHE_MISS */
325 { 0x1000f0, 0x40000a }, /* PM_INST_IMC_MATCH_CMPL */
326 { 0x1000f8, 0x200008 }, /* PM_GCT_EMPTY_CYC */
327 { 0x1000fc, 0x400006 }, /* PM_LSU_DERAT_MISS_CYC */
328 { 0x20000e, 0x400007 }, /* PM_LSU_DERAT_MISS */
329 { 0x200012, 0x300012 }, /* PM_INST_DISP */
330 { 0x2000f2, 0x3000f2 }, /* PM_INST_DISP */
331 { 0x2000f8, 0x300010 }, /* PM_EXT_INT */
332 { 0x2000fe, 0x300056 }, /* PM_DATA_FROM_L2MISS */
333 { 0x2d0030, 0x30001a }, /* PM_MRK_FPU_FIN */
334 { 0x30000a, 0x400018 }, /* PM_MRK_INST_FIN */
335 { 0x3000f6, 0x40000e }, /* PM_L1_DCACHE_RELOAD_VALID */
336 { 0x3000fe, 0x400056 }, /* PM_DATA_FROM_L3MISS */
340 * This could be made more efficient with a binary search on
341 * a presorted list, if necessary
343 static int find_alternatives_list(u64 event)
348 for (i = 0; i < ARRAY_SIZE(event_alternatives); ++i) {
349 if (event < event_alternatives[i][0])
351 for (j = 0; j < MAX_ALT; ++j) {
352 alt = event_alternatives[i][j];
353 if (!alt || event < alt)
362 static int p6_get_alternatives(u64 event, unsigned int flags, u64 alt[])
365 unsigned int psel, pmc;
366 unsigned int nalt = 1;
370 nlim = p6_limited_pmc_event(event);
372 /* check the alternatives table */
373 i = find_alternatives_list(event);
375 /* copy out alternatives from list */
376 for (j = 0; j < MAX_ALT; ++j) {
377 aevent = event_alternatives[i][j];
381 alt[nalt++] = aevent;
382 nlim += p6_limited_pmc_event(aevent);
386 /* Check for alternative ways of computing sum events */
387 /* PMCSEL 0x32 counter N == PMCSEL 0x34 counter 5-N */
388 psel = event & (PM_PMCSEL_MSK & ~1); /* ignore edge bit */
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);
394 /* PMCSEL 0x38 counter N == PMCSEL 0x3a counter N+/-2 */
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);
400 if (flags & PPMU_ONLY_COUNT_RUN) {
402 * We're only counting in RUN state,
403 * so PM_CYC is equivalent to PM_RUN_CYC,
404 * PM_INST_CMPL === PM_RUN_INST_CMPL, PM_PURR === PM_RUN_PURR.
405 * This doesn't include alternatives that don't provide
406 * any extra flexibility in assigning PMCs (e.g.
407 * 0x10000a for PM_RUN_CYC vs. 0x1e for PM_CYC).
408 * Note that even with these additional alternatives
409 * we never end up with more than 4 alternatives for any event.
412 for (i = 0; i < nalt; ++i) {
414 case 0x1e: /* PM_CYC */
415 alt[j++] = 0x600005; /* PM_RUN_CYC */
418 case 0x10000a: /* PM_RUN_CYC */
419 alt[j++] = 0x1e; /* PM_CYC */
421 case 2: /* PM_INST_CMPL */
422 alt[j++] = 0x500009; /* PM_RUN_INST_CMPL */
425 case 0x500009: /* PM_RUN_INST_CMPL */
426 alt[j++] = 2; /* PM_INST_CMPL */
428 case 0x10000e: /* PM_PURR */
429 alt[j++] = 0x4000f4; /* PM_RUN_PURR */
431 case 0x4000f4: /* PM_RUN_PURR */
432 alt[j++] = 0x10000e; /* PM_PURR */
439 if (!(flags & PPMU_LIMITED_PMC_OK) && nlim) {
440 /* remove the limited PMC events */
442 for (i = 0; i < nalt; ++i) {
443 if (!p6_limited_pmc_event(alt[i])) {
449 } else if ((flags & PPMU_LIMITED_PMC_REQD) && nlim < nalt) {
450 /* remove all but the limited PMC events */
452 for (i = 0; i < nalt; ++i) {
453 if (p6_limited_pmc_event(alt[i])) {
464 static void p6_disable_pmc(unsigned int pmc, unsigned long mmcr[])
466 /* Set PMCxSEL to 0 to disable PMCx */
468 mmcr[1] &= ~(0xffUL << MMCR1_PMCSEL_SH(pmc));
471 static int power6_generic_events[] = {
472 [PERF_COUNT_HW_CPU_CYCLES] = 0x1e,
473 [PERF_COUNT_HW_INSTRUCTIONS] = 2,
474 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x280030, /* LD_REF_L1 */
475 [PERF_COUNT_HW_CACHE_MISSES] = 0x30000c, /* LD_MISS_L1 */
476 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x410a0, /* BR_PRED */
477 [PERF_COUNT_HW_BRANCH_MISSES] = 0x400052, /* BR_MPRED */
480 #define C(x) PERF_COUNT_HW_CACHE_##x
483 * Table of generalized cache-related events.
484 * 0 means not supported, -1 means nonsensical, other values
486 * The "DTLB" and "ITLB" events relate to the DERAT and IERAT.
488 static int power6_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
489 [C(L1D)] = { /* RESULT_ACCESS RESULT_MISS */
490 [C(OP_READ)] = { 0x80082, 0x80080 },
491 [C(OP_WRITE)] = { 0x80086, 0x80088 },
492 [C(OP_PREFETCH)] = { 0x810a4, 0 },
494 [C(L1I)] = { /* RESULT_ACCESS RESULT_MISS */
495 [C(OP_READ)] = { 0, 0x100056 },
496 [C(OP_WRITE)] = { -1, -1 },
497 [C(OP_PREFETCH)] = { 0x4008c, 0 },
499 [C(LL)] = { /* RESULT_ACCESS RESULT_MISS */
500 [C(OP_READ)] = { 0x150730, 0x250532 },
501 [C(OP_WRITE)] = { 0x250432, 0x150432 },
502 [C(OP_PREFETCH)] = { 0x810a6, 0 },
504 [C(DTLB)] = { /* RESULT_ACCESS RESULT_MISS */
505 [C(OP_READ)] = { 0, 0x20000e },
506 [C(OP_WRITE)] = { -1, -1 },
507 [C(OP_PREFETCH)] = { -1, -1 },
509 [C(ITLB)] = { /* RESULT_ACCESS RESULT_MISS */
510 [C(OP_READ)] = { 0, 0x420ce },
511 [C(OP_WRITE)] = { -1, -1 },
512 [C(OP_PREFETCH)] = { -1, -1 },
514 [C(BPU)] = { /* RESULT_ACCESS RESULT_MISS */
515 [C(OP_READ)] = { 0x430e6, 0x400052 },
516 [C(OP_WRITE)] = { -1, -1 },
517 [C(OP_PREFETCH)] = { -1, -1 },
521 static struct power_pmu power6_pmu = {
524 .max_alternatives = MAX_ALT,
525 .add_fields = 0x1555,
526 .test_adder = 0x3000,
527 .compute_mmcr = p6_compute_mmcr,
528 .get_constraint = p6_get_constraint,
529 .get_alternatives = p6_get_alternatives,
530 .disable_pmc = p6_disable_pmc,
531 .limited_pmc_event = p6_limited_pmc_event,
532 .flags = PPMU_LIMITED_PMC5_6 | PPMU_ALT_SIPR,
533 .n_generic = ARRAY_SIZE(power6_generic_events),
534 .generic_events = power6_generic_events,
535 .cache_events = &power6_cache_events,
538 static int init_power6_pmu(void)
540 if (strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power6"))
543 return register_power_pmu(&power6_pmu);
546 arch_initcall(init_power6_pmu);