2 * Performance counter support for POWER5 (not POWER5++) processors.
4 * Copyright 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>
16 * Bits in event code for POWER5 (not POWER5++)
18 #define PM_PMC_SH 20 /* 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 16 /* TTMMUX number and setting - unit select */
22 #define PM_UNIT_MSK 0xf
23 #define PM_BYTE_SH 12 /* Byte number of event bus to use */
25 #define PM_GRS_SH 8 /* Storage subsystem mux select */
27 #define PM_BUSEVENT_MSK 0x80 /* Set if event uses event bus */
28 #define PM_PMCSEL_MSK 0x7f
30 /* Values in PM_UNIT field */
40 #define PM_LASTUNIT 0xc
43 * Bits in MMCR1 for POWER5
45 #define MMCR1_TTM0SEL_SH 62
46 #define MMCR1_TTM1SEL_SH 60
47 #define MMCR1_TTM2SEL_SH 58
48 #define MMCR1_TTM3SEL_SH 56
49 #define MMCR1_TTMSEL_MSK 3
50 #define MMCR1_TD_CP_DBG0SEL_SH 54
51 #define MMCR1_TD_CP_DBG1SEL_SH 52
52 #define MMCR1_TD_CP_DBG2SEL_SH 50
53 #define MMCR1_TD_CP_DBG3SEL_SH 48
54 #define MMCR1_GRS_L2SEL_SH 46
55 #define MMCR1_GRS_L2SEL_MSK 3
56 #define MMCR1_GRS_L3SEL_SH 44
57 #define MMCR1_GRS_L3SEL_MSK 3
58 #define MMCR1_GRS_MCSEL_SH 41
59 #define MMCR1_GRS_MCSEL_MSK 7
60 #define MMCR1_GRS_FABSEL_SH 39
61 #define MMCR1_GRS_FABSEL_MSK 3
62 #define MMCR1_PMC1_ADDER_SEL_SH 35
63 #define MMCR1_PMC2_ADDER_SEL_SH 34
64 #define MMCR1_PMC3_ADDER_SEL_SH 33
65 #define MMCR1_PMC4_ADDER_SEL_SH 32
66 #define MMCR1_PMC1SEL_SH 25
67 #define MMCR1_PMC2SEL_SH 17
68 #define MMCR1_PMC3SEL_SH 9
69 #define MMCR1_PMC4SEL_SH 1
70 #define MMCR1_PMCSEL_SH(n) (MMCR1_PMC1SEL_SH - (n) * 8)
71 #define MMCR1_PMCSEL_MSK 0x7f
78 * Layout of constraint bits:
79 * 6666555555555544444444443333333333222222222211111111110000000000
80 * 3210987654321098765432109876543210987654321098765432109876543210
81 * <><>[ ><><>< ><> [ >[ >[ >< >< >< >< ><><><><><><>
82 * T0T1 NC G0G1G2 G3 UC PS1PS2 B0 B1 B2 B3 P6P5P4P3P2P1
84 * T0 - TTM0 constraint
85 * 54-55: TTM0SEL value (0=FPU, 2=IFU, 3=ISU1) 0xc0_0000_0000_0000
87 * T1 - TTM1 constraint
88 * 52-53: TTM1SEL value (0=IDU, 3=GRS) 0x30_0000_0000_0000
90 * NC - number of counters
91 * 51: NC error 0x0008_0000_0000_0000
92 * 48-50: number of events needing PMC1-4 0x0007_0000_0000_0000
94 * G0..G3 - GRS mux constraints
95 * 46-47: GRS_L2SEL value
96 * 44-45: GRS_L3SEL value
97 * 41-44: GRS_MCSEL value
98 * 39-40: GRS_FABSEL value
99 * Note that these match up with their bit positions in MMCR1
101 * UC - unit constraint: can't have all three of FPU|IFU|ISU1, ISU0, IDU|GRS
102 * 37: UC3 error 0x20_0000_0000
103 * 36: FPU|IFU|ISU1 events needed 0x10_0000_0000
104 * 35: ISU0 events needed 0x08_0000_0000
105 * 34: IDU|GRS events needed 0x04_0000_0000
108 * 33: PS1 error 0x2_0000_0000
109 * 31-32: count of events needing PMC1/2 0x1_8000_0000
112 * 30: PS2 error 0x4000_0000
113 * 28-29: count of events needing PMC3/4 0x3000_0000
116 * 24-27: Byte 0 event source 0x0f00_0000
117 * Encoding as for the event code
120 * 20-23, 16-19, 12-15: Byte 1, 2, 3 event sources
123 * 0-11: Count of events needing PMC1..PMC6
126 static const int grsel_shift[8] = {
127 MMCR1_GRS_L2SEL_SH, MMCR1_GRS_L2SEL_SH, MMCR1_GRS_L2SEL_SH,
128 MMCR1_GRS_L3SEL_SH, MMCR1_GRS_L3SEL_SH, MMCR1_GRS_L3SEL_SH,
129 MMCR1_GRS_MCSEL_SH, MMCR1_GRS_FABSEL_SH
132 /* Masks and values for using events from the various units */
133 static u64 unit_cons[PM_LASTUNIT+1][2] = {
134 [PM_FPU] = { 0xc0002000000000ull, 0x00001000000000ull },
135 [PM_ISU0] = { 0x00002000000000ull, 0x00000800000000ull },
136 [PM_ISU1] = { 0xc0002000000000ull, 0xc0001000000000ull },
137 [PM_IFU] = { 0xc0002000000000ull, 0x80001000000000ull },
138 [PM_IDU] = { 0x30002000000000ull, 0x00000400000000ull },
139 [PM_GRS] = { 0x30002000000000ull, 0x30000400000000ull },
142 static int power5_get_constraint(u64 event, u64 *maskp, u64 *valp)
144 int pmc, byte, unit, sh;
146 u64 mask = 0, value = 0;
149 pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
157 grp = (pmc - 1) >> 1;
158 else if (event != 0x500009 && event != 0x600005)
161 if (event & PM_BUSEVENT_MSK) {
162 unit = (event >> PM_UNIT_SH) & PM_UNIT_MSK;
163 if (unit > PM_LASTUNIT)
165 if (unit == PM_ISU0_ALT)
167 mask |= unit_cons[unit][0];
168 value |= unit_cons[unit][1];
169 byte = (event >> PM_BYTE_SH) & PM_BYTE_MSK;
173 /* Map LSU1 low word (bytes 4-7) to unit LSU1+1 */
177 if (unit == PM_GRS) {
179 fmask = (bit == 6)? 7: 3;
180 sh = grsel_shift[bit];
181 mask |= (u64)fmask << sh;
182 value |= (u64)((event >> PM_GRS_SH) & fmask) << sh;
185 * Bus events on bytes 0 and 2 can be counted
186 * on PMC1/2; bytes 1 and 3 on PMC3/4.
190 /* Set byte lane select field */
191 mask |= 0xfULL << (24 - 4 * byte);
192 value |= (u64)unit << (24 - 4 * byte);
195 /* increment PMC1/2 field */
196 mask |= 0x200000000ull;
197 value |= 0x080000000ull;
198 } else if (grp == 1) {
199 /* increment PMC3/4 field */
200 mask |= 0x40000000ull;
201 value |= 0x10000000ull;
204 /* need a counter from PMC1-4 set */
205 mask |= 0x8000000000000ull;
206 value |= 0x1000000000000ull;
213 #define MAX_ALT 3 /* at most 3 alternatives for any event */
215 static const unsigned int event_alternatives[][MAX_ALT] = {
216 { 0x120e4, 0x400002 }, /* PM_GRP_DISP_REJECT */
217 { 0x410c7, 0x441084 }, /* PM_THRD_L2MISS_BOTH_CYC */
218 { 0x100005, 0x600005 }, /* PM_RUN_CYC */
219 { 0x100009, 0x200009, 0x500009 }, /* PM_INST_CMPL */
220 { 0x300009, 0x400009 }, /* PM_INST_DISP */
224 * Scan the alternatives table for a match and return the
225 * index into the alternatives table if found, else -1.
227 static int find_alternative(u64 event)
231 for (i = 0; i < ARRAY_SIZE(event_alternatives); ++i) {
232 if (event < event_alternatives[i][0])
234 for (j = 0; j < MAX_ALT && event_alternatives[i][j]; ++j)
235 if (event == event_alternatives[i][j])
241 static const unsigned char bytedecode_alternatives[4][4] = {
242 /* PMC 1 */ { 0x21, 0x23, 0x25, 0x27 },
243 /* PMC 2 */ { 0x07, 0x17, 0x0e, 0x1e },
244 /* PMC 3 */ { 0x20, 0x22, 0x24, 0x26 },
245 /* PMC 4 */ { 0x07, 0x17, 0x0e, 0x1e }
249 * Some direct events for decodes of event bus byte 3 have alternative
250 * PMCSEL values on other counters. This returns the alternative
251 * event code for those that do, or -1 otherwise.
253 static s64 find_alternative_bdecode(u64 event)
255 int pmc, altpmc, pp, j;
257 pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
258 if (pmc == 0 || pmc > 4)
260 altpmc = 5 - pmc; /* 1 <-> 4, 2 <-> 3 */
261 pp = event & PM_PMCSEL_MSK;
262 for (j = 0; j < 4; ++j) {
263 if (bytedecode_alternatives[pmc - 1][j] == pp) {
264 return (event & ~(PM_PMC_MSKS | PM_PMCSEL_MSK)) |
265 (altpmc << PM_PMC_SH) |
266 bytedecode_alternatives[altpmc - 1][j];
272 static int power5_get_alternatives(u64 event, unsigned int flags, u64 alt[])
279 i = find_alternative(event);
281 for (j = 0; j < MAX_ALT; ++j) {
282 ae = event_alternatives[i][j];
283 if (ae && ae != event)
287 ae = find_alternative_bdecode(event);
295 * Map of which direct events on which PMCs are marked instruction events.
296 * Indexed by PMCSEL value, bit i (LE) set if PMC i is a marked event.
297 * Bit 0 is set if it is marked for all PMCs.
298 * The 0x80 bit indicates a byte decode PMCSEL value.
300 static unsigned char direct_event_is_marked[0x28] = {
302 0x1f, /* 01 PM_IOPS_CMPL */
303 0x2, /* 02 PM_MRK_GRP_DISP */
304 0xe, /* 03 PM_MRK_ST_CMPL, PM_MRK_ST_GPS, PM_MRK_ST_CMPL_INT */
306 0x1c, /* 05 PM_MRK_BRU_FIN, PM_MRK_INST_FIN, PM_MRK_CRU_FIN */
309 0, 0, 0,/* 08 - 0a */
310 0x18, /* 0b PM_THRESH_TIMEO, PM_MRK_GRP_TIMEO */
316 0x14, /* 11 PM_MRK_GRP_BR_REDIR, PM_MRK_GRP_IC_MISS */
318 0x10, /* 13 PM_MRK_GRP_CMPL */
319 0x1f, /* 14 PM_GRP_MRK, PM_MRK_{FXU,FPU,LSU}_FIN */
320 0x2, /* 15 PM_MRK_GRP_ISSUED */
338 * Returns 1 if event counts things relating to marked instructions
339 * and thus needs the MMCRA_SAMPLE_ENABLE bit set, or 0 if not.
341 static int power5_marked_instr_event(u64 event)
347 pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
348 psel = event & PM_PMCSEL_MSK;
353 if (psel < sizeof(direct_event_is_marked)) {
354 if (direct_event_is_marked[psel] & (1 << pmc))
356 if (direct_event_is_marked[psel] & 0x80)
358 else if (psel == 0x08)
360 else if (psel == 0x10)
362 else if (psel == 0x1b && (pmc == 1 || pmc == 3))
364 } else if ((psel & 0x58) == 0x40)
367 if (!(event & PM_BUSEVENT_MSK))
370 byte = (event >> PM_BYTE_SH) & PM_BYTE_MSK;
371 unit = (event >> PM_UNIT_SH) & PM_UNIT_MSK;
372 if (unit == PM_LSU0) {
373 /* byte 1 bits 0-7, byte 2 bits 0,2-4,6 */
375 } else if (unit == PM_LSU1 && byte >= 4) {
377 /* byte 4 bits 1,3,5,7, byte 5 bits 6-7, byte 7 bits 0-4,6 */
382 return (mask >> (byte * 8 + bit)) & 1;
385 static int power5_compute_mmcr(u64 event[], int n_ev,
386 unsigned int hwc[], u64 mmcr[])
390 unsigned int pmc, unit, byte, psel;
391 unsigned int ttm, grp;
392 int i, isbus, bit, grsel;
393 unsigned int pmc_inuse = 0;
394 unsigned int pmc_grp_use[2];
395 unsigned char busbyte[4];
396 unsigned char unituse[16];
402 /* First pass to count resource use */
403 pmc_grp_use[0] = pmc_grp_use[1] = 0;
404 memset(busbyte, 0, sizeof(busbyte));
405 memset(unituse, 0, sizeof(unituse));
406 for (i = 0; i < n_ev; ++i) {
407 pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK;
411 if (pmc_inuse & (1 << (pmc - 1)))
413 pmc_inuse |= 1 << (pmc - 1);
414 /* count 1/2 vs 3/4 use */
416 ++pmc_grp_use[(pmc - 1) >> 1];
418 if (event[i] & PM_BUSEVENT_MSK) {
419 unit = (event[i] >> PM_UNIT_SH) & PM_UNIT_MSK;
420 byte = (event[i] >> PM_BYTE_SH) & PM_BYTE_MSK;
421 if (unit > PM_LASTUNIT)
423 if (unit == PM_ISU0_ALT)
432 ++pmc_grp_use[byte & 1];
433 if (busbyte[byte] && busbyte[byte] != unit)
435 busbyte[byte] = unit;
439 if (pmc_grp_use[0] > 2 || pmc_grp_use[1] > 2)
443 * Assign resources and set multiplexer selects.
445 * PM_ISU0 can go either on TTM0 or TTM1, but that's the only
446 * choice we have to deal with.
448 if (unituse[PM_ISU0] &
449 (unituse[PM_FPU] | unituse[PM_IFU] | unituse[PM_ISU1])) {
450 unituse[PM_ISU0_ALT] = 1; /* move ISU to TTM1 */
451 unituse[PM_ISU0] = 0;
453 /* Set TTM[01]SEL fields. */
455 for (i = PM_FPU; i <= PM_ISU1; ++i) {
460 mmcr1 |= (u64)i << MMCR1_TTM0SEL_SH;
463 for (; i <= PM_GRS; ++i) {
468 mmcr1 |= (u64)(i & 3) << MMCR1_TTM1SEL_SH;
473 /* Set byte lane select fields, TTM[23]SEL and GRS_*SEL. */
474 for (byte = 0; byte < 4; ++byte) {
475 unit = busbyte[byte];
478 if (unit == PM_ISU0 && unituse[PM_ISU0_ALT]) {
479 /* get ISU0 through TTM1 rather than TTM0 */
481 } else if (unit == PM_LSU1 + 1) {
482 /* select lower word of LSU1 for this byte */
483 mmcr1 |= 1ull << (MMCR1_TTM3SEL_SH + 3 - byte);
486 mmcr1 |= (u64)ttm << (MMCR1_TD_CP_DBG0SEL_SH - 2 * byte);
489 /* Second pass: assign PMCs, set PMCxSEL and PMCx_ADDER_SEL fields */
490 for (i = 0; i < n_ev; ++i) {
491 pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK;
492 unit = (event[i] >> PM_UNIT_SH) & PM_UNIT_MSK;
493 byte = (event[i] >> PM_BYTE_SH) & PM_BYTE_MSK;
494 psel = event[i] & PM_PMCSEL_MSK;
495 isbus = event[i] & PM_BUSEVENT_MSK;
497 /* Bus event or any-PMC direct event */
498 for (pmc = 0; pmc < 4; ++pmc) {
499 if (pmc_inuse & (1 << pmc))
501 grp = (pmc >> 1) & 1;
503 if (grp == (byte & 1))
505 } else if (pmc_grp_use[grp] < 2) {
510 pmc_inuse |= 1 << pmc;
511 } else if (pmc <= 4) {
514 if ((psel == 8 || psel == 0x10) && isbus && (byte & 2))
515 /* add events on higher-numbered bus */
516 mmcr1 |= 1ull << (MMCR1_PMC1_ADDER_SEL_SH - pmc);
518 /* Instructions or run cycles on PMC5/6 */
521 if (isbus && unit == PM_GRS) {
523 grsel = (event[i] >> PM_GRS_SH) & PM_GRS_MSK;
524 mmcr1 |= (u64)grsel << grsel_shift[bit];
526 if (power5_marked_instr_event(event[i]))
527 mmcra |= MMCRA_SAMPLE_ENABLE;
529 mmcr1 |= psel << MMCR1_PMCSEL_SH(pmc);
533 /* Return MMCRx values */
536 mmcr[0] = MMCR0_PMC1CE;
537 if (pmc_inuse & 0x3e)
538 mmcr[0] |= MMCR0_PMCjCE;
544 static void power5_disable_pmc(unsigned int pmc, u64 mmcr[])
547 mmcr[1] &= ~(0x7fUL << MMCR1_PMCSEL_SH(pmc));
550 static int power5_generic_events[] = {
551 [PERF_COUNT_HW_CPU_CYCLES] = 0xf,
552 [PERF_COUNT_HW_INSTRUCTIONS] = 0x100009,
553 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4c1090, /* LD_REF_L1 */
554 [PERF_COUNT_HW_CACHE_MISSES] = 0x3c1088, /* LD_MISS_L1 */
555 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x230e4, /* BR_ISSUED */
556 [PERF_COUNT_HW_BRANCH_MISSES] = 0x230e5, /* BR_MPRED_CR */
559 #define C(x) PERF_COUNT_HW_CACHE_##x
562 * Table of generalized cache-related events.
563 * 0 means not supported, -1 means nonsensical, other values
566 static int power5_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
567 [C(L1D)] = { /* RESULT_ACCESS RESULT_MISS */
568 [C(OP_READ)] = { 0x4c1090, 0x3c1088 },
569 [C(OP_WRITE)] = { 0x3c1090, 0xc10c3 },
570 [C(OP_PREFETCH)] = { 0xc70e7, 0 },
572 [C(L1I)] = { /* RESULT_ACCESS RESULT_MISS */
573 [C(OP_READ)] = { 0, 0 },
574 [C(OP_WRITE)] = { -1, -1 },
575 [C(OP_PREFETCH)] = { 0, 0 },
577 [C(LL)] = { /* RESULT_ACCESS RESULT_MISS */
578 [C(OP_READ)] = { 0, 0x3c309b },
579 [C(OP_WRITE)] = { 0, 0 },
580 [C(OP_PREFETCH)] = { 0xc50c3, 0 },
582 [C(DTLB)] = { /* RESULT_ACCESS RESULT_MISS */
583 [C(OP_READ)] = { 0x2c4090, 0x800c4 },
584 [C(OP_WRITE)] = { -1, -1 },
585 [C(OP_PREFETCH)] = { -1, -1 },
587 [C(ITLB)] = { /* RESULT_ACCESS RESULT_MISS */
588 [C(OP_READ)] = { 0, 0x800c0 },
589 [C(OP_WRITE)] = { -1, -1 },
590 [C(OP_PREFETCH)] = { -1, -1 },
592 [C(BPU)] = { /* RESULT_ACCESS RESULT_MISS */
593 [C(OP_READ)] = { 0x230e4, 0x230e5 },
594 [C(OP_WRITE)] = { -1, -1 },
595 [C(OP_PREFETCH)] = { -1, -1 },
599 struct power_pmu power5_pmu = {
601 .max_alternatives = MAX_ALT,
602 .add_fields = 0x7000090000555ull,
603 .test_adder = 0x3000490000000ull,
604 .compute_mmcr = power5_compute_mmcr,
605 .get_constraint = power5_get_constraint,
606 .get_alternatives = power5_get_alternatives,
607 .disable_pmc = power5_disable_pmc,
608 .n_generic = ARRAY_SIZE(power5_generic_events),
609 .generic_events = power5_generic_events,
610 .cache_events = &power5_cache_events,