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>
13 #include <linux/string.h>
15 #include <asm/cputable.h>
18 * Bits in event code for POWER5 (not POWER5++)
20 #define PM_PMC_SH 20 /* PMC number (1-based) for direct events */
21 #define PM_PMC_MSK 0xf
22 #define PM_PMC_MSKS (PM_PMC_MSK << PM_PMC_SH)
23 #define PM_UNIT_SH 16 /* TTMMUX number and setting - unit select */
24 #define PM_UNIT_MSK 0xf
25 #define PM_BYTE_SH 12 /* Byte number of event bus to use */
27 #define PM_GRS_SH 8 /* Storage subsystem mux select */
29 #define PM_BUSEVENT_MSK 0x80 /* Set if event uses event bus */
30 #define PM_PMCSEL_MSK 0x7f
32 /* Values in PM_UNIT field */
42 #define PM_LASTUNIT 0xc
45 * Bits in MMCR1 for POWER5
47 #define MMCR1_TTM0SEL_SH 62
48 #define MMCR1_TTM1SEL_SH 60
49 #define MMCR1_TTM2SEL_SH 58
50 #define MMCR1_TTM3SEL_SH 56
51 #define MMCR1_TTMSEL_MSK 3
52 #define MMCR1_TD_CP_DBG0SEL_SH 54
53 #define MMCR1_TD_CP_DBG1SEL_SH 52
54 #define MMCR1_TD_CP_DBG2SEL_SH 50
55 #define MMCR1_TD_CP_DBG3SEL_SH 48
56 #define MMCR1_GRS_L2SEL_SH 46
57 #define MMCR1_GRS_L2SEL_MSK 3
58 #define MMCR1_GRS_L3SEL_SH 44
59 #define MMCR1_GRS_L3SEL_MSK 3
60 #define MMCR1_GRS_MCSEL_SH 41
61 #define MMCR1_GRS_MCSEL_MSK 7
62 #define MMCR1_GRS_FABSEL_SH 39
63 #define MMCR1_GRS_FABSEL_MSK 3
64 #define MMCR1_PMC1_ADDER_SEL_SH 35
65 #define MMCR1_PMC2_ADDER_SEL_SH 34
66 #define MMCR1_PMC3_ADDER_SEL_SH 33
67 #define MMCR1_PMC4_ADDER_SEL_SH 32
68 #define MMCR1_PMC1SEL_SH 25
69 #define MMCR1_PMC2SEL_SH 17
70 #define MMCR1_PMC3SEL_SH 9
71 #define MMCR1_PMC4SEL_SH 1
72 #define MMCR1_PMCSEL_SH(n) (MMCR1_PMC1SEL_SH - (n) * 8)
73 #define MMCR1_PMCSEL_MSK 0x7f
80 * Layout of constraint bits:
81 * 6666555555555544444444443333333333222222222211111111110000000000
82 * 3210987654321098765432109876543210987654321098765432109876543210
83 * <><>[ ><><>< ><> [ >[ >[ >< >< >< >< ><><><><><><>
84 * T0T1 NC G0G1G2 G3 UC PS1PS2 B0 B1 B2 B3 P6P5P4P3P2P1
86 * T0 - TTM0 constraint
87 * 54-55: TTM0SEL value (0=FPU, 2=IFU, 3=ISU1) 0xc0_0000_0000_0000
89 * T1 - TTM1 constraint
90 * 52-53: TTM1SEL value (0=IDU, 3=GRS) 0x30_0000_0000_0000
92 * NC - number of counters
93 * 51: NC error 0x0008_0000_0000_0000
94 * 48-50: number of events needing PMC1-4 0x0007_0000_0000_0000
96 * G0..G3 - GRS mux constraints
97 * 46-47: GRS_L2SEL value
98 * 44-45: GRS_L3SEL value
99 * 41-44: GRS_MCSEL value
100 * 39-40: GRS_FABSEL value
101 * Note that these match up with their bit positions in MMCR1
103 * UC - unit constraint: can't have all three of FPU|IFU|ISU1, ISU0, IDU|GRS
104 * 37: UC3 error 0x20_0000_0000
105 * 36: FPU|IFU|ISU1 events needed 0x10_0000_0000
106 * 35: ISU0 events needed 0x08_0000_0000
107 * 34: IDU|GRS events needed 0x04_0000_0000
110 * 33: PS1 error 0x2_0000_0000
111 * 31-32: count of events needing PMC1/2 0x1_8000_0000
114 * 30: PS2 error 0x4000_0000
115 * 28-29: count of events needing PMC3/4 0x3000_0000
118 * 24-27: Byte 0 event source 0x0f00_0000
119 * Encoding as for the event code
122 * 20-23, 16-19, 12-15: Byte 1, 2, 3 event sources
125 * 0-11: Count of events needing PMC1..PMC6
128 static const int grsel_shift
[8] = {
129 MMCR1_GRS_L2SEL_SH
, MMCR1_GRS_L2SEL_SH
, MMCR1_GRS_L2SEL_SH
,
130 MMCR1_GRS_L3SEL_SH
, MMCR1_GRS_L3SEL_SH
, MMCR1_GRS_L3SEL_SH
,
131 MMCR1_GRS_MCSEL_SH
, MMCR1_GRS_FABSEL_SH
134 /* Masks and values for using events from the various units */
135 static unsigned long unit_cons
[PM_LASTUNIT
+1][2] = {
136 [PM_FPU
] = { 0xc0002000000000ul
, 0x00001000000000ul
},
137 [PM_ISU0
] = { 0x00002000000000ul
, 0x00000800000000ul
},
138 [PM_ISU1
] = { 0xc0002000000000ul
, 0xc0001000000000ul
},
139 [PM_IFU
] = { 0xc0002000000000ul
, 0x80001000000000ul
},
140 [PM_IDU
] = { 0x30002000000000ul
, 0x00000400000000ul
},
141 [PM_GRS
] = { 0x30002000000000ul
, 0x30000400000000ul
},
144 static int power5_get_constraint(u64 event
, unsigned long *maskp
,
147 int pmc
, byte
, unit
, sh
;
149 unsigned long mask
= 0, value
= 0;
152 pmc
= (event
>> PM_PMC_SH
) & PM_PMC_MSK
;
160 grp
= (pmc
- 1) >> 1;
161 else if (event
!= 0x500009 && event
!= 0x600005)
164 if (event
& PM_BUSEVENT_MSK
) {
165 unit
= (event
>> PM_UNIT_SH
) & PM_UNIT_MSK
;
166 if (unit
> PM_LASTUNIT
)
168 if (unit
== PM_ISU0_ALT
)
170 mask
|= unit_cons
[unit
][0];
171 value
|= unit_cons
[unit
][1];
172 byte
= (event
>> PM_BYTE_SH
) & PM_BYTE_MSK
;
176 /* Map LSU1 low word (bytes 4-7) to unit LSU1+1 */
180 if (unit
== PM_GRS
) {
182 fmask
= (bit
== 6)? 7: 3;
183 sh
= grsel_shift
[bit
];
184 mask
|= (unsigned long)fmask
<< sh
;
185 value
|= (unsigned long)((event
>> PM_GRS_SH
) & fmask
)
189 * Bus events on bytes 0 and 2 can be counted
190 * on PMC1/2; bytes 1 and 3 on PMC3/4.
194 /* Set byte lane select field */
195 mask
|= 0xfUL
<< (24 - 4 * byte
);
196 value
|= (unsigned long)unit
<< (24 - 4 * byte
);
199 /* increment PMC1/2 field */
200 mask
|= 0x200000000ul
;
201 value
|= 0x080000000ul
;
202 } else if (grp
== 1) {
203 /* increment PMC3/4 field */
204 mask
|= 0x40000000ul
;
205 value
|= 0x10000000ul
;
208 /* need a counter from PMC1-4 set */
209 mask
|= 0x8000000000000ul
;
210 value
|= 0x1000000000000ul
;
217 #define MAX_ALT 3 /* at most 3 alternatives for any event */
219 static const unsigned int event_alternatives
[][MAX_ALT
] = {
220 { 0x120e4, 0x400002 }, /* PM_GRP_DISP_REJECT */
221 { 0x410c7, 0x441084 }, /* PM_THRD_L2MISS_BOTH_CYC */
222 { 0x100005, 0x600005 }, /* PM_RUN_CYC */
223 { 0x100009, 0x200009, 0x500009 }, /* PM_INST_CMPL */
224 { 0x300009, 0x400009 }, /* PM_INST_DISP */
228 * Scan the alternatives table for a match and return the
229 * index into the alternatives table if found, else -1.
231 static int find_alternative(u64 event
)
235 for (i
= 0; i
< ARRAY_SIZE(event_alternatives
); ++i
) {
236 if (event
< event_alternatives
[i
][0])
238 for (j
= 0; j
< MAX_ALT
&& event_alternatives
[i
][j
]; ++j
)
239 if (event
== event_alternatives
[i
][j
])
245 static const unsigned char bytedecode_alternatives
[4][4] = {
246 /* PMC 1 */ { 0x21, 0x23, 0x25, 0x27 },
247 /* PMC 2 */ { 0x07, 0x17, 0x0e, 0x1e },
248 /* PMC 3 */ { 0x20, 0x22, 0x24, 0x26 },
249 /* PMC 4 */ { 0x07, 0x17, 0x0e, 0x1e }
253 * Some direct events for decodes of event bus byte 3 have alternative
254 * PMCSEL values on other counters. This returns the alternative
255 * event code for those that do, or -1 otherwise.
257 static s64
find_alternative_bdecode(u64 event
)
259 int pmc
, altpmc
, pp
, j
;
261 pmc
= (event
>> PM_PMC_SH
) & PM_PMC_MSK
;
262 if (pmc
== 0 || pmc
> 4)
264 altpmc
= 5 - pmc
; /* 1 <-> 4, 2 <-> 3 */
265 pp
= event
& PM_PMCSEL_MSK
;
266 for (j
= 0; j
< 4; ++j
) {
267 if (bytedecode_alternatives
[pmc
- 1][j
] == pp
) {
268 return (event
& ~(PM_PMC_MSKS
| PM_PMCSEL_MSK
)) |
269 (altpmc
<< PM_PMC_SH
) |
270 bytedecode_alternatives
[altpmc
- 1][j
];
276 static int power5_get_alternatives(u64 event
, unsigned int flags
, u64 alt
[])
283 i
= find_alternative(event
);
285 for (j
= 0; j
< MAX_ALT
; ++j
) {
286 ae
= event_alternatives
[i
][j
];
287 if (ae
&& ae
!= event
)
291 ae
= find_alternative_bdecode(event
);
299 * Map of which direct events on which PMCs are marked instruction events.
300 * Indexed by PMCSEL value, bit i (LE) set if PMC i is a marked event.
301 * Bit 0 is set if it is marked for all PMCs.
302 * The 0x80 bit indicates a byte decode PMCSEL value.
304 static unsigned char direct_event_is_marked
[0x28] = {
306 0x1f, /* 01 PM_IOPS_CMPL */
307 0x2, /* 02 PM_MRK_GRP_DISP */
308 0xe, /* 03 PM_MRK_ST_CMPL, PM_MRK_ST_GPS, PM_MRK_ST_CMPL_INT */
310 0x1c, /* 05 PM_MRK_BRU_FIN, PM_MRK_INST_FIN, PM_MRK_CRU_FIN */
313 0, 0, 0,/* 08 - 0a */
314 0x18, /* 0b PM_THRESH_TIMEO, PM_MRK_GRP_TIMEO */
320 0x14, /* 11 PM_MRK_GRP_BR_REDIR, PM_MRK_GRP_IC_MISS */
322 0x10, /* 13 PM_MRK_GRP_CMPL */
323 0x1f, /* 14 PM_GRP_MRK, PM_MRK_{FXU,FPU,LSU}_FIN */
324 0x2, /* 15 PM_MRK_GRP_ISSUED */
342 * Returns 1 if event counts things relating to marked instructions
343 * and thus needs the MMCRA_SAMPLE_ENABLE bit set, or 0 if not.
345 static int power5_marked_instr_event(u64 event
)
351 pmc
= (event
>> PM_PMC_SH
) & PM_PMC_MSK
;
352 psel
= event
& PM_PMCSEL_MSK
;
357 if (psel
< sizeof(direct_event_is_marked
)) {
358 if (direct_event_is_marked
[psel
] & (1 << pmc
))
360 if (direct_event_is_marked
[psel
] & 0x80)
362 else if (psel
== 0x08)
364 else if (psel
== 0x10)
366 else if (psel
== 0x1b && (pmc
== 1 || pmc
== 3))
368 } else if ((psel
& 0x58) == 0x40)
371 if (!(event
& PM_BUSEVENT_MSK
))
374 byte
= (event
>> PM_BYTE_SH
) & PM_BYTE_MSK
;
375 unit
= (event
>> PM_UNIT_SH
) & PM_UNIT_MSK
;
376 if (unit
== PM_LSU0
) {
377 /* byte 1 bits 0-7, byte 2 bits 0,2-4,6 */
379 } else if (unit
== PM_LSU1
&& byte
>= 4) {
381 /* byte 4 bits 1,3,5,7, byte 5 bits 6-7, byte 7 bits 0-4,6 */
386 return (mask
>> (byte
* 8 + bit
)) & 1;
389 static int power5_compute_mmcr(u64 event
[], int n_ev
,
390 unsigned int hwc
[], unsigned long mmcr
[])
392 unsigned long mmcr1
= 0;
393 unsigned long mmcra
= 0;
394 unsigned int pmc
, unit
, byte
, psel
;
395 unsigned int ttm
, grp
;
396 int i
, isbus
, bit
, grsel
;
397 unsigned int pmc_inuse
= 0;
398 unsigned int pmc_grp_use
[2];
399 unsigned char busbyte
[4];
400 unsigned char unituse
[16];
406 /* First pass to count resource use */
407 pmc_grp_use
[0] = pmc_grp_use
[1] = 0;
408 memset(busbyte
, 0, sizeof(busbyte
));
409 memset(unituse
, 0, sizeof(unituse
));
410 for (i
= 0; i
< n_ev
; ++i
) {
411 pmc
= (event
[i
] >> PM_PMC_SH
) & PM_PMC_MSK
;
415 if (pmc_inuse
& (1 << (pmc
- 1)))
417 pmc_inuse
|= 1 << (pmc
- 1);
418 /* count 1/2 vs 3/4 use */
420 ++pmc_grp_use
[(pmc
- 1) >> 1];
422 if (event
[i
] & PM_BUSEVENT_MSK
) {
423 unit
= (event
[i
] >> PM_UNIT_SH
) & PM_UNIT_MSK
;
424 byte
= (event
[i
] >> PM_BYTE_SH
) & PM_BYTE_MSK
;
425 if (unit
> PM_LASTUNIT
)
427 if (unit
== PM_ISU0_ALT
)
436 ++pmc_grp_use
[byte
& 1];
437 if (busbyte
[byte
] && busbyte
[byte
] != unit
)
439 busbyte
[byte
] = unit
;
443 if (pmc_grp_use
[0] > 2 || pmc_grp_use
[1] > 2)
447 * Assign resources and set multiplexer selects.
449 * PM_ISU0 can go either on TTM0 or TTM1, but that's the only
450 * choice we have to deal with.
452 if (unituse
[PM_ISU0
] &
453 (unituse
[PM_FPU
] | unituse
[PM_IFU
] | unituse
[PM_ISU1
])) {
454 unituse
[PM_ISU0_ALT
] = 1; /* move ISU to TTM1 */
455 unituse
[PM_ISU0
] = 0;
457 /* Set TTM[01]SEL fields. */
459 for (i
= PM_FPU
; i
<= PM_ISU1
; ++i
) {
464 mmcr1
|= (unsigned long)i
<< MMCR1_TTM0SEL_SH
;
467 for (; i
<= PM_GRS
; ++i
) {
472 mmcr1
|= (unsigned long)(i
& 3) << MMCR1_TTM1SEL_SH
;
477 /* Set byte lane select fields, TTM[23]SEL and GRS_*SEL. */
478 for (byte
= 0; byte
< 4; ++byte
) {
479 unit
= busbyte
[byte
];
482 if (unit
== PM_ISU0
&& unituse
[PM_ISU0_ALT
]) {
483 /* get ISU0 through TTM1 rather than TTM0 */
485 } else if (unit
== PM_LSU1
+ 1) {
486 /* select lower word of LSU1 for this byte */
487 mmcr1
|= 1ul << (MMCR1_TTM3SEL_SH
+ 3 - byte
);
490 mmcr1
|= (unsigned long)ttm
491 << (MMCR1_TD_CP_DBG0SEL_SH
- 2 * byte
);
494 /* Second pass: assign PMCs, set PMCxSEL and PMCx_ADDER_SEL fields */
495 for (i
= 0; i
< n_ev
; ++i
) {
496 pmc
= (event
[i
] >> PM_PMC_SH
) & PM_PMC_MSK
;
497 unit
= (event
[i
] >> PM_UNIT_SH
) & PM_UNIT_MSK
;
498 byte
= (event
[i
] >> PM_BYTE_SH
) & PM_BYTE_MSK
;
499 psel
= event
[i
] & PM_PMCSEL_MSK
;
500 isbus
= event
[i
] & PM_BUSEVENT_MSK
;
502 /* Bus event or any-PMC direct event */
503 for (pmc
= 0; pmc
< 4; ++pmc
) {
504 if (pmc_inuse
& (1 << pmc
))
506 grp
= (pmc
>> 1) & 1;
508 if (grp
== (byte
& 1))
510 } else if (pmc_grp_use
[grp
] < 2) {
515 pmc_inuse
|= 1 << pmc
;
516 } else if (pmc
<= 4) {
519 if ((psel
== 8 || psel
== 0x10) && isbus
&& (byte
& 2))
520 /* add events on higher-numbered bus */
521 mmcr1
|= 1ul << (MMCR1_PMC1_ADDER_SEL_SH
- pmc
);
523 /* Instructions or run cycles on PMC5/6 */
526 if (isbus
&& unit
== PM_GRS
) {
528 grsel
= (event
[i
] >> PM_GRS_SH
) & PM_GRS_MSK
;
529 mmcr1
|= (unsigned long)grsel
<< grsel_shift
[bit
];
531 if (power5_marked_instr_event(event
[i
]))
532 mmcra
|= MMCRA_SAMPLE_ENABLE
;
534 mmcr1
|= psel
<< MMCR1_PMCSEL_SH(pmc
);
538 /* Return MMCRx values */
541 mmcr
[0] = MMCR0_PMC1CE
;
542 if (pmc_inuse
& 0x3e)
543 mmcr
[0] |= MMCR0_PMCjCE
;
549 static void power5_disable_pmc(unsigned int pmc
, unsigned long mmcr
[])
552 mmcr
[1] &= ~(0x7fUL
<< MMCR1_PMCSEL_SH(pmc
));
555 static int power5_generic_events
[] = {
556 [PERF_COUNT_HW_CPU_CYCLES
] = 0xf,
557 [PERF_COUNT_HW_INSTRUCTIONS
] = 0x100009,
558 [PERF_COUNT_HW_CACHE_REFERENCES
] = 0x4c1090, /* LD_REF_L1 */
559 [PERF_COUNT_HW_CACHE_MISSES
] = 0x3c1088, /* LD_MISS_L1 */
560 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS
] = 0x230e4, /* BR_ISSUED */
561 [PERF_COUNT_HW_BRANCH_MISSES
] = 0x230e5, /* BR_MPRED_CR */
564 #define C(x) PERF_COUNT_HW_CACHE_##x
567 * Table of generalized cache-related events.
568 * 0 means not supported, -1 means nonsensical, other values
571 static int power5_cache_events
[C(MAX
)][C(OP_MAX
)][C(RESULT_MAX
)] = {
572 [C(L1D
)] = { /* RESULT_ACCESS RESULT_MISS */
573 [C(OP_READ
)] = { 0x4c1090, 0x3c1088 },
574 [C(OP_WRITE
)] = { 0x3c1090, 0xc10c3 },
575 [C(OP_PREFETCH
)] = { 0xc70e7, 0 },
577 [C(L1I
)] = { /* RESULT_ACCESS RESULT_MISS */
578 [C(OP_READ
)] = { 0, 0 },
579 [C(OP_WRITE
)] = { -1, -1 },
580 [C(OP_PREFETCH
)] = { 0, 0 },
582 [C(LL
)] = { /* RESULT_ACCESS RESULT_MISS */
583 [C(OP_READ
)] = { 0, 0x3c309b },
584 [C(OP_WRITE
)] = { 0, 0 },
585 [C(OP_PREFETCH
)] = { 0xc50c3, 0 },
587 [C(DTLB
)] = { /* RESULT_ACCESS RESULT_MISS */
588 [C(OP_READ
)] = { 0x2c4090, 0x800c4 },
589 [C(OP_WRITE
)] = { -1, -1 },
590 [C(OP_PREFETCH
)] = { -1, -1 },
592 [C(ITLB
)] = { /* RESULT_ACCESS RESULT_MISS */
593 [C(OP_READ
)] = { 0, 0x800c0 },
594 [C(OP_WRITE
)] = { -1, -1 },
595 [C(OP_PREFETCH
)] = { -1, -1 },
597 [C(BPU
)] = { /* RESULT_ACCESS RESULT_MISS */
598 [C(OP_READ
)] = { 0x230e4, 0x230e5 },
599 [C(OP_WRITE
)] = { -1, -1 },
600 [C(OP_PREFETCH
)] = { -1, -1 },
604 static struct power_pmu power5_pmu
= {
607 .max_alternatives
= MAX_ALT
,
608 .add_fields
= 0x7000090000555ul
,
609 .test_adder
= 0x3000490000000ul
,
610 .compute_mmcr
= power5_compute_mmcr
,
611 .get_constraint
= power5_get_constraint
,
612 .get_alternatives
= power5_get_alternatives
,
613 .disable_pmc
= power5_disable_pmc
,
614 .n_generic
= ARRAY_SIZE(power5_generic_events
),
615 .generic_events
= power5_generic_events
,
616 .cache_events
= &power5_cache_events
,
619 static int init_power5_pmu(void)
621 if (strcmp(cur_cpu_spec
->oprofile_cpu_type
, "ppc64/power5"))
624 return register_power_pmu(&power5_pmu
);
627 arch_initcall(init_power5_pmu
);