1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Performance counter support for POWER6 processors.
5 * Copyright 2008-2009 Paul Mackerras, IBM Corporation.
7 #include <linux/kernel.h>
8 #include <linux/perf_event.h>
9 #include <linux/string.h>
11 #include <asm/cputable.h>
14 * Bits in event code for POWER6
16 #define PM_PMC_SH 20 /* PMC number (1-based) for direct events */
17 #define PM_PMC_MSK 0x7
18 #define PM_PMC_MSKS (PM_PMC_MSK << PM_PMC_SH)
19 #define PM_UNIT_SH 16 /* Unit event comes (TTMxSEL encoding) */
20 #define PM_UNIT_MSK 0xf
21 #define PM_UNIT_MSKS (PM_UNIT_MSK << PM_UNIT_SH)
22 #define PM_LLAV 0x8000 /* Load lookahead match value */
23 #define PM_LLA 0x4000 /* Load lookahead match enable */
24 #define PM_BYTE_SH 12 /* Byte of event bus to use */
26 #define PM_SUBUNIT_SH 8 /* Subunit event comes from (NEST_SEL enc.) */
27 #define PM_SUBUNIT_MSK 7
28 #define PM_SUBUNIT_MSKS (PM_SUBUNIT_MSK << PM_SUBUNIT_SH)
29 #define PM_PMCSEL_MSK 0xff /* PMCxSEL value */
30 #define PM_BUSEVENT_MSK 0xf3700
33 * Bits in MMCR1 for POWER6
35 #define MMCR1_TTM0SEL_SH 60
36 #define MMCR1_TTMSEL_SH(n) (MMCR1_TTM0SEL_SH - (n) * 4)
37 #define MMCR1_TTMSEL_MSK 0xf
38 #define MMCR1_TTMSEL(m, n) (((m) >> MMCR1_TTMSEL_SH(n)) & MMCR1_TTMSEL_MSK)
39 #define MMCR1_NESTSEL_SH 45
40 #define MMCR1_NESTSEL_MSK 0x7
41 #define MMCR1_NESTSEL(m) (((m) >> MMCR1_NESTSEL_SH) & MMCR1_NESTSEL_MSK)
42 #define MMCR1_PMC1_LLA (1ul << 44)
43 #define MMCR1_PMC1_LLA_VALUE (1ul << 39)
44 #define MMCR1_PMC1_ADDR_SEL (1ul << 35)
45 #define MMCR1_PMC1SEL_SH 24
46 #define MMCR1_PMCSEL_SH(n) (MMCR1_PMC1SEL_SH - (n) * 8)
47 #define MMCR1_PMCSEL_MSK 0xff
50 * Map of which direct events on which PMCs are marked instruction events.
51 * Indexed by PMCSEL value >> 1.
52 * Bottom 4 bits are a map of which PMCs are interesting,
53 * top 4 bits say what sort of event:
54 * 0 = direct marked event,
55 * 1 = byte decode event,
56 * 4 = add/and event (PMC1 -> bits 0 & 4),
57 * 5 = add/and event (PMC1 -> bits 1 & 5),
58 * 6 = add/and event (PMC1 -> bits 2 & 6),
59 * 7 = add/and event (PMC1 -> bits 3 & 7).
61 static unsigned char direct_event_is_marked
[0x60 >> 1] = {
65 0x07, /* 06 PM_MRK_ST_CMPL, PM_MRK_ST_GPS, PM_MRK_ST_CMPL_INT */
66 0x04, /* 08 PM_MRK_DFU_FIN */
67 0x06, /* 0a PM_MRK_IFU_FIN, PM_MRK_INST_FIN */
70 0x02, /* 10 PM_MRK_INST_DISP */
71 0x08, /* 12 PM_MRK_LSU_DERAT_MISS */
74 0x0c, /* 18 PM_THRESH_TIMEO, PM_MRK_INST_FIN */
75 0x0f, /* 1a PM_MRK_INST_DISP, PM_MRK_{FXU,FPU,LSU}_FIN */
76 0x01, /* 1c PM_MRK_INST_ISSUED */
82 0x15, /* 28 PM_MRK_DATA_FROM_L2MISS, PM_MRK_DATA_FROM_L3MISS */
93 0x08, /* 3e PM_MRK_INST_TIMEO */
103 0x05, /* 52 PM_MRK_BR_TAKEN, PM_MRK_BR_MPRED */
104 0x1c, /* 54 PM_MRK_PTEG_FROM_L3MISS, PM_MRK_PTEG_FROM_L2MISS */
105 0x02, /* 56 PM_MRK_LD_MISS_L1 */
113 * Masks showing for each unit which bits are marked events.
114 * These masks are in LE order, i.e. 0x00000001 is byte 0, bit 0.
116 static u32 marked_bus_events
[16] = {
117 0x01000000, /* direct events set 1: byte 3 bit 0 */
118 0x00010000, /* direct events set 2: byte 2 bit 0 */
119 0, 0, 0, 0, /* IDU, IFU, nest: nothing */
120 0x00000088, /* VMX set 1: byte 0 bits 3, 7 */
121 0x000000c0, /* VMX set 2: byte 0 bits 4-7 */
122 0x04010000, /* LSU set 1: byte 2 bit 0, byte 3 bit 2 */
123 0xff010000u
, /* LSU set 2: byte 2 bit 0, all of byte 3 */
125 0x00000010, /* VMX set 3: byte 0 bit 4 */
127 0x00000022, /* BFP set 2: byte 0 bits 1, 5 */
132 * Returns 1 if event counts things relating to marked instructions
133 * and thus needs the MMCRA_SAMPLE_ENABLE bit set, or 0 if not.
135 static int power6_marked_instr_event(u64 event
)
137 int pmc
, psel
, ptype
;
141 pmc
= (event
>> PM_PMC_SH
) & PM_PMC_MSK
;
142 psel
= (event
& PM_PMCSEL_MSK
) >> 1; /* drop edge/level bit */
147 if (psel
< sizeof(direct_event_is_marked
)) {
148 ptype
= direct_event_is_marked
[psel
];
149 if (pmc
== 0 || !(ptype
& (1 << (pmc
- 1))))
157 bit
= ptype
^ (pmc
- 1);
158 } else if ((psel
& 0x48) == 0x40)
161 if (!(event
& PM_BUSEVENT_MSK
) || bit
== -1)
164 byte
= (event
>> PM_BYTE_SH
) & PM_BYTE_MSK
;
165 unit
= (event
>> PM_UNIT_SH
) & PM_UNIT_MSK
;
166 mask
= marked_bus_events
[unit
];
167 return (mask
>> (byte
* 8 + bit
)) & 1;
171 * Assign PMC numbers and compute MMCR1 value for a set of events
173 static int p6_compute_mmcr(u64 event
[], int n_ev
,
174 unsigned int hwc
[], unsigned long mmcr
[], struct perf_event
*pevents
[])
176 unsigned long mmcr1
= 0;
177 unsigned long mmcra
= MMCRA_SDAR_DCACHE_MISS
| MMCRA_SDAR_ERAT_MISS
;
179 unsigned int pmc
, ev
, b
, u
, s
, psel
;
180 unsigned int ttmset
= 0;
181 unsigned int pmc_inuse
= 0;
185 for (i
= 0; i
< n_ev
; ++i
) {
186 pmc
= (event
[i
] >> PM_PMC_SH
) & PM_PMC_MSK
;
188 if (pmc_inuse
& (1 << (pmc
- 1)))
189 return -1; /* collision! */
190 pmc_inuse
|= 1 << (pmc
- 1);
193 for (i
= 0; i
< n_ev
; ++i
) {
195 pmc
= (ev
>> PM_PMC_SH
) & PM_PMC_MSK
;
199 /* can go on any PMC; find a free one */
200 for (pmc
= 0; pmc
< 4; ++pmc
)
201 if (!(pmc_inuse
& (1 << pmc
)))
205 pmc_inuse
|= 1 << pmc
;
208 psel
= ev
& PM_PMCSEL_MSK
;
209 if (ev
& PM_BUSEVENT_MSK
) {
210 /* this event uses the event bus */
211 b
= (ev
>> PM_BYTE_SH
) & PM_BYTE_MSK
;
212 u
= (ev
>> PM_UNIT_SH
) & PM_UNIT_MSK
;
213 /* check for conflict on this byte of event bus */
214 if ((ttmset
& (1 << b
)) && MMCR1_TTMSEL(mmcr1
, b
) != u
)
216 mmcr1
|= (unsigned long)u
<< MMCR1_TTMSEL_SH(b
);
219 /* Nest events have a further mux */
220 s
= (ev
>> PM_SUBUNIT_SH
) & PM_SUBUNIT_MSK
;
221 if ((ttmset
& 0x10) &&
222 MMCR1_NESTSEL(mmcr1
) != s
)
225 mmcr1
|= (unsigned long)s
<< MMCR1_NESTSEL_SH
;
227 if (0x30 <= psel
&& psel
<= 0x3d) {
228 /* these need the PMCx_ADDR_SEL bits */
230 mmcr1
|= MMCR1_PMC1_ADDR_SEL
>> pmc
;
232 /* bus select values are different for PMC3/4 */
233 if (pmc
>= 2 && (psel
& 0x90) == 0x80)
237 mmcr1
|= MMCR1_PMC1_LLA
>> pmc
;
239 mmcr1
|= MMCR1_PMC1_LLA_VALUE
>> pmc
;
241 if (power6_marked_instr_event(event
[i
]))
242 mmcra
|= MMCRA_SAMPLE_ENABLE
;
244 mmcr1
|= (unsigned long)psel
<< MMCR1_PMCSEL_SH(pmc
);
248 mmcr
[0] = MMCR0_PMC1CE
;
250 mmcr
[0] |= MMCR0_PMCjCE
;
257 * Layout of constraint bits:
259 * 0-1 add field: number of uses of PMC1 (max 1)
260 * 2-3, 4-5, 6-7, 8-9, 10-11: ditto for PMC2, 3, 4, 5, 6
261 * 12-15 add field: number of uses of PMC1-4 (max 4)
262 * 16-19 select field: unit on byte 0 of event bus
263 * 20-23, 24-27, 28-31 ditto for bytes 1, 2, 3
264 * 32-34 select field: nest (subunit) event selector
266 static int p6_get_constraint(u64 event
, unsigned long *maskp
,
269 int pmc
, byte
, sh
, subunit
;
270 unsigned long mask
= 0, value
= 0;
272 pmc
= (event
>> PM_PMC_SH
) & PM_PMC_MSK
;
274 if (pmc
> 4 && !(event
== 0x500009 || event
== 0x600005))
280 if (event
& PM_BUSEVENT_MSK
) {
281 byte
= (event
>> PM_BYTE_SH
) & PM_BYTE_MSK
;
282 sh
= byte
* 4 + (16 - PM_UNIT_SH
);
283 mask
|= PM_UNIT_MSKS
<< sh
;
284 value
|= (unsigned long)(event
& PM_UNIT_MSKS
) << sh
;
285 if ((event
& PM_UNIT_MSKS
) == (5 << PM_UNIT_SH
)) {
286 subunit
= (event
>> PM_SUBUNIT_SH
) & PM_SUBUNIT_MSK
;
287 mask
|= (unsigned long)PM_SUBUNIT_MSK
<< 32;
288 value
|= (unsigned long)subunit
<< 32;
292 mask
|= 0x8000; /* add field for count of PMC1-4 uses */
300 static int p6_limited_pmc_event(u64 event
)
302 int pmc
= (event
>> PM_PMC_SH
) & PM_PMC_MSK
;
304 return pmc
== 5 || pmc
== 6;
307 #define MAX_ALT 4 /* at most 4 alternatives for any event */
309 static const unsigned int event_alternatives
[][MAX_ALT
] = {
310 { 0x0130e8, 0x2000f6, 0x3000fc }, /* PM_PTEG_RELOAD_VALID */
311 { 0x080080, 0x10000d, 0x30000c, 0x4000f0 }, /* PM_LD_MISS_L1 */
312 { 0x080088, 0x200054, 0x3000f0 }, /* PM_ST_MISS_L1 */
313 { 0x10000a, 0x2000f4, 0x600005 }, /* PM_RUN_CYC */
314 { 0x10000b, 0x2000f5 }, /* PM_RUN_COUNT */
315 { 0x10000e, 0x400010 }, /* PM_PURR */
316 { 0x100010, 0x4000f8 }, /* PM_FLUSH */
317 { 0x10001a, 0x200010 }, /* PM_MRK_INST_DISP */
318 { 0x100026, 0x3000f8 }, /* PM_TB_BIT_TRANS */
319 { 0x100054, 0x2000f0 }, /* PM_ST_FIN */
320 { 0x100056, 0x2000fc }, /* PM_L1_ICACHE_MISS */
321 { 0x1000f0, 0x40000a }, /* PM_INST_IMC_MATCH_CMPL */
322 { 0x1000f8, 0x200008 }, /* PM_GCT_EMPTY_CYC */
323 { 0x1000fc, 0x400006 }, /* PM_LSU_DERAT_MISS_CYC */
324 { 0x20000e, 0x400007 }, /* PM_LSU_DERAT_MISS */
325 { 0x200012, 0x300012 }, /* PM_INST_DISP */
326 { 0x2000f2, 0x3000f2 }, /* PM_INST_DISP */
327 { 0x2000f8, 0x300010 }, /* PM_EXT_INT */
328 { 0x2000fe, 0x300056 }, /* PM_DATA_FROM_L2MISS */
329 { 0x2d0030, 0x30001a }, /* PM_MRK_FPU_FIN */
330 { 0x30000a, 0x400018 }, /* PM_MRK_INST_FIN */
331 { 0x3000f6, 0x40000e }, /* PM_L1_DCACHE_RELOAD_VALID */
332 { 0x3000fe, 0x400056 }, /* PM_DATA_FROM_L3MISS */
336 * This could be made more efficient with a binary search on
337 * a presorted list, if necessary
339 static int find_alternatives_list(u64 event
)
344 for (i
= 0; i
< ARRAY_SIZE(event_alternatives
); ++i
) {
345 if (event
< event_alternatives
[i
][0])
347 for (j
= 0; j
< MAX_ALT
; ++j
) {
348 alt
= event_alternatives
[i
][j
];
349 if (!alt
|| event
< alt
)
358 static int p6_get_alternatives(u64 event
, unsigned int flags
, u64 alt
[])
361 unsigned int psel
, pmc
;
362 unsigned int nalt
= 1;
366 nlim
= p6_limited_pmc_event(event
);
368 /* check the alternatives table */
369 i
= find_alternatives_list(event
);
371 /* copy out alternatives from list */
372 for (j
= 0; j
< MAX_ALT
; ++j
) {
373 aevent
= event_alternatives
[i
][j
];
377 alt
[nalt
++] = aevent
;
378 nlim
+= p6_limited_pmc_event(aevent
);
382 /* Check for alternative ways of computing sum events */
383 /* PMCSEL 0x32 counter N == PMCSEL 0x34 counter 5-N */
384 psel
= event
& (PM_PMCSEL_MSK
& ~1); /* ignore edge bit */
385 pmc
= (event
>> PM_PMC_SH
) & PM_PMC_MSK
;
386 if (pmc
&& (psel
== 0x32 || psel
== 0x34))
387 alt
[nalt
++] = ((event
^ 0x6) & ~PM_PMC_MSKS
) |
388 ((5 - pmc
) << PM_PMC_SH
);
390 /* PMCSEL 0x38 counter N == PMCSEL 0x3a counter N+/-2 */
391 if (pmc
&& (psel
== 0x38 || psel
== 0x3a))
392 alt
[nalt
++] = ((event
^ 0x2) & ~PM_PMC_MSKS
) |
393 ((pmc
> 2? pmc
- 2: pmc
+ 2) << PM_PMC_SH
);
396 if (flags
& PPMU_ONLY_COUNT_RUN
) {
398 * We're only counting in RUN state,
399 * so PM_CYC is equivalent to PM_RUN_CYC,
400 * PM_INST_CMPL === PM_RUN_INST_CMPL, PM_PURR === PM_RUN_PURR.
401 * This doesn't include alternatives that don't provide
402 * any extra flexibility in assigning PMCs (e.g.
403 * 0x10000a for PM_RUN_CYC vs. 0x1e for PM_CYC).
404 * Note that even with these additional alternatives
405 * we never end up with more than 4 alternatives for any event.
408 for (i
= 0; i
< nalt
; ++i
) {
410 case 0x1e: /* PM_CYC */
411 alt
[j
++] = 0x600005; /* PM_RUN_CYC */
414 case 0x10000a: /* PM_RUN_CYC */
415 alt
[j
++] = 0x1e; /* PM_CYC */
417 case 2: /* PM_INST_CMPL */
418 alt
[j
++] = 0x500009; /* PM_RUN_INST_CMPL */
421 case 0x500009: /* PM_RUN_INST_CMPL */
422 alt
[j
++] = 2; /* PM_INST_CMPL */
424 case 0x10000e: /* PM_PURR */
425 alt
[j
++] = 0x4000f4; /* PM_RUN_PURR */
427 case 0x4000f4: /* PM_RUN_PURR */
428 alt
[j
++] = 0x10000e; /* PM_PURR */
435 if (!(flags
& PPMU_LIMITED_PMC_OK
) && nlim
) {
436 /* remove the limited PMC events */
438 for (i
= 0; i
< nalt
; ++i
) {
439 if (!p6_limited_pmc_event(alt
[i
])) {
445 } else if ((flags
& PPMU_LIMITED_PMC_REQD
) && nlim
< nalt
) {
446 /* remove all but the limited PMC events */
448 for (i
= 0; i
< nalt
; ++i
) {
449 if (p6_limited_pmc_event(alt
[i
])) {
460 static void p6_disable_pmc(unsigned int pmc
, unsigned long mmcr
[])
462 /* Set PMCxSEL to 0 to disable PMCx */
464 mmcr
[1] &= ~(0xffUL
<< MMCR1_PMCSEL_SH(pmc
));
467 static int power6_generic_events
[] = {
468 [PERF_COUNT_HW_CPU_CYCLES
] = 0x1e,
469 [PERF_COUNT_HW_INSTRUCTIONS
] = 2,
470 [PERF_COUNT_HW_CACHE_REFERENCES
] = 0x280030, /* LD_REF_L1 */
471 [PERF_COUNT_HW_CACHE_MISSES
] = 0x30000c, /* LD_MISS_L1 */
472 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS
] = 0x410a0, /* BR_PRED */
473 [PERF_COUNT_HW_BRANCH_MISSES
] = 0x400052, /* BR_MPRED */
476 #define C(x) PERF_COUNT_HW_CACHE_##x
479 * Table of generalized cache-related events.
480 * 0 means not supported, -1 means nonsensical, other values
482 * The "DTLB" and "ITLB" events relate to the DERAT and IERAT.
484 static int power6_cache_events
[C(MAX
)][C(OP_MAX
)][C(RESULT_MAX
)] = {
485 [C(L1D
)] = { /* RESULT_ACCESS RESULT_MISS */
486 [C(OP_READ
)] = { 0x280030, 0x80080 },
487 [C(OP_WRITE
)] = { 0x180032, 0x80088 },
488 [C(OP_PREFETCH
)] = { 0x810a4, 0 },
490 [C(L1I
)] = { /* RESULT_ACCESS RESULT_MISS */
491 [C(OP_READ
)] = { 0, 0x100056 },
492 [C(OP_WRITE
)] = { -1, -1 },
493 [C(OP_PREFETCH
)] = { 0x4008c, 0 },
495 [C(LL
)] = { /* RESULT_ACCESS RESULT_MISS */
496 [C(OP_READ
)] = { 0x150730, 0x250532 },
497 [C(OP_WRITE
)] = { 0x250432, 0x150432 },
498 [C(OP_PREFETCH
)] = { 0x810a6, 0 },
500 [C(DTLB
)] = { /* RESULT_ACCESS RESULT_MISS */
501 [C(OP_READ
)] = { 0, 0x20000e },
502 [C(OP_WRITE
)] = { -1, -1 },
503 [C(OP_PREFETCH
)] = { -1, -1 },
505 [C(ITLB
)] = { /* RESULT_ACCESS RESULT_MISS */
506 [C(OP_READ
)] = { 0, 0x420ce },
507 [C(OP_WRITE
)] = { -1, -1 },
508 [C(OP_PREFETCH
)] = { -1, -1 },
510 [C(BPU
)] = { /* RESULT_ACCESS RESULT_MISS */
511 [C(OP_READ
)] = { 0x430e6, 0x400052 },
512 [C(OP_WRITE
)] = { -1, -1 },
513 [C(OP_PREFETCH
)] = { -1, -1 },
515 [C(NODE
)] = { /* RESULT_ACCESS RESULT_MISS */
516 [C(OP_READ
)] = { -1, -1 },
517 [C(OP_WRITE
)] = { -1, -1 },
518 [C(OP_PREFETCH
)] = { -1, -1 },
522 static struct power_pmu power6_pmu
= {
525 .max_alternatives
= MAX_ALT
,
526 .add_fields
= 0x1555,
527 .test_adder
= 0x3000,
528 .compute_mmcr
= p6_compute_mmcr
,
529 .get_constraint
= p6_get_constraint
,
530 .get_alternatives
= p6_get_alternatives
,
531 .disable_pmc
= p6_disable_pmc
,
532 .limited_pmc_event
= p6_limited_pmc_event
,
533 .flags
= PPMU_LIMITED_PMC5_6
| PPMU_ALT_SIPR
,
534 .n_generic
= ARRAY_SIZE(power6_generic_events
),
535 .generic_events
= power6_generic_events
,
536 .cache_events
= &power6_cache_events
,
539 int init_power6_pmu(void)
541 if (!cur_cpu_spec
->oprofile_cpu_type
||
542 strcmp(cur_cpu_spec
->oprofile_cpu_type
, "ppc64/power6"))
545 return register_power_pmu(&power6_pmu
);