WIP FPC-III support
[linux/fpc-iii.git] / arch / powerpc / perf / power6-pmu.c
blob189974478e9fcce8d59d7d9c89aebd1c039028a3
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * Performance counter support for POWER6 processors.
5 * Copyright 2008-2009 Paul Mackerras, IBM Corporation.
6 */
7 #include <linux/kernel.h>
8 #include <linux/perf_event.h>
9 #include <linux/string.h>
10 #include <asm/reg.h>
11 #include <asm/cputable.h>
13 #include "internal.h"
16 * Bits in event code for POWER6
18 #define PM_PMC_SH 20 /* PMC number (1-based) for direct events */
19 #define PM_PMC_MSK 0x7
20 #define PM_PMC_MSKS (PM_PMC_MSK << PM_PMC_SH)
21 #define PM_UNIT_SH 16 /* Unit event comes (TTMxSEL encoding) */
22 #define PM_UNIT_MSK 0xf
23 #define PM_UNIT_MSKS (PM_UNIT_MSK << PM_UNIT_SH)
24 #define PM_LLAV 0x8000 /* Load lookahead match value */
25 #define PM_LLA 0x4000 /* Load lookahead match enable */
26 #define PM_BYTE_SH 12 /* Byte of event bus to use */
27 #define PM_BYTE_MSK 3
28 #define PM_SUBUNIT_SH 8 /* Subunit event comes from (NEST_SEL enc.) */
29 #define PM_SUBUNIT_MSK 7
30 #define PM_SUBUNIT_MSKS (PM_SUBUNIT_MSK << PM_SUBUNIT_SH)
31 #define PM_PMCSEL_MSK 0xff /* PMCxSEL value */
32 #define PM_BUSEVENT_MSK 0xf3700
35 * Bits in MMCR1 for POWER6
37 #define MMCR1_TTM0SEL_SH 60
38 #define MMCR1_TTMSEL_SH(n) (MMCR1_TTM0SEL_SH - (n) * 4)
39 #define MMCR1_TTMSEL_MSK 0xf
40 #define MMCR1_TTMSEL(m, n) (((m) >> MMCR1_TTMSEL_SH(n)) & MMCR1_TTMSEL_MSK)
41 #define MMCR1_NESTSEL_SH 45
42 #define MMCR1_NESTSEL_MSK 0x7
43 #define MMCR1_NESTSEL(m) (((m) >> MMCR1_NESTSEL_SH) & MMCR1_NESTSEL_MSK)
44 #define MMCR1_PMC1_LLA (1ul << 44)
45 #define MMCR1_PMC1_LLA_VALUE (1ul << 39)
46 #define MMCR1_PMC1_ADDR_SEL (1ul << 35)
47 #define MMCR1_PMC1SEL_SH 24
48 #define MMCR1_PMCSEL_SH(n) (MMCR1_PMC1SEL_SH - (n) * 8)
49 #define MMCR1_PMCSEL_MSK 0xff
52 * Map of which direct events on which PMCs are marked instruction events.
53 * Indexed by PMCSEL value >> 1.
54 * Bottom 4 bits are a map of which PMCs are interesting,
55 * top 4 bits say what sort of event:
56 * 0 = direct marked event,
57 * 1 = byte decode event,
58 * 4 = add/and event (PMC1 -> bits 0 & 4),
59 * 5 = add/and event (PMC1 -> bits 1 & 5),
60 * 6 = add/and event (PMC1 -> bits 2 & 6),
61 * 7 = add/and event (PMC1 -> bits 3 & 7).
63 static unsigned char direct_event_is_marked[0x60 >> 1] = {
64 0, /* 00 */
65 0, /* 02 */
66 0, /* 04 */
67 0x07, /* 06 PM_MRK_ST_CMPL, PM_MRK_ST_GPS, PM_MRK_ST_CMPL_INT */
68 0x04, /* 08 PM_MRK_DFU_FIN */
69 0x06, /* 0a PM_MRK_IFU_FIN, PM_MRK_INST_FIN */
70 0, /* 0c */
71 0, /* 0e */
72 0x02, /* 10 PM_MRK_INST_DISP */
73 0x08, /* 12 PM_MRK_LSU_DERAT_MISS */
74 0, /* 14 */
75 0, /* 16 */
76 0x0c, /* 18 PM_THRESH_TIMEO, PM_MRK_INST_FIN */
77 0x0f, /* 1a PM_MRK_INST_DISP, PM_MRK_{FXU,FPU,LSU}_FIN */
78 0x01, /* 1c PM_MRK_INST_ISSUED */
79 0, /* 1e */
80 0, /* 20 */
81 0, /* 22 */
82 0, /* 24 */
83 0, /* 26 */
84 0x15, /* 28 PM_MRK_DATA_FROM_L2MISS, PM_MRK_DATA_FROM_L3MISS */
85 0, /* 2a */
86 0, /* 2c */
87 0, /* 2e */
88 0x4f, /* 30 */
89 0x7f, /* 32 */
90 0x4f, /* 34 */
91 0x5f, /* 36 */
92 0x6f, /* 38 */
93 0x4f, /* 3a */
94 0, /* 3c */
95 0x08, /* 3e PM_MRK_INST_TIMEO */
96 0x1f, /* 40 */
97 0x1f, /* 42 */
98 0x1f, /* 44 */
99 0x1f, /* 46 */
100 0x1f, /* 48 */
101 0x1f, /* 4a */
102 0x1f, /* 4c */
103 0x1f, /* 4e */
104 0, /* 50 */
105 0x05, /* 52 PM_MRK_BR_TAKEN, PM_MRK_BR_MPRED */
106 0x1c, /* 54 PM_MRK_PTEG_FROM_L3MISS, PM_MRK_PTEG_FROM_L2MISS */
107 0x02, /* 56 PM_MRK_LD_MISS_L1 */
108 0, /* 58 */
109 0, /* 5a */
110 0, /* 5c */
111 0, /* 5e */
115 * Masks showing for each unit which bits are marked events.
116 * These masks are in LE order, i.e. 0x00000001 is byte 0, bit 0.
118 static u32 marked_bus_events[16] = {
119 0x01000000, /* direct events set 1: byte 3 bit 0 */
120 0x00010000, /* direct events set 2: byte 2 bit 0 */
121 0, 0, 0, 0, /* IDU, IFU, nest: nothing */
122 0x00000088, /* VMX set 1: byte 0 bits 3, 7 */
123 0x000000c0, /* VMX set 2: byte 0 bits 4-7 */
124 0x04010000, /* LSU set 1: byte 2 bit 0, byte 3 bit 2 */
125 0xff010000u, /* LSU set 2: byte 2 bit 0, all of byte 3 */
126 0, /* LSU set 3 */
127 0x00000010, /* VMX set 3: byte 0 bit 4 */
128 0, /* BFP set 1 */
129 0x00000022, /* BFP set 2: byte 0 bits 1, 5 */
130 0, 0
134 * Returns 1 if event counts things relating to marked instructions
135 * and thus needs the MMCRA_SAMPLE_ENABLE bit set, or 0 if not.
137 static int power6_marked_instr_event(u64 event)
139 int pmc, psel, ptype;
140 int bit, byte, unit;
141 u32 mask;
143 pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
144 psel = (event & PM_PMCSEL_MSK) >> 1; /* drop edge/level bit */
145 if (pmc >= 5)
146 return 0;
148 bit = -1;
149 if (psel < sizeof(direct_event_is_marked)) {
150 ptype = direct_event_is_marked[psel];
151 if (pmc == 0 || !(ptype & (1 << (pmc - 1))))
152 return 0;
153 ptype >>= 4;
154 if (ptype == 0)
155 return 1;
156 if (ptype == 1)
157 bit = 0;
158 else
159 bit = ptype ^ (pmc - 1);
160 } else if ((psel & 0x48) == 0x40)
161 bit = psel & 7;
163 if (!(event & PM_BUSEVENT_MSK) || bit == -1)
164 return 0;
166 byte = (event >> PM_BYTE_SH) & PM_BYTE_MSK;
167 unit = (event >> PM_UNIT_SH) & PM_UNIT_MSK;
168 mask = marked_bus_events[unit];
169 return (mask >> (byte * 8 + bit)) & 1;
173 * Assign PMC numbers and compute MMCR1 value for a set of events
175 static int p6_compute_mmcr(u64 event[], int n_ev,
176 unsigned int hwc[], struct mmcr_regs *mmcr, struct perf_event *pevents[])
178 unsigned long mmcr1 = 0;
179 unsigned long mmcra = MMCRA_SDAR_DCACHE_MISS | MMCRA_SDAR_ERAT_MISS;
180 int i;
181 unsigned int pmc, ev, b, u, s, psel;
182 unsigned int ttmset = 0;
183 unsigned int pmc_inuse = 0;
185 if (n_ev > 6)
186 return -1;
187 for (i = 0; i < n_ev; ++i) {
188 pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK;
189 if (pmc) {
190 if (pmc_inuse & (1 << (pmc - 1)))
191 return -1; /* collision! */
192 pmc_inuse |= 1 << (pmc - 1);
195 for (i = 0; i < n_ev; ++i) {
196 ev = event[i];
197 pmc = (ev >> PM_PMC_SH) & PM_PMC_MSK;
198 if (pmc) {
199 --pmc;
200 } else {
201 /* can go on any PMC; find a free one */
202 for (pmc = 0; pmc < 4; ++pmc)
203 if (!(pmc_inuse & (1 << pmc)))
204 break;
205 if (pmc >= 4)
206 return -1;
207 pmc_inuse |= 1 << pmc;
209 hwc[i] = pmc;
210 psel = ev & PM_PMCSEL_MSK;
211 if (ev & PM_BUSEVENT_MSK) {
212 /* this event uses the event bus */
213 b = (ev >> PM_BYTE_SH) & PM_BYTE_MSK;
214 u = (ev >> PM_UNIT_SH) & PM_UNIT_MSK;
215 /* check for conflict on this byte of event bus */
216 if ((ttmset & (1 << b)) && MMCR1_TTMSEL(mmcr1, b) != u)
217 return -1;
218 mmcr1 |= (unsigned long)u << MMCR1_TTMSEL_SH(b);
219 ttmset |= 1 << b;
220 if (u == 5) {
221 /* Nest events have a further mux */
222 s = (ev >> PM_SUBUNIT_SH) & PM_SUBUNIT_MSK;
223 if ((ttmset & 0x10) &&
224 MMCR1_NESTSEL(mmcr1) != s)
225 return -1;
226 ttmset |= 0x10;
227 mmcr1 |= (unsigned long)s << MMCR1_NESTSEL_SH;
229 if (0x30 <= psel && psel <= 0x3d) {
230 /* these need the PMCx_ADDR_SEL bits */
231 if (b >= 2)
232 mmcr1 |= MMCR1_PMC1_ADDR_SEL >> pmc;
234 /* bus select values are different for PMC3/4 */
235 if (pmc >= 2 && (psel & 0x90) == 0x80)
236 psel ^= 0x20;
238 if (ev & PM_LLA) {
239 mmcr1 |= MMCR1_PMC1_LLA >> pmc;
240 if (ev & PM_LLAV)
241 mmcr1 |= MMCR1_PMC1_LLA_VALUE >> pmc;
243 if (power6_marked_instr_event(event[i]))
244 mmcra |= MMCRA_SAMPLE_ENABLE;
245 if (pmc < 4)
246 mmcr1 |= (unsigned long)psel << MMCR1_PMCSEL_SH(pmc);
248 mmcr->mmcr0 = 0;
249 if (pmc_inuse & 1)
250 mmcr->mmcr0 = MMCR0_PMC1CE;
251 if (pmc_inuse & 0xe)
252 mmcr->mmcr0 |= MMCR0_PMCjCE;
253 mmcr->mmcr1 = mmcr1;
254 mmcr->mmcra = mmcra;
255 return 0;
259 * Layout of constraint bits:
261 * 0-1 add field: number of uses of PMC1 (max 1)
262 * 2-3, 4-5, 6-7, 8-9, 10-11: ditto for PMC2, 3, 4, 5, 6
263 * 12-15 add field: number of uses of PMC1-4 (max 4)
264 * 16-19 select field: unit on byte 0 of event bus
265 * 20-23, 24-27, 28-31 ditto for bytes 1, 2, 3
266 * 32-34 select field: nest (subunit) event selector
268 static int p6_get_constraint(u64 event, unsigned long *maskp,
269 unsigned long *valp)
271 int pmc, byte, sh, subunit;
272 unsigned long mask = 0, value = 0;
274 pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
275 if (pmc) {
276 if (pmc > 4 && !(event == 0x500009 || event == 0x600005))
277 return -1;
278 sh = (pmc - 1) * 2;
279 mask |= 2 << sh;
280 value |= 1 << sh;
282 if (event & PM_BUSEVENT_MSK) {
283 byte = (event >> PM_BYTE_SH) & PM_BYTE_MSK;
284 sh = byte * 4 + (16 - PM_UNIT_SH);
285 mask |= PM_UNIT_MSKS << sh;
286 value |= (unsigned long)(event & PM_UNIT_MSKS) << sh;
287 if ((event & PM_UNIT_MSKS) == (5 << PM_UNIT_SH)) {
288 subunit = (event >> PM_SUBUNIT_SH) & PM_SUBUNIT_MSK;
289 mask |= (unsigned long)PM_SUBUNIT_MSK << 32;
290 value |= (unsigned long)subunit << 32;
293 if (pmc <= 4) {
294 mask |= 0x8000; /* add field for count of PMC1-4 uses */
295 value |= 0x1000;
297 *maskp = mask;
298 *valp = value;
299 return 0;
302 static int p6_limited_pmc_event(u64 event)
304 int pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
306 return pmc == 5 || pmc == 6;
309 #define MAX_ALT 4 /* at most 4 alternatives for any event */
311 static const unsigned int event_alternatives[][MAX_ALT] = {
312 { 0x0130e8, 0x2000f6, 0x3000fc }, /* PM_PTEG_RELOAD_VALID */
313 { 0x080080, 0x10000d, 0x30000c, 0x4000f0 }, /* PM_LD_MISS_L1 */
314 { 0x080088, 0x200054, 0x3000f0 }, /* PM_ST_MISS_L1 */
315 { 0x10000a, 0x2000f4, 0x600005 }, /* PM_RUN_CYC */
316 { 0x10000b, 0x2000f5 }, /* PM_RUN_COUNT */
317 { 0x10000e, 0x400010 }, /* PM_PURR */
318 { 0x100010, 0x4000f8 }, /* PM_FLUSH */
319 { 0x10001a, 0x200010 }, /* PM_MRK_INST_DISP */
320 { 0x100026, 0x3000f8 }, /* PM_TB_BIT_TRANS */
321 { 0x100054, 0x2000f0 }, /* PM_ST_FIN */
322 { 0x100056, 0x2000fc }, /* PM_L1_ICACHE_MISS */
323 { 0x1000f0, 0x40000a }, /* PM_INST_IMC_MATCH_CMPL */
324 { 0x1000f8, 0x200008 }, /* PM_GCT_EMPTY_CYC */
325 { 0x1000fc, 0x400006 }, /* PM_LSU_DERAT_MISS_CYC */
326 { 0x20000e, 0x400007 }, /* PM_LSU_DERAT_MISS */
327 { 0x200012, 0x300012 }, /* PM_INST_DISP */
328 { 0x2000f2, 0x3000f2 }, /* PM_INST_DISP */
329 { 0x2000f8, 0x300010 }, /* PM_EXT_INT */
330 { 0x2000fe, 0x300056 }, /* PM_DATA_FROM_L2MISS */
331 { 0x2d0030, 0x30001a }, /* PM_MRK_FPU_FIN */
332 { 0x30000a, 0x400018 }, /* PM_MRK_INST_FIN */
333 { 0x3000f6, 0x40000e }, /* PM_L1_DCACHE_RELOAD_VALID */
334 { 0x3000fe, 0x400056 }, /* PM_DATA_FROM_L3MISS */
338 * This could be made more efficient with a binary search on
339 * a presorted list, if necessary
341 static int find_alternatives_list(u64 event)
343 int i, j;
344 unsigned int alt;
346 for (i = 0; i < ARRAY_SIZE(event_alternatives); ++i) {
347 if (event < event_alternatives[i][0])
348 return -1;
349 for (j = 0; j < MAX_ALT; ++j) {
350 alt = event_alternatives[i][j];
351 if (!alt || event < alt)
352 break;
353 if (event == alt)
354 return i;
357 return -1;
360 static int p6_get_alternatives(u64 event, unsigned int flags, u64 alt[])
362 int i, j, nlim;
363 unsigned int psel, pmc;
364 unsigned int nalt = 1;
365 u64 aevent;
367 alt[0] = event;
368 nlim = p6_limited_pmc_event(event);
370 /* check the alternatives table */
371 i = find_alternatives_list(event);
372 if (i >= 0) {
373 /* copy out alternatives from list */
374 for (j = 0; j < MAX_ALT; ++j) {
375 aevent = event_alternatives[i][j];
376 if (!aevent)
377 break;
378 if (aevent != event)
379 alt[nalt++] = aevent;
380 nlim += p6_limited_pmc_event(aevent);
383 } else {
384 /* Check for alternative ways of computing sum events */
385 /* PMCSEL 0x32 counter N == PMCSEL 0x34 counter 5-N */
386 psel = event & (PM_PMCSEL_MSK & ~1); /* ignore edge bit */
387 pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
388 if (pmc && (psel == 0x32 || psel == 0x34))
389 alt[nalt++] = ((event ^ 0x6) & ~PM_PMC_MSKS) |
390 ((5 - pmc) << PM_PMC_SH);
392 /* PMCSEL 0x38 counter N == PMCSEL 0x3a counter N+/-2 */
393 if (pmc && (psel == 0x38 || psel == 0x3a))
394 alt[nalt++] = ((event ^ 0x2) & ~PM_PMC_MSKS) |
395 ((pmc > 2? pmc - 2: pmc + 2) << PM_PMC_SH);
398 if (flags & PPMU_ONLY_COUNT_RUN) {
400 * We're only counting in RUN state,
401 * so PM_CYC is equivalent to PM_RUN_CYC,
402 * PM_INST_CMPL === PM_RUN_INST_CMPL, PM_PURR === PM_RUN_PURR.
403 * This doesn't include alternatives that don't provide
404 * any extra flexibility in assigning PMCs (e.g.
405 * 0x10000a for PM_RUN_CYC vs. 0x1e for PM_CYC).
406 * Note that even with these additional alternatives
407 * we never end up with more than 4 alternatives for any event.
409 j = nalt;
410 for (i = 0; i < nalt; ++i) {
411 switch (alt[i]) {
412 case 0x1e: /* PM_CYC */
413 alt[j++] = 0x600005; /* PM_RUN_CYC */
414 ++nlim;
415 break;
416 case 0x10000a: /* PM_RUN_CYC */
417 alt[j++] = 0x1e; /* PM_CYC */
418 break;
419 case 2: /* PM_INST_CMPL */
420 alt[j++] = 0x500009; /* PM_RUN_INST_CMPL */
421 ++nlim;
422 break;
423 case 0x500009: /* PM_RUN_INST_CMPL */
424 alt[j++] = 2; /* PM_INST_CMPL */
425 break;
426 case 0x10000e: /* PM_PURR */
427 alt[j++] = 0x4000f4; /* PM_RUN_PURR */
428 break;
429 case 0x4000f4: /* PM_RUN_PURR */
430 alt[j++] = 0x10000e; /* PM_PURR */
431 break;
434 nalt = j;
437 if (!(flags & PPMU_LIMITED_PMC_OK) && nlim) {
438 /* remove the limited PMC events */
439 j = 0;
440 for (i = 0; i < nalt; ++i) {
441 if (!p6_limited_pmc_event(alt[i])) {
442 alt[j] = alt[i];
443 ++j;
446 nalt = j;
447 } else if ((flags & PPMU_LIMITED_PMC_REQD) && nlim < nalt) {
448 /* remove all but the limited PMC events */
449 j = 0;
450 for (i = 0; i < nalt; ++i) {
451 if (p6_limited_pmc_event(alt[i])) {
452 alt[j] = alt[i];
453 ++j;
456 nalt = j;
459 return nalt;
462 static void p6_disable_pmc(unsigned int pmc, struct mmcr_regs *mmcr)
464 /* Set PMCxSEL to 0 to disable PMCx */
465 if (pmc <= 3)
466 mmcr->mmcr1 &= ~(0xffUL << MMCR1_PMCSEL_SH(pmc));
469 static int power6_generic_events[] = {
470 [PERF_COUNT_HW_CPU_CYCLES] = 0x1e,
471 [PERF_COUNT_HW_INSTRUCTIONS] = 2,
472 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x280030, /* LD_REF_L1 */
473 [PERF_COUNT_HW_CACHE_MISSES] = 0x30000c, /* LD_MISS_L1 */
474 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x410a0, /* BR_PRED */
475 [PERF_COUNT_HW_BRANCH_MISSES] = 0x400052, /* BR_MPRED */
478 #define C(x) PERF_COUNT_HW_CACHE_##x
481 * Table of generalized cache-related events.
482 * 0 means not supported, -1 means nonsensical, other values
483 * are event codes.
484 * The "DTLB" and "ITLB" events relate to the DERAT and IERAT.
486 static u64 power6_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
487 [C(L1D)] = { /* RESULT_ACCESS RESULT_MISS */
488 [C(OP_READ)] = { 0x280030, 0x80080 },
489 [C(OP_WRITE)] = { 0x180032, 0x80088 },
490 [C(OP_PREFETCH)] = { 0x810a4, 0 },
492 [C(L1I)] = { /* RESULT_ACCESS RESULT_MISS */
493 [C(OP_READ)] = { 0, 0x100056 },
494 [C(OP_WRITE)] = { -1, -1 },
495 [C(OP_PREFETCH)] = { 0x4008c, 0 },
497 [C(LL)] = { /* RESULT_ACCESS RESULT_MISS */
498 [C(OP_READ)] = { 0x150730, 0x250532 },
499 [C(OP_WRITE)] = { 0x250432, 0x150432 },
500 [C(OP_PREFETCH)] = { 0x810a6, 0 },
502 [C(DTLB)] = { /* RESULT_ACCESS RESULT_MISS */
503 [C(OP_READ)] = { 0, 0x20000e },
504 [C(OP_WRITE)] = { -1, -1 },
505 [C(OP_PREFETCH)] = { -1, -1 },
507 [C(ITLB)] = { /* RESULT_ACCESS RESULT_MISS */
508 [C(OP_READ)] = { 0, 0x420ce },
509 [C(OP_WRITE)] = { -1, -1 },
510 [C(OP_PREFETCH)] = { -1, -1 },
512 [C(BPU)] = { /* RESULT_ACCESS RESULT_MISS */
513 [C(OP_READ)] = { 0x430e6, 0x400052 },
514 [C(OP_WRITE)] = { -1, -1 },
515 [C(OP_PREFETCH)] = { -1, -1 },
517 [C(NODE)] = { /* RESULT_ACCESS RESULT_MISS */
518 [C(OP_READ)] = { -1, -1 },
519 [C(OP_WRITE)] = { -1, -1 },
520 [C(OP_PREFETCH)] = { -1, -1 },
524 static struct power_pmu power6_pmu = {
525 .name = "POWER6",
526 .n_counter = 6,
527 .max_alternatives = MAX_ALT,
528 .add_fields = 0x1555,
529 .test_adder = 0x3000,
530 .compute_mmcr = p6_compute_mmcr,
531 .get_constraint = p6_get_constraint,
532 .get_alternatives = p6_get_alternatives,
533 .disable_pmc = p6_disable_pmc,
534 .limited_pmc_event = p6_limited_pmc_event,
535 .flags = PPMU_LIMITED_PMC5_6 | PPMU_ALT_SIPR,
536 .n_generic = ARRAY_SIZE(power6_generic_events),
537 .generic_events = power6_generic_events,
538 .cache_events = &power6_cache_events,
541 int init_power6_pmu(void)
543 if (!cur_cpu_spec->oprofile_cpu_type ||
544 strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power6"))
545 return -ENODEV;
547 return register_power_pmu(&power6_pmu);