1 /* Performance event support for sparc64. 2 * 3 * Copyright (C) 2009, 2010 David S. Miller <davem@davemloft.net> 4 * 5 * This code is based almost entirely upon the x86 perf event 6 * code, which is: 7 * 8 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de> 9 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar 10 * Copyright (C) 2009 Jaswinder Singh Rajput 11 * Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter 12 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> 13 */ 14 15 #include <linux/perf_event.h> 16 #include <linux/kprobes.h> 17 #include <linux/ftrace.h> 18 #include <linux/kernel.h> 19 #include <linux/kdebug.h> 20 #include <linux/mutex.h> 21 22 #include <asm/stacktrace.h> 23 #include <asm/cpudata.h> 24 #include <asm/uaccess.h> 25 #include <asm/atomic.h> 26 #include <asm/nmi.h> 27 #include <asm/pcr.h> 28 29 #include "kstack.h" 30 31 /* Sparc64 chips have two performance counters, 32-bits each, with 32 * overflow interrupts generated on transition from 0xffffffff to 0. 33 * The counters are accessed in one go using a 64-bit register. 34 * 35 * Both counters are controlled using a single control register. The 36 * only way to stop all sampling is to clear all of the context (user, 37 * supervisor, hypervisor) sampling enable bits. But these bits apply 38 * to both counters, thus the two counters can't be enabled/disabled 39 * individually. 40 * 41 * The control register has two event fields, one for each of the two 42 * counters. It's thus nearly impossible to have one counter going 43 * while keeping the other one stopped. Therefore it is possible to 44 * get overflow interrupts for counters not currently "in use" and 45 * that condition must be checked in the overflow interrupt handler. 46 * 47 * So we use a hack, in that we program inactive counters with the 48 * "sw_count0" and "sw_count1" events. These count how many times 49 * the instruction "sethi %hi(0xfc000), %g0" is executed. It's an 50 * unusual way to encode a NOP and therefore will not trigger in 51 * normal code. 52 */ 53 54 #define MAX_HWEVENTS 2 55 #define MAX_PERIOD ((1UL << 32) - 1) 56 57 #define PIC_UPPER_INDEX 0 58 #define PIC_LOWER_INDEX 1 59 #define PIC_NO_INDEX -1 60 61 struct cpu_hw_events { 62 /* Number of events currently scheduled onto this cpu. 63 * This tells how many entries in the arrays below 64 * are valid. 65 */ 66 int n_events; 67 68 /* Number of new events added since the last hw_perf_disable(). 69 * This works because the perf event layer always adds new 70 * events inside of a perf_{disable,enable}() sequence. 71 */ 72 int n_added; 73 74 /* Array of events current scheduled on this cpu. */ 75 struct perf_event *event[MAX_HWEVENTS]; 76 77 /* Array of encoded longs, specifying the %pcr register 78 * encoding and the mask of PIC counters this even can 79 * be scheduled on. See perf_event_encode() et al. 80 */ 81 unsigned long events[MAX_HWEVENTS]; 82 83 /* The current counter index assigned to an event. When the 84 * event hasn't been programmed into the cpu yet, this will 85 * hold PIC_NO_INDEX. The event->hw.idx value tells us where 86 * we ought to schedule the event. 87 */ 88 int current_idx[MAX_HWEVENTS]; 89 90 /* Software copy of %pcr register on this cpu. */ 91 u64 pcr; 92 93 /* Enabled/disable state. */ 94 int enabled; 95 96 unsigned int group_flag; 97 }; 98 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = { .enabled = 1, }; 99 100 /* An event map describes the characteristics of a performance 101 * counter event. In particular it gives the encoding as well as 102 * a mask telling which counters the event can be measured on. 103 */ 104 struct perf_event_map { 105 u16 encoding; 106 u8 pic_mask; 107 #define PIC_NONE 0x00 108 #define PIC_UPPER 0x01 109 #define PIC_LOWER 0x02 110 }; 111 112 /* Encode a perf_event_map entry into a long. */ 113 static unsigned long perf_event_encode(const struct perf_event_map *pmap) 114 { 115 return ((unsigned long) pmap->encoding << 16) | pmap->pic_mask; 116 } 117 118 static u8 perf_event_get_msk(unsigned long val) 119 { 120 return val & 0xff; 121 } 122 123 static u64 perf_event_get_enc(unsigned long val) 124 { 125 return val >> 16; 126 } 127 128 #define C(x) PERF_COUNT_HW_CACHE_##x 129 130 #define CACHE_OP_UNSUPPORTED 0xfffe 131 #define CACHE_OP_NONSENSE 0xffff 132 133 typedef struct perf_event_map cache_map_t 134 [PERF_COUNT_HW_CACHE_MAX] 135 [PERF_COUNT_HW_CACHE_OP_MAX] 136 [PERF_COUNT_HW_CACHE_RESULT_MAX]; 137 138 struct sparc_pmu { 139 const struct perf_event_map *(*event_map)(int); 140 const cache_map_t *cache_map; 141 int max_events; 142 int upper_shift; 143 int lower_shift; 144 int event_mask; 145 int hv_bit; 146 int irq_bit; 147 int upper_nop; 148 int lower_nop; 149 }; 150 151 static const struct perf_event_map ultra3_perfmon_event_map[] = { 152 [PERF_COUNT_HW_CPU_CYCLES] = { 0x0000, PIC_UPPER | PIC_LOWER }, 153 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x0001, PIC_UPPER | PIC_LOWER }, 154 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0009, PIC_LOWER }, 155 [PERF_COUNT_HW_CACHE_MISSES] = { 0x0009, PIC_UPPER }, 156 }; 157 158 static const struct perf_event_map *ultra3_event_map(int event_id) 159 { 160 return &ultra3_perfmon_event_map[event_id]; 161 } 162 163 static const cache_map_t ultra3_cache_map = { 164 [C(L1D)] = { 165 [C(OP_READ)] = { 166 [C(RESULT_ACCESS)] = { 0x09, PIC_LOWER, }, 167 [C(RESULT_MISS)] = { 0x09, PIC_UPPER, }, 168 }, 169 [C(OP_WRITE)] = { 170 [C(RESULT_ACCESS)] = { 0x0a, PIC_LOWER }, 171 [C(RESULT_MISS)] = { 0x0a, PIC_UPPER }, 172 }, 173 [C(OP_PREFETCH)] = { 174 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 175 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED }, 176 }, 177 }, 178 [C(L1I)] = { 179 [C(OP_READ)] = { 180 [C(RESULT_ACCESS)] = { 0x09, PIC_LOWER, }, 181 [C(RESULT_MISS)] = { 0x09, PIC_UPPER, }, 182 }, 183 [ C(OP_WRITE) ] = { 184 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE }, 185 [ C(RESULT_MISS) ] = { CACHE_OP_NONSENSE }, 186 }, 187 [ C(OP_PREFETCH) ] = { 188 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 189 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 190 }, 191 }, 192 [C(LL)] = { 193 [C(OP_READ)] = { 194 [C(RESULT_ACCESS)] = { 0x0c, PIC_LOWER, }, 195 [C(RESULT_MISS)] = { 0x0c, PIC_UPPER, }, 196 }, 197 [C(OP_WRITE)] = { 198 [C(RESULT_ACCESS)] = { 0x0c, PIC_LOWER }, 199 [C(RESULT_MISS)] = { 0x0c, PIC_UPPER }, 200 }, 201 [C(OP_PREFETCH)] = { 202 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 203 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED }, 204 }, 205 }, 206 [C(DTLB)] = { 207 [C(OP_READ)] = { 208 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 209 [C(RESULT_MISS)] = { 0x12, PIC_UPPER, }, 210 }, 211 [ C(OP_WRITE) ] = { 212 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 213 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 214 }, 215 [ C(OP_PREFETCH) ] = { 216 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 217 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 218 }, 219 }, 220 [C(ITLB)] = { 221 [C(OP_READ)] = { 222 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 223 [C(RESULT_MISS)] = { 0x11, PIC_UPPER, }, 224 }, 225 [ C(OP_WRITE) ] = { 226 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 227 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 228 }, 229 [ C(OP_PREFETCH) ] = { 230 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 231 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 232 }, 233 }, 234 [C(BPU)] = { 235 [C(OP_READ)] = { 236 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 237 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED }, 238 }, 239 [ C(OP_WRITE) ] = { 240 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 241 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 242 }, 243 [ C(OP_PREFETCH) ] = { 244 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 245 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 246 }, 247 }, 248 }; 249 250 static const struct sparc_pmu ultra3_pmu = { 251 .event_map = ultra3_event_map, 252 .cache_map = &ultra3_cache_map, 253 .max_events = ARRAY_SIZE(ultra3_perfmon_event_map), 254 .upper_shift = 11, 255 .lower_shift = 4, 256 .event_mask = 0x3f, 257 .upper_nop = 0x1c, 258 .lower_nop = 0x14, 259 }; 260 261 /* Niagara1 is very limited. The upper PIC is hard-locked to count 262 * only instructions, so it is free running which creates all kinds of 263 * problems. Some hardware designs make one wonder if the creator 264 * even looked at how this stuff gets used by software. 265 */ 266 static const struct perf_event_map niagara1_perfmon_event_map[] = { 267 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, PIC_UPPER }, 268 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x00, PIC_UPPER }, 269 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0, PIC_NONE }, 270 [PERF_COUNT_HW_CACHE_MISSES] = { 0x03, PIC_LOWER }, 271 }; 272 273 static const struct perf_event_map *niagara1_event_map(int event_id) 274 { 275 return &niagara1_perfmon_event_map[event_id]; 276 } 277 278 static const cache_map_t niagara1_cache_map = { 279 [C(L1D)] = { 280 [C(OP_READ)] = { 281 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 282 [C(RESULT_MISS)] = { 0x03, PIC_LOWER, }, 283 }, 284 [C(OP_WRITE)] = { 285 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 286 [C(RESULT_MISS)] = { 0x03, PIC_LOWER, }, 287 }, 288 [C(OP_PREFETCH)] = { 289 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 290 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED }, 291 }, 292 }, 293 [C(L1I)] = { 294 [C(OP_READ)] = { 295 [C(RESULT_ACCESS)] = { 0x00, PIC_UPPER }, 296 [C(RESULT_MISS)] = { 0x02, PIC_LOWER, }, 297 }, 298 [ C(OP_WRITE) ] = { 299 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE }, 300 [ C(RESULT_MISS) ] = { CACHE_OP_NONSENSE }, 301 }, 302 [ C(OP_PREFETCH) ] = { 303 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 304 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 305 }, 306 }, 307 [C(LL)] = { 308 [C(OP_READ)] = { 309 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 310 [C(RESULT_MISS)] = { 0x07, PIC_LOWER, }, 311 }, 312 [C(OP_WRITE)] = { 313 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 314 [C(RESULT_MISS)] = { 0x07, PIC_LOWER, }, 315 }, 316 [C(OP_PREFETCH)] = { 317 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 318 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED }, 319 }, 320 }, 321 [C(DTLB)] = { 322 [C(OP_READ)] = { 323 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 324 [C(RESULT_MISS)] = { 0x05, PIC_LOWER, }, 325 }, 326 [ C(OP_WRITE) ] = { 327 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 328 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 329 }, 330 [ C(OP_PREFETCH) ] = { 331 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 332 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 333 }, 334 }, 335 [C(ITLB)] = { 336 [C(OP_READ)] = { 337 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 338 [C(RESULT_MISS)] = { 0x04, PIC_LOWER, }, 339 }, 340 [ C(OP_WRITE) ] = { 341 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 342 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 343 }, 344 [ C(OP_PREFETCH) ] = { 345 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 346 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 347 }, 348 }, 349 [C(BPU)] = { 350 [C(OP_READ)] = { 351 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 352 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED }, 353 }, 354 [ C(OP_WRITE) ] = { 355 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 356 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 357 }, 358 [ C(OP_PREFETCH) ] = { 359 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 360 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 361 }, 362 }, 363 }; 364 365 static const struct sparc_pmu niagara1_pmu = { 366 .event_map = niagara1_event_map, 367 .cache_map = &niagara1_cache_map, 368 .max_events = ARRAY_SIZE(niagara1_perfmon_event_map), 369 .upper_shift = 0, 370 .lower_shift = 4, 371 .event_mask = 0x7, 372 .upper_nop = 0x0, 373 .lower_nop = 0x0, 374 }; 375 376 static const struct perf_event_map niagara2_perfmon_event_map[] = { 377 [PERF_COUNT_HW_CPU_CYCLES] = { 0x02ff, PIC_UPPER | PIC_LOWER }, 378 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x02ff, PIC_UPPER | PIC_LOWER }, 379 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0208, PIC_UPPER | PIC_LOWER }, 380 [PERF_COUNT_HW_CACHE_MISSES] = { 0x0302, PIC_UPPER | PIC_LOWER }, 381 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x0201, PIC_UPPER | PIC_LOWER }, 382 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x0202, PIC_UPPER | PIC_LOWER }, 383 }; 384 385 static const struct perf_event_map *niagara2_event_map(int event_id) 386 { 387 return &niagara2_perfmon_event_map[event_id]; 388 } 389 390 static const cache_map_t niagara2_cache_map = { 391 [C(L1D)] = { 392 [C(OP_READ)] = { 393 [C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, }, 394 [C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, }, 395 }, 396 [C(OP_WRITE)] = { 397 [C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, }, 398 [C(RESULT_MISS)] = { 0x0302, PIC_UPPER | PIC_LOWER, }, 399 }, 400 [C(OP_PREFETCH)] = { 401 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 402 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED }, 403 }, 404 }, 405 [C(L1I)] = { 406 [C(OP_READ)] = { 407 [C(RESULT_ACCESS)] = { 0x02ff, PIC_UPPER | PIC_LOWER, }, 408 [C(RESULT_MISS)] = { 0x0301, PIC_UPPER | PIC_LOWER, }, 409 }, 410 [ C(OP_WRITE) ] = { 411 [ C(RESULT_ACCESS) ] = { CACHE_OP_NONSENSE }, 412 [ C(RESULT_MISS) ] = { CACHE_OP_NONSENSE }, 413 }, 414 [ C(OP_PREFETCH) ] = { 415 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 416 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 417 }, 418 }, 419 [C(LL)] = { 420 [C(OP_READ)] = { 421 [C(RESULT_ACCESS)] = { 0x0208, PIC_UPPER | PIC_LOWER, }, 422 [C(RESULT_MISS)] = { 0x0330, PIC_UPPER | PIC_LOWER, }, 423 }, 424 [C(OP_WRITE)] = { 425 [C(RESULT_ACCESS)] = { 0x0210, PIC_UPPER | PIC_LOWER, }, 426 [C(RESULT_MISS)] = { 0x0320, PIC_UPPER | PIC_LOWER, }, 427 }, 428 [C(OP_PREFETCH)] = { 429 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 430 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED }, 431 }, 432 }, 433 [C(DTLB)] = { 434 [C(OP_READ)] = { 435 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 436 [C(RESULT_MISS)] = { 0x0b08, PIC_UPPER | PIC_LOWER, }, 437 }, 438 [ C(OP_WRITE) ] = { 439 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 440 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 441 }, 442 [ C(OP_PREFETCH) ] = { 443 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 444 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 445 }, 446 }, 447 [C(ITLB)] = { 448 [C(OP_READ)] = { 449 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 450 [C(RESULT_MISS)] = { 0xb04, PIC_UPPER | PIC_LOWER, }, 451 }, 452 [ C(OP_WRITE) ] = { 453 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 454 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 455 }, 456 [ C(OP_PREFETCH) ] = { 457 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 458 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 459 }, 460 }, 461 [C(BPU)] = { 462 [C(OP_READ)] = { 463 [C(RESULT_ACCESS)] = { CACHE_OP_UNSUPPORTED }, 464 [C(RESULT_MISS)] = { CACHE_OP_UNSUPPORTED }, 465 }, 466 [ C(OP_WRITE) ] = { 467 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 468 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 469 }, 470 [ C(OP_PREFETCH) ] = { 471 [ C(RESULT_ACCESS) ] = { CACHE_OP_UNSUPPORTED }, 472 [ C(RESULT_MISS) ] = { CACHE_OP_UNSUPPORTED }, 473 }, 474 }, 475 }; 476 477 static const struct sparc_pmu niagara2_pmu = { 478 .event_map = niagara2_event_map, 479 .cache_map = &niagara2_cache_map, 480 .max_events = ARRAY_SIZE(niagara2_perfmon_event_map), 481 .upper_shift = 19, 482 .lower_shift = 6, 483 .event_mask = 0xfff, 484 .hv_bit = 0x8, 485 .irq_bit = 0x30, 486 .upper_nop = 0x220, 487 .lower_nop = 0x220, 488 }; 489 490 static const struct sparc_pmu *sparc_pmu __read_mostly; 491 492 static u64 event_encoding(u64 event_id, int idx) 493 { 494 if (idx == PIC_UPPER_INDEX) 495 event_id <<= sparc_pmu->upper_shift; 496 else 497 event_id <<= sparc_pmu->lower_shift; 498 return event_id; 499 } 500 501 static u64 mask_for_index(int idx) 502 { 503 return event_encoding(sparc_pmu->event_mask, idx); 504 } 505 506 static u64 nop_for_index(int idx) 507 { 508 return event_encoding(idx == PIC_UPPER_INDEX ? 509 sparc_pmu->upper_nop : 510 sparc_pmu->lower_nop, idx); 511 } 512 513 static inline void sparc_pmu_enable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx) 514 { 515 u64 val, mask = mask_for_index(idx); 516 517 val = cpuc->pcr; 518 val &= ~mask; 519 val |= hwc->config; 520 cpuc->pcr = val; 521 522 pcr_ops->write(cpuc->pcr); 523 } 524 525 static inline void sparc_pmu_disable_event(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc, int idx) 526 { 527 u64 mask = mask_for_index(idx); 528 u64 nop = nop_for_index(idx); 529 u64 val; 530 531 val = cpuc->pcr; 532 val &= ~mask; 533 val |= nop; 534 cpuc->pcr = val; 535 536 pcr_ops->write(cpuc->pcr); 537 } 538 539 static u32 read_pmc(int idx) 540 { 541 u64 val; 542 543 read_pic(val); 544 if (idx == PIC_UPPER_INDEX) 545 val >>= 32; 546 547 return val & 0xffffffff; 548 } 549 550 static void write_pmc(int idx, u64 val) 551 { 552 u64 shift, mask, pic; 553 554 shift = 0; 555 if (idx == PIC_UPPER_INDEX) 556 shift = 32; 557 558 mask = ((u64) 0xffffffff) << shift; 559 val <<= shift; 560 561 read_pic(pic); 562 pic &= ~mask; 563 pic |= val; 564 write_pic(pic); 565 } 566 567 static u64 sparc_perf_event_update(struct perf_event *event, 568 struct hw_perf_event *hwc, int idx) 569 { 570 int shift = 64 - 32; 571 u64 prev_raw_count, new_raw_count; 572 s64 delta; 573 574 again: 575 prev_raw_count = local64_read(&hwc->prev_count); 576 new_raw_count = read_pmc(idx); 577 578 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count, 579 new_raw_count) != prev_raw_count) 580 goto again; 581 582 delta = (new_raw_count << shift) - (prev_raw_count << shift); 583 delta >>= shift; 584 585 local64_add(delta, &event->count); 586 local64_sub(delta, &hwc->period_left); 587 588 return new_raw_count; 589 } 590 591 static int sparc_perf_event_set_period(struct perf_event *event, 592 struct hw_perf_event *hwc, int idx) 593 { 594 s64 left = local64_read(&hwc->period_left); 595 s64 period = hwc->sample_period; 596 int ret = 0; 597 598 if (unlikely(left <= -period)) { 599 left = period; 600 local64_set(&hwc->period_left, left); 601 hwc->last_period = period; 602 ret = 1; 603 } 604 605 if (unlikely(left <= 0)) { 606 left += period; 607 local64_set(&hwc->period_left, left); 608 hwc->last_period = period; 609 ret = 1; 610 } 611 if (left > MAX_PERIOD) 612 left = MAX_PERIOD; 613 614 local64_set(&hwc->prev_count, (u64)-left); 615 616 write_pmc(idx, (u64)(-left) & 0xffffffff); 617 618 perf_event_update_userpage(event); 619 620 return ret; 621 } 622 623 /* If performance event entries have been added, move existing 624 * events around (if necessary) and then assign new entries to 625 * counters. 626 */ 627 static u64 maybe_change_configuration(struct cpu_hw_events *cpuc, u64 pcr) 628 { 629 int i; 630 631 if (!cpuc->n_added) 632 goto out; 633 634 /* Read in the counters which are moving. */ 635 for (i = 0; i < cpuc->n_events; i++) { 636 struct perf_event *cp = cpuc->event[i]; 637 638 if (cpuc->current_idx[i] != PIC_NO_INDEX && 639 cpuc->current_idx[i] != cp->hw.idx) { 640 sparc_perf_event_update(cp, &cp->hw, 641 cpuc->current_idx[i]); 642 cpuc->current_idx[i] = PIC_NO_INDEX; 643 } 644 } 645 646 /* Assign to counters all unassigned events. */ 647 for (i = 0; i < cpuc->n_events; i++) { 648 struct perf_event *cp = cpuc->event[i]; 649 struct hw_perf_event *hwc = &cp->hw; 650 int idx = hwc->idx; 651 u64 enc; 652 653 if (cpuc->current_idx[i] != PIC_NO_INDEX) 654 continue; 655 656 sparc_perf_event_set_period(cp, hwc, idx); 657 cpuc->current_idx[i] = idx; 658 659 enc = perf_event_get_enc(cpuc->events[i]); 660 pcr &= ~mask_for_index(idx); 661 pcr |= event_encoding(enc, idx); 662 } 663 out: 664 return pcr; 665 } 666 667 void hw_perf_enable(void) 668 { 669 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); 670 u64 pcr; 671 672 if (cpuc->enabled) 673 return; 674 675 cpuc->enabled = 1; 676 barrier(); 677 678 pcr = cpuc->pcr; 679 if (!cpuc->n_events) { 680 pcr = 0; 681 } else { 682 pcr = maybe_change_configuration(cpuc, pcr); 683 684 /* We require that all of the events have the same 685 * configuration, so just fetch the settings from the 686 * first entry. 687 */ 688 cpuc->pcr = pcr | cpuc->event[0]->hw.config_base; 689 } 690 691 pcr_ops->write(cpuc->pcr); 692 } 693 694 void hw_perf_disable(void) 695 { 696 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); 697 u64 val; 698 699 if (!cpuc->enabled) 700 return; 701 702 cpuc->enabled = 0; 703 cpuc->n_added = 0; 704 705 val = cpuc->pcr; 706 val &= ~(PCR_UTRACE | PCR_STRACE | 707 sparc_pmu->hv_bit | sparc_pmu->irq_bit); 708 cpuc->pcr = val; 709 710 pcr_ops->write(cpuc->pcr); 711 } 712 713 static void sparc_pmu_disable(struct perf_event *event) 714 { 715 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); 716 struct hw_perf_event *hwc = &event->hw; 717 unsigned long flags; 718 int i; 719 720 local_irq_save(flags); 721 perf_disable(); 722 723 for (i = 0; i < cpuc->n_events; i++) { 724 if (event == cpuc->event[i]) { 725 int idx = cpuc->current_idx[i]; 726 727 /* Shift remaining entries down into 728 * the existing slot. 729 */ 730 while (++i < cpuc->n_events) { 731 cpuc->event[i - 1] = cpuc->event[i]; 732 cpuc->events[i - 1] = cpuc->events[i]; 733 cpuc->current_idx[i - 1] = 734 cpuc->current_idx[i]; 735 } 736 737 /* Absorb the final count and turn off the 738 * event. 739 */ 740 sparc_pmu_disable_event(cpuc, hwc, idx); 741 barrier(); 742 sparc_perf_event_update(event, hwc, idx); 743 744 perf_event_update_userpage(event); 745 746 cpuc->n_events--; 747 break; 748 } 749 } 750 751 perf_enable(); 752 local_irq_restore(flags); 753 } 754 755 static int active_event_index(struct cpu_hw_events *cpuc, 756 struct perf_event *event) 757 { 758 int i; 759 760 for (i = 0; i < cpuc->n_events; i++) { 761 if (cpuc->event[i] == event) 762 break; 763 } 764 BUG_ON(i == cpuc->n_events); 765 return cpuc->current_idx[i]; 766 } 767 768 static void sparc_pmu_read(struct perf_event *event) 769 { 770 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); 771 int idx = active_event_index(cpuc, event); 772 struct hw_perf_event *hwc = &event->hw; 773 774 sparc_perf_event_update(event, hwc, idx); 775 } 776 777 static void sparc_pmu_unthrottle(struct perf_event *event) 778 { 779 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); 780 int idx = active_event_index(cpuc, event); 781 struct hw_perf_event *hwc = &event->hw; 782 783 sparc_pmu_enable_event(cpuc, hwc, idx); 784 } 785 786 static atomic_t active_events = ATOMIC_INIT(0); 787 static DEFINE_MUTEX(pmc_grab_mutex); 788 789 static void perf_stop_nmi_watchdog(void *unused) 790 { 791 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); 792 793 stop_nmi_watchdog(NULL); 794 cpuc->pcr = pcr_ops->read(); 795 } 796 797 void perf_event_grab_pmc(void) 798 { 799 if (atomic_inc_not_zero(&active_events)) 800 return; 801 802 mutex_lock(&pmc_grab_mutex); 803 if (atomic_read(&active_events) == 0) { 804 if (atomic_read(&nmi_active) > 0) { 805 on_each_cpu(perf_stop_nmi_watchdog, NULL, 1); 806 BUG_ON(atomic_read(&nmi_active) != 0); 807 } 808 atomic_inc(&active_events); 809 } 810 mutex_unlock(&pmc_grab_mutex); 811 } 812 813 void perf_event_release_pmc(void) 814 { 815 if (atomic_dec_and_mutex_lock(&active_events, &pmc_grab_mutex)) { 816 if (atomic_read(&nmi_active) == 0) 817 on_each_cpu(start_nmi_watchdog, NULL, 1); 818 mutex_unlock(&pmc_grab_mutex); 819 } 820 } 821 822 static const struct perf_event_map *sparc_map_cache_event(u64 config) 823 { 824 unsigned int cache_type, cache_op, cache_result; 825 const struct perf_event_map *pmap; 826 827 if (!sparc_pmu->cache_map) 828 return ERR_PTR(-ENOENT); 829 830 cache_type = (config >> 0) & 0xff; 831 if (cache_type >= PERF_COUNT_HW_CACHE_MAX) 832 return ERR_PTR(-EINVAL); 833 834 cache_op = (config >> 8) & 0xff; 835 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX) 836 return ERR_PTR(-EINVAL); 837 838 cache_result = (config >> 16) & 0xff; 839 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX) 840 return ERR_PTR(-EINVAL); 841 842 pmap = &((*sparc_pmu->cache_map)[cache_type][cache_op][cache_result]); 843 844 if (pmap->encoding == CACHE_OP_UNSUPPORTED) 845 return ERR_PTR(-ENOENT); 846 847 if (pmap->encoding == CACHE_OP_NONSENSE) 848 return ERR_PTR(-EINVAL); 849 850 return pmap; 851 } 852 853 static void hw_perf_event_destroy(struct perf_event *event) 854 { 855 perf_event_release_pmc(); 856 } 857 858 /* Make sure all events can be scheduled into the hardware at 859 * the same time. This is simplified by the fact that we only 860 * need to support 2 simultaneous HW events. 861 * 862 * As a side effect, the evts[]->hw.idx values will be assigned 863 * on success. These are pending indexes. When the events are 864 * actually programmed into the chip, these values will propagate 865 * to the per-cpu cpuc->current_idx[] slots, see the code in 866 * maybe_change_configuration() for details. 867 */ 868 static int sparc_check_constraints(struct perf_event **evts, 869 unsigned long *events, int n_ev) 870 { 871 u8 msk0 = 0, msk1 = 0; 872 int idx0 = 0; 873 874 /* This case is possible when we are invoked from 875 * hw_perf_group_sched_in(). 876 */ 877 if (!n_ev) 878 return 0; 879 880 if (n_ev > perf_max_events) 881 return -1; 882 883 msk0 = perf_event_get_msk(events[0]); 884 if (n_ev == 1) { 885 if (msk0 & PIC_LOWER) 886 idx0 = 1; 887 goto success; 888 } 889 BUG_ON(n_ev != 2); 890 msk1 = perf_event_get_msk(events[1]); 891 892 /* If both events can go on any counter, OK. */ 893 if (msk0 == (PIC_UPPER | PIC_LOWER) && 894 msk1 == (PIC_UPPER | PIC_LOWER)) 895 goto success; 896 897 /* If one event is limited to a specific counter, 898 * and the other can go on both, OK. 899 */ 900 if ((msk0 == PIC_UPPER || msk0 == PIC_LOWER) && 901 msk1 == (PIC_UPPER | PIC_LOWER)) { 902 if (msk0 & PIC_LOWER) 903 idx0 = 1; 904 goto success; 905 } 906 907 if ((msk1 == PIC_UPPER || msk1 == PIC_LOWER) && 908 msk0 == (PIC_UPPER | PIC_LOWER)) { 909 if (msk1 & PIC_UPPER) 910 idx0 = 1; 911 goto success; 912 } 913 914 /* If the events are fixed to different counters, OK. */ 915 if ((msk0 == PIC_UPPER && msk1 == PIC_LOWER) || 916 (msk0 == PIC_LOWER && msk1 == PIC_UPPER)) { 917 if (msk0 & PIC_LOWER) 918 idx0 = 1; 919 goto success; 920 } 921 922 /* Otherwise, there is a conflict. */ 923 return -1; 924 925 success: 926 evts[0]->hw.idx = idx0; 927 if (n_ev == 2) 928 evts[1]->hw.idx = idx0 ^ 1; 929 return 0; 930 } 931 932 static int check_excludes(struct perf_event **evts, int n_prev, int n_new) 933 { 934 int eu = 0, ek = 0, eh = 0; 935 struct perf_event *event; 936 int i, n, first; 937 938 n = n_prev + n_new; 939 if (n <= 1) 940 return 0; 941 942 first = 1; 943 for (i = 0; i < n; i++) { 944 event = evts[i]; 945 if (first) { 946 eu = event->attr.exclude_user; 947 ek = event->attr.exclude_kernel; 948 eh = event->attr.exclude_hv; 949 first = 0; 950 } else if (event->attr.exclude_user != eu || 951 event->attr.exclude_kernel != ek || 952 event->attr.exclude_hv != eh) { 953 return -EAGAIN; 954 } 955 } 956 957 return 0; 958 } 959 960 static int collect_events(struct perf_event *group, int max_count, 961 struct perf_event *evts[], unsigned long *events, 962 int *current_idx) 963 { 964 struct perf_event *event; 965 int n = 0; 966 967 if (!is_software_event(group)) { 968 if (n >= max_count) 969 return -1; 970 evts[n] = group; 971 events[n] = group->hw.event_base; 972 current_idx[n++] = PIC_NO_INDEX; 973 } 974 list_for_each_entry(event, &group->sibling_list, group_entry) { 975 if (!is_software_event(event) && 976 event->state != PERF_EVENT_STATE_OFF) { 977 if (n >= max_count) 978 return -1; 979 evts[n] = event; 980 events[n] = event->hw.event_base; 981 current_idx[n++] = PIC_NO_INDEX; 982 } 983 } 984 return n; 985 } 986 987 static int sparc_pmu_enable(struct perf_event *event) 988 { 989 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); 990 int n0, ret = -EAGAIN; 991 unsigned long flags; 992 993 local_irq_save(flags); 994 perf_disable(); 995 996 n0 = cpuc->n_events; 997 if (n0 >= perf_max_events) 998 goto out; 999 1000 cpuc->event[n0] = event; 1001 cpuc->events[n0] = event->hw.event_base; 1002 cpuc->current_idx[n0] = PIC_NO_INDEX; 1003 1004 /* 1005 * If group events scheduling transaction was started, 1006 * skip the schedulability test here, it will be peformed 1007 * at commit time(->commit_txn) as a whole 1008 */ 1009 if (cpuc->group_flag & PERF_EVENT_TXN) 1010 goto nocheck; 1011 1012 if (check_excludes(cpuc->event, n0, 1)) 1013 goto out; 1014 if (sparc_check_constraints(cpuc->event, cpuc->events, n0 + 1)) 1015 goto out; 1016 1017 nocheck: 1018 cpuc->n_events++; 1019 cpuc->n_added++; 1020 1021 ret = 0; 1022 out: 1023 perf_enable(); 1024 local_irq_restore(flags); 1025 return ret; 1026 } 1027 1028 static int __hw_perf_event_init(struct perf_event *event) 1029 { 1030 struct perf_event_attr *attr = &event->attr; 1031 struct perf_event *evts[MAX_HWEVENTS]; 1032 struct hw_perf_event *hwc = &event->hw; 1033 unsigned long events[MAX_HWEVENTS]; 1034 int current_idx_dmy[MAX_HWEVENTS]; 1035 const struct perf_event_map *pmap; 1036 int n; 1037 1038 if (atomic_read(&nmi_active) < 0) 1039 return -ENODEV; 1040 1041 pmap = NULL; 1042 if (attr->type == PERF_TYPE_HARDWARE) { 1043 if (attr->config >= sparc_pmu->max_events) 1044 return -EINVAL; 1045 pmap = sparc_pmu->event_map(attr->config); 1046 } else if (attr->type == PERF_TYPE_HW_CACHE) { 1047 pmap = sparc_map_cache_event(attr->config); 1048 if (IS_ERR(pmap)) 1049 return PTR_ERR(pmap); 1050 } else if (attr->type != PERF_TYPE_RAW) 1051 return -EOPNOTSUPP; 1052 1053 if (pmap) { 1054 hwc->event_base = perf_event_encode(pmap); 1055 } else { 1056 /* User gives us "(encoding << 16) | pic_mask" for 1057 * PERF_TYPE_RAW events. 1058 */ 1059 hwc->event_base = attr->config; 1060 } 1061 1062 /* We save the enable bits in the config_base. */ 1063 hwc->config_base = sparc_pmu->irq_bit; 1064 if (!attr->exclude_user) 1065 hwc->config_base |= PCR_UTRACE; 1066 if (!attr->exclude_kernel) 1067 hwc->config_base |= PCR_STRACE; 1068 if (!attr->exclude_hv) 1069 hwc->config_base |= sparc_pmu->hv_bit; 1070 1071 n = 0; 1072 if (event->group_leader != event) { 1073 n = collect_events(event->group_leader, 1074 perf_max_events - 1, 1075 evts, events, current_idx_dmy); 1076 if (n < 0) 1077 return -EINVAL; 1078 } 1079 events[n] = hwc->event_base; 1080 evts[n] = event; 1081 1082 if (check_excludes(evts, n, 1)) 1083 return -EINVAL; 1084 1085 if (sparc_check_constraints(evts, events, n + 1)) 1086 return -EINVAL; 1087 1088 hwc->idx = PIC_NO_INDEX; 1089 1090 /* Try to do all error checking before this point, as unwinding 1091 * state after grabbing the PMC is difficult. 1092 */ 1093 perf_event_grab_pmc(); 1094 event->destroy = hw_perf_event_destroy; 1095 1096 if (!hwc->sample_period) { 1097 hwc->sample_period = MAX_PERIOD; 1098 hwc->last_period = hwc->sample_period; 1099 local64_set(&hwc->period_left, hwc->sample_period); 1100 } 1101 1102 return 0; 1103 } 1104 1105 /* 1106 * Start group events scheduling transaction 1107 * Set the flag to make pmu::enable() not perform the 1108 * schedulability test, it will be performed at commit time 1109 */ 1110 static void sparc_pmu_start_txn(const struct pmu *pmu) 1111 { 1112 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events); 1113 1114 cpuhw->group_flag |= PERF_EVENT_TXN; 1115 } 1116 1117 /* 1118 * Stop group events scheduling transaction 1119 * Clear the flag and pmu::enable() will perform the 1120 * schedulability test. 1121 */ 1122 static void sparc_pmu_cancel_txn(const struct pmu *pmu) 1123 { 1124 struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events); 1125 1126 cpuhw->group_flag &= ~PERF_EVENT_TXN; 1127 } 1128 1129 /* 1130 * Commit group events scheduling transaction 1131 * Perform the group schedulability test as a whole 1132 * Return 0 if success 1133 */ 1134 static int sparc_pmu_commit_txn(const struct pmu *pmu) 1135 { 1136 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); 1137 int n; 1138 1139 if (!sparc_pmu) 1140 return -EINVAL; 1141 1142 cpuc = &__get_cpu_var(cpu_hw_events); 1143 n = cpuc->n_events; 1144 if (check_excludes(cpuc->event, 0, n)) 1145 return -EINVAL; 1146 if (sparc_check_constraints(cpuc->event, cpuc->events, n)) 1147 return -EAGAIN; 1148 1149 cpuc->group_flag &= ~PERF_EVENT_TXN; 1150 return 0; 1151 } 1152 1153 static const struct pmu pmu = { 1154 .enable = sparc_pmu_enable, 1155 .disable = sparc_pmu_disable, 1156 .read = sparc_pmu_read, 1157 .unthrottle = sparc_pmu_unthrottle, 1158 .start_txn = sparc_pmu_start_txn, 1159 .cancel_txn = sparc_pmu_cancel_txn, 1160 .commit_txn = sparc_pmu_commit_txn, 1161 }; 1162 1163 const struct pmu *hw_perf_event_init(struct perf_event *event) 1164 { 1165 int err = __hw_perf_event_init(event); 1166 1167 if (err) 1168 return ERR_PTR(err); 1169 return &pmu; 1170 } 1171 1172 void perf_event_print_debug(void) 1173 { 1174 unsigned long flags; 1175 u64 pcr, pic; 1176 int cpu; 1177 1178 if (!sparc_pmu) 1179 return; 1180 1181 local_irq_save(flags); 1182 1183 cpu = smp_processor_id(); 1184 1185 pcr = pcr_ops->read(); 1186 read_pic(pic); 1187 1188 pr_info("\n"); 1189 pr_info("CPU#%d: PCR[%016llx] PIC[%016llx]\n", 1190 cpu, pcr, pic); 1191 1192 local_irq_restore(flags); 1193 } 1194 1195 static int __kprobes perf_event_nmi_handler(struct notifier_block *self, 1196 unsigned long cmd, void *__args) 1197 { 1198 struct die_args *args = __args; 1199 struct perf_sample_data data; 1200 struct cpu_hw_events *cpuc; 1201 struct pt_regs *regs; 1202 int i; 1203 1204 if (!atomic_read(&active_events)) 1205 return NOTIFY_DONE; 1206 1207 switch (cmd) { 1208 case DIE_NMI: 1209 break; 1210 1211 default: 1212 return NOTIFY_DONE; 1213 } 1214 1215 regs = args->regs; 1216 1217 perf_sample_data_init(&data, 0); 1218 1219 cpuc = &__get_cpu_var(cpu_hw_events); 1220 1221 /* If the PMU has the TOE IRQ enable bits, we need to do a 1222 * dummy write to the %pcr to clear the overflow bits and thus 1223 * the interrupt. 1224 * 1225 * Do this before we peek at the counters to determine 1226 * overflow so we don't lose any events. 1227 */ 1228 if (sparc_pmu->irq_bit) 1229 pcr_ops->write(cpuc->pcr); 1230 1231 for (i = 0; i < cpuc->n_events; i++) { 1232 struct perf_event *event = cpuc->event[i]; 1233 int idx = cpuc->current_idx[i]; 1234 struct hw_perf_event *hwc; 1235 u64 val; 1236 1237 hwc = &event->hw; 1238 val = sparc_perf_event_update(event, hwc, idx); 1239 if (val & (1ULL << 31)) 1240 continue; 1241 1242 data.period = event->hw.last_period; 1243 if (!sparc_perf_event_set_period(event, hwc, idx)) 1244 continue; 1245 1246 if (perf_event_overflow(event, 1, &data, regs)) 1247 sparc_pmu_disable_event(cpuc, hwc, idx); 1248 } 1249 1250 return NOTIFY_STOP; 1251 } 1252 1253 static __read_mostly struct notifier_block perf_event_nmi_notifier = { 1254 .notifier_call = perf_event_nmi_handler, 1255 }; 1256 1257 static bool __init supported_pmu(void) 1258 { 1259 if (!strcmp(sparc_pmu_type, "ultra3") || 1260 !strcmp(sparc_pmu_type, "ultra3+") || 1261 !strcmp(sparc_pmu_type, "ultra3i") || 1262 !strcmp(sparc_pmu_type, "ultra4+")) { 1263 sparc_pmu = &ultra3_pmu; 1264 return true; 1265 } 1266 if (!strcmp(sparc_pmu_type, "niagara")) { 1267 sparc_pmu = &niagara1_pmu; 1268 return true; 1269 } 1270 if (!strcmp(sparc_pmu_type, "niagara2")) { 1271 sparc_pmu = &niagara2_pmu; 1272 return true; 1273 } 1274 return false; 1275 } 1276 1277 void __init init_hw_perf_events(void) 1278 { 1279 pr_info("Performance events: "); 1280 1281 if (!supported_pmu()) { 1282 pr_cont("No support for PMU type '%s'\n", sparc_pmu_type); 1283 return; 1284 } 1285 1286 pr_cont("Supported PMU type is '%s'\n", sparc_pmu_type); 1287 1288 /* All sparc64 PMUs currently have 2 events. */ 1289 perf_max_events = 2; 1290 1291 register_die_notifier(&perf_event_nmi_notifier); 1292 } 1293 1294 static inline void callchain_store(struct perf_callchain_entry *entry, u64 ip) 1295 { 1296 if (entry->nr < PERF_MAX_STACK_DEPTH) 1297 entry->ip[entry->nr++] = ip; 1298 } 1299 1300 static void perf_callchain_kernel(struct pt_regs *regs, 1301 struct perf_callchain_entry *entry) 1302 { 1303 unsigned long ksp, fp; 1304 #ifdef CONFIG_FUNCTION_GRAPH_TRACER 1305 int graph = 0; 1306 #endif 1307 1308 callchain_store(entry, PERF_CONTEXT_KERNEL); 1309 callchain_store(entry, regs->tpc); 1310 1311 ksp = regs->u_regs[UREG_I6]; 1312 fp = ksp + STACK_BIAS; 1313 do { 1314 struct sparc_stackf *sf; 1315 struct pt_regs *regs; 1316 unsigned long pc; 1317 1318 if (!kstack_valid(current_thread_info(), fp)) 1319 break; 1320 1321 sf = (struct sparc_stackf *) fp; 1322 regs = (struct pt_regs *) (sf + 1); 1323 1324 if (kstack_is_trap_frame(current_thread_info(), regs)) { 1325 if (user_mode(regs)) 1326 break; 1327 pc = regs->tpc; 1328 fp = regs->u_regs[UREG_I6] + STACK_BIAS; 1329 } else { 1330 pc = sf->callers_pc; 1331 fp = (unsigned long)sf->fp + STACK_BIAS; 1332 } 1333 callchain_store(entry, pc); 1334 #ifdef CONFIG_FUNCTION_GRAPH_TRACER 1335 if ((pc + 8UL) == (unsigned long) &return_to_handler) { 1336 int index = current->curr_ret_stack; 1337 if (current->ret_stack && index >= graph) { 1338 pc = current->ret_stack[index - graph].ret; 1339 callchain_store(entry, pc); 1340 graph++; 1341 } 1342 } 1343 #endif 1344 } while (entry->nr < PERF_MAX_STACK_DEPTH); 1345 } 1346 1347 static void perf_callchain_user_64(struct pt_regs *regs, 1348 struct perf_callchain_entry *entry) 1349 { 1350 unsigned long ufp; 1351 1352 callchain_store(entry, PERF_CONTEXT_USER); 1353 callchain_store(entry, regs->tpc); 1354 1355 ufp = regs->u_regs[UREG_I6] + STACK_BIAS; 1356 do { 1357 struct sparc_stackf *usf, sf; 1358 unsigned long pc; 1359 1360 usf = (struct sparc_stackf *) ufp; 1361 if (__copy_from_user_inatomic(&sf, usf, sizeof(sf))) 1362 break; 1363 1364 pc = sf.callers_pc; 1365 ufp = (unsigned long)sf.fp + STACK_BIAS; 1366 callchain_store(entry, pc); 1367 } while (entry->nr < PERF_MAX_STACK_DEPTH); 1368 } 1369 1370 static void perf_callchain_user_32(struct pt_regs *regs, 1371 struct perf_callchain_entry *entry) 1372 { 1373 unsigned long ufp; 1374 1375 callchain_store(entry, PERF_CONTEXT_USER); 1376 callchain_store(entry, regs->tpc); 1377 1378 ufp = regs->u_regs[UREG_I6] & 0xffffffffUL; 1379 do { 1380 struct sparc_stackf32 *usf, sf; 1381 unsigned long pc; 1382 1383 usf = (struct sparc_stackf32 *) ufp; 1384 if (__copy_from_user_inatomic(&sf, usf, sizeof(sf))) 1385 break; 1386 1387 pc = sf.callers_pc; 1388 ufp = (unsigned long)sf.fp; 1389 callchain_store(entry, pc); 1390 } while (entry->nr < PERF_MAX_STACK_DEPTH); 1391 } 1392 1393 /* Like powerpc we can't get PMU interrupts within the PMU handler, 1394 * so no need for separate NMI and IRQ chains as on x86. 1395 */ 1396 static DEFINE_PER_CPU(struct perf_callchain_entry, callchain); 1397 1398 struct perf_callchain_entry *perf_callchain(struct pt_regs *regs) 1399 { 1400 struct perf_callchain_entry *entry = &__get_cpu_var(callchain); 1401 1402 entry->nr = 0; 1403 if (!user_mode(regs)) { 1404 stack_trace_flush(); 1405 perf_callchain_kernel(regs, entry); 1406 if (current->mm) 1407 regs = task_pt_regs(current); 1408 else 1409 regs = NULL; 1410 } 1411 if (regs) { 1412 flushw_user(); 1413 if (test_thread_flag(TIF_32BIT)) 1414 perf_callchain_user_32(regs, entry); 1415 else 1416 perf_callchain_user_64(regs, entry); 1417 } 1418 return entry; 1419 } 1420