| 1 | /* |
| 2 | * Copyright (C) 2012 Fusion-io |
| 3 | * |
| 4 | * This program is free software; you can redistribute it and/or |
| 5 | * modify it under the terms of the GNU General Public |
| 6 | * License v2 as published by the Free Software Foundation. |
| 7 | * |
| 8 | * This program is distributed in the hope that it will be useful, |
| 9 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 10 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 11 | * GNU General Public License for more details. |
| 12 | * |
| 13 | * You should have received a copy of the GNU General Public License |
| 14 | * along with this program; if not, write to the Free Software |
| 15 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 16 | * |
| 17 | * Parts of this file were imported from Jens Axboe's blktrace sources (also GPL) |
| 18 | */ |
| 19 | #include <sys/types.h> |
| 20 | #include <sys/stat.h> |
| 21 | #include <fcntl.h> |
| 22 | #include <unistd.h> |
| 23 | #include <stdlib.h> |
| 24 | #include <stdio.h> |
| 25 | #include <math.h> |
| 26 | #include <inttypes.h> |
| 27 | #include <string.h> |
| 28 | #include <asm/types.h> |
| 29 | #include <errno.h> |
| 30 | #include <sys/mman.h> |
| 31 | #include <time.h> |
| 32 | #include <math.h> |
| 33 | #include <dirent.h> |
| 34 | |
| 35 | #include "plot.h" |
| 36 | #include "blkparse.h" |
| 37 | #include "list.h" |
| 38 | #include "tracers.h" |
| 39 | #include "../blktrace_api.h" |
| 40 | |
| 41 | #define IO_HASH_TABLE_BITS 11 |
| 42 | #define IO_HASH_TABLE_SIZE (1 << IO_HASH_TABLE_BITS) |
| 43 | static struct list_head io_hash_table[IO_HASH_TABLE_SIZE]; |
| 44 | static u64 ios_in_flight = 0; |
| 45 | |
| 46 | #define PROCESS_HASH_TABLE_BITS 7 |
| 47 | #define PROCESS_HASH_TABLE_SIZE (1 << PROCESS_HASH_TABLE_BITS) |
| 48 | static struct list_head process_hash_table[PROCESS_HASH_TABLE_SIZE]; |
| 49 | |
| 50 | extern int plot_io_action; |
| 51 | extern int io_per_process; |
| 52 | |
| 53 | #define BLK_DATADIR(a) (((a) >> BLK_TC_SHIFT) & (BLK_TC_READ | BLK_TC_WRITE)) |
| 54 | #define BLK_TA_MASK (((1 << BLK_TC_SHIFT) - 1) & ~__BLK_TA_CGROUP) |
| 55 | |
| 56 | struct pending_io { |
| 57 | /* sector offset of this IO */ |
| 58 | u64 sector; |
| 59 | |
| 60 | /* dev_t for this IO */ |
| 61 | u32 device; |
| 62 | |
| 63 | /* time this IO was dispatched */ |
| 64 | u64 dispatch_time; |
| 65 | /* time this IO was finished */ |
| 66 | u64 completion_time; |
| 67 | struct list_head hash_list; |
| 68 | /* process which queued this IO */ |
| 69 | u32 pid; |
| 70 | }; |
| 71 | |
| 72 | struct pid_map { |
| 73 | struct list_head hash_list; |
| 74 | u32 pid; |
| 75 | int index; |
| 76 | char name[0]; |
| 77 | }; |
| 78 | |
| 79 | u64 get_record_time(struct trace *trace) |
| 80 | { |
| 81 | return trace->io->time; |
| 82 | } |
| 83 | |
| 84 | void init_io_hash_table(void) |
| 85 | { |
| 86 | int i; |
| 87 | struct list_head *head; |
| 88 | |
| 89 | for (i = 0; i < IO_HASH_TABLE_SIZE; i++) { |
| 90 | head = io_hash_table + i; |
| 91 | INIT_LIST_HEAD(head); |
| 92 | } |
| 93 | } |
| 94 | |
| 95 | /* taken from the kernel hash.h */ |
| 96 | static inline u64 hash_sector(u64 val) |
| 97 | { |
| 98 | u64 hash = val; |
| 99 | |
| 100 | /* Sigh, gcc can't optimise this alone like it does for 32 bits. */ |
| 101 | u64 n = hash; |
| 102 | n <<= 18; |
| 103 | hash -= n; |
| 104 | n <<= 33; |
| 105 | hash -= n; |
| 106 | n <<= 3; |
| 107 | hash += n; |
| 108 | n <<= 3; |
| 109 | hash -= n; |
| 110 | n <<= 4; |
| 111 | hash += n; |
| 112 | n <<= 2; |
| 113 | hash += n; |
| 114 | |
| 115 | /* High bits are more random, so use them. */ |
| 116 | return hash >> (64 - IO_HASH_TABLE_BITS); |
| 117 | } |
| 118 | |
| 119 | static int io_hash_table_insert(struct pending_io *ins_pio) |
| 120 | { |
| 121 | u64 sector = ins_pio->sector; |
| 122 | u32 dev = ins_pio->device; |
| 123 | int slot = hash_sector(sector); |
| 124 | struct list_head *head; |
| 125 | struct pending_io *pio; |
| 126 | |
| 127 | head = io_hash_table + slot; |
| 128 | list_for_each_entry(pio, head, hash_list) { |
| 129 | if (pio->sector == sector && pio->device == dev) |
| 130 | return -EEXIST; |
| 131 | } |
| 132 | list_add_tail(&ins_pio->hash_list, head); |
| 133 | return 0; |
| 134 | } |
| 135 | |
| 136 | static struct pending_io *io_hash_table_search(u64 sector, u32 dev) |
| 137 | { |
| 138 | int slot = hash_sector(sector); |
| 139 | struct list_head *head; |
| 140 | struct pending_io *pio; |
| 141 | |
| 142 | head = io_hash_table + slot; |
| 143 | list_for_each_entry(pio, head, hash_list) { |
| 144 | if (pio->sector == sector && pio->device == dev) |
| 145 | return pio; |
| 146 | } |
| 147 | return NULL; |
| 148 | } |
| 149 | |
| 150 | static struct pending_io *hash_queued_io(struct blk_io_trace *io) |
| 151 | { |
| 152 | struct pending_io *pio; |
| 153 | int ret; |
| 154 | |
| 155 | pio = calloc(1, sizeof(*pio)); |
| 156 | pio->sector = io->sector; |
| 157 | pio->device = io->device; |
| 158 | pio->pid = io->pid; |
| 159 | |
| 160 | ret = io_hash_table_insert(pio); |
| 161 | if (ret < 0) { |
| 162 | /* crud, the IO is there already */ |
| 163 | free(pio); |
| 164 | return NULL; |
| 165 | } |
| 166 | return pio; |
| 167 | } |
| 168 | |
| 169 | static struct pending_io *hash_dispatched_io(struct blk_io_trace *io) |
| 170 | { |
| 171 | struct pending_io *pio; |
| 172 | |
| 173 | pio = io_hash_table_search(io->sector, io->device); |
| 174 | if (!pio) { |
| 175 | pio = hash_queued_io(io); |
| 176 | if (!pio) |
| 177 | return NULL; |
| 178 | } |
| 179 | pio->dispatch_time = io->time; |
| 180 | return pio; |
| 181 | } |
| 182 | |
| 183 | static struct pending_io *hash_completed_io(struct blk_io_trace *io) |
| 184 | { |
| 185 | struct pending_io *pio; |
| 186 | |
| 187 | pio = io_hash_table_search(io->sector, io->device); |
| 188 | |
| 189 | if (!pio) |
| 190 | return NULL; |
| 191 | return pio; |
| 192 | } |
| 193 | |
| 194 | void init_process_hash_table(void) |
| 195 | { |
| 196 | int i; |
| 197 | struct list_head *head; |
| 198 | |
| 199 | for (i = 0; i < PROCESS_HASH_TABLE_SIZE; i++) { |
| 200 | head = process_hash_table + i; |
| 201 | INIT_LIST_HEAD(head); |
| 202 | } |
| 203 | } |
| 204 | |
| 205 | static u32 hash_pid(u32 pid) |
| 206 | { |
| 207 | u32 hash = pid; |
| 208 | |
| 209 | hash ^= pid >> 3; |
| 210 | hash ^= pid >> 3; |
| 211 | hash ^= pid >> 4; |
| 212 | hash ^= pid >> 6; |
| 213 | return (hash & (PROCESS_HASH_TABLE_SIZE - 1)); |
| 214 | } |
| 215 | |
| 216 | static struct pid_map *process_hash_search(u32 pid) |
| 217 | { |
| 218 | int slot = hash_pid(pid); |
| 219 | struct list_head *head; |
| 220 | struct pid_map *pm; |
| 221 | |
| 222 | head = process_hash_table + slot; |
| 223 | list_for_each_entry(pm, head, hash_list) { |
| 224 | if (pm->pid == pid) |
| 225 | return pm; |
| 226 | } |
| 227 | return NULL; |
| 228 | } |
| 229 | |
| 230 | static struct pid_map *process_hash_insert(u32 pid, char *name) |
| 231 | { |
| 232 | int slot = hash_pid(pid); |
| 233 | struct pid_map *pm; |
| 234 | int old_index = 0; |
| 235 | char buf[16]; |
| 236 | |
| 237 | pm = process_hash_search(pid); |
| 238 | if (pm) { |
| 239 | /* Entry exists and name shouldn't be changed? */ |
| 240 | if (!name || !strcmp(name, pm->name)) |
| 241 | return pm; |
| 242 | list_del(&pm->hash_list); |
| 243 | old_index = pm->index; |
| 244 | free(pm); |
| 245 | } |
| 246 | if (!name) { |
| 247 | sprintf(buf, "[%u]", pid); |
| 248 | name = buf; |
| 249 | } |
| 250 | pm = malloc(sizeof(struct pid_map) + strlen(name) + 1); |
| 251 | pm->pid = pid; |
| 252 | pm->index = old_index; |
| 253 | strcpy(pm->name, name); |
| 254 | list_add_tail(&pm->hash_list, process_hash_table + slot); |
| 255 | |
| 256 | return pm; |
| 257 | } |
| 258 | |
| 259 | static void handle_notify(struct trace *trace) |
| 260 | { |
| 261 | struct blk_io_trace *io = trace->io; |
| 262 | void *payload = (char *)io + sizeof(*io); |
| 263 | int pdu_len = io->pdu_len; |
| 264 | u32 two32[2]; |
| 265 | |
| 266 | if (io->action & __BLK_TN_CGROUP) { |
| 267 | payload += sizeof(struct blk_io_cgroup_payload); |
| 268 | pdu_len -= sizeof(struct blk_io_cgroup_payload); |
| 269 | } |
| 270 | if ((io->action & ~__BLK_TN_CGROUP) == BLK_TN_PROCESS) { |
| 271 | if (io_per_process) |
| 272 | process_hash_insert(io->pid, payload); |
| 273 | return; |
| 274 | } |
| 275 | |
| 276 | if ((io->action & ~__BLK_TN_CGROUP) != BLK_TN_TIMESTAMP) |
| 277 | return; |
| 278 | |
| 279 | if (pdu_len != sizeof(two32)) |
| 280 | return; |
| 281 | |
| 282 | memcpy(two32, payload, sizeof(two32)); |
| 283 | trace->start_timestamp = io->time; |
| 284 | trace->abs_start_time.tv_sec = two32[0]; |
| 285 | trace->abs_start_time.tv_nsec = two32[1]; |
| 286 | if (trace->abs_start_time.tv_nsec < 0) { |
| 287 | trace->abs_start_time.tv_sec--; |
| 288 | trace->abs_start_time.tv_nsec += 1000000000; |
| 289 | } |
| 290 | } |
| 291 | |
| 292 | int next_record(struct trace *trace) |
| 293 | { |
| 294 | int skip = trace->io->pdu_len; |
| 295 | u64 offset; |
| 296 | |
| 297 | trace->cur += sizeof(*trace->io) + skip; |
| 298 | offset = trace->cur - trace->start; |
| 299 | if (offset >= trace->len) |
| 300 | return 1; |
| 301 | |
| 302 | trace->io = (struct blk_io_trace *)trace->cur; |
| 303 | return 0; |
| 304 | } |
| 305 | |
| 306 | void first_record(struct trace *trace) |
| 307 | { |
| 308 | trace->cur = trace->start; |
| 309 | trace->io = (struct blk_io_trace *)trace->cur; |
| 310 | } |
| 311 | |
| 312 | static int is_io_event(struct blk_io_trace *test) |
| 313 | { |
| 314 | char *message; |
| 315 | if (!(test->action & BLK_TC_ACT(BLK_TC_NOTIFY))) |
| 316 | return 1; |
| 317 | if ((test->action & ~__BLK_TN_CGROUP) == BLK_TN_MESSAGE) { |
| 318 | int len = test->pdu_len; |
| 319 | |
| 320 | message = (char *)(test + 1); |
| 321 | if (test->action & __BLK_TN_CGROUP) { |
| 322 | len -= sizeof(struct blk_io_cgroup_payload); |
| 323 | message += sizeof(struct blk_io_cgroup_payload); |
| 324 | } |
| 325 | if (len < 3) |
| 326 | return 0; |
| 327 | if (strncmp(message, "fio ", 4) == 0) { |
| 328 | return 1; |
| 329 | } |
| 330 | } |
| 331 | return 0; |
| 332 | } |
| 333 | |
| 334 | u64 find_last_time(struct trace *trace) |
| 335 | { |
| 336 | char *p = trace->start + trace->len; |
| 337 | struct blk_io_trace *test; |
| 338 | int search_len = 0; |
| 339 | u64 found = 0; |
| 340 | |
| 341 | if (trace->len < sizeof(*trace->io)) |
| 342 | return 0; |
| 343 | p -= sizeof(*trace->io); |
| 344 | while (p >= trace->start) { |
| 345 | test = (struct blk_io_trace *)p; |
| 346 | if (CHECK_MAGIC(test) && is_io_event(test)) { |
| 347 | u64 offset = p - trace->start; |
| 348 | if (offset + sizeof(*test) + test->pdu_len == trace->len) { |
| 349 | return test->time; |
| 350 | } |
| 351 | } |
| 352 | p--; |
| 353 | search_len++; |
| 354 | if (search_len > 8192) { |
| 355 | break; |
| 356 | } |
| 357 | } |
| 358 | |
| 359 | /* searching backwards didn't work out, we'll have to scan the file */ |
| 360 | first_record(trace); |
| 361 | while (1) { |
| 362 | if (is_io_event(trace->io)) |
| 363 | found = trace->io->time; |
| 364 | if (next_record(trace)) |
| 365 | break; |
| 366 | } |
| 367 | first_record(trace); |
| 368 | return found; |
| 369 | } |
| 370 | |
| 371 | static int parse_fio_bank_message(struct trace *trace, u64 *bank_ret, u64 *offset_ret, |
| 372 | u64 *num_banks_ret) |
| 373 | { |
| 374 | char *s; |
| 375 | char *next; |
| 376 | char *message; |
| 377 | struct blk_io_trace *test = trace->io; |
| 378 | int len = test->pdu_len; |
| 379 | u64 bank; |
| 380 | u64 offset; |
| 381 | u64 num_banks; |
| 382 | |
| 383 | if (!(test->action & BLK_TC_ACT(BLK_TC_NOTIFY))) |
| 384 | return -1; |
| 385 | if ((test->action & ~__BLK_TN_CGROUP) != BLK_TN_MESSAGE) |
| 386 | return -1; |
| 387 | |
| 388 | message = (char *)(test + 1); |
| 389 | if (test->action & __BLK_TN_CGROUP) { |
| 390 | len -= sizeof(struct blk_io_cgroup_payload); |
| 391 | message += sizeof(struct blk_io_cgroup_payload); |
| 392 | } |
| 393 | /* the message is fio rw bank offset num_banks */ |
| 394 | if (len < 3) |
| 395 | return -1; |
| 396 | if (strncmp(message, "fio r ", 6) != 0) |
| 397 | return -1; |
| 398 | |
| 399 | message = strndup(message, len); |
| 400 | s = strchr(message, ' '); |
| 401 | if (!s) |
| 402 | goto out; |
| 403 | s++; |
| 404 | s = strchr(s, ' '); |
| 405 | if (!s) |
| 406 | goto out; |
| 407 | |
| 408 | bank = strtoll(s, &next, 10); |
| 409 | if (s == next) |
| 410 | goto out; |
| 411 | s = next; |
| 412 | |
| 413 | offset = strtoll(s, &next, 10); |
| 414 | if (s == next) |
| 415 | goto out; |
| 416 | s = next; |
| 417 | |
| 418 | num_banks = strtoll(s, &next, 10); |
| 419 | if (s == next) |
| 420 | goto out; |
| 421 | |
| 422 | *bank_ret = bank; |
| 423 | *offset_ret = offset; |
| 424 | *num_banks_ret = num_banks; |
| 425 | |
| 426 | return 0; |
| 427 | out: |
| 428 | free(message); |
| 429 | return -1; |
| 430 | } |
| 431 | |
| 432 | static struct dev_info *lookup_dev(struct trace *trace, struct blk_io_trace *io) |
| 433 | { |
| 434 | u32 dev = io->device; |
| 435 | int i; |
| 436 | struct dev_info *di = NULL; |
| 437 | |
| 438 | for (i = 0; i < trace->num_devices; i++) { |
| 439 | if (trace->devices[i].device == dev) { |
| 440 | di = trace->devices + i; |
| 441 | goto found; |
| 442 | } |
| 443 | } |
| 444 | i = trace->num_devices++; |
| 445 | if (i >= MAX_DEVICES_PER_TRACE) { |
| 446 | fprintf(stderr, "Trace contains too many devices (%d)\n", i); |
| 447 | exit(1); |
| 448 | } |
| 449 | di = trace->devices + i; |
| 450 | di->device = dev; |
| 451 | found: |
| 452 | return di; |
| 453 | } |
| 454 | |
| 455 | static void map_devices(struct trace *trace) |
| 456 | { |
| 457 | struct dev_info *di; |
| 458 | u64 found; |
| 459 | u64 map_start = 0; |
| 460 | int i; |
| 461 | |
| 462 | first_record(trace); |
| 463 | while (1) { |
| 464 | if (!(trace->io->action & BLK_TC_ACT(BLK_TC_NOTIFY))) { |
| 465 | di = lookup_dev(trace, trace->io); |
| 466 | found = trace->io->sector << 9; |
| 467 | if (found < di->min) |
| 468 | di->min = found; |
| 469 | |
| 470 | found += trace->io->bytes; |
| 471 | if (di->max < found) |
| 472 | di->max = found; |
| 473 | } |
| 474 | if (next_record(trace)) |
| 475 | break; |
| 476 | } |
| 477 | first_record(trace); |
| 478 | for (i = 0; i < trace->num_devices; i++) { |
| 479 | di = trace->devices + i; |
| 480 | di->map = map_start; |
| 481 | map_start += di->max - di->min; |
| 482 | } |
| 483 | } |
| 484 | |
| 485 | static u64 map_io(struct trace *trace, struct blk_io_trace *io) |
| 486 | { |
| 487 | struct dev_info *di = lookup_dev(trace, io); |
| 488 | u64 val = trace->io->sector << 9; |
| 489 | return di->map + val - di->min; |
| 490 | } |
| 491 | |
| 492 | void find_extreme_offsets(struct trace *trace, u64 *min_ret, u64 *max_ret, u64 *max_bank_ret, |
| 493 | u64 *max_offset_ret) |
| 494 | { |
| 495 | u64 found = 0; |
| 496 | u64 max = 0, min = ~(u64)0; |
| 497 | u64 max_bank = 0; |
| 498 | u64 max_bank_offset = 0; |
| 499 | u64 num_banks = 0; |
| 500 | |
| 501 | map_devices(trace); |
| 502 | |
| 503 | first_record(trace); |
| 504 | while (1) { |
| 505 | if (!(trace->io->action & BLK_TC_ACT(BLK_TC_NOTIFY))) { |
| 506 | found = map_io(trace, trace->io); |
| 507 | if (found < min) |
| 508 | min = found; |
| 509 | |
| 510 | found += trace->io->bytes; |
| 511 | if (max < found) |
| 512 | max = found; |
| 513 | } else { |
| 514 | u64 bank; |
| 515 | u64 offset; |
| 516 | if (!parse_fio_bank_message(trace, &bank, |
| 517 | &offset, &num_banks)) { |
| 518 | if (bank > max_bank) |
| 519 | max_bank = bank; |
| 520 | if (offset > max_bank_offset) |
| 521 | max_bank_offset = offset; |
| 522 | } |
| 523 | } |
| 524 | if (next_record(trace)) |
| 525 | break; |
| 526 | } |
| 527 | first_record(trace); |
| 528 | *min_ret = min; |
| 529 | *max_ret = max; |
| 530 | *max_bank_ret = max_bank; |
| 531 | *max_offset_ret = max_bank_offset; |
| 532 | } |
| 533 | |
| 534 | static void check_io_types(struct trace *trace) |
| 535 | { |
| 536 | struct blk_io_trace *io = trace->io; |
| 537 | int action = io->action & BLK_TA_MASK; |
| 538 | |
| 539 | if (!(io->action & BLK_TC_ACT(BLK_TC_NOTIFY))) { |
| 540 | switch (action) { |
| 541 | case __BLK_TA_COMPLETE: |
| 542 | trace->found_completion = 1; |
| 543 | break; |
| 544 | case __BLK_TA_ISSUE: |
| 545 | trace->found_issue = 1; |
| 546 | break; |
| 547 | case __BLK_TA_QUEUE: |
| 548 | trace->found_queue = 1; |
| 549 | break; |
| 550 | }; |
| 551 | } |
| 552 | } |
| 553 | |
| 554 | |
| 555 | int filter_outliers(struct trace *trace, u64 min_offset, u64 max_offset, |
| 556 | u64 *yzoom_min, u64 *yzoom_max) |
| 557 | { |
| 558 | int hits[11]; |
| 559 | u64 max_per_bucket[11]; |
| 560 | u64 min_per_bucket[11]; |
| 561 | u64 bytes_per_bucket = (max_offset - min_offset + 1) / 10; |
| 562 | int slot; |
| 563 | int fat_count = 0; |
| 564 | |
| 565 | memset(hits, 0, sizeof(int) * 11); |
| 566 | memset(max_per_bucket, 0, sizeof(u64) * 11); |
| 567 | memset(min_per_bucket, 0xff, sizeof(u64) * 11); |
| 568 | first_record(trace); |
| 569 | while (1) { |
| 570 | check_io_types(trace); |
| 571 | if (!(trace->io->action & BLK_TC_ACT(BLK_TC_NOTIFY)) && |
| 572 | (trace->io->action & BLK_TA_MASK) == __BLK_TA_QUEUE) { |
| 573 | u64 off = map_io(trace, trace->io) - min_offset; |
| 574 | |
| 575 | slot = (int)(off / bytes_per_bucket); |
| 576 | hits[slot]++; |
| 577 | if (off < min_per_bucket[slot]) |
| 578 | min_per_bucket[slot] = off; |
| 579 | |
| 580 | off += trace->io->bytes; |
| 581 | slot = (int)(off / bytes_per_bucket); |
| 582 | hits[slot]++; |
| 583 | if (off > max_per_bucket[slot]) |
| 584 | max_per_bucket[slot] = off; |
| 585 | } |
| 586 | if (next_record(trace)) |
| 587 | break; |
| 588 | } |
| 589 | first_record(trace); |
| 590 | for (slot = 0; slot < 11; slot++) { |
| 591 | if (hits[slot] > fat_count) { |
| 592 | fat_count = hits[slot]; |
| 593 | } |
| 594 | } |
| 595 | |
| 596 | *yzoom_max = max_offset; |
| 597 | for (slot = 10; slot >= 0; slot--) { |
| 598 | double d = hits[slot]; |
| 599 | |
| 600 | if (d >= (double)fat_count * .05) { |
| 601 | *yzoom_max = max_per_bucket[slot] + min_offset; |
| 602 | break; |
| 603 | } |
| 604 | } |
| 605 | |
| 606 | *yzoom_min = min_offset; |
| 607 | for (slot = 0; slot < 10; slot++) { |
| 608 | double d = hits[slot]; |
| 609 | |
| 610 | if (d >= (double)fat_count * .05) { |
| 611 | *yzoom_min = min_per_bucket[slot] + min_offset; |
| 612 | break; |
| 613 | } |
| 614 | } |
| 615 | return 0; |
| 616 | } |
| 617 | |
| 618 | static char footer[] = ".blktrace.0"; |
| 619 | static int footer_len = sizeof(footer) - 1; |
| 620 | |
| 621 | static int match_trace(char *name, int *len) |
| 622 | { |
| 623 | int match_len; |
| 624 | int footer_start; |
| 625 | |
| 626 | match_len = strlen(name); |
| 627 | if (match_len <= footer_len) |
| 628 | return 0; |
| 629 | |
| 630 | footer_start = match_len - footer_len; |
| 631 | if (strcmp(name + footer_start, footer) != 0) |
| 632 | return 0; |
| 633 | |
| 634 | if (len) |
| 635 | *len = match_len; |
| 636 | return 1; |
| 637 | } |
| 638 | |
| 639 | struct tracelist { |
| 640 | struct tracelist *next; |
| 641 | char *name; |
| 642 | }; |
| 643 | |
| 644 | static struct tracelist *traces_list(char *dir_name, int *len) |
| 645 | { |
| 646 | int count = 0; |
| 647 | struct tracelist *traces = NULL; |
| 648 | int dlen = strlen(dir_name); |
| 649 | DIR *dir = opendir(dir_name); |
| 650 | if (!dir) |
| 651 | return NULL; |
| 652 | |
| 653 | while (1) { |
| 654 | int n = 0; |
| 655 | struct tracelist *tl; |
| 656 | struct dirent *d = readdir(dir); |
| 657 | if (!d) |
| 658 | break; |
| 659 | |
| 660 | if (!match_trace(d->d_name, &n)) |
| 661 | continue; |
| 662 | |
| 663 | n += dlen + 1; /* dir + '/' + file */ |
| 664 | /* Allocate space for tracelist + filename */ |
| 665 | tl = calloc(1, sizeof(struct tracelist) + (sizeof(char) * (n + 1))); |
| 666 | if (!tl) { |
| 667 | closedir(dir); |
| 668 | return NULL; |
| 669 | } |
| 670 | tl->next = traces; |
| 671 | tl->name = (char *)(tl + 1); |
| 672 | snprintf(tl->name, n, "%s/%s", dir_name, d->d_name); |
| 673 | traces = tl; |
| 674 | count++; |
| 675 | } |
| 676 | |
| 677 | closedir(dir); |
| 678 | |
| 679 | if (len) |
| 680 | *len = count; |
| 681 | |
| 682 | return traces; |
| 683 | } |
| 684 | |
| 685 | static void traces_free(struct tracelist *traces) |
| 686 | { |
| 687 | while (traces) { |
| 688 | struct tracelist *tl = traces; |
| 689 | traces = traces->next; |
| 690 | free(tl); |
| 691 | } |
| 692 | } |
| 693 | |
| 694 | static int dump_traces(struct tracelist *traces, int count, char *dumpfile) |
| 695 | { |
| 696 | struct tracelist *tl; |
| 697 | char **argv = NULL; |
| 698 | int argc = 0; |
| 699 | int i; |
| 700 | int err = 0; |
| 701 | |
| 702 | argc = count * 2; /* {"-i", trace } */ |
| 703 | argc += 4; /* See below */ |
| 704 | argv = calloc(argc + 1, sizeof(char *)); |
| 705 | if (!argv) |
| 706 | return -errno; |
| 707 | |
| 708 | i = 0; |
| 709 | argv[i++] = "blkparse"; |
| 710 | argv[i++] = "-O"; |
| 711 | argv[i++] = "-d"; |
| 712 | argv[i++] = dumpfile; |
| 713 | for (tl = traces; tl != NULL; tl = tl->next) { |
| 714 | argv[i++] = "-i"; |
| 715 | argv[i++] = tl->name; |
| 716 | } |
| 717 | |
| 718 | err = run_program(argc, argv, 1, NULL, NULL); |
| 719 | if (err) |
| 720 | fprintf(stderr, "%s exited with %d, expected 0\n", argv[0], err); |
| 721 | free(argv); |
| 722 | return err; |
| 723 | } |
| 724 | |
| 725 | static char *find_trace_file(char *filename) |
| 726 | { |
| 727 | int ret; |
| 728 | struct stat st; |
| 729 | char *dot; |
| 730 | int found_dir = 0; |
| 731 | char *dumpfile; |
| 732 | int len = strlen(filename); |
| 733 | |
| 734 | /* look for an exact match of whatever they pass in. |
| 735 | * If it is a file, assume it is the dump file. |
| 736 | * If a directory, remember that it existed so we |
| 737 | * can combine traces in that directory later |
| 738 | */ |
| 739 | ret = stat(filename, &st); |
| 740 | if (ret == 0) { |
| 741 | if (S_ISREG(st.st_mode)) |
| 742 | return strdup(filename); |
| 743 | |
| 744 | if (S_ISDIR(st.st_mode)) |
| 745 | found_dir = 1; |
| 746 | } |
| 747 | |
| 748 | if (found_dir) { |
| 749 | int i; |
| 750 | /* Eat up trailing '/'s */ |
| 751 | for (i = len - 1; filename[i] == '/'; i--) |
| 752 | filename[i] = '\0'; |
| 753 | } |
| 754 | |
| 755 | /* |
| 756 | * try tacking .dump onto the end and see if that already |
| 757 | * has been generated |
| 758 | */ |
| 759 | ret = asprintf(&dumpfile, "%s.dump", filename); |
| 760 | if (ret == -1) { |
| 761 | perror("Error building dump file name"); |
| 762 | return NULL; |
| 763 | } |
| 764 | ret = stat(dumpfile, &st); |
| 765 | if (ret == 0) |
| 766 | return dumpfile; |
| 767 | |
| 768 | /* |
| 769 | * try to generate the .dump from all the traces in |
| 770 | * a single dir. |
| 771 | */ |
| 772 | if (found_dir) { |
| 773 | int count; |
| 774 | struct tracelist *traces = traces_list(filename, &count); |
| 775 | if (traces) { |
| 776 | ret = dump_traces(traces, count, dumpfile); |
| 777 | traces_free(traces); |
| 778 | if (ret == 0) |
| 779 | return dumpfile; |
| 780 | } |
| 781 | } |
| 782 | free(dumpfile); |
| 783 | |
| 784 | /* |
| 785 | * try to generate the .dump from all the blktrace |
| 786 | * files for a named trace |
| 787 | */ |
| 788 | dot = strrchr(filename, '.'); |
| 789 | if (!dot || strcmp(".dump", dot) != 0) { |
| 790 | struct tracelist trace = {0 ,NULL}; |
| 791 | if (dot && dot != filename) |
| 792 | len = dot - filename; |
| 793 | |
| 794 | ret = asprintf(&trace.name, "%*s.blktrace.0", len, filename); |
| 795 | if (ret == -1) |
| 796 | return NULL; |
| 797 | ret = asprintf(&dumpfile, "%*s.dump", len, filename); |
| 798 | if (ret == -1) { |
| 799 | free(trace.name); |
| 800 | return NULL; |
| 801 | } |
| 802 | |
| 803 | ret = dump_traces(&trace, 1, dumpfile); |
| 804 | if (ret == 0) { |
| 805 | free(trace.name); |
| 806 | return dumpfile; |
| 807 | } |
| 808 | free(trace.name); |
| 809 | free(dumpfile); |
| 810 | } |
| 811 | return NULL; |
| 812 | } |
| 813 | struct trace *open_trace(char *filename) |
| 814 | { |
| 815 | int fd; |
| 816 | char *p; |
| 817 | struct stat st; |
| 818 | int ret; |
| 819 | struct trace *trace; |
| 820 | char *found_filename; |
| 821 | |
| 822 | trace = calloc(1, sizeof(*trace)); |
| 823 | if (!trace) { |
| 824 | fprintf(stderr, "unable to allocate memory for trace\n"); |
| 825 | return NULL; |
| 826 | } |
| 827 | |
| 828 | found_filename = find_trace_file(filename); |
| 829 | if (!found_filename) { |
| 830 | fprintf(stderr, "Unable to find trace file %s\n", filename); |
| 831 | goto fail; |
| 832 | } |
| 833 | filename = found_filename; |
| 834 | |
| 835 | fd = open(filename, O_RDONLY); |
| 836 | if (fd < 0) { |
| 837 | fprintf(stderr, "Unable to open trace file %s err %s\n", filename, strerror(errno)); |
| 838 | goto fail; |
| 839 | } |
| 840 | ret = fstat(fd, &st); |
| 841 | if (ret < 0) { |
| 842 | fprintf(stderr, "stat failed on %s err %s\n", filename, strerror(errno)); |
| 843 | goto fail_fd; |
| 844 | } |
| 845 | p = mmap(NULL, st.st_size, PROT_READ, MAP_PRIVATE, fd, 0); |
| 846 | if (p == MAP_FAILED) { |
| 847 | fprintf(stderr, "Unable to mmap trace file %s, err %s\n", filename, strerror(errno)); |
| 848 | goto fail_fd; |
| 849 | } |
| 850 | trace->fd = fd; |
| 851 | trace->len = st.st_size; |
| 852 | trace->start = p; |
| 853 | trace->cur = p; |
| 854 | trace->io = (struct blk_io_trace *)p; |
| 855 | return trace; |
| 856 | |
| 857 | fail_fd: |
| 858 | close(fd); |
| 859 | fail: |
| 860 | free(trace); |
| 861 | return NULL; |
| 862 | } |
| 863 | static inline int tput_event(struct trace *trace) |
| 864 | { |
| 865 | if (trace->found_completion) |
| 866 | return __BLK_TA_COMPLETE; |
| 867 | if (trace->found_issue) |
| 868 | return __BLK_TA_ISSUE; |
| 869 | if (trace->found_queue) |
| 870 | return __BLK_TA_QUEUE; |
| 871 | |
| 872 | return __BLK_TA_COMPLETE; |
| 873 | } |
| 874 | |
| 875 | int action_char_to_num(char action) |
| 876 | { |
| 877 | switch (action) { |
| 878 | case 'Q': |
| 879 | return __BLK_TA_QUEUE; |
| 880 | case 'D': |
| 881 | return __BLK_TA_ISSUE; |
| 882 | case 'C': |
| 883 | return __BLK_TA_COMPLETE; |
| 884 | } |
| 885 | return -1; |
| 886 | } |
| 887 | |
| 888 | static inline int io_event(struct trace *trace) |
| 889 | { |
| 890 | if (plot_io_action) |
| 891 | return plot_io_action; |
| 892 | if (trace->found_queue) |
| 893 | return __BLK_TA_QUEUE; |
| 894 | if (trace->found_issue) |
| 895 | return __BLK_TA_ISSUE; |
| 896 | if (trace->found_completion) |
| 897 | return __BLK_TA_COMPLETE; |
| 898 | |
| 899 | return __BLK_TA_COMPLETE; |
| 900 | } |
| 901 | |
| 902 | void add_tput(struct trace *trace, struct graph_line_data *writes_gld, |
| 903 | struct graph_line_data *reads_gld) |
| 904 | { |
| 905 | struct blk_io_trace *io = trace->io; |
| 906 | struct graph_line_data *gld; |
| 907 | int action = io->action & BLK_TA_MASK; |
| 908 | int seconds; |
| 909 | |
| 910 | if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY)) |
| 911 | return; |
| 912 | |
| 913 | if (action != tput_event(trace)) |
| 914 | return; |
| 915 | |
| 916 | if (BLK_DATADIR(io->action) & BLK_TC_READ) |
| 917 | gld = reads_gld; |
| 918 | else |
| 919 | gld = writes_gld; |
| 920 | |
| 921 | seconds = SECONDS(io->time); |
| 922 | gld->data[seconds].sum += io->bytes; |
| 923 | |
| 924 | gld->data[seconds].count = 1; |
| 925 | if (gld->data[seconds].sum > gld->max) |
| 926 | gld->max = gld->data[seconds].sum; |
| 927 | } |
| 928 | |
| 929 | #define GDD_PTR_ALLOC_STEP 16 |
| 930 | |
| 931 | static struct pid_map *get_pid_map(struct trace_file *tf, u32 pid) |
| 932 | { |
| 933 | struct pid_map *pm; |
| 934 | |
| 935 | if (!io_per_process) { |
| 936 | if (!tf->io_plots) |
| 937 | tf->io_plots = 1; |
| 938 | return NULL; |
| 939 | } |
| 940 | |
| 941 | pm = process_hash_insert(pid, NULL); |
| 942 | /* New entry? */ |
| 943 | if (!pm->index) { |
| 944 | if (tf->io_plots == tf->io_plots_allocated) { |
| 945 | tf->io_plots_allocated += GDD_PTR_ALLOC_STEP; |
| 946 | tf->gdd_reads = realloc(tf->gdd_reads, tf->io_plots_allocated * sizeof(struct graph_dot_data *)); |
| 947 | if (!tf->gdd_reads) |
| 948 | abort(); |
| 949 | tf->gdd_writes = realloc(tf->gdd_writes, tf->io_plots_allocated * sizeof(struct graph_dot_data *)); |
| 950 | if (!tf->gdd_writes) |
| 951 | abort(); |
| 952 | memset(tf->gdd_reads + tf->io_plots_allocated - GDD_PTR_ALLOC_STEP, |
| 953 | 0, GDD_PTR_ALLOC_STEP * sizeof(struct graph_dot_data *)); |
| 954 | memset(tf->gdd_writes + tf->io_plots_allocated - GDD_PTR_ALLOC_STEP, |
| 955 | 0, GDD_PTR_ALLOC_STEP * sizeof(struct graph_dot_data *)); |
| 956 | } |
| 957 | pm->index = tf->io_plots++; |
| 958 | |
| 959 | return pm; |
| 960 | } |
| 961 | return pm; |
| 962 | } |
| 963 | |
| 964 | void add_io(struct trace *trace, struct trace_file *tf) |
| 965 | { |
| 966 | struct blk_io_trace *io = trace->io; |
| 967 | int action = io->action & BLK_TA_MASK; |
| 968 | u64 offset; |
| 969 | int index; |
| 970 | char *label; |
| 971 | struct pid_map *pm; |
| 972 | |
| 973 | if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY)) |
| 974 | return; |
| 975 | |
| 976 | if (action != io_event(trace)) |
| 977 | return; |
| 978 | |
| 979 | offset = map_io(trace, io); |
| 980 | |
| 981 | pm = get_pid_map(tf, io->pid); |
| 982 | if (!pm) { |
| 983 | index = 0; |
| 984 | label = ""; |
| 985 | } else { |
| 986 | index = pm->index; |
| 987 | label = pm->name; |
| 988 | } |
| 989 | if (BLK_DATADIR(io->action) & BLK_TC_READ) { |
| 990 | if (!tf->gdd_reads[index]) |
| 991 | tf->gdd_reads[index] = alloc_dot_data(tf->min_seconds, tf->max_seconds, tf->min_offset, tf->max_offset, tf->stop_seconds, pick_color(), strdup(label)); |
| 992 | set_gdd_bit(tf->gdd_reads[index], offset, io->bytes, io->time); |
| 993 | } else if (BLK_DATADIR(io->action) & BLK_TC_WRITE) { |
| 994 | if (!tf->gdd_writes[index]) |
| 995 | tf->gdd_writes[index] = alloc_dot_data(tf->min_seconds, tf->max_seconds, tf->min_offset, tf->max_offset, tf->stop_seconds, pick_color(), strdup(label)); |
| 996 | set_gdd_bit(tf->gdd_writes[index], offset, io->bytes, io->time); |
| 997 | } |
| 998 | } |
| 999 | |
| 1000 | void add_pending_io(struct trace *trace, struct graph_line_data *gld) |
| 1001 | { |
| 1002 | unsigned int seconds; |
| 1003 | struct blk_io_trace *io = trace->io; |
| 1004 | int action = io->action & BLK_TA_MASK; |
| 1005 | double avg; |
| 1006 | struct pending_io *pio; |
| 1007 | |
| 1008 | if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY)) |
| 1009 | return; |
| 1010 | |
| 1011 | if (action == __BLK_TA_QUEUE) { |
| 1012 | if (io->sector == 0) |
| 1013 | return; |
| 1014 | /* |
| 1015 | * If D (issue) events are available, use them for I/O |
| 1016 | * accounting. Nothing needs to be done for Q. |
| 1017 | */ |
| 1018 | if (trace->found_issue) |
| 1019 | return; |
| 1020 | /* |
| 1021 | * If there are no D or C events, then all that can be |
| 1022 | * done is to account the Q event (and make sure not to |
| 1023 | * add the I/O to the hash, because it will never be |
| 1024 | * removed). |
| 1025 | */ |
| 1026 | if (!trace->found_completion) |
| 1027 | goto account_io; |
| 1028 | /* |
| 1029 | * When there are no ISSUE events, count depth and |
| 1030 | * latency from queue events. |
| 1031 | */ |
| 1032 | pio = hash_queued_io(trace->io); |
| 1033 | if (pio) { |
| 1034 | pio->dispatch_time = io->time; |
| 1035 | goto account_io; |
| 1036 | } |
| 1037 | return; |
| 1038 | } |
| 1039 | if (action == __BLK_TA_REQUEUE) { |
| 1040 | if (ios_in_flight > 0) |
| 1041 | ios_in_flight--; |
| 1042 | return; |
| 1043 | } |
| 1044 | if (action != __BLK_TA_ISSUE) |
| 1045 | return; |
| 1046 | |
| 1047 | pio = hash_dispatched_io(trace->io); |
| 1048 | if (!pio) |
| 1049 | return; |
| 1050 | |
| 1051 | if (!trace->found_completion) { |
| 1052 | list_del(&pio->hash_list); |
| 1053 | free(pio); |
| 1054 | } |
| 1055 | |
| 1056 | account_io: |
| 1057 | ios_in_flight++; |
| 1058 | |
| 1059 | seconds = SECONDS(io->time); |
| 1060 | gld->data[seconds].sum += ios_in_flight; |
| 1061 | gld->data[seconds].count++; |
| 1062 | |
| 1063 | avg = (double)gld->data[seconds].sum / gld->data[seconds].count; |
| 1064 | if (gld->max < (u64)avg) { |
| 1065 | gld->max = avg; |
| 1066 | } |
| 1067 | } |
| 1068 | |
| 1069 | void add_completed_io(struct trace *trace, |
| 1070 | struct graph_line_data *latency_gld) |
| 1071 | { |
| 1072 | struct blk_io_trace *io = trace->io; |
| 1073 | int seconds; |
| 1074 | int action = io->action & BLK_TA_MASK; |
| 1075 | struct pending_io *pio; |
| 1076 | double avg; |
| 1077 | u64 latency; |
| 1078 | |
| 1079 | if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY)) |
| 1080 | return; |
| 1081 | |
| 1082 | if (action != __BLK_TA_COMPLETE) |
| 1083 | return; |
| 1084 | |
| 1085 | seconds = SECONDS(io->time); |
| 1086 | |
| 1087 | pio = hash_completed_io(trace->io); |
| 1088 | if (!pio) |
| 1089 | return; |
| 1090 | |
| 1091 | if (ios_in_flight > 0) |
| 1092 | ios_in_flight--; |
| 1093 | if (io->time >= pio->dispatch_time) { |
| 1094 | latency = io->time - pio->dispatch_time; |
| 1095 | latency_gld->data[seconds].sum += latency; |
| 1096 | latency_gld->data[seconds].count++; |
| 1097 | } |
| 1098 | |
| 1099 | list_del(&pio->hash_list); |
| 1100 | free(pio); |
| 1101 | |
| 1102 | avg = (double)latency_gld->data[seconds].sum / |
| 1103 | latency_gld->data[seconds].count; |
| 1104 | if (latency_gld->max < (u64)avg) { |
| 1105 | latency_gld->max = avg; |
| 1106 | } |
| 1107 | } |
| 1108 | |
| 1109 | void add_iop(struct trace *trace, struct graph_line_data *gld) |
| 1110 | { |
| 1111 | struct blk_io_trace *io = trace->io; |
| 1112 | int action = io->action & BLK_TA_MASK; |
| 1113 | int seconds; |
| 1114 | |
| 1115 | if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY)) |
| 1116 | return; |
| 1117 | |
| 1118 | /* iops and tput use the same events */ |
| 1119 | if (action != tput_event(trace)) |
| 1120 | return; |
| 1121 | |
| 1122 | seconds = SECONDS(io->time); |
| 1123 | gld->data[seconds].sum += 1; |
| 1124 | gld->data[seconds].count = 1; |
| 1125 | if (gld->data[seconds].sum > gld->max) |
| 1126 | gld->max = gld->data[seconds].sum; |
| 1127 | } |
| 1128 | |
| 1129 | void check_record(struct trace *trace) |
| 1130 | { |
| 1131 | handle_notify(trace); |
| 1132 | } |