| 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., 59 Temple Place, Suite 330, Boston, MA 02111-1307 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 | |
| 34 | #include "plot.h" |
| 35 | #include "blkparse.h" |
| 36 | #include "list.h" |
| 37 | #include "tracers.h" |
| 38 | |
| 39 | #define IO_HASH_TABLE_BITS 11 |
| 40 | #define IO_HASH_TABLE_SIZE (1 << IO_HASH_TABLE_BITS) |
| 41 | static struct list_head io_hash_table[IO_HASH_TABLE_SIZE]; |
| 42 | static u64 ios_in_flight = 0; |
| 43 | |
| 44 | #define PROCESS_HASH_TABLE_BITS 7 |
| 45 | #define PROCESS_HASH_TABLE_SIZE (1 << PROCESS_HASH_TABLE_BITS) |
| 46 | static struct list_head process_hash_table[PROCESS_HASH_TABLE_SIZE]; |
| 47 | |
| 48 | extern int plot_io_action; |
| 49 | extern int io_per_process; |
| 50 | |
| 51 | /* |
| 52 | * Trace categories |
| 53 | */ |
| 54 | enum { |
| 55 | BLK_TC_READ = 1 << 0, /* reads */ |
| 56 | BLK_TC_WRITE = 1 << 1, /* writes */ |
| 57 | BLK_TC_FLUSH = 1 << 2, /* flush */ |
| 58 | BLK_TC_SYNC = 1 << 3, /* sync */ |
| 59 | BLK_TC_QUEUE = 1 << 4, /* queueing/merging */ |
| 60 | BLK_TC_REQUEUE = 1 << 5, /* requeueing */ |
| 61 | BLK_TC_ISSUE = 1 << 6, /* issue */ |
| 62 | BLK_TC_COMPLETE = 1 << 7, /* completions */ |
| 63 | BLK_TC_FS = 1 << 8, /* fs requests */ |
| 64 | BLK_TC_PC = 1 << 9, /* pc requests */ |
| 65 | BLK_TC_NOTIFY = 1 << 10, /* special message */ |
| 66 | BLK_TC_AHEAD = 1 << 11, /* readahead */ |
| 67 | BLK_TC_META = 1 << 12, /* metadata */ |
| 68 | BLK_TC_DISCARD = 1 << 13, /* discard requests */ |
| 69 | BLK_TC_DRV_DATA = 1 << 14, /* binary driver data */ |
| 70 | BLK_TC_FUA = 1 << 15, /* fua requests */ |
| 71 | |
| 72 | BLK_TC_END = 1 << 15, /* we've run out of bits! */ |
| 73 | }; |
| 74 | |
| 75 | #define BLK_TC_SHIFT (16) |
| 76 | #define BLK_TC_ACT(act) ((act) << BLK_TC_SHIFT) |
| 77 | #define BLK_DATADIR(a) (((a) >> BLK_TC_SHIFT) & (BLK_TC_READ | BLK_TC_WRITE)) |
| 78 | |
| 79 | /* |
| 80 | * Basic trace actions |
| 81 | */ |
| 82 | enum { |
| 83 | __BLK_TA_QUEUE = 1, /* queued */ |
| 84 | __BLK_TA_BACKMERGE, /* back merged to existing rq */ |
| 85 | __BLK_TA_FRONTMERGE, /* front merge to existing rq */ |
| 86 | __BLK_TA_GETRQ, /* allocated new request */ |
| 87 | __BLK_TA_SLEEPRQ, /* sleeping on rq allocation */ |
| 88 | __BLK_TA_REQUEUE, /* request requeued */ |
| 89 | __BLK_TA_ISSUE, /* sent to driver */ |
| 90 | __BLK_TA_COMPLETE, /* completed by driver */ |
| 91 | __BLK_TA_PLUG, /* queue was plugged */ |
| 92 | __BLK_TA_UNPLUG_IO, /* queue was unplugged by io */ |
| 93 | __BLK_TA_UNPLUG_TIMER, /* queue was unplugged by timer */ |
| 94 | __BLK_TA_INSERT, /* insert request */ |
| 95 | __BLK_TA_SPLIT, /* bio was split */ |
| 96 | __BLK_TA_BOUNCE, /* bio was bounced */ |
| 97 | __BLK_TA_REMAP, /* bio was remapped */ |
| 98 | __BLK_TA_ABORT, /* request aborted */ |
| 99 | __BLK_TA_DRV_DATA, /* binary driver data */ |
| 100 | }; |
| 101 | |
| 102 | #define BLK_TA_MASK ((1 << BLK_TC_SHIFT) - 1) |
| 103 | |
| 104 | /* |
| 105 | * Notify events. |
| 106 | */ |
| 107 | enum blktrace_notify { |
| 108 | __BLK_TN_PROCESS = 0, /* establish pid/name mapping */ |
| 109 | __BLK_TN_TIMESTAMP, /* include system clock */ |
| 110 | __BLK_TN_MESSAGE, /* Character string message */ |
| 111 | }; |
| 112 | |
| 113 | /* |
| 114 | * Trace actions in full. Additionally, read or write is masked |
| 115 | */ |
| 116 | #define BLK_TA_QUEUE (__BLK_TA_QUEUE | BLK_TC_ACT(BLK_TC_QUEUE)) |
| 117 | #define BLK_TA_BACKMERGE (__BLK_TA_BACKMERGE | BLK_TC_ACT(BLK_TC_QUEUE)) |
| 118 | #define BLK_TA_FRONTMERGE (__BLK_TA_FRONTMERGE | BLK_TC_ACT(BLK_TC_QUEUE)) |
| 119 | #define BLK_TA_GETRQ (__BLK_TA_GETRQ | BLK_TC_ACT(BLK_TC_QUEUE)) |
| 120 | #define BLK_TA_SLEEPRQ (__BLK_TA_SLEEPRQ | BLK_TC_ACT(BLK_TC_QUEUE)) |
| 121 | #define BLK_TA_REQUEUE (__BLK_TA_REQUEUE | BLK_TC_ACT(BLK_TC_REQUEUE)) |
| 122 | #define BLK_TA_ISSUE (__BLK_TA_ISSUE | BLK_TC_ACT(BLK_TC_ISSUE)) |
| 123 | #define BLK_TA_COMPLETE (__BLK_TA_COMPLETE| BLK_TC_ACT(BLK_TC_COMPLETE)) |
| 124 | #define BLK_TA_PLUG (__BLK_TA_PLUG | BLK_TC_ACT(BLK_TC_QUEUE)) |
| 125 | #define BLK_TA_UNPLUG_IO (__BLK_TA_UNPLUG_IO | BLK_TC_ACT(BLK_TC_QUEUE)) |
| 126 | #define BLK_TA_UNPLUG_TIMER (__BLK_TA_UNPLUG_TIMER | BLK_TC_ACT(BLK_TC_QUEUE)) |
| 127 | #define BLK_TA_INSERT (__BLK_TA_INSERT | BLK_TC_ACT(BLK_TC_QUEUE)) |
| 128 | #define BLK_TA_SPLIT (__BLK_TA_SPLIT) |
| 129 | #define BLK_TA_BOUNCE (__BLK_TA_BOUNCE) |
| 130 | #define BLK_TA_REMAP (__BLK_TA_REMAP | BLK_TC_ACT(BLK_TC_QUEUE)) |
| 131 | #define BLK_TA_ABORT (__BLK_TA_ABORT | BLK_TC_ACT(BLK_TC_QUEUE)) |
| 132 | #define BLK_TA_DRV_DATA (__BLK_TA_DRV_DATA | BLK_TC_ACT(BLK_TC_DRV_DATA)) |
| 133 | |
| 134 | #define BLK_TN_PROCESS (__BLK_TN_PROCESS | BLK_TC_ACT(BLK_TC_NOTIFY)) |
| 135 | #define BLK_TN_TIMESTAMP (__BLK_TN_TIMESTAMP | BLK_TC_ACT(BLK_TC_NOTIFY)) |
| 136 | #define BLK_TN_MESSAGE (__BLK_TN_MESSAGE | BLK_TC_ACT(BLK_TC_NOTIFY)) |
| 137 | |
| 138 | #define BLK_IO_TRACE_MAGIC 0x65617400 |
| 139 | #define BLK_IO_TRACE_VERSION 0x07 |
| 140 | /* |
| 141 | * The trace itself |
| 142 | */ |
| 143 | struct blk_io_trace { |
| 144 | __u32 magic; /* MAGIC << 8 | version */ |
| 145 | __u32 sequence; /* event number */ |
| 146 | __u64 time; /* in nanoseconds */ |
| 147 | __u64 sector; /* disk offset */ |
| 148 | __u32 bytes; /* transfer length */ |
| 149 | __u32 action; /* what happened */ |
| 150 | __u32 pid; /* who did it */ |
| 151 | __u32 device; /* device identifier (dev_t) */ |
| 152 | __u32 cpu; /* on what cpu did it happen */ |
| 153 | __u16 error; /* completion error */ |
| 154 | __u16 pdu_len; /* length of data after this trace */ |
| 155 | }; |
| 156 | |
| 157 | struct pending_io { |
| 158 | /* sector offset of this IO */ |
| 159 | u64 sector; |
| 160 | |
| 161 | /* time this IO was dispatched */ |
| 162 | u64 dispatch_time; |
| 163 | /* time this IO was finished */ |
| 164 | u64 completion_time; |
| 165 | struct list_head hash_list; |
| 166 | /* process which queued this IO */ |
| 167 | u32 pid; |
| 168 | }; |
| 169 | |
| 170 | struct pid_map { |
| 171 | struct list_head hash_list; |
| 172 | u32 pid; |
| 173 | int index; |
| 174 | char name[0]; |
| 175 | }; |
| 176 | |
| 177 | #define MINORBITS 20 |
| 178 | #define MINORMASK ((1 << MINORBITS) - 1) |
| 179 | #define SECONDS(x) ((unsigned long long)(x) / 1000000000) |
| 180 | #define NANO_SECONDS(x) ((unsigned long long)(x) % 1000000000) |
| 181 | #define DOUBLE_TO_NANO_ULL(d) ((unsigned long long)((d) * 1000000000)) |
| 182 | #define CHECK_MAGIC(t) (((t)->magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) |
| 183 | |
| 184 | void init_io_hash_table(void) |
| 185 | { |
| 186 | int i; |
| 187 | struct list_head *head; |
| 188 | |
| 189 | for (i = 0; i < IO_HASH_TABLE_SIZE; i++) { |
| 190 | head = io_hash_table + i; |
| 191 | INIT_LIST_HEAD(head); |
| 192 | } |
| 193 | } |
| 194 | |
| 195 | /* taken from the kernel hash.h */ |
| 196 | static inline u64 hash_sector(u64 val) |
| 197 | { |
| 198 | u64 hash = val; |
| 199 | |
| 200 | /* Sigh, gcc can't optimise this alone like it does for 32 bits. */ |
| 201 | u64 n = hash; |
| 202 | n <<= 18; |
| 203 | hash -= n; |
| 204 | n <<= 33; |
| 205 | hash -= n; |
| 206 | n <<= 3; |
| 207 | hash += n; |
| 208 | n <<= 3; |
| 209 | hash -= n; |
| 210 | n <<= 4; |
| 211 | hash += n; |
| 212 | n <<= 2; |
| 213 | hash += n; |
| 214 | |
| 215 | /* High bits are more random, so use them. */ |
| 216 | return hash >> (64 - IO_HASH_TABLE_BITS); |
| 217 | } |
| 218 | |
| 219 | static int io_hash_table_insert(struct pending_io *ins_pio) |
| 220 | { |
| 221 | u64 sector = ins_pio->sector; |
| 222 | int slot = hash_sector(sector); |
| 223 | struct list_head *head; |
| 224 | struct pending_io *pio; |
| 225 | |
| 226 | head = io_hash_table + slot; |
| 227 | list_for_each_entry(pio, head, hash_list) { |
| 228 | if (pio->sector == sector) |
| 229 | return -EEXIST; |
| 230 | } |
| 231 | list_add_tail(&ins_pio->hash_list, head); |
| 232 | return 0; |
| 233 | } |
| 234 | |
| 235 | static struct pending_io *io_hash_table_search(u64 sector) |
| 236 | { |
| 237 | int slot = hash_sector(sector); |
| 238 | struct list_head *head; |
| 239 | struct pending_io *pio; |
| 240 | |
| 241 | head = io_hash_table + slot; |
| 242 | list_for_each_entry(pio, head, hash_list) { |
| 243 | if (pio->sector == sector) |
| 244 | return pio; |
| 245 | } |
| 246 | return NULL; |
| 247 | } |
| 248 | |
| 249 | static int hash_queued_io(struct blk_io_trace *io) |
| 250 | { |
| 251 | struct pending_io *pio; |
| 252 | int ret; |
| 253 | |
| 254 | pio = calloc(1, sizeof(*pio)); |
| 255 | pio->sector = io->sector; |
| 256 | pio->pid = io->pid; |
| 257 | |
| 258 | ret = io_hash_table_insert(pio); |
| 259 | if (ret < 0) { |
| 260 | /* crud, the IO is there already */ |
| 261 | free(pio); |
| 262 | return ret; |
| 263 | } |
| 264 | return 0; |
| 265 | } |
| 266 | |
| 267 | static struct pending_io *hash_dispatched_io(struct blk_io_trace *io) |
| 268 | { |
| 269 | struct pending_io *pio; |
| 270 | |
| 271 | pio = io_hash_table_search(io->sector); |
| 272 | if (!pio) |
| 273 | hash_queued_io(io); |
| 274 | pio->dispatch_time = io->time; |
| 275 | return pio; |
| 276 | } |
| 277 | |
| 278 | static struct pending_io *hash_completed_io(struct blk_io_trace *io) |
| 279 | { |
| 280 | struct pending_io *pio; |
| 281 | |
| 282 | pio = io_hash_table_search(io->sector); |
| 283 | |
| 284 | if (!pio) |
| 285 | return NULL; |
| 286 | return pio; |
| 287 | } |
| 288 | |
| 289 | void init_process_hash_table(void) |
| 290 | { |
| 291 | int i; |
| 292 | struct list_head *head; |
| 293 | |
| 294 | for (i = 0; i < PROCESS_HASH_TABLE_SIZE; i++) { |
| 295 | head = process_hash_table + i; |
| 296 | INIT_LIST_HEAD(head); |
| 297 | } |
| 298 | } |
| 299 | |
| 300 | static u32 hash_pid(u32 pid) |
| 301 | { |
| 302 | u32 hash = pid; |
| 303 | |
| 304 | hash ^= pid >> 3; |
| 305 | hash ^= pid >> 3; |
| 306 | hash ^= pid >> 4; |
| 307 | hash ^= pid >> 6; |
| 308 | return (hash & (PROCESS_HASH_TABLE_SIZE - 1)); |
| 309 | } |
| 310 | |
| 311 | static struct pid_map *process_hash_search(u32 pid) |
| 312 | { |
| 313 | int slot = hash_pid(pid); |
| 314 | struct list_head *head; |
| 315 | struct pid_map *pm; |
| 316 | |
| 317 | head = process_hash_table + slot; |
| 318 | list_for_each_entry(pm, head, hash_list) { |
| 319 | if (pm->pid == pid) |
| 320 | return pm; |
| 321 | } |
| 322 | return NULL; |
| 323 | } |
| 324 | |
| 325 | static struct pid_map *process_hash_insert(u32 pid, char *name) |
| 326 | { |
| 327 | int slot = hash_pid(pid); |
| 328 | struct pid_map *pm; |
| 329 | int old_index = 0; |
| 330 | char buf[16]; |
| 331 | |
| 332 | pm = process_hash_search(pid); |
| 333 | if (pm) { |
| 334 | /* Entry exists and name shouldn't be changed? */ |
| 335 | if (!name || !strcmp(name, pm->name)) |
| 336 | return pm; |
| 337 | list_del(&pm->hash_list); |
| 338 | old_index = pm->index; |
| 339 | free(pm); |
| 340 | } |
| 341 | if (!name) { |
| 342 | sprintf(buf, "[%u]", pid); |
| 343 | name = buf; |
| 344 | } |
| 345 | pm = malloc(sizeof(struct pid_map) + strlen(name) + 1); |
| 346 | pm->pid = pid; |
| 347 | pm->index = old_index; |
| 348 | strcpy(pm->name, name); |
| 349 | list_add_tail(&pm->hash_list, process_hash_table + slot); |
| 350 | |
| 351 | return pm; |
| 352 | } |
| 353 | |
| 354 | static void handle_notify(struct trace *trace) |
| 355 | { |
| 356 | struct blk_io_trace *io = trace->io; |
| 357 | void *payload = (char *)io + sizeof(*io); |
| 358 | u32 two32[2]; |
| 359 | |
| 360 | if (io->action == BLK_TN_PROCESS) { |
| 361 | if (io_per_process) |
| 362 | process_hash_insert(io->pid, payload); |
| 363 | return; |
| 364 | } |
| 365 | |
| 366 | if (io->action != BLK_TN_TIMESTAMP) |
| 367 | return; |
| 368 | |
| 369 | if (io->pdu_len != sizeof(two32)) |
| 370 | return; |
| 371 | |
| 372 | memcpy(two32, payload, sizeof(two32)); |
| 373 | trace->start_timestamp = io->time; |
| 374 | trace->abs_start_time.tv_sec = two32[0]; |
| 375 | trace->abs_start_time.tv_nsec = two32[1]; |
| 376 | if (trace->abs_start_time.tv_nsec < 0) { |
| 377 | trace->abs_start_time.tv_sec--; |
| 378 | trace->abs_start_time.tv_nsec += 1000000000; |
| 379 | } |
| 380 | } |
| 381 | |
| 382 | int next_record(struct trace *trace) |
| 383 | { |
| 384 | int skip = trace->io->pdu_len; |
| 385 | u64 offset; |
| 386 | |
| 387 | trace->cur += sizeof(*trace->io) + skip; |
| 388 | offset = trace->cur - trace->start; |
| 389 | if (offset >= trace->len) |
| 390 | return 1; |
| 391 | |
| 392 | trace->io = (struct blk_io_trace *)trace->cur; |
| 393 | return 0; |
| 394 | } |
| 395 | |
| 396 | void first_record(struct trace *trace) |
| 397 | { |
| 398 | trace->cur = trace->start; |
| 399 | trace->io = (struct blk_io_trace *)trace->cur; |
| 400 | } |
| 401 | |
| 402 | int is_io_event(struct blk_io_trace *test) |
| 403 | { |
| 404 | char *message; |
| 405 | if (!(test->action & BLK_TC_ACT(BLK_TC_NOTIFY))) |
| 406 | return 1; |
| 407 | if (test->action == BLK_TN_MESSAGE) { |
| 408 | int len = test->pdu_len; |
| 409 | if (len < 3) |
| 410 | return 0; |
| 411 | message = (char *)(test + 1); |
| 412 | if (strncmp(message, "fio ", 4) == 0) { |
| 413 | return 1; |
| 414 | } |
| 415 | } |
| 416 | return 0; |
| 417 | } |
| 418 | |
| 419 | u64 find_last_time(struct trace *trace) |
| 420 | { |
| 421 | char *p = trace->start + trace->len; |
| 422 | struct blk_io_trace *test; |
| 423 | int search_len = 0; |
| 424 | u64 found = 0; |
| 425 | |
| 426 | if (trace->len < sizeof(*trace->io)) |
| 427 | return 0; |
| 428 | p -= sizeof(*trace->io); |
| 429 | while (p >= trace->start) { |
| 430 | test = (struct blk_io_trace *)p; |
| 431 | if (CHECK_MAGIC(test) && is_io_event(test)) { |
| 432 | u64 offset = p - trace->start; |
| 433 | if (offset + sizeof(*test) + test->pdu_len == trace->len) { |
| 434 | return test->time; |
| 435 | } |
| 436 | } |
| 437 | p--; |
| 438 | search_len++; |
| 439 | if (search_len > 8192) { |
| 440 | break; |
| 441 | } |
| 442 | } |
| 443 | |
| 444 | /* searching backwards didn't work out, we'll have to scan the file */ |
| 445 | first_record(trace); |
| 446 | while (1) { |
| 447 | if (is_io_event(trace->io)) |
| 448 | found = trace->io->time; |
| 449 | if (next_record(trace)) |
| 450 | break; |
| 451 | } |
| 452 | first_record(trace); |
| 453 | return found; |
| 454 | } |
| 455 | |
| 456 | int parse_fio_bank_message(struct trace *trace, u64 *bank_ret, u64 *offset_ret, |
| 457 | u64 *num_banks_ret) |
| 458 | { |
| 459 | char *s; |
| 460 | char *next; |
| 461 | char *message; |
| 462 | struct blk_io_trace *test = trace->io; |
| 463 | int len = test->pdu_len; |
| 464 | u64 bank; |
| 465 | u64 offset; |
| 466 | u64 num_banks; |
| 467 | |
| 468 | if (!(test->action & BLK_TC_ACT(BLK_TC_NOTIFY))) |
| 469 | return -1; |
| 470 | if (test->action != BLK_TN_MESSAGE) |
| 471 | return -1; |
| 472 | |
| 473 | /* the message is fio rw bank offset num_banks */ |
| 474 | if (len < 3) |
| 475 | return -1; |
| 476 | message = (char *)(test + 1); |
| 477 | if (strncmp(message, "fio r ", 6) != 0) |
| 478 | return -1; |
| 479 | |
| 480 | message = strndup(message, len); |
| 481 | s = strchr(message, ' '); |
| 482 | if (!s) |
| 483 | goto out; |
| 484 | s++; |
| 485 | s = strchr(s, ' '); |
| 486 | if (!s) |
| 487 | goto out; |
| 488 | |
| 489 | bank = strtoll(s, &next, 10); |
| 490 | if (s == next) |
| 491 | goto out; |
| 492 | s = next; |
| 493 | |
| 494 | offset = strtoll(s, &next, 10); |
| 495 | if (s == next) |
| 496 | goto out; |
| 497 | s = next; |
| 498 | |
| 499 | num_banks = strtoll(s, &next, 10); |
| 500 | if (s == next) |
| 501 | goto out; |
| 502 | |
| 503 | *bank_ret = bank; |
| 504 | *offset_ret = offset; |
| 505 | *num_banks_ret = num_banks; |
| 506 | |
| 507 | return 0; |
| 508 | out: |
| 509 | free(message); |
| 510 | return -1; |
| 511 | } |
| 512 | |
| 513 | void find_extreme_offsets(struct trace *trace, u64 *min_ret, u64 *max_ret, u64 *max_bank_ret, |
| 514 | u64 *max_offset_ret) |
| 515 | { |
| 516 | u64 found = 0; |
| 517 | u64 max = 0, min = ~(u64)0; |
| 518 | u64 max_bank = 0; |
| 519 | u64 max_bank_offset = 0; |
| 520 | u64 num_banks = 0; |
| 521 | first_record(trace); |
| 522 | while (1) { |
| 523 | if (!(trace->io->action & BLK_TC_ACT(BLK_TC_NOTIFY))) { |
| 524 | found = trace->io->sector << 9; |
| 525 | if (found < min) |
| 526 | min = found; |
| 527 | |
| 528 | found += trace->io->bytes; |
| 529 | if (max < found) |
| 530 | max = found; |
| 531 | } else { |
| 532 | u64 bank; |
| 533 | u64 offset; |
| 534 | if (!parse_fio_bank_message(trace, &bank, |
| 535 | &offset, &num_banks)) { |
| 536 | if (bank > max_bank) |
| 537 | max_bank = bank; |
| 538 | if (offset > max_bank_offset) |
| 539 | max_bank_offset = offset; |
| 540 | } |
| 541 | } |
| 542 | if (next_record(trace)) |
| 543 | break; |
| 544 | } |
| 545 | first_record(trace); |
| 546 | *min_ret = min; |
| 547 | *max_ret = max; |
| 548 | *max_bank_ret = max_bank; |
| 549 | *max_offset_ret = max_bank_offset; |
| 550 | } |
| 551 | |
| 552 | static void check_io_types(struct trace *trace) |
| 553 | { |
| 554 | struct blk_io_trace *io = trace->io; |
| 555 | int action = io->action & BLK_TA_MASK; |
| 556 | |
| 557 | if (!(io->action & BLK_TC_ACT(BLK_TC_NOTIFY))) { |
| 558 | switch (action) { |
| 559 | case __BLK_TA_COMPLETE: |
| 560 | trace->found_completion = 1; |
| 561 | break; |
| 562 | case __BLK_TA_ISSUE: |
| 563 | trace->found_issue = 1; |
| 564 | break; |
| 565 | case __BLK_TA_QUEUE: |
| 566 | trace->found_queue = 1; |
| 567 | break; |
| 568 | }; |
| 569 | } |
| 570 | } |
| 571 | |
| 572 | |
| 573 | int filter_outliers(struct trace *trace, u64 min_offset, u64 max_offset, |
| 574 | u64 *yzoom_min, u64 *yzoom_max) |
| 575 | { |
| 576 | int hits[11]; |
| 577 | u64 max_per_bucket[11]; |
| 578 | u64 min_per_bucket[11]; |
| 579 | u64 bytes_per_bucket = (max_offset - min_offset + 1) / 10; |
| 580 | int slot; |
| 581 | int fat_count = 0; |
| 582 | |
| 583 | memset(hits, 0, sizeof(int) * 11); |
| 584 | memset(max_per_bucket, 0, sizeof(u64) * 11); |
| 585 | memset(min_per_bucket, 0xff, sizeof(u64) * 11); |
| 586 | first_record(trace); |
| 587 | while (1) { |
| 588 | check_io_types(trace); |
| 589 | if (!(trace->io->action & BLK_TC_ACT(BLK_TC_NOTIFY)) && |
| 590 | (trace->io->action & BLK_TA_MASK) == __BLK_TA_QUEUE) { |
| 591 | u64 off = (trace->io->sector << 9) - min_offset; |
| 592 | |
| 593 | slot = (int)(off / bytes_per_bucket); |
| 594 | hits[slot]++; |
| 595 | if (off < min_per_bucket[slot]) |
| 596 | min_per_bucket[slot] = off; |
| 597 | |
| 598 | off += trace->io->bytes; |
| 599 | slot = (int)(off / bytes_per_bucket); |
| 600 | hits[slot]++; |
| 601 | if (off > max_per_bucket[slot]) |
| 602 | max_per_bucket[slot] = off; |
| 603 | } |
| 604 | if (next_record(trace)) |
| 605 | break; |
| 606 | } |
| 607 | first_record(trace); |
| 608 | for (slot = 0; slot < 11; slot++) { |
| 609 | if (hits[slot] > fat_count) { |
| 610 | fat_count = hits[slot]; |
| 611 | } |
| 612 | } |
| 613 | |
| 614 | *yzoom_max = max_offset; |
| 615 | for (slot = 10; slot >= 0; slot--) { |
| 616 | double d = hits[slot]; |
| 617 | |
| 618 | if (d >= (double)fat_count * .05) { |
| 619 | *yzoom_max = max_per_bucket[slot] + min_offset; |
| 620 | break; |
| 621 | } |
| 622 | } |
| 623 | |
| 624 | *yzoom_min = min_offset; |
| 625 | for (slot = 0; slot < 10; slot++) { |
| 626 | double d = hits[slot]; |
| 627 | |
| 628 | if (d >= (double)fat_count * .05) { |
| 629 | *yzoom_min = min_per_bucket[slot] + min_offset; |
| 630 | break; |
| 631 | } |
| 632 | } |
| 633 | return 0; |
| 634 | } |
| 635 | |
| 636 | static char *find_trace_file(char *filename) |
| 637 | { |
| 638 | int ret; |
| 639 | struct stat st; |
| 640 | char line[1024]; |
| 641 | char *dot; |
| 642 | char *try; |
| 643 | |
| 644 | ret = stat(filename, &st); |
| 645 | if (ret == 0) |
| 646 | return strdup(filename); |
| 647 | |
| 648 | snprintf(line, 1024, "%s.%s", filename, "dump"); |
| 649 | ret = stat(line, &st); |
| 650 | if (ret == 0) |
| 651 | return strdup(line); |
| 652 | |
| 653 | try = strdup(filename); |
| 654 | dot = strrchr(try, '.'); |
| 655 | if (!dot || strcmp(".dump", dot) != 0) { |
| 656 | if (dot && dot != try) |
| 657 | *dot = '\0'; |
| 658 | snprintf(line, 1024, "%s%s", try, ".blktrace.0"); |
| 659 | ret = stat(line, &st); |
| 660 | if (ret == 0) { |
| 661 | blktrace_to_dump(try); |
| 662 | snprintf(line, 1024, "%s.%s", try, "dump"); |
| 663 | ret = stat(line, &st); |
| 664 | if (ret == 0) { |
| 665 | free(try); |
| 666 | return strdup(line); |
| 667 | } |
| 668 | } |
| 669 | } |
| 670 | free(try); |
| 671 | return NULL; |
| 672 | } |
| 673 | struct trace *open_trace(char *filename) |
| 674 | { |
| 675 | int fd; |
| 676 | char *p; |
| 677 | struct stat st; |
| 678 | int ret; |
| 679 | struct trace *trace; |
| 680 | char *found_filename; |
| 681 | |
| 682 | trace = calloc(1, sizeof(*trace)); |
| 683 | if (!trace) { |
| 684 | fprintf(stderr, "unable to allocate memory for trace\n"); |
| 685 | return NULL; |
| 686 | } |
| 687 | |
| 688 | found_filename = find_trace_file(filename); |
| 689 | if (!found_filename) { |
| 690 | fprintf(stderr, "Unable to find trace file %s\n", filename); |
| 691 | goto fail; |
| 692 | } |
| 693 | filename = found_filename; |
| 694 | |
| 695 | fd = open(filename, O_RDONLY); |
| 696 | if (fd < 0) { |
| 697 | fprintf(stderr, "Unable to open trace file %s err %s\n", filename, strerror(errno)); |
| 698 | goto fail; |
| 699 | } |
| 700 | ret = fstat(fd, &st); |
| 701 | if (ret < 0) { |
| 702 | fprintf(stderr, "stat failed on %s err %s\n", filename, strerror(errno)); |
| 703 | goto fail_fd; |
| 704 | } |
| 705 | p = mmap(NULL, st.st_size, PROT_READ, MAP_PRIVATE, fd, 0); |
| 706 | if (p == MAP_FAILED) { |
| 707 | fprintf(stderr, "Unable to mmap trace file %s, err %s\n", filename, strerror(errno)); |
| 708 | goto fail_fd; |
| 709 | } |
| 710 | trace->fd = fd; |
| 711 | trace->len = st.st_size; |
| 712 | trace->start = p; |
| 713 | trace->cur = p; |
| 714 | trace->io = (struct blk_io_trace *)p; |
| 715 | return trace; |
| 716 | |
| 717 | fail_fd: |
| 718 | close(fd); |
| 719 | fail: |
| 720 | free(trace); |
| 721 | return NULL; |
| 722 | } |
| 723 | static inline int tput_event(struct trace *trace) |
| 724 | { |
| 725 | if (trace->found_completion) |
| 726 | return __BLK_TA_COMPLETE; |
| 727 | if (trace->found_issue) |
| 728 | return __BLK_TA_ISSUE; |
| 729 | if (trace->found_queue) |
| 730 | return __BLK_TA_QUEUE; |
| 731 | |
| 732 | return __BLK_TA_COMPLETE; |
| 733 | } |
| 734 | |
| 735 | int action_char_to_num(char action) |
| 736 | { |
| 737 | switch (action) { |
| 738 | case 'Q': |
| 739 | return __BLK_TA_QUEUE; |
| 740 | case 'D': |
| 741 | return __BLK_TA_ISSUE; |
| 742 | case 'C': |
| 743 | return __BLK_TA_COMPLETE; |
| 744 | } |
| 745 | return -1; |
| 746 | } |
| 747 | |
| 748 | static inline int io_event(struct trace *trace) |
| 749 | { |
| 750 | if (plot_io_action) |
| 751 | return plot_io_action; |
| 752 | if (trace->found_queue) |
| 753 | return __BLK_TA_QUEUE; |
| 754 | if (trace->found_issue) |
| 755 | return __BLK_TA_ISSUE; |
| 756 | if (trace->found_completion) |
| 757 | return __BLK_TA_COMPLETE; |
| 758 | |
| 759 | return __BLK_TA_COMPLETE; |
| 760 | } |
| 761 | |
| 762 | void add_tput(struct trace *trace, struct graph_line_data *gld) |
| 763 | { |
| 764 | struct blk_io_trace *io = trace->io; |
| 765 | int action = io->action & BLK_TA_MASK; |
| 766 | int seconds; |
| 767 | |
| 768 | if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY)) |
| 769 | return; |
| 770 | |
| 771 | if (action != tput_event(trace)) |
| 772 | return; |
| 773 | |
| 774 | seconds = SECONDS(io->time); |
| 775 | if (seconds > gld->max_seconds) |
| 776 | return; |
| 777 | |
| 778 | gld->data[seconds].sum += io->bytes; |
| 779 | gld->data[seconds].count = 1; |
| 780 | if (gld->data[seconds].sum > gld->max) |
| 781 | gld->max = gld->data[seconds].sum; |
| 782 | } |
| 783 | |
| 784 | #define GDD_PTR_ALLOC_STEP 16 |
| 785 | |
| 786 | static struct pid_map *get_pid_map(struct trace_file *tf, u32 pid) |
| 787 | { |
| 788 | struct pid_map *pm; |
| 789 | |
| 790 | if (!io_per_process) { |
| 791 | if (!tf->io_plots) |
| 792 | tf->io_plots = 1; |
| 793 | return NULL; |
| 794 | } |
| 795 | |
| 796 | pm = process_hash_insert(pid, NULL); |
| 797 | /* New entry? */ |
| 798 | if (!pm->index) { |
| 799 | if (tf->io_plots == tf->io_plots_allocated) { |
| 800 | tf->io_plots_allocated += GDD_PTR_ALLOC_STEP; |
| 801 | tf->gdd_reads = realloc(tf->gdd_reads, tf->io_plots_allocated * sizeof(struct graph_dot_data *)); |
| 802 | if (!tf->gdd_reads) |
| 803 | abort(); |
| 804 | tf->gdd_writes = realloc(tf->gdd_writes, tf->io_plots_allocated * sizeof(struct graph_dot_data *)); |
| 805 | if (!tf->gdd_writes) |
| 806 | abort(); |
| 807 | memset(tf->gdd_reads + tf->io_plots_allocated - GDD_PTR_ALLOC_STEP, |
| 808 | 0, GDD_PTR_ALLOC_STEP * sizeof(struct graph_dot_data *)); |
| 809 | memset(tf->gdd_writes + tf->io_plots_allocated - GDD_PTR_ALLOC_STEP, |
| 810 | 0, GDD_PTR_ALLOC_STEP * sizeof(struct graph_dot_data *)); |
| 811 | } |
| 812 | pm->index = tf->io_plots++; |
| 813 | |
| 814 | return pm; |
| 815 | } |
| 816 | return pm; |
| 817 | } |
| 818 | |
| 819 | void add_io(struct trace *trace, struct trace_file *tf) |
| 820 | { |
| 821 | struct blk_io_trace *io = trace->io; |
| 822 | int action = io->action & BLK_TA_MASK; |
| 823 | u64 offset; |
| 824 | int index; |
| 825 | char *label; |
| 826 | struct pid_map *pm; |
| 827 | |
| 828 | if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY)) |
| 829 | return; |
| 830 | |
| 831 | if (action != io_event(trace)) |
| 832 | return; |
| 833 | |
| 834 | offset = io->sector << 9; |
| 835 | |
| 836 | pm = get_pid_map(tf, io->pid); |
| 837 | if (!pm) { |
| 838 | index = 0; |
| 839 | label = ""; |
| 840 | } else { |
| 841 | index = pm->index; |
| 842 | label = pm->name; |
| 843 | } |
| 844 | if (BLK_DATADIR(io->action) & BLK_TC_READ) { |
| 845 | if (!tf->gdd_reads[index]) |
| 846 | 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)); |
| 847 | set_gdd_bit(tf->gdd_reads[index], offset, io->bytes, io->time); |
| 848 | } else if (BLK_DATADIR(io->action) & BLK_TC_WRITE) { |
| 849 | if (!tf->gdd_writes[index]) |
| 850 | 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)); |
| 851 | set_gdd_bit(tf->gdd_writes[index], offset, io->bytes, io->time); |
| 852 | } |
| 853 | } |
| 854 | |
| 855 | void add_pending_io(struct trace *trace, struct graph_line_data *gld) |
| 856 | { |
| 857 | int seconds; |
| 858 | struct blk_io_trace *io = trace->io; |
| 859 | int action = io->action & BLK_TA_MASK; |
| 860 | double avg; |
| 861 | struct pending_io *pio; |
| 862 | |
| 863 | if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY)) |
| 864 | return; |
| 865 | |
| 866 | if (action == __BLK_TA_QUEUE) { |
| 867 | if (trace->found_issue || trace->found_completion) |
| 868 | hash_queued_io(trace->io); |
| 869 | return; |
| 870 | } |
| 871 | if (action != __BLK_TA_ISSUE) |
| 872 | return; |
| 873 | |
| 874 | seconds = SECONDS(io->time); |
| 875 | if (seconds > gld->max_seconds) |
| 876 | return; |
| 877 | |
| 878 | pio = hash_dispatched_io(trace->io); |
| 879 | if (!pio) |
| 880 | return; |
| 881 | |
| 882 | if (!trace->found_completion) { |
| 883 | list_del(&pio->hash_list); |
| 884 | free(pio); |
| 885 | } |
| 886 | |
| 887 | ios_in_flight++; |
| 888 | |
| 889 | gld->data[seconds].sum += ios_in_flight; |
| 890 | gld->data[seconds].count++; |
| 891 | |
| 892 | avg = (double)gld->data[seconds].sum / gld->data[seconds].count; |
| 893 | if (gld->max < (u64)avg) { |
| 894 | gld->max = avg; |
| 895 | } |
| 896 | } |
| 897 | |
| 898 | void add_completed_io(struct trace *trace, |
| 899 | struct graph_line_data *latency_gld) |
| 900 | { |
| 901 | struct blk_io_trace *io = trace->io; |
| 902 | int seconds; |
| 903 | int action = io->action & BLK_TA_MASK; |
| 904 | struct pending_io *pio; |
| 905 | double avg; |
| 906 | u64 latency; |
| 907 | |
| 908 | if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY)) |
| 909 | return; |
| 910 | |
| 911 | if (action != __BLK_TA_COMPLETE) |
| 912 | return; |
| 913 | |
| 914 | seconds = SECONDS(io->time); |
| 915 | |
| 916 | pio = hash_completed_io(trace->io); |
| 917 | if (!pio) |
| 918 | return; |
| 919 | |
| 920 | if (ios_in_flight > 0) |
| 921 | ios_in_flight--; |
| 922 | if (io->time >= pio->dispatch_time) { |
| 923 | latency = io->time - pio->dispatch_time; |
| 924 | latency_gld->data[seconds].sum += latency; |
| 925 | latency_gld->data[seconds].count++; |
| 926 | } |
| 927 | |
| 928 | list_del(&pio->hash_list); |
| 929 | free(pio); |
| 930 | |
| 931 | avg = (double)latency_gld->data[seconds].sum / |
| 932 | latency_gld->data[seconds].count; |
| 933 | if (latency_gld->max < (u64)avg) { |
| 934 | latency_gld->max = avg; |
| 935 | } |
| 936 | } |
| 937 | |
| 938 | void add_iop(struct trace *trace, struct graph_line_data *gld) |
| 939 | { |
| 940 | struct blk_io_trace *io = trace->io; |
| 941 | int action = io->action & BLK_TA_MASK; |
| 942 | int seconds; |
| 943 | |
| 944 | if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY)) |
| 945 | return; |
| 946 | |
| 947 | /* iops and tput use the same events */ |
| 948 | if (action != tput_event(trace)) |
| 949 | return; |
| 950 | |
| 951 | seconds = SECONDS(io->time); |
| 952 | if (seconds > gld->max_seconds) |
| 953 | return; |
| 954 | |
| 955 | gld->data[seconds].sum += 1; |
| 956 | gld->data[seconds].count = 1; |
| 957 | if (gld->data[seconds].sum > gld->max) |
| 958 | gld->max = gld->data[seconds].sum; |
| 959 | } |
| 960 | |
| 961 | void check_record(struct trace *trace) |
| 962 | { |
| 963 | handle_notify(trace); |
| 964 | } |