iowatcher/blkparse.c

   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
  40 #define IO_HASH_TABLE_BITS  11
  41 #define IO_HASH_TABLE_SIZE (1 << IO_HASH_TABLE_BITS)
  42 static struct list_head io_hash_table[IO_HASH_TABLE_SIZE];
  43 static u64 ios_in_flight = 0;
  44
  45 #define PROCESS_HASH_TABLE_BITS 7
  46 #define PROCESS_HASH_TABLE_SIZE (1 << PROCESS_HASH_TABLE_BITS)
  47 static struct list_head process_hash_table[PROCESS_HASH_TABLE_SIZE];
  48
  49 extern int plot_io_action;
  50 extern int io_per_process;
  51
  52 static const int line_len = 1024;
  53 static char line[1024];
  54
  55 /*
  56  * Trace categories
  57  */
  58 enum {
  59         BLK_TC_READ     = 1 << 0,       /* reads */
  60         BLK_TC_WRITE    = 1 << 1,       /* writes */
  61         BLK_TC_FLUSH    = 1 << 2,       /* flush */
  62         BLK_TC_SYNC     = 1 << 3,       /* sync */
  63         BLK_TC_QUEUE    = 1 << 4,       /* queueing/merging */
  64         BLK_TC_REQUEUE  = 1 << 5,       /* requeueing */
  65         BLK_TC_ISSUE    = 1 << 6,       /* issue */
  66         BLK_TC_COMPLETE = 1 << 7,       /* completions */
  67         BLK_TC_FS       = 1 << 8,       /* fs requests */
  68         BLK_TC_PC       = 1 << 9,       /* pc requests */
  69         BLK_TC_NOTIFY   = 1 << 10,      /* special message */
  70         BLK_TC_AHEAD    = 1 << 11,      /* readahead */
  71         BLK_TC_META     = 1 << 12,      /* metadata */
  72         BLK_TC_DISCARD  = 1 << 13,      /* discard requests */
  73         BLK_TC_DRV_DATA = 1 << 14,      /* binary driver data */
  74         BLK_TC_FUA      = 1 << 15,      /* fua requests */
  75
  76         BLK_TC_END      = 1 << 15,      /* we've run out of bits! */
  77 };
  78
  79 #define BLK_TC_SHIFT            (16)
  80 #define BLK_TC_ACT(act)         ((act) << BLK_TC_SHIFT)
  81 #define BLK_DATADIR(a) (((a) >> BLK_TC_SHIFT) & (BLK_TC_READ | BLK_TC_WRITE))
  82
  83 /*
  84  * Basic trace actions
  85  */
  86 enum {
  87         __BLK_TA_QUEUE = 1,             /* queued */
  88         __BLK_TA_BACKMERGE,             /* back merged to existing rq */
  89         __BLK_TA_FRONTMERGE,            /* front merge to existing rq */
  90         __BLK_TA_GETRQ,                 /* allocated new request */
  91         __BLK_TA_SLEEPRQ,               /* sleeping on rq allocation */
  92         __BLK_TA_REQUEUE,               /* request requeued */
  93         __BLK_TA_ISSUE,                 /* sent to driver */
  94         __BLK_TA_COMPLETE,              /* completed by driver */
  95         __BLK_TA_PLUG,                  /* queue was plugged */
  96         __BLK_TA_UNPLUG_IO,             /* queue was unplugged by io */
  97         __BLK_TA_UNPLUG_TIMER,          /* queue was unplugged by timer */
  98         __BLK_TA_INSERT,                /* insert request */
  99         __BLK_TA_SPLIT,                 /* bio was split */
 100         __BLK_TA_BOUNCE,                /* bio was bounced */
 101         __BLK_TA_REMAP,                 /* bio was remapped */
 102         __BLK_TA_ABORT,                 /* request aborted */
 103         __BLK_TA_DRV_DATA,              /* binary driver data */
 104 };
 105
 106 #define BLK_TA_MASK ((1 << BLK_TC_SHIFT) - 1)
 107
 108 /*
 109  * Notify events.
 110  */
 111 enum blktrace_notify {
 112         __BLK_TN_PROCESS = 0,           /* establish pid/name mapping */
 113         __BLK_TN_TIMESTAMP,             /* include system clock */
 114         __BLK_TN_MESSAGE,               /* Character string message */
 115 };
 116
 117 /*
 118  * Trace actions in full. Additionally, read or write is masked
 119  */
 120 #define BLK_TA_QUEUE            (__BLK_TA_QUEUE | BLK_TC_ACT(BLK_TC_QUEUE))
 121 #define BLK_TA_BACKMERGE        (__BLK_TA_BACKMERGE | BLK_TC_ACT(BLK_TC_QUEUE))
 122 #define BLK_TA_FRONTMERGE       (__BLK_TA_FRONTMERGE | BLK_TC_ACT(BLK_TC_QUEUE))
 123 #define BLK_TA_GETRQ            (__BLK_TA_GETRQ | BLK_TC_ACT(BLK_TC_QUEUE))
 124 #define BLK_TA_SLEEPRQ          (__BLK_TA_SLEEPRQ | BLK_TC_ACT(BLK_TC_QUEUE))
 125 #define BLK_TA_REQUEUE          (__BLK_TA_REQUEUE | BLK_TC_ACT(BLK_TC_REQUEUE))
 126 #define BLK_TA_ISSUE            (__BLK_TA_ISSUE | BLK_TC_ACT(BLK_TC_ISSUE))
 127 #define BLK_TA_COMPLETE         (__BLK_TA_COMPLETE| BLK_TC_ACT(BLK_TC_COMPLETE))
 128 #define BLK_TA_PLUG             (__BLK_TA_PLUG | BLK_TC_ACT(BLK_TC_QUEUE))
 129 #define BLK_TA_UNPLUG_IO        (__BLK_TA_UNPLUG_IO | BLK_TC_ACT(BLK_TC_QUEUE))
 130 #define BLK_TA_UNPLUG_TIMER     (__BLK_TA_UNPLUG_TIMER | BLK_TC_ACT(BLK_TC_QUEUE))
 131 #define BLK_TA_INSERT           (__BLK_TA_INSERT | BLK_TC_ACT(BLK_TC_QUEUE))
 132 #define BLK_TA_SPLIT            (__BLK_TA_SPLIT)
 133 #define BLK_TA_BOUNCE           (__BLK_TA_BOUNCE)
 134 #define BLK_TA_REMAP            (__BLK_TA_REMAP | BLK_TC_ACT(BLK_TC_QUEUE))
 135 #define BLK_TA_ABORT            (__BLK_TA_ABORT | BLK_TC_ACT(BLK_TC_QUEUE))
 136 #define BLK_TA_DRV_DATA         (__BLK_TA_DRV_DATA | BLK_TC_ACT(BLK_TC_DRV_DATA))
 137
 138 #define BLK_TN_PROCESS          (__BLK_TN_PROCESS | BLK_TC_ACT(BLK_TC_NOTIFY))
 139 #define BLK_TN_TIMESTAMP        (__BLK_TN_TIMESTAMP | BLK_TC_ACT(BLK_TC_NOTIFY))
 140 #define BLK_TN_MESSAGE          (__BLK_TN_MESSAGE | BLK_TC_ACT(BLK_TC_NOTIFY))
 141
 142 #define BLK_IO_TRACE_MAGIC      0x65617400
 143 #define BLK_IO_TRACE_VERSION    0x07
 144 /*
 145  * The trace itself
 146  */
 147 struct blk_io_trace {
 148         __u32 magic;            /* MAGIC << 8 | version */
 149         __u32 sequence;         /* event number */
 150         __u64 time;             /* in nanoseconds */
 151         __u64 sector;           /* disk offset */
 152         __u32 bytes;            /* transfer length */
 153         __u32 action;           /* what happened */
 154         __u32 pid;              /* who did it */
 155         __u32 device;           /* device identifier (dev_t) */
 156         __u32 cpu;              /* on what cpu did it happen */
 157         __u16 error;            /* completion error */
 158         __u16 pdu_len;          /* length of data after this trace */
 159 };
 160
 161 struct pending_io {
 162         /* sector offset of this IO */
 163         u64 sector;
 164
 165         /* dev_t for this IO */
 166         u32 device;
 167
 168         /* time this IO was dispatched */
 169         u64 dispatch_time;
 170         /* time this IO was finished */
 171         u64 completion_time;
 172         struct list_head hash_list;
 173         /* process which queued this IO */
 174         u32 pid;
 175 };
 176
 177 struct pid_map {
 178         struct list_head hash_list;
 179         u32 pid;
 180         int index;
 181         char name[0];
 182 };
 183
 184 u64 get_record_time(struct trace *trace)
 185 {
 186         return trace->io->time;
 187 }
 188
 189 void init_io_hash_table(void)
 190 {
 191         int i;
 192         struct list_head *head;
 193
 194         for (i = 0; i < IO_HASH_TABLE_SIZE; i++) {
 195                 head = io_hash_table + i;
 196                 INIT_LIST_HEAD(head);
 197         }
 198 }
 199
 200 /* taken from the kernel hash.h */
 201 static inline u64 hash_sector(u64 val)
 202 {
 203         u64 hash = val;
 204
 205         /*  Sigh, gcc can't optimise this alone like it does for 32 bits. */
 206         u64 n = hash;
 207         n <<= 18;
 208         hash -= n;
 209         n <<= 33;
 210         hash -= n;
 211         n <<= 3;
 212         hash += n;
 213         n <<= 3;
 214         hash -= n;
 215         n <<= 4;
 216         hash += n;
 217         n <<= 2;
 218         hash += n;
 219
 220         /* High bits are more random, so use them. */
 221         return hash >> (64 - IO_HASH_TABLE_BITS);
 222 }
 223
 224 static int io_hash_table_insert(struct pending_io *ins_pio)
 225 {
 226         u64 sector = ins_pio->sector;
 227         u32 dev = ins_pio->device;
 228         int slot = hash_sector(sector);
 229         struct list_head *head;
 230         struct pending_io *pio;
 231
 232         head = io_hash_table + slot;
 233         list_for_each_entry(pio, head, hash_list) {
 234                 if (pio->sector == sector && pio->device == dev)
 235                         return -EEXIST;
 236         }
 237         list_add_tail(&ins_pio->hash_list, head);
 238         return 0;
 239 }
 240
 241 static struct pending_io *io_hash_table_search(u64 sector, u32 dev)
 242 {
 243         int slot = hash_sector(sector);
 244         struct list_head *head;
 245         struct pending_io *pio;
 246
 247         head = io_hash_table + slot;
 248         list_for_each_entry(pio, head, hash_list) {
 249                 if (pio->sector == sector && pio->device == dev)
 250                         return pio;
 251         }
 252         return NULL;
 253 }
 254
 255 static struct pending_io *hash_queued_io(struct blk_io_trace *io)
 256 {
 257         struct pending_io *pio;
 258         int ret;
 259
 260         pio = calloc(1, sizeof(*pio));
 261         pio->sector = io->sector;
 262         pio->device = io->device;
 263         pio->pid = io->pid;
 264
 265         ret = io_hash_table_insert(pio);
 266         if (ret < 0) {
 267                 /* crud, the IO is there already */
 268                 free(pio);
 269                 return NULL;
 270         }
 271         return pio;
 272 }
 273
 274 static struct pending_io *hash_dispatched_io(struct blk_io_trace *io)
 275 {
 276         struct pending_io *pio;
 277
 278         pio = io_hash_table_search(io->sector, io->device);
 279         if (!pio) {
 280                 pio = hash_queued_io(io);
 281                 if (!pio)
 282                         return NULL;
 283         }
 284         pio->dispatch_time = io->time;
 285         return pio;
 286 }
 287
 288 static struct pending_io *hash_completed_io(struct blk_io_trace *io)
 289 {
 290         struct pending_io *pio;
 291
 292         pio = io_hash_table_search(io->sector, io->device);
 293
 294         if (!pio)
 295                 return NULL;
 296         return pio;
 297 }
 298
 299 void init_process_hash_table(void)
 300 {
 301         int i;
 302         struct list_head *head;
 303
 304         for (i = 0; i < PROCESS_HASH_TABLE_SIZE; i++) {
 305                 head = process_hash_table + i;
 306                 INIT_LIST_HEAD(head);
 307         }
 308 }
 309
 310 static u32 hash_pid(u32 pid)
 311 {
 312         u32 hash = pid;
 313
 314         hash ^= pid >> 3;
 315         hash ^= pid >> 3;
 316         hash ^= pid >> 4;
 317         hash ^= pid >> 6;
 318         return (hash & (PROCESS_HASH_TABLE_SIZE - 1));
 319 }
 320
 321 static struct pid_map *process_hash_search(u32 pid)
 322 {
 323         int slot = hash_pid(pid);
 324         struct list_head *head;
 325         struct pid_map *pm;
 326
 327         head = process_hash_table + slot;
 328         list_for_each_entry(pm, head, hash_list) {
 329                 if (pm->pid == pid)
 330                         return pm;
 331         }
 332         return NULL;
 333 }
 334
 335 static struct pid_map *process_hash_insert(u32 pid, char *name)
 336 {
 337         int slot = hash_pid(pid);
 338         struct pid_map *pm;
 339         int old_index = 0;
 340         char buf[16];
 341
 342         pm = process_hash_search(pid);
 343         if (pm) {
 344                 /* Entry exists and name shouldn't be changed? */
 345                 if (!name || !strcmp(name, pm->name))
 346                         return pm;
 347                 list_del(&pm->hash_list);
 348                 old_index = pm->index;
 349                 free(pm);
 350         }
 351         if (!name) {
 352                 sprintf(buf, "[%u]", pid);
 353                 name = buf;
 354         }
 355         pm = malloc(sizeof(struct pid_map) + strlen(name) + 1);
 356         pm->pid = pid;
 357         pm->index = old_index;
 358         strcpy(pm->name, name);
 359         list_add_tail(&pm->hash_list, process_hash_table + slot);
 360
 361         return pm;
 362 }
 363
 364 static void handle_notify(struct trace *trace)
 365 {
 366         struct blk_io_trace *io = trace->io;
 367         void *payload = (char *)io + sizeof(*io);
 368         u32 two32[2];
 369
 370         if (io->action == BLK_TN_PROCESS) {
 371                 if (io_per_process)
 372                         process_hash_insert(io->pid, payload);
 373                 return;
 374         }
 375
 376         if (io->action != BLK_TN_TIMESTAMP)
 377                 return;
 378
 379         if (io->pdu_len != sizeof(two32))
 380                 return;
 381
 382         memcpy(two32, payload, sizeof(two32));
 383         trace->start_timestamp = io->time;
 384         trace->abs_start_time.tv_sec = two32[0];
 385         trace->abs_start_time.tv_nsec = two32[1];
 386         if (trace->abs_start_time.tv_nsec < 0) {
 387                 trace->abs_start_time.tv_sec--;
 388                 trace->abs_start_time.tv_nsec += 1000000000;
 389         }
 390 }
 391
 392 int next_record(struct trace *trace)
 393 {
 394         int skip = trace->io->pdu_len;
 395         u64 offset;
 396
 397         trace->cur += sizeof(*trace->io) + skip;
 398         offset = trace->cur - trace->start;
 399         if (offset >= trace->len)
 400                 return 1;
 401
 402         trace->io = (struct blk_io_trace *)trace->cur;
 403         return 0;
 404 }
 405
 406 void first_record(struct trace *trace)
 407 {
 408         trace->cur = trace->start;
 409         trace->io = (struct blk_io_trace *)trace->cur;
 410 }
 411
 412 int is_io_event(struct blk_io_trace *test)
 413 {
 414         char *message;
 415         if (!(test->action & BLK_TC_ACT(BLK_TC_NOTIFY)))
 416                 return 1;
 417         if (test->action == BLK_TN_MESSAGE) {
 418                 int len = test->pdu_len;
 419                 if (len < 3)
 420                         return 0;
 421                 message = (char *)(test + 1);
 422                 if (strncmp(message, "fio ", 4) == 0) {
 423                         return 1;
 424                 }
 425         }
 426         return 0;
 427 }
 428
 429 u64 find_last_time(struct trace *trace)
 430 {
 431         char *p = trace->start + trace->len;
 432         struct blk_io_trace *test;
 433         int search_len = 0;
 434         u64 found = 0;
 435
 436         if (trace->len < sizeof(*trace->io))
 437                 return 0;
 438         p -= sizeof(*trace->io);
 439         while (p >= trace->start) {
 440                 test = (struct blk_io_trace *)p;
 441                 if (CHECK_MAGIC(test) && is_io_event(test)) {
 442                         u64 offset = p - trace->start;
 443                         if (offset + sizeof(*test) + test->pdu_len == trace->len) {
 444                                 return test->time;
 445                         }
 446                 }
 447                 p--;
 448                 search_len++;
 449                 if (search_len > 8192) {
 450                         break;
 451                 }
 452         }
 453
 454         /* searching backwards didn't work out, we'll have to scan the file */
 455         first_record(trace);
 456         while (1) {
 457                 if (is_io_event(trace->io))
 458                         found = trace->io->time;
 459                 if (next_record(trace))
 460                         break;
 461         }
 462         first_record(trace);
 463         return found;
 464 }
 465
 466 int parse_fio_bank_message(struct trace *trace, u64 *bank_ret, u64 *offset_ret,
 467                            u64 *num_banks_ret)
 468 {
 469         char *s;
 470         char *next;
 471         char *message;
 472         struct blk_io_trace *test = trace->io;
 473         int len = test->pdu_len;
 474         u64 bank;
 475         u64 offset;
 476         u64 num_banks;
 477
 478         if (!(test->action & BLK_TC_ACT(BLK_TC_NOTIFY)))
 479                 return -1;
 480         if (test->action != BLK_TN_MESSAGE)
 481                 return -1;
 482
 483         /* the message is fio rw bank offset num_banks */
 484         if (len < 3)
 485                 return -1;
 486         message = (char *)(test + 1);
 487         if (strncmp(message, "fio r ", 6) != 0)
 488                 return -1;
 489
 490         message = strndup(message, len);
 491         s = strchr(message, ' ');
 492         if (!s)
 493                 goto out;
 494         s++;
 495         s = strchr(s, ' ');
 496         if (!s)
 497                 goto out;
 498
 499         bank = strtoll(s, &next, 10);
 500         if (s == next)
 501                 goto out;
 502         s = next;
 503
 504         offset = strtoll(s, &next, 10);
 505         if (s == next)
 506                 goto out;
 507         s = next;
 508
 509         num_banks = strtoll(s, &next, 10);
 510         if (s == next)
 511                 goto out;
 512
 513         *bank_ret = bank;
 514         *offset_ret = offset;
 515         *num_banks_ret = num_banks;
 516
 517         return 0;
 518 out:
 519         free(message);
 520         return -1;
 521 }
 522
 523 static struct dev_info *lookup_dev(struct trace *trace, struct blk_io_trace *io)
 524 {
 525         u32 dev = io->device;
 526         int i;
 527         struct dev_info *di = NULL;
 528
 529         for (i = 0; i < trace->num_devices; i++) {
 530                 if (trace->devices[i].device == dev) {
 531                         di = trace->devices + i;
 532                         goto found;
 533                 }
 534         }
 535         i = trace->num_devices++;
 536         if (i >= MAX_DEVICES_PER_TRACE) {
 537                 fprintf(stderr, "Trace contains too many devices (%d)\n", i);
 538                 exit(1);
 539         }
 540         di = trace->devices + i;
 541         di->device = dev;
 542 found:
 543         return di;
 544 }
 545
 546 static void map_devices(struct trace *trace)
 547 {
 548         struct dev_info *di;
 549         u64 found;
 550         u64 map_start = 0;
 551         int i;
 552
 553         first_record(trace);
 554         while (1) {
 555                 if (!(trace->io->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
 556                         di = lookup_dev(trace, trace->io);
 557                         found = trace->io->sector << 9;
 558                         if (found < di->min)
 559                                 di->min = found;
 560
 561                         found += trace->io->bytes;
 562                         if (di->max < found)
 563                                 di->max = found;
 564                 }
 565                 if (next_record(trace))
 566                         break;
 567         }
 568         first_record(trace);
 569         for (i = 0; i < trace->num_devices; i++) {
 570                 di = trace->devices + i;
 571                 di->map = map_start;
 572                 map_start += di->max - di->min;
 573         }
 574 }
 575
 576 u64 map_io(struct trace *trace, struct blk_io_trace *io)
 577 {
 578         struct dev_info *di = lookup_dev(trace, io);
 579         u64 val = trace->io->sector << 9;
 580         return di->map + val - di->min;
 581 }
 582
 583 void find_extreme_offsets(struct trace *trace, u64 *min_ret, u64 *max_ret, u64 *max_bank_ret,
 584                           u64 *max_offset_ret)
 585 {
 586         u64 found = 0;
 587         u64 max = 0, min = ~(u64)0;
 588         u64 max_bank = 0;
 589         u64 max_bank_offset = 0;
 590         u64 num_banks = 0;
 591
 592         map_devices(trace);
 593
 594         first_record(trace);
 595         while (1) {
 596                 if (!(trace->io->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
 597                         found = map_io(trace, trace->io);
 598                         if (found < min)
 599                                 min = found;
 600
 601                         found += trace->io->bytes;
 602                         if (max < found)
 603                                 max = found;
 604                 } else {
 605                         u64 bank;
 606                         u64 offset;
 607                         if (!parse_fio_bank_message(trace, &bank,
 608                                                     &offset, &num_banks)) {
 609                                 if (bank > max_bank)
 610                                         max_bank = bank;
 611                                 if (offset > max_bank_offset)
 612                                         max_bank_offset = offset;
 613                         }
 614                 }
 615                 if (next_record(trace))
 616                         break;
 617         }
 618         first_record(trace);
 619         *min_ret = min;
 620         *max_ret = max;
 621         *max_bank_ret = max_bank;
 622         *max_offset_ret = max_bank_offset;
 623 }
 624
 625 static void check_io_types(struct trace *trace)
 626 {
 627         struct blk_io_trace *io = trace->io;
 628         int action = io->action & BLK_TA_MASK;
 629
 630         if (!(io->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
 631                 switch (action) {
 632                 case __BLK_TA_COMPLETE:
 633                         trace->found_completion = 1;
 634                         break;
 635                 case __BLK_TA_ISSUE:
 636                         trace->found_issue = 1;
 637                         break;
 638                 case __BLK_TA_QUEUE:
 639                         trace->found_queue = 1;
 640                         break;
 641                 };
 642         }
 643 }
 644
 645
 646 int filter_outliers(struct trace *trace, u64 min_offset, u64 max_offset,
 647                     u64 *yzoom_min, u64 *yzoom_max)
 648 {
 649         int hits[11];
 650         u64 max_per_bucket[11];
 651         u64 min_per_bucket[11];
 652         u64 bytes_per_bucket = (max_offset - min_offset + 1) / 10;
 653         int slot;
 654         int fat_count = 0;
 655
 656         memset(hits, 0, sizeof(int) * 11);
 657         memset(max_per_bucket, 0, sizeof(u64) * 11);
 658         memset(min_per_bucket, 0xff, sizeof(u64) * 11);
 659         first_record(trace);
 660         while (1) {
 661                 check_io_types(trace);
 662                 if (!(trace->io->action & BLK_TC_ACT(BLK_TC_NOTIFY)) &&
 663                     (trace->io->action & BLK_TA_MASK) == __BLK_TA_QUEUE) {
 664                         u64 off = map_io(trace, trace->io) - min_offset;
 665
 666                         slot = (int)(off / bytes_per_bucket);
 667                         hits[slot]++;
 668                         if (off < min_per_bucket[slot])
 669                                 min_per_bucket[slot] = off;
 670
 671                         off += trace->io->bytes;
 672                         slot = (int)(off / bytes_per_bucket);
 673                         hits[slot]++;
 674                         if (off > max_per_bucket[slot])
 675                                 max_per_bucket[slot] = off;
 676                 }
 677                 if (next_record(trace))
 678                         break;
 679         }
 680         first_record(trace);
 681         for (slot = 0; slot < 11; slot++) {
 682                 if (hits[slot] > fat_count) {
 683                         fat_count = hits[slot];
 684                 }
 685         }
 686
 687         *yzoom_max = max_offset;
 688         for (slot = 10; slot >= 0; slot--) {
 689                 double d = hits[slot];
 690
 691                 if (d >= (double)fat_count * .05) {
 692                         *yzoom_max = max_per_bucket[slot] + min_offset;
 693                         break;
 694                 }
 695         }
 696
 697         *yzoom_min = min_offset;
 698         for (slot = 0; slot < 10; slot++) {
 699                 double d = hits[slot];
 700
 701                 if (d >= (double)fat_count * .05) {
 702                         *yzoom_min = min_per_bucket[slot] + min_offset;
 703                         break;
 704                 }
 705         }
 706         return 0;
 707 }
 708
 709 static char footer[] = ".blktrace.0";
 710 static int footer_len = sizeof(footer) - 1;
 711
 712 static int match_trace(char *name, int *len)
 713 {
 714         int match_len;
 715         int footer_start;
 716
 717         match_len = strlen(name);
 718         if (match_len <= footer_len)
 719                 return 0;
 720
 721         footer_start = match_len - footer_len;
 722         if (strcmp(name + footer_start, footer) != 0)
 723                 return 0;
 724
 725         if (len)
 726                 *len = match_len;
 727         return 1;
 728 }
 729
 730 struct tracelist {
 731         struct tracelist *next;
 732         char *name;
 733 };
 734
 735 static struct tracelist *traces_list(char *dir_name, int *len)
 736 {
 737         int count = 0;
 738         struct tracelist *traces = NULL;
 739         DIR *dir = opendir(dir_name);
 740         if (!dir)
 741                 return NULL;
 742
 743         while (1) {
 744                 int len;
 745                 struct tracelist *tl;
 746                 struct dirent *d = readdir(dir);
 747                 if (!d)
 748                         break;
 749
 750                 if (!match_trace(d->d_name, &len))
 751                         continue;
 752
 753                 /* Allocate space for tracelist + filename */
 754                 tl = calloc(1, sizeof(struct tracelist) + (sizeof(char) * (len + 1)));
 755                 if (!tl)
 756                         return NULL;
 757                 tl->next = traces;
 758                 tl->name = (char *)(tl + 1);
 759                 strncpy(tl->name, d->d_name, len);
 760                 traces = tl;
 761                 count++;
 762         }
 763
 764         closedir(dir);
 765
 766         if (len)
 767                 *len = count;
 768
 769         return traces;
 770 }
 771
 772 static void traces_free(struct tracelist *traces)
 773 {
 774         while (traces) {
 775                 struct tracelist *tl = traces;
 776                 traces = traces->next;
 777                 free(tl);
 778         }
 779 }
 780
 781 static char *combine_blktrace_devs(char *dir_name)
 782 {
 783         struct tracelist *traces = NULL;
 784         struct tracelist *tl;
 785         char *ret = NULL;
 786         char **argv = NULL;
 787         char *dumpfile;
 788         int argc = 0;
 789         int i;
 790         int err;
 791
 792         if (!asprintf(&dumpfile, "%s.dump", dir_name))
 793                 goto out;
 794
 795         traces = traces_list(dir_name, &argc);
 796         if (!traces)
 797                 goto out;
 798
 799         argc *= 2; /* {"-i", trace } */
 800         argc += 6; /* See below */
 801         argv = calloc(argc + 1, sizeof(char *));
 802         if (!argv)
 803                 goto out;
 804
 805         i = 0;
 806         argv[i++] = "blkparse";
 807         argv[i++] = "-O";
 808         argv[i++] = "-D";
 809         argv[i++] = dir_name;
 810         argv[i++] = "-d";
 811         argv[i++] = dumpfile;
 812         for (tl = traces; tl != NULL; tl = tl->next) {
 813                 argv[i++] = "-i";
 814                 argv[i++] = tl->name;
 815         }
 816
 817         err = run_program2(argc, argv);
 818         free(argv);
 819         if (err) {
 820                 fprintf(stderr, "blkparse failed with exit code %d\n", err);
 821                 exit(1);
 822         }
 823         ret = dumpfile;
 824 out:
 825         traces_free(traces);
 826         return ret;
 827 }
 828
 829 static char *find_trace_file(char *filename)
 830 {
 831         int ret;
 832         struct stat st;
 833         char *dot;
 834         char *try;
 835         int found_dir = 0;
 836
 837         /* look for an exact match of whatever they pass in.
 838          * If it is a file, assume it is the dump file.
 839          * If a directory, remember that it existed so we
 840          * can combine traces in that directory later
 841          */
 842         ret = stat(filename, &st);
 843         if (ret == 0) {
 844                 if (S_ISREG(st.st_mode))
 845                         return strdup(filename);
 846
 847                 if (S_ISDIR(st.st_mode))
 848                         found_dir = 1;
 849         }
 850
 851         if (found_dir) {
 852                 int i;
 853                 /* Eat up trailing '/'s */
 854                 for (i = strlen(filename) - 1; filename[i] == '/'; i--)
 855                         filename[i] = '\0';
 856         }
 857
 858         /*
 859          * try tacking .dump onto the end and see if that already
 860          * has been generated
 861          */
 862         snprintf(line, line_len, "%s.%s", filename, "dump");
 863         ret = stat(line, &st);
 864         if (ret == 0)
 865                 return strdup(line);
 866
 867         /*
 868          * try to generate the .dump from all the traces in
 869          * a single dir.
 870          */
 871         if (found_dir) {
 872                 try = combine_blktrace_devs(filename);
 873                 if (try)
 874                         return try;
 875         }
 876
 877         /*
 878          * try to generate the .dump from all the blktrace
 879          * files for a named trace
 880          */
 881         try = strdup(filename);
 882         dot = strrchr(try, '.');
 883         if (!dot || strcmp(".dump", dot) != 0) {
 884                 if (dot && dot != try)
 885                         *dot = '\0';
 886                 snprintf(line, line_len, "%s%s", try, ".blktrace.0");
 887                 ret = stat(line, &st);
 888                 if (ret == 0) {
 889                         blktrace_to_dump(try);
 890                         snprintf(line, line_len, "%s.%s", try, "dump");
 891                         ret = stat(line, &st);
 892                         if (ret == 0) {
 893                                 free(try);
 894                                 return strdup(line);
 895                         }
 896                 }
 897         }
 898         free(try);
 899         return NULL;
 900 }
 901 struct trace *open_trace(char *filename)
 902 {
 903         int fd;
 904         char *p;
 905         struct stat st;
 906         int ret;
 907         struct trace *trace;
 908         char *found_filename;
 909
 910         trace = calloc(1, sizeof(*trace));
 911         if (!trace) {
 912                 fprintf(stderr, "unable to allocate memory for trace\n");
 913                 return NULL;
 914         }
 915
 916         found_filename = find_trace_file(filename);
 917         if (!found_filename) {
 918                 fprintf(stderr, "Unable to find trace file %s\n", filename);
 919                 goto fail;
 920         }
 921         filename = found_filename;
 922
 923         fd = open(filename, O_RDONLY);
 924         if (fd < 0) {
 925                 fprintf(stderr, "Unable to open trace file %s err %s\n", filename, strerror(errno));
 926                 goto fail;
 927         }
 928         ret = fstat(fd, &st);
 929         if (ret < 0) {
 930                 fprintf(stderr, "stat failed on %s err %s\n", filename, strerror(errno));
 931                 goto fail_fd;
 932         }
 933         p = mmap(NULL, st.st_size, PROT_READ, MAP_PRIVATE, fd, 0);
 934         if (p == MAP_FAILED) {
 935                 fprintf(stderr, "Unable to mmap trace file %s, err %s\n", filename, strerror(errno));
 936                 goto fail_fd;
 937         }
 938         trace->fd = fd;
 939         trace->len = st.st_size;
 940         trace->start = p;
 941         trace->cur = p;
 942         trace->io = (struct blk_io_trace *)p;
 943         return trace;
 944
 945 fail_fd:
 946         close(fd);
 947 fail:
 948         free(trace);
 949         return NULL;
 950 }
 951 static inline int tput_event(struct trace *trace)
 952 {
 953         if (trace->found_completion)
 954                 return __BLK_TA_COMPLETE;
 955         if (trace->found_issue)
 956                 return __BLK_TA_ISSUE;
 957         if (trace->found_queue)
 958                 return __BLK_TA_QUEUE;
 959
 960         return __BLK_TA_COMPLETE;
 961 }
 962
 963 int action_char_to_num(char action)
 964 {
 965         switch (action) {
 966         case 'Q':
 967                 return __BLK_TA_QUEUE;
 968         case 'D':
 969                 return __BLK_TA_ISSUE;
 970         case 'C':
 971                 return __BLK_TA_COMPLETE;
 972         }
 973         return -1;
 974 }
 975
 976 static inline int io_event(struct trace *trace)
 977 {
 978         if (plot_io_action)
 979                 return plot_io_action;
 980         if (trace->found_queue)
 981                 return __BLK_TA_QUEUE;
 982         if (trace->found_issue)
 983                 return __BLK_TA_ISSUE;
 984         if (trace->found_completion)
 985                 return __BLK_TA_COMPLETE;
 986
 987         return __BLK_TA_COMPLETE;
 988 }
 989
 990 void add_tput(struct trace *trace, struct graph_line_data *writes_gld,
 991               struct graph_line_data *reads_gld)
 992 {
 993         struct blk_io_trace *io = trace->io;
 994         struct graph_line_data *gld;
 995         int action = io->action & BLK_TA_MASK;
 996         int seconds;
 997
 998         if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY))
 999                 return;
1000
1001         if (action != tput_event(trace))
1002                 return;
1003
1004         if (BLK_DATADIR(io->action) & BLK_TC_READ)
1005                 gld = reads_gld;
1006         else
1007                 gld = writes_gld;
1008
1009         seconds = SECONDS(io->time);
1010         gld->data[seconds].sum += io->bytes;
1011
1012         gld->data[seconds].count = 1;
1013         if (gld->data[seconds].sum > gld->max)
1014                 gld->max = gld->data[seconds].sum;
1015 }
1016
1017 #define GDD_PTR_ALLOC_STEP 16
1018
1019 static struct pid_map *get_pid_map(struct trace_file *tf, u32 pid)
1020 {
1021         struct pid_map *pm;
1022
1023         if (!io_per_process) {
1024                 if (!tf->io_plots)
1025                         tf->io_plots = 1;
1026                 return NULL;
1027         }
1028
1029         pm = process_hash_insert(pid, NULL);
1030         /* New entry? */
1031         if (!pm->index) {
1032                 if (tf->io_plots == tf->io_plots_allocated) {
1033                         tf->io_plots_allocated += GDD_PTR_ALLOC_STEP;
1034                         tf->gdd_reads = realloc(tf->gdd_reads, tf->io_plots_allocated * sizeof(struct graph_dot_data *));
1035                         if (!tf->gdd_reads)
1036                                 abort();
1037                         tf->gdd_writes = realloc(tf->gdd_writes, tf->io_plots_allocated * sizeof(struct graph_dot_data *));
1038                         if (!tf->gdd_writes)
1039                                 abort();
1040                         memset(tf->gdd_reads + tf->io_plots_allocated - GDD_PTR_ALLOC_STEP,
1041                                0, GDD_PTR_ALLOC_STEP * sizeof(struct graph_dot_data *));
1042                         memset(tf->gdd_writes + tf->io_plots_allocated - GDD_PTR_ALLOC_STEP,
1043                                0, GDD_PTR_ALLOC_STEP * sizeof(struct graph_dot_data *));
1044                 }
1045                 pm->index = tf->io_plots++;
1046
1047                 return pm;
1048         }
1049         return pm;
1050 }
1051
1052 void add_io(struct trace *trace, struct trace_file *tf)
1053 {
1054         struct blk_io_trace *io = trace->io;
1055         int action = io->action & BLK_TA_MASK;
1056         u64 offset;
1057         int index;
1058         char *label;
1059         struct pid_map *pm;
1060
1061         if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY))
1062                 return;
1063
1064         if (action != io_event(trace))
1065                 return;
1066
1067         offset = map_io(trace, io);
1068
1069         pm = get_pid_map(tf, io->pid);
1070         if (!pm) {
1071                 index = 0;
1072                 label = "";
1073         } else {
1074                 index = pm->index;
1075                 label = pm->name;
1076         }
1077         if (BLK_DATADIR(io->action) & BLK_TC_READ) {
1078                 if (!tf->gdd_reads[index])
1079                         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));
1080                 set_gdd_bit(tf->gdd_reads[index], offset, io->bytes, io->time);
1081         } else if (BLK_DATADIR(io->action) & BLK_TC_WRITE) {
1082                 if (!tf->gdd_writes[index])
1083                         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));
1084                 set_gdd_bit(tf->gdd_writes[index], offset, io->bytes, io->time);
1085         }
1086 }
1087
1088 void add_pending_io(struct trace *trace, struct graph_line_data *gld)
1089 {
1090         unsigned int seconds;
1091         struct blk_io_trace *io = trace->io;
1092         int action = io->action & BLK_TA_MASK;
1093         double avg;
1094         struct pending_io *pio;
1095
1096         if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY))
1097                 return;
1098
1099         if (action == __BLK_TA_QUEUE) {
1100                 if (trace->found_issue || trace->found_completion)
1101                         hash_queued_io(trace->io);
1102                 return;
1103         }
1104         if (action == __BLK_TA_REQUEUE) {
1105                 if (ios_in_flight > 0)
1106                         ios_in_flight--;
1107                 return;
1108         }
1109         if (action != __BLK_TA_ISSUE)
1110                 return;
1111
1112         pio = hash_dispatched_io(trace->io);
1113         if (!pio)
1114                 return;
1115
1116         if (!trace->found_completion) {
1117                 list_del(&pio->hash_list);
1118                 free(pio);
1119         }
1120
1121         ios_in_flight++;
1122
1123         seconds = SECONDS(io->time);
1124         gld->data[seconds].sum += ios_in_flight;
1125         gld->data[seconds].count++;
1126
1127         avg = (double)gld->data[seconds].sum / gld->data[seconds].count;
1128         if (gld->max < (u64)avg) {
1129                 gld->max = avg;
1130         }
1131 }
1132
1133 void add_completed_io(struct trace *trace,
1134                       struct graph_line_data *latency_gld)
1135 {
1136         struct blk_io_trace *io = trace->io;
1137         int seconds;
1138         int action = io->action & BLK_TA_MASK;
1139         struct pending_io *pio;
1140         double avg;
1141         u64 latency;
1142
1143         if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY))
1144                 return;
1145
1146         if (action != __BLK_TA_COMPLETE)
1147                 return;
1148
1149         seconds = SECONDS(io->time);
1150
1151         pio = hash_completed_io(trace->io);
1152         if (!pio)
1153                 return;
1154
1155         if (ios_in_flight > 0)
1156                 ios_in_flight--;
1157         if (io->time >= pio->dispatch_time) {
1158                 latency = io->time - pio->dispatch_time;
1159                 latency_gld->data[seconds].sum += latency;
1160                 latency_gld->data[seconds].count++;
1161         }
1162
1163         list_del(&pio->hash_list);
1164         free(pio);
1165
1166         avg = (double)latency_gld->data[seconds].sum /
1167                 latency_gld->data[seconds].count;
1168         if (latency_gld->max < (u64)avg) {
1169                 latency_gld->max = avg;
1170         }
1171 }
1172
1173 void add_iop(struct trace *trace, struct graph_line_data *gld)
1174 {
1175         struct blk_io_trace *io = trace->io;
1176         int action = io->action & BLK_TA_MASK;
1177         int seconds;
1178
1179         if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY))
1180                 return;
1181
1182         /* iops and tput use the same events */
1183         if (action != tput_event(trace))
1184                 return;
1185
1186         seconds = SECONDS(io->time);
1187         gld->data[seconds].sum += 1;
1188         gld->data[seconds].count = 1;
1189         if (gld->data[seconds].sum > gld->max)
1190                 gld->max = gld->data[seconds].sum;
1191 }
1192
1193 void check_record(struct trace *trace)
1194 {
1195         handle_notify(trace);
1196 }