kernel/blk-trace-2.6.16-rc2-git-B2

   1 diff --git a/block/Kconfig b/block/Kconfig
   2 index 377f6dd..27eaed9 100644
   3 --- a/block/Kconfig
   4 +++ b/block/Kconfig
   5 @@ -11,4 +11,15 @@ config LBD
   6           your machine, or if you want to have a raid or loopback device
   7           bigger than 2TB.  Otherwise say N.
   8
   9 +config BLK_DEV_IO_TRACE
  10 +       bool "Support for tracing block io actions"
  11 +       select RELAYFS_FS
  12 +       help
  13 +         Say Y here, if you want to be able to trace the block layer actions
  14 +         on a given queue. Tracing allows you to see any traffic happening
  15 +         on a block device queue. For more information (and the user space
  16 +         support tools needed), fetch the blktrace app from:
  17 +
  18 +         git://brick.kernel.dk/data/git/blktrace.git
  19 +
  20  source block/Kconfig.iosched
  21 diff --git a/block/Makefile b/block/Makefile
  22 index 7e4f93e..c05de0e 100644
  23 --- a/block/Makefile
  24 +++ b/block/Makefile
  25 @@ -8,3 +8,5 @@ obj-$(CONFIG_IOSCHED_NOOP)      += noop-iosch
  26  obj-$(CONFIG_IOSCHED_AS)       += as-iosched.o
  27  obj-$(CONFIG_IOSCHED_DEADLINE) += deadline-iosched.o
  28  obj-$(CONFIG_IOSCHED_CFQ)      += cfq-iosched.o
  29 +
  30 +obj-$(CONFIG_BLK_DEV_IO_TRACE) += blktrace.o
  31 diff --git a/block/blktrace.c b/block/blktrace.c
  32 new file mode 100644
  33 index 0000000..3b03587
  34 --- /dev/null
  35 +++ b/block/blktrace.c
  36 @@ -0,0 +1,499 @@
  37 +#include <linux/config.h>
  38 +#include <linux/kernel.h>
  39 +#include <linux/blkdev.h>
  40 +#include <linux/blktrace_api.h>
  41 +#include <linux/percpu.h>
  42 +#include <linux/init.h>
  43 +#include <linux/mutex.h>
  44 +#include <asm/uaccess.h>
  45 +
  46 +static DEFINE_PER_CPU(unsigned long long, blk_trace_cpu_offset) = { 0, };
  47 +static unsigned int blktrace_seq = 1;
  48 +
  49 +static void trace_note_tsk(struct blk_trace *bt, struct task_struct *tsk)
  50 +{
  51 +       struct blk_io_trace *t;
  52 +
  53 +       t = relay_reserve(bt->rchan, sizeof(*t) + sizeof(tsk->comm));
  54 +       if (t) {
  55 +               t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
  56 +               t->device = bt->dev;
  57 +               t->action = BLK_TC_ACT(BLK_TC_NOTIFY);
  58 +               t->pid = tsk->pid;
  59 +               t->cpu = smp_processor_id();
  60 +               t->pdu_len = sizeof(tsk->comm);
  61 +               memcpy((void *) t + sizeof(*t), tsk->comm, t->pdu_len);
  62 +               tsk->btrace_seq = blktrace_seq;
  63 +       }
  64 +}
  65 +
  66 +static int act_log_check(struct blk_trace *bt, u32 what, sector_t sector,
  67 +                        pid_t pid)
  68 +{
  69 +       if (((bt->act_mask << BLK_TC_SHIFT) & what) == 0)
  70 +               return 1;
  71 +       if (sector < bt->start_lba || sector > bt->end_lba)
  72 +               return 1;
  73 +       if (bt->pid && pid != bt->pid)
  74 +               return 1;
  75 +
  76 +       return 0;
  77 +}
  78 +
  79 +/*
  80 + * Data direction bit lookup
  81 + */
  82 +static u32 ddir_act[2] __read_mostly = { BLK_TC_ACT(BLK_TC_READ), BLK_TC_ACT(BLK_TC_WRITE) };
  83 +
  84 +/*
  85 + * Bio action bits of interest
  86 + */
  87 +static u32 bio_act[3] __read_mostly = { 0, BLK_TC_ACT(BLK_TC_BARRIER), BLK_TC_ACT(BLK_TC_SYNC) };
  88 +
  89 +/*
  90 + * More could be added as needed, taking care to increment the decrementer
  91 + * to get correct indexing
  92 + */
  93 +#define trace_barrier_bit(rw)  \
  94 +       (((rw) & (1 << BIO_RW_BARRIER)) >> (BIO_RW_BARRIER - 0))
  95 +#define trace_sync_bit(rw)     \
  96 +       (((rw) & (1 << BIO_RW_SYNC)) >> (BIO_RW_SYNC - 1))
  97 +
  98 +/*
  99 + * The worker for the various blk_add_trace*() types. Fills out a
 100 + * blk_io_trace structure and places it in a per-cpu subbuffer.
 101 + */
 102 +void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
 103 +                    int rw, u32 what, int error, int pdu_len, void *pdu_data)
 104 +{
 105 +       struct task_struct *tsk = current;
 106 +       struct blk_io_trace *t;
 107 +       unsigned long flags;
 108 +       unsigned long *sequence;
 109 +       unsigned long seq;
 110 +       u64 cpu_time;
 111 +       pid_t pid;
 112 +       int cpu;
 113 +
 114 +       if (unlikely(bt->trace_state != Blktrace_running))
 115 +               return;
 116 +
 117 +       what |= ddir_act[rw & WRITE];
 118 +       what |= bio_act[trace_barrier_bit(rw)];
 119 +       what |= bio_act[trace_sync_bit(rw)];
 120 +
 121 +       pid = tsk->pid;
 122 +       if (unlikely(act_log_check(bt, what, sector, pid)))
 123 +               return;
 124 +
 125 +       /*
 126 +        * A word about the locking here - we disable interrupts to reserve
 127 +        * some space in the relayfs per-cpu buffer, to prevent an irq
 128 +        * from coming in and stepping on our toes. Once reserved, it's
 129 +        * enough to get preemption disabled to prevent read of this data
 130 +        * before we are through filling it. get_cpu()/put_cpu() does this
 131 +        * for us
 132 +        */
 133 +       local_irq_save(flags);
 134 +
 135 +       if (unlikely(tsk->btrace_seq != blktrace_seq))
 136 +               trace_note_tsk(bt, tsk);
 137 +
 138 +       t = relay_reserve(bt->rchan, sizeof(*t) + pdu_len);
 139 +       if (unlikely(!t)) {
 140 +               local_irq_restore(flags);
 141 +               return;
 142 +       }
 143 +
 144 +       cpu = get_cpu();
 145 +
 146 +       sequence = per_cpu_ptr(bt->sequence, cpu);
 147 +       seq = ++(*sequence);
 148 +       cpu_time = sched_clock() - per_cpu(blk_trace_cpu_offset, cpu);
 149 +
 150 +       local_irq_restore(flags);
 151 +
 152 +       t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
 153 +       t->sequence = seq;
 154 +       t->time = cpu_time;
 155 +       t->sector = sector;
 156 +       t->bytes = bytes;
 157 +       t->action = what;
 158 +       t->pid = pid;
 159 +       t->device = bt->dev;
 160 +       t->cpu = cpu;
 161 +       t->error = error;
 162 +       t->pdu_len = pdu_len;
 163 +
 164 +       if (pdu_len)
 165 +               memcpy((void *) t + sizeof(*t), pdu_data, pdu_len);
 166 +
 167 +       put_cpu();
 168 +}
 169 +
 170 +EXPORT_SYMBOL_GPL(__blk_add_trace);
 171 +
 172 +static struct dentry *blk_tree_root;
 173 +static struct mutex blk_tree_mutex;
 174 +
 175 +static inline void blk_remove_root(void)
 176 +{
 177 +       if (relayfs_remove_dir(blk_tree_root) != -ENOTEMPTY)
 178 +               blk_tree_root = NULL;
 179 +}
 180 +
 181 +static void blk_remove_tree(struct dentry *dir)
 182 +{
 183 +       mutex_lock(&blk_tree_mutex);
 184 +       relayfs_remove_dir(dir);
 185 +       blk_remove_root();
 186 +       mutex_unlock(&blk_tree_mutex);
 187 +}
 188 +
 189 +static struct dentry *blk_create_tree(const char *blk_name)
 190 +{
 191 +       struct dentry *dir = NULL;
 192 +
 193 +       mutex_lock(&blk_tree_mutex);
 194 +
 195 +       if (!blk_tree_root) {
 196 +               blk_tree_root = relayfs_create_dir("block", NULL);
 197 +               if (!blk_tree_root)
 198 +                       goto err;
 199 +       }
 200 +
 201 +       dir = relayfs_create_dir(blk_name, blk_tree_root);
 202 +       if (!dir)
 203 +               blk_remove_root();
 204 +
 205 +err:
 206 +       mutex_unlock(&blk_tree_mutex);
 207 +       return dir;
 208 +}
 209 +
 210 +void blk_trace_cleanup(struct blk_trace *bt)
 211 +{
 212 +       relay_close(bt->rchan);
 213 +       relayfs_remove_file(bt->dropped_file);
 214 +       blk_remove_tree(bt->dir);
 215 +       free_percpu(bt->sequence);
 216 +       kfree(bt);
 217 +}
 218 +
 219 +static int blk_trace_remove(request_queue_t *q)
 220 +{
 221 +       struct blk_trace *bt;
 222 +
 223 +       bt = xchg(&q->blk_trace, NULL);
 224 +       if (!bt)
 225 +               return -EINVAL;
 226 +
 227 +       if (bt->trace_state == Blktrace_setup ||
 228 +           bt->trace_state == Blktrace_stopped)
 229 +               blk_trace_cleanup(bt);
 230 +
 231 +       return 0;
 232 +}
 233 +
 234 +static int blk_dropped_open(struct inode *inode, struct file *filp)
 235 +{
 236 +       filp->private_data = inode->u.generic_ip;
 237 +
 238 +       return 0;
 239 +}
 240 +
 241 +static ssize_t blk_dropped_read(struct file *filp, char __user *buffer,
 242 +                               size_t count, loff_t *ppos)
 243 +{
 244 +       struct blk_trace *bt = filp->private_data;
 245 +       char buf[16];
 246 +
 247 +       snprintf(buf, sizeof(buf), "%u\n", atomic_read(&bt->dropped));
 248 +
 249 +       return simple_read_from_buffer(buffer, count, ppos, buf, strlen(buf));
 250 +}
 251 +
 252 +static struct file_operations blk_dropped_fops = {
 253 +       .owner =        THIS_MODULE,
 254 +       .open =         blk_dropped_open,
 255 +       .read =         blk_dropped_read,
 256 +};
 257 +
 258 +/*
 259 + * Keep track of how many times we encountered a full subbuffer, to aid
 260 + * the user space app in telling how many lost events there were.
 261 + */
 262 +static int blk_subbuf_start_callback(struct rchan_buf *buf, void *subbuf,
 263 +                                    void *prev_subbuf, size_t prev_padding)
 264 +{
 265 +       struct blk_trace *bt;
 266 +
 267 +       if (!relay_buf_full(buf))
 268 +               return 1;
 269 +
 270 +       bt = buf->chan->private_data;
 271 +       atomic_inc(&bt->dropped);
 272 +       return 0;
 273 +}
 274 +
 275 +static struct rchan_callbacks blk_relay_callbacks = {
 276 +       .subbuf_start = blk_subbuf_start_callback,
 277 +};
 278 +
 279 +/*
 280 + * Setup everything required to start tracing
 281 + */
 282 +static int blk_trace_setup(request_queue_t *q, struct block_device *bdev,
 283 +                          char __user *arg)
 284 +{
 285 +       struct blk_user_trace_setup buts;
 286 +       struct blk_trace *old_bt, *bt = NULL;
 287 +       struct dentry *dir = NULL;
 288 +       char b[BDEVNAME_SIZE];
 289 +       int ret, i;
 290 +
 291 +       if (copy_from_user(&buts, arg, sizeof(buts)))
 292 +               return -EFAULT;
 293 +
 294 +       if (!buts.buf_size || !buts.buf_nr)
 295 +               return -EINVAL;
 296 +
 297 +       strcpy(buts.name, bdevname(bdev, b));
 298 +
 299 +       /*
 300 +        * some device names have larger paths - convert the slashes
 301 +        * to underscores for this to work as expected
 302 +        */
 303 +       for (i = 0; i < strlen(buts.name); i++)
 304 +               if (buts.name[i] == '/')
 305 +                       buts.name[i] = '_';
 306 +
 307 +       if (copy_to_user(arg, &buts, sizeof(buts)))
 308 +               return -EFAULT;
 309 +
 310 +       ret = -ENOMEM;
 311 +       bt = kzalloc(sizeof(*bt), GFP_KERNEL);
 312 +       if (!bt)
 313 +               goto err;
 314 +
 315 +       bt->sequence = alloc_percpu(unsigned long);
 316 +       if (!bt->sequence)
 317 +               goto err;
 318 +
 319 +       ret = -ENOENT;
 320 +       dir = blk_create_tree(buts.name);
 321 +       if (!dir)
 322 +               goto err;
 323 +
 324 +       bt->dir = dir;
 325 +       bt->dev = bdev->bd_dev;
 326 +       atomic_set(&bt->dropped, 0);
 327 +
 328 +       ret = -EIO;
 329 +       bt->dropped_file = relayfs_create_file("dropped", dir, 0, &blk_dropped_fops, bt);
 330 +       if (!bt->dropped_file)
 331 +               goto err;
 332 +
 333 +       bt->rchan = relay_open("trace", dir, buts.buf_size, buts.buf_nr, &blk_relay_callbacks);
 334 +       if (!bt->rchan)
 335 +               goto err;
 336 +       bt->rchan->private_data = bt;
 337 +
 338 +       bt->act_mask = buts.act_mask;
 339 +       if (!bt->act_mask)
 340 +               bt->act_mask = (u16) -1;
 341 +
 342 +       bt->start_lba = buts.start_lba;
 343 +       bt->end_lba = buts.end_lba;
 344 +       if (!bt->end_lba)
 345 +               bt->end_lba = -1ULL;
 346 +
 347 +       bt->pid = buts.pid;
 348 +       bt->trace_state = Blktrace_setup;
 349 +
 350 +       ret = -EBUSY;
 351 +       old_bt = xchg(&q->blk_trace, bt);
 352 +       if (old_bt) {
 353 +               xchg(&q->blk_trace, old_bt);
 354 +               goto err;
 355 +       }
 356 +
 357 +       return 0;
 358 +err:
 359 +       if (bt && bt->dropped_file)
 360 +               relayfs_remove_file(bt->dropped_file);
 361 +       if (dir)
 362 +               blk_remove_tree(dir);
 363 +       if (bt) {
 364 +               if (bt->sequence)
 365 +                       free_percpu(bt->sequence);
 366 +               kfree(bt);
 367 +       }
 368 +       return ret;
 369 +}
 370 +
 371 +static int blk_trace_startstop(request_queue_t *q, int start)
 372 +{
 373 +       struct blk_trace *bt;
 374 +       int ret;
 375 +
 376 +       if ((bt = q->blk_trace) == NULL)
 377 +               return -EINVAL;
 378 +
 379 +       /*
 380 +        * For starting a trace, we can transition from a setup or stopped
 381 +        * trace. For stopping a trace, the state must be running
 382 +        */
 383 +       ret = -EINVAL;
 384 +       if (start) {
 385 +               if (bt->trace_state == Blktrace_setup ||
 386 +                   bt->trace_state == Blktrace_stopped) {
 387 +                       blktrace_seq++;
 388 +                       smp_mb();
 389 +                       bt->trace_state = Blktrace_running;
 390 +                       ret = 0;
 391 +               }
 392 +       } else {
 393 +               if (bt->trace_state == Blktrace_running) {
 394 +                       bt->trace_state = Blktrace_stopped;
 395 +                       ret = 0;
 396 +               }
 397 +       }
 398 +
 399 +       return ret;
 400 +}
 401 +
 402 +/**
 403 + * blk_trace_ioctl: - handle the ioctls associated with tracing
 404 + * @bdev:      the block device
 405 + * @cmd:       the ioctl cmd
 406 + * @arg:       the argument data, if any
 407 + *
 408 + **/
 409 +int blk_trace_ioctl(struct block_device *bdev, unsigned cmd, char __user *arg)
 410 +{
 411 +       request_queue_t *q;
 412 +       int ret, start = 0;
 413 +
 414 +       q = bdev_get_queue(bdev);
 415 +       if (!q)
 416 +               return -ENXIO;
 417 +
 418 +       down(&bdev->bd_sem);
 419 +
 420 +       switch (cmd) {
 421 +       case BLKTRACESETUP:
 422 +               ret = blk_trace_setup(q, bdev, arg);
 423 +               break;
 424 +       case BLKTRACESTART:
 425 +               start = 1;
 426 +       case BLKTRACESTOP:
 427 +               ret = blk_trace_startstop(q, start);
 428 +               break;
 429 +       case BLKTRACETEARDOWN:
 430 +               ret = blk_trace_remove(q);
 431 +               break;
 432 +       default:
 433 +               ret = -ENOTTY;
 434 +               break;
 435 +       }
 436 +
 437 +       up(&bdev->bd_sem);
 438 +       return ret;
 439 +}
 440 +
 441 +/**
 442 + * blk_trace_shutdown: - stop and cleanup trace structures
 443 + * @q:    the request queue associated with the device
 444 + *
 445 + **/
 446 +void blk_trace_shutdown(request_queue_t *q)
 447 +{
 448 +       blk_trace_startstop(q, 0);
 449 +       blk_trace_remove(q);
 450 +}
 451 +
 452 +/*
 453 + * Average offset over two calls to sched_clock() with a gettimeofday()
 454 + * in the middle
 455 + */
 456 +static void blk_check_time(unsigned long long *t)
 457 +{
 458 +       unsigned long long a, b;
 459 +       struct timeval tv;
 460 +
 461 +       a = sched_clock();
 462 +       do_gettimeofday(&tv);
 463 +       b = sched_clock();
 464 +
 465 +       *t = tv.tv_sec * 1000000000 + tv.tv_usec * 1000;
 466 +       *t -= (a + b) / 2;
 467 +}
 468 +
 469 +static void blk_trace_check_cpu_time(void *data)
 470 +{
 471 +       unsigned long long *t;
 472 +       int cpu = get_cpu();
 473 +
 474 +       t = &per_cpu(blk_trace_cpu_offset, cpu);
 475 +
 476 +       /*
 477 +        * Just call it twice, hopefully the second call will be cache hot
 478 +        * and a little more precise
 479 +        */
 480 +       blk_check_time(t);
 481 +       blk_check_time(t);
 482 +
 483 +       put_cpu();
 484 +}
 485 +
 486 +/*
 487 + * Call blk_trace_check_cpu_time() on each CPU to calibrate our inter-CPU
 488 + * timings
 489 + */
 490 +static void blk_trace_calibrate_offsets(void)
 491 +{
 492 +       unsigned long flags;
 493 +
 494 +       smp_call_function(blk_trace_check_cpu_time, NULL, 1, 1);
 495 +       local_irq_save(flags);
 496 +       blk_trace_check_cpu_time(NULL);
 497 +       local_irq_restore(flags);
 498 +}
 499 +
 500 +static void blk_trace_set_ht_offsets(void)
 501 +{
 502 +#if defined(CONFIG_SCHED_SMT)
 503 +       int cpu, i;
 504 +
 505 +       /*
 506 +        * now make sure HT siblings have the same time offset
 507 +        */
 508 +       preempt_disable();
 509 +       for_each_online_cpu(cpu) {
 510 +               unsigned long long *cpu_off, *sibling_off;
 511 +
 512 +               for_each_cpu_mask(i, cpu_sibling_map[cpu]) {
 513 +                       if (i == cpu)
 514 +                               continue;
 515 +
 516 +                       cpu_off = &per_cpu(blk_trace_cpu_offset, cpu);
 517 +                       sibling_off = &per_cpu(blk_trace_cpu_offset, i);
 518 +                       *sibling_off = *cpu_off;
 519 +               }
 520 +       }
 521 +       preempt_enable();
 522 +#endif
 523 +}
 524 +
 525 +static __init int blk_trace_init(void)
 526 +{
 527 +       mutex_init(&blk_tree_mutex);
 528 +       blk_trace_calibrate_offsets();
 529 +       blk_trace_set_ht_offsets();
 530 +
 531 +       return 0;
 532 +}
 533 +
 534 +module_init(blk_trace_init);
 535 +
 536 diff --git a/block/elevator.c b/block/elevator.c
 537 index 24b702d..0c9fafe 100644
 538 --- a/block/elevator.c
 539 +++ b/block/elevator.c
 540 @@ -33,6 +33,7 @@
 541  #include <linux/init.h>
 542  #include <linux/compiler.h>
 543  #include <linux/delay.h>
 544 +#include <linux/blktrace_api.h>
 545
 546  #include <asm/uaccess.h>
 547
 548 @@ -315,6 +316,8 @@ void elv_insert(request_queue_t *q, stru
 549         struct list_head *pos;
 550         unsigned ordseq;
 551
 552 +       blk_add_trace_rq(q, rq, BLK_TA_INSERT);
 553 +
 554         rq->q = q;
 555
 556         switch (where) {
 557 @@ -481,6 +484,7 @@ struct request *elv_next_request(request
 558                          * not be passed by new incoming requests
 559                          */
 560                         rq->flags |= REQ_STARTED;
 561 +                       blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
 562                 }
 563
 564                 if (!q->boundary_rq || q->boundary_rq == rq) {
 565 diff --git a/block/ioctl.c b/block/ioctl.c
 566 index e110949..7acb56c 100644
 567 --- a/block/ioctl.c
 568 +++ b/block/ioctl.c
 569 @@ -5,6 +5,7 @@
 570  #include <linux/backing-dev.h>
 571  #include <linux/buffer_head.h>
 572  #include <linux/smp_lock.h>
 573 +#include <linux/blktrace_api.h>
 574  #include <asm/uaccess.h>
 575
 576  static int blkpg_ioctl(struct block_device *bdev, struct blkpg_ioctl_arg __user *arg)
 577 @@ -189,6 +190,11 @@ static int blkdev_locked_ioctl(struct fi
 578                 return put_ulong(arg, bdev->bd_inode->i_size >> 9);
 579         case BLKGETSIZE64:
 580                 return put_u64(arg, bdev->bd_inode->i_size);
 581 +       case BLKTRACESTART:
 582 +       case BLKTRACESTOP:
 583 +       case BLKTRACESETUP:
 584 +       case BLKTRACETEARDOWN:
 585 +               return blk_trace_ioctl(bdev, cmd, (char __user *) arg);
 586         }
 587         return -ENOIOCTLCMD;
 588  }
 589 diff --git a/block/ll_rw_blk.c b/block/ll_rw_blk.c
 590 index 03d9c82..41387f5 100644
 591 --- a/block/ll_rw_blk.c
 592 +++ b/block/ll_rw_blk.c
 593 @@ -28,6 +28,7 @@
 594  #include <linux/writeback.h>
 595  #include <linux/interrupt.h>
 596  #include <linux/cpu.h>
 597 +#include <linux/blktrace_api.h>
 598
 599  /*
 600   * for max sense size
 601 @@ -1551,8 +1552,10 @@ void blk_plug_device(request_queue_t *q)
 602         if (test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags))
 603                 return;
 604
 605 -       if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
 606 +       if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags)) {
 607                 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
 608 +               blk_add_trace_generic(q, NULL, 0, BLK_TA_PLUG);
 609 +       }
 610  }
 611
 612  EXPORT_SYMBOL(blk_plug_device);
 613 @@ -1616,14 +1619,21 @@ static void blk_backing_dev_unplug(struc
 614         /*
 615          * devices don't necessarily have an ->unplug_fn defined
 616          */
 617 -       if (q->unplug_fn)
 618 +       if (q->unplug_fn) {
 619 +               blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
 620 +                                       q->rq.count[READ] + q->rq.count[WRITE]);
 621 +
 622                 q->unplug_fn(q);
 623 +       }
 624  }
 625
 626  static void blk_unplug_work(void *data)
 627  {
 628         request_queue_t *q = data;
 629
 630 +       blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
 631 +                               q->rq.count[READ] + q->rq.count[WRITE]);
 632 +
 633         q->unplug_fn(q);
 634  }
 635
 636 @@ -1631,6 +1641,9 @@ static void blk_unplug_timeout(unsigned
 637  {
 638         request_queue_t *q = (request_queue_t *)data;
 639
 640 +       blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_TIMER, NULL,
 641 +                               q->rq.count[READ] + q->rq.count[WRITE]);
 642 +
 643         kblockd_schedule_work(&q->unplug_work);
 644  }
 645
 646 @@ -1753,6 +1766,9 @@ void blk_cleanup_queue(request_queue_t *
 647         if (q->queue_tags)
 648                 __blk_queue_free_tags(q);
 649
 650 +       if (q->blk_trace)
 651 +               blk_trace_shutdown(q);
 652 +
 653         kmem_cache_free(requestq_cachep, q);
 654  }
 655
 656 @@ -2104,6 +2120,8 @@ rq_starved:
 657
 658         rq_init(q, rq);
 659         rq->rl = rl;
 660 +
 661 +       blk_add_trace_generic(q, bio, rw, BLK_TA_GETRQ);
 662  out:
 663         return rq;
 664  }
 665 @@ -2132,6 +2150,8 @@ static struct request *get_request_wait(
 666                 if (!rq) {
 667                         struct io_context *ioc;
 668
 669 +                       blk_add_trace_generic(q, bio, rw, BLK_TA_SLEEPRQ);
 670 +
 671                         __generic_unplug_device(q);
 672                         spin_unlock_irq(q->queue_lock);
 673                         io_schedule();
 674 @@ -2185,6 +2205,8 @@ EXPORT_SYMBOL(blk_get_request);
 675   */
 676  void blk_requeue_request(request_queue_t *q, struct request *rq)
 677  {
 678 +       blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
 679 +
 680         if (blk_rq_tagged(rq))
 681                 blk_queue_end_tag(q, rq);
 682
 683 @@ -2819,6 +2841,8 @@ static int __make_request(request_queue_
 684                         if (!q->back_merge_fn(q, req, bio))
 685                                 break;
 686
 687 +                       blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
 688 +
 689                         req->biotail->bi_next = bio;
 690                         req->biotail = bio;
 691                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
 692 @@ -2834,6 +2858,8 @@ static int __make_request(request_queue_
 693                         if (!q->front_merge_fn(q, req, bio))
 694                                 break;
 695
 696 +                       blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
 697 +
 698                         bio->bi_next = req->bio;
 699                         req->bio = bio;
 700
 701 @@ -2951,6 +2977,7 @@ void generic_make_request(struct bio *bi
 702         request_queue_t *q;
 703         sector_t maxsector;
 704         int ret, nr_sectors = bio_sectors(bio);
 705 +       dev_t old_dev;
 706
 707         might_sleep();
 708         /* Test device or partition size, when known. */
 709 @@ -2977,6 +3004,8 @@ void generic_make_request(struct bio *bi
 710          * NOTE: we don't repeat the blk_size check for each new device.
 711          * Stacking drivers are expected to know what they are doing.
 712          */
 713 +       maxsector = -1;
 714 +       old_dev = 0;
 715         do {
 716                 char b[BDEVNAME_SIZE];
 717
 718 @@ -3009,6 +3038,15 @@ end_io:
 719                  */
 720                 blk_partition_remap(bio);
 721
 722 +               if (maxsector != -1)
 723 +                       blk_add_trace_remap(q, bio, old_dev, bio->bi_sector,
 724 +                                           maxsector);
 725 +
 726 +               blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
 727 +
 728 +               maxsector = bio->bi_sector;
 729 +               old_dev = bio->bi_bdev->bd_dev;
 730 +
 731                 ret = q->make_request_fn(q, bio);
 732         } while (ret);
 733  }
 734 @@ -3128,6 +3166,8 @@ static int __end_that_request_first(stru
 735         int total_bytes, bio_nbytes, error, next_idx = 0;
 736         struct bio *bio;
 737
 738 +       blk_add_trace_rq(req->q, req, BLK_TA_COMPLETE);
 739 +
 740         /*
 741          * extend uptodate bool to allow < 0 value to be direct io error
 742          */
 743 diff --git a/drivers/block/cciss.c b/drivers/block/cciss.c
 744 index 0d65394..234048e 100644
 745 --- a/drivers/block/cciss.c
 746 +++ b/drivers/block/cciss.c
 747 @@ -38,6 +38,7 @@
 748  #include <linux/hdreg.h>
 749  #include <linux/spinlock.h>
 750  #include <linux/compat.h>
 751 +#include <linux/blktrace_api.h>
 752  #include <asm/uaccess.h>
 753  #include <asm/io.h>
 754
 755 @@ -2331,6 +2332,7 @@ static inline void complete_command( ctl
 756
 757         cmd->rq->completion_data = cmd;
 758         cmd->rq->errors = status;
 759 +       blk_add_trace_rq(cmd->rq->q, cmd->rq, BLK_TA_COMPLETE);
 760         blk_complete_request(cmd->rq);
 761  }
 762
 763 diff --git a/drivers/md/dm.c b/drivers/md/dm.c
 764 index e9adeb9..c8f3aa2 100644
 765 --- a/drivers/md/dm.c
 766 +++ b/drivers/md/dm.c
 767 @@ -17,6 +17,7 @@
 768  #include <linux/mempool.h>
 769  #include <linux/slab.h>
 770  #include <linux/idr.h>
 771 +#include <linux/blktrace_api.h>
 772
 773  static const char *_name = DM_NAME;
 774
 775 @@ -334,6 +335,8 @@ static void dec_pending(struct dm_io *io
 776                         /* nudge anyone waiting on suspend queue */
 777                         wake_up(&io->md->wait);
 778
 779 +               blk_add_trace_bio(io->md->queue, io->bio, BLK_TA_COMPLETE);
 780 +
 781                 bio_endio(io->bio, io->bio->bi_size, io->error);
 782                 free_io(io->md, io);
 783         }
 784 @@ -392,6 +395,7 @@ static void __map_bio(struct dm_target *
 785                       struct target_io *tio)
 786  {
 787         int r;
 788 +       sector_t sector;
 789
 790         /*
 791          * Sanity checks.
 792 @@ -407,10 +411,17 @@ static void __map_bio(struct dm_target *
 793          * this io.
 794          */
 795         atomic_inc(&tio->io->io_count);
 796 +       sector = clone->bi_sector;
 797         r = ti->type->map(ti, clone, &tio->info);
 798 -       if (r > 0)
 799 +       if (r > 0) {
 800                 /* the bio has been remapped so dispatch it */
 801 +
 802 +               blk_add_trace_remap(bdev_get_queue(clone->bi_bdev), clone,
 803 +                                   tio->io->bio->bi_bdev->bd_dev, sector,
 804 +                                   clone->bi_sector);
 805 +
 806                 generic_make_request(clone);
 807 +       }
 808
 809         else if (r < 0) {
 810                 /* error the io and bail out */
 811 diff --git a/fs/bio.c b/fs/bio.c
 812 index 1f3bb50..0dd0d81 100644
 813 --- a/fs/bio.c
 814 +++ b/fs/bio.c
 815 @@ -25,6 +25,7 @@
 816  #include <linux/module.h>
 817  #include <linux/mempool.h>
 818  #include <linux/workqueue.h>
 819 +#include <linux/blktrace_api.h>
 820  #include <scsi/sg.h>           /* for struct sg_iovec */
 821
 822  #define BIO_POOL_SIZE 256
 823 @@ -1095,6 +1096,9 @@ struct bio_pair *bio_split(struct bio *b
 824         if (!bp)
 825                 return bp;
 826
 827 +       blk_add_trace_pdu_int(bdev_get_queue(bi->bi_bdev), BLK_TA_SPLIT, bi,
 828 +                               bi->bi_sector + first_sectors);
 829 +
 830         BUG_ON(bi->bi_vcnt != 1);
 831         BUG_ON(bi->bi_idx != 0);
 832         atomic_set(&bp->cnt, 3);
 833 diff --git a/fs/compat_ioctl.c b/fs/compat_ioctl.c
 834 index 057e602..893d600 100644
 835 --- a/fs/compat_ioctl.c
 836 +++ b/fs/compat_ioctl.c
 837 @@ -72,6 +72,7 @@
 838  #include <linux/i2c-dev.h>
 839  #include <linux/wireless.h>
 840  #include <linux/atalk.h>
 841 +#include <linux/blktrace_api.h>
 842
 843  #include <net/sock.h>          /* siocdevprivate_ioctl */
 844  #include <net/bluetooth/bluetooth.h>
 845 diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
 846 index 860e7a4..266ce9d 100644
 847 --- a/include/linux/blkdev.h
 848 +++ b/include/linux/blkdev.h
 849 @@ -22,6 +22,7 @@ typedef struct request_queue request_que
 850  struct elevator_queue;
 851  typedef struct elevator_queue elevator_t;
 852  struct request_pm_state;
 853 +struct blk_trace;
 854
 855  #define BLKDEV_MIN_RQ  4
 856  #define BLKDEV_MAX_RQ  128     /* Default maximum */
 857 @@ -416,6 +417,8 @@ struct request_queue
 858         unsigned int            sg_reserved_size;
 859         int                     node;
 860
 861 +       struct blk_trace        *blk_trace;
 862 +
 863         /*
 864          * reserved for flush operations
 865          */
 866 diff --git a/include/linux/blktrace_api.h b/include/linux/blktrace_api.h
 867 new file mode 100644
 868 index 0000000..fca40ef
 869 --- /dev/null
 870 +++ b/include/linux/blktrace_api.h
 871 @@ -0,0 +1,277 @@
 872 +#ifndef BLKTRACE_H
 873 +#define BLKTRACE_H
 874 +
 875 +#include <linux/config.h>
 876 +#include <linux/blkdev.h>
 877 +#include <linux/relayfs_fs.h>
 878 +
 879 +/*
 880 + * Trace categories
 881 + */
 882 +enum blktrace_cat {
 883 +       BLK_TC_READ     = 1 << 0,       /* reads */
 884 +       BLK_TC_WRITE    = 1 << 1,       /* writes */
 885 +       BLK_TC_BARRIER  = 1 << 2,       /* barrier */
 886 +       BLK_TC_SYNC     = 1 << 3,       /* barrier */
 887 +       BLK_TC_QUEUE    = 1 << 4,       /* queueing/merging */
 888 +       BLK_TC_REQUEUE  = 1 << 5,       /* requeueing */
 889 +       BLK_TC_ISSUE    = 1 << 6,       /* issue */
 890 +       BLK_TC_COMPLETE = 1 << 7,       /* completions */
 891 +       BLK_TC_FS       = 1 << 8,       /* fs requests */
 892 +       BLK_TC_PC       = 1 << 9,       /* pc requests */
 893 +       BLK_TC_NOTIFY   = 1 << 10,      /* special message */
 894 +
 895 +       BLK_TC_END      = 1 << 15,      /* only 16-bits, reminder */
 896 +};
 897 +
 898 +#define BLK_TC_SHIFT           (16)
 899 +#define BLK_TC_ACT(act)                ((act) << BLK_TC_SHIFT)
 900 +
 901 +/*
 902 + * Basic trace actions
 903 + */
 904 +enum blktrace_act {
 905 +       __BLK_TA_QUEUE = 1,             /* queued */
 906 +       __BLK_TA_BACKMERGE,             /* back merged to existing rq */
 907 +       __BLK_TA_FRONTMERGE,            /* front merge to existing rq */
 908 +       __BLK_TA_GETRQ,                 /* allocated new request */
 909 +       __BLK_TA_SLEEPRQ,               /* sleeping on rq allocation */
 910 +       __BLK_TA_REQUEUE,               /* request requeued */
 911 +       __BLK_TA_ISSUE,                 /* sent to driver */
 912 +       __BLK_TA_COMPLETE,              /* completed by driver */
 913 +       __BLK_TA_PLUG,                  /* queue was plugged */
 914 +       __BLK_TA_UNPLUG_IO,             /* queue was unplugged by io */
 915 +       __BLK_TA_UNPLUG_TIMER,          /* queue was unplugged by timer */
 916 +       __BLK_TA_INSERT,                /* insert request */
 917 +       __BLK_TA_SPLIT,                 /* bio was split */
 918 +       __BLK_TA_BOUNCE,                /* bio was bounced */
 919 +       __BLK_TA_REMAP,                 /* bio was remapped */
 920 +};
 921 +
 922 +/*
 923 + * Trace actions in full. Additionally, read or write is masked
 924 + */
 925 +#define BLK_TA_QUEUE           (__BLK_TA_QUEUE | BLK_TC_ACT(BLK_TC_QUEUE))
 926 +#define BLK_TA_BACKMERGE       (__BLK_TA_BACKMERGE | BLK_TC_ACT(BLK_TC_QUEUE))
 927 +#define BLK_TA_FRONTMERGE      (__BLK_TA_FRONTMERGE | BLK_TC_ACT(BLK_TC_QUEUE))
 928 +#define        BLK_TA_GETRQ            (__BLK_TA_GETRQ | BLK_TC_ACT(BLK_TC_QUEUE))
 929 +#define        BLK_TA_SLEEPRQ          (__BLK_TA_SLEEPRQ | BLK_TC_ACT(BLK_TC_QUEUE))
 930 +#define        BLK_TA_REQUEUE          (__BLK_TA_REQUEUE | BLK_TC_ACT(BLK_TC_REQUEUE))
 931 +#define BLK_TA_ISSUE           (__BLK_TA_ISSUE | BLK_TC_ACT(BLK_TC_ISSUE))
 932 +#define BLK_TA_COMPLETE                (__BLK_TA_COMPLETE| BLK_TC_ACT(BLK_TC_COMPLETE))
 933 +#define BLK_TA_PLUG            (__BLK_TA_PLUG | BLK_TC_ACT(BLK_TC_QUEUE))
 934 +#define BLK_TA_UNPLUG_IO       (__BLK_TA_UNPLUG_IO | BLK_TC_ACT(BLK_TC_QUEUE))
 935 +#define BLK_TA_UNPLUG_TIMER    (__BLK_TA_UNPLUG_TIMER | BLK_TC_ACT(BLK_TC_QUEUE))
 936 +#define BLK_TA_INSERT          (__BLK_TA_INSERT | BLK_TC_ACT(BLK_TC_QUEUE))
 937 +#define BLK_TA_SPLIT           (__BLK_TA_SPLIT)
 938 +#define BLK_TA_BOUNCE          (__BLK_TA_BOUNCE)
 939 +#define BLK_TA_REMAP           (__BLK_TA_REMAP | BLK_TC_ACT(BLK_TC_QUEUE))
 940 +
 941 +#define BLK_IO_TRACE_MAGIC     0x65617400
 942 +#define BLK_IO_TRACE_VERSION   0x07
 943 +
 944 +/*
 945 + * The trace itself
 946 + */
 947 +struct blk_io_trace {
 948 +       u32 magic;              /* MAGIC << 8 | version */
 949 +       u32 sequence;           /* event number */
 950 +       u64 time;               /* in microseconds */
 951 +       u64 sector;             /* disk offset */
 952 +       u32 bytes;              /* transfer length */
 953 +       u32 action;             /* what happened */
 954 +       u32 pid;                /* who did it */
 955 +       u32 device;             /* device number */
 956 +       u32 cpu;                /* on what cpu did it happen */
 957 +       u16 error;              /* completion error */
 958 +       u16 pdu_len;            /* length of data after this trace */
 959 +};
 960 +
 961 +/*
 962 + * The remap event
 963 + */
 964 +struct blk_io_trace_remap {
 965 +       u32 device;
 966 +       u32 __pad;
 967 +       u64 sector;
 968 +};
 969 +
 970 +enum {
 971 +       Blktrace_setup = 1,
 972 +       Blktrace_running,
 973 +       Blktrace_stopped,
 974 +};
 975 +
 976 +struct blk_trace {
 977 +       int trace_state;
 978 +       struct dentry *dir;
 979 +       struct rchan *rchan;
 980 +       struct dentry *dropped_file;
 981 +       atomic_t dropped;
 982 +       unsigned long *sequence;
 983 +       u32 dev;
 984 +       u16 act_mask;
 985 +       u64 start_lba;
 986 +       u64 end_lba;
 987 +       u32 pid;
 988 +};
 989 +
 990 +/*
 991 + * User setup structure passed with BLKTRACESTART
 992 + */
 993 +struct blk_user_trace_setup {
 994 +       char name[BDEVNAME_SIZE];       /* output */
 995 +       u16 act_mask;                   /* input */
 996 +       u32 buf_size;                   /* input */
 997 +       u32 buf_nr;                     /* input */
 998 +       u64 start_lba;
 999 +       u64 end_lba;
1000 +       u32 pid;
1001 +};
1002 +
1003 +#if defined(CONFIG_BLK_DEV_IO_TRACE)
1004 +extern int blk_trace_ioctl(struct block_device *, unsigned, char __user *);
1005 +extern void blk_trace_shutdown(request_queue_t *);
1006 +extern void __blk_add_trace(struct blk_trace *, sector_t, int, int, u32, int, int, void *);
1007 +
1008 +/**
1009 + * blk_add_trace_rq - Add a trace for a request oriented action
1010 + * @q:         queue the io is for
1011 + * @rq:                the source request
1012 + * @what:      the action
1013 + *
1014 + * Description:
1015 + *     Records an action against a request. Will log the bio offset + size.
1016 + *
1017 + **/
1018 +static inline void blk_add_trace_rq(struct request_queue *q, struct request *rq,
1019 +                                   u32 what)
1020 +{
1021 +       struct blk_trace *bt = q->blk_trace;
1022 +       int rw = rq->flags & 0x07;
1023 +
1024 +       if (likely(!bt))
1025 +               return;
1026 +
1027 +       if (blk_pc_request(rq)) {
1028 +               what |= BLK_TC_ACT(BLK_TC_PC);
1029 +               __blk_add_trace(bt, 0, rq->data_len, rw, what, rq->errors, sizeof(rq->cmd), rq->cmd);
1030 +       } else  {
1031 +               what |= BLK_TC_ACT(BLK_TC_FS);
1032 +               __blk_add_trace(bt, rq->hard_sector, rq->hard_nr_sectors << 9, rw, what, rq->errors, 0, NULL);
1033 +       }
1034 +}
1035 +
1036 +/**
1037 + * blk_add_trace_bio - Add a trace for a bio oriented action
1038 + * @q:         queue the io is for
1039 + * @bio:       the source bio
1040 + * @what:      the action
1041 + *
1042 + * Description:
1043 + *     Records an action against a bio. Will log the bio offset + size.
1044 + *
1045 + **/
1046 +static inline void blk_add_trace_bio(struct request_queue *q, struct bio *bio,
1047 +                                    u32 what)
1048 +{
1049 +       struct blk_trace *bt = q->blk_trace;
1050 +
1051 +       if (likely(!bt))
1052 +               return;
1053 +
1054 +       __blk_add_trace(bt, bio->bi_sector, bio->bi_size, bio->bi_rw, what, !bio_flagged(bio, BIO_UPTODATE), 0, NULL);
1055 +}
1056 +
1057 +/**
1058 + * blk_add_trace_generic - Add a trace for a generic action
1059 + * @q:         queue the io is for
1060 + * @bio:       the source bio
1061 + * @rw:                the data direction
1062 + * @what:      the action
1063 + *
1064 + * Description:
1065 + *     Records a simple trace
1066 + *
1067 + **/
1068 +static inline void blk_add_trace_generic(struct request_queue *q,
1069 +                                        struct bio *bio, int rw, u32 what)
1070 +{
1071 +       struct blk_trace *bt = q->blk_trace;
1072 +
1073 +       if (likely(!bt))
1074 +               return;
1075 +
1076 +       if (bio)
1077 +               blk_add_trace_bio(q, bio, what);
1078 +       else
1079 +               __blk_add_trace(bt, 0, 0, rw, what, 0, 0, NULL);
1080 +}
1081 +
1082 +/**
1083 + * blk_add_trace_pdu_int - Add a trace for a bio with an integer payload
1084 + * @q:         queue the io is for
1085 + * @what:      the action
1086 + * @bio:       the source bio
1087 + * @pdu:       the integer payload
1088 + *
1089 + * Description:
1090 + *     Adds a trace with some integer payload. This might be an unplug
1091 + *     option given as the action, with the depth at unplug time given
1092 + *     as the payload
1093 + *
1094 + **/
1095 +static inline void blk_add_trace_pdu_int(struct request_queue *q, u32 what,
1096 +                                        struct bio *bio, unsigned int pdu)
1097 +{
1098 +       struct blk_trace *bt = q->blk_trace;
1099 +       u64 rpdu = cpu_to_be64(pdu);
1100 +
1101 +       if (likely(!bt))
1102 +               return;
1103 +
1104 +       if (bio)
1105 +               __blk_add_trace(bt, bio->bi_sector, bio->bi_size, bio->bi_rw, what, !bio_flagged(bio, BIO_UPTODATE), sizeof(rpdu), &rpdu);
1106 +       else
1107 +               __blk_add_trace(bt, 0, 0, 0, what, 0, sizeof(rpdu), &rpdu);
1108 +}
1109 +
1110 +/**
1111 + * blk_add_trace_remap - Add a trace for a remap operation
1112 + * @q:         queue the io is for
1113 + * @bio:       the source bio
1114 + * @dev:       target device
1115 + * @from:      source sector
1116 + * @to:                target sector
1117 + *
1118 + * Description:
1119 + *     Device mapper or raid target sometimes need to split a bio because
1120 + *     it spans a stripe (or similar). Add a trace for that action.
1121 + *
1122 + **/
1123 +static inline void blk_add_trace_remap(struct request_queue *q, struct bio *bio,
1124 +                                      dev_t dev, sector_t from, sector_t to)
1125 +{
1126 +       struct blk_trace *bt = q->blk_trace;
1127 +       struct blk_io_trace_remap r;
1128 +
1129 +       if (likely(!bt))
1130 +               return;
1131 +
1132 +       r.device = cpu_to_be32(dev);
1133 +       r.sector = cpu_to_be64(to);
1134 +
1135 +       __blk_add_trace(bt, from, bio->bi_size, bio->bi_rw, BLK_TA_REMAP, !bio_flagged(bio, BIO_UPTODATE), sizeof(r), &r);
1136 +}
1137 +
1138 +#else /* !CONFIG_BLK_DEV_IO_TRACE */
1139 +#define blk_trace_ioctl(bdev, cmd, arg)                (-ENOTTY)
1140 +#define blk_trace_shutdown(q)                  do { } while (0)
1141 +#define blk_add_trace_rq(q, rq, what)          do { } while (0)
1142 +#define blk_add_trace_bio(q, rq, what)         do { } while (0)
1143 +#define blk_add_trace_generic(q, rq, rw, what) do { } while (0)
1144 +#define blk_add_trace_pdu_int(q, what, bio, pdu)       do { } while (0)
1145 +#define blk_add_trace_remap(q, bio, dev, f, t) do {} while (0)
1146 +#endif /* CONFIG_BLK_DEV_IO_TRACE */
1147 +
1148 +#endif
1149 diff --git a/include/linux/compat_ioctl.h b/include/linux/compat_ioctl.h
1150 index 8fad50f..a3ed64b 100644
1151 --- a/include/linux/compat_ioctl.h
1152 +++ b/include/linux/compat_ioctl.h
1153 @@ -97,6 +97,10 @@ COMPATIBLE_IOCTL(BLKRRPART)
1154  COMPATIBLE_IOCTL(BLKFLSBUF)
1155  COMPATIBLE_IOCTL(BLKSECTSET)
1156  COMPATIBLE_IOCTL(BLKSSZGET)
1157 +COMPATIBLE_IOCTL(BLKTRACESTART)
1158 +COMPATIBLE_IOCTL(BLKTRACESTOP)
1159 +COMPATIBLE_IOCTL(BLKTRACESETUP)
1160 +COMPATIBLE_IOCTL(BLKTRACETEARDOWN)
1161  ULONG_IOCTL(BLKRASET)
1162  ULONG_IOCTL(BLKFRASET)
1163  /* RAID */
1164 diff --git a/include/linux/fs.h b/include/linux/fs.h
1165 index e059da9..c7a63cd 100644
1166 --- a/include/linux/fs.h
1167 +++ b/include/linux/fs.h
1168 @@ -196,6 +196,10 @@ extern int dir_notify_enable;
1169  #define BLKBSZGET  _IOR(0x12,112,size_t)
1170  #define BLKBSZSET  _IOW(0x12,113,size_t)
1171  #define BLKGETSIZE64 _IOR(0x12,114,size_t)     /* return device size in bytes (u64 *arg) */
1172 +#define BLKTRACESETUP _IOWR(0x12,115,struct blk_user_trace_setup)
1173 +#define BLKTRACESTART _IO(0x12,116)
1174 +#define BLKTRACESTOP _IO(0x12,117)
1175 +#define BLKTRACETEARDOWN _IO(0x12,118)
1176
1177  #define BMAP_IOCTL 1           /* obsolete - kept for compatibility */
1178  #define FIBMAP    _IO(0x00,1)  /* bmap access */
1179 diff --git a/include/linux/sched.h b/include/linux/sched.h
1180 index 0cfcd1c..69cd119 100644
1181 --- a/include/linux/sched.h
1182 +++ b/include/linux/sched.h
1183 @@ -708,6 +708,7 @@ struct task_struct {
1184         prio_array_t *array;
1185
1186         unsigned short ioprio;
1187 +       unsigned int btrace_seq;
1188
1189         unsigned long sleep_avg;
1190         unsigned long long timestamp, last_ran;
1191 diff --git a/kernel/fork.c b/kernel/fork.c
1192 index 8e88b37..60f838f 100644
1193 --- a/kernel/fork.c
1194 +++ b/kernel/fork.c
1195 @@ -179,6 +179,7 @@ static struct task_struct *dup_task_stru
1196         /* One for us, one for whoever does the "release_task()" (usually parent) */
1197         atomic_set(&tsk->usage,2);
1198         atomic_set(&tsk->fs_excl, 0);
1199 +       tsk->btrace_seq = 0;
1200         return tsk;
1201  }
1202
1203 diff --git a/mm/highmem.c b/mm/highmem.c
1204 index ce2e7e8..d0ea1ee 100644
1205 --- a/mm/highmem.c
1206 +++ b/mm/highmem.c
1207 @@ -26,6 +26,7 @@
1208  #include <linux/init.h>
1209  #include <linux/hash.h>
1210  #include <linux/highmem.h>
1211 +#include <linux/blktrace_api.h>
1212  #include <asm/tlbflush.h>
1213
1214  static mempool_t *page_pool, *isa_page_pool;
1215 @@ -483,6 +484,8 @@ void blk_queue_bounce(request_queue_t *q
1216                 pool = isa_page_pool;
1217         }
1218
1219 +       blk_add_trace_bio(q, *bio_orig, BLK_TA_BOUNCE);
1220 +
1221         /*
1222          * slow path
1223          */