kernel/blk-trace-2.6.16-rc1-git-T0

   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..d12d166
  34 --- /dev/null
  35 +++ b/block/blktrace.c
  36 @@ -0,0 +1,360 @@
  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 +
  48 +void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
  49 +                    int rw, u32 what, int error, int pdu_len, void *pdu_data)
  50 +{
  51 +       struct blk_io_trace t;
  52 +       unsigned long flags;
  53 +       void *rbuf;
  54 +       pid_t pid;
  55 +       int cpu;
  56 +
  57 +       if (rw & (1 << BIO_RW_BARRIER))
  58 +               what |= BLK_TC_ACT(BLK_TC_BARRIER);
  59 +       if (rw & (1 << BIO_RW_SYNC))
  60 +               what |= BLK_TC_ACT(BLK_TC_SYNC);
  61 +
  62 +       if (rw & WRITE)
  63 +               what |= BLK_TC_ACT(BLK_TC_WRITE);
  64 +       else
  65 +               what |= BLK_TC_ACT(BLK_TC_READ);
  66 +
  67 +       if (((bt->act_mask << BLK_TC_SHIFT) & what) == 0)
  68 +               return;
  69 +       if (sector < bt->start_lba || sector > bt->end_lba)
  70 +               return;
  71 +
  72 +       pid = current->pid;
  73 +       if (bt->pid && pid != bt->pid)
  74 +               return;
  75 +
  76 +       t.magic         = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
  77 +
  78 +       t.device        = bt->dev;
  79 +       t.sector        = sector;
  80 +       t.bytes         = bytes;
  81 +       t.action        = what;
  82 +       t.error         = error;
  83 +       t.pdu_len       = pdu_len;
  84 +
  85 +       t.pid           = pid;
  86 +       memcpy(t.comm, current->comm, sizeof(t.comm));
  87 +
  88 +       /*
  89 +        * need to serialize this part completely to prevent multiple CPUs
  90 +        * from misordering events
  91 +        */
  92 +       spin_lock_irqsave(&bt->lock, flags);
  93 +
  94 +       t.sequence      = ++bt->sequence;
  95 +
  96 +       cpu = smp_processor_id();
  97 +       t.cpu           = cpu;
  98 +       t.time          = sched_clock() - per_cpu(blk_trace_cpu_offset, cpu);
  99 +
 100 +       rbuf = relay_reserve(bt->rchan, sizeof(t) + pdu_len);
 101 +       if (rbuf) {
 102 +               memcpy(rbuf, &t, sizeof(t));
 103 +               if (pdu_len)
 104 +                       memcpy(rbuf + sizeof(t), pdu_data, pdu_len);
 105 +       }
 106 +
 107 +       spin_unlock_irqrestore(&bt->lock, flags);
 108 +}
 109 +
 110 +EXPORT_SYMBOL_GPL(__blk_add_trace);
 111 +
 112 +static struct dentry *blk_tree_root;
 113 +static struct mutex blk_tree_mutex;
 114 +
 115 +static inline void blk_remove_root(void)
 116 +{
 117 +       if (relayfs_remove_dir(blk_tree_root) != -ENOTEMPTY)
 118 +               blk_tree_root = NULL;
 119 +}
 120 +
 121 +static void blk_remove_tree(struct dentry *dir)
 122 +{
 123 +       mutex_lock(&blk_tree_mutex);
 124 +       relayfs_remove_dir(dir);
 125 +       blk_remove_root();
 126 +       mutex_unlock(&blk_tree_mutex);
 127 +}
 128 +
 129 +static struct dentry *blk_create_tree(const char *blk_name)
 130 +{
 131 +       struct dentry *dir = NULL;
 132 +
 133 +       mutex_lock(&blk_tree_mutex);
 134 +
 135 +       if (!blk_tree_root) {
 136 +               blk_tree_root = relayfs_create_dir("block", NULL);
 137 +               if (!blk_tree_root)
 138 +                       goto err;
 139 +       }
 140 +
 141 +       dir = relayfs_create_dir(blk_name, blk_tree_root);
 142 +       if (!dir)
 143 +               blk_remove_root();
 144 +
 145 +err:
 146 +       mutex_unlock(&blk_tree_mutex);
 147 +       return dir;
 148 +}
 149 +
 150 +void blk_cleanup_trace(struct blk_trace *bt)
 151 +{
 152 +       relay_close(bt->rchan);
 153 +       relayfs_remove_file(bt->dropped_file);
 154 +       blk_remove_tree(bt->dir);
 155 +       kfree(bt);
 156 +}
 157 +
 158 +int blk_stop_trace(struct block_device *bdev)
 159 +{
 160 +       request_queue_t *q = bdev_get_queue(bdev);
 161 +       struct blk_trace *bt = NULL;
 162 +       int ret = -EINVAL;
 163 +
 164 +       if (!q)
 165 +               return -ENXIO;
 166 +
 167 +       down(&bdev->bd_sem);
 168 +
 169 +       if (q->blk_trace) {
 170 +               bt = q->blk_trace;
 171 +               q->blk_trace = NULL;
 172 +               ret = 0;
 173 +       }
 174 +
 175 +       up(&bdev->bd_sem);
 176 +
 177 +       if (bt)
 178 +               blk_cleanup_trace(bt);
 179 +
 180 +       return ret;
 181 +}
 182 +
 183 +static int blk_dropped_open(struct inode *inode, struct file *filp)
 184 +{
 185 +       filp->private_data = inode->u.generic_ip;
 186 +
 187 +       return 0;
 188 +}
 189 +
 190 +static ssize_t blk_dropped_read(struct file *filp, char __user *buffer,
 191 +                               size_t count, loff_t *ppos)
 192 +{
 193 +       struct blk_trace *bt = filp->private_data;
 194 +       char buf[16];
 195 +
 196 +       snprintf(buf, sizeof(buf), "%u\n", atomic_read(&bt->dropped));
 197 +
 198 +       return simple_read_from_buffer(buffer, count, ppos, buf, strlen(buf));
 199 +}
 200 +
 201 +static struct file_operations blk_dropped_fops = {
 202 +       .owner =        THIS_MODULE,
 203 +       .open =         blk_dropped_open,
 204 +       .read =         blk_dropped_read,
 205 +};
 206 +
 207 +static int blk_subbuf_start_callback (struct rchan_buf *buf,
 208 +                                     void *subbuf,
 209 +                                     void *prev_subbuf,
 210 +                                     size_t prev_padding)
 211 +{
 212 +       if (relay_buf_full(buf)) {
 213 +               struct blk_trace *bt = buf->chan->private_data;
 214 +
 215 +               atomic_inc(&bt->dropped);
 216 +               return 0;
 217 +       }
 218 +
 219 +       return 1;
 220 +}
 221 +
 222 +static struct rchan_callbacks blk_relay_callbacks = {
 223 +       .subbuf_start = blk_subbuf_start_callback,
 224 +};
 225 +
 226 +int blk_start_trace(struct block_device *bdev, char __user *arg)
 227 +{
 228 +       request_queue_t *q = bdev_get_queue(bdev);
 229 +       struct blk_user_trace_setup buts;
 230 +       struct blk_trace *bt = NULL;
 231 +       struct dentry *dir = NULL;
 232 +       char b[BDEVNAME_SIZE];
 233 +       int ret, i;
 234 +
 235 +       if (!q)
 236 +               return -ENXIO;
 237 +
 238 +       if (copy_from_user(&buts, arg, sizeof(buts)))
 239 +               return -EFAULT;
 240 +
 241 +       if (!buts.buf_size || !buts.buf_nr)
 242 +               return -EINVAL;
 243 +
 244 +       strcpy(buts.name, bdevname(bdev, b));
 245 +
 246 +       /*
 247 +        * some device names have larger paths - convert the slashes
 248 +        * to underscores for this to work as expected
 249 +        */
 250 +       for (i = 0; i < strlen(buts.name); i++)
 251 +               if (buts.name[i] == '/')
 252 +                       buts.name[i] = '_';
 253 +
 254 +       if (copy_to_user(arg, &buts, sizeof(buts)))
 255 +               return -EFAULT;
 256 +
 257 +       down(&bdev->bd_sem);
 258 +       ret = -EBUSY;
 259 +       if (q->blk_trace)
 260 +               goto err;
 261 +
 262 +       ret = -ENOMEM;
 263 +       bt = kmalloc(sizeof(*bt), GFP_KERNEL);
 264 +       if (!bt)
 265 +               goto err;
 266 +
 267 +       ret = -ENOENT;
 268 +       dir = blk_create_tree(buts.name);
 269 +       if (!dir)
 270 +               goto err;
 271 +
 272 +       bt->dir = dir;
 273 +       bt->dev = bdev->bd_dev;
 274 +       bt->sequence = 0;
 275 +       spin_lock_init(&bt->lock);
 276 +       atomic_set(&bt->dropped, 0);
 277 +
 278 +       ret = -EIO;
 279 +       bt->dropped_file = relayfs_create_file("dropped", dir, 0, &blk_dropped_fops, bt);
 280 +       if (!bt->dropped_file)
 281 +               goto err;
 282 +
 283 +       bt->rchan = relay_open("trace", dir, buts.buf_size, buts.buf_nr, &blk_relay_callbacks);
 284 +       if (!bt->rchan)
 285 +               goto err;
 286 +       bt->rchan->private_data = bt;
 287 +
 288 +       bt->act_mask = buts.act_mask;
 289 +       if (!bt->act_mask)
 290 +               bt->act_mask = (u16) -1;
 291 +
 292 +       bt->start_lba = buts.start_lba;
 293 +       bt->end_lba = buts.end_lba;
 294 +       if (!bt->end_lba)
 295 +               bt->end_lba = -1ULL;
 296 +
 297 +       bt->pid = buts.pid;
 298 +
 299 +       q->blk_trace = bt;
 300 +       up(&bdev->bd_sem);
 301 +       return 0;
 302 +err:
 303 +       up(&bdev->bd_sem);
 304 +       if (bt && bt->dropped_file)
 305 +               relayfs_remove_file(bt->dropped_file);
 306 +       if (dir)
 307 +               blk_remove_tree(dir);
 308 +       if (bt)
 309 +               kfree(bt);
 310 +       return ret;
 311 +}
 312 +
 313 +/*
 314 + * Average offset over two calls to sched_clock() with a gettimeofday()
 315 + * in the middle
 316 + */
 317 +static void blk_check_time(unsigned long long *t)
 318 +{
 319 +       unsigned long long a, b;
 320 +       struct timeval tv;
 321 +
 322 +       a = sched_clock();
 323 +       do_gettimeofday(&tv);
 324 +       b = sched_clock();
 325 +
 326 +       *t = tv.tv_sec * 1000000000 + tv.tv_usec * 1000;
 327 +       *t -= (a + b) / 2;
 328 +}
 329 +
 330 +static void blk_trace_check_cpu_time(void *data)
 331 +{
 332 +       unsigned long long *t;
 333 +       int cpu = get_cpu();
 334 +
 335 +       t = &per_cpu(blk_trace_cpu_offset, cpu);
 336 +
 337 +       /*
 338 +        * Just call it twice, hopefully the second call will be cache hot
 339 +        * and a little more precise
 340 +        */
 341 +       blk_check_time(t);
 342 +       blk_check_time(t);
 343 +
 344 +       put_cpu();
 345 +}
 346 +
 347 +/*
 348 + * Call blk_trace_check_cpu_time() on each CPU to calibrate our inter-CPU
 349 + * timings
 350 + */
 351 +static void blk_trace_calibrate_offsets(void)
 352 +{
 353 +       unsigned long flags;
 354 +
 355 +       smp_call_function(blk_trace_check_cpu_time, NULL, 1, 1);
 356 +       local_irq_save(flags);
 357 +       blk_trace_check_cpu_time(NULL);
 358 +       local_irq_restore(flags);
 359 +}
 360 +
 361 +static void blk_trace_set_ht_offsets(void)
 362 +{
 363 +#if defined(CONFIG_SCHED_SMT)
 364 +       int cpu, i;
 365 +
 366 +       /*
 367 +        * now make sure HT siblings have the same time offset
 368 +        */
 369 +       preempt_disable();
 370 +       for_each_online_cpu(cpu) {
 371 +               unsigned long long *cpu_off, *sibling_off;
 372 +
 373 +               for_each_cpu_mask(i, cpu_sibling_map[cpu]) {
 374 +                       if (i == cpu)
 375 +                               continue;
 376 +
 377 +                       cpu_off = &per_cpu(blk_trace_cpu_offset, cpu);
 378 +                       sibling_off = &per_cpu(blk_trace_cpu_offset, i);
 379 +                       *sibling_off = *cpu_off;
 380 +               }
 381 +       }
 382 +       preempt_enable();
 383 +#endif
 384 +}
 385 +
 386 +static __init int blk_trace_init(void)
 387 +{
 388 +       mutex_init(&blk_tree_mutex);
 389 +       blk_trace_calibrate_offsets();
 390 +       blk_trace_set_ht_offsets();
 391 +
 392 +       return 0;
 393 +}
 394 +
 395 +module_init(blk_trace_init);
 396 +
 397 diff --git a/block/elevator.c b/block/elevator.c
 398 index c9f424d..793c686 100644
 399 --- a/block/elevator.c
 400 +++ b/block/elevator.c
 401 @@ -33,6 +33,7 @@
 402  #include <linux/init.h>
 403  #include <linux/compiler.h>
 404  #include <linux/delay.h>
 405 +#include <linux/blktrace_api.h>
 406
 407  #include <asm/uaccess.h>
 408
 409 @@ -333,6 +334,8 @@ void __elv_add_request(request_queue_t *
 410         struct list_head *pos;
 411         unsigned ordseq;
 412
 413 +       blk_add_trace_rq(q, rq, BLK_TA_INSERT);
 414 +
 415         if (q->ordcolor)
 416                 rq->flags |= REQ_ORDERED_COLOR;
 417
 418 @@ -491,6 +494,7 @@ struct request *elv_next_request(request
 419                          * not be passed by new incoming requests
 420                          */
 421                         rq->flags |= REQ_STARTED;
 422 +                       blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
 423                 }
 424
 425                 if (!q->boundary_rq || q->boundary_rq == rq) {
 426 diff --git a/block/ioctl.c b/block/ioctl.c
 427 index e110949..63e67a2 100644
 428 --- a/block/ioctl.c
 429 +++ b/block/ioctl.c
 430 @@ -5,6 +5,7 @@
 431  #include <linux/backing-dev.h>
 432  #include <linux/buffer_head.h>
 433  #include <linux/smp_lock.h>
 434 +#include <linux/blktrace_api.h>
 435  #include <asm/uaccess.h>
 436
 437  static int blkpg_ioctl(struct block_device *bdev, struct blkpg_ioctl_arg __user *arg)
 438 @@ -189,6 +190,10 @@ static int blkdev_locked_ioctl(struct fi
 439                 return put_ulong(arg, bdev->bd_inode->i_size >> 9);
 440         case BLKGETSIZE64:
 441                 return put_u64(arg, bdev->bd_inode->i_size);
 442 +       case BLKSTARTTRACE:
 443 +               return blk_start_trace(bdev, (char __user *) arg);
 444 +       case BLKSTOPTRACE:
 445 +               return blk_stop_trace(bdev);
 446         }
 447         return -ENOIOCTLCMD;
 448  }
 449 diff --git a/block/ll_rw_blk.c b/block/ll_rw_blk.c
 450 index 8e27d0a..bfcde0f 100644
 451 --- a/block/ll_rw_blk.c
 452 +++ b/block/ll_rw_blk.c
 453 @@ -28,6 +28,7 @@
 454  #include <linux/writeback.h>
 455  #include <linux/interrupt.h>
 456  #include <linux/cpu.h>
 457 +#include <linux/blktrace_api.h>
 458
 459  /*
 460   * for max sense size
 461 @@ -1555,8 +1556,10 @@ void blk_plug_device(request_queue_t *q)
 462         if (test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags))
 463                 return;
 464
 465 -       if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
 466 +       if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags)) {
 467                 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
 468 +               blk_add_trace_generic(q, NULL, 0, BLK_TA_PLUG);
 469 +       }
 470  }
 471
 472  EXPORT_SYMBOL(blk_plug_device);
 473 @@ -1620,14 +1623,21 @@ static void blk_backing_dev_unplug(struc
 474         /*
 475          * devices don't necessarily have an ->unplug_fn defined
 476          */
 477 -       if (q->unplug_fn)
 478 +       if (q->unplug_fn) {
 479 +               blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
 480 +                                       q->rq.count[READ] + q->rq.count[WRITE]);
 481 +
 482                 q->unplug_fn(q);
 483 +       }
 484  }
 485
 486  static void blk_unplug_work(void *data)
 487  {
 488         request_queue_t *q = data;
 489
 490 +       blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
 491 +                               q->rq.count[READ] + q->rq.count[WRITE]);
 492 +
 493         q->unplug_fn(q);
 494  }
 495
 496 @@ -1635,6 +1645,9 @@ static void blk_unplug_timeout(unsigned
 497  {
 498         request_queue_t *q = (request_queue_t *)data;
 499
 500 +       blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_TIMER, NULL,
 501 +                               q->rq.count[READ] + q->rq.count[WRITE]);
 502 +
 503         kblockd_schedule_work(&q->unplug_work);
 504  }
 505
 506 @@ -1757,6 +1770,11 @@ void blk_cleanup_queue(request_queue_t *
 507         if (q->queue_tags)
 508                 __blk_queue_free_tags(q);
 509
 510 +       if (q->blk_trace) {
 511 +               blk_cleanup_trace(q->blk_trace);
 512 +               q->blk_trace = NULL;
 513 +       }
 514 +
 515         kmem_cache_free(requestq_cachep, q);
 516  }
 517
 518 @@ -2108,6 +2126,8 @@ rq_starved:
 519
 520         rq_init(q, rq);
 521         rq->rl = rl;
 522 +
 523 +       blk_add_trace_generic(q, bio, rw, BLK_TA_GETRQ);
 524  out:
 525         return rq;
 526  }
 527 @@ -2136,6 +2156,8 @@ static struct request *get_request_wait(
 528                 if (!rq) {
 529                         struct io_context *ioc;
 530
 531 +                       blk_add_trace_generic(q, bio, rw, BLK_TA_SLEEPRQ);
 532 +
 533                         __generic_unplug_device(q);
 534                         spin_unlock_irq(q->queue_lock);
 535                         io_schedule();
 536 @@ -2189,6 +2211,8 @@ EXPORT_SYMBOL(blk_get_request);
 537   */
 538  void blk_requeue_request(request_queue_t *q, struct request *rq)
 539  {
 540 +       blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
 541 +
 542         if (blk_rq_tagged(rq))
 543                 blk_queue_end_tag(q, rq);
 544
 545 @@ -2820,6 +2844,8 @@ static int __make_request(request_queue_
 546                         if (!q->back_merge_fn(q, req, bio))
 547                                 break;
 548
 549 +                       blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
 550 +
 551                         req->biotail->bi_next = bio;
 552                         req->biotail = bio;
 553                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
 554 @@ -2835,6 +2861,8 @@ static int __make_request(request_queue_
 555                         if (!q->front_merge_fn(q, req, bio))
 556                                 break;
 557
 558 +                       blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
 559 +
 560                         bio->bi_next = req->bio;
 561                         req->bio = bio;
 562
 563 @@ -2952,6 +2980,7 @@ void generic_make_request(struct bio *bi
 564         request_queue_t *q;
 565         sector_t maxsector;
 566         int ret, nr_sectors = bio_sectors(bio);
 567 +       dev_t old_dev;
 568
 569         might_sleep();
 570         /* Test device or partition size, when known. */
 571 @@ -2978,6 +3007,8 @@ void generic_make_request(struct bio *bi
 572          * NOTE: we don't repeat the blk_size check for each new device.
 573          * Stacking drivers are expected to know what they are doing.
 574          */
 575 +       maxsector = -1;
 576 +       old_dev = 0;
 577         do {
 578                 char b[BDEVNAME_SIZE];
 579
 580 @@ -3010,6 +3041,15 @@ end_io:
 581                  */
 582                 blk_partition_remap(bio);
 583
 584 +               if (maxsector != -1)
 585 +                       blk_add_trace_remap(q, bio, old_dev, bio->bi_sector,
 586 +                                           maxsector);
 587 +
 588 +               blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
 589 +
 590 +               maxsector = bio->bi_sector;
 591 +               old_dev = bio->bi_bdev->bd_dev;
 592 +
 593                 ret = q->make_request_fn(q, bio);
 594         } while (ret);
 595  }
 596 @@ -3129,6 +3169,8 @@ static int __end_that_request_first(stru
 597         int total_bytes, bio_nbytes, error, next_idx = 0;
 598         struct bio *bio;
 599
 600 +       blk_add_trace_rq(req->q, req, BLK_TA_COMPLETE);
 601 +
 602         /*
 603          * extend uptodate bool to allow < 0 value to be direct io error
 604          */
 605 diff --git a/drivers/block/cciss.c b/drivers/block/cciss.c
 606 index 12d7b9b..880892a 100644
 607 --- a/drivers/block/cciss.c
 608 +++ b/drivers/block/cciss.c
 609 @@ -38,6 +38,7 @@
 610  #include <linux/hdreg.h>
 611  #include <linux/spinlock.h>
 612  #include <linux/compat.h>
 613 +#include <linux/blktrace_api.h>
 614  #include <asm/uaccess.h>
 615  #include <asm/io.h>
 616
 617 @@ -2330,6 +2331,7 @@ static inline void complete_command( ctl
 618
 619         cmd->rq->completion_data = cmd;
 620         cmd->rq->errors = status;
 621 +       blk_add_trace_rq(cmd->rq->q, cmd->rq, BLK_TA_COMPLETE);
 622         blk_complete_request(cmd->rq);
 623  }
 624
 625 diff --git a/drivers/md/dm.c b/drivers/md/dm.c
 626 index 8c16359..8c979c2 100644
 627 --- a/drivers/md/dm.c
 628 +++ b/drivers/md/dm.c
 629 @@ -17,6 +17,7 @@
 630  #include <linux/mempool.h>
 631  #include <linux/slab.h>
 632  #include <linux/idr.h>
 633 +#include <linux/blktrace_api.h>
 634
 635  static const char *_name = DM_NAME;
 636
 637 @@ -303,6 +304,8 @@ static void dec_pending(struct dm_io *io
 638                         /* nudge anyone waiting on suspend queue */
 639                         wake_up(&io->md->wait);
 640
 641 +               blk_add_trace_bio(io->md->queue, io->bio, BLK_TA_COMPLETE);
 642 +
 643                 bio_endio(io->bio, io->bio->bi_size, io->error);
 644                 free_io(io->md, io);
 645         }
 646 @@ -361,6 +364,7 @@ static void __map_bio(struct dm_target *
 647                       struct target_io *tio)
 648  {
 649         int r;
 650 +       sector_t sector;
 651
 652         /*
 653          * Sanity checks.
 654 @@ -376,10 +380,17 @@ static void __map_bio(struct dm_target *
 655          * this io.
 656          */
 657         atomic_inc(&tio->io->io_count);
 658 +       sector = clone->bi_sector;
 659         r = ti->type->map(ti, clone, &tio->info);
 660 -       if (r > 0)
 661 +       if (r > 0) {
 662                 /* the bio has been remapped so dispatch it */
 663 +
 664 +               blk_add_trace_remap(bdev_get_queue(clone->bi_bdev), clone,
 665 +                                   tio->io->bio->bi_bdev->bd_dev, sector,
 666 +                                   clone->bi_sector);
 667 +
 668                 generic_make_request(clone);
 669 +       }
 670
 671         else if (r < 0) {
 672                 /* error the io and bail out */
 673 diff --git a/fs/bio.c b/fs/bio.c
 674 index bbc442b..8a1b0b6 100644
 675 --- a/fs/bio.c
 676 +++ b/fs/bio.c
 677 @@ -25,6 +25,7 @@
 678  #include <linux/module.h>
 679  #include <linux/mempool.h>
 680  #include <linux/workqueue.h>
 681 +#include <linux/blktrace_api.h>
 682  #include <scsi/sg.h>           /* for struct sg_iovec */
 683
 684  #define BIO_POOL_SIZE 256
 685 @@ -1094,6 +1095,9 @@ struct bio_pair *bio_split(struct bio *b
 686         if (!bp)
 687                 return bp;
 688
 689 +       blk_add_trace_pdu_int(bdev_get_queue(bi->bi_bdev), BLK_TA_SPLIT, bi,
 690 +                               bi->bi_sector + first_sectors);
 691 +
 692         BUG_ON(bi->bi_vcnt != 1);
 693         BUG_ON(bi->bi_idx != 0);
 694         atomic_set(&bp->cnt, 3);
 695 diff --git a/fs/compat_ioctl.c b/fs/compat_ioctl.c
 696 index 5dd0207..010e02b 100644
 697 --- a/fs/compat_ioctl.c
 698 +++ b/fs/compat_ioctl.c
 699 @@ -72,6 +72,7 @@
 700  #include <linux/i2c-dev.h>
 701  #include <linux/wireless.h>
 702  #include <linux/atalk.h>
 703 +#include <linux/blktrace_api.h>
 704
 705  #include <net/sock.h>          /* siocdevprivate_ioctl */
 706  #include <net/bluetooth/bluetooth.h>
 707 diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
 708 index 02a585f..195c3b9 100644
 709 --- a/include/linux/blkdev.h
 710 +++ b/include/linux/blkdev.h
 711 @@ -22,6 +22,7 @@ typedef struct request_queue request_que
 712  struct elevator_queue;
 713  typedef struct elevator_queue elevator_t;
 714  struct request_pm_state;
 715 +struct blk_trace;
 716
 717  #define BLKDEV_MIN_RQ  4
 718  #define BLKDEV_MAX_RQ  128     /* Default maximum */
 719 @@ -416,6 +417,8 @@ struct request_queue
 720         unsigned int            sg_reserved_size;
 721         int                     node;
 722
 723 +       struct blk_trace        *blk_trace;
 724 +
 725         /*
 726          * reserved for flush operations
 727          */
 728 diff --git a/include/linux/blktrace_api.h b/include/linux/blktrace_api.h
 729 new file mode 100644
 730 index 0000000..d6b4317
 731 --- /dev/null
 732 +++ b/include/linux/blktrace_api.h
 733 @@ -0,0 +1,216 @@
 734 +#ifndef BLKTRACE_H
 735 +#define BLKTRACE_H
 736 +
 737 +#include <linux/config.h>
 738 +#include <linux/blkdev.h>
 739 +#include <linux/relayfs_fs.h>
 740 +
 741 +/*
 742 + * Trace categories
 743 + */
 744 +enum blktrace_cat {
 745 +       BLK_TC_READ     = 1 << 0,       /* reads */
 746 +       BLK_TC_WRITE    = 1 << 1,       /* writes */
 747 +       BLK_TC_BARRIER  = 1 << 2,       /* barrier */
 748 +       BLK_TC_SYNC     = 1 << 3,       /* barrier */
 749 +       BLK_TC_QUEUE    = 1 << 4,       /* queueing/merging */
 750 +       BLK_TC_REQUEUE  = 1 << 5,       /* requeueing */
 751 +       BLK_TC_ISSUE    = 1 << 6,       /* issue */
 752 +       BLK_TC_COMPLETE = 1 << 7,       /* completions */
 753 +       BLK_TC_FS       = 1 << 8,       /* fs requests */
 754 +       BLK_TC_PC       = 1 << 9,       /* pc requests */
 755 +
 756 +       BLK_TC_END      = 1 << 15,      /* only 16-bits, reminder */
 757 +};
 758 +
 759 +#define BLK_TC_SHIFT           (16)
 760 +#define BLK_TC_ACT(act)                ((act) << BLK_TC_SHIFT)
 761 +
 762 +/*
 763 + * Basic trace actions
 764 + */
 765 +enum blktrace_act {
 766 +       __BLK_TA_QUEUE = 1,             /* queued */
 767 +       __BLK_TA_BACKMERGE,             /* back merged to existing rq */
 768 +       __BLK_TA_FRONTMERGE,            /* front merge to existing rq */
 769 +       __BLK_TA_GETRQ,                 /* allocated new request */
 770 +       __BLK_TA_SLEEPRQ,               /* sleeping on rq allocation */
 771 +       __BLK_TA_REQUEUE,               /* request requeued */
 772 +       __BLK_TA_ISSUE,                 /* sent to driver */
 773 +       __BLK_TA_COMPLETE,              /* completed by driver */
 774 +       __BLK_TA_PLUG,                  /* queue was plugged */
 775 +       __BLK_TA_UNPLUG_IO,             /* queue was unplugged by io */
 776 +       __BLK_TA_UNPLUG_TIMER,          /* queue was unplugged by timer */
 777 +       __BLK_TA_INSERT,                /* insert request */
 778 +       __BLK_TA_SPLIT,                 /* bio was split */
 779 +       __BLK_TA_BOUNCE,                /* bio was bounced */
 780 +       __BLK_TA_REMAP,                 /* bio was remapped */
 781 +};
 782 +
 783 +/*
 784 + * Trace actions in full. Additionally, read or write is masked
 785 + */
 786 +#define BLK_TA_QUEUE           (__BLK_TA_QUEUE | BLK_TC_ACT(BLK_TC_QUEUE))
 787 +#define BLK_TA_BACKMERGE       (__BLK_TA_BACKMERGE | BLK_TC_ACT(BLK_TC_QUEUE))
 788 +#define BLK_TA_FRONTMERGE      (__BLK_TA_FRONTMERGE | BLK_TC_ACT(BLK_TC_QUEUE))
 789 +#define        BLK_TA_GETRQ            (__BLK_TA_GETRQ | BLK_TC_ACT(BLK_TC_QUEUE))
 790 +#define        BLK_TA_SLEEPRQ          (__BLK_TA_SLEEPRQ | BLK_TC_ACT(BLK_TC_QUEUE))
 791 +#define        BLK_TA_REQUEUE          (__BLK_TA_REQUEUE | BLK_TC_ACT(BLK_TC_REQUEUE))
 792 +#define BLK_TA_ISSUE           (__BLK_TA_ISSUE | BLK_TC_ACT(BLK_TC_ISSUE))
 793 +#define BLK_TA_COMPLETE                (__BLK_TA_COMPLETE| BLK_TC_ACT(BLK_TC_COMPLETE))
 794 +#define BLK_TA_PLUG            (__BLK_TA_PLUG | BLK_TC_ACT(BLK_TC_QUEUE))
 795 +#define BLK_TA_UNPLUG_IO       (__BLK_TA_UNPLUG_IO | BLK_TC_ACT(BLK_TC_QUEUE))
 796 +#define BLK_TA_UNPLUG_TIMER    (__BLK_TA_UNPLUG_TIMER | BLK_TC_ACT(BLK_TC_QUEUE))
 797 +#define BLK_TA_INSERT          (__BLK_TA_INSERT | BLK_TC_ACT(BLK_TC_QUEUE))
 798 +#define BLK_TA_SPLIT           (__BLK_TA_SPLIT)
 799 +#define BLK_TA_BOUNCE          (__BLK_TA_BOUNCE)
 800 +#define BLK_TA_REMAP           (__BLK_TA_REMAP | BLK_TC_ACT(BLK_TC_QUEUE))
 801 +
 802 +#define BLK_IO_TRACE_MAGIC     0x65617400
 803 +#define BLK_IO_TRACE_VERSION   0x05
 804 +
 805 +/*
 806 + * The trace itself
 807 + */
 808 +struct blk_io_trace {
 809 +       u32 magic;              /* MAGIC << 8 | version */
 810 +       u32 sequence;           /* event number */
 811 +       u64 time;               /* in microseconds */
 812 +       u64 sector;             /* disk offset */
 813 +       u32 bytes;              /* transfer length */
 814 +       u32 action;             /* what happened */
 815 +       u32 pid;                /* who did it */
 816 +       u32 cpu;                /* on what cpu did it happen */
 817 +       u16 error;              /* completion error */
 818 +       u16 pdu_len;            /* length of data after this trace */
 819 +       u32 device;             /* device number */
 820 +       char comm[16];          /* task command name (TASK_COMM_LEN) */
 821 +};
 822 +
 823 +/*
 824 + * The remap event
 825 + */
 826 +struct blk_io_trace_remap {
 827 +       u32 device;
 828 +       u32 __pad;
 829 +       u64 sector;
 830 +};
 831 +
 832 +struct blk_trace {
 833 +       struct dentry *dir;
 834 +       struct rchan *rchan;
 835 +       struct dentry *dropped_file;
 836 +       atomic_t dropped;
 837 +       spinlock_t lock;
 838 +       unsigned long sequence;
 839 +       u32 dev;
 840 +       u16 act_mask;
 841 +       u64 start_lba;
 842 +       u64 end_lba;
 843 +       u32 pid;
 844 +};
 845 +
 846 +/*
 847 + * User setup structure passed with BLKSTARTTRACE
 848 + */
 849 +struct blk_user_trace_setup {
 850 +       char name[BDEVNAME_SIZE];       /* output */
 851 +       u16 act_mask;                   /* input */
 852 +       u32 buf_size;                   /* input */
 853 +       u32 buf_nr;                     /* input */
 854 +       u64 start_lba;
 855 +       u64 end_lba;
 856 +       u32 pid;
 857 +};
 858 +
 859 +#if defined(CONFIG_BLK_DEV_IO_TRACE)
 860 +extern int blk_start_trace(struct block_device *, char __user *);
 861 +extern int blk_stop_trace(struct block_device *);
 862 +extern void blk_cleanup_trace(struct blk_trace *);
 863 +extern void __blk_add_trace(struct blk_trace *, sector_t, int, int, u32, int, int, void *);
 864 +
 865 +static inline void blk_add_trace_rq(struct request_queue *q, struct request *rq,
 866 +                                   u32 what)
 867 +{
 868 +       struct blk_trace *bt = q->blk_trace;
 869 +       int rw = rq->flags & 0x07;
 870 +
 871 +       if (likely(!bt))
 872 +               return;
 873 +
 874 +       if (blk_pc_request(rq)) {
 875 +               what |= BLK_TC_ACT(BLK_TC_PC);
 876 +               __blk_add_trace(bt, 0, rq->data_len, rw, what, rq->errors, sizeof(rq->cmd), rq->cmd);
 877 +       } else  {
 878 +               what |= BLK_TC_ACT(BLK_TC_FS);
 879 +               __blk_add_trace(bt, rq->hard_sector, rq->hard_nr_sectors << 9, rw, what, rq->errors, 0, NULL);
 880 +       }
 881 +}
 882 +
 883 +static inline void blk_add_trace_bio(struct request_queue *q, struct bio *bio,
 884 +                                    u32 what)
 885 +{
 886 +       struct blk_trace *bt = q->blk_trace;
 887 +
 888 +       if (likely(!bt))
 889 +               return;
 890 +
 891 +       __blk_add_trace(bt, bio->bi_sector, bio->bi_size, bio->bi_rw, what, !bio_flagged(bio, BIO_UPTODATE), 0, NULL);
 892 +}
 893 +
 894 +static inline void blk_add_trace_generic(struct request_queue *q,
 895 +                                        struct bio *bio, int rw, u32 what)
 896 +{
 897 +       struct blk_trace *bt = q->blk_trace;
 898 +
 899 +       if (likely(!bt))
 900 +               return;
 901 +
 902 +       if (bio)
 903 +               blk_add_trace_bio(q, bio, what);
 904 +       else
 905 +               __blk_add_trace(bt, 0, 0, rw, what, 0, 0, NULL);
 906 +}
 907 +
 908 +static inline void blk_add_trace_pdu_int(struct request_queue *q, u32 what,
 909 +                                        struct bio *bio, unsigned int pdu)
 910 +{
 911 +       struct blk_trace *bt = q->blk_trace;
 912 +       u64 rpdu = cpu_to_be64(pdu);
 913 +
 914 +       if (likely(!bt))
 915 +               return;
 916 +
 917 +       if (bio)
 918 +               __blk_add_trace(bt, bio->bi_sector, bio->bi_size, bio->bi_rw, what, !bio_flagged(bio, BIO_UPTODATE), sizeof(rpdu), &rpdu);
 919 +       else
 920 +               __blk_add_trace(bt, 0, 0, 0, what, 0, sizeof(rpdu), &rpdu);
 921 +}
 922 +
 923 +static inline void blk_add_trace_remap(struct request_queue *q, struct bio *bio,
 924 +                                      dev_t dev, sector_t from, sector_t to)
 925 +{
 926 +       struct blk_trace *bt = q->blk_trace;
 927 +       struct blk_io_trace_remap r;
 928 +
 929 +       if (likely(!bt))
 930 +               return;
 931 +
 932 +       r.device = cpu_to_be32(dev);
 933 +       r.sector = cpu_to_be64(to);
 934 +
 935 +       __blk_add_trace(bt, from, bio->bi_size, bio->bi_rw, BLK_TA_REMAP, !bio_flagged(bio, BIO_UPTODATE), sizeof(r), &r);
 936 +}
 937 +
 938 +#else /* !CONFIG_BLK_DEV_IO_TRACE */
 939 +#define blk_start_trace(bdev, arg)             (-EINVAL)
 940 +#define blk_stop_trace(bdev)                   (-EINVAL)
 941 +#define blk_cleanup_trace(bt)                  do { } while (0)
 942 +#define blk_add_trace_rq(q, rq, what)          do { } while (0)
 943 +#define blk_add_trace_bio(q, rq, what)         do { } while (0)
 944 +#define blk_add_trace_generic(q, rq, rw, what) do { } while (0)
 945 +#define blk_add_trace_pdu_int(q, what, bio, pdu)       do { } while (0)
 946 +#define blk_add_trace_remap(q, bio, dev, f, t) do {} while (0)
 947 +#endif /* CONFIG_BLK_DEV_IO_TRACE */
 948 +
 949 +#endif
 950 diff --git a/include/linux/compat_ioctl.h b/include/linux/compat_ioctl.h
 951 index 8fad50f..5bed09a 100644
 952 --- a/include/linux/compat_ioctl.h
 953 +++ b/include/linux/compat_ioctl.h
 954 @@ -97,6 +97,8 @@ COMPATIBLE_IOCTL(BLKRRPART)
 955  COMPATIBLE_IOCTL(BLKFLSBUF)
 956  COMPATIBLE_IOCTL(BLKSECTSET)
 957  COMPATIBLE_IOCTL(BLKSSZGET)
 958 +COMPATIBLE_IOCTL(BLKSTARTTRACE)
 959 +COMPATIBLE_IOCTL(BLKSTOPTRACE)
 960  ULONG_IOCTL(BLKRASET)
 961  ULONG_IOCTL(BLKFRASET)
 962  /* RAID */
 963 diff --git a/include/linux/fs.h b/include/linux/fs.h
 964 index b77f260..5575284 100644
 965 --- a/include/linux/fs.h
 966 +++ b/include/linux/fs.h
 967 @@ -196,6 +196,8 @@ extern int dir_notify_enable;
 968  #define BLKBSZGET  _IOR(0x12,112,size_t)
 969  #define BLKBSZSET  _IOW(0x12,113,size_t)
 970  #define BLKGETSIZE64 _IOR(0x12,114,size_t)     /* return device size in bytes (u64 *arg) */
 971 +#define BLKSTARTTRACE _IOWR(0x12,115,struct blk_user_trace_setup)
 972 +#define BLKSTOPTRACE _IO(0x12,116)
 973
 974  #define BMAP_IOCTL 1           /* obsolete - kept for compatibility */
 975  #define FIBMAP    _IO(0x00,1)  /* bmap access */
 976 diff --git a/mm/highmem.c b/mm/highmem.c
 977 index ce2e7e8..d0ea1ee 100644
 978 --- a/mm/highmem.c
 979 +++ b/mm/highmem.c
 980 @@ -26,6 +26,7 @@
 981  #include <linux/init.h>
 982  #include <linux/hash.h>
 983  #include <linux/highmem.h>
 984 +#include <linux/blktrace_api.h>
 985  #include <asm/tlbflush.h>
 986
 987  static mempool_t *page_pool, *isa_page_pool;
 988 @@ -483,6 +484,8 @@ void blk_queue_bounce(request_queue_t *q
 989                 pool = isa_page_pool;
 990         }
 991
 992 +       blk_add_trace_bio(q, *bio_orig, BLK_TA_BOUNCE);
 993 +
 994         /*
 995          * slow path
 996          */