kernel/blk-trace-2.6.16-rc1-git-U2

   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..6bfcdb8
  34 --- /dev/null
  35 +++ b/block/blktrace.c
  36 @@ -0,0 +1,382 @@
  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 +/*
  49 + * The worker for the various blk_add_trace*() types. Fills out a
  50 + * blk_io_trace structure and places it in a per-cpu subbuffer.
  51 + */
  52 +void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
  53 +                    int rw, u32 what, int error, int pdu_len, void *pdu_data)
  54 +{
  55 +       struct blk_io_trace *t;
  56 +       unsigned long flags;
  57 +       unsigned long *sequence;
  58 +       pid_t pid;
  59 +       int cpu;
  60 +
  61 +       if (rw & (1 << BIO_RW_BARRIER))
  62 +               what |= BLK_TC_ACT(BLK_TC_BARRIER);
  63 +       if (rw & (1 << BIO_RW_SYNC))
  64 +               what |= BLK_TC_ACT(BLK_TC_SYNC);
  65 +
  66 +       if (rw & WRITE)
  67 +               what |= BLK_TC_ACT(BLK_TC_WRITE);
  68 +       else
  69 +               what |= BLK_TC_ACT(BLK_TC_READ);
  70 +
  71 +       if (((bt->act_mask << BLK_TC_SHIFT) & what) == 0)
  72 +               return;
  73 +       if (sector < bt->start_lba || sector > bt->end_lba)
  74 +               return;
  75 +
  76 +       pid = current->pid;
  77 +       if (bt->pid && pid != bt->pid)
  78 +               return;
  79 +
  80 +       /*
  81 +        * A word about the locking here - we disable interrupts to reserve
  82 +        * some space in the relayfs per-cpu buffer, to prevent an irq
  83 +        * from coming in and stepping on our toes. Once reserved, it's
  84 +        * enough to get preemption disabled to prevent read of this data
  85 +        * before we are through filling it. get_cpu()/put_cpu() does this
  86 +        * for us
  87 +        */
  88 +       local_irq_save(flags);
  89 +
  90 +       t = relay_reserve(bt->rchan, sizeof(*t) + pdu_len);
  91 +       if (unlikely(!t)) {
  92 +               local_irq_restore(flags);
  93 +               return;
  94 +       }
  95 +
  96 +       cpu = get_cpu();
  97 +
  98 +       sequence = per_cpu_ptr(bt->sequence, cpu);
  99 +       t->sequence = ++(*sequence);
 100 +       t->time = sched_clock() - per_cpu(blk_trace_cpu_offset, cpu);
 101 +       t->cpu = cpu;
 102 +
 103 +       local_irq_restore(flags);
 104 +
 105 +       t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
 106 +       t->device = bt->dev;
 107 +       t->sector = sector;
 108 +       t->bytes = bytes;
 109 +       t->action = what;
 110 +       t->error = error;
 111 +       t->pdu_len = pdu_len;
 112 +
 113 +       t->pid = pid;
 114 +       strncpy(t->comm, current->comm, sizeof(t->comm));
 115 +
 116 +       if (pdu_len)
 117 +               memcpy((void *) t + sizeof(*t), pdu_data, pdu_len);
 118 +
 119 +       put_cpu();
 120 +}
 121 +
 122 +EXPORT_SYMBOL_GPL(__blk_add_trace);
 123 +
 124 +static struct dentry *blk_tree_root;
 125 +static struct mutex blk_tree_mutex;
 126 +
 127 +static inline void blk_remove_root(void)
 128 +{
 129 +       if (relayfs_remove_dir(blk_tree_root) != -ENOTEMPTY)
 130 +               blk_tree_root = NULL;
 131 +}
 132 +
 133 +static void blk_remove_tree(struct dentry *dir)
 134 +{
 135 +       mutex_lock(&blk_tree_mutex);
 136 +       relayfs_remove_dir(dir);
 137 +       blk_remove_root();
 138 +       mutex_unlock(&blk_tree_mutex);
 139 +}
 140 +
 141 +static struct dentry *blk_create_tree(const char *blk_name)
 142 +{
 143 +       struct dentry *dir = NULL;
 144 +
 145 +       mutex_lock(&blk_tree_mutex);
 146 +
 147 +       if (!blk_tree_root) {
 148 +               blk_tree_root = relayfs_create_dir("block", NULL);
 149 +               if (!blk_tree_root)
 150 +                       goto err;
 151 +       }
 152 +
 153 +       dir = relayfs_create_dir(blk_name, blk_tree_root);
 154 +       if (!dir)
 155 +               blk_remove_root();
 156 +
 157 +err:
 158 +       mutex_unlock(&blk_tree_mutex);
 159 +       return dir;
 160 +}
 161 +
 162 +void blk_cleanup_trace(struct blk_trace *bt)
 163 +{
 164 +       relay_close(bt->rchan);
 165 +       relayfs_remove_file(bt->dropped_file);
 166 +       blk_remove_tree(bt->dir);
 167 +       kfree(bt);
 168 +}
 169 +
 170 +int blk_stop_trace(struct block_device *bdev)
 171 +{
 172 +       request_queue_t *q = bdev_get_queue(bdev);
 173 +       struct blk_trace *bt = NULL;
 174 +       int ret = -EINVAL;
 175 +
 176 +       if (!q)
 177 +               return -ENXIO;
 178 +
 179 +       down(&bdev->bd_sem);
 180 +
 181 +       if (q->blk_trace) {
 182 +               bt = q->blk_trace;
 183 +               q->blk_trace = NULL;
 184 +               ret = 0;
 185 +       }
 186 +
 187 +       up(&bdev->bd_sem);
 188 +
 189 +       if (bt)
 190 +               blk_cleanup_trace(bt);
 191 +
 192 +       return ret;
 193 +}
 194 +
 195 +static int blk_dropped_open(struct inode *inode, struct file *filp)
 196 +{
 197 +       filp->private_data = inode->u.generic_ip;
 198 +
 199 +       return 0;
 200 +}
 201 +
 202 +static ssize_t blk_dropped_read(struct file *filp, char __user *buffer,
 203 +                               size_t count, loff_t *ppos)
 204 +{
 205 +       struct blk_trace *bt = filp->private_data;
 206 +       char buf[16];
 207 +
 208 +       snprintf(buf, sizeof(buf), "%u\n", atomic_read(&bt->dropped));
 209 +
 210 +       return simple_read_from_buffer(buffer, count, ppos, buf, strlen(buf));
 211 +}
 212 +
 213 +static struct file_operations blk_dropped_fops = {
 214 +       .owner =        THIS_MODULE,
 215 +       .open =         blk_dropped_open,
 216 +       .read =         blk_dropped_read,
 217 +};
 218 +
 219 +/*
 220 + * Keep track of how many times we encountered a full subbuffer, to aid
 221 + * the user space app in telling how many lost events there were.
 222 + */
 223 +static int blk_subbuf_start_callback(struct rchan_buf *buf, void *subbuf,
 224 +                                    void *prev_subbuf, size_t prev_padding)
 225 +{
 226 +       struct blk_trace *bt;
 227 +
 228 +       if (!relay_buf_full(buf))
 229 +               return 1;
 230 +
 231 +       bt = buf->chan->private_data;
 232 +       atomic_inc(&bt->dropped);
 233 +       return 0;
 234 +}
 235 +
 236 +static struct rchan_callbacks blk_relay_callbacks = {
 237 +       .subbuf_start = blk_subbuf_start_callback,
 238 +};
 239 +
 240 +/*
 241 + * Setup everything required to start tracing
 242 + */
 243 +int blk_start_trace(struct block_device *bdev, char __user *arg)
 244 +{
 245 +       request_queue_t *q = bdev_get_queue(bdev);
 246 +       struct blk_user_trace_setup buts;
 247 +       struct blk_trace *bt = NULL;
 248 +       struct dentry *dir = NULL;
 249 +       char b[BDEVNAME_SIZE];
 250 +       int ret, i;
 251 +
 252 +       if (!q)
 253 +               return -ENXIO;
 254 +
 255 +       if (copy_from_user(&buts, arg, sizeof(buts)))
 256 +               return -EFAULT;
 257 +
 258 +       if (!buts.buf_size || !buts.buf_nr)
 259 +               return -EINVAL;
 260 +
 261 +       strcpy(buts.name, bdevname(bdev, b));
 262 +
 263 +       /*
 264 +        * some device names have larger paths - convert the slashes
 265 +        * to underscores for this to work as expected
 266 +        */
 267 +       for (i = 0; i < strlen(buts.name); i++)
 268 +               if (buts.name[i] == '/')
 269 +                       buts.name[i] = '_';
 270 +
 271 +       if (copy_to_user(arg, &buts, sizeof(buts)))
 272 +               return -EFAULT;
 273 +
 274 +       down(&bdev->bd_sem);
 275 +       ret = -EBUSY;
 276 +       if (q->blk_trace)
 277 +               goto err;
 278 +
 279 +       ret = -ENOMEM;
 280 +       bt = kzalloc(sizeof(*bt), GFP_KERNEL);
 281 +       if (!bt)
 282 +               goto err;
 283 +
 284 +       bt->sequence = alloc_percpu(unsigned long);
 285 +       if (!bt->sequence)
 286 +               goto err;
 287 +
 288 +       ret = -ENOENT;
 289 +       dir = blk_create_tree(buts.name);
 290 +       if (!dir)
 291 +               goto err;
 292 +
 293 +       bt->dir = dir;
 294 +       bt->dev = bdev->bd_dev;
 295 +       atomic_set(&bt->dropped, 0);
 296 +
 297 +       ret = -EIO;
 298 +       bt->dropped_file = relayfs_create_file("dropped", dir, 0, &blk_dropped_fops, bt);
 299 +       if (!bt->dropped_file)
 300 +               goto err;
 301 +
 302 +       bt->rchan = relay_open("trace", dir, buts.buf_size, buts.buf_nr, &blk_relay_callbacks);
 303 +       if (!bt->rchan)
 304 +               goto err;
 305 +       bt->rchan->private_data = bt;
 306 +
 307 +       bt->act_mask = buts.act_mask;
 308 +       if (!bt->act_mask)
 309 +               bt->act_mask = (u16) -1;
 310 +
 311 +       bt->start_lba = buts.start_lba;
 312 +       bt->end_lba = buts.end_lba;
 313 +       if (!bt->end_lba)
 314 +               bt->end_lba = -1ULL;
 315 +
 316 +       bt->pid = buts.pid;
 317 +
 318 +       q->blk_trace = bt;
 319 +       up(&bdev->bd_sem);
 320 +       return 0;
 321 +err:
 322 +       up(&bdev->bd_sem);
 323 +       if (bt && bt->dropped_file)
 324 +               relayfs_remove_file(bt->dropped_file);
 325 +       if (dir)
 326 +               blk_remove_tree(dir);
 327 +       if (bt) {
 328 +               if (bt->sequence)
 329 +                       free_percpu(bt->sequence);
 330 +               kfree(bt);
 331 +       }
 332 +       return ret;
 333 +}
 334 +
 335 +/*
 336 + * Average offset over two calls to sched_clock() with a gettimeofday()
 337 + * in the middle
 338 + */
 339 +static void blk_check_time(unsigned long long *t)
 340 +{
 341 +       unsigned long long a, b;
 342 +       struct timeval tv;
 343 +
 344 +       a = sched_clock();
 345 +       do_gettimeofday(&tv);
 346 +       b = sched_clock();
 347 +
 348 +       *t = tv.tv_sec * 1000000000 + tv.tv_usec * 1000;
 349 +       *t -= (a + b) / 2;
 350 +}
 351 +
 352 +static void blk_trace_check_cpu_time(void *data)
 353 +{
 354 +       unsigned long long *t;
 355 +       int cpu = get_cpu();
 356 +
 357 +       t = &per_cpu(blk_trace_cpu_offset, cpu);
 358 +
 359 +       /*
 360 +        * Just call it twice, hopefully the second call will be cache hot
 361 +        * and a little more precise
 362 +        */
 363 +       blk_check_time(t);
 364 +       blk_check_time(t);
 365 +
 366 +       put_cpu();
 367 +}
 368 +
 369 +/*
 370 + * Call blk_trace_check_cpu_time() on each CPU to calibrate our inter-CPU
 371 + * timings
 372 + */
 373 +static void blk_trace_calibrate_offsets(void)
 374 +{
 375 +       unsigned long flags;
 376 +
 377 +       smp_call_function(blk_trace_check_cpu_time, NULL, 1, 1);
 378 +       local_irq_save(flags);
 379 +       blk_trace_check_cpu_time(NULL);
 380 +       local_irq_restore(flags);
 381 +}
 382 +
 383 +static void blk_trace_set_ht_offsets(void)
 384 +{
 385 +#if defined(CONFIG_SCHED_SMT)
 386 +       int cpu, i;
 387 +
 388 +       /*
 389 +        * now make sure HT siblings have the same time offset
 390 +        */
 391 +       preempt_disable();
 392 +       for_each_online_cpu(cpu) {
 393 +               unsigned long long *cpu_off, *sibling_off;
 394 +
 395 +               for_each_cpu_mask(i, cpu_sibling_map[cpu]) {
 396 +                       if (i == cpu)
 397 +                               continue;
 398 +
 399 +                       cpu_off = &per_cpu(blk_trace_cpu_offset, cpu);
 400 +                       sibling_off = &per_cpu(blk_trace_cpu_offset, i);
 401 +                       *sibling_off = *cpu_off;
 402 +               }
 403 +       }
 404 +       preempt_enable();
 405 +#endif
 406 +}
 407 +
 408 +static __init int blk_trace_init(void)
 409 +{
 410 +       mutex_init(&blk_tree_mutex);
 411 +       blk_trace_calibrate_offsets();
 412 +       blk_trace_set_ht_offsets();
 413 +
 414 +       return 0;
 415 +}
 416 +
 417 +module_init(blk_trace_init);
 418 +
 419 diff --git a/block/elevator.c b/block/elevator.c
 420 index 96a61e0..af75304 100644
 421 --- a/block/elevator.c
 422 +++ b/block/elevator.c
 423 @@ -33,6 +33,7 @@
 424  #include <linux/init.h>
 425  #include <linux/compiler.h>
 426  #include <linux/delay.h>
 427 +#include <linux/blktrace_api.h>
 428
 429  #include <asm/uaccess.h>
 430
 431 @@ -316,6 +317,8 @@ void __elv_add_request(request_queue_t *
 432         struct list_head *pos;
 433         unsigned ordseq;
 434
 435 +       blk_add_trace_rq(q, rq, BLK_TA_INSERT);
 436 +
 437         if (q->ordcolor)
 438                 rq->flags |= REQ_ORDERED_COLOR;
 439
 440 @@ -474,6 +477,7 @@ struct request *elv_next_request(request
 441                          * not be passed by new incoming requests
 442                          */
 443                         rq->flags |= REQ_STARTED;
 444 +                       blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
 445                 }
 446
 447                 if (!q->boundary_rq || q->boundary_rq == rq) {
 448 diff --git a/block/ioctl.c b/block/ioctl.c
 449 index e110949..63e67a2 100644
 450 --- a/block/ioctl.c
 451 +++ b/block/ioctl.c
 452 @@ -5,6 +5,7 @@
 453  #include <linux/backing-dev.h>
 454  #include <linux/buffer_head.h>
 455  #include <linux/smp_lock.h>
 456 +#include <linux/blktrace_api.h>
 457  #include <asm/uaccess.h>
 458
 459  static int blkpg_ioctl(struct block_device *bdev, struct blkpg_ioctl_arg __user *arg)
 460 @@ -189,6 +190,10 @@ static int blkdev_locked_ioctl(struct fi
 461                 return put_ulong(arg, bdev->bd_inode->i_size >> 9);
 462         case BLKGETSIZE64:
 463                 return put_u64(arg, bdev->bd_inode->i_size);
 464 +       case BLKSTARTTRACE:
 465 +               return blk_start_trace(bdev, (char __user *) arg);
 466 +       case BLKSTOPTRACE:
 467 +               return blk_stop_trace(bdev);
 468         }
 469         return -ENOIOCTLCMD;
 470  }
 471 diff --git a/block/ll_rw_blk.c b/block/ll_rw_blk.c
 472 index f9fc07e..4295103 100644
 473 --- a/block/ll_rw_blk.c
 474 +++ b/block/ll_rw_blk.c
 475 @@ -28,6 +28,7 @@
 476  #include <linux/writeback.h>
 477  #include <linux/interrupt.h>
 478  #include <linux/cpu.h>
 479 +#include <linux/blktrace_api.h>
 480
 481  /*
 482   * for max sense size
 483 @@ -1557,8 +1558,10 @@ void blk_plug_device(request_queue_t *q)
 484         if (test_bit(QUEUE_FLAG_STOPPED, &q->queue_flags))
 485                 return;
 486
 487 -       if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags))
 488 +       if (!test_and_set_bit(QUEUE_FLAG_PLUGGED, &q->queue_flags)) {
 489                 mod_timer(&q->unplug_timer, jiffies + q->unplug_delay);
 490 +               blk_add_trace_generic(q, NULL, 0, BLK_TA_PLUG);
 491 +       }
 492  }
 493
 494  EXPORT_SYMBOL(blk_plug_device);
 495 @@ -1622,14 +1625,21 @@ static void blk_backing_dev_unplug(struc
 496         /*
 497          * devices don't necessarily have an ->unplug_fn defined
 498          */
 499 -       if (q->unplug_fn)
 500 +       if (q->unplug_fn) {
 501 +               blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
 502 +                                       q->rq.count[READ] + q->rq.count[WRITE]);
 503 +
 504                 q->unplug_fn(q);
 505 +       }
 506  }
 507
 508  static void blk_unplug_work(void *data)
 509  {
 510         request_queue_t *q = data;
 511
 512 +       blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
 513 +                               q->rq.count[READ] + q->rq.count[WRITE]);
 514 +
 515         q->unplug_fn(q);
 516  }
 517
 518 @@ -1637,6 +1647,9 @@ static void blk_unplug_timeout(unsigned
 519  {
 520         request_queue_t *q = (request_queue_t *)data;
 521
 522 +       blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_TIMER, NULL,
 523 +                               q->rq.count[READ] + q->rq.count[WRITE]);
 524 +
 525         kblockd_schedule_work(&q->unplug_work);
 526  }
 527
 528 @@ -1759,6 +1772,11 @@ void blk_cleanup_queue(request_queue_t *
 529         if (q->queue_tags)
 530                 __blk_queue_free_tags(q);
 531
 532 +       if (q->blk_trace) {
 533 +               blk_cleanup_trace(q->blk_trace);
 534 +               q->blk_trace = NULL;
 535 +       }
 536 +
 537         kmem_cache_free(requestq_cachep, q);
 538  }
 539
 540 @@ -2110,6 +2128,8 @@ rq_starved:
 541
 542         rq_init(q, rq);
 543         rq->rl = rl;
 544 +
 545 +       blk_add_trace_generic(q, bio, rw, BLK_TA_GETRQ);
 546  out:
 547         return rq;
 548  }
 549 @@ -2138,6 +2158,8 @@ static struct request *get_request_wait(
 550                 if (!rq) {
 551                         struct io_context *ioc;
 552
 553 +                       blk_add_trace_generic(q, bio, rw, BLK_TA_SLEEPRQ);
 554 +
 555                         __generic_unplug_device(q);
 556                         spin_unlock_irq(q->queue_lock);
 557                         io_schedule();
 558 @@ -2191,6 +2213,8 @@ EXPORT_SYMBOL(blk_get_request);
 559   */
 560  void blk_requeue_request(request_queue_t *q, struct request *rq)
 561  {
 562 +       blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
 563 +
 564         if (blk_rq_tagged(rq))
 565                 blk_queue_end_tag(q, rq);
 566
 567 @@ -2825,6 +2849,8 @@ static int __make_request(request_queue_
 568                         if (!q->back_merge_fn(q, req, bio))
 569                                 break;
 570
 571 +                       blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
 572 +
 573                         req->biotail->bi_next = bio;
 574                         req->biotail = bio;
 575                         req->nr_sectors = req->hard_nr_sectors += nr_sectors;
 576 @@ -2840,6 +2866,8 @@ static int __make_request(request_queue_
 577                         if (!q->front_merge_fn(q, req, bio))
 578                                 break;
 579
 580 +                       blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
 581 +
 582                         bio->bi_next = req->bio;
 583                         req->bio = bio;
 584
 585 @@ -2957,6 +2985,7 @@ void generic_make_request(struct bio *bi
 586         request_queue_t *q;
 587         sector_t maxsector;
 588         int ret, nr_sectors = bio_sectors(bio);
 589 +       dev_t old_dev;
 590
 591         might_sleep();
 592         /* Test device or partition size, when known. */
 593 @@ -2983,6 +3012,8 @@ void generic_make_request(struct bio *bi
 594          * NOTE: we don't repeat the blk_size check for each new device.
 595          * Stacking drivers are expected to know what they are doing.
 596          */
 597 +       maxsector = -1;
 598 +       old_dev = 0;
 599         do {
 600                 char b[BDEVNAME_SIZE];
 601
 602 @@ -3015,6 +3046,15 @@ end_io:
 603                  */
 604                 blk_partition_remap(bio);
 605
 606 +               if (maxsector != -1)
 607 +                       blk_add_trace_remap(q, bio, old_dev, bio->bi_sector,
 608 +                                           maxsector);
 609 +
 610 +               blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
 611 +
 612 +               maxsector = bio->bi_sector;
 613 +               old_dev = bio->bi_bdev->bd_dev;
 614 +
 615                 ret = q->make_request_fn(q, bio);
 616         } while (ret);
 617  }
 618 @@ -3134,6 +3174,8 @@ static int __end_that_request_first(stru
 619         int total_bytes, bio_nbytes, error, next_idx = 0;
 620         struct bio *bio;
 621
 622 +       blk_add_trace_rq(req->q, req, BLK_TA_COMPLETE);
 623 +
 624         /*
 625          * extend uptodate bool to allow < 0 value to be direct io error
 626          */
 627 diff --git a/drivers/block/cciss.c b/drivers/block/cciss.c
 628 index 12d7b9b..880892a 100644
 629 --- a/drivers/block/cciss.c
 630 +++ b/drivers/block/cciss.c
 631 @@ -38,6 +38,7 @@
 632  #include <linux/hdreg.h>
 633  #include <linux/spinlock.h>
 634  #include <linux/compat.h>
 635 +#include <linux/blktrace_api.h>
 636  #include <asm/uaccess.h>
 637  #include <asm/io.h>
 638
 639 @@ -2330,6 +2331,7 @@ static inline void complete_command( ctl
 640
 641         cmd->rq->completion_data = cmd;
 642         cmd->rq->errors = status;
 643 +       blk_add_trace_rq(cmd->rq->q, cmd->rq, BLK_TA_COMPLETE);
 644         blk_complete_request(cmd->rq);
 645  }
 646
 647 diff --git a/drivers/md/dm.c b/drivers/md/dm.c
 648 index e9adeb9..c8f3aa2 100644
 649 --- a/drivers/md/dm.c
 650 +++ b/drivers/md/dm.c
 651 @@ -17,6 +17,7 @@
 652  #include <linux/mempool.h>
 653  #include <linux/slab.h>
 654  #include <linux/idr.h>
 655 +#include <linux/blktrace_api.h>
 656
 657  static const char *_name = DM_NAME;
 658
 659 @@ -334,6 +335,8 @@ static void dec_pending(struct dm_io *io
 660                         /* nudge anyone waiting on suspend queue */
 661                         wake_up(&io->md->wait);
 662
 663 +               blk_add_trace_bio(io->md->queue, io->bio, BLK_TA_COMPLETE);
 664 +
 665                 bio_endio(io->bio, io->bio->bi_size, io->error);
 666                 free_io(io->md, io);
 667         }
 668 @@ -392,6 +395,7 @@ static void __map_bio(struct dm_target *
 669                       struct target_io *tio)
 670  {
 671         int r;
 672 +       sector_t sector;
 673
 674         /*
 675          * Sanity checks.
 676 @@ -407,10 +411,17 @@ static void __map_bio(struct dm_target *
 677          * this io.
 678          */
 679         atomic_inc(&tio->io->io_count);
 680 +       sector = clone->bi_sector;
 681         r = ti->type->map(ti, clone, &tio->info);
 682 -       if (r > 0)
 683 +       if (r > 0) {
 684                 /* the bio has been remapped so dispatch it */
 685 +
 686 +               blk_add_trace_remap(bdev_get_queue(clone->bi_bdev), clone,
 687 +                                   tio->io->bio->bi_bdev->bd_dev, sector,
 688 +                                   clone->bi_sector);
 689 +
 690                 generic_make_request(clone);
 691 +       }
 692
 693         else if (r < 0) {
 694                 /* error the io and bail out */
 695 diff --git a/fs/bio.c b/fs/bio.c
 696 index 1f3bb50..0dd0d81 100644
 697 --- a/fs/bio.c
 698 +++ b/fs/bio.c
 699 @@ -25,6 +25,7 @@
 700  #include <linux/module.h>
 701  #include <linux/mempool.h>
 702  #include <linux/workqueue.h>
 703 +#include <linux/blktrace_api.h>
 704  #include <scsi/sg.h>           /* for struct sg_iovec */
 705
 706  #define BIO_POOL_SIZE 256
 707 @@ -1095,6 +1096,9 @@ struct bio_pair *bio_split(struct bio *b
 708         if (!bp)
 709                 return bp;
 710
 711 +       blk_add_trace_pdu_int(bdev_get_queue(bi->bi_bdev), BLK_TA_SPLIT, bi,
 712 +                               bi->bi_sector + first_sectors);
 713 +
 714         BUG_ON(bi->bi_vcnt != 1);
 715         BUG_ON(bi->bi_idx != 0);
 716         atomic_set(&bp->cnt, 3);
 717 diff --git a/fs/compat_ioctl.c b/fs/compat_ioctl.c
 718 index 5dd0207..010e02b 100644
 719 --- a/fs/compat_ioctl.c
 720 +++ b/fs/compat_ioctl.c
 721 @@ -72,6 +72,7 @@
 722  #include <linux/i2c-dev.h>
 723  #include <linux/wireless.h>
 724  #include <linux/atalk.h>
 725 +#include <linux/blktrace_api.h>
 726
 727  #include <net/sock.h>          /* siocdevprivate_ioctl */
 728  #include <net/bluetooth/bluetooth.h>
 729 diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
 730 index 860e7a4..266ce9d 100644
 731 --- a/include/linux/blkdev.h
 732 +++ b/include/linux/blkdev.h
 733 @@ -22,6 +22,7 @@ typedef struct request_queue request_que
 734  struct elevator_queue;
 735  typedef struct elevator_queue elevator_t;
 736  struct request_pm_state;
 737 +struct blk_trace;
 738
 739  #define BLKDEV_MIN_RQ  4
 740  #define BLKDEV_MAX_RQ  128     /* Default maximum */
 741 @@ -416,6 +417,8 @@ struct request_queue
 742         unsigned int            sg_reserved_size;
 743         int                     node;
 744
 745 +       struct blk_trace        *blk_trace;
 746 +
 747         /*
 748          * reserved for flush operations
 749          */
 750 diff --git a/include/linux/blktrace_api.h b/include/linux/blktrace_api.h
 751 new file mode 100644
 752 index 0000000..b0fc5cf
 753 --- /dev/null
 754 +++ b/include/linux/blktrace_api.h
 755 @@ -0,0 +1,272 @@
 756 +#ifndef BLKTRACE_H
 757 +#define BLKTRACE_H
 758 +
 759 +#include <linux/config.h>
 760 +#include <linux/blkdev.h>
 761 +#include <linux/relayfs_fs.h>
 762 +
 763 +/*
 764 + * Trace categories
 765 + */
 766 +enum blktrace_cat {
 767 +       BLK_TC_READ     = 1 << 0,       /* reads */
 768 +       BLK_TC_WRITE    = 1 << 1,       /* writes */
 769 +       BLK_TC_BARRIER  = 1 << 2,       /* barrier */
 770 +       BLK_TC_SYNC     = 1 << 3,       /* barrier */
 771 +       BLK_TC_QUEUE    = 1 << 4,       /* queueing/merging */
 772 +       BLK_TC_REQUEUE  = 1 << 5,       /* requeueing */
 773 +       BLK_TC_ISSUE    = 1 << 6,       /* issue */
 774 +       BLK_TC_COMPLETE = 1 << 7,       /* completions */
 775 +       BLK_TC_FS       = 1 << 8,       /* fs requests */
 776 +       BLK_TC_PC       = 1 << 9,       /* pc requests */
 777 +
 778 +       BLK_TC_END      = 1 << 15,      /* only 16-bits, reminder */
 779 +};
 780 +
 781 +#define BLK_TC_SHIFT           (16)
 782 +#define BLK_TC_ACT(act)                ((act) << BLK_TC_SHIFT)
 783 +
 784 +/*
 785 + * Basic trace actions
 786 + */
 787 +enum blktrace_act {
 788 +       __BLK_TA_QUEUE = 1,             /* queued */
 789 +       __BLK_TA_BACKMERGE,             /* back merged to existing rq */
 790 +       __BLK_TA_FRONTMERGE,            /* front merge to existing rq */
 791 +       __BLK_TA_GETRQ,                 /* allocated new request */
 792 +       __BLK_TA_SLEEPRQ,               /* sleeping on rq allocation */
 793 +       __BLK_TA_REQUEUE,               /* request requeued */
 794 +       __BLK_TA_ISSUE,                 /* sent to driver */
 795 +       __BLK_TA_COMPLETE,              /* completed by driver */
 796 +       __BLK_TA_PLUG,                  /* queue was plugged */
 797 +       __BLK_TA_UNPLUG_IO,             /* queue was unplugged by io */
 798 +       __BLK_TA_UNPLUG_TIMER,          /* queue was unplugged by timer */
 799 +       __BLK_TA_INSERT,                /* insert request */
 800 +       __BLK_TA_SPLIT,                 /* bio was split */
 801 +       __BLK_TA_BOUNCE,                /* bio was bounced */
 802 +       __BLK_TA_REMAP,                 /* bio was remapped */
 803 +};
 804 +
 805 +/*
 806 + * Trace actions in full. Additionally, read or write is masked
 807 + */
 808 +#define BLK_TA_QUEUE           (__BLK_TA_QUEUE | BLK_TC_ACT(BLK_TC_QUEUE))
 809 +#define BLK_TA_BACKMERGE       (__BLK_TA_BACKMERGE | BLK_TC_ACT(BLK_TC_QUEUE))
 810 +#define BLK_TA_FRONTMERGE      (__BLK_TA_FRONTMERGE | BLK_TC_ACT(BLK_TC_QUEUE))
 811 +#define        BLK_TA_GETRQ            (__BLK_TA_GETRQ | BLK_TC_ACT(BLK_TC_QUEUE))
 812 +#define        BLK_TA_SLEEPRQ          (__BLK_TA_SLEEPRQ | BLK_TC_ACT(BLK_TC_QUEUE))
 813 +#define        BLK_TA_REQUEUE          (__BLK_TA_REQUEUE | BLK_TC_ACT(BLK_TC_REQUEUE))
 814 +#define BLK_TA_ISSUE           (__BLK_TA_ISSUE | BLK_TC_ACT(BLK_TC_ISSUE))
 815 +#define BLK_TA_COMPLETE                (__BLK_TA_COMPLETE| BLK_TC_ACT(BLK_TC_COMPLETE))
 816 +#define BLK_TA_PLUG            (__BLK_TA_PLUG | BLK_TC_ACT(BLK_TC_QUEUE))
 817 +#define BLK_TA_UNPLUG_IO       (__BLK_TA_UNPLUG_IO | BLK_TC_ACT(BLK_TC_QUEUE))
 818 +#define BLK_TA_UNPLUG_TIMER    (__BLK_TA_UNPLUG_TIMER | BLK_TC_ACT(BLK_TC_QUEUE))
 819 +#define BLK_TA_INSERT          (__BLK_TA_INSERT | BLK_TC_ACT(BLK_TC_QUEUE))
 820 +#define BLK_TA_SPLIT           (__BLK_TA_SPLIT)
 821 +#define BLK_TA_BOUNCE          (__BLK_TA_BOUNCE)
 822 +#define BLK_TA_REMAP           (__BLK_TA_REMAP | BLK_TC_ACT(BLK_TC_QUEUE))
 823 +
 824 +#define BLK_IO_TRACE_MAGIC     0x65617400
 825 +#define BLK_IO_TRACE_VERSION   0x06
 826 +
 827 +/*
 828 + * The trace itself
 829 + */
 830 +struct blk_io_trace {
 831 +       u32 magic;              /* MAGIC << 8 | version */
 832 +       u32 sequence;           /* event number */
 833 +       u64 time;               /* in microseconds */
 834 +       u64 sector;             /* disk offset */
 835 +       u32 bytes;              /* transfer length */
 836 +       u32 action;             /* what happened */
 837 +       u32 pid;                /* who did it */
 838 +       u32 cpu;                /* on what cpu did it happen */
 839 +       u16 error;              /* completion error */
 840 +       u16 pdu_len;            /* length of data after this trace */
 841 +       u32 device;             /* device number */
 842 +       char comm[16];          /* task command name (TASK_COMM_LEN) */
 843 +};
 844 +
 845 +/*
 846 + * The remap event
 847 + */
 848 +struct blk_io_trace_remap {
 849 +       u32 device;
 850 +       u32 __pad;
 851 +       u64 sector;
 852 +};
 853 +
 854 +struct blk_trace {
 855 +       struct dentry *dir;
 856 +       struct rchan *rchan;
 857 +       struct dentry *dropped_file;
 858 +       atomic_t dropped;
 859 +       unsigned long *sequence;
 860 +       u32 dev;
 861 +       u16 act_mask;
 862 +       u64 start_lba;
 863 +       u64 end_lba;
 864 +       u32 pid;
 865 +};
 866 +
 867 +/*
 868 + * User setup structure passed with BLKSTARTTRACE
 869 + */
 870 +struct blk_user_trace_setup {
 871 +       char name[BDEVNAME_SIZE];       /* output */
 872 +       u16 act_mask;                   /* input */
 873 +       u32 buf_size;                   /* input */
 874 +       u32 buf_nr;                     /* input */
 875 +       u64 start_lba;
 876 +       u64 end_lba;
 877 +       u32 pid;
 878 +};
 879 +
 880 +#if defined(CONFIG_BLK_DEV_IO_TRACE)
 881 +extern int blk_start_trace(struct block_device *, char __user *);
 882 +extern int blk_stop_trace(struct block_device *);
 883 +extern void blk_cleanup_trace(struct blk_trace *);
 884 +extern void __blk_add_trace(struct blk_trace *, sector_t, int, int, u32, int, int, void *);
 885 +
 886 +/**
 887 + * blk_add_trace_rq - Add a trace for a request oriented action
 888 + * @q:         queue the io is for
 889 + * @rq:                the source request
 890 + * @what:      the action
 891 + *
 892 + * Description:
 893 + *     Records an action against a request. Will log the bio offset + size.
 894 + *
 895 + **/
 896 +static inline void blk_add_trace_rq(struct request_queue *q, struct request *rq,
 897 +                                   u32 what)
 898 +{
 899 +       struct blk_trace *bt = q->blk_trace;
 900 +       int rw = rq->flags & 0x07;
 901 +
 902 +       if (likely(!bt))
 903 +               return;
 904 +
 905 +       if (blk_pc_request(rq)) {
 906 +               what |= BLK_TC_ACT(BLK_TC_PC);
 907 +               __blk_add_trace(bt, 0, rq->data_len, rw, what, rq->errors, sizeof(rq->cmd), rq->cmd);
 908 +       } else  {
 909 +               what |= BLK_TC_ACT(BLK_TC_FS);
 910 +               __blk_add_trace(bt, rq->hard_sector, rq->hard_nr_sectors << 9, rw, what, rq->errors, 0, NULL);
 911 +       }
 912 +}
 913 +
 914 +/**
 915 + * blk_add_trace_bio - Add a trace for a bio oriented action
 916 + * @q:         queue the io is for
 917 + * @bio:       the source bio
 918 + * @what:      the action
 919 + *
 920 + * Description:
 921 + *     Records an action against a bio. Will log the bio offset + size.
 922 + *
 923 + **/
 924 +static inline void blk_add_trace_bio(struct request_queue *q, struct bio *bio,
 925 +                                    u32 what)
 926 +{
 927 +       struct blk_trace *bt = q->blk_trace;
 928 +
 929 +       if (likely(!bt))
 930 +               return;
 931 +
 932 +       __blk_add_trace(bt, bio->bi_sector, bio->bi_size, bio->bi_rw, what, !bio_flagged(bio, BIO_UPTODATE), 0, NULL);
 933 +}
 934 +
 935 +/**
 936 + * blk_add_trace_generic - Add a trace for a generic action
 937 + * @q:         queue the io is for
 938 + * @bio:       the source bio
 939 + * @rw:                the data direction
 940 + * @what:      the action
 941 + *
 942 + * Description:
 943 + *     Records a simple trace
 944 + *
 945 + **/
 946 +static inline void blk_add_trace_generic(struct request_queue *q,
 947 +                                        struct bio *bio, int rw, u32 what)
 948 +{
 949 +       struct blk_trace *bt = q->blk_trace;
 950 +
 951 +       if (likely(!bt))
 952 +               return;
 953 +
 954 +       if (bio)
 955 +               blk_add_trace_bio(q, bio, what);
 956 +       else
 957 +               __blk_add_trace(bt, 0, 0, rw, what, 0, 0, NULL);
 958 +}
 959 +
 960 +/**
 961 + * blk_add_trace_pdu_int - Add a trace for a bio with an integer payload
 962 + * @q:         queue the io is for
 963 + * @what:      the action
 964 + * @bio:       the source bio
 965 + * @pdu:       the integer payload
 966 + *
 967 + * Description:
 968 + *     Adds a trace with some integer payload. This might be an unplug
 969 + *     option given as the action, with the depth at unplug time given
 970 + *     as the payload
 971 + *
 972 + **/
 973 +static inline void blk_add_trace_pdu_int(struct request_queue *q, u32 what,
 974 +                                        struct bio *bio, unsigned int pdu)
 975 +{
 976 +       struct blk_trace *bt = q->blk_trace;
 977 +       u64 rpdu = cpu_to_be64(pdu);
 978 +
 979 +       if (likely(!bt))
 980 +               return;
 981 +
 982 +       if (bio)
 983 +               __blk_add_trace(bt, bio->bi_sector, bio->bi_size, bio->bi_rw, what, !bio_flagged(bio, BIO_UPTODATE), sizeof(rpdu), &rpdu);
 984 +       else
 985 +               __blk_add_trace(bt, 0, 0, 0, what, 0, sizeof(rpdu), &rpdu);
 986 +}
 987 +
 988 +/**
 989 + * blk_add_trace_remap - Add a trace for a remap operation
 990 + * @q:         queue the io is for
 991 + * @bio:       the source bio
 992 + * @dev:       target device
 993 + * @from:      source sector
 994 + * @to:                target sector
 995 + *
 996 + * Description:
 997 + *     Device mapper or raid target sometimes need to split a bio because
 998 + *     it spans a stripe (or similar). Add a trace for that action.
 999 + *
1000 + **/
1001 +static inline void blk_add_trace_remap(struct request_queue *q, struct bio *bio,
1002 +                                      dev_t dev, sector_t from, sector_t to)
1003 +{
1004 +       struct blk_trace *bt = q->blk_trace;
1005 +       struct blk_io_trace_remap r;
1006 +
1007 +       if (likely(!bt))
1008 +               return;
1009 +
1010 +       r.device = cpu_to_be32(dev);
1011 +       r.sector = cpu_to_be64(to);
1012 +
1013 +       __blk_add_trace(bt, from, bio->bi_size, bio->bi_rw, BLK_TA_REMAP, !bio_flagged(bio, BIO_UPTODATE), sizeof(r), &r);
1014 +}
1015 +
1016 +#else /* !CONFIG_BLK_DEV_IO_TRACE */
1017 +#define blk_start_trace(bdev, arg)             (-EINVAL)
1018 +#define blk_stop_trace(bdev)                   (-EINVAL)
1019 +#define blk_cleanup_trace(bt)                  do { } while (0)
1020 +#define blk_add_trace_rq(q, rq, what)          do { } while (0)
1021 +#define blk_add_trace_bio(q, rq, what)         do { } while (0)
1022 +#define blk_add_trace_generic(q, rq, rw, what) do { } while (0)
1023 +#define blk_add_trace_pdu_int(q, what, bio, pdu)       do { } while (0)
1024 +#define blk_add_trace_remap(q, bio, dev, f, t) do {} while (0)
1025 +#endif /* CONFIG_BLK_DEV_IO_TRACE */
1026 +
1027 +#endif
1028 diff --git a/include/linux/compat_ioctl.h b/include/linux/compat_ioctl.h
1029 index 8fad50f..5bed09a 100644
1030 --- a/include/linux/compat_ioctl.h
1031 +++ b/include/linux/compat_ioctl.h
1032 @@ -97,6 +97,8 @@ COMPATIBLE_IOCTL(BLKRRPART)
1033  COMPATIBLE_IOCTL(BLKFLSBUF)
1034  COMPATIBLE_IOCTL(BLKSECTSET)
1035  COMPATIBLE_IOCTL(BLKSSZGET)
1036 +COMPATIBLE_IOCTL(BLKSTARTTRACE)
1037 +COMPATIBLE_IOCTL(BLKSTOPTRACE)
1038  ULONG_IOCTL(BLKRASET)
1039  ULONG_IOCTL(BLKFRASET)
1040  /* RAID */
1041 diff --git a/include/linux/fs.h b/include/linux/fs.h
1042 index e059da9..35c6b45 100644
1043 --- a/include/linux/fs.h
1044 +++ b/include/linux/fs.h
1045 @@ -196,6 +196,8 @@ extern int dir_notify_enable;
1046  #define BLKBSZGET  _IOR(0x12,112,size_t)
1047  #define BLKBSZSET  _IOW(0x12,113,size_t)
1048  #define BLKGETSIZE64 _IOR(0x12,114,size_t)     /* return device size in bytes (u64 *arg) */
1049 +#define BLKSTARTTRACE _IOWR(0x12,115,struct blk_user_trace_setup)
1050 +#define BLKSTOPTRACE _IO(0x12,116)
1051
1052  #define BMAP_IOCTL 1           /* obsolete - kept for compatibility */
1053  #define FIBMAP    _IO(0x00,1)  /* bmap access */
1054 diff --git a/mm/highmem.c b/mm/highmem.c
1055 index ce2e7e8..d0ea1ee 100644
1056 --- a/mm/highmem.c
1057 +++ b/mm/highmem.c
1058 @@ -26,6 +26,7 @@
1059  #include <linux/init.h>
1060  #include <linux/hash.h>
1061  #include <linux/highmem.h>
1062 +#include <linux/blktrace_api.h>
1063  #include <asm/tlbflush.h>
1064
1065  static mempool_t *page_pool, *isa_page_pool;
1066 @@ -483,6 +484,8 @@ void blk_queue_bounce(request_queue_t *q
1067                 pool = isa_page_pool;
1068         }
1069
1070 +       blk_add_trace_bio(q, *bio_orig, BLK_TA_BOUNCE);
1071 +
1072         /*
1073          * slow path
1074          */