block/blktrace.c

   1 /*
   2  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License version 2 as
   6  * published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  11  * GNU General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public License
  14  * along with this program; if not, write to the Free Software
  15  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  16  *
  17  */
  18 #include <linux/kernel.h>
  19 #include <linux/blkdev.h>
  20 #include <linux/blktrace_api.h>
  21 #include <linux/percpu.h>
  22 #include <linux/init.h>
  23 #include <linux/mutex.h>
  24 #include <linux/debugfs.h>
  25 #include <linux/time.h>
  26 #include <asm/uaccess.h>
  27
  28 static DEFINE_PER_CPU(unsigned long long, blk_trace_cpu_offset) = { 0, };
  29 static unsigned int blktrace_seq __read_mostly = 1;
  30
  31 /*
  32  * Send out a notify message.
  33  */
  34 static void trace_note(struct blk_trace *bt, pid_t pid, int action,
  35                        const void *data, size_t len)
  36 {
  37         struct blk_io_trace *t;
  38
  39         t = relay_reserve(bt->rchan, sizeof(*t) + len);
  40         if (t) {
  41                 const int cpu = smp_processor_id();
  42
  43                 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
  44                 t->time = cpu_clock(cpu) - per_cpu(blk_trace_cpu_offset, cpu);
  45                 t->device = bt->dev;
  46                 t->action = action;
  47                 t->pid = pid;
  48                 t->cpu = cpu;
  49                 t->pdu_len = len;
  50                 memcpy((void *) t + sizeof(*t), data, len);
  51         }
  52 }
  53
  54 /*
  55  * Send out a notify for this process, if we haven't done so since a trace
  56  * started
  57  */
  58 static void trace_note_tsk(struct blk_trace *bt, struct task_struct *tsk)
  59 {
  60         tsk->btrace_seq = blktrace_seq;
  61         trace_note(bt, tsk->pid, BLK_TN_PROCESS, tsk->comm, sizeof(tsk->comm));
  62 }
  63
  64 static void trace_note_time(struct blk_trace *bt)
  65 {
  66         struct timespec now;
  67         unsigned long flags;
  68         u32 words[2];
  69
  70         getnstimeofday(&now);
  71         words[0] = now.tv_sec;
  72         words[1] = now.tv_nsec;
  73
  74         local_irq_save(flags);
  75         trace_note(bt, 0, BLK_TN_TIMESTAMP, words, sizeof(words));
  76         local_irq_restore(flags);
  77 }
  78
  79 static int act_log_check(struct blk_trace *bt, u32 what, sector_t sector,
  80                          pid_t pid)
  81 {
  82         if (((bt->act_mask << BLK_TC_SHIFT) & what) == 0)
  83                 return 1;
  84         if (sector < bt->start_lba || sector > bt->end_lba)
  85                 return 1;
  86         if (bt->pid && pid != bt->pid)
  87                 return 1;
  88
  89         return 0;
  90 }
  91
  92 /*
  93  * Data direction bit lookup
  94  */
  95 static u32 ddir_act[2] __read_mostly = { BLK_TC_ACT(BLK_TC_READ), BLK_TC_ACT(BLK_TC_WRITE) };
  96
  97 /*
  98  * Bio action bits of interest
  99  */
 100 static u32 bio_act[9] __read_mostly = { 0, BLK_TC_ACT(BLK_TC_BARRIER), BLK_TC_ACT(BLK_TC_SYNC), 0, BLK_TC_ACT(BLK_TC_AHEAD), 0, 0, 0, BLK_TC_ACT(BLK_TC_META) };
 101
 102 /*
 103  * More could be added as needed, taking care to increment the decrementer
 104  * to get correct indexing
 105  */
 106 #define trace_barrier_bit(rw)   \
 107         (((rw) & (1 << BIO_RW_BARRIER)) >> (BIO_RW_BARRIER - 0))
 108 #define trace_sync_bit(rw)      \
 109         (((rw) & (1 << BIO_RW_SYNC)) >> (BIO_RW_SYNC - 1))
 110 #define trace_ahead_bit(rw)     \
 111         (((rw) & (1 << BIO_RW_AHEAD)) << (2 - BIO_RW_AHEAD))
 112 #define trace_meta_bit(rw)      \
 113         (((rw) & (1 << BIO_RW_META)) >> (BIO_RW_META - 3))
 114
 115 /*
 116  * The worker for the various blk_add_trace*() types. Fills out a
 117  * blk_io_trace structure and places it in a per-cpu subbuffer.
 118  */
 119 void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
 120                      int rw, u32 what, int error, int pdu_len, void *pdu_data)
 121 {
 122         struct task_struct *tsk = current;
 123         struct blk_io_trace *t;
 124         unsigned long flags;
 125         unsigned long *sequence;
 126         pid_t pid;
 127         int cpu;
 128
 129         if (unlikely(bt->trace_state != Blktrace_running))
 130                 return;
 131
 132         what |= ddir_act[rw & WRITE];
 133         what |= bio_act[trace_barrier_bit(rw)];
 134         what |= bio_act[trace_sync_bit(rw)];
 135         what |= bio_act[trace_ahead_bit(rw)];
 136         what |= bio_act[trace_meta_bit(rw)];
 137
 138         pid = tsk->pid;
 139         if (unlikely(act_log_check(bt, what, sector, pid)))
 140                 return;
 141
 142         /*
 143          * A word about the locking here - we disable interrupts to reserve
 144          * some space in the relay per-cpu buffer, to prevent an irq
 145          * from coming in and stepping on our toes. Once reserved, it's
 146          * enough to get preemption disabled to prevent read of this data
 147          * before we are through filling it. get_cpu()/put_cpu() does this
 148          * for us
 149          */
 150         local_irq_save(flags);
 151
 152         if (unlikely(tsk->btrace_seq != blktrace_seq))
 153                 trace_note_tsk(bt, tsk);
 154
 155         t = relay_reserve(bt->rchan, sizeof(*t) + pdu_len);
 156         if (t) {
 157                 cpu = smp_processor_id();
 158                 sequence = per_cpu_ptr(bt->sequence, cpu);
 159
 160                 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
 161                 t->sequence = ++(*sequence);
 162                 t->time = cpu_clock(cpu) - per_cpu(blk_trace_cpu_offset, cpu);
 163                 t->sector = sector;
 164                 t->bytes = bytes;
 165                 t->action = what;
 166                 t->pid = pid;
 167                 t->device = bt->dev;
 168                 t->cpu = cpu;
 169                 t->error = error;
 170                 t->pdu_len = pdu_len;
 171
 172                 if (pdu_len)
 173                         memcpy((void *) t + sizeof(*t), pdu_data, pdu_len);
 174         }
 175
 176         local_irq_restore(flags);
 177 }
 178
 179 EXPORT_SYMBOL_GPL(__blk_add_trace);
 180
 181 static struct dentry *blk_tree_root;
 182 static struct mutex blk_tree_mutex;
 183 static unsigned int root_users;
 184
 185 static inline void blk_remove_root(void)
 186 {
 187         if (blk_tree_root) {
 188                 debugfs_remove(blk_tree_root);
 189                 blk_tree_root = NULL;
 190         }
 191 }
 192
 193 static void blk_remove_tree(struct dentry *dir)
 194 {
 195         mutex_lock(&blk_tree_mutex);
 196         debugfs_remove(dir);
 197         if (--root_users == 0)
 198                 blk_remove_root();
 199         mutex_unlock(&blk_tree_mutex);
 200 }
 201
 202 static struct dentry *blk_create_tree(const char *blk_name)
 203 {
 204         struct dentry *dir = NULL;
 205         int created = 0;
 206
 207         mutex_lock(&blk_tree_mutex);
 208
 209         if (!blk_tree_root) {
 210                 blk_tree_root = debugfs_create_dir("block", NULL);
 211                 if (!blk_tree_root)
 212                         goto err;
 213                 created = 1;
 214         }
 215
 216         dir = debugfs_create_dir(blk_name, blk_tree_root);
 217         if (dir)
 218                 root_users++;
 219         else {
 220                 /* Delete root only if we created it */
 221                 if (created)
 222                         blk_remove_root();
 223         }
 224
 225 err:
 226         mutex_unlock(&blk_tree_mutex);
 227         return dir;
 228 }
 229
 230 static void blk_trace_cleanup(struct blk_trace *bt)
 231 {
 232         relay_close(bt->rchan);
 233         debugfs_remove(bt->dropped_file);
 234         blk_remove_tree(bt->dir);
 235         free_percpu(bt->sequence);
 236         kfree(bt);
 237 }
 238
 239 static int blk_trace_remove(struct request_queue *q)
 240 {
 241         struct blk_trace *bt;
 242
 243         bt = xchg(&q->blk_trace, NULL);
 244         if (!bt)
 245                 return -EINVAL;
 246
 247         if (bt->trace_state == Blktrace_setup ||
 248             bt->trace_state == Blktrace_stopped)
 249                 blk_trace_cleanup(bt);
 250
 251         return 0;
 252 }
 253
 254 static int blk_dropped_open(struct inode *inode, struct file *filp)
 255 {
 256         filp->private_data = inode->i_private;
 257
 258         return 0;
 259 }
 260
 261 static ssize_t blk_dropped_read(struct file *filp, char __user *buffer,
 262                                 size_t count, loff_t *ppos)
 263 {
 264         struct blk_trace *bt = filp->private_data;
 265         char buf[16];
 266
 267         snprintf(buf, sizeof(buf), "%u\n", atomic_read(&bt->dropped));
 268
 269         return simple_read_from_buffer(buffer, count, ppos, buf, strlen(buf));
 270 }
 271
 272 static const struct file_operations blk_dropped_fops = {
 273         .owner =        THIS_MODULE,
 274         .open =         blk_dropped_open,
 275         .read =         blk_dropped_read,
 276 };
 277
 278 /*
 279  * Keep track of how many times we encountered a full subbuffer, to aid
 280  * the user space app in telling how many lost events there were.
 281  */
 282 static int blk_subbuf_start_callback(struct rchan_buf *buf, void *subbuf,
 283                                      void *prev_subbuf, size_t prev_padding)
 284 {
 285         struct blk_trace *bt;
 286
 287         if (!relay_buf_full(buf))
 288                 return 1;
 289
 290         bt = buf->chan->private_data;
 291         atomic_inc(&bt->dropped);
 292         return 0;
 293 }
 294
 295 static int blk_remove_buf_file_callback(struct dentry *dentry)
 296 {
 297         debugfs_remove(dentry);
 298         return 0;
 299 }
 300
 301 static struct dentry *blk_create_buf_file_callback(const char *filename,
 302                                                    struct dentry *parent,
 303                                                    int mode,
 304                                                    struct rchan_buf *buf,
 305                                                    int *is_global)
 306 {
 307         return debugfs_create_file(filename, mode, parent, buf,
 308                                         &relay_file_operations);
 309 }
 310
 311 static struct rchan_callbacks blk_relay_callbacks = {
 312         .subbuf_start           = blk_subbuf_start_callback,
 313         .create_buf_file        = blk_create_buf_file_callback,
 314         .remove_buf_file        = blk_remove_buf_file_callback,
 315 };
 316
 317 /*
 318  * Setup everything required to start tracing
 319  */
 320 int do_blk_trace_setup(struct request_queue *q, struct block_device *bdev,
 321                         struct blk_user_trace_setup *buts)
 322 {
 323         struct blk_trace *old_bt, *bt = NULL;
 324         struct dentry *dir = NULL;
 325         char b[BDEVNAME_SIZE];
 326         int ret, i;
 327
 328         if (!buts->buf_size || !buts->buf_nr)
 329                 return -EINVAL;
 330
 331         strcpy(buts->name, bdevname(bdev, b));
 332
 333         /*
 334          * some device names have larger paths - convert the slashes
 335          * to underscores for this to work as expected
 336          */
 337         for (i = 0; i < strlen(buts->name); i++)
 338                 if (buts->name[i] == '/')
 339                         buts->name[i] = '_';
 340
 341         ret = -ENOMEM;
 342         bt = kzalloc(sizeof(*bt), GFP_KERNEL);
 343         if (!bt)
 344                 goto err;
 345
 346         bt->sequence = alloc_percpu(unsigned long);
 347         if (!bt->sequence)
 348                 goto err;
 349
 350         ret = -ENOENT;
 351         dir = blk_create_tree(buts->name);
 352         if (!dir)
 353                 goto err;
 354
 355         bt->dir = dir;
 356         bt->dev = bdev->bd_dev;
 357         atomic_set(&bt->dropped, 0);
 358
 359         ret = -EIO;
 360         bt->dropped_file = debugfs_create_file("dropped", 0444, dir, bt, &blk_dropped_fops);
 361         if (!bt->dropped_file)
 362                 goto err;
 363
 364         bt->rchan = relay_open("trace", dir, buts->buf_size,
 365                                 buts->buf_nr, &blk_relay_callbacks, bt);
 366         if (!bt->rchan)
 367                 goto err;
 368
 369         bt->act_mask = buts->act_mask;
 370         if (!bt->act_mask)
 371                 bt->act_mask = (u16) -1;
 372
 373         bt->start_lba = buts->start_lba;
 374         bt->end_lba = buts->end_lba;
 375         if (!bt->end_lba)
 376                 bt->end_lba = -1ULL;
 377
 378         bt->pid = buts->pid;
 379         bt->trace_state = Blktrace_setup;
 380
 381         ret = -EBUSY;
 382         old_bt = xchg(&q->blk_trace, bt);
 383         if (old_bt) {
 384                 (void) xchg(&q->blk_trace, old_bt);
 385                 goto err;
 386         }
 387
 388         return 0;
 389 err:
 390         if (dir)
 391                 blk_remove_tree(dir);
 392         if (bt) {
 393                 if (bt->dropped_file)
 394                         debugfs_remove(bt->dropped_file);
 395                 free_percpu(bt->sequence);
 396                 if (bt->rchan)
 397                         relay_close(bt->rchan);
 398                 kfree(bt);
 399         }
 400         return ret;
 401 }
 402
 403 static int blk_trace_setup(struct request_queue *q, struct block_device *bdev,
 404                            char __user *arg)
 405 {
 406         struct blk_user_trace_setup buts;
 407         int ret;
 408
 409         ret = copy_from_user(&buts, arg, sizeof(buts));
 410         if (ret)
 411                 return -EFAULT;
 412
 413         ret = do_blk_trace_setup(q, bdev, &buts);
 414         if (ret)
 415                 return ret;
 416
 417         if (copy_to_user(arg, &buts, sizeof(buts)))
 418                 return -EFAULT;
 419
 420         return 0;
 421 }
 422
 423 static int blk_trace_startstop(struct request_queue *q, int start)
 424 {
 425         struct blk_trace *bt;
 426         int ret;
 427
 428         if ((bt = q->blk_trace) == NULL)
 429                 return -EINVAL;
 430
 431         /*
 432          * For starting a trace, we can transition from a setup or stopped
 433          * trace. For stopping a trace, the state must be running
 434          */
 435         ret = -EINVAL;
 436         if (start) {
 437                 if (bt->trace_state == Blktrace_setup ||
 438                     bt->trace_state == Blktrace_stopped) {
 439                         blktrace_seq++;
 440                         smp_mb();
 441                         bt->trace_state = Blktrace_running;
 442
 443                         trace_note_time(bt);
 444                         ret = 0;
 445                 }
 446         } else {
 447                 if (bt->trace_state == Blktrace_running) {
 448                         bt->trace_state = Blktrace_stopped;
 449                         relay_flush(bt->rchan);
 450                         ret = 0;
 451                 }
 452         }
 453
 454         return ret;
 455 }
 456
 457 /**
 458  * blk_trace_ioctl: - handle the ioctls associated with tracing
 459  * @bdev:       the block device
 460  * @cmd:        the ioctl cmd
 461  * @arg:        the argument data, if any
 462  *
 463  **/
 464 int blk_trace_ioctl(struct block_device *bdev, unsigned cmd, char __user *arg)
 465 {
 466         struct request_queue *q;
 467         int ret, start = 0;
 468
 469         q = bdev_get_queue(bdev);
 470         if (!q)
 471                 return -ENXIO;
 472
 473         mutex_lock(&bdev->bd_mutex);
 474
 475         switch (cmd) {
 476         case BLKTRACESETUP:
 477                 ret = blk_trace_setup(q, bdev, arg);
 478                 break;
 479         case BLKTRACESTART:
 480                 start = 1;
 481         case BLKTRACESTOP:
 482                 ret = blk_trace_startstop(q, start);
 483                 break;
 484         case BLKTRACETEARDOWN:
 485                 ret = blk_trace_remove(q);
 486                 break;
 487         default:
 488                 ret = -ENOTTY;
 489                 break;
 490         }
 491
 492         mutex_unlock(&bdev->bd_mutex);
 493         return ret;
 494 }
 495
 496 /**
 497  * blk_trace_shutdown: - stop and cleanup trace structures
 498  * @q:    the request queue associated with the device
 499  *
 500  **/
 501 void blk_trace_shutdown(struct request_queue *q)
 502 {
 503         if (q->blk_trace) {
 504                 blk_trace_startstop(q, 0);
 505                 blk_trace_remove(q);
 506         }
 507 }
 508
 509 /*
 510  * Average offset over two calls to cpu_clock() with a gettimeofday()
 511  * in the middle
 512  */
 513 static void blk_check_time(unsigned long long *t, int this_cpu)
 514 {
 515         unsigned long long a, b;
 516         struct timeval tv;
 517
 518         a = cpu_clock(this_cpu);
 519         do_gettimeofday(&tv);
 520         b = cpu_clock(this_cpu);
 521
 522         *t = tv.tv_sec * 1000000000 + tv.tv_usec * 1000;
 523         *t -= (a + b) / 2;
 524 }
 525
 526 /*
 527  * calibrate our inter-CPU timings
 528  */
 529 static void blk_trace_check_cpu_time(void *data)
 530 {
 531         unsigned long long *t;
 532         int this_cpu = get_cpu();
 533
 534         t = &per_cpu(blk_trace_cpu_offset, this_cpu);
 535
 536         /*
 537          * Just call it twice, hopefully the second call will be cache hot
 538          * and a little more precise
 539          */
 540         blk_check_time(t, this_cpu);
 541         blk_check_time(t, this_cpu);
 542
 543         put_cpu();
 544 }
 545
 546 static void blk_trace_set_ht_offsets(void)
 547 {
 548 #if defined(CONFIG_SCHED_SMT)
 549         int cpu, i;
 550
 551         /*
 552          * now make sure HT siblings have the same time offset
 553          */
 554         preempt_disable();
 555         for_each_online_cpu(cpu) {
 556                 unsigned long long *cpu_off, *sibling_off;
 557
 558                 for_each_cpu_mask(i, per_cpu(cpu_sibling_map, cpu)) {
 559                         if (i == cpu)
 560                                 continue;
 561
 562                         cpu_off = &per_cpu(blk_trace_cpu_offset, cpu);
 563                         sibling_off = &per_cpu(blk_trace_cpu_offset, i);
 564                         *sibling_off = *cpu_off;
 565                 }
 566         }
 567         preempt_enable();
 568 #endif
 569 }
 570
 571 static __init int blk_trace_init(void)
 572 {
 573         mutex_init(&blk_tree_mutex);
 574         on_each_cpu(blk_trace_check_cpu_time, NULL, 1, 1);
 575         blk_trace_set_ht_offsets();
 576
 577         return 0;
 578 }
 579
 580 module_init(blk_trace_init);
 581