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
 206         mutex_lock(&blk_tree_mutex);
 207
 208         if (!blk_tree_root) {
 209                 blk_tree_root = debugfs_create_dir("block", NULL);
 210                 if (!blk_tree_root)
 211                         goto err;
 212         }
 213
 214         dir = debugfs_create_dir(blk_name, blk_tree_root);
 215         if (dir)
 216                 root_users++;
 217         else
 218                 blk_remove_root();
 219
 220 err:
 221         mutex_unlock(&blk_tree_mutex);
 222         return dir;
 223 }
 224
 225 static void blk_trace_cleanup(struct blk_trace *bt)
 226 {
 227         relay_close(bt->rchan);
 228         debugfs_remove(bt->dropped_file);
 229         blk_remove_tree(bt->dir);
 230         free_percpu(bt->sequence);
 231         kfree(bt);
 232 }
 233
 234 static int blk_trace_remove(struct request_queue *q)
 235 {
 236         struct blk_trace *bt;
 237
 238         bt = xchg(&q->blk_trace, NULL);
 239         if (!bt)
 240                 return -EINVAL;
 241
 242         if (bt->trace_state == Blktrace_setup ||
 243             bt->trace_state == Blktrace_stopped)
 244                 blk_trace_cleanup(bt);
 245
 246         return 0;
 247 }
 248
 249 static int blk_dropped_open(struct inode *inode, struct file *filp)
 250 {
 251         filp->private_data = inode->i_private;
 252
 253         return 0;
 254 }
 255
 256 static ssize_t blk_dropped_read(struct file *filp, char __user *buffer,
 257                                 size_t count, loff_t *ppos)
 258 {
 259         struct blk_trace *bt = filp->private_data;
 260         char buf[16];
 261
 262         snprintf(buf, sizeof(buf), "%u\n", atomic_read(&bt->dropped));
 263
 264         return simple_read_from_buffer(buffer, count, ppos, buf, strlen(buf));
 265 }
 266
 267 static const struct file_operations blk_dropped_fops = {
 268         .owner =        THIS_MODULE,
 269         .open =         blk_dropped_open,
 270         .read =         blk_dropped_read,
 271 };
 272
 273 /*
 274  * Keep track of how many times we encountered a full subbuffer, to aid
 275  * the user space app in telling how many lost events there were.
 276  */
 277 static int blk_subbuf_start_callback(struct rchan_buf *buf, void *subbuf,
 278                                      void *prev_subbuf, size_t prev_padding)
 279 {
 280         struct blk_trace *bt;
 281
 282         if (!relay_buf_full(buf))
 283                 return 1;
 284
 285         bt = buf->chan->private_data;
 286         atomic_inc(&bt->dropped);
 287         return 0;
 288 }
 289
 290 static int blk_remove_buf_file_callback(struct dentry *dentry)
 291 {
 292         debugfs_remove(dentry);
 293         return 0;
 294 }
 295
 296 static struct dentry *blk_create_buf_file_callback(const char *filename,
 297                                                    struct dentry *parent,
 298                                                    int mode,
 299                                                    struct rchan_buf *buf,
 300                                                    int *is_global)
 301 {
 302         return debugfs_create_file(filename, mode, parent, buf,
 303                                         &relay_file_operations);
 304 }
 305
 306 static struct rchan_callbacks blk_relay_callbacks = {
 307         .subbuf_start           = blk_subbuf_start_callback,
 308         .create_buf_file        = blk_create_buf_file_callback,
 309         .remove_buf_file        = blk_remove_buf_file_callback,
 310 };
 311
 312 /*
 313  * Setup everything required to start tracing
 314  */
 315 int do_blk_trace_setup(struct request_queue *q, struct block_device *bdev,
 316                         struct blk_user_trace_setup *buts)
 317 {
 318         struct blk_trace *old_bt, *bt = NULL;
 319         struct dentry *dir = NULL;
 320         char b[BDEVNAME_SIZE];
 321         int ret, i;
 322
 323         if (!buts->buf_size || !buts->buf_nr)
 324                 return -EINVAL;
 325
 326         strcpy(buts->name, bdevname(bdev, b));
 327
 328         /*
 329          * some device names have larger paths - convert the slashes
 330          * to underscores for this to work as expected
 331          */
 332         for (i = 0; i < strlen(buts->name); i++)
 333                 if (buts->name[i] == '/')
 334                         buts->name[i] = '_';
 335
 336         ret = -ENOMEM;
 337         bt = kzalloc(sizeof(*bt), GFP_KERNEL);
 338         if (!bt)
 339                 goto err;
 340
 341         bt->sequence = alloc_percpu(unsigned long);
 342         if (!bt->sequence)
 343                 goto err;
 344
 345         ret = -ENOENT;
 346         dir = blk_create_tree(buts->name);
 347         if (!dir)
 348                 goto err;
 349
 350         bt->dir = dir;
 351         bt->dev = bdev->bd_dev;
 352         atomic_set(&bt->dropped, 0);
 353
 354         ret = -EIO;
 355         bt->dropped_file = debugfs_create_file("dropped", 0444, dir, bt, &blk_dropped_fops);
 356         if (!bt->dropped_file)
 357                 goto err;
 358
 359         bt->rchan = relay_open("trace", dir, buts->buf_size,
 360                                 buts->buf_nr, &blk_relay_callbacks, bt);
 361         if (!bt->rchan)
 362                 goto err;
 363
 364         bt->act_mask = buts->act_mask;
 365         if (!bt->act_mask)
 366                 bt->act_mask = (u16) -1;
 367
 368         bt->start_lba = buts->start_lba;
 369         bt->end_lba = buts->end_lba;
 370         if (!bt->end_lba)
 371                 bt->end_lba = -1ULL;
 372
 373         bt->pid = buts->pid;
 374         bt->trace_state = Blktrace_setup;
 375
 376         ret = -EBUSY;
 377         old_bt = xchg(&q->blk_trace, bt);
 378         if (old_bt) {
 379                 (void) xchg(&q->blk_trace, old_bt);
 380                 goto err;
 381         }
 382
 383         return 0;
 384 err:
 385         if (dir)
 386                 blk_remove_tree(dir);
 387         if (bt) {
 388                 if (bt->dropped_file)
 389                         debugfs_remove(bt->dropped_file);
 390                 free_percpu(bt->sequence);
 391                 if (bt->rchan)
 392                         relay_close(bt->rchan);
 393                 kfree(bt);
 394         }
 395         return ret;
 396 }
 397
 398 static int blk_trace_setup(struct request_queue *q, struct block_device *bdev,
 399                            char __user *arg)
 400 {
 401         struct blk_user_trace_setup buts;
 402         int ret;
 403
 404         ret = copy_from_user(&buts, arg, sizeof(buts));
 405         if (ret)
 406                 return -EFAULT;
 407
 408         ret = do_blk_trace_setup(q, bdev, &buts);
 409         if (ret)
 410                 return ret;
 411
 412         if (copy_to_user(arg, &buts, sizeof(buts)))
 413                 return -EFAULT;
 414
 415         return 0;
 416 }
 417
 418 static int blk_trace_startstop(struct request_queue *q, int start)
 419 {
 420         struct blk_trace *bt;
 421         int ret;
 422
 423         if ((bt = q->blk_trace) == NULL)
 424                 return -EINVAL;
 425
 426         /*
 427          * For starting a trace, we can transition from a setup or stopped
 428          * trace. For stopping a trace, the state must be running
 429          */
 430         ret = -EINVAL;
 431         if (start) {
 432                 if (bt->trace_state == Blktrace_setup ||
 433                     bt->trace_state == Blktrace_stopped) {
 434                         blktrace_seq++;
 435                         smp_mb();
 436                         bt->trace_state = Blktrace_running;
 437
 438                         trace_note_time(bt);
 439                         ret = 0;
 440                 }
 441         } else {
 442                 if (bt->trace_state == Blktrace_running) {
 443                         bt->trace_state = Blktrace_stopped;
 444                         relay_flush(bt->rchan);
 445                         ret = 0;
 446                 }
 447         }
 448
 449         return ret;
 450 }
 451
 452 /**
 453  * blk_trace_ioctl: - handle the ioctls associated with tracing
 454  * @bdev:       the block device
 455  * @cmd:        the ioctl cmd
 456  * @arg:        the argument data, if any
 457  *
 458  **/
 459 int blk_trace_ioctl(struct block_device *bdev, unsigned cmd, char __user *arg)
 460 {
 461         struct request_queue *q;
 462         int ret, start = 0;
 463
 464         q = bdev_get_queue(bdev);
 465         if (!q)
 466                 return -ENXIO;
 467
 468         mutex_lock(&bdev->bd_mutex);
 469
 470         switch (cmd) {
 471         case BLKTRACESETUP:
 472                 ret = blk_trace_setup(q, bdev, arg);
 473                 break;
 474         case BLKTRACESTART:
 475                 start = 1;
 476         case BLKTRACESTOP:
 477                 ret = blk_trace_startstop(q, start);
 478                 break;
 479         case BLKTRACETEARDOWN:
 480                 ret = blk_trace_remove(q);
 481                 break;
 482         default:
 483                 ret = -ENOTTY;
 484                 break;
 485         }
 486
 487         mutex_unlock(&bdev->bd_mutex);
 488         return ret;
 489 }
 490
 491 /**
 492  * blk_trace_shutdown: - stop and cleanup trace structures
 493  * @q:    the request queue associated with the device
 494  *
 495  **/
 496 void blk_trace_shutdown(struct request_queue *q)
 497 {
 498         if (q->blk_trace) {
 499                 blk_trace_startstop(q, 0);
 500                 blk_trace_remove(q);
 501         }
 502 }
 503
 504 /*
 505  * Average offset over two calls to cpu_clock() with a gettimeofday()
 506  * in the middle
 507  */
 508 static void blk_check_time(unsigned long long *t, int this_cpu)
 509 {
 510         unsigned long long a, b;
 511         struct timeval tv;
 512
 513         a = cpu_clock(this_cpu);
 514         do_gettimeofday(&tv);
 515         b = cpu_clock(this_cpu);
 516
 517         *t = tv.tv_sec * 1000000000 + tv.tv_usec * 1000;
 518         *t -= (a + b) / 2;
 519 }
 520
 521 /*
 522  * calibrate our inter-CPU timings
 523  */
 524 static void blk_trace_check_cpu_time(void *data)
 525 {
 526         unsigned long long *t;
 527         int this_cpu = get_cpu();
 528
 529         t = &per_cpu(blk_trace_cpu_offset, this_cpu);
 530
 531         /*
 532          * Just call it twice, hopefully the second call will be cache hot
 533          * and a little more precise
 534          */
 535         blk_check_time(t, this_cpu);
 536         blk_check_time(t, this_cpu);
 537
 538         put_cpu();
 539 }
 540
 541 static void blk_trace_set_ht_offsets(void)
 542 {
 543 #if defined(CONFIG_SCHED_SMT)
 544         int cpu, i;
 545
 546         /*
 547          * now make sure HT siblings have the same time offset
 548          */
 549         preempt_disable();
 550         for_each_online_cpu(cpu) {
 551                 unsigned long long *cpu_off, *sibling_off;
 552
 553                 for_each_cpu_mask(i, cpu_sibling_map[cpu]) {
 554                         if (i == cpu)
 555                                 continue;
 556
 557                         cpu_off = &per_cpu(blk_trace_cpu_offset, cpu);
 558                         sibling_off = &per_cpu(blk_trace_cpu_offset, i);
 559                         *sibling_off = *cpu_off;
 560                 }
 561         }
 562         preempt_enable();
 563 #endif
 564 }
 565
 566 static __init int blk_trace_init(void)
 567 {
 568         mutex_init(&blk_tree_mutex);
 569         on_each_cpu(blk_trace_check_cpu_time, NULL, 1, 1);
 570         blk_trace_set_ht_offsets();
 571
 572         return 0;
 573 }
 574
 575 module_init(blk_trace_init);
 576