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 unsigned int blktrace_seq __read_mostly = 1;
  29
  30 /*
  31  * Send out a notify message.
  32  */
  33 static void trace_note(struct blk_trace *bt, pid_t pid, int action,
  34                        const void *data, size_t len)
  35 {
  36         struct blk_io_trace *t;
  37
  38         t = relay_reserve(bt->rchan, sizeof(*t) + len);
  39         if (t) {
  40                 const int cpu = smp_processor_id();
  41
  42                 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
  43                 t->time = ktime_to_ns(ktime_get());
  44                 t->device = bt->dev;
  45                 t->action = action;
  46                 t->pid = pid;
  47                 t->cpu = cpu;
  48                 t->pdu_len = len;
  49                 memcpy((void *) t + sizeof(*t), data, len);
  50         }
  51 }
  52
  53 /*
  54  * Send out a notify for this process, if we haven't done so since a trace
  55  * started
  56  */
  57 static void trace_note_tsk(struct blk_trace *bt, struct task_struct *tsk)
  58 {
  59         tsk->btrace_seq = blktrace_seq;
  60         trace_note(bt, tsk->pid, BLK_TN_PROCESS, tsk->comm, sizeof(tsk->comm));
  61 }
  62
  63 static void trace_note_time(struct blk_trace *bt)
  64 {
  65         struct timespec now;
  66         unsigned long flags;
  67         u32 words[2];
  68
  69         getnstimeofday(&now);
  70         words[0] = now.tv_sec;
  71         words[1] = now.tv_nsec;
  72
  73         local_irq_save(flags);
  74         trace_note(bt, 0, BLK_TN_TIMESTAMP, words, sizeof(words));
  75         local_irq_restore(flags);
  76 }
  77
  78 void __trace_note_message(struct blk_trace *bt, const char *fmt, ...)
  79 {
  80         int n;
  81         va_list args;
  82         unsigned long flags;
  83         char *buf;
  84
  85         local_irq_save(flags);
  86         buf = per_cpu_ptr(bt->msg_data, smp_processor_id());
  87         va_start(args, fmt);
  88         n = vscnprintf(buf, BLK_TN_MAX_MSG, fmt, args);
  89         va_end(args);
  90
  91         trace_note(bt, 0, BLK_TN_MESSAGE, buf, n);
  92         local_irq_restore(flags);
  93 }
  94 EXPORT_SYMBOL_GPL(__trace_note_message);
  95
  96 static int act_log_check(struct blk_trace *bt, u32 what, sector_t sector,
  97                          pid_t pid)
  98 {
  99         if (((bt->act_mask << BLK_TC_SHIFT) & what) == 0)
 100                 return 1;
 101         if (sector < bt->start_lba || sector > bt->end_lba)
 102                 return 1;
 103         if (bt->pid && pid != bt->pid)
 104                 return 1;
 105
 106         return 0;
 107 }
 108
 109 /*
 110  * Data direction bit lookup
 111  */
 112 static u32 ddir_act[2] __read_mostly = { BLK_TC_ACT(BLK_TC_READ), BLK_TC_ACT(BLK_TC_WRITE) };
 113
 114 /*
 115  * Bio action bits of interest
 116  */
 117 static u32 bio_act[17] __read_mostly = {
 118         [1] = BLK_TC_ACT(BLK_TC_BARRIER),
 119         [2] = BLK_TC_ACT(BLK_TC_SYNC),
 120         [4] = BLK_TC_ACT(BLK_TC_AHEAD),
 121         [8] = BLK_TC_ACT(BLK_TC_META),
 122         [16] = BLK_TC_ACT(BLK_TC_DISCARD)
 123 };
 124
 125 /*
 126  * More could be added as needed, taking care to increment the decrementer
 127  * to get correct indexing
 128  */
 129 #define trace_barrier_bit(rw)   \
 130         (((rw) & (1 << BIO_RW_BARRIER)) >> (BIO_RW_BARRIER - 0))
 131 #define trace_sync_bit(rw)      \
 132         (((rw) & (1 << BIO_RW_SYNC)) >> (BIO_RW_SYNC - 1))
 133 #define trace_ahead_bit(rw)     \
 134         (((rw) & (1 << BIO_RW_AHEAD)) << (2 - BIO_RW_AHEAD))
 135 #define trace_meta_bit(rw)      \
 136         (((rw) & (1 << BIO_RW_META)) >> (BIO_RW_META - 3))
 137 #define trace_discard_bit(rw)   \
 138         (((rw) & (1 << BIO_RW_DISCARD)) >> (BIO_RW_DISCARD - 4))
 139
 140 /*
 141  * The worker for the various blk_add_trace*() types. Fills out a
 142  * blk_io_trace structure and places it in a per-cpu subbuffer.
 143  */
 144 void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
 145                      int rw, u32 what, int error, int pdu_len, void *pdu_data)
 146 {
 147         struct task_struct *tsk = current;
 148         struct blk_io_trace *t;
 149         unsigned long flags;
 150         unsigned long *sequence;
 151         pid_t pid;
 152         int cpu;
 153
 154         if (unlikely(bt->trace_state != Blktrace_running))
 155                 return;
 156
 157         what |= ddir_act[rw & WRITE];
 158         what |= bio_act[trace_barrier_bit(rw)];
 159         what |= bio_act[trace_sync_bit(rw)];
 160         what |= bio_act[trace_ahead_bit(rw)];
 161         what |= bio_act[trace_meta_bit(rw)];
 162         what |= bio_act[trace_discard_bit(rw)];
 163
 164         pid = tsk->pid;
 165         if (unlikely(act_log_check(bt, what, sector, pid)))
 166                 return;
 167
 168         /*
 169          * A word about the locking here - we disable interrupts to reserve
 170          * some space in the relay per-cpu buffer, to prevent an irq
 171          * from coming in and stepping on our toes.
 172          */
 173         local_irq_save(flags);
 174
 175         if (unlikely(tsk->btrace_seq != blktrace_seq))
 176                 trace_note_tsk(bt, tsk);
 177
 178         t = relay_reserve(bt->rchan, sizeof(*t) + pdu_len);
 179         if (t) {
 180                 cpu = smp_processor_id();
 181                 sequence = per_cpu_ptr(bt->sequence, cpu);
 182
 183                 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
 184                 t->sequence = ++(*sequence);
 185                 t->time = ktime_to_ns(ktime_get());
 186                 t->sector = sector;
 187                 t->bytes = bytes;
 188                 t->action = what;
 189                 t->pid = pid;
 190                 t->device = bt->dev;
 191                 t->cpu = cpu;
 192                 t->error = error;
 193                 t->pdu_len = pdu_len;
 194
 195                 if (pdu_len)
 196                         memcpy((void *) t + sizeof(*t), pdu_data, pdu_len);
 197         }
 198
 199         local_irq_restore(flags);
 200 }
 201
 202 EXPORT_SYMBOL_GPL(__blk_add_trace);
 203
 204 static struct dentry *blk_tree_root;
 205 static DEFINE_MUTEX(blk_tree_mutex);
 206 static unsigned int root_users;
 207
 208 static inline void blk_remove_root(void)
 209 {
 210         if (blk_tree_root) {
 211                 debugfs_remove(blk_tree_root);
 212                 blk_tree_root = NULL;
 213         }
 214 }
 215
 216 static void blk_remove_tree(struct dentry *dir)
 217 {
 218         mutex_lock(&blk_tree_mutex);
 219         debugfs_remove(dir);
 220         if (--root_users == 0)
 221                 blk_remove_root();
 222         mutex_unlock(&blk_tree_mutex);
 223 }
 224
 225 static struct dentry *blk_create_tree(const char *blk_name)
 226 {
 227         struct dentry *dir = NULL;
 228         int created = 0;
 229
 230         mutex_lock(&blk_tree_mutex);
 231
 232         if (!blk_tree_root) {
 233                 blk_tree_root = debugfs_create_dir("block", NULL);
 234                 if (!blk_tree_root)
 235                         goto err;
 236                 created = 1;
 237         }
 238
 239         dir = debugfs_create_dir(blk_name, blk_tree_root);
 240         if (dir)
 241                 root_users++;
 242         else {
 243                 /* Delete root only if we created it */
 244                 if (created)
 245                         blk_remove_root();
 246         }
 247
 248 err:
 249         mutex_unlock(&blk_tree_mutex);
 250         return dir;
 251 }
 252
 253 static void blk_trace_cleanup(struct blk_trace *bt)
 254 {
 255         relay_close(bt->rchan);
 256         debugfs_remove(bt->msg_file);
 257         debugfs_remove(bt->dropped_file);
 258         blk_remove_tree(bt->dir);
 259         free_percpu(bt->sequence);
 260         free_percpu(bt->msg_data);
 261         kfree(bt);
 262 }
 263
 264 int blk_trace_remove(struct request_queue *q)
 265 {
 266         struct blk_trace *bt;
 267
 268         bt = xchg(&q->blk_trace, NULL);
 269         if (!bt)
 270                 return -EINVAL;
 271
 272         if (bt->trace_state == Blktrace_setup ||
 273             bt->trace_state == Blktrace_stopped)
 274                 blk_trace_cleanup(bt);
 275
 276         return 0;
 277 }
 278 EXPORT_SYMBOL_GPL(blk_trace_remove);
 279
 280 static int blk_dropped_open(struct inode *inode, struct file *filp)
 281 {
 282         filp->private_data = inode->i_private;
 283
 284         return 0;
 285 }
 286
 287 static ssize_t blk_dropped_read(struct file *filp, char __user *buffer,
 288                                 size_t count, loff_t *ppos)
 289 {
 290         struct blk_trace *bt = filp->private_data;
 291         char buf[16];
 292
 293         snprintf(buf, sizeof(buf), "%u\n", atomic_read(&bt->dropped));
 294
 295         return simple_read_from_buffer(buffer, count, ppos, buf, strlen(buf));
 296 }
 297
 298 static const struct file_operations blk_dropped_fops = {
 299         .owner =        THIS_MODULE,
 300         .open =         blk_dropped_open,
 301         .read =         blk_dropped_read,
 302 };
 303
 304 static int blk_msg_open(struct inode *inode, struct file *filp)
 305 {
 306         filp->private_data = inode->i_private;
 307
 308         return 0;
 309 }
 310
 311 static ssize_t blk_msg_write(struct file *filp, const char __user *buffer,
 312                                 size_t count, loff_t *ppos)
 313 {
 314         char *msg;
 315         struct blk_trace *bt;
 316
 317         if (count > BLK_TN_MAX_MSG)
 318                 return -EINVAL;
 319
 320         msg = kmalloc(count, GFP_KERNEL);
 321         if (msg == NULL)
 322                 return -ENOMEM;
 323
 324         if (copy_from_user(msg, buffer, count)) {
 325                 kfree(msg);
 326                 return -EFAULT;
 327         }
 328
 329         bt = filp->private_data;
 330         __trace_note_message(bt, "%s", msg);
 331         kfree(msg);
 332
 333         return count;
 334 }
 335
 336 static const struct file_operations blk_msg_fops = {
 337         .owner =        THIS_MODULE,
 338         .open =         blk_msg_open,
 339         .write =        blk_msg_write,
 340 };
 341
 342 /*
 343  * Keep track of how many times we encountered a full subbuffer, to aid
 344  * the user space app in telling how many lost events there were.
 345  */
 346 static int blk_subbuf_start_callback(struct rchan_buf *buf, void *subbuf,
 347                                      void *prev_subbuf, size_t prev_padding)
 348 {
 349         struct blk_trace *bt;
 350
 351         if (!relay_buf_full(buf))
 352                 return 1;
 353
 354         bt = buf->chan->private_data;
 355         atomic_inc(&bt->dropped);
 356         return 0;
 357 }
 358
 359 static int blk_remove_buf_file_callback(struct dentry *dentry)
 360 {
 361         debugfs_remove(dentry);
 362         return 0;
 363 }
 364
 365 static struct dentry *blk_create_buf_file_callback(const char *filename,
 366                                                    struct dentry *parent,
 367                                                    int mode,
 368                                                    struct rchan_buf *buf,
 369                                                    int *is_global)
 370 {
 371         return debugfs_create_file(filename, mode, parent, buf,
 372                                         &relay_file_operations);
 373 }
 374
 375 static struct rchan_callbacks blk_relay_callbacks = {
 376         .subbuf_start           = blk_subbuf_start_callback,
 377         .create_buf_file        = blk_create_buf_file_callback,
 378         .remove_buf_file        = blk_remove_buf_file_callback,
 379 };
 380
 381 /*
 382  * Setup everything required to start tracing
 383  */
 384 int do_blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
 385                         struct blk_user_trace_setup *buts)
 386 {
 387         struct blk_trace *old_bt, *bt = NULL;
 388         struct dentry *dir = NULL;
 389         int ret, i;
 390
 391         if (!buts->buf_size || !buts->buf_nr)
 392                 return -EINVAL;
 393
 394         strcpy(buts->name, name);
 395
 396         /*
 397          * some device names have larger paths - convert the slashes
 398          * to underscores for this to work as expected
 399          */
 400         for (i = 0; i < strlen(buts->name); i++)
 401                 if (buts->name[i] == '/')
 402                         buts->name[i] = '_';
 403
 404         ret = -ENOMEM;
 405         bt = kzalloc(sizeof(*bt), GFP_KERNEL);
 406         if (!bt)
 407                 goto err;
 408
 409         bt->sequence = alloc_percpu(unsigned long);
 410         if (!bt->sequence)
 411                 goto err;
 412
 413         bt->msg_data = __alloc_percpu(BLK_TN_MAX_MSG);
 414         if (!bt->msg_data)
 415                 goto err;
 416
 417         ret = -ENOENT;
 418         dir = blk_create_tree(buts->name);
 419         if (!dir)
 420                 goto err;
 421
 422         bt->dir = dir;
 423         bt->dev = dev;
 424         atomic_set(&bt->dropped, 0);
 425
 426         ret = -EIO;
 427         bt->dropped_file = debugfs_create_file("dropped", 0444, dir, bt, &blk_dropped_fops);
 428         if (!bt->dropped_file)
 429                 goto err;
 430
 431         bt->msg_file = debugfs_create_file("msg", 0222, dir, bt, &blk_msg_fops);
 432         if (!bt->msg_file)
 433                 goto err;
 434
 435         bt->rchan = relay_open("trace", dir, buts->buf_size,
 436                                 buts->buf_nr, &blk_relay_callbacks, bt);
 437         if (!bt->rchan)
 438                 goto err;
 439
 440         bt->act_mask = buts->act_mask;
 441         if (!bt->act_mask)
 442                 bt->act_mask = (u16) -1;
 443
 444         bt->start_lba = buts->start_lba;
 445         bt->end_lba = buts->end_lba;
 446         if (!bt->end_lba)
 447                 bt->end_lba = -1ULL;
 448
 449         bt->pid = buts->pid;
 450         bt->trace_state = Blktrace_setup;
 451
 452         ret = -EBUSY;
 453         old_bt = xchg(&q->blk_trace, bt);
 454         if (old_bt) {
 455                 (void) xchg(&q->blk_trace, old_bt);
 456                 goto err;
 457         }
 458
 459         return 0;
 460 err:
 461         if (dir)
 462                 blk_remove_tree(dir);
 463         if (bt) {
 464                 if (bt->msg_file)
 465                         debugfs_remove(bt->msg_file);
 466                 if (bt->dropped_file)
 467                         debugfs_remove(bt->dropped_file);
 468                 free_percpu(bt->sequence);
 469                 free_percpu(bt->msg_data);
 470                 if (bt->rchan)
 471                         relay_close(bt->rchan);
 472                 kfree(bt);
 473         }
 474         return ret;
 475 }
 476
 477 int blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
 478                     char __user *arg)
 479 {
 480         struct blk_user_trace_setup buts;
 481         int ret;
 482
 483         ret = copy_from_user(&buts, arg, sizeof(buts));
 484         if (ret)
 485                 return -EFAULT;
 486
 487         ret = do_blk_trace_setup(q, name, dev, &buts);
 488         if (ret)
 489                 return ret;
 490
 491         if (copy_to_user(arg, &buts, sizeof(buts)))
 492                 return -EFAULT;
 493
 494         return 0;
 495 }
 496 EXPORT_SYMBOL_GPL(blk_trace_setup);
 497
 498 int blk_trace_startstop(struct request_queue *q, int start)
 499 {
 500         struct blk_trace *bt;
 501         int ret;
 502
 503         if ((bt = q->blk_trace) == NULL)
 504                 return -EINVAL;
 505
 506         /*
 507          * For starting a trace, we can transition from a setup or stopped
 508          * trace. For stopping a trace, the state must be running
 509          */
 510         ret = -EINVAL;
 511         if (start) {
 512                 if (bt->trace_state == Blktrace_setup ||
 513                     bt->trace_state == Blktrace_stopped) {
 514                         blktrace_seq++;
 515                         smp_mb();
 516                         bt->trace_state = Blktrace_running;
 517
 518                         trace_note_time(bt);
 519                         ret = 0;
 520                 }
 521         } else {
 522                 if (bt->trace_state == Blktrace_running) {
 523                         bt->trace_state = Blktrace_stopped;
 524                         relay_flush(bt->rchan);
 525                         ret = 0;
 526                 }
 527         }
 528
 529         return ret;
 530 }
 531 EXPORT_SYMBOL_GPL(blk_trace_startstop);
 532
 533 /**
 534  * blk_trace_ioctl: - handle the ioctls associated with tracing
 535  * @bdev:       the block device
 536  * @cmd:        the ioctl cmd
 537  * @arg:        the argument data, if any
 538  *
 539  **/
 540 int blk_trace_ioctl(struct block_device *bdev, unsigned cmd, char __user *arg)
 541 {
 542         struct request_queue *q;
 543         int ret, start = 0;
 544         char b[BDEVNAME_SIZE];
 545
 546         q = bdev_get_queue(bdev);
 547         if (!q)
 548                 return -ENXIO;
 549
 550         mutex_lock(&bdev->bd_mutex);
 551
 552         switch (cmd) {
 553         case BLKTRACESETUP:
 554                 bdevname(bdev, b);
 555                 ret = blk_trace_setup(q, b, bdev->bd_dev, arg);
 556                 break;
 557         case BLKTRACESTART:
 558                 start = 1;
 559         case BLKTRACESTOP:
 560                 ret = blk_trace_startstop(q, start);
 561                 break;
 562         case BLKTRACETEARDOWN:
 563                 ret = blk_trace_remove(q);
 564                 break;
 565         default:
 566                 ret = -ENOTTY;
 567                 break;
 568         }
 569
 570         mutex_unlock(&bdev->bd_mutex);
 571         return ret;
 572 }
 573
 574 /**
 575  * blk_trace_shutdown: - stop and cleanup trace structures
 576  * @q:    the request queue associated with the device
 577  *
 578  **/
 579 void blk_trace_shutdown(struct request_queue *q)
 580 {
 581         if (q->blk_trace) {
 582                 blk_trace_startstop(q, 0);
 583                 blk_trace_remove(q);
 584         }
 585 }