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