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 static int act_log_check(struct blk_trace *bt, u32 what, sector_t sector,
  79                          pid_t pid)
  80 {
  81         if (((bt->act_mask << BLK_TC_SHIFT) & what) == 0)
  82                 return 1;
  83         if (sector < bt->start_lba || sector > bt->end_lba)
  84                 return 1;
  85         if (bt->pid && pid != bt->pid)
  86                 return 1;
  87
  88         return 0;
  89 }
  90
  91 /*
  92  * Data direction bit lookup
  93  */
  94 static u32 ddir_act[2] __read_mostly = { BLK_TC_ACT(BLK_TC_READ), BLK_TC_ACT(BLK_TC_WRITE) };
  95
  96 /*
  97  * Bio action bits of interest
  98  */
  99 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) };
 100
 101 /*
 102  * More could be added as needed, taking care to increment the decrementer
 103  * to get correct indexing
 104  */
 105 #define trace_barrier_bit(rw)   \
 106         (((rw) & (1 << BIO_RW_BARRIER)) >> (BIO_RW_BARRIER - 0))
 107 #define trace_sync_bit(rw)      \
 108         (((rw) & (1 << BIO_RW_SYNC)) >> (BIO_RW_SYNC - 1))
 109 #define trace_ahead_bit(rw)     \
 110         (((rw) & (1 << BIO_RW_AHEAD)) << (2 - BIO_RW_AHEAD))
 111 #define trace_meta_bit(rw)      \
 112         (((rw) & (1 << BIO_RW_META)) >> (BIO_RW_META - 3))
 113
 114 /*
 115  * The worker for the various blk_add_trace*() types. Fills out a
 116  * blk_io_trace structure and places it in a per-cpu subbuffer.
 117  */
 118 void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
 119                      int rw, u32 what, int error, int pdu_len, void *pdu_data)
 120 {
 121         struct task_struct *tsk = current;
 122         struct blk_io_trace *t;
 123         unsigned long flags;
 124         unsigned long *sequence;
 125         pid_t pid;
 126         int cpu;
 127
 128         if (unlikely(bt->trace_state != Blktrace_running))
 129                 return;
 130
 131         what |= ddir_act[rw & WRITE];
 132         what |= bio_act[trace_barrier_bit(rw)];
 133         what |= bio_act[trace_sync_bit(rw)];
 134         what |= bio_act[trace_ahead_bit(rw)];
 135         what |= bio_act[trace_meta_bit(rw)];
 136
 137         pid = tsk->pid;
 138         if (unlikely(act_log_check(bt, what, sector, pid)))
 139                 return;
 140
 141         /*
 142          * A word about the locking here - we disable interrupts to reserve
 143          * some space in the relay per-cpu buffer, to prevent an irq
 144          * from coming in and stepping on our toes. Once reserved, it's
 145          * enough to get preemption disabled to prevent read of this data
 146          * before we are through filling it. get_cpu()/put_cpu() does this
 147          * for us
 148          */
 149         local_irq_save(flags);
 150
 151         if (unlikely(tsk->btrace_seq != blktrace_seq))
 152                 trace_note_tsk(bt, tsk);
 153
 154         t = relay_reserve(bt->rchan, sizeof(*t) + pdu_len);
 155         if (t) {
 156                 cpu = smp_processor_id();
 157                 sequence = per_cpu_ptr(bt->sequence, cpu);
 158
 159                 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
 160                 t->sequence = ++(*sequence);
 161                 t->time = ktime_to_ns(ktime_get());
 162                 t->sector = sector;
 163                 t->bytes = bytes;
 164                 t->action = what;
 165                 t->pid = pid;
 166                 t->device = bt->dev;
 167                 t->cpu = cpu;
 168                 t->error = error;
 169                 t->pdu_len = pdu_len;
 170
 171                 if (pdu_len)
 172                         memcpy((void *) t + sizeof(*t), pdu_data, pdu_len);
 173         }
 174
 175         local_irq_restore(flags);
 176 }
 177
 178 EXPORT_SYMBOL_GPL(__blk_add_trace);
 179
 180 static struct dentry *blk_tree_root;
 181 static DEFINE_MUTEX(blk_tree_mutex);
 182 static unsigned int root_users;
 183
 184 static inline void blk_remove_root(void)
 185 {
 186         if (blk_tree_root) {
 187                 debugfs_remove(blk_tree_root);
 188                 blk_tree_root = NULL;
 189         }
 190 }
 191
 192 static void blk_remove_tree(struct dentry *dir)
 193 {
 194         mutex_lock(&blk_tree_mutex);
 195         debugfs_remove(dir);
 196         if (--root_users == 0)
 197                 blk_remove_root();
 198         mutex_unlock(&blk_tree_mutex);
 199 }
 200
 201 static struct dentry *blk_create_tree(const char *blk_name)
 202 {
 203         struct dentry *dir = NULL;
 204         int created = 0;
 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                 created = 1;
 213         }
 214
 215         dir = debugfs_create_dir(blk_name, blk_tree_root);
 216         if (dir)
 217                 root_users++;
 218         else {
 219                 /* Delete root only if we created it */
 220                 if (created)
 221                         blk_remove_root();
 222         }
 223
 224 err:
 225         mutex_unlock(&blk_tree_mutex);
 226         return dir;
 227 }
 228
 229 static void blk_trace_cleanup(struct blk_trace *bt)
 230 {
 231         relay_close(bt->rchan);
 232         debugfs_remove(bt->dropped_file);
 233         blk_remove_tree(bt->dir);
 234         free_percpu(bt->sequence);
 235         kfree(bt);
 236 }
 237
 238 static int blk_trace_remove(struct request_queue *q)
 239 {
 240         struct blk_trace *bt;
 241
 242         bt = xchg(&q->blk_trace, NULL);
 243         if (!bt)
 244                 return -EINVAL;
 245
 246         if (bt->trace_state == Blktrace_setup ||
 247             bt->trace_state == Blktrace_stopped)
 248                 blk_trace_cleanup(bt);
 249
 250         return 0;
 251 }
 252
 253 static int blk_dropped_open(struct inode *inode, struct file *filp)
 254 {
 255         filp->private_data = inode->i_private;
 256
 257         return 0;
 258 }
 259
 260 static ssize_t blk_dropped_read(struct file *filp, char __user *buffer,
 261                                 size_t count, loff_t *ppos)
 262 {
 263         struct blk_trace *bt = filp->private_data;
 264         char buf[16];
 265
 266         snprintf(buf, sizeof(buf), "%u\n", atomic_read(&bt->dropped));
 267
 268         return simple_read_from_buffer(buffer, count, ppos, buf, strlen(buf));
 269 }
 270
 271 static const struct file_operations blk_dropped_fops = {
 272         .owner =        THIS_MODULE,
 273         .open =         blk_dropped_open,
 274         .read =         blk_dropped_read,
 275 };
 276
 277 /*
 278  * Keep track of how many times we encountered a full subbuffer, to aid
 279  * the user space app in telling how many lost events there were.
 280  */
 281 static int blk_subbuf_start_callback(struct rchan_buf *buf, void *subbuf,
 282                                      void *prev_subbuf, size_t prev_padding)
 283 {
 284         struct blk_trace *bt;
 285
 286         if (!relay_buf_full(buf))
 287                 return 1;
 288
 289         bt = buf->chan->private_data;
 290         atomic_inc(&bt->dropped);
 291         return 0;
 292 }
 293
 294 static int blk_remove_buf_file_callback(struct dentry *dentry)
 295 {
 296         debugfs_remove(dentry);
 297         return 0;
 298 }
 299
 300 static struct dentry *blk_create_buf_file_callback(const char *filename,
 301                                                    struct dentry *parent,
 302                                                    int mode,
 303                                                    struct rchan_buf *buf,
 304                                                    int *is_global)
 305 {
 306         return debugfs_create_file(filename, mode, parent, buf,
 307                                         &relay_file_operations);
 308 }
 309
 310 static struct rchan_callbacks blk_relay_callbacks = {
 311         .subbuf_start           = blk_subbuf_start_callback,
 312         .create_buf_file        = blk_create_buf_file_callback,
 313         .remove_buf_file        = blk_remove_buf_file_callback,
 314 };
 315
 316 /*
 317  * Setup everything required to start tracing
 318  */
 319 int do_blk_trace_setup(struct request_queue *q, struct block_device *bdev,
 320                         struct blk_user_trace_setup *buts)
 321 {
 322         struct blk_trace *old_bt, *bt = NULL;
 323         struct dentry *dir = NULL;
 324         char b[BDEVNAME_SIZE];
 325         int ret, i;
 326
 327         if (!buts->buf_size || !buts->buf_nr)
 328                 return -EINVAL;
 329
 330         strcpy(buts->name, bdevname(bdev, b));
 331
 332         /*
 333          * some device names have larger paths - convert the slashes
 334          * to underscores for this to work as expected
 335          */
 336         for (i = 0; i < strlen(buts->name); i++)
 337                 if (buts->name[i] == '/')
 338                         buts->name[i] = '_';
 339
 340         ret = -ENOMEM;
 341         bt = kzalloc(sizeof(*bt), GFP_KERNEL);
 342         if (!bt)
 343                 goto err;
 344
 345         bt->sequence = alloc_percpu(unsigned long);
 346         if (!bt->sequence)
 347                 goto err;
 348
 349         ret = -ENOENT;
 350         dir = blk_create_tree(buts->name);
 351         if (!dir)
 352                 goto err;
 353
 354         bt->dir = dir;
 355         bt->dev = bdev->bd_dev;
 356         atomic_set(&bt->dropped, 0);
 357
 358         ret = -EIO;
 359         bt->dropped_file = debugfs_create_file("dropped", 0444, dir, bt, &blk_dropped_fops);
 360         if (!bt->dropped_file)
 361                 goto err;
 362
 363         bt->rchan = relay_open("trace", dir, buts->buf_size,
 364                                 buts->buf_nr, &blk_relay_callbacks, bt);
 365         if (!bt->rchan)
 366                 goto err;
 367
 368         bt->act_mask = buts->act_mask;
 369         if (!bt->act_mask)
 370                 bt->act_mask = (u16) -1;
 371
 372         bt->start_lba = buts->start_lba;
 373         bt->end_lba = buts->end_lba;
 374         if (!bt->end_lba)
 375                 bt->end_lba = -1ULL;
 376
 377         bt->pid = buts->pid;
 378         bt->trace_state = Blktrace_setup;
 379
 380         ret = -EBUSY;
 381         old_bt = xchg(&q->blk_trace, bt);
 382         if (old_bt) {
 383                 (void) xchg(&q->blk_trace, old_bt);
 384                 goto err;
 385         }
 386
 387         return 0;
 388 err:
 389         if (dir)
 390                 blk_remove_tree(dir);
 391         if (bt) {
 392                 if (bt->dropped_file)
 393                         debugfs_remove(bt->dropped_file);
 394                 free_percpu(bt->sequence);
 395                 if (bt->rchan)
 396                         relay_close(bt->rchan);
 397                 kfree(bt);
 398         }
 399         return ret;
 400 }
 401
 402 static int blk_trace_setup(struct request_queue *q, struct block_device *bdev,
 403                            char __user *arg)
 404 {
 405         struct blk_user_trace_setup buts;
 406         int ret;
 407
 408         ret = copy_from_user(&buts, arg, sizeof(buts));
 409         if (ret)
 410                 return -EFAULT;
 411
 412         ret = do_blk_trace_setup(q, bdev, &buts);
 413         if (ret)
 414                 return ret;
 415
 416         if (copy_to_user(arg, &buts, sizeof(buts)))
 417                 return -EFAULT;
 418
 419         return 0;
 420 }
 421
 422 static int blk_trace_startstop(struct request_queue *q, int start)
 423 {
 424         struct blk_trace *bt;
 425         int ret;
 426
 427         if ((bt = q->blk_trace) == NULL)
 428                 return -EINVAL;
 429
 430         /*
 431          * For starting a trace, we can transition from a setup or stopped
 432          * trace. For stopping a trace, the state must be running
 433          */
 434         ret = -EINVAL;
 435         if (start) {
 436                 if (bt->trace_state == Blktrace_setup ||
 437                     bt->trace_state == Blktrace_stopped) {
 438                         blktrace_seq++;
 439                         smp_mb();
 440                         bt->trace_state = Blktrace_running;
 441
 442                         trace_note_time(bt);
 443                         ret = 0;
 444                 }
 445         } else {
 446                 if (bt->trace_state == Blktrace_running) {
 447                         bt->trace_state = Blktrace_stopped;
 448                         relay_flush(bt->rchan);
 449                         ret = 0;
 450                 }
 451         }
 452
 453         return ret;
 454 }
 455
 456 /**
 457  * blk_trace_ioctl: - handle the ioctls associated with tracing
 458  * @bdev:       the block device
 459  * @cmd:        the ioctl cmd
 460  * @arg:        the argument data, if any
 461  *
 462  **/
 463 int blk_trace_ioctl(struct block_device *bdev, unsigned cmd, char __user *arg)
 464 {
 465         struct request_queue *q;
 466         int ret, start = 0;
 467
 468         q = bdev_get_queue(bdev);
 469         if (!q)
 470                 return -ENXIO;
 471
 472         mutex_lock(&bdev->bd_mutex);
 473
 474         switch (cmd) {
 475         case BLKTRACESETUP:
 476                 ret = blk_trace_setup(q, bdev, arg);
 477                 break;
 478         case BLKTRACESTART:
 479                 start = 1;
 480         case BLKTRACESTOP:
 481                 ret = blk_trace_startstop(q, start);
 482                 break;
 483         case BLKTRACETEARDOWN:
 484                 ret = blk_trace_remove(q);
 485                 break;
 486         default:
 487                 ret = -ENOTTY;
 488                 break;
 489         }
 490
 491         mutex_unlock(&bdev->bd_mutex);
 492         return ret;
 493 }
 494
 495 /**
 496  * blk_trace_shutdown: - stop and cleanup trace structures
 497  * @q:    the request queue associated with the device
 498  *
 499  **/
 500 void blk_trace_shutdown(struct request_queue *q)
 501 {
 502         if (q->blk_trace) {
 503                 blk_trace_startstop(q, 0);
 504                 blk_trace_remove(q);
 505         }
 506 }