block/blktrace.c

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