lustre/kernel_patches/patches/ext3-delete_thread-2.4.18.patch

   1
   2 Create a service thread to handle delete and truncate of inodes, to avoid
   3 long latency while truncating very large files.
   4
   5
   6  fs/ext3/inode.c            |  116 ++++++++++++++++++++++
   7  fs/ext3/super.c            |  231 +++++++++++++++++++++++++++++++++++++++++++++
   8  include/linux/ext3_fs.h    |    5
   9  include/linux/ext3_fs_sb.h |   10 +
  10  4 files changed, 362 insertions(+)
  11
  12 --- linux-2.4.18-18.8.0-l15/fs/ext3/super.c~ext3-delete_thread-2.4.18   Tue Jun  3 17:26:21 2003
  13 +++ linux-2.4.18-18.8.0-l15-adilger/fs/ext3/super.c     Wed Jul  2 23:49:40 2003
  14 @@ -396,6 +396,220 @@ static void dump_orphan_list(struct supe
  15         }
  16  }
  17
  18 +#ifdef EXT3_DELETE_THREAD
  19 +/*
  20 + * Delete inodes in a loop until there are no more to be deleted.
  21 + * Normally, we run in the background doing the deletes and sleeping again,
  22 + * and clients just add new inodes to be deleted onto the end of the list.
  23 + * If someone is concerned about free space (e.g. block allocation or similar)
  24 + * then they can sleep on s_delete_waiter_queue and be woken up when space
  25 + * has been freed.
  26 + */
  27 +int ext3_delete_thread(void *data)
  28 +{
  29 +       struct super_block *sb = data;
  30 +       struct ext3_sb_info *sbi = EXT3_SB(sb);
  31 +       struct task_struct *tsk = current;
  32 +
  33 +       /* Almost like daemonize, but not quite */
  34 +       exit_mm(current);
  35 +       tsk->session = 1;
  36 +       tsk->pgrp = 1;
  37 +       tsk->tty = NULL;
  38 +       exit_files(current);
  39 +       reparent_to_init();
  40 +
  41 +       sprintf(tsk->comm, "kdelext3-%s", kdevname(sb->s_dev));
  42 +       sigfillset(&tsk->blocked);
  43 +
  44 +       /*tsk->flags |= PF_KERNTHREAD;*/
  45 +
  46 +       INIT_LIST_HEAD(&sbi->s_delete_list);
  47 +       wake_up(&sbi->s_delete_waiter_queue);
  48 +       ext3_debug("delete thread on %s started\n", kdevname(sb->s_dev));
  49 +
  50 +       /* main loop */
  51 +       for (;;) {
  52 +               wait_event_interruptible(sbi->s_delete_thread_queue,
  53 +                                        !list_empty(&sbi->s_delete_list) ||
  54 +                                        !test_opt(sb, ASYNCDEL));
  55 +               ext3_debug("%s woken up: %lu inodes, %lu blocks\n",
  56 +                          tsk->comm,sbi->s_delete_inodes,sbi->s_delete_blocks);
  57 +
  58 +               spin_lock(&sbi->s_delete_lock);
  59 +               if (list_empty(&sbi->s_delete_list)) {
  60 +                       clear_opt(sbi->s_mount_opt, ASYNCDEL);
  61 +                       memset(&sbi->s_delete_list, 0,
  62 +                              sizeof(sbi->s_delete_list));
  63 +                       spin_unlock(&sbi->s_delete_lock);
  64 +                       ext3_debug("delete thread on %s exiting\n",
  65 +                                  kdevname(sb->s_dev));
  66 +                       wake_up(&sbi->s_delete_waiter_queue);
  67 +                       break;
  68 +               }
  69 +
  70 +               while (!list_empty(&sbi->s_delete_list)) {
  71 +                       struct inode *inode=list_entry(sbi->s_delete_list.next,
  72 +                                                      struct inode, i_dentry);
  73 +                       unsigned long blocks = inode->i_blocks >>
  74 +                                                       (inode->i_blkbits - 9);
  75 +
  76 +                       list_del_init(&inode->i_dentry);
  77 +                       spin_unlock(&sbi->s_delete_lock);
  78 +                       ext3_debug("%s delete ino %lu blk %lu\n",
  79 +                                  tsk->comm, inode->i_ino, blocks);
  80 +
  81 +                       iput(inode);
  82 +
  83 +                       spin_lock(&sbi->s_delete_lock);
  84 +                       sbi->s_delete_blocks -= blocks;
  85 +                       sbi->s_delete_inodes--;
  86 +               }
  87 +               if (sbi->s_delete_blocks != 0 || sbi->s_delete_inodes != 0) {
  88 +                       ext3_warning(sb, __FUNCTION__,
  89 +                                    "%lu blocks, %lu inodes on list?\n",
  90 +                                    sbi->s_delete_blocks,sbi->s_delete_inodes);
  91 +                       sbi->s_delete_blocks = 0;
  92 +                       sbi->s_delete_inodes = 0;
  93 +               }
  94 +               spin_unlock(&sbi->s_delete_lock);
  95 +               wake_up(&sbi->s_delete_waiter_queue);
  96 +       }
  97 +
  98 +       return 0;
  99 +}
 100 +
 101 +static void ext3_start_delete_thread(struct super_block *sb)
 102 +{
 103 +       struct ext3_sb_info *sbi = EXT3_SB(sb);
 104 +       int rc;
 105 +
 106 +       spin_lock_init(&sbi->s_delete_lock);
 107 +       init_waitqueue_head(&sbi->s_delete_thread_queue);
 108 +       init_waitqueue_head(&sbi->s_delete_waiter_queue);
 109 +
 110 +       if (!test_opt(sb, ASYNCDEL))
 111 +               return;
 112 +
 113 +       rc = kernel_thread(ext3_delete_thread, sb, CLONE_VM | CLONE_FILES);
 114 +       if (rc < 0)
 115 +               printk(KERN_ERR "EXT3-fs: cannot start delete thread: rc %d\n",
 116 +                      rc);
 117 +       else
 118 +               wait_event(sbi->s_delete_waiter_queue, sbi->s_delete_list.next);
 119 +}
 120 +
 121 +static void ext3_stop_delete_thread(struct ext3_sb_info *sbi)
 122 +{
 123 +       if (sbi->s_delete_list.next == 0)       /* thread never started */
 124 +               return;
 125 +
 126 +       clear_opt(sbi->s_mount_opt, ASYNCDEL);
 127 +       wake_up(&sbi->s_delete_thread_queue);
 128 +       wait_event(sbi->s_delete_waiter_queue, list_empty(&sbi->s_delete_list));
 129 +}
 130 +
 131 +/* Instead of playing games with the inode flags, destruction, etc we just
 132 + * create a new inode locally and put it on a list for the truncate thread.
 133 + * We need large parts of the inode struct in order to complete the
 134 + * truncate and unlink, so we may as well just have a real inode to do it.
 135 + *
 136 + * If we have any problem deferring the delete, just delete it right away.
 137 + * If we defer it, we also mark how many blocks it would free, so that we
 138 + * can keep the statfs data correct, and we know if we should sleep on the
 139 + * delete thread when we run out of space.
 140 + */
 141 +static void ext3_delete_inode_thread(struct inode *old_inode)
 142 +{
 143 +       struct ext3_sb_info *sbi = EXT3_SB(old_inode->i_sb);
 144 +       struct ext3_inode_info *nei, *oei = EXT3_I(old_inode);
 145 +       struct inode *new_inode;
 146 +       unsigned long blocks = old_inode->i_blocks >> (old_inode->i_blkbits-9);
 147 +
 148 +       if (is_bad_inode(old_inode)) {
 149 +               clear_inode(old_inode);
 150 +               return;
 151 +       }
 152 +
 153 +       if (!test_opt(old_inode->i_sb, ASYNCDEL) || !sbi->s_delete_list.next)
 154 +               goto out_delete;
 155 +
 156 +       /* We may want to delete the inode immediately and not defer it */
 157 +       if (IS_SYNC(old_inode) || blocks <= EXT3_NDIR_BLOCKS)
 158 +               goto out_delete;
 159 +
 160 +       /* We can't use the delete thread as-is during real orphan recovery,
 161 +        * as we add to the orphan list here, causing ext3_orphan_cleanup()
 162 +        * to loop endlessly.  It would be nice to do so, but needs work.
 163 +        */
 164 +       if (oei->i_state & EXT3_STATE_DELETE ||
 165 +           sbi->s_mount_state & EXT3_ORPHAN_FS) {
 166 +               ext3_debug("doing deferred inode %lu delete (%lu blocks)\n",
 167 +                          old_inode->i_ino, blocks);
 168 +               goto out_delete;
 169 +       }
 170 +
 171 +       /* We can iget this inode again here, because our caller has unhashed
 172 +        * old_inode, so new_inode will be in a different inode struct.
 173 +        *
 174 +        * We need to ensure that the i_orphan pointers in the other inodes
 175 +        * point at the new inode copy instead of the old one so the orphan
 176 +        * list doesn't get corrupted when the old orphan inode is freed.
 177 +        */
 178 +       down(&sbi->s_orphan_lock);
 179 +
 180 +       sbi->s_mount_state |= EXT3_ORPHAN_FS;
 181 +       new_inode = iget(old_inode->i_sb, old_inode->i_ino);
 182 +       sbi->s_mount_state &= ~EXT3_ORPHAN_FS;
 183 +       if (is_bad_inode(new_inode)) {
 184 +               printk(KERN_WARNING "read bad inode %lu\n", old_inode->i_ino);
 185 +               iput(new_inode);
 186 +               new_inode = NULL;
 187 +       }
 188 +       if (!new_inode) {
 189 +               up(&sbi->s_orphan_lock);
 190 +               ext3_debug("delete inode %lu directly (bad read)\n",
 191 +                          old_inode->i_ino);
 192 +               goto out_delete;
 193 +       }
 194 +       J_ASSERT(new_inode != old_inode);
 195 +
 196 +       J_ASSERT(!list_empty(&oei->i_orphan));
 197 +
 198 +       nei = EXT3_I(new_inode);
 199 +       /* Ugh.  We need to insert new_inode into the same spot on the list
 200 +        * as old_inode was, to ensure the in-memory orphan list is still
 201 +        * in the same order as the on-disk orphan list (badness otherwise).
 202 +        */
 203 +       nei->i_orphan = oei->i_orphan;
 204 +       nei->i_orphan.next->prev = &nei->i_orphan;
 205 +       nei->i_orphan.prev->next = &nei->i_orphan;
 206 +       nei->i_state |= EXT3_STATE_DELETE;
 207 +       up(&sbi->s_orphan_lock);
 208 +
 209 +       clear_inode(old_inode);
 210 +
 211 +       spin_lock(&sbi->s_delete_lock);
 212 +       J_ASSERT(list_empty(&new_inode->i_dentry));
 213 +       list_add_tail(&new_inode->i_dentry, &sbi->s_delete_list);
 214 +       sbi->s_delete_blocks += blocks;
 215 +       sbi->s_delete_inodes++;
 216 +       spin_unlock(&sbi->s_delete_lock);
 217 +
 218 +       ext3_debug("delete inode %lu (%lu blocks) by thread\n",
 219 +                  new_inode->i_ino, blocks);
 220 +
 221 +       wake_up(&sbi->s_delete_thread_queue);
 222 +       return;
 223 +
 224 +out_delete:
 225 +       ext3_delete_inode(old_inode);
 226 +}
 227 +#else
 228 +#define ext3_start_delete_thread(sbi) do {} while(0)
 229 +#define ext3_stop_delete_thread(sbi) do {} while(0)
 230 +#endif /* EXT3_DELETE_THREAD */
 231 +
 232  void ext3_put_super (struct super_block * sb)
 233  {
 234         struct ext3_sb_info *sbi = EXT3_SB(sb);
 235 @@ -403,6 +617,7 @@ void ext3_put_super (struct super_block
 236         kdev_t j_dev = sbi->s_journal->j_dev;
 237         int i;
 238
 239 +       ext3_stop_delete_thread(sbi);
 240         ext3_xattr_put_super(sb);
 241         journal_destroy(sbi->s_journal);
 242         if (!(sb->s_flags & MS_RDONLY)) {
 243 @@ -451,7 +666,11 @@ static struct super_operations ext3_sops
 244         write_inode:    ext3_write_inode,       /* BKL not held.  Don't need */
 245         dirty_inode:    ext3_dirty_inode,       /* BKL not held.  We take it */
 246         put_inode:      ext3_put_inode,         /* BKL not held.  Don't need */
 247 +#ifdef EXT3_DELETE_THREAD
 248 +       delete_inode:   ext3_delete_inode_thread,/* BKL not held. We take it */
 249 +#else
 250         delete_inode:   ext3_delete_inode,      /* BKL not held.  We take it */
 251 +#endif
 252         put_super:      ext3_put_super,         /* BKL held */
 253         write_super:    ext3_write_super,       /* BKL held */
 254         write_super_lockfs: ext3_write_super_lockfs, /* BKL not held. Take it */
 255 @@ -511,6 +730,14 @@ static int parse_options (char * options
 256              this_char = strtok (NULL, ",")) {
 257                 if ((value = strchr (this_char, '=')) != NULL)
 258                         *value++ = 0;
 259 +#ifdef EXT3_DELETE_THREAD
 260 +               if (!strcmp(this_char, "asyncdel"))
 261 +                       set_opt(*mount_options, ASYNCDEL);
 262 +               else if (!strcmp(this_char, "noasyncdel"))
 263 +                       clear_opt(*mount_options, ASYNCDEL);
 264 +               else
 265 +#endif
 266 +
 267                 if (!strcmp (this_char, "bsddf"))
 268                         clear_opt (*mount_options, MINIX_DF);
 269                 else if (!strcmp (this_char, "nouid32")) {
 270 @@ -1206,6 +1433,7 @@ struct super_block * ext3_read_super (st
 271         }
 272
 273         ext3_setup_super (sb, es, sb->s_flags & MS_RDONLY);
 274 +       ext3_start_delete_thread(sb);
 275         /*
 276          * akpm: core read_super() calls in here with the superblock locked.
 277          * That deadlocks, because orphan cleanup needs to lock the superblock
 278 @@ -1648,6 +1876,9 @@ int ext3_remount (struct super_block * s
 279         if (!parse_options(data, &tmp, sbi, &tmp, 1))
 280                 return -EINVAL;
 281
 282 +       if (!test_opt(sb, ASYNCDEL) || (*flags & MS_RDONLY))
 283 +               ext3_stop_delete_thread(sbi);
 284 +
 285         if (sbi->s_mount_opt & EXT3_MOUNT_ABORT)
 286                 ext3_abort(sb, __FUNCTION__, "Abort forced by user");
 287
 288 --- linux/fs/ext3/file.c.orig   Fri Jan 17 10:57:31 2003
 289 +++ linux/fs/ext3/file.c        Mon Jun 30 13:28:52 2003
 290 @@ -121,7 +121,11 @@ struct file_operations ext3_file_operati
 291  };
 292
 293  struct inode_operations ext3_file_inode_operations = {
 294 +#ifdef EXT3_DELETE_THREAD
 295 +       truncate:       ext3_truncate_thread,   /* BKL held */
 296 +#else
 297         truncate:       ext3_truncate,          /* BKL held */
 298 +#endif
 299         setattr:        ext3_setattr,           /* BKL held */
 300  };
 301
 302 --- linux-2.4.18-18.8.0-l15/fs/ext3/inode.c~ext3-delete_thread-2.4.18   Wed Jul  2 23:13:58 2003
 303 +++ linux-2.4.18-18.8.0-l15-adilger/fs/ext3/inode.c     Wed Jul  2 23:50:29 2003
 304 @@ -2004,6 +2004,118 @@ out_stop:
 305         ext3_journal_stop(handle, inode);
 306  }
 307
 308 +#ifdef EXT3_DELETE_THREAD
 309 +/* Move blocks from to-be-truncated inode over to a new inode, and delete
 310 + * that one from the delete thread instead.  This avoids a lot of latency
 311 + * when truncating large files.
 312 + *
 313 + * If we have any problem deferring the truncate, just truncate it right away.
 314 + * If we defer it, we also mark how many blocks it would free, so that we
 315 + * can keep the statfs data correct, and we know if we should sleep on the
 316 + * delete thread when we run out of space.
 317 + */
 318 +void ext3_truncate_thread(struct inode *old_inode)
 319 +{
 320 +       struct ext3_sb_info *sbi = EXT3_SB(old_inode->i_sb);
 321 +       struct ext3_inode_info *nei, *oei = EXT3_I(old_inode);
 322 +       struct inode *new_inode;
 323 +       handle_t *handle;
 324 +       unsigned long blocks = old_inode->i_blocks >> (old_inode->i_blkbits-9);
 325 +
 326 +       if (!test_opt(old_inode->i_sb, ASYNCDEL) || !sbi->s_delete_list.next)
 327 +               goto out_truncate;
 328 +
 329 +       /* XXX This is a temporary limitation for code simplicity.
 330 +        *     We could truncate to arbitrary sizes at some later time.
 331 +        */
 332 +       if (old_inode->i_size != 0)
 333 +               goto out_truncate;
 334 +
 335 +       /* We may want to truncate the inode immediately and not defer it */
 336 +       if (IS_SYNC(old_inode) || blocks <= EXT3_NDIR_BLOCKS ||
 337 +           old_inode->i_size > oei->i_disksize)
 338 +               goto out_truncate;
 339 +
 340 +       /* We can't use the delete thread as-is during real orphan recovery,
 341 +        * as we add to the orphan list here, causing ext3_orphan_cleanup()
 342 +        * to loop endlessly.  It would be nice to do so, but needs work.
 343 +        */
 344 +       if (oei->i_state & EXT3_STATE_DELETE ||
 345 +           sbi->s_mount_state & EXT3_ORPHAN_FS) {
 346 +               ext3_debug("doing deferred inode %lu delete (%lu blocks)\n",
 347 +                          old_inode->i_ino, blocks);
 348 +               goto out_truncate;
 349 +       }
 350 +
 351 +       ext3_discard_prealloc(old_inode);
 352 +
 353 +       /* old_inode   = 1
 354 +        * new_inode   = sb + GDT + ibitmap
 355 +        * orphan list = 1 inode/superblock for add, 2 inodes for del
 356 +        * quota files = 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
 357 +        */
 358 +       handle = ext3_journal_start(old_inode, 7);
 359 +       if (IS_ERR(handle))
 360 +               goto out_truncate;
 361 +
 362 +       new_inode = ext3_new_inode(handle, old_inode, old_inode->i_mode);
 363 +       if (IS_ERR(new_inode)) {
 364 +               ext3_debug("truncate inode %lu directly (no new inodes)\n",
 365 +                          old_inode->i_ino);
 366 +               goto out_journal;
 367 +       }
 368 +
 369 +       nei = EXT3_I(new_inode);
 370 +
 371 +       down_write(&oei->truncate_sem);
 372 +       new_inode->i_size = old_inode->i_size;
 373 +       new_inode->i_blocks = old_inode->i_blocks;
 374 +       new_inode->i_uid = old_inode->i_uid;
 375 +       new_inode->i_gid = old_inode->i_gid;
 376 +       new_inode->i_nlink = 0;
 377 +
 378 +       /* FIXME when we do arbitrary truncates */
 379 +       old_inode->i_blocks = oei->i_file_acl ? old_inode->i_blksize / 512 : 0;
 380 +       old_inode->i_mtime = old_inode->i_ctime = CURRENT_TIME;
 381 +
 382 +       memcpy(nei->i_data, oei->i_data, sizeof(nei->i_data));
 383 +       memset(oei->i_data, 0, sizeof(oei->i_data));
 384 +
 385 +       nei->i_disksize = oei->i_disksize;
 386 +       nei->i_state |= EXT3_STATE_DELETE;
 387 +       up_write(&oei->truncate_sem);
 388 +
 389 +       if (ext3_orphan_add(handle, new_inode) < 0)
 390 +               goto out_journal;
 391 +
 392 +       if (ext3_orphan_del(handle, old_inode) < 0) {
 393 +               ext3_orphan_del(handle, new_inode);
 394 +               iput(new_inode);
 395 +               goto out_journal;
 396 +       }
 397 +
 398 +       ext3_journal_stop(handle, old_inode);
 399 +
 400 +       spin_lock(&sbi->s_delete_lock);
 401 +       J_ASSERT(list_empty(&new_inode->i_dentry));
 402 +       list_add_tail(&new_inode->i_dentry, &sbi->s_delete_list);
 403 +       sbi->s_delete_blocks += blocks;
 404 +       sbi->s_delete_inodes++;
 405 +       spin_unlock(&sbi->s_delete_lock);
 406 +
 407 +       ext3_debug("delete inode %lu (%lu blocks) by thread\n",
 408 +                  new_inode->i_ino, blocks);
 409 +
 410 +       wake_up(&sbi->s_delete_thread_queue);
 411 +       return;
 412 +
 413 +out_journal:
 414 +       ext3_journal_stop(handle, old_inode);
 415 +out_truncate:
 416 +       ext3_truncate(old_inode);
 417 +}
 418 +#endif /* EXT3_DELETE_THREAD */
 419 +
 420  /*
 421   * ext3_get_inode_loc returns with an extra refcount against the
 422   * inode's underlying buffer_head on success.
 423 --- linux-2.4.18-18.8.0-l15/include/linux/ext3_fs.h~ext3-delete_thread-2.4.18   Tue Jun  3 17:26:20 2003
 424 +++ linux-2.4.18-18.8.0-l15-adilger/include/linux/ext3_fs.h     Wed Jul  2 23:19:09 2003
 425 @@ -190,6 +190,7 @@ struct ext3_group_desc
 426   */
 427  #define EXT3_STATE_JDATA               0x00000001 /* journaled data exists */
 428  #define EXT3_STATE_NEW                 0x00000002 /* inode is newly created */
 429 +#define EXT3_STATE_DELETE              0x00000010 /* deferred delete inode */
 430
 431  /*
 432   * ioctl commands
 433 @@ -317,6 +318,7 @@ struct ext3_inode {
 434  #define EXT3_MOUNT_UPDATE_JOURNAL      0x1000  /* Update the journal format */
 435  #define EXT3_MOUNT_NO_UID32            0x2000  /* Disable 32-bit UIDs */
 436  #define EXT3_MOUNT_INDEX               0x4000  /* Enable directory index */
 437 +#define EXT3_MOUNT_ASYNCDEL            0x20000 /* Delayed deletion */
 438
 439  /* Compatibility, for having both ext2_fs.h and ext3_fs.h included at once */
 440  #ifndef _LINUX_EXT2_FS_H
 441 @@ -651,6 +653,9 @@ extern void ext3_discard_prealloc (struc
 442  extern void ext3_dirty_inode(struct inode *);
 443  extern int ext3_change_inode_journal_flag(struct inode *, int);
 444  extern void ext3_truncate (struct inode *);
 445 +#ifdef EXT3_DELETE_THREAD
 446 +extern void ext3_truncate_thread(struct inode *inode);
 447 +#endif
 448
 449  /* ioctl.c */
 450  extern int ext3_ioctl (struct inode *, struct file *, unsigned int,
 451 --- linux-2.4.18-18.8.0-l15/include/linux/ext3_fs_sb.h~ext3-delete_thread-2.4.18        Tue Jun  3 17:26:21 2003
 452 +++ linux-2.4.18-18.8.0-l15-adilger/include/linux/ext3_fs_sb.h  Wed Jul  2 23:19:09 2003
 453 @@ -29,6 +29,8 @@
 454
 455  #define EXT3_MAX_GROUP_LOADED  32
 456
 457 +#define EXT3_DELETE_THREAD
 458 +
 459  /*
 460   * third extended-fs super-block data in memory
 461   */
 462 @@ -74,6 +76,14 @@ struct ext3_sb_info {
 463         struct timer_list turn_ro_timer;        /* For turning read-only (crash simulation) */
 464         wait_queue_head_t ro_wait_queue;        /* For people waiting for the fs to go read-only */
 465  #endif
 466 +#ifdef EXT3_DELETE_THREAD
 467 +       spinlock_t s_delete_lock;
 468 +       struct list_head s_delete_list;
 469 +       unsigned long s_delete_blocks;
 470 +       unsigned long s_delete_inodes;
 471 +       wait_queue_head_t s_delete_thread_queue;
 472 +       wait_queue_head_t s_delete_waiter_queue;
 473 +#endif
 474  };
 475
 476  #endif /* _LINUX_EXT3_FS_SB */
 477
 478 _