lustre/kernel_patches/patches/ext3-delete_thread-2.4.20.patch

   1  fs/ext3/file.c             |    4
   2  fs/ext3/inode.c            |  116 ++++++++++++++++++++++
   3  fs/ext3/super.c            |  230 +++++++++++++++++++++++++++++++++++++++++++++
   4  include/linux/ext3_fs.h    |    5
   5  include/linux/ext3_fs_sb.h |   10 +
   6  5 files changed, 365 insertions(+)
   7
   8 --- linux/fs/ext3/super.c~ext3-delete_thread-2.4.20     Thu Jul 10 14:11:32 2003
   9 +++ linux-mmonroe/fs/ext3/super.c       Thu Jul 10 14:11:33 2003
  10 @@ -400,6 +400,220 @@ static void dump_orphan_list(struct supe
  11         }
  12  }
  13
  14 +#ifdef EXT3_DELETE_THREAD
  15 +/*
  16 + * Delete inodes in a loop until there are no more to be deleted.
  17 + * Normally, we run in the background doing the deletes and sleeping again,
  18 + * and clients just add new inodes to be deleted onto the end of the list.
  19 + * If someone is concerned about free space (e.g. block allocation or similar)
  20 + * then they can sleep on s_delete_waiter_queue and be woken up when space
  21 + * has been freed.
  22 + */
  23 +int ext3_delete_thread(void *data)
  24 +{
  25 +       struct super_block *sb = data;
  26 +       struct ext3_sb_info *sbi = EXT3_SB(sb);
  27 +       struct task_struct *tsk = current;
  28 +
  29 +       /* Almost like daemonize, but not quite */
  30 +       exit_mm(current);
  31 +       tsk->session = 1;
  32 +       tsk->pgrp = 1;
  33 +       tsk->tty = NULL;
  34 +       exit_files(current);
  35 +       reparent_to_init();
  36 +
  37 +       sprintf(tsk->comm, "kdelext3-%s", kdevname(sb->s_dev));
  38 +       sigfillset(&tsk->blocked);
  39 +
  40 +       /*tsk->flags |= PF_KERNTHREAD;*/
  41 +
  42 +       INIT_LIST_HEAD(&sbi->s_delete_list);
  43 +       wake_up(&sbi->s_delete_waiter_queue);
  44 +       ext3_debug("delete thread on %s started\n", kdevname(sb->s_dev));
  45 +
  46 +       /* main loop */
  47 +       for (;;) {
  48 +               wait_event_interruptible(sbi->s_delete_thread_queue,
  49 +                                        !list_empty(&sbi->s_delete_list) ||
  50 +                                        !test_opt(sb, ASYNCDEL));
  51 +               ext3_debug("%s woken up: %lu inodes, %lu blocks\n",
  52 +                          tsk->comm,sbi->s_delete_inodes,sbi->s_delete_blocks);
  53 +
  54 +               spin_lock(&sbi->s_delete_lock);
  55 +               if (list_empty(&sbi->s_delete_list)) {
  56 +                       clear_opt(sbi->s_mount_opt, ASYNCDEL);
  57 +                       memset(&sbi->s_delete_list, 0,
  58 +                              sizeof(sbi->s_delete_list));
  59 +                       spin_unlock(&sbi->s_delete_lock);
  60 +                       ext3_debug("delete thread on %s exiting\n",
  61 +                                  kdevname(sb->s_dev));
  62 +                       wake_up(&sbi->s_delete_waiter_queue);
  63 +                       break;
  64 +               }
  65 +
  66 +               while (!list_empty(&sbi->s_delete_list)) {
  67 +                       struct inode *inode=list_entry(sbi->s_delete_list.next,
  68 +                                                      struct inode, i_dentry);
  69 +                       unsigned long blocks = inode->i_blocks >>
  70 +                                                       (inode->i_blkbits - 9);
  71 +
  72 +                       list_del_init(&inode->i_dentry);
  73 +                       spin_unlock(&sbi->s_delete_lock);
  74 +                       ext3_debug("%s delete ino %lu blk %lu\n",
  75 +                                  tsk->comm, inode->i_ino, blocks);
  76 +
  77 +                       iput(inode);
  78 +
  79 +                       spin_lock(&sbi->s_delete_lock);
  80 +                       sbi->s_delete_blocks -= blocks;
  81 +                       sbi->s_delete_inodes--;
  82 +               }
  83 +               if (sbi->s_delete_blocks != 0 || sbi->s_delete_inodes != 0) {
  84 +                       ext3_warning(sb, __FUNCTION__,
  85 +                                    "%lu blocks, %lu inodes on list?\n",
  86 +                                    sbi->s_delete_blocks,sbi->s_delete_inodes);
  87 +                       sbi->s_delete_blocks = 0;
  88 +                       sbi->s_delete_inodes = 0;
  89 +               }
  90 +               spin_unlock(&sbi->s_delete_lock);
  91 +               wake_up(&sbi->s_delete_waiter_queue);
  92 +       }
  93 +
  94 +       return 0;
  95 +}
  96 +
  97 +static void ext3_start_delete_thread(struct super_block *sb)
  98 +{
  99 +       struct ext3_sb_info *sbi = EXT3_SB(sb);
 100 +       int rc;
 101 +
 102 +       spin_lock_init(&sbi->s_delete_lock);
 103 +       init_waitqueue_head(&sbi->s_delete_thread_queue);
 104 +       init_waitqueue_head(&sbi->s_delete_waiter_queue);
 105 +
 106 +       if (!test_opt(sb, ASYNCDEL))
 107 +               return;
 108 +
 109 +       rc = kernel_thread(ext3_delete_thread, sb, CLONE_VM | CLONE_FILES);
 110 +       if (rc < 0)
 111 +               printk(KERN_ERR "EXT3-fs: cannot start delete thread: rc %d\n",
 112 +                      rc);
 113 +       else
 114 +               wait_event(sbi->s_delete_waiter_queue, sbi->s_delete_list.next);
 115 +}
 116 +
 117 +static void ext3_stop_delete_thread(struct ext3_sb_info *sbi)
 118 +{
 119 +       if (sbi->s_delete_list.next == 0)       /* thread never started */
 120 +               return;
 121 +
 122 +       clear_opt(sbi->s_mount_opt, ASYNCDEL);
 123 +       wake_up(&sbi->s_delete_thread_queue);
 124 +       wait_event(sbi->s_delete_waiter_queue, list_empty(&sbi->s_delete_list));
 125 +}
 126 +
 127 +/* Instead of playing games with the inode flags, destruction, etc we just
 128 + * create a new inode locally and put it on a list for the truncate thread.
 129 + * We need large parts of the inode struct in order to complete the
 130 + * truncate and unlink, so we may as well just have a real inode to do it.
 131 + *
 132 + * If we have any problem deferring the delete, just delete it right away.
 133 + * If we defer it, we also mark how many blocks it would free, so that we
 134 + * can keep the statfs data correct, and we know if we should sleep on the
 135 + * delete thread when we run out of space.
 136 + */
 137 +static void ext3_delete_inode_thread(struct inode *old_inode)
 138 +{
 139 +       struct ext3_sb_info *sbi = EXT3_SB(old_inode->i_sb);
 140 +       struct ext3_inode_info *nei, *oei = EXT3_I(old_inode);
 141 +       struct inode *new_inode;
 142 +       unsigned long blocks = old_inode->i_blocks >> (old_inode->i_blkbits-9);
 143 +
 144 +       if (is_bad_inode(old_inode)) {
 145 +               clear_inode(old_inode);
 146 +               return;
 147 +       }
 148 +
 149 +       if (!test_opt(old_inode->i_sb, ASYNCDEL) || !sbi->s_delete_list.next)
 150 +               goto out_delete;
 151 +
 152 +       /* We may want to delete the inode immediately and not defer it */
 153 +       if (IS_SYNC(old_inode) || blocks <= EXT3_NDIR_BLOCKS)
 154 +               goto out_delete;
 155 +
 156 +       /* We can't use the delete thread as-is during real orphan recovery,
 157 +        * as we add to the orphan list here, causing ext3_orphan_cleanup()
 158 +        * to loop endlessly.  It would be nice to do so, but needs work.
 159 +        */
 160 +       if (oei->i_state & EXT3_STATE_DELETE ||
 161 +           sbi->s_mount_state & EXT3_ORPHAN_FS) {
 162 +               ext3_debug("doing deferred inode %lu delete (%lu blocks)\n",
 163 +                          old_inode->i_ino, blocks);
 164 +               goto out_delete;
 165 +       }
 166 +
 167 +       /* We can iget this inode again here, because our caller has unhashed
 168 +        * old_inode, so new_inode will be in a different inode struct.
 169 +        *
 170 +        * We need to ensure that the i_orphan pointers in the other inodes
 171 +        * point at the new inode copy instead of the old one so the orphan
 172 +        * list doesn't get corrupted when the old orphan inode is freed.
 173 +        */
 174 +       down(&sbi->s_orphan_lock);
 175 +
 176 +       sbi->s_mount_state |= EXT3_ORPHAN_FS;
 177 +       new_inode = iget(old_inode->i_sb, old_inode->i_ino);
 178 +       sbi->s_mount_state &= ~EXT3_ORPHAN_FS;
 179 +       if (is_bad_inode(new_inode)) {
 180 +               printk(KERN_WARNING "read bad inode %lu\n", old_inode->i_ino);
 181 +               iput(new_inode);
 182 +               new_inode = NULL;
 183 +       }
 184 +       if (!new_inode) {
 185 +               up(&sbi->s_orphan_lock);
 186 +               ext3_debug("delete inode %lu directly (bad read)\n",
 187 +                          old_inode->i_ino);
 188 +               goto out_delete;
 189 +       }
 190 +       J_ASSERT(new_inode != old_inode);
 191 +
 192 +       J_ASSERT(!list_empty(&oei->i_orphan));
 193 +
 194 +       nei = EXT3_I(new_inode);
 195 +       /* Ugh.  We need to insert new_inode into the same spot on the list
 196 +        * as old_inode was, to ensure the in-memory orphan list is still
 197 +        * in the same order as the on-disk orphan list (badness otherwise).
 198 +        */
 199 +       nei->i_orphan = oei->i_orphan;
 200 +       nei->i_orphan.next->prev = &nei->i_orphan;
 201 +       nei->i_orphan.prev->next = &nei->i_orphan;
 202 +       nei->i_state |= EXT3_STATE_DELETE;
 203 +       up(&sbi->s_orphan_lock);
 204 +
 205 +       clear_inode(old_inode);
 206 +
 207 +       spin_lock(&sbi->s_delete_lock);
 208 +       J_ASSERT(list_empty(&new_inode->i_dentry));
 209 +       list_add_tail(&new_inode->i_dentry, &sbi->s_delete_list);
 210 +       sbi->s_delete_blocks += blocks;
 211 +       sbi->s_delete_inodes++;
 212 +       spin_unlock(&sbi->s_delete_lock);
 213 +
 214 +       ext3_debug("delete inode %lu (%lu blocks) by thread\n",
 215 +                  new_inode->i_ino, blocks);
 216 +
 217 +       wake_up(&sbi->s_delete_thread_queue);
 218 +       return;
 219 +
 220 +out_delete:
 221 +       ext3_delete_inode(old_inode);
 222 +}
 223 +#else
 224 +#define ext3_start_delete_thread(sbi) do {} while(0)
 225 +#define ext3_stop_delete_thread(sbi) do {} while(0)
 226 +#endif /* EXT3_DELETE_THREAD */
 227 +
 228  void ext3_put_super (struct super_block * sb)
 229  {
 230         struct ext3_sb_info *sbi = EXT3_SB(sb);
 231 @@ -407,6 +621,7 @@ void ext3_put_super (struct super_block
 232         kdev_t j_dev = sbi->s_journal->j_dev;
 233         int i;
 234
 235 +       ext3_stop_delete_thread(sbi);
 236         ext3_xattr_put_super(sb);
 237         journal_destroy(sbi->s_journal);
 238         if (!(sb->s_flags & MS_RDONLY)) {
 239 @@ -455,7 +670,11 @@ static struct super_operations ext3_sops
 240         write_inode:    ext3_write_inode,       /* BKL not held.  Don't need */
 241         dirty_inode:    ext3_dirty_inode,       /* BKL not held.  We take it */
 242         put_inode:      ext3_put_inode,         /* BKL not held.  Don't need */
 243 +#ifdef EXT3_DELETE_THREAD
 244 +       delete_inode:   ext3_delete_inode_thread,/* BKL not held. We take it */
 245 +#else
 246         delete_inode:   ext3_delete_inode,      /* BKL not held.  We take it */
 247 +#endif
 248         put_super:      ext3_put_super,         /* BKL held */
 249         write_super:    ext3_write_super,       /* BKL held */
 250         sync_fs:        ext3_sync_fs,
 251 @@ -524,6 +743,13 @@ static int parse_options (char * options
 252                         clear_opt (*mount_options, XATTR_USER);
 253                 else
 254  #endif
 255 +#ifdef EXT3_DELETE_THREAD
 256 +               if (!strcmp(this_char, "asyncdel"))
 257 +                       set_opt(*mount_options, ASYNCDEL);
 258 +               else if (!strcmp(this_char, "noasyncdel"))
 259 +                       clear_opt(*mount_options, ASYNCDEL);
 260 +               else
 261 +#endif
 262                 if (!strcmp (this_char, "bsddf"))
 263                         clear_opt (*mount_options, MINIX_DF);
 264                 else if (!strcmp (this_char, "nouid32")) {
 265 @@ -1223,6 +1449,7 @@ struct super_block * ext3_read_super (st
 266         }
 267
 268         ext3_setup_super (sb, es, sb->s_flags & MS_RDONLY);
 269 +       ext3_start_delete_thread(sb);
 270         /*
 271          * akpm: core read_super() calls in here with the superblock locked.
 272          * That deadlocks, because orphan cleanup needs to lock the superblock
 273 @@ -1678,6 +1905,9 @@ int ext3_remount (struct super_block * s
 274         if (!parse_options(data, &tmp, sbi, &tmp, 1))
 275                 return -EINVAL;
 276
 277 +       if (!test_opt(sb, ASYNCDEL) || (*flags & MS_RDONLY))
 278 +               ext3_stop_delete_thread(sbi);
 279 +
 280         if (sbi->s_mount_opt & EXT3_MOUNT_ABORT)
 281                 ext3_abort(sb, __FUNCTION__, "Abort forced by user");
 282
 283 --- linux/fs/ext3/inode.c~ext3-delete_thread-2.4.20     Thu Jul 10 14:11:29 2003
 284 +++ linux-mmonroe/fs/ext3/inode.c       Thu Jul 10 14:11:33 2003
 285 @@ -2013,6 +2013,118 @@ out_stop:
 286         ext3_journal_stop(handle, inode);
 287  }
 288
 289 +#ifdef EXT3_DELETE_THREAD
 290 +/* Move blocks from to-be-truncated inode over to a new inode, and delete
 291 + * that one from the delete thread instead.  This avoids a lot of latency
 292 + * when truncating large files.
 293 + *
 294 + * If we have any problem deferring the truncate, just truncate it right away.
 295 + * If we defer it, we also mark how many blocks it would free, so that we
 296 + * can keep the statfs data correct, and we know if we should sleep on the
 297 + * delete thread when we run out of space.
 298 + */
 299 +void ext3_truncate_thread(struct inode *old_inode)
 300 +{
 301 +       struct ext3_sb_info *sbi = EXT3_SB(old_inode->i_sb);
 302 +       struct ext3_inode_info *nei, *oei = EXT3_I(old_inode);
 303 +       struct inode *new_inode;
 304 +       handle_t *handle;
 305 +       unsigned long blocks = old_inode->i_blocks >> (old_inode->i_blkbits-9);
 306 +
 307 +       if (!test_opt(old_inode->i_sb, ASYNCDEL) || !sbi->s_delete_list.next)
 308 +               goto out_truncate;
 309 +
 310 +       /* XXX This is a temporary limitation for code simplicity.
 311 +        *     We could truncate to arbitrary sizes at some later time.
 312 +        */
 313 +       if (old_inode->i_size != 0)
 314 +               goto out_truncate;
 315 +
 316 +       /* We may want to truncate the inode immediately and not defer it */
 317 +       if (IS_SYNC(old_inode) || blocks <= EXT3_NDIR_BLOCKS ||
 318 +           old_inode->i_size > oei->i_disksize)
 319 +               goto out_truncate;
 320 +
 321 +       /* We can't use the delete thread as-is during real orphan recovery,
 322 +        * as we add to the orphan list here, causing ext3_orphan_cleanup()
 323 +        * to loop endlessly.  It would be nice to do so, but needs work.
 324 +        */
 325 +       if (oei->i_state & EXT3_STATE_DELETE ||
 326 +           sbi->s_mount_state & EXT3_ORPHAN_FS) {
 327 +               ext3_debug("doing deferred inode %lu delete (%lu blocks)\n",
 328 +                          old_inode->i_ino, blocks);
 329 +               goto out_truncate;
 330 +       }
 331 +
 332 +       ext3_discard_prealloc(old_inode);
 333 +
 334 +       /* old_inode   = 1
 335 +        * new_inode   = sb + GDT + ibitmap
 336 +        * orphan list = 1 inode/superblock for add, 2 inodes for del
 337 +        * quota files = 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
 338 +        */
 339 +       handle = ext3_journal_start(old_inode, 7);
 340 +       if (IS_ERR(handle))
 341 +               goto out_truncate;
 342 +
 343 +       new_inode = ext3_new_inode(handle, old_inode, old_inode->i_mode);
 344 +       if (IS_ERR(new_inode)) {
 345 +               ext3_debug("truncate inode %lu directly (no new inodes)\n",
 346 +                          old_inode->i_ino);
 347 +               goto out_journal;
 348 +       }
 349 +
 350 +       nei = EXT3_I(new_inode);
 351 +
 352 +       down_write(&oei->truncate_sem);
 353 +       new_inode->i_size = old_inode->i_size;
 354 +       new_inode->i_blocks = old_inode->i_blocks;
 355 +       new_inode->i_uid = old_inode->i_uid;
 356 +       new_inode->i_gid = old_inode->i_gid;
 357 +       new_inode->i_nlink = 0;
 358 +
 359 +       /* FIXME when we do arbitrary truncates */
 360 +       old_inode->i_blocks = oei->i_file_acl ? old_inode->i_blksize / 512 : 0;
 361 +       old_inode->i_mtime = old_inode->i_ctime = CURRENT_TIME;
 362 +
 363 +       memcpy(nei->i_data, oei->i_data, sizeof(nei->i_data));
 364 +       memset(oei->i_data, 0, sizeof(oei->i_data));
 365 +
 366 +       nei->i_disksize = oei->i_disksize;
 367 +       nei->i_state |= EXT3_STATE_DELETE;
 368 +       up_write(&oei->truncate_sem);
 369 +
 370 +       if (ext3_orphan_add(handle, new_inode) < 0)
 371 +               goto out_journal;
 372 +
 373 +       if (ext3_orphan_del(handle, old_inode) < 0) {
 374 +               ext3_orphan_del(handle, new_inode);
 375 +               iput(new_inode);
 376 +               goto out_journal;
 377 +       }
 378 +
 379 +       ext3_journal_stop(handle, old_inode);
 380 +
 381 +       spin_lock(&sbi->s_delete_lock);
 382 +       J_ASSERT(list_empty(&new_inode->i_dentry));
 383 +       list_add_tail(&new_inode->i_dentry, &sbi->s_delete_list);
 384 +       sbi->s_delete_blocks += blocks;
 385 +       sbi->s_delete_inodes++;
 386 +       spin_unlock(&sbi->s_delete_lock);
 387 +
 388 +       ext3_debug("delete inode %lu (%lu blocks) by thread\n",
 389 +                  new_inode->i_ino, blocks);
 390 +
 391 +       wake_up(&sbi->s_delete_thread_queue);
 392 +       return;
 393 +
 394 +out_journal:
 395 +       ext3_journal_stop(handle, old_inode);
 396 +out_truncate:
 397 +       ext3_truncate(old_inode);
 398 +}
 399 +#endif /* EXT3_DELETE_THREAD */
 400 +
 401  /*
 402   * ext3_get_inode_loc returns with an extra refcount against the
 403   * inode's underlying buffer_head on success.
 404 --- linux/fs/ext3/file.c~ext3-delete_thread-2.4.20      Thu Jul 10 14:11:21 2003
 405 +++ linux-mmonroe/fs/ext3/file.c        Thu Jul 10 14:12:17 2003
 406 @@ -125,7 +125,11 @@ struct file_operations ext3_file_operati
 407  };
 408
 409  struct inode_operations ext3_file_inode_operations = {
 410 +#ifdef EXT3_DELETE_THREAD
 411 +       truncate:       ext3_truncate_thread,   /* BKL held */
 412 +#else
 413         truncate:       ext3_truncate,          /* BKL held */
 414 +#endif
 415         setattr:        ext3_setattr,           /* BKL held */
 416         setxattr:       ext3_setxattr,          /* BKL held */
 417         getxattr:       ext3_getxattr,          /* BKL held */
 418 --- linux/include/linux/ext3_fs.h~ext3-delete_thread-2.4.20     Thu Jul 10 14:11:26 2003
 419 +++ linux-mmonroe/include/linux/ext3_fs.h       Thu Jul 10 14:11:33 2003
 420 @@ -193,6 +193,7 @@ struct ext3_group_desc
 421   */
 422  #define EXT3_STATE_JDATA               0x00000001 /* journaled data exists */
 423  #define EXT3_STATE_NEW                 0x00000002 /* inode is newly created */
 424 +#define EXT3_STATE_DELETE              0x00000010 /* deferred delete inode */
 425
 426  /*
 427   * ioctl commands
 428 @@ -320,6 +321,7 @@ struct ext3_inode {
 429  #define EXT3_MOUNT_UPDATE_JOURNAL      0x1000  /* Update the journal format */
 430  #define EXT3_MOUNT_NO_UID32            0x2000  /* Disable 32-bit UIDs */
 431  #define EXT3_MOUNT_XATTR_USER          0x4000  /* Extended user attributes */
 432 +#define EXT3_MOUNT_ASYNCDEL            0x20000 /* Delayed deletion */
 433
 434  /* Compatibility, for having both ext2_fs.h and ext3_fs.h included at once */
 435  #ifndef _LINUX_EXT2_FS_H
 436 @@ -694,6 +696,9 @@ extern void ext3_discard_prealloc (struc
 437  extern void ext3_dirty_inode(struct inode *);
 438  extern int ext3_change_inode_journal_flag(struct inode *, int);
 439  extern void ext3_truncate (struct inode *);
 440 +#ifdef EXT3_DELETE_THREAD
 441 +extern void ext3_truncate_thread(struct inode *inode);
 442 +#endif
 443
 444  /* ioctl.c */
 445  extern int ext3_ioctl (struct inode *, struct file *, unsigned int,
 446 --- linux/include/linux/ext3_fs_sb.h~ext3-delete_thread-2.4.20  Thu Jul 10 14:11:32 2003
 447 +++ linux-mmonroe/include/linux/ext3_fs_sb.h    Thu Jul 10 14:11:33 2003
 448 @@ -29,6 +29,8 @@
 449
 450  #define EXT3_MAX_GROUP_LOADED  8
 451
 452 +#define EXT3_DELETE_THREAD
 453 +
 454  /*
 455   * third extended-fs super-block data in memory
 456   */
 457 @@ -76,6 +78,14 @@ struct ext3_sb_info {
 458         struct timer_list turn_ro_timer;        /* For turning read-only (crash simulation) */
 459         wait_queue_head_t ro_wait_queue;        /* For people waiting for the fs to go read-only */
 460  #endif
 461 +#ifdef EXT3_DELETE_THREAD
 462 +       spinlock_t s_delete_lock;
 463 +       struct list_head s_delete_list;
 464 +       unsigned long s_delete_blocks;
 465 +       unsigned long s_delete_inodes;
 466 +       wait_queue_head_t s_delete_thread_queue;
 467 +       wait_queue_head_t s_delete_waiter_queue;
 468 +#endif
 469  };
 470
 471  #endif /* _LINUX_EXT3_FS_SB */
 472
 473 _