lustre/kernel_patches/patches/ext3-delete_thread-2.4.20-hp.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 Index: linux/fs/ext3/super.c
   9 ===================================================================
  10 --- linux.orig/fs/ext3/super.c  Mon Feb  2 20:57:35 2004
  11 +++ linux/fs/ext3/super.c       Mon Feb  2 20:58:05 2004
  12 @@ -400,6 +400,221 @@
  13         }
  14  }
  15
  16 +#ifdef EXT3_DELETE_THREAD
  17 +/*
  18 + * Delete inodes in a loop until there are no more to be deleted.
  19 + * Normally, we run in the background doing the deletes and sleeping again,
  20 + * and clients just add new inodes to be deleted onto the end of the list.
  21 + * If someone is concerned about free space (e.g. block allocation or similar)
  22 + * then they can sleep on s_delete_waiter_queue and be woken up when space
  23 + * has been freed.
  24 + */
  25 +int ext3_delete_thread(void *data)
  26 +{
  27 +       struct super_block *sb = data;
  28 +       struct ext3_sb_info *sbi = EXT3_SB(sb);
  29 +       struct task_struct *tsk = current;
  30 +
  31 +       /* Almost like daemonize, but not quite */
  32 +       exit_mm(current);
  33 +       tsk->session = 1;
  34 +       tsk->pgrp = 1;
  35 +       tsk->tty = NULL;
  36 +       exit_files(current);
  37 +       reparent_to_init();
  38 +
  39 +       sprintf(tsk->comm, "kdelext3-%s", kdevname(sb->s_dev));
  40 +       sigfillset(&tsk->blocked);
  41 +
  42 +       /*tsk->flags |= PF_KERNTHREAD;*/
  43 +
  44 +       INIT_LIST_HEAD(&sbi->s_delete_list);
  45 +       wake_up(&sbi->s_delete_waiter_queue);
  46 +       ext3_debug("delete thread on %s started\n", kdevname(sb->s_dev));
  47 +
  48 +       /* main loop */
  49 +       for (;;) {
  50 +               wait_event_interruptible(sbi->s_delete_thread_queue,
  51 +                                        !list_empty(&sbi->s_delete_list) ||
  52 +                                        !test_opt(sb, ASYNCDEL));
  53 +               ext3_debug("%s woken up: %lu inodes, %lu blocks\n",
  54 +                          tsk->comm,sbi->s_delete_inodes,sbi->s_delete_blocks);
  55 +
  56 +               spin_lock(&sbi->s_delete_lock);
  57 +               if (list_empty(&sbi->s_delete_list)) {
  58 +                       clear_opt(sbi->s_mount_opt, ASYNCDEL);
  59 +                       memset(&sbi->s_delete_list, 0,
  60 +                              sizeof(sbi->s_delete_list));
  61 +                       spin_unlock(&sbi->s_delete_lock);
  62 +                       ext3_debug("delete thread on %s exiting\n",
  63 +                                  kdevname(sb->s_dev));
  64 +                       wake_up(&sbi->s_delete_waiter_queue);
  65 +                       break;
  66 +               }
  67 +
  68 +               while (!list_empty(&sbi->s_delete_list)) {
  69 +                       struct inode *inode=list_entry(sbi->s_delete_list.next,
  70 +                                                      struct inode, i_dentry);
  71 +                       unsigned long blocks = inode->i_blocks >>
  72 +                                                       (inode->i_blkbits - 9);
  73 +
  74 +                       list_del_init(&inode->i_dentry);
  75 +                       spin_unlock(&sbi->s_delete_lock);
  76 +                       ext3_debug("%s delete ino %lu blk %lu\n",
  77 +                                  tsk->comm, inode->i_ino, blocks);
  78 +
  79 +                       iput(inode);
  80 +
  81 +                       spin_lock(&sbi->s_delete_lock);
  82 +                       sbi->s_delete_blocks -= blocks;
  83 +                       sbi->s_delete_inodes--;
  84 +               }
  85 +               if (sbi->s_delete_blocks != 0 || sbi->s_delete_inodes != 0) {
  86 +                       ext3_warning(sb, __FUNCTION__,
  87 +                                    "%lu blocks, %lu inodes on list?\n",
  88 +                                    sbi->s_delete_blocks,sbi->s_delete_inodes);
  89 +                       sbi->s_delete_blocks = 0;
  90 +                       sbi->s_delete_inodes = 0;
  91 +               }
  92 +               spin_unlock(&sbi->s_delete_lock);
  93 +               wake_up(&sbi->s_delete_waiter_queue);
  94 +       }
  95 +
  96 +       return 0;
  97 +}
  98 +
  99 +static void ext3_start_delete_thread(struct super_block *sb)
 100 +{
 101 +       struct ext3_sb_info *sbi = EXT3_SB(sb);
 102 +       int rc;
 103 +
 104 +       spin_lock_init(&sbi->s_delete_lock);
 105 +       init_waitqueue_head(&sbi->s_delete_thread_queue);
 106 +       init_waitqueue_head(&sbi->s_delete_waiter_queue);
 107 +
 108 +       if (!test_opt(sb, ASYNCDEL))
 109 +               return;
 110 +
 111 +       rc = kernel_thread(ext3_delete_thread, sb, CLONE_VM | CLONE_FILES);
 112 +       if (rc < 0)
 113 +               printk(KERN_ERR "EXT3-fs: cannot start delete thread: rc %d\n",
 114 +                      rc);
 115 +       else
 116 +               wait_event(sbi->s_delete_waiter_queue, sbi->s_delete_list.next);
 117 +}
 118 +
 119 +static void ext3_stop_delete_thread(struct ext3_sb_info *sbi)
 120 +{
 121 +       if (sbi->s_delete_list.next == 0)       /* thread never started */
 122 +               return;
 123 +
 124 +       clear_opt(sbi->s_mount_opt, ASYNCDEL);
 125 +       wake_up(&sbi->s_delete_thread_queue);
 126 +       wait_event(sbi->s_delete_waiter_queue,
 127 +                       sbi->s_delete_list.next == 0 && sbi->s_delete_inodes == 0);
 128 +}
 129 +
 130 +/* Instead of playing games with the inode flags, destruction, etc we just
 131 + * create a new inode locally and put it on a list for the truncate thread.
 132 + * We need large parts of the inode struct in order to complete the
 133 + * truncate and unlink, so we may as well just have a real inode to do it.
 134 + *
 135 + * If we have any problem deferring the delete, just delete it right away.
 136 + * If we defer it, we also mark how many blocks it would free, so that we
 137 + * can keep the statfs data correct, and we know if we should sleep on the
 138 + * delete thread when we run out of space.
 139 + */
 140 +static void ext3_delete_inode_thread(struct inode *old_inode)
 141 +{
 142 +       struct ext3_sb_info *sbi = EXT3_SB(old_inode->i_sb);
 143 +       struct ext3_inode_info *nei, *oei = EXT3_I(old_inode);
 144 +       struct inode *new_inode;
 145 +       unsigned long blocks = old_inode->i_blocks >> (old_inode->i_blkbits-9);
 146 +
 147 +       if (is_bad_inode(old_inode)) {
 148 +               clear_inode(old_inode);
 149 +               return;
 150 +       }
 151 +
 152 +       if (!test_opt(old_inode->i_sb, ASYNCDEL) || !sbi->s_delete_list.next)
 153 +               goto out_delete;
 154 +
 155 +       /* We may want to delete the inode immediately and not defer it */
 156 +       if (IS_SYNC(old_inode) || blocks <= EXT3_NDIR_BLOCKS)
 157 +               goto out_delete;
 158 +
 159 +       /* We can't use the delete thread as-is during real orphan recovery,
 160 +        * as we add to the orphan list here, causing ext3_orphan_cleanup()
 161 +        * to loop endlessly.  It would be nice to do so, but needs work.
 162 +        */
 163 +       if (oei->i_state & EXT3_STATE_DELETE ||
 164 +           sbi->s_mount_state & EXT3_ORPHAN_FS) {
 165 +               ext3_debug("doing deferred inode %lu delete (%lu blocks)\n",
 166 +                          old_inode->i_ino, blocks);
 167 +               goto out_delete;
 168 +       }
 169 +
 170 +       /* We can iget this inode again here, because our caller has unhashed
 171 +        * old_inode, so new_inode will be in a different inode struct.
 172 +        *
 173 +        * We need to ensure that the i_orphan pointers in the other inodes
 174 +        * point at the new inode copy instead of the old one so the orphan
 175 +        * list doesn't get corrupted when the old orphan inode is freed.
 176 +        */
 177 +       down(&sbi->s_orphan_lock);
 178 +
 179 +       sbi->s_mount_state |= EXT3_ORPHAN_FS;
 180 +       new_inode = iget(old_inode->i_sb, old_inode->i_ino);
 181 +       sbi->s_mount_state &= ~EXT3_ORPHAN_FS;
 182 +       if (is_bad_inode(new_inode)) {
 183 +               printk(KERN_WARNING "read bad inode %lu\n", old_inode->i_ino);
 184 +               iput(new_inode);
 185 +               new_inode = NULL;
 186 +       }
 187 +       if (!new_inode) {
 188 +               up(&sbi->s_orphan_lock);
 189 +               ext3_debug("delete inode %lu directly (bad read)\n",
 190 +                          old_inode->i_ino);
 191 +               goto out_delete;
 192 +       }
 193 +       J_ASSERT(new_inode != old_inode);
 194 +
 195 +       J_ASSERT(!list_empty(&oei->i_orphan));
 196 +
 197 +       nei = EXT3_I(new_inode);
 198 +       /* Ugh.  We need to insert new_inode into the same spot on the list
 199 +        * as old_inode was, to ensure the in-memory orphan list is still
 200 +        * in the same order as the on-disk orphan list (badness otherwise).
 201 +        */
 202 +       nei->i_orphan = oei->i_orphan;
 203 +       nei->i_orphan.next->prev = &nei->i_orphan;
 204 +       nei->i_orphan.prev->next = &nei->i_orphan;
 205 +       nei->i_state |= EXT3_STATE_DELETE;
 206 +       up(&sbi->s_orphan_lock);
 207 +
 208 +       clear_inode(old_inode);
 209 +
 210 +       spin_lock(&sbi->s_delete_lock);
 211 +       J_ASSERT(list_empty(&new_inode->i_dentry));
 212 +       list_add_tail(&new_inode->i_dentry, &sbi->s_delete_list);
 213 +       sbi->s_delete_blocks += blocks;
 214 +       sbi->s_delete_inodes++;
 215 +       spin_unlock(&sbi->s_delete_lock);
 216 +
 217 +       ext3_debug("delete inode %lu (%lu blocks) by thread\n",
 218 +                  new_inode->i_ino, blocks);
 219 +
 220 +       wake_up(&sbi->s_delete_thread_queue);
 221 +       return;
 222 +
 223 +out_delete:
 224 +       ext3_delete_inode(old_inode);
 225 +}
 226 +#else
 227 +#define ext3_start_delete_thread(sbi) do {} while(0)
 228 +#define ext3_stop_delete_thread(sbi) do {} while(0)
 229 +#endif /* EXT3_DELETE_THREAD */
 230 +
 231  void ext3_put_super (struct super_block * sb)
 232  {
 233         struct ext3_sb_info *sbi = EXT3_SB(sb);
 234 @@ -407,6 +622,7 @@
 235         kdev_t j_dev = sbi->s_journal->j_dev;
 236         int i;
 237
 238 +       J_ASSERT(sbi->s_delete_inodes == 0);
 239         ext3_xattr_put_super(sb);
 240         journal_destroy(sbi->s_journal);
 241         if (!(sb->s_flags & MS_RDONLY)) {
 242 @@ -455,7 +671,11 @@
 243         write_inode:    ext3_write_inode,       /* BKL not held.  Don't need */
 244         dirty_inode:    ext3_dirty_inode,       /* BKL not held.  We take it */
 245         put_inode:      ext3_put_inode,         /* BKL not held.  Don't need */
 246 -       delete_inode:   ext3_delete_inode,      /* BKL not held.  We take it */
 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         sync_fs:        ext3_sync_fs,
 255 @@ -524,6 +744,13 @@
 256                         clear_opt (*mount_options, XATTR_USER);
 257                 else
 258  #endif
 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                 if (!strcmp (this_char, "bsddf"))
 267                         clear_opt (*mount_options, MINIX_DF);
 268                 else if (!strcmp (this_char, "nouid32")) {
 269 @@ -1223,6 +1450,7 @@
 270         }
 271
 272         ext3_setup_super (sb, es, sb->s_flags & MS_RDONLY);
 273 +       ext3_start_delete_thread(sb);
 274         /*
 275          * akpm: core read_super() calls in here with the superblock locked.
 276          * That deadlocks, because orphan cleanup needs to lock the superblock
 277 @@ -1614,7 +1842,12 @@
 278  static int ext3_sync_fs(struct super_block *sb)
 279  {
 280         tid_t target;
 281 -
 282 +
 283 +       if (atomic_read(&sb->s_active) == 0) {
 284 +               /* fs is being umounted: time to stop delete thread */
 285 +               ext3_stop_delete_thread(EXT3_SB(sb));
 286 +       }
 287 +
 288         sb->s_dirt = 0;
 289         target = log_start_commit(EXT3_SB(sb)->s_journal, NULL);
 290         log_wait_commit(EXT3_SB(sb)->s_journal, target);
 291 @@ -1678,6 +1911,9 @@
 292         if (!parse_options(data, &tmp, sbi, &tmp, 1))
 293                 return -EINVAL;
 294
 295 +       if (!test_opt(sb, ASYNCDEL) || (*flags & MS_RDONLY))
 296 +               ext3_stop_delete_thread(sbi);
 297 +
 298         if (sbi->s_mount_opt & EXT3_MOUNT_ABORT)
 299                 ext3_abort(sb, __FUNCTION__, "Abort forced by user");
 300
 301 Index: linux/fs/ext3/inode.c
 302 ===================================================================
 303 --- linux.orig/fs/ext3/inode.c  Mon Feb  2 20:57:35 2004
 304 +++ linux/fs/ext3/inode.c       Mon Feb  2 20:58:05 2004
 305 @@ -2500,6 +2500,118 @@
 306         return err;
 307  }
 308
 309 +#ifdef EXT3_DELETE_THREAD
 310 +/* Move blocks from to-be-truncated inode over to a new inode, and delete
 311 + * that one from the delete thread instead.  This avoids a lot of latency
 312 + * when truncating large files.
 313 + *
 314 + * If we have any problem deferring the truncate, just truncate it right away.
 315 + * If we defer it, we also mark how many blocks it would free, so that we
 316 + * can keep the statfs data correct, and we know if we should sleep on the
 317 + * delete thread when we run out of space.
 318 + */
 319 +void ext3_truncate_thread(struct inode *old_inode)
 320 +{
 321 +       struct ext3_sb_info *sbi = EXT3_SB(old_inode->i_sb);
 322 +       struct ext3_inode_info *nei, *oei = EXT3_I(old_inode);
 323 +       struct inode *new_inode;
 324 +       handle_t *handle;
 325 +       unsigned long blocks = old_inode->i_blocks >> (old_inode->i_blkbits-9);
 326 +
 327 +       if (!test_opt(old_inode->i_sb, ASYNCDEL) || !sbi->s_delete_list.next)
 328 +               goto out_truncate;
 329 +
 330 +       /* XXX This is a temporary limitation for code simplicity.
 331 +        *     We could truncate to arbitrary sizes at some later time.
 332 +        */
 333 +       if (old_inode->i_size != 0)
 334 +               goto out_truncate;
 335 +
 336 +       /* We may want to truncate the inode immediately and not defer it */
 337 +       if (IS_SYNC(old_inode) || blocks <= EXT3_NDIR_BLOCKS ||
 338 +           old_inode->i_size > oei->i_disksize)
 339 +               goto out_truncate;
 340 +
 341 +       /* We can't use the delete thread as-is during real orphan recovery,
 342 +        * as we add to the orphan list here, causing ext3_orphan_cleanup()
 343 +        * to loop endlessly.  It would be nice to do so, but needs work.
 344 +        */
 345 +       if (oei->i_state & EXT3_STATE_DELETE ||
 346 +           sbi->s_mount_state & EXT3_ORPHAN_FS) {
 347 +               ext3_debug("doing deferred inode %lu delete (%lu blocks)\n",
 348 +                          old_inode->i_ino, blocks);
 349 +               goto out_truncate;
 350 +       }
 351 +
 352 +       ext3_discard_prealloc(old_inode);
 353 +
 354 +       /* old_inode   = 1
 355 +        * new_inode   = sb + GDT + ibitmap
 356 +        * orphan list = 1 inode/superblock for add, 2 inodes for del
 357 +        * quota files = 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
 358 +        */
 359 +       handle = ext3_journal_start(old_inode, 7);
 360 +       if (IS_ERR(handle))
 361 +               goto out_truncate;
 362 +
 363 +       new_inode = ext3_new_inode(handle, old_inode, old_inode->i_mode);
 364 +       if (IS_ERR(new_inode)) {
 365 +               ext3_debug("truncate inode %lu directly (no new inodes)\n",
 366 +                          old_inode->i_ino);
 367 +               goto out_journal;
 368 +       }
 369 +
 370 +       nei = EXT3_I(new_inode);
 371 +
 372 +       down_write(&oei->truncate_sem);
 373 +       new_inode->i_size = old_inode->i_size;
 374 +       new_inode->i_blocks = old_inode->i_blocks;
 375 +       new_inode->i_uid = old_inode->i_uid;
 376 +       new_inode->i_gid = old_inode->i_gid;
 377 +       new_inode->i_nlink = 0;
 378 +
 379 +       /* FIXME when we do arbitrary truncates */
 380 +       old_inode->i_blocks = oei->i_file_acl ? old_inode->i_blksize / 512 : 0;
 381 +       old_inode->i_mtime = old_inode->i_ctime = CURRENT_TIME;
 382 +
 383 +       memcpy(nei->i_data, oei->i_data, sizeof(nei->i_data));
 384 +       memset(oei->i_data, 0, sizeof(oei->i_data));
 385 +
 386 +       nei->i_disksize = oei->i_disksize;
 387 +       nei->i_state |= EXT3_STATE_DELETE;
 388 +       up_write(&oei->truncate_sem);
 389 +
 390 +       if (ext3_orphan_add(handle, new_inode) < 0)
 391 +               goto out_journal;
 392 +
 393 +       if (ext3_orphan_del(handle, old_inode) < 0) {
 394 +               ext3_orphan_del(handle, new_inode);
 395 +               iput(new_inode);
 396 +               goto out_journal;
 397 +       }
 398 +
 399 +       ext3_journal_stop(handle, old_inode);
 400 +
 401 +       spin_lock(&sbi->s_delete_lock);
 402 +       J_ASSERT(list_empty(&new_inode->i_dentry));
 403 +       list_add_tail(&new_inode->i_dentry, &sbi->s_delete_list);
 404 +       sbi->s_delete_blocks += blocks;
 405 +       sbi->s_delete_inodes++;
 406 +       spin_unlock(&sbi->s_delete_lock);
 407 +
 408 +       ext3_debug("delete inode %lu (%lu blocks) by thread\n",
 409 +                  new_inode->i_ino, blocks);
 410 +
 411 +       wake_up(&sbi->s_delete_thread_queue);
 412 +       return;
 413 +
 414 +out_journal:
 415 +       ext3_journal_stop(handle, old_inode);
 416 +out_truncate:
 417 +       ext3_truncate(old_inode);
 418 +}
 419 +#endif /* EXT3_DELETE_THREAD */
 420 +
 421  /*
 422   * On success, We end up with an outstanding reference count against
 423   * iloc->bh.  This _must_ be cleaned up later.
 424 Index: linux/fs/ext3/file.c
 425 ===================================================================
 426 --- linux.orig/fs/ext3/file.c   Mon Feb  2 20:57:34 2004
 427 +++ linux/fs/ext3/file.c        Mon Feb  2 20:58:05 2004
 428 @@ -125,7 +125,11 @@
 429  };
 430
 431  struct inode_operations ext3_file_inode_operations = {
 432 +#ifdef EXT3_DELETE_THREAD
 433 +       truncate:       ext3_truncate_thread,   /* BKL held */
 434 +#else
 435         truncate:       ext3_truncate,          /* BKL held */
 436 +#endif
 437         setattr:        ext3_setattr,           /* BKL held */
 438         setxattr:       ext3_setxattr,          /* BKL held */
 439         getxattr:       ext3_getxattr,          /* BKL held */
 440 Index: linux/include/linux/ext3_fs.h
 441 ===================================================================
 442 --- linux.orig/include/linux/ext3_fs.h  Mon Feb  2 20:57:35 2004
 443 +++ linux/include/linux/ext3_fs.h       Mon Feb  2 20:58:05 2004
 444 @@ -193,6 +193,7 @@
 445   */
 446  #define EXT3_STATE_JDATA               0x00000001 /* journaled data exists */
 447  #define EXT3_STATE_NEW                 0x00000002 /* inode is newly created */
 448 +#define EXT3_STATE_DELETE              0x00000010 /* deferred delete inode */
 449
 450  /*
 451   * ioctl commands
 452 @@ -320,6 +321,7 @@
 453  #define EXT3_MOUNT_UPDATE_JOURNAL      0x1000  /* Update the journal format */
 454  #define EXT3_MOUNT_NO_UID32            0x2000  /* Disable 32-bit UIDs */
 455  #define EXT3_MOUNT_XATTR_USER          0x4000  /* Extended user attributes */
 456 +#define EXT3_MOUNT_ASYNCDEL            0x20000 /* Delayed deletion */
 457
 458  /* Compatibility, for having both ext2_fs.h and ext3_fs.h included at once */
 459  #ifndef _LINUX_EXT2_FS_H
 460 @@ -694,6 +696,9 @@
 461  extern void ext3_dirty_inode(struct inode *);
 462  extern int ext3_change_inode_journal_flag(struct inode *, int);
 463  extern void ext3_truncate (struct inode *);
 464 +#ifdef EXT3_DELETE_THREAD
 465 +extern void ext3_truncate_thread(struct inode *inode);
 466 +#endif
 467
 468  /* ioctl.c */
 469  extern int ext3_ioctl (struct inode *, struct file *, unsigned int,
 470 Index: linux/include/linux/ext3_fs_sb.h
 471 ===================================================================
 472 --- linux.orig/include/linux/ext3_fs_sb.h       Mon Feb  2 20:57:35 2004
 473 +++ linux/include/linux/ext3_fs_sb.h    Mon Feb  2 20:58:05 2004
 474 @@ -29,6 +29,8 @@
 475
 476  #define EXT3_MAX_GROUP_LOADED  8
 477
 478 +#define EXT3_DELETE_THREAD
 479 +
 480  /*
 481   * third extended-fs super-block data in memory
 482   */
 483 @@ -76,6 +78,14 @@
 484         struct timer_list turn_ro_timer;        /* For turning read-only (crash simulation) */
 485         wait_queue_head_t ro_wait_queue;        /* For people waiting for the fs to go read-only */
 486  #endif
 487 +#ifdef EXT3_DELETE_THREAD
 488 +       spinlock_t s_delete_lock;
 489 +       struct list_head s_delete_list;
 490 +       unsigned long s_delete_blocks;
 491 +       unsigned long s_delete_inodes;
 492 +       wait_queue_head_t s_delete_thread_queue;
 493 +       wait_queue_head_t s_delete_waiter_queue;
 494 +#endif
 495  };
 496
 497  #endif /* _LINUX_EXT3_FS_SB */