lustre/kernel_patches/patches/ext3-delete_thread-2.4.29.patch

   1 Index: linux-2.4.29/fs/ext3/super.c
   2 ===================================================================
   3 --- linux-2.4.29.orig/fs/ext3/super.c   2005-05-03 15:53:33.047533872 +0300
   4 +++ linux-2.4.29/fs/ext3/super.c        2005-05-03 15:54:47.192262160 +0300
   5 @@ -400,6 +400,127 @@
   6         }
   7  }
   8
   9 +#ifdef EXT3_DELETE_THREAD
  10 +/*
  11 + * Delete inodes in a loop until there are no more to be deleted.
  12 + * Normally, we run in the background doing the deletes and sleeping again,
  13 + * and clients just add new inodes to be deleted onto the end of the list.
  14 + * If someone is concerned about free space (e.g. block allocation or similar)
  15 + * then they can sleep on s_delete_waiter_queue and be woken up when space
  16 + * has been freed.
  17 + */
  18 +int ext3_delete_thread(void *data)
  19 +{
  20 +       struct super_block *sb = data;
  21 +       struct ext3_sb_info *sbi = EXT3_SB(sb);
  22 +       struct task_struct *tsk = current;
  23 +
  24 +       /* Almost like daemonize, but not quite */
  25 +       exit_mm(current);
  26 +       tsk->session = 1;
  27 +       tsk->pgrp = 1;
  28 +       tsk->tty = NULL;
  29 +       exit_files(current);
  30 +       reparent_to_init();
  31 +
  32 +       sprintf(tsk->comm, "kdelext3-%s", kdevname(sb->s_dev));
  33 +       sigfillset(&tsk->blocked);
  34 +
  35 +       /*tsk->flags |= PF_KERNTHREAD;*/
  36 +
  37 +       INIT_LIST_HEAD(&sbi->s_delete_list);
  38 +       wake_up(&sbi->s_delete_waiter_queue);
  39 +       ext3_debug("delete thread on %s started\n", kdevname(sb->s_dev));
  40 +
  41 +       /* main loop */
  42 +       for (;;) {
  43 +               wait_event_interruptible(sbi->s_delete_thread_queue,
  44 +                                        !list_empty(&sbi->s_delete_list) ||
  45 +                                        !test_opt(sb, ASYNCDEL));
  46 +               ext3_debug("%s woken up: %lu inodes, %lu blocks\n",
  47 +                          tsk->comm,sbi->s_delete_inodes,sbi->s_delete_blocks);
  48 +
  49 +               spin_lock(&sbi->s_delete_lock);
  50 +               if (list_empty(&sbi->s_delete_list)) {
  51 +                       clear_opt(sbi->s_mount_opt, ASYNCDEL);
  52 +                       memset(&sbi->s_delete_list, 0,
  53 +                              sizeof(sbi->s_delete_list));
  54 +                       spin_unlock(&sbi->s_delete_lock);
  55 +                       ext3_debug("delete thread on %s exiting\n",
  56 +                                  kdevname(sb->s_dev));
  57 +                       wake_up(&sbi->s_delete_waiter_queue);
  58 +                       break;
  59 +               }
  60 +
  61 +               while (!list_empty(&sbi->s_delete_list)) {
  62 +                       struct inode *inode=list_entry(sbi->s_delete_list.next,
  63 +                                                      struct inode, i_devices);
  64 +                       unsigned long blocks = inode->i_blocks >>
  65 +                                                       (inode->i_blkbits - 9);
  66 +
  67 +                       list_del_init(&inode->i_devices);
  68 +                       spin_unlock(&sbi->s_delete_lock);
  69 +                       ext3_debug("%s delete ino %lu blk %lu\n",
  70 +                                  tsk->comm, inode->i_ino, blocks);
  71 +
  72 +                       J_ASSERT(EXT3_I(inode)->i_state & EXT3_STATE_DELETE);
  73 +                       J_ASSERT(inode->i_nlink == 1);
  74 +                       inode->i_nlink = 0;
  75 +                       iput(inode);
  76 +
  77 +                       spin_lock(&sbi->s_delete_lock);
  78 +                       sbi->s_delete_blocks -= blocks;
  79 +                       sbi->s_delete_inodes--;
  80 +               }
  81 +               if (sbi->s_delete_blocks != 0 || sbi->s_delete_inodes != 0) {
  82 +                       ext3_warning(sb, __FUNCTION__,
  83 +                                    "%lu blocks, %lu inodes on list?\n",
  84 +                                    sbi->s_delete_blocks,sbi->s_delete_inodes);
  85 +                       sbi->s_delete_blocks = 0;
  86 +                       sbi->s_delete_inodes = 0;
  87 +               }
  88 +               spin_unlock(&sbi->s_delete_lock);
  89 +               wake_up(&sbi->s_delete_waiter_queue);
  90 +       }
  91 +
  92 +       return 0;
  93 +}
  94 +
  95 +static void ext3_start_delete_thread(struct super_block *sb)
  96 +{
  97 +       struct ext3_sb_info *sbi = EXT3_SB(sb);
  98 +       int rc;
  99 +
 100 +       spin_lock_init(&sbi->s_delete_lock);
 101 +       init_waitqueue_head(&sbi->s_delete_thread_queue);
 102 +       init_waitqueue_head(&sbi->s_delete_waiter_queue);
 103 +
 104 +       if (!test_opt(sb, ASYNCDEL))
 105 +               return;
 106 +
 107 +       rc = kernel_thread(ext3_delete_thread, sb, CLONE_VM | CLONE_FILES);
 108 +       if (rc < 0)
 109 +               printk(KERN_ERR "EXT3-fs: cannot start delete thread: rc %d\n",
 110 +                      rc);
 111 +       else
 112 +               wait_event(sbi->s_delete_waiter_queue, sbi->s_delete_list.next);
 113 +}
 114 +
 115 +static void ext3_stop_delete_thread(struct ext3_sb_info *sbi)
 116 +{
 117 +       if (sbi->s_delete_list.next == 0)       /* thread never started */
 118 +               return;
 119 +
 120 +       clear_opt(sbi->s_mount_opt, ASYNCDEL);
 121 +       wake_up(&sbi->s_delete_thread_queue);
 122 +       wait_event(sbi->s_delete_waiter_queue,
 123 +                       sbi->s_delete_list.next == 0 && sbi->s_delete_inodes == 0);
 124 +}
 125 +#else
 126 +#define ext3_start_delete_thread(sbi) do {} while(0)
 127 +#define ext3_stop_delete_thread(sbi) do {} while(0)
 128 +#endif /* EXT3_DELETE_THREAD */
 129 +
 130  void ext3_put_super (struct super_block * sb)
 131  {
 132         struct ext3_sb_info *sbi = EXT3_SB(sb);
 133 @@ -407,6 +528,9 @@
 134         kdev_t j_dev = sbi->s_journal->j_dev;
 135         int i;
 136
 137 +#ifdef EXT3_DELETE_THREAD
 138 +       J_ASSERT(sbi->s_delete_inodes == 0);
 139 +#endif
 140         ext3_xattr_put_super(sb);
 141         journal_destroy(sbi->s_journal);
 142         if (!(sb->s_flags & MS_RDONLY)) {
 143 @@ -526,6 +650,13 @@
 144                         clear_opt (*mount_options, XATTR_USER);
 145                 else
 146  #endif
 147 +#ifdef EXT3_DELETE_THREAD
 148 +               if (!strcmp(this_char, "asyncdel"))
 149 +                       set_opt(*mount_options, ASYNCDEL);
 150 +               else if (!strcmp(this_char, "noasyncdel"))
 151 +                       clear_opt(*mount_options, ASYNCDEL);
 152 +               else
 153 +#endif
 154                 if (!strcmp (this_char, "bsddf"))
 155                         clear_opt (*mount_options, MINIX_DF);
 156                 else if (!strcmp (this_char, "nouid32")) {
 157 @@ -1244,6 +1375,7 @@
 158         }
 159
 160         ext3_setup_super (sb, es, sb->s_flags & MS_RDONLY);
 161 +       ext3_start_delete_thread(sb);
 162         EXT3_SB(sb)->s_mount_state |= EXT3_ORPHAN_FS;
 163         ext3_orphan_cleanup(sb, es);
 164         EXT3_SB(sb)->s_mount_state &= ~EXT3_ORPHAN_FS;
 165 @@ -1626,7 +1758,12 @@
 166  static int ext3_sync_fs(struct super_block *sb)
 167  {
 168         tid_t target;
 169 -
 170 +
 171 +       if (atomic_read(&sb->s_active) == 0) {
 172 +               /* fs is being umounted: time to stop delete thread */
 173 +               ext3_stop_delete_thread(EXT3_SB(sb));
 174 +       }
 175 +
 176         sb->s_dirt = 0;
 177         target = log_start_commit(EXT3_SB(sb)->s_journal, NULL);
 178         log_wait_commit(EXT3_SB(sb)->s_journal, target);
 179 @@ -1690,6 +1827,9 @@
 180         if (!parse_options(data, &tmp, sbi, &tmp, 1))
 181                 return -EINVAL;
 182
 183 +       if (!test_opt(sb, ASYNCDEL) || (*flags & MS_RDONLY))
 184 +               ext3_stop_delete_thread(sbi);
 185 +
 186         if (sbi->s_mount_opt & EXT3_MOUNT_ABORT)
 187                 ext3_abort(sb, __FUNCTION__, "Abort forced by user");
 188
 189 Index: linux-2.4.29/fs/ext3/inode.c
 190 ===================================================================
 191 --- linux-2.4.29.orig/fs/ext3/inode.c   2005-05-03 15:53:36.555000656 +0300
 192 +++ linux-2.4.29/fs/ext3/inode.c        2005-05-03 15:53:56.901907456 +0300
 193 @@ -2562,6 +2562,118 @@
 194         return err;
 195  }
 196
 197 +#ifdef EXT3_DELETE_THREAD
 198 +/* Move blocks from to-be-truncated inode over to a new inode, and delete
 199 + * that one from the delete thread instead.  This avoids a lot of latency
 200 + * when truncating large files.
 201 + *
 202 + * If we have any problem deferring the truncate, just truncate it right away.
 203 + * If we defer it, we also mark how many blocks it would free, so that we
 204 + * can keep the statfs data correct, and we know if we should sleep on the
 205 + * delete thread when we run out of space.
 206 + */
 207 +void ext3_truncate_thread(struct inode *old_inode)
 208 +{
 209 +       struct ext3_sb_info *sbi = EXT3_SB(old_inode->i_sb);
 210 +       struct ext3_inode_info *nei, *oei = EXT3_I(old_inode);
 211 +       struct inode *new_inode;
 212 +       handle_t *handle;
 213 +       unsigned long blocks = old_inode->i_blocks >> (old_inode->i_blkbits-9);
 214 +
 215 +       if (!test_opt(old_inode->i_sb, ASYNCDEL) || !sbi->s_delete_list.next)
 216 +               goto out_truncate;
 217 +
 218 +       /* XXX This is a temporary limitation for code simplicity.
 219 +        *     We could truncate to arbitrary sizes at some later time.
 220 +        */
 221 +       if (old_inode->i_size != 0)
 222 +               goto out_truncate;
 223 +
 224 +       /* We may want to truncate the inode immediately and not defer it */
 225 +       if (IS_SYNC(old_inode) || blocks <= EXT3_NDIR_BLOCKS ||
 226 +           old_inode->i_size > oei->i_disksize)
 227 +               goto out_truncate;
 228 +
 229 +       /* We can't use the delete thread as-is during real orphan recovery,
 230 +        * as we add to the orphan list here, causing ext3_orphan_cleanup()
 231 +        * to loop endlessly.  It would be nice to do so, but needs work.
 232 +        */
 233 +       if (oei->i_state & EXT3_STATE_DELETE ||
 234 +           sbi->s_mount_state & EXT3_ORPHAN_FS) {
 235 +               ext3_debug("doing deferred inode %lu delete (%lu blocks)\n",
 236 +                          old_inode->i_ino, blocks);
 237 +               goto out_truncate;
 238 +       }
 239 +
 240 +       ext3_discard_prealloc(old_inode);
 241 +
 242 +       /* old_inode   = 1
 243 +        * new_inode   = sb + GDT + ibitmap
 244 +        * orphan list = 1 inode/superblock for add, 2 inodes for del
 245 +        * quota files = 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
 246 +        */
 247 +       handle = ext3_journal_start(old_inode, 7);
 248 +       if (IS_ERR(handle))
 249 +               goto out_truncate;
 250 +
 251 +       new_inode = ext3_new_inode(handle, old_inode, old_inode->i_mode);
 252 +       if (IS_ERR(new_inode)) {
 253 +               ext3_debug("truncate inode %lu directly (no new inodes)\n",
 254 +                          old_inode->i_ino);
 255 +               goto out_journal;
 256 +       }
 257 +
 258 +       nei = EXT3_I(new_inode);
 259 +
 260 +       down_write(&oei->truncate_sem);
 261 +       new_inode->i_size = old_inode->i_size;
 262 +       new_inode->i_blocks = old_inode->i_blocks;
 263 +       new_inode->i_uid = old_inode->i_uid;
 264 +       new_inode->i_gid = old_inode->i_gid;
 265 +       new_inode->i_nlink = 1;
 266 +
 267 +       /* FIXME when we do arbitrary truncates */
 268 +       old_inode->i_blocks = oei->i_file_acl ? old_inode->i_blksize / 512 : 0;
 269 +       old_inode->i_mtime = old_inode->i_ctime = CURRENT_TIME;
 270 +
 271 +       memcpy(nei->i_data, oei->i_data, sizeof(nei->i_data));
 272 +       memset(oei->i_data, 0, sizeof(oei->i_data));
 273 +
 274 +       nei->i_disksize = oei->i_disksize;
 275 +       nei->i_state |= EXT3_STATE_DELETE;
 276 +       up_write(&oei->truncate_sem);
 277 +
 278 +       if (ext3_orphan_add(handle, new_inode) < 0)
 279 +               goto out_journal;
 280 +
 281 +       if (ext3_orphan_del(handle, old_inode) < 0) {
 282 +               ext3_orphan_del(handle, new_inode);
 283 +               iput(new_inode);
 284 +               goto out_journal;
 285 +       }
 286 +
 287 +       ext3_journal_stop(handle, old_inode);
 288 +
 289 +       spin_lock(&sbi->s_delete_lock);
 290 +       J_ASSERT(list_empty(&new_inode->i_devices));
 291 +       list_add_tail(&new_inode->i_devices, &sbi->s_delete_list);
 292 +       sbi->s_delete_blocks += blocks;
 293 +       sbi->s_delete_inodes++;
 294 +       spin_unlock(&sbi->s_delete_lock);
 295 +
 296 +       ext3_debug("delete inode %lu (%lu blocks) by thread\n",
 297 +                  new_inode->i_ino, blocks);
 298 +
 299 +       wake_up(&sbi->s_delete_thread_queue);
 300 +       return;
 301 +
 302 +out_journal:
 303 +       ext3_journal_stop(handle, old_inode);
 304 +out_truncate:
 305 +       ext3_truncate(old_inode);
 306 +}
 307 +#endif /* EXT3_DELETE_THREAD */
 308 +
 309  /*
 310   * On success, We end up with an outstanding reference count against
 311   * iloc->bh.  This _must_ be cleaned up later.
 312 Index: linux-2.4.29/fs/ext3/file.c
 313 ===================================================================
 314 --- linux-2.4.29.orig/fs/ext3/file.c    2005-04-07 19:31:00.000000000 +0300
 315 +++ linux-2.4.29/fs/ext3/file.c 2005-05-03 15:53:56.902907304 +0300
 316 @@ -123,7 +123,11 @@
 317  };
 318
 319  struct inode_operations ext3_file_inode_operations = {
 320 +#ifdef EXT3_DELETE_THREAD
 321 +       truncate:       ext3_truncate_thread,   /* BKL held */
 322 +#else
 323         truncate:       ext3_truncate,          /* BKL held */
 324 +#endif
 325         setattr:        ext3_setattr,           /* BKL held */
 326         setxattr:       ext3_setxattr,          /* BKL held */
 327         getxattr:       ext3_getxattr,          /* BKL held */
 328 Index: linux-2.4.29/fs/ext3/namei.c
 329 ===================================================================
 330 --- linux-2.4.29.orig/fs/ext3/namei.c   2005-05-03 15:53:33.044534328 +0300
 331 +++ linux-2.4.29/fs/ext3/namei.c        2005-05-03 15:53:56.905906848 +0300
 332 @@ -838,6 +838,40 @@
 333         return retval;
 334  }
 335
 336 +#ifdef EXT3_DELETE_THREAD
 337 +static int ext3_try_to_delay_deletion(struct inode *inode)
 338 +{
 339 +       struct ext3_sb_info *sbi = EXT3_SB(inode->i_sb);
 340 +       struct ext3_inode_info *ei = EXT3_I(inode);
 341 +       unsigned long blocks;
 342 +
 343 +       if (!test_opt(inode->i_sb, ASYNCDEL))
 344 +               return 0;
 345 +
 346 +       /* We may want to delete the inode immediately and not defer it */
 347 +       blocks = inode->i_blocks >> (inode->i_blkbits - 9);
 348 +       if (IS_SYNC(inode) || blocks <= EXT3_NDIR_BLOCKS)
 349 +               return 0;
 350 +
 351 +       inode->i_nlink = 1;
 352 +       atomic_inc(&inode->i_count);
 353 +       ei->i_state |= EXT3_STATE_DELETE;
 354 +
 355 +       spin_lock(&sbi->s_delete_lock);
 356 +       J_ASSERT(list_empty(&inode->i_devices));
 357 +       list_add_tail(&inode->i_devices, &sbi->s_delete_list);
 358 +       sbi->s_delete_blocks += blocks;
 359 +       sbi->s_delete_inodes++;
 360 +       spin_unlock(&sbi->s_delete_lock);
 361 +
 362 +       wake_up(&sbi->s_delete_thread_queue);
 363 +
 364 +       return 0;
 365 +}
 366 +#else
 367 +#define ext3_try_to_delay_deletion(inode) do {} while (0)
 368 +#endif
 369 +
 370  static int ext3_unlink(struct inode * dir, struct dentry *dentry)
 371  {
 372         int retval;
 373 @@ -878,8 +912,10 @@
 374         dir->u.ext3_i.i_flags &= ~EXT3_INDEX_FL;
 375         ext3_mark_inode_dirty(handle, dir);
 376         inode->i_nlink--;
 377 -       if (!inode->i_nlink)
 378 +       if (!inode->i_nlink) {
 379 +               ext3_try_to_delay_deletion(inode);
 380                 ext3_orphan_add(handle, inode);
 381 +       }
 382         inode->i_ctime = dir->i_ctime;
 383         ext3_mark_inode_dirty(handle, inode);
 384         retval = 0;
 385 Index: linux-2.4.29/include/linux/ext3_fs.h
 386 ===================================================================
 387 --- linux-2.4.29.orig/include/linux/ext3_fs.h   2005-05-03 15:53:37.124914016 +0300
 388 +++ linux-2.4.29/include/linux/ext3_fs.h        2005-05-03 15:53:56.907906544 +0300
 389 @@ -188,6 +188,7 @@
 390   */
 391  #define EXT3_STATE_JDATA               0x00000001 /* journaled data exists */
 392  #define EXT3_STATE_NEW                 0x00000002 /* inode is newly created */
 393 +#define EXT3_STATE_DELETE              0x00000010 /* deferred delete inode */
 394
 395  /*
 396   * ioctl commands
 397 @@ -315,6 +316,7 @@
 398  #define EXT3_MOUNT_UPDATE_JOURNAL      0x1000  /* Update the journal format */
 399  #define EXT3_MOUNT_NO_UID32            0x2000  /* Disable 32-bit UIDs */
 400  #define EXT3_MOUNT_XATTR_USER          0x4000  /* Extended user attributes */
 401 +#define EXT3_MOUNT_ASYNCDEL            0x20000 /* Delayed deletion */
 402
 403  /* Compatibility, for having both ext2_fs.h and ext3_fs.h included at once */
 404  #ifndef _LINUX_EXT2_FS_H
 405 @@ -639,6 +641,9 @@
 406  extern void ext3_dirty_inode(struct inode *);
 407  extern int ext3_change_inode_journal_flag(struct inode *, int);
 408  extern void ext3_truncate (struct inode *);
 409 +#ifdef EXT3_DELETE_THREAD
 410 +extern void ext3_truncate_thread(struct inode *inode);
 411 +#endif
 412  extern void ext3_set_inode_flags(struct inode *);
 413
 414  /* ioctl.c */
 415 Index: linux-2.4.29/include/linux/ext3_fs_sb.h
 416 ===================================================================
 417 --- linux-2.4.29.orig/include/linux/ext3_fs_sb.h        2005-05-03 15:53:33.048533720 +0300
 418 +++ linux-2.4.29/include/linux/ext3_fs_sb.h     2005-05-03 15:53:56.909906240 +0300
 419 @@ -29,6 +29,8 @@
 420
 421  #define EXT3_MAX_GROUP_LOADED  8
 422
 423 +#define EXT3_DELETE_THREAD
 424 +
 425  /*
 426   * third extended-fs super-block data in memory
 427   */
 428 @@ -74,6 +76,14 @@
 429         struct timer_list turn_ro_timer;        /* For turning read-only (crash simulation) */
 430         wait_queue_head_t ro_wait_queue;        /* For people waiting for the fs to go read-only */
 431  #endif
 432 +#ifdef EXT3_DELETE_THREAD
 433 +       spinlock_t s_delete_lock;
 434 +       struct list_head s_delete_list;
 435 +       unsigned long s_delete_blocks;
 436 +       unsigned long s_delete_inodes;
 437 +       wait_queue_head_t s_delete_thread_queue;
 438 +       wait_queue_head_t s_delete_waiter_queue;
 439 +#endif
 440  };
 441
 442  #endif /* _LINUX_EXT3_FS_SB */