lustre/kernel_patches/patches/ext3-delete_thread-2.4.24.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-2.4.24/fs/ext3/super.c
   9 ===================================================================
  10 --- linux-2.4.24.orig/fs/ext3/super.c   2004-01-12 20:36:31.000000000 +0300
  11 +++ linux-2.4.24/fs/ext3/super.c        2004-01-13 16:27:43.000000000 +0300
  12 @@ -400,6 +400,127 @@
  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_devices);
  71 +                       unsigned long blocks = inode->i_blocks >>
  72 +                                                       (inode->i_blkbits - 9);
  73 +
  74 +                       list_del_init(&inode->i_devices);
  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 +                       J_ASSERT(EXT3_I(inode)->i_state & EXT3_STATE_DELETE);
  80 +                       J_ASSERT(inode->i_nlink == 1);
  81 +                       inode->i_nlink = 0;
  82 +                       iput(inode);
  83 +
  84 +                       spin_lock(&sbi->s_delete_lock);
  85 +                       sbi->s_delete_blocks -= blocks;
  86 +                       sbi->s_delete_inodes--;
  87 +               }
  88 +               if (sbi->s_delete_blocks != 0 || sbi->s_delete_inodes != 0) {
  89 +                       ext3_warning(sb, __FUNCTION__,
  90 +                                    "%lu blocks, %lu inodes on list?\n",
  91 +                                    sbi->s_delete_blocks,sbi->s_delete_inodes);
  92 +                       sbi->s_delete_blocks = 0;
  93 +                       sbi->s_delete_inodes = 0;
  94 +               }
  95 +               spin_unlock(&sbi->s_delete_lock);
  96 +               wake_up(&sbi->s_delete_waiter_queue);
  97 +       }
  98 +
  99 +       return 0;
 100 +}
 101 +
 102 +static void ext3_start_delete_thread(struct super_block *sb)
 103 +{
 104 +       struct ext3_sb_info *sbi = EXT3_SB(sb);
 105 +       int rc;
 106 +
 107 +       spin_lock_init(&sbi->s_delete_lock);
 108 +       init_waitqueue_head(&sbi->s_delete_thread_queue);
 109 +       init_waitqueue_head(&sbi->s_delete_waiter_queue);
 110 +
 111 +       if (!test_opt(sb, ASYNCDEL))
 112 +               return;
 113 +
 114 +       rc = kernel_thread(ext3_delete_thread, sb, CLONE_VM | CLONE_FILES);
 115 +       if (rc < 0)
 116 +               printk(KERN_ERR "EXT3-fs: cannot start delete thread: rc %d\n",
 117 +                      rc);
 118 +       else
 119 +               wait_event(sbi->s_delete_waiter_queue, sbi->s_delete_list.next);
 120 +}
 121 +
 122 +static void ext3_stop_delete_thread(struct ext3_sb_info *sbi)
 123 +{
 124 +       if (sbi->s_delete_list.next == 0)       /* thread never started */
 125 +               return;
 126 +
 127 +       clear_opt(sbi->s_mount_opt, ASYNCDEL);
 128 +       wake_up(&sbi->s_delete_thread_queue);
 129 +       wait_event(sbi->s_delete_waiter_queue,
 130 +                       sbi->s_delete_list.next == 0 && sbi->s_delete_inodes == 0);
 131 +}
 132 +#else
 133 +#define ext3_start_delete_thread(sbi) do {} while(0)
 134 +#define ext3_stop_delete_thread(sbi) do {} while(0)
 135 +#endif /* EXT3_DELETE_THREAD */
 136 +
 137  void ext3_put_super (struct super_block * sb)
 138  {
 139         struct ext3_sb_info *sbi = EXT3_SB(sb);
 140 @@ -407,6 +529,9 @@
 141         kdev_t j_dev = sbi->s_journal->j_dev;
 142         int i;
 143
 144 +#ifdef EXT3_DELETE_THREAD
 145 +       J_ASSERT(sbi->s_delete_inodes == 0);
 146 +#endif
 147         ext3_xattr_put_super(sb);
 148         journal_destroy(sbi->s_journal);
 149         if (!(sb->s_flags & MS_RDONLY)) {
 150 @@ -527,6 +650,13 @@
 151                         clear_opt (*mount_options, XATTR_USER);
 152                 else
 153  #endif
 154 +#ifdef EXT3_DELETE_THREAD
 155 +               if (!strcmp(this_char, "asyncdel"))
 156 +                       set_opt(*mount_options, ASYNCDEL);
 157 +               else if (!strcmp(this_char, "noasyncdel"))
 158 +                       clear_opt(*mount_options, ASYNCDEL);
 159 +               else
 160 +#endif
 161                 if (!strcmp (this_char, "bsddf"))
 162                         clear_opt (*mount_options, MINIX_DF);
 163                 else if (!strcmp (this_char, "nouid32")) {
 164 @@ -1227,6 +1357,7 @@
 165         }
 166
 167         ext3_setup_super (sb, es, sb->s_flags & MS_RDONLY);
 168 +       ext3_start_delete_thread(sb);
 169         /*
 170          * akpm: core read_super() calls in here with the superblock locked.
 171          * That deadlocks, because orphan cleanup needs to lock the superblock
 172 @@ -1618,7 +1749,12 @@
 173  static int ext3_sync_fs(struct super_block *sb)
 174  {
 175         tid_t target;
 176 -
 177 +
 178 +       if (atomic_read(&sb->s_active) == 0) {
 179 +               /* fs is being umounted: time to stop delete thread */
 180 +               ext3_stop_delete_thread(EXT3_SB(sb));
 181 +       }
 182 +
 183         sb->s_dirt = 0;
 184         target = log_start_commit(EXT3_SB(sb)->s_journal, NULL);
 185         log_wait_commit(EXT3_SB(sb)->s_journal, target);
 186 @@ -1682,6 +1818,9 @@
 187         if (!parse_options(data, &tmp, sbi, &tmp, 1))
 188                 return -EINVAL;
 189
 190 +       if (!test_opt(sb, ASYNCDEL) || (*flags & MS_RDONLY))
 191 +               ext3_stop_delete_thread(sbi);
 192 +
 193         if (sbi->s_mount_opt & EXT3_MOUNT_ABORT)
 194                 ext3_abort(sb, __FUNCTION__, "Abort forced by user");
 195
 196 Index: linux-2.4.24/fs/ext3/inode.c
 197 ===================================================================
 198 --- linux-2.4.24.orig/fs/ext3/inode.c   2004-01-12 20:36:31.000000000 +0300
 199 +++ linux-2.4.24/fs/ext3/inode.c        2004-01-12 20:36:32.000000000 +0300
 200 @@ -2551,6 +2551,118 @@
 201         return err;
 202  }
 203
 204 +#ifdef EXT3_DELETE_THREAD
 205 +/* Move blocks from to-be-truncated inode over to a new inode, and delete
 206 + * that one from the delete thread instead.  This avoids a lot of latency
 207 + * when truncating large files.
 208 + *
 209 + * If we have any problem deferring the truncate, just truncate it right away.
 210 + * If we defer it, we also mark how many blocks it would free, so that we
 211 + * can keep the statfs data correct, and we know if we should sleep on the
 212 + * delete thread when we run out of space.
 213 + */
 214 +void ext3_truncate_thread(struct inode *old_inode)
 215 +{
 216 +       struct ext3_sb_info *sbi = EXT3_SB(old_inode->i_sb);
 217 +       struct ext3_inode_info *nei, *oei = EXT3_I(old_inode);
 218 +       struct inode *new_inode;
 219 +       handle_t *handle;
 220 +       unsigned long blocks = old_inode->i_blocks >> (old_inode->i_blkbits-9);
 221 +
 222 +       if (!test_opt(old_inode->i_sb, ASYNCDEL) || !sbi->s_delete_list.next)
 223 +               goto out_truncate;
 224 +
 225 +       /* XXX This is a temporary limitation for code simplicity.
 226 +        *     We could truncate to arbitrary sizes at some later time.
 227 +        */
 228 +       if (old_inode->i_size != 0)
 229 +               goto out_truncate;
 230 +
 231 +       /* We may want to truncate the inode immediately and not defer it */
 232 +       if (IS_SYNC(old_inode) || blocks <= EXT3_NDIR_BLOCKS ||
 233 +           old_inode->i_size > oei->i_disksize)
 234 +               goto out_truncate;
 235 +
 236 +       /* We can't use the delete thread as-is during real orphan recovery,
 237 +        * as we add to the orphan list here, causing ext3_orphan_cleanup()
 238 +        * to loop endlessly.  It would be nice to do so, but needs work.
 239 +        */
 240 +       if (oei->i_state & EXT3_STATE_DELETE ||
 241 +           sbi->s_mount_state & EXT3_ORPHAN_FS) {
 242 +               ext3_debug("doing deferred inode %lu delete (%lu blocks)\n",
 243 +                          old_inode->i_ino, blocks);
 244 +               goto out_truncate;
 245 +       }
 246 +
 247 +       ext3_discard_prealloc(old_inode);
 248 +
 249 +       /* old_inode   = 1
 250 +        * new_inode   = sb + GDT + ibitmap
 251 +        * orphan list = 1 inode/superblock for add, 2 inodes for del
 252 +        * quota files = 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
 253 +        */
 254 +       handle = ext3_journal_start(old_inode, 7);
 255 +       if (IS_ERR(handle))
 256 +               goto out_truncate;
 257 +
 258 +       new_inode = ext3_new_inode(handle, old_inode, old_inode->i_mode);
 259 +       if (IS_ERR(new_inode)) {
 260 +               ext3_debug("truncate inode %lu directly (no new inodes)\n",
 261 +                          old_inode->i_ino);
 262 +               goto out_journal;
 263 +       }
 264 +
 265 +       nei = EXT3_I(new_inode);
 266 +
 267 +       down_write(&oei->truncate_sem);
 268 +       new_inode->i_size = old_inode->i_size;
 269 +       new_inode->i_blocks = old_inode->i_blocks;
 270 +       new_inode->i_uid = old_inode->i_uid;
 271 +       new_inode->i_gid = old_inode->i_gid;
 272 +       new_inode->i_nlink = 1;
 273 +
 274 +       /* FIXME when we do arbitrary truncates */
 275 +       old_inode->i_blocks = oei->i_file_acl ? old_inode->i_blksize / 512 : 0;
 276 +       old_inode->i_mtime = old_inode->i_ctime = CURRENT_TIME;
 277 +
 278 +       memcpy(nei->i_data, oei->i_data, sizeof(nei->i_data));
 279 +       memset(oei->i_data, 0, sizeof(oei->i_data));
 280 +
 281 +       nei->i_disksize = oei->i_disksize;
 282 +       nei->i_state |= EXT3_STATE_DELETE;
 283 +       up_write(&oei->truncate_sem);
 284 +
 285 +       if (ext3_orphan_add(handle, new_inode) < 0)
 286 +               goto out_journal;
 287 +
 288 +       if (ext3_orphan_del(handle, old_inode) < 0) {
 289 +               ext3_orphan_del(handle, new_inode);
 290 +               iput(new_inode);
 291 +               goto out_journal;
 292 +       }
 293 +
 294 +       ext3_journal_stop(handle, old_inode);
 295 +
 296 +       spin_lock(&sbi->s_delete_lock);
 297 +       J_ASSERT(list_empty(&new_inode->i_devices));
 298 +       list_add_tail(&new_inode->i_devices, &sbi->s_delete_list);
 299 +       sbi->s_delete_blocks += blocks;
 300 +       sbi->s_delete_inodes++;
 301 +       spin_unlock(&sbi->s_delete_lock);
 302 +
 303 +       ext3_debug("delete inode %lu (%lu blocks) by thread\n",
 304 +                  new_inode->i_ino, blocks);
 305 +
 306 +       wake_up(&sbi->s_delete_thread_queue);
 307 +       return;
 308 +
 309 +out_journal:
 310 +       ext3_journal_stop(handle, old_inode);
 311 +out_truncate:
 312 +       ext3_truncate(old_inode);
 313 +}
 314 +#endif /* EXT3_DELETE_THREAD */
 315 +
 316  /*
 317   * On success, We end up with an outstanding reference count against
 318   * iloc->bh.  This _must_ be cleaned up later.
 319 Index: linux-2.4.24/fs/ext3/file.c
 320 ===================================================================
 321 --- linux-2.4.24.orig/fs/ext3/file.c    2004-01-12 20:36:29.000000000 +0300
 322 +++ linux-2.4.24/fs/ext3/file.c 2004-01-12 20:36:32.000000000 +0300
 323 @@ -126,7 +126,11 @@
 324  };
 325
 326  struct inode_operations ext3_file_inode_operations = {
 327 +#ifdef EXT3_DELETE_THREAD
 328 +       truncate:       ext3_truncate_thread,   /* BKL held */
 329 +#else
 330         truncate:       ext3_truncate,          /* BKL held */
 331 +#endif
 332         setattr:        ext3_setattr,           /* BKL held */
 333         setxattr:       ext3_setxattr,          /* BKL held */
 334         getxattr:       ext3_getxattr,          /* BKL held */
 335 Index: linux-2.4.24/fs/ext3/namei.c
 336 ===================================================================
 337 --- linux-2.4.24.orig/fs/ext3/namei.c   2004-01-12 20:36:31.000000000 +0300
 338 +++ linux-2.4.24/fs/ext3/namei.c        2004-01-12 20:36:32.000000000 +0300
 339 @@ -1936,6 +1936,40 @@
 340         return retval;
 341  }
 342
 343 +#ifdef EXT3_DELETE_THREAD
 344 +static int ext3_try_to_delay_deletion(struct inode *inode)
 345 +{
 346 +       struct ext3_sb_info *sbi = EXT3_SB(inode->i_sb);
 347 +       struct ext3_inode_info *ei = EXT3_I(inode);
 348 +       unsigned long blocks;
 349 +
 350 +       if (!test_opt(inode->i_sb, ASYNCDEL))
 351 +               return 0;
 352 +
 353 +       /* We may want to delete the inode immediately and not defer it */
 354 +       blocks = inode->i_blocks >> (inode->i_blkbits - 9);
 355 +       if (IS_SYNC(inode) || blocks <= EXT3_NDIR_BLOCKS)
 356 +               return 0;
 357 +
 358 +       inode->i_nlink = 1;
 359 +       atomic_inc(&inode->i_count);
 360 +       ei->i_state |= EXT3_STATE_DELETE;
 361 +
 362 +       spin_lock(&sbi->s_delete_lock);
 363 +       J_ASSERT(list_empty(&inode->i_devices));
 364 +       list_add_tail(&inode->i_devices, &sbi->s_delete_list);
 365 +       sbi->s_delete_blocks += blocks;
 366 +       sbi->s_delete_inodes++;
 367 +       spin_unlock(&sbi->s_delete_lock);
 368 +
 369 +       wake_up(&sbi->s_delete_thread_queue);
 370 +
 371 +       return 0;
 372 +}
 373 +#else
 374 +#define ext3_try_to_delay_deletion(inode) do {} while (0)
 375 +#endif
 376 +
 377  static int ext3_unlink(struct inode * dir, struct dentry *dentry)
 378  {
 379         int retval;
 380 @@ -1977,8 +2007,10 @@
 381         ext3_update_dx_flag(dir);
 382         ext3_mark_inode_dirty(handle, dir);
 383         inode->i_nlink--;
 384 -       if (!inode->i_nlink)
 385 +       if (!inode->i_nlink) {
 386 +               ext3_try_to_delay_deletion(inode);
 387                 ext3_orphan_add(handle, inode);
 388 +       }
 389         inode->i_ctime = dir->i_ctime;
 390         ext3_mark_inode_dirty(handle, inode);
 391         retval = 0;
 392 Index: linux-2.4.24/include/linux/ext3_fs.h
 393 ===================================================================
 394 --- linux-2.4.24.orig/include/linux/ext3_fs.h   2004-01-12 20:36:31.000000000 +0300
 395 +++ linux-2.4.24/include/linux/ext3_fs.h        2004-01-12 20:36:32.000000000 +0300
 396 @@ -193,6 +193,7 @@
 397   */
 398  #define EXT3_STATE_JDATA               0x00000001 /* journaled data exists */
 399  #define EXT3_STATE_NEW                 0x00000002 /* inode is newly created */
 400 +#define EXT3_STATE_DELETE              0x00000010 /* deferred delete inode */
 401
 402  /*
 403   * ioctl commands
 404 @@ -320,6 +321,7 @@
 405  #define EXT3_MOUNT_UPDATE_JOURNAL      0x1000  /* Update the journal format */
 406  #define EXT3_MOUNT_NO_UID32            0x2000  /* Disable 32-bit UIDs */
 407  #define EXT3_MOUNT_XATTR_USER          0x4000  /* Extended user attributes */
 408 +#define EXT3_MOUNT_ASYNCDEL            0x20000 /* Delayed deletion */
 409
 410  /* Compatibility, for having both ext2_fs.h and ext3_fs.h included at once */
 411  #ifndef _LINUX_EXT2_FS_H
 412 @@ -697,6 +699,9 @@
 413  extern void ext3_dirty_inode(struct inode *);
 414  extern int ext3_change_inode_journal_flag(struct inode *, int);
 415  extern void ext3_truncate (struct inode *);
 416 +#ifdef EXT3_DELETE_THREAD
 417 +extern void ext3_truncate_thread(struct inode *inode);
 418 +#endif
 419  extern void ext3_set_inode_flags(struct inode *);
 420
 421  /* ioctl.c */
 422 Index: linux-2.4.24/include/linux/ext3_fs_sb.h
 423 ===================================================================
 424 --- linux-2.4.24.orig/include/linux/ext3_fs_sb.h        2004-01-12 20:36:31.000000000 +0300
 425 +++ linux-2.4.24/include/linux/ext3_fs_sb.h     2004-01-12 20:36:32.000000000 +0300
 426 @@ -29,6 +29,8 @@
 427
 428  #define EXT3_MAX_GROUP_LOADED  8
 429
 430 +#define EXT3_DELETE_THREAD
 431 +
 432  /*
 433   * third extended-fs super-block data in memory
 434   */
 435 @@ -76,6 +78,14 @@
 436         struct timer_list turn_ro_timer;        /* For turning read-only (crash simulation) */
 437         wait_queue_head_t ro_wait_queue;        /* For people waiting for the fs to go read-only */
 438  #endif
 439 +#ifdef EXT3_DELETE_THREAD
 440 +       spinlock_t s_delete_lock;
 441 +       struct list_head s_delete_list;
 442 +       unsigned long s_delete_blocks;
 443 +       unsigned long s_delete_inodes;
 444 +       wait_queue_head_t s_delete_thread_queue;
 445 +       wait_queue_head_t s_delete_waiter_queue;
 446 +#endif
 447  };
 448
 449  #endif /* _LINUX_EXT3_FS_SB */