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

   1  0 files changed
   2
   3 --- linux-2.4.18-chaos52/fs/ext3/super.c~ext3-delete_thread-2.4.18      2003-06-01 03:24:13.000000000 +0800
   4 +++ linux-2.4.18-chaos52-root/fs/ext3/super.c   2003-06-03 17:01:49.000000000 +0800
   5 @@ -398,6 +398,210 @@ static void dump_orphan_list(struct supe
   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("EXT3-fs: delete thread on %s started\n",
  40 +              kdevname(sb->s_dev));
  41 +
  42 +       /* main loop */
  43 +       for (;;) {
  44 +               sleep_on(&sbi->s_delete_thread_queue);
  45 +               ext3_debug("%s woken up: %lu inodes, %lu blocks\n",
  46 +                          tsk->comm,sbi->s_delete_inodes,sbi->s_delete_blocks);
  47 +
  48 +               spin_lock(&sbi->s_delete_lock);
  49 +               if (list_empty(&sbi->s_delete_list)) {
  50 +                       memset(&sbi->s_delete_list, 0,
  51 +                              sizeof(sbi->s_delete_list));
  52 +                       spin_unlock(&sbi->s_delete_lock);
  53 +                       ext3_debug("ext3 delete thread on %s exiting\n",
  54 +                              kdevname(sb->s_dev));
  55 +                       wake_up(&sbi->s_delete_waiter_queue);
  56 +                       break;
  57 +               }
  58 +
  59 +               while (!list_empty(&sbi->s_delete_list)) {
  60 +                       struct inode *inode=list_entry(sbi->s_delete_list.next,
  61 +                                                      struct inode, i_dentry);
  62 +                       unsigned long blocks = inode->i_blocks >>
  63 +                                                       (inode->i_blkbits - 9);
  64 +
  65 +                       list_del_init(&inode->i_dentry);
  66 +                       spin_unlock(&sbi->s_delete_lock);
  67 +                       ext3_debug("%s delete ino %lu blk %lu\n",
  68 +                                  tsk->comm, inode->i_ino, blocks);
  69 +
  70 +                       iput(inode);
  71 +
  72 +                       spin_lock(&sbi->s_delete_lock);
  73 +                       sbi->s_delete_blocks -= blocks;
  74 +                       sbi->s_delete_inodes--;
  75 +               }
  76 +               if (sbi->s_delete_blocks != 0 || sbi->s_delete_inodes != 0)
  77 +                       ext3_warning(sb, __FUNCTION__,
  78 +                                    "%lu blocks, %lu inodes on list?\n",
  79 +                                    sbi->s_delete_blocks,sbi->s_delete_inodes);
  80 +               sbi->s_delete_blocks = 0;
  81 +               sbi->s_delete_inodes = 0;
  82 +               spin_unlock(&sbi->s_delete_lock);
  83 +               wake_up(&sbi->s_delete_waiter_queue);
  84 +       }
  85 +
  86 +       return 0;
  87 +}
  88 +
  89 +static void ext3_start_delete_thread(struct super_block *sb)
  90 +{
  91 +       struct ext3_sb_info *sbi = EXT3_SB(sb);
  92 +       int rc;
  93 +
  94 +       spin_lock_init(&sbi->s_delete_lock);
  95 +       memset(&sbi->s_delete_list, 0, sizeof(sbi->s_delete_list));
  96 +       init_waitqueue_head(&sbi->s_delete_thread_queue);
  97 +       init_waitqueue_head(&sbi->s_delete_waiter_queue);
  98 +       sbi->s_delete_blocks = 0;
  99 +       sbi->s_delete_inodes = 0;
 100 +
 101 +       rc = kernel_thread(ext3_delete_thread, sb, CLONE_VM | CLONE_FILES);
 102 +       if (rc < 0)
 103 +               printk(KERN_ERR "EXT3-fs: cannot start delete thread: rc %d\n",
 104 +                      rc);
 105 +       else
 106 +               wait_event(sbi->s_delete_waiter_queue, sbi->s_delete_list.next);
 107 +}
 108 +
 109 +static void ext3_stop_delete_thread(struct ext3_sb_info *sbi)
 110 +{
 111 +       wake_up(&sbi->s_delete_thread_queue);
 112 +       wait_event(sbi->s_delete_waiter_queue, list_empty(&sbi->s_delete_list));
 113 +}
 114 +
 115 +/* Instead of playing games with the inode flags, destruction, etc we just
 116 + * create a new inode locally and put it on a list for the truncate thread.
 117 + * We need large parts of the inode struct in order to complete the
 118 + * truncate and unlink, so we may as well just have a real inode to do it.
 119 + *
 120 + * If we have any problem deferring the delete, just delete it right away.
 121 + * If we defer it, we also mark how many blocks it would free, so that we
 122 + * can keep the statfs data correct, and we know if we should sleep on the
 123 + * truncate thread when we run out of space.
 124 + *
 125 + * In 2.5 this can be done much more cleanly by just registering a "drop"
 126 + * method in the super_operations struct.
 127 + */
 128 +static void ext3_delete_inode_thread(struct inode *old_inode)
 129 +{
 130 +       struct ext3_sb_info *sbi = EXT3_SB(old_inode->i_sb);
 131 +       struct inode *new_inode;
 132 +       unsigned long blocks = old_inode->i_blocks >> (old_inode->i_blkbits-9);
 133 +
 134 +       if (is_bad_inode(old_inode)) {
 135 +               clear_inode(old_inode);
 136 +               return;
 137 +       }
 138 +
 139 +       if (!test_opt (old_inode->i_sb, ASYNCDEL)) {
 140 +               ext3_delete_inode(old_inode);
 141 +               return;
 142 +       }
 143 +
 144 +       /* We may want to delete the inode immediately and not defer it */
 145 +       if (IS_SYNC(old_inode) || blocks <= EXT3_NDIR_BLOCKS ||
 146 +           !sbi->s_delete_list.next) {
 147 +               ext3_delete_inode(old_inode);
 148 +               return;
 149 +       }
 150 +
 151 +       if (EXT3_I(old_inode)->i_state & EXT3_STATE_DELETE) {
 152 +               ext3_debug("doing deferred inode %lu delete (%lu blocks)\n",
 153 +                          old_inode->i_ino, blocks);
 154 +               ext3_delete_inode(old_inode);
 155 +               return;
 156 +       }
 157 +
 158 +       /* We can iget this inode again here, because our caller has unhashed
 159 +        * old_inode, so new_inode will be in a different inode struct.
 160 +        *
 161 +        * We need to ensure that the i_orphan pointers in the other inodes
 162 +        * point at the new inode copy instead of the old one so the orphan
 163 +        * list doesn't get corrupted when the old orphan inode is freed.
 164 +        */
 165 +       down(&sbi->s_orphan_lock);
 166 +
 167 +       EXT3_SB(old_inode->i_sb)->s_mount_state |= EXT3_ORPHAN_FS;
 168 +       new_inode = iget(old_inode->i_sb, old_inode->i_ino);
 169 +       EXT3_SB(old_inode->i_sb)->s_mount_state &= ~EXT3_ORPHAN_FS;
 170 +       if (is_bad_inode(new_inode)) {
 171 +               printk(KERN_WARNING "read bad inode %lu\n", old_inode->i_ino);
 172 +               iput(new_inode);
 173 +               new_inode = NULL;
 174 +       }
 175 +       if (!new_inode) {
 176 +               up(&sbi->s_orphan_lock);
 177 +               ext3_debug(KERN_DEBUG "delete inode %lu directly (bad read)\n",
 178 +                          old_inode->i_ino);
 179 +               ext3_delete_inode(old_inode);
 180 +               return;
 181 +       }
 182 +       J_ASSERT(new_inode != old_inode);
 183 +
 184 +       J_ASSERT(!list_empty(&EXT3_I(old_inode)->i_orphan));
 185 +       /* Ugh.  We need to insert new_inode into the same spot on the list
 186 +        * as old_inode was, to ensure the in-memory orphan list is still
 187 +        * in the same order as the on-disk orphan list (badness otherwise).
 188 +        */
 189 +       EXT3_I(new_inode)->i_orphan = EXT3_I(old_inode)->i_orphan;
 190 +       EXT3_I(new_inode)->i_orphan.next->prev = &EXT3_I(new_inode)->i_orphan;
 191 +       EXT3_I(new_inode)->i_orphan.prev->next = &EXT3_I(new_inode)->i_orphan;
 192 +       EXT3_I(new_inode)->i_state |= EXT3_STATE_DELETE;
 193 +       up(&sbi->s_orphan_lock);
 194 +
 195 +       clear_inode(old_inode);
 196 +
 197 +       ext3_debug("delete inode %lu (%lu blocks) by thread\n",
 198 +                  new_inode->i_ino, blocks);
 199 +       spin_lock(&sbi->s_delete_lock);
 200 +       J_ASSERT(list_empty(&new_inode->i_dentry));
 201 +       list_add_tail(&new_inode->i_dentry, &sbi->s_delete_list);
 202 +       sbi->s_delete_blocks += blocks;
 203 +       sbi->s_delete_inodes++;
 204 +       spin_unlock(&sbi->s_delete_lock);
 205 +
 206 +       wake_up(&sbi->s_delete_thread_queue);
 207 +}
 208 +#else
 209 +#define ext3_start_delete_thread(sbi) do {} while(0)
 210 +#define ext3_stop_delete_thread(sbi) do {} while(0)
 211 +#endif /* EXT3_DELETE_THREAD */
 212 +
 213  void ext3_put_super (struct super_block * sb)
 214  {
 215         struct ext3_sb_info *sbi = EXT3_SB(sb);
 216 @@ -405,6 +609,7 @@ void ext3_put_super (struct super_block
 217         kdev_t j_dev = sbi->s_journal->j_dev;
 218         int i;
 219
 220 +       ext3_stop_delete_thread(sbi);
 221         ext3_xattr_put_super(sb);
 222         journal_destroy(sbi->s_journal);
 223         if (!(sb->s_flags & MS_RDONLY)) {
 224 @@ -453,7 +658,11 @@ static struct super_operations ext3_sops
 225         write_inode:    ext3_write_inode,       /* BKL not held.  Don't need */
 226         dirty_inode:    ext3_dirty_inode,       /* BKL not held.  We take it */
 227         put_inode:      ext3_put_inode,         /* BKL not held.  Don't need */
 228 +#ifdef EXT3_DELETE_THREAD
 229 +       delete_inode:   ext3_delete_inode_thread,/* BKL not held. We take it */
 230 +#else
 231         delete_inode:   ext3_delete_inode,      /* BKL not held.  We take it */
 232 +#endif
 233         put_super:      ext3_put_super,         /* BKL held */
 234         write_super:    ext3_write_super,       /* BKL held */
 235         sync_fs:        ext3_sync_fs,
 236 @@ -514,6 +723,12 @@ static int parse_options (char * options
 237              this_char = strtok (NULL, ",")) {
 238                 if ((value = strchr (this_char, '=')) != NULL)
 239                         *value++ = 0;
 240 +#ifdef EXT3_DELETE_THREAD
 241 +               if (!strcmp(this_char, "asyncdel"))
 242 +                       set_opt(*mount_options, ASYNCDEL);
 243 +               else
 244 +#endif
 245 +
 246                 if (!strcmp (this_char, "bsddf"))
 247                         clear_opt (*mount_options, MINIX_DF);
 248                 else if (!strcmp (this_char, "nouid32")) {
 249 @@ -1209,6 +1424,7 @@ struct super_block * ext3_read_super (st
 250         }
 251
 252         ext3_setup_super (sb, es, sb->s_flags & MS_RDONLY);
 253 +       ext3_start_delete_thread(sb);
 254         /*
 255          * akpm: core read_super() calls in here with the superblock locked.
 256          * That deadlocks, because orphan cleanup needs to lock the superblock
 257 --- linux-2.4.18-chaos52/include/linux/ext3_fs.h~ext3-delete_thread-2.4.18      2003-06-01 03:24:11.000000000 +0800
 258 +++ linux-2.4.18-chaos52-root/include/linux/ext3_fs.h   2003-06-03 17:03:28.000000000 +0800
 259 @@ -190,6 +190,7 @@ struct ext3_group_desc
 260   */
 261  #define EXT3_STATE_JDATA               0x00000001 /* journaled data exists */
 262  #define EXT3_STATE_NEW                 0x00000002 /* inode is newly created */
 263 +#define EXT3_STATE_DELETE              0x00000010 /* deferred delete inode */
 264
 265  /*
 266   * ioctl commands
 267 @@ -317,6 +318,7 @@ struct ext3_inode {
 268  #define EXT3_MOUNT_UPDATE_JOURNAL      0x1000  /* Update the journal format */
 269  #define EXT3_MOUNT_NO_UID32            0x2000  /* Disable 32-bit UIDs */
 270  #define EXT3_MOUNT_INDEX               0x4000  /* Enable directory index */
 271 +#define EXT3_MOUNT_ASYNCDEL            0x20000 /* Delayed deletion */
 272
 273  /* Compatibility, for having both ext2_fs.h and ext3_fs.h included at once */
 274  #ifndef _LINUX_EXT2_FS_H
 275 --- linux-2.4.18-chaos52/include/linux/ext3_fs_sb.h~ext3-delete_thread-2.4.18   2003-06-01 03:24:13.000000000 +0800
 276 +++ linux-2.4.18-chaos52-root/include/linux/ext3_fs_sb.h        2003-06-03 16:59:24.000000000 +0800
 277 @@ -29,6 +29,8 @@
 278
 279  #define EXT3_MAX_GROUP_LOADED  32
 280
 281 +#define EXT3_DELETE_THREAD
 282 +
 283  /*
 284   * third extended-fs super-block data in memory
 285   */
 286 @@ -74,6 +76,14 @@ struct ext3_sb_info {
 287         struct timer_list turn_ro_timer;        /* For turning read-only (crash simulation) */
 288         wait_queue_head_t ro_wait_queue;        /* For people waiting for the fs to go read-only */
 289  #endif
 290 +#ifdef EXT3_DELETE_THREAD
 291 +       spinlock_t s_delete_lock;
 292 +       struct list_head s_delete_list;
 293 +       unsigned long s_delete_blocks;
 294 +       unsigned long s_delete_inodes;
 295 +       wait_queue_head_t s_delete_thread_queue;
 296 +       wait_queue_head_t s_delete_waiter_queue;
 297 +#endif
 298  };
 299
 300  #endif /* _LINUX_EXT3_FS_SB */
 301
 302 _