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

   1  fs/ext3/ialloc.c        |   47 ++++++++++++++++++++++-
   2  fs/ext3/inode.c         |   96 +++++++++++++++++++++++++++++++++++++-----------
   3  include/linux/ext3_fs.h |    2 +
   4  3 files changed, 121 insertions(+), 24 deletions(-)
   5
   6 --- linux-2.4.20/fs/ext3/ialloc.c~ext3-noread-2.4.20    2003-05-16 12:21:39.000000000 +0800
   7 +++ linux-2.4.20-root/fs/ext3/ialloc.c  2003-05-16 12:21:46.000000000 +0800
   8 @@ -289,6 +289,37 @@ error_return:
   9  }
  10
  11  /*
  12 + * @block_group: block group of inode
  13 + * @offset: relative offset of inode within @block_group
  14 + *
  15 + * Check whether any of the inodes in this disk block are in use.
  16 + *
  17 + * Caller must be holding superblock lock (group/bitmap read lock in future).
  18 + */
  19 +int ext3_itable_block_used(struct super_block *sb, unsigned int block_group,
  20 +                          int offset)
  21 +{
  22 +       int bitmap_nr = load_inode_bitmap(sb, block_group);
  23 +       int inodes_per_block;
  24 +       unsigned long inum, iend;
  25 +       struct buffer_head *ibitmap;
  26 +
  27 +       if (bitmap_nr < 0)
  28 +               return 1;
  29 +
  30 +       inodes_per_block = sb->s_blocksize / EXT3_SB(sb)->s_inode_size;
  31 +       inum = offset & ~(inodes_per_block - 1);
  32 +       iend = inum + inodes_per_block;
  33 +       ibitmap = EXT3_SB(sb)->s_inode_bitmap[bitmap_nr];
  34 +       for (; inum < iend; inum++) {
  35 +               if (inum != offset && ext3_test_bit(inum, ibitmap->b_data))
  36 +                       return 1;
  37 +       }
  38 +
  39 +       return 0;
  40 +}
  41 +
  42 +/*
  43   * There are two policies for allocating an inode.  If the new inode is
  44   * a directory, then a forward search is made for a block group with both
  45   * free space and a low directory-to-inode ratio; if that fails, then of
  46 @@ -310,6 +341,7 @@ struct inode * ext3_new_inode (handle_t
  47         struct ext3_group_desc * gdp;
  48         struct ext3_group_desc * tmp;
  49         struct ext3_super_block * es;
  50 +       struct ext3_iloc iloc;
  51         int err = 0;
  52
  53         /* Cannot create files in a deleted directory */
  54 @@ -510,8 +542,19 @@ repeat:
  55         inode->i_generation = sb->u.ext3_sb.s_next_generation++;
  56
  57         inode->u.ext3_i.i_state = EXT3_STATE_NEW;
  58 -       err = ext3_mark_inode_dirty(handle, inode);
  59 -       if (err) goto fail;
  60 +       err = ext3_get_inode_loc_new(inode, &iloc, 1);
  61 +       if (err) goto fail;
  62 +       BUFFER_TRACE(iloc->bh, "get_write_access");
  63 +       err = ext3_journal_get_write_access(handle, iloc.bh);
  64 +       if (err) {
  65 +               brelse(iloc.bh);
  66 +               iloc.bh = NULL;
  67 +               goto fail;
  68 +       }
  69 +       err = ext3_mark_iloc_dirty(handle, inode, &iloc);
  70 +       if (err) goto fail;
  71 +
  72 +
  73
  74         unlock_super (sb);
  75         if(DQUOT_ALLOC_INODE(inode)) {
  76 --- linux-2.4.20/fs/ext3/inode.c~ext3-noread-2.4.20     2003-05-16 12:21:41.000000000 +0800
  77 +++ linux-2.4.20-root/fs/ext3/inode.c   2003-05-16 12:22:15.000000000 +0800
  78 @@ -2013,14 +2013,19 @@ out_stop:
  79         ext3_journal_stop(handle, inode);
  80  }
  81
  82 -/*
  83 - * ext3_get_inode_loc returns with an extra refcount against the
  84 - * inode's underlying buffer_head on success.
  85 - */
  86 -
  87 -int ext3_get_inode_loc (struct inode *inode, struct ext3_iloc *iloc)
  88 +#define NUM_INODE_PREREAD      16
  89 +
  90 +/*
  91 + * ext3_get_inode_loc returns with an extra refcount against the inode's
  92 + * underlying buffer_head on success.  If this is for a new inode allocation
  93 + * (new is non-zero) then we may be able to optimize away the read if there
  94 + * are no other in-use inodes in this inode table block.  If we need to do
  95 + * a read, then read in a whole chunk of blocks to avoid blocking again soon
  96 + * if we are doing lots of creates/updates.
  97 + */
  98 +int ext3_get_inode_loc_new(struct inode *inode, struct ext3_iloc *iloc, int new)
  99  {
 100 -       struct buffer_head *bh = 0;
 101 +       struct buffer_head *bh[NUM_INODE_PREREAD];
 102         unsigned long block;
 103         unsigned long block_group;
 104         unsigned long group_desc;
 105 @@ -2045,31 +2050,73 @@ int ext3_get_inode_loc (struct inode *in
 106         }
 107         group_desc = block_group >> EXT3_DESC_PER_BLOCK_BITS(inode->i_sb);
 108         desc = block_group & (EXT3_DESC_PER_BLOCK(inode->i_sb) - 1);
 109 -       bh = inode->i_sb->u.ext3_sb.s_group_desc[group_desc];
 110 -       if (!bh) {
 111 +       if (!(inode->i_sb->u.ext3_sb.s_group_desc[group_desc])) {
 112                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
 113                             "Descriptor not loaded");
 114                 goto bad_inode;
 115         }
 116
 117 -       gdp = (struct ext3_group_desc *) bh->b_data;
 118 +       gdp = (struct ext3_group_desc *)(inode->i_sb->u.ext3_sb.s_group_desc[group_desc]->b_data);
 119         /*
 120          * Figure out the offset within the block group inode table
 121          */
 122 -       offset = ((inode->i_ino - 1) % EXT3_INODES_PER_GROUP(inode->i_sb)) *
 123 -               EXT3_INODE_SIZE(inode->i_sb);
 124 +       offset = ((inode->i_ino - 1) % EXT3_INODES_PER_GROUP(inode->i_sb));
 125 +
 126         block = le32_to_cpu(gdp[desc].bg_inode_table) +
 127 -               (offset >> EXT3_BLOCK_SIZE_BITS(inode->i_sb));
 128 -       if (!(bh = sb_bread(inode->i_sb, block))) {
 129 -               ext3_error (inode->i_sb, "ext3_get_inode_loc",
 130 -                           "unable to read inode block - "
 131 -                           "inode=%lu, block=%lu", inode->i_ino, block);
 132 -               goto bad_inode;
 133 -       }
 134 -       offset &= (EXT3_BLOCK_SIZE(inode->i_sb) - 1);
 135 +               (offset * EXT3_INODE_SIZE(inode->i_sb) >> EXT3_BLOCK_SIZE_BITS(inode->i_sb));
 136
 137 -       iloc->bh = bh;
 138 -       iloc->raw_inode = (struct ext3_inode *) (bh->b_data + offset);
 139 +       bh[0] = sb_getblk(inode->i_sb, block);
 140 +       if (buffer_uptodate(bh[0]))
 141 +               goto done;
 142 +
 143 +       /* If we don't really need to read this block, and it isn't already
 144 +        * in memory, then we just zero it out.  Otherwise, we keep the
 145 +        * current block contents (deleted inode data) for posterity.
 146 +        */
 147 +       if (new && !ext3_itable_block_used(inode->i_sb, block_group, offset)) {
 148 +               lock_buffer(bh[0]);
 149 +               memset(bh[0]->b_data, 0, bh[0]->b_size);
 150 +               mark_buffer_uptodate(bh[0], 1);
 151 +               unlock_buffer(bh[0]);
 152 +       } else {
 153 +               unsigned long block_end, itable_end;
 154 +               int count = 1;
 155 +
 156 +               itable_end = le32_to_cpu(gdp[desc].bg_inode_table) +
 157 +                               inode->i_sb->u.ext3_sb.s_itb_per_group;
 158 +               block_end = block + NUM_INODE_PREREAD;
 159 +               if (block_end > itable_end)
 160 +                       block_end = itable_end;
 161 +
 162 +               for (++block; block < block_end; block++) {
 163 +                       bh[count] = sb_getblk(inode->i_sb, block);
 164 +                       if (count && (buffer_uptodate(bh[count]) ||
 165 +                                     buffer_locked(bh[count]))) {
 166 +                               __brelse(bh[count]);
 167 +                       } else
 168 +                               count++;
 169 +               }
 170 +
 171 +               ll_rw_block(READ, count, bh);
 172 +
 173 +               /* Release all but the block we actually need (bh[0]) */
 174 +               while (--count > 0)
 175 +                       __brelse(bh[count]);
 176 +
 177 +               wait_on_buffer(bh[0]);
 178 +               if (!buffer_uptodate(bh[0])) {
 179 +                       ext3_error(inode->i_sb, __FUNCTION__,
 180 +                                  "unable to read inode block - "
 181 +                                  "inode=%lu, block=%lu", inode->i_ino,
 182 +                                  bh[0]->b_blocknr);
 183 +                       goto bad_inode;
 184 +               }
 185 +       }
 186 +done:
 187 +       offset = (offset * EXT3_INODE_SIZE(inode->i_sb)) & (EXT3_BLOCK_SIZE(inode->i_sb) - 1);
 188 +
 189 +       iloc->bh = bh[0];
 190 +       iloc->raw_inode = (struct ext3_inode *)(bh[0]->b_data + offset);
 191         iloc->block_group = block_group;
 192
 193         return 0;
 194 @@ -2078,6 +2125,11 @@ int ext3_get_inode_loc (struct inode *in
 195         return -EIO;
 196  }
 197
 198 +int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
 199 +{
 200 +       return ext3_get_inode_loc_new(inode, iloc, 0);
 201 +}
 202 +
 203  void ext3_read_inode(struct inode * inode)
 204  {
 205         struct ext3_iloc iloc;
 206 --- linux-2.4.20/include/linux/ext3_fs.h~ext3-noread-2.4.20     2003-05-16 12:21:39.000000000 +0800
 207 +++ linux-2.4.20-root/include/linux/ext3_fs.h   2003-05-16 12:21:46.000000000 +0800
 208 @@ -683,6 +683,8 @@ extern int ext3_forget(handle_t *, int,
 209  extern struct buffer_head * ext3_getblk (handle_t *, struct inode *, long, int, int *);
 210  extern struct buffer_head * ext3_bread (handle_t *, struct inode *, int, int, int *);
 211
 212 +extern int ext3_itable_block_used(struct super_block *sb, unsigned int, int);
 213 +extern int ext3_get_inode_loc_new(struct inode *, struct ext3_iloc *, int);
 214  extern int  ext3_get_inode_loc (struct inode *, struct ext3_iloc *);
 215  extern void ext3_read_inode (struct inode *);
 216  extern void ext3_write_inode (struct inode *, int);
 217
 218 _