ldiskfs/kernel_patches/patches/ext3-mballoc3-sles10.patch

   1 Index: linux-2.6.16.46-0.14/include/linux/ext3_fs_i.h
   2 ===================================================================
   3 --- linux-2.6.16.46-0.14.orig/include/linux/ext3_fs_i.h
   4 +++ linux-2.6.16.46-0.14/include/linux/ext3_fs_i.h
   5 @@ -135,6 +135,10 @@ struct ext3_inode_info {
   6         struct inode vfs_inode;
   7
   8         __u32 i_cached_extent[4];
   9 +
  10 +       /* mballoc */
  11 +       struct list_head i_prealloc_list;
  12 +       spinlock_t i_prealloc_lock;
  13  };
  14
  15  #endif /* _LINUX_EXT3_FS_I */
  16 Index: linux-2.6.16.46-0.14/include/linux/ext3_fs_sb.h
  17 ===================================================================
  18 --- linux-2.6.16.46-0.14.orig/include/linux/ext3_fs_sb.h
  19 +++ linux-2.6.16.46-0.14/include/linux/ext3_fs_sb.h
  20 @@ -21,8 +21,15 @@
  21  #include <linux/wait.h>
  22  #include <linux/blockgroup_lock.h>
  23  #include <linux/percpu_counter.h>
  24 +#include <linux/list.h>
  25  #endif
  26  #include <linux/rbtree.h>
  27 +#include <linux/proc_fs.h>
  28 +
  29 +struct ext3_buddy_group_blocks;
  30 +struct ext3_locality_group;
  31 +struct ext3_mb_history;
  32 +#define EXT3_BB_MAX_BLOCKS
  33
  34  /*
  35   * third extended-fs super-block data in memory
  36 Index: linux-2.6.16.46-0.14/include/linux/ext3_fs.h
  37 ===================================================================
  38 --- linux-2.6.16.46-0.14.orig/include/linux/ext3_fs.h
  39 +++ linux-2.6.16.46-0.14/include/linux/ext3_fs.h
  40 @@ -407,6 +407,7 @@ struct ext3_inode {
  41  #define EXT3_MOUNT_IOPEN_NOPRIV                0x800000/* Make iopen world-readable */
  42  #define EXT3_MOUNT_EXTENTS             0x1000000/* Extents support */
  43  #define EXT3_MOUNT_EXTDEBUG            0x2000000/* Extents debug */
  44 +#define EXT3_MOUNT_MBALLOC             0x4000000/* Buddy allocation support */
  45
  46  /* Compatibility, for having both ext2_fs.h and ext3_fs.h included at once */
  47  #ifndef clear_opt
  48 @@ -784,8 +785,9 @@ struct dir_private_info {
  49  extern int ext3_bg_has_super(struct super_block *sb, int group);
  50  extern unsigned long ext3_bg_num_gdb(struct super_block *sb, int group);
  51  extern int ext3_new_block (handle_t *, struct inode *, unsigned long, int *);
  52 +extern int ext3_new_block_old (handle_t *, struct inode *, unsigned long, int *);
  53  extern void ext3_free_blocks (handle_t *, struct inode *, unsigned long,
  54 -                             unsigned long);
  55 +                             unsigned long, int);
  56  extern void ext3_free_blocks_sb (handle_t *, struct super_block *,
  57                                  unsigned long, unsigned long, int *);
  58  extern unsigned long ext3_count_free_blocks (struct super_block *);
  59 Index: linux-2.6.16.46-0.14/fs/ext3/super.c
  60 ===================================================================
  61 --- linux-2.6.16.46-0.14.orig/fs/ext3/super.c
  62 +++ linux-2.6.16.46-0.14/fs/ext3/super.c
  63 @@ -681,6 +681,7 @@ enum {
  64         Opt_iopen, Opt_noiopen, Opt_iopen_nopriv,
  65         Opt_grpquota,
  66         Opt_extents, Opt_noextents, Opt_extdebug,
  67 +       Opt_mballoc, Opt_nomballoc, Opt_stripe,
  68  };
  69
  70  static match_table_t tokens = {
  71 @@ -740,6 +741,9 @@ static match_table_t tokens = {
  72         {Opt_extents, "extents"},
  73         {Opt_noextents, "noextents"},
  74         {Opt_extdebug, "extdebug"},
  75 +       {Opt_mballoc, "mballoc"},
  76 +       {Opt_nomballoc, "nomballoc"},
  77 +       {Opt_stripe, "stripe=%u"},
  78         {Opt_err, NULL},
  79         {Opt_resize, "resize"},
  80  };
  81 @@ -1077,6 +1081,19 @@ clear_qf_name:
  82                 case Opt_extdebug:
  83                         set_opt (sbi->s_mount_opt, EXTDEBUG);
  84                         break;
  85 +               case Opt_mballoc:
  86 +                       set_opt(sbi->s_mount_opt, MBALLOC);
  87 +                       break;
  88 +               case Opt_nomballoc:
  89 +                       clear_opt(sbi->s_mount_opt, MBALLOC);
  90 +                       break;
  91 +               case Opt_stripe:
  92 +                       if (match_int(&args[0], &option))
  93 +                               return 0;
  94 +                       if (option < 0)
  95 +                               return 0;
  96 +                       sbi->s_stripe = option;
  97 +                       break;
  98                 default:
  99                         printk (KERN_ERR
 100                                 "EXT3-fs: Unrecognized mount option \"%s\" "
 101 @@ -1858,7 +1858,7 @@ static int ext3_fill_super (struct super
 102                 ext3_count_dirs(sb));
 103
 104         ext3_ext_init(sb);
 105 -
 106 +       ext3_mb_init(sb, needs_recovery);
 107         lock_kernel();
 108         return 0;
 109
 110 Index: linux-2.6.16.46-0.14/fs/ext3/extents.c
 111 ===================================================================
 112 --- linux-2.6.16.46-0.14.orig/fs/ext3/extents.c
 113 +++ linux-2.6.16.46-0.14/fs/ext3/extents.c
 114 @@ -819,7 +819,7 @@ cleanup:
 115                 for (i = 0; i < depth; i++) {
 116                         if (!ablocks[i])
 117                                 continue;
 118 -                       ext3_free_blocks(handle, tree->inode, ablocks[i], 1);
 119 +                       ext3_free_blocks(handle, tree->inode, ablocks[i], 1, 1);
 120                 }
 121         }
 122         kfree(ablocks);
 123 @@ -1629,7 +1629,7 @@ int ext3_ext_rm_idx(handle_t *handle, st
 124                   path->p_idx->ei_leaf);
 125         bh = sb_find_get_block(tree->inode->i_sb, path->p_idx->ei_leaf);
 126         ext3_forget(handle, 1, tree->inode, bh, path->p_idx->ei_leaf);
 127 -       ext3_free_blocks(handle, tree->inode, path->p_idx->ei_leaf, 1);
 128 +       ext3_free_blocks(handle, tree->inode, path->p_idx->ei_leaf, 1, 1);
 129         return err;
 130  }
 131
 132 @@ -2128,10 +2128,12 @@ ext3_remove_blocks(struct ext3_extents_t
 133         int needed = ext3_remove_blocks_credits(tree, ex, from, to);
 134         handle_t *handle = ext3_journal_start(tree->inode, needed);
 135         struct buffer_head *bh;
 136 -       int i;
 137 +       int i, metadata = 0;
 138
 139         if (IS_ERR(handle))
 140                 return PTR_ERR(handle);
 141 +       if (S_ISDIR(tree->inode->i_mode) || S_ISLNK(tree->inode->i_mode))
 142 +               metadata = 1;
 143         if (from >= ex->ee_block && to == ex->ee_block + ex->ee_len - 1) {
 144                 /* tail removal */
 145                 unsigned long num, start;
 146 @@ -2143,7 +2145,7 @@ ext3_remove_blocks(struct ext3_extents_t
 147                         bh = sb_find_get_block(tree->inode->i_sb, start + i);
 148                         ext3_forget(handle, 0, tree->inode, bh, start + i);
 149                 }
 150 -               ext3_free_blocks(handle, tree->inode, start, num);
 151 +               ext3_free_blocks(handle, tree->inode, start, num, metadata);
 152         } else if (from == ex->ee_block && to <= ex->ee_block + ex->ee_len - 1) {
 153                 printk("strange request: removal %lu-%lu from %u:%u\n",
 154                        from, to, ex->ee_block, ex->ee_len);
 155 @@ -2234,11 +2236,8 @@ int ext3_ext_get_block(handle_t *handle,
 156         struct ext3_extent *ex;
 157         int goal, newblock, err = 0, depth;
 158         struct ext3_extents_tree tree;
 159 -       unsigned long next;
 160 -       int allocated = 0;
 161 -
 162 -       /* until we have multiblock allocation */
 163 -       max_blocks = 1;
 164 +       unsigned long allocated = 0;
 165 +       struct ext3_allocation_request ar;
 166
 167         clear_buffer_new(bh_result);
 168         ext3_init_tree_desc(&tree, inode);
 169 @@ -2310,18 +2309,36 @@ int ext3_ext_get_block(handle_t *handle,
 170                 goto out2;
 171         }
 172
 173 +       /* find neighbour allocated blocks */
 174 +       ar.lleft = iblock;
 175 +       err = ext3_ext_search_left(&tree, path, &ar.lleft, &ar.pleft);
 176 +       if (err)
 177 +               goto out2;
 178 +       ar.lright = iblock;
 179 +       err = ext3_ext_search_right(&tree, path, &ar.lright, &ar.pright);
 180 +       if (err)
 181 +               goto out2;
 182 +
 183         /* find next allocated block so that we know how many
 184          * blocks we can allocate without ovelapping next extent */
 185 -       EXT_ASSERT(iblock >= ex->ee_block + ex->ee_len);
 186 -       next = ext3_ext_next_allocated_block(path);
 187 -       EXT_ASSERT(next > iblock);
 188 -       allocated = next - iblock;
 189 +       EXT_ASSERT(ar.pright == 0 || ar.lright > iblock);
 190 +       if (ar.pright == 0)
 191 +               allocated = EXT_MAX_BLOCK - iblock;
 192 +       else
 193 +               allocated = ar.lright - iblock;
 194         if (allocated > max_blocks)
 195                 allocated = max_blocks;
 196
 197         /* allocate new block */
 198 -       goal = ext3_ext_find_goal(inode, path, iblock);
 199 -       newblock = ext3_new_block(handle, inode, goal, &err);
 200 +       ar.inode = inode;
 201 +       ar.goal = ext3_ext_find_goal(inode, path, iblock);
 202 +       ar.logical = iblock;
 203 +       ar.len = allocated;
 204 +       if (S_ISREG(inode->i_mode))
 205 +               ar.flags = EXT3_MB_HINT_DATA;
 206 +       else
 207 +               ar.flags = 0;
 208 +       newblock = ext3_mb_new_blocks(handle, &ar, &err);
 209         if (!newblock)
 210                 goto out2;
 211         ext_debug(&tree, "allocate new block: goal %d, found %d\n",
 212 @@ -2331,11 +2348,14 @@ int ext3_ext_get_block(handle_t *handle,
 213         newex.ee_block = iblock;
 214         newex.ee_start = newblock;
 215         newex.ee_start_hi = 0;
 216 -       newex.ee_len = 1;
 217 +       newex.ee_len = ar.len;
 218         err = ext3_ext_insert_extent(handle, &tree, path, &newex);
 219         if (err) {
 220                 /* free data blocks we just allocated */
 221 -               ext3_free_blocks(handle, inode, newex.ee_start, newex.ee_len);
 222 +               /* not a good idea to call discard here directly,
 223 +                * but otherwise we'd need to call it every free() */
 224 +               ext3_mb_discard_inode_preallocations(inode);
 225 +               ext3_free_blocks(handle, inode, newex.ee_start, newex.ee_len, 0);
 226                 goto out2;
 227         }
 228
 229 @@ -2344,6 +2364,7 @@ int ext3_ext_get_block(handle_t *handle,
 230
 231         /* previous routine could use block we allocated */
 232         newblock = newex.ee_start;
 233 +       allocated = newex.ee_len;
 234         set_buffer_new(bh_result);
 235
 236         ext3_ext_put_in_cache(&tree, newex.ee_block, newex.ee_len,
 237 @@ -2396,6 +2417,9 @@ void ext3_ext_truncate(struct inode * in
 238         down(&EXT3_I(inode)->truncate_sem);
 239         ext3_ext_invalidate_cache(&tree);
 240
 241 +       /* it's important to discard preallocations under truncate_sem */
 242 +       ext3_mb_discard_inode_preallocations(inode);
 243 +
 244         /*
 245          * TODO: optimization is possible here
 246          * probably we need not scaning at all,
 247 Index: linux-2.6.18.8/fs/ext3/Makefile
 248 ===================================================================
 249 --- linux-2.6.18.8.orig/fs/ext3/Makefile
 250 +++ linux-2.6.18.8/fs/ext3/Makefile
 251 @@ -5,7 +5,7 @@
 252  obj-$(CONFIG_EXT3_FS) += ext3.o
 253
 254  ext3-y := balloc.o bitmap.o dir.o file.o fsync.o ialloc.o inode.o iopen.o \
 255 -          ioctl.o namei.o super.o symlink.o hash.o resize.o extents.o
 256 +          ioctl.o namei.o super.o symlink.o hash.o resize.o extents.o mballoc.o
 257
 258  ext3-$(CONFIG_EXT3_FS_XATTR)    += xattr.o xattr_user.o xattr_trusted.o
 259  ext3-$(CONFIG_EXT3_FS_POSIX_ACL) += acl.o
 260 Index: linux-2.6.16.46-0.14/fs/ext3/xattr.c
 261 ===================================================================
 262 --- linux-2.6.16.46-0.14.orig/fs/ext3/xattr.c
 263 +++ linux-2.6.16.46-0.14/fs/ext3/xattr.c
 264 @@ -484,7 +484,7 @@ ext3_xattr_release_block(handle_t *handl
 265                 ea_bdebug(bh, "refcount now=0; freeing");
 266                 if (ce)
 267                         mb_cache_entry_free(ce);
 268 -               ext3_free_blocks(handle, inode, bh->b_blocknr, 1);
 269 +               ext3_free_blocks(handle, inode, bh->b_blocknr, 1, 1);
 270                 get_bh(bh);
 271                 ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
 272         } else {
 273 @@ -804,7 +804,7 @@ inserted:
 274                         new_bh = sb_getblk(sb, block);
 275                         if (!new_bh) {
 276  getblk_failed:
 277 -                               ext3_free_blocks(handle, inode, block, 1);
 278 +                               ext3_free_blocks(handle, inode, block, 1, 1);
 279                                 error = -EIO;
 280                                 goto cleanup;
 281                         }
 282 Index: linux-2.6.16.46-0.14/fs/ext3/balloc.c
 283 ===================================================================
 284 --- linux-2.6.16.46-0.14.orig/fs/ext3/balloc.c
 285 +++ linux-2.6.16.46-0.14/fs/ext3/balloc.c
 286 @@ -80,7 +80,7 @@ struct ext3_group_desc * ext3_get_group_
 287   *
 288   * Return buffer_head on success or NULL in case of failure.
 289   */
 290 -static struct buffer_head *
 291 +struct buffer_head *
 292  read_block_bitmap(struct super_block *sb, unsigned int block_group)
 293  {
 294         struct ext3_group_desc * desc;
 295 @@ -296,6 +296,8 @@ void ext3_discard_reservation(struct ino
 296         struct ext3_reserve_window_node *rsv;
 297         spinlock_t *rsv_lock = &EXT3_SB(inode->i_sb)->s_rsv_window_lock;
 298
 299 +       ext3_mb_discard_inode_preallocations(inode);
 300 +
 301         if (!block_i)
 302                 return;
 303
 304 @@ -491,21 +493,25 @@ error_return:
 305         return;
 306  }
 307
 308 -/* Free given blocks, update quota and i_blocks field */
 309 -void ext3_free_blocks(handle_t *handle, struct inode *inode,
 310 -                       unsigned long block, unsigned long count)
 311 +void ext3_free_blocks(handle_t *handle, struct inode * inode,
 312 +               unsigned long block, unsigned long count, int metadata)
 313  {
 314 -       struct super_block * sb;
 315 -       int dquot_freed_blocks;
 316 +       struct super_block *sb;
 317 +       int freed;
 318 +
 319 +       /* this isn't the right place to decide whether block is metadata
 320 +        * inode.c/extents.c knows better, but for safety ... */
 321 +       if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
 322 +                       ext3_should_journal_data(inode))
 323 +               metadata = 1;
 324
 325         sb = inode->i_sb;
 326 -       if (!sb) {
 327 -               printk ("ext3_free_blocks: nonexistent device");
 328 -               return;
 329 -       }
 330 -       ext3_free_blocks_sb(handle, sb, block, count, &dquot_freed_blocks);
 331 -       if (dquot_freed_blocks)
 332 -               DQUOT_FREE_BLOCK(inode, dquot_freed_blocks);
 333 +       if (!test_opt(sb, MBALLOC) || !EXT3_SB(sb)->s_group_info)
 334 +               ext3_free_blocks_sb(handle, sb, block, count, &freed);
 335 +       else
 336 +               ext3_mb_free_blocks(handle, inode, block, count, metadata, &freed);
 337 +       if (freed)
 338 +               DQUOT_FREE_BLOCK(inode, freed);
 339         return;
 340  }
 341
 342 @@ -1154,7 +1160,7 @@ int ext3_should_retry_alloc(struct super
 343   * bitmap, and then for any free bit if that fails.
 344   * This function also updates quota and i_blocks field.
 345   */
 346 -int ext3_new_block(handle_t *handle, struct inode *inode,
 347 +int ext3_new_block_old(handle_t *handle, struct inode *inode,
 348                         unsigned long goal, int *errp)
 349  {
 350         struct buffer_head *bitmap_bh = NULL;
 351 Index: linux-2.6.16.46-0.14/fs/ext3/inode.c
 352 ===================================================================
 353 --- linux-2.6.16.46-0.14.orig/fs/ext3/inode.c
 354 +++ linux-2.6.16.46-0.14/fs/ext3/inode.c
 355 @@ -569,7 +569,7 @@ static int ext3_alloc_branch(handle_t *h
 356                 ext3_journal_forget(handle, branch[i].bh);
 357         }
 358         for (i = 0; i < keys; i++)
 359 -               ext3_free_blocks(handle, inode, le32_to_cpu(branch[i].key), 1);
 360 +               ext3_free_blocks(handle, inode, le32_to_cpu(branch[i].key), 1, 0);
 361         return err;
 362  }
 363
 364 @@ -1866,7 +1866,7 @@ ext3_clear_blocks(handle_t *handle, stru
 365                 }
 366         }
 367
 368 -       ext3_free_blocks(handle, inode, block_to_free, count);
 369 +       ext3_free_blocks(handle, inode, block_to_free, count, 0);
 370  }
 371
 372  /**
 373 @@ -2039,7 +2039,7 @@ static void ext3_free_branches(handle_t
 374                                 ext3_journal_test_restart(handle, inode);
 375                         }
 376
 377 -                       ext3_free_blocks(handle, inode, nr, 1);
 378 +                       ext3_free_blocks(handle, inode, nr, 1, 1);
 379
 380                         if (parent_bh) {
 381                                 /*