lustre/kernel_patches/patches/ext3-mballoc3-rhel4.patch

   1 Index: linux-2.6.9-full/include/linux/ext3_fs_i.h
   2 ===================================================================
   3 --- linux-2.6.9-full.orig/include/linux/ext3_fs_i.h     2007-03-28 01:29:38.000000000 +0400
   4 +++ linux-2.6.9-full/include/linux/ext3_fs_i.h  2007-03-28 15:45:41.000000000 +0400
   5 @@ -130,6 +130,10 @@ 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-2.6.9-full/include/linux/ext3_fs_sb.h
  17 ===================================================================
  18 --- linux-2.6.9-full.orig/include/linux/ext3_fs_sb.h    2007-03-28 15:42:16.000000000 +0400
  19 +++ linux-2.6.9-full/include/linux/ext3_fs_sb.h 2007-03-28 15:45:41.000000000 +0400
  20 @@ -23,9 +23,16 @@
  21  #define EXT_INCLUDE
  22  #include <linux/blockgroup_lock.h>
  23  #include <linux/percpu_counter.h>
  24 +#include <linux/list.h>
  25  #endif
  26  #endif
  27  #include <linux/rbtree.h>
  28 +#include <linux/proc_fs.h>
  29 +
  30 +struct ext3_buddy_group_blocks;
  31 +struct ext3_locality_group;
  32 +struct ext3_mb_history;
  33 +#define EXT3_BB_MAX_BLOCKS
  34
  35  /*
  36   * third extended-fs super-block data in memory
  37 Index: linux-2.6.9-full/include/linux/ext3_fs.h
  38 ===================================================================
  39 --- linux-2.6.9-full.orig/include/linux/ext3_fs.h       2007-03-28 15:45:07.000000000 +0400
  40 +++ linux-2.6.9-full/include/linux/ext3_fs.h    2007-03-28 15:45:41.000000000 +0400
  41 @@ -389,6 +389,7 @@ struct ext3_inode {
  42  #define EXT3_MOUNT_IOPEN_NOPRIV                0x100000/* Make iopen world-readable */
  43  #define EXT3_MOUNT_EXTENTS             0x200000/* Extents support */
  44  #define EXT3_MOUNT_EXTDEBUG            0x400000/* Extents debug */
  45 +#define EXT3_MOUNT_MBALLOC             0x800000/* Buddy allocation support */
  46
  47  /* Compatibility, for having both ext2_fs.h and ext3_fs.h included at once */
  48  #ifndef clear_opt
  49 @@ -749,8 +750,9 @@ struct dir_private_info {
  50  extern int ext3_bg_has_super(struct super_block *sb, int group);
  51  extern unsigned long ext3_bg_num_gdb(struct super_block *sb, int group);
  52  extern int ext3_new_block (handle_t *, struct inode *, unsigned long, int *);
  53 +extern int ext3_new_block_old (handle_t *, struct inode *, unsigned long, int *);
  54  extern void ext3_free_blocks (handle_t *, struct inode *, unsigned long,
  55 -                             unsigned long);
  56 +                             unsigned long, int);
  57  extern void ext3_free_blocks_sb (handle_t *, struct super_block *,
  58                                  unsigned long, unsigned long, int *);
  59  extern unsigned long ext3_count_free_blocks (struct super_block *);
  60 Index: linux-2.6.9-full/fs/ext3/super.c
  61 ===================================================================
  62 --- linux-2.6.9-full.orig/fs/ext3/super.c       2007-03-28 15:42:16.000000000 +0400
  63 +++ linux-2.6.9-full/fs/ext3/super.c    2007-03-28 15:45:41.000000000 +0400
  64 @@ -600,6 +600,7 @@ enum {
  65         Opt_ignore, Opt_barrier, Opt_err, Opt_resize,
  66         Opt_iopen, Opt_noiopen, Opt_iopen_nopriv,
  67         Opt_extents, Opt_noextents, Opt_extdebug,
  68 +       Opt_mballoc, Opt_nomballoc, Opt_stripe,
  69  };
  70
  71  static match_table_t tokens = {
  72 @@ -653,6 +654,9 @@ static match_table_t tokens = {
  73         {Opt_noextents, "noextents"},
  74         {Opt_extdebug, "extdebug"},
  75         {Opt_barrier, "barrier=%u"},
  76 +       {Opt_mballoc, "mballoc"},
  77 +       {Opt_nomballoc, "nomballoc"},
  78 +       {Opt_stripe, "stripe=%u"},
  79         {Opt_err, NULL},
  80         {Opt_resize, "resize"},
  81  };
  82 @@ -965,6 +969,19 @@ clear_qf_name:
  83                 case Opt_extdebug:
  84                         set_opt (sbi->s_mount_opt, EXTDEBUG);
  85                         break;
  86 +               case Opt_mballoc:
  87 +                       set_opt(sbi->s_mount_opt, MBALLOC);
  88 +                       break;
  89 +               case Opt_nomballoc:
  90 +                       clear_opt(sbi->s_mount_opt, MBALLOC);
  91 +                       break;
  92 +               case Opt_stripe:
  93 +                       if (match_int(&args[0], &option))
  94 +                               return 0;
  95 +                       if (option < 0)
  96 +                               return 0;
  97 +                       sbi->s_stripe = option;
  98 +                       break;
  99                 default:
 100                         printk (KERN_ERR
 101                                 "EXT3-fs: Unrecognized mount option \"%s\" "
 102 @@ -1654,6 +1671,7 @@ static int ext3_fill_super (struct super
 103                 ext3_count_dirs(sb));
 104
 105         ext3_ext_init(sb);
 106 +       ext3_mb_init(sb, needs_recovery);
 107
 108         return 0;
 109
 110 Index: linux-2.6.9-full/fs/ext3/extents.c
 111 ===================================================================
 112 --- linux-2.6.9-full.orig/fs/ext3/extents.c     2007-03-28 01:29:41.000000000 +0400
 113 +++ linux-2.6.9-full/fs/ext3/extents.c  2007-03-28 15:45:41.000000000 +0400
 114 @@ -779,7 +779,7 @@ 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 @@ -1586,7 +1586,7 @@ 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 @@ -2071,10 +2071,12 @@ 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 @@ -2086,7 +2088,7 @@ 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 @@ -2177,11 +2179,8 @@ 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 @@ -2253,18 +2252,33 @@ 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 +       ar.flags = EXT3_MB_HINT_DATA;
 205 +       newblock = ext3_mb_new_blocks(handle, &ar, &err);
 206         if (!newblock)
 207                 goto out2;
 208         ext_debug(&tree, "allocate new block: goal %d, found %d\n",
 209 @@ -2274,11 +2288,14 @@ int ext3_ext_get_block(handle_t *handle,
 210         newex.ee_block = iblock;
 211         newex.ee_start = newblock;
 212         newex.ee_start_hi = 0;
 213 -       newex.ee_len = 1;
 214 +       newex.ee_len = ar.len;
 215         err = ext3_ext_insert_extent(handle, &tree, path, &newex);
 216         if (err) {
 217                 /* free data blocks we just allocated */
 218 -               ext3_free_blocks(handle, inode, newex.ee_start, newex.ee_len);
 219 +               /* not a good idea to call discard here directly,
 220 +                * but otherwise we'd need to call it every free() */
 221 +               ext3_mb_discard_inode_preallocations(inode);
 222 +               ext3_free_blocks(handle, inode, newex.ee_start, newex.ee_len, 0);
 223                 goto out2;
 224         }
 225
 226 @@ -2287,6 +2304,7 @@ int ext3_ext_get_block(handle_t *handle,
 227
 228         /* previous routine could use block we allocated */
 229         newblock = newex.ee_start;
 230 +       allocated = newex.ee_len;
 231         set_buffer_new(bh_result);
 232
 233         ext3_ext_put_in_cache(&tree, newex.ee_block, newex.ee_len,
 234 @@ -2339,6 +2357,9 @@ void ext3_ext_truncate(struct inode * in
 235         down(&EXT3_I(inode)->truncate_sem);
 236         ext3_ext_invalidate_cache(&tree);
 237
 238 +       /* it's important to discard preallocations under truncate_sem */
 239 +       ext3_mb_discard_inode_preallocations(inode);
 240 +
 241         /*
 242          * TODO: optimization is possible here
 243          * probably we need not scaning at all,
 244 Index: linux-2.6.9-full/fs/ext3/Makefile
 245 ===================================================================
 246 --- linux-2.6.9-full.orig/fs/ext3/Makefile      2007-03-28 01:29:38.000000000 +0400
 247 +++ linux-2.6.9-full/fs/ext3/Makefile   2007-03-28 15:45:41.000000000 +0400
 248 @@ -6,7 +6,7 @@ obj-$(CONFIG_EXT3_FS) += ext3.o
 249
 250  ext3-y := balloc.o bitmap.o dir.o file.o fsync.o ialloc.o inode.o iopen.o \
 251            ioctl.o namei.o super.o symlink.o hash.o resize.o \
 252 -          extents.o
 253 +          extents.o mballoc.o
 254
 255  ext3-$(CONFIG_EXT3_FS_XATTR)    += xattr.o xattr_user.o xattr_trusted.o
 256  ext3-$(CONFIG_EXT3_FS_POSIX_ACL) += acl.o
 257 Index: linux-2.6.9-full/fs/ext3/xattr.c
 258 ===================================================================
 259 --- linux-2.6.9-full.orig/fs/ext3/xattr.c       2006-05-18 23:57:04.000000000 +0400
 260 +++ linux-2.6.9-full/fs/ext3/xattr.c    2007-03-28 15:45:41.000000000 +0400
 261 @@ -1281,7 +1281,7 @@ ext3_xattr_set_handle2(handle_t *handle,
 262                         new_bh = sb_getblk(sb, block);
 263                         if (!new_bh) {
 264  getblk_failed:
 265 -                               ext3_free_blocks(handle, inode, block, 1);
 266 +                               ext3_free_blocks(handle, inode, block, 1, 1);
 267                                 error = -EIO;
 268                                 goto cleanup;
 269                         }
 270 @@ -1328,7 +1328,7 @@ getblk_failed:
 271                         if (ce)
 272                                 mb_cache_entry_free(ce);
 273                         ea_bdebug(old_bh, "freeing");
 274 -                       ext3_free_blocks(handle, inode, old_bh->b_blocknr, 1);
 275 +                       ext3_free_blocks(handle, inode, old_bh->b_blocknr, 1, 1);
 276
 277                         /* ext3_forget() calls bforget() for us, but we
 278                            let our caller release old_bh, so we need to
 279 @@ -1427,7 +1427,7 @@ ext3_xattr_delete_inode(handle_t *handle
 280         if (HDR(bh)->h_refcount == cpu_to_le32(1)) {
 281                 if (ce)
 282                         mb_cache_entry_free(ce);
 283 -               ext3_free_blocks(handle, inode, EXT3_I(inode)->i_file_acl, 1);
 284 +               ext3_free_blocks(handle, inode, EXT3_I(inode)->i_file_acl, 1, 1);
 285                 get_bh(bh);
 286                 ext3_forget(handle, 1, inode, bh, EXT3_I(inode)->i_file_acl);
 287         } else {
 288 Index: linux-2.6.9-full/fs/ext3/balloc.c
 289 ===================================================================
 290 --- linux-2.6.9-full.orig/fs/ext3/balloc.c      2006-03-10 18:20:03.000000000 +0300
 291 +++ linux-2.6.9-full/fs/ext3/balloc.c   2007-03-28 15:45:41.000000000 +0400
 292 @@ -79,7 +79,7 @@ struct ext3_group_desc * ext3_get_group_
 293   *
 294   * Return buffer_head on success or NULL in case of failure.
 295   */
 296 -static struct buffer_head *
 297 +struct buffer_head *
 298  read_block_bitmap(struct super_block *sb, unsigned int block_group)
 299  {
 300         struct ext3_group_desc * desc;
 301 @@ -267,6 +267,8 @@ void ext3_discard_reservation(struct ino
 302         struct reserve_window_node *rsv = &ei->i_rsv_window;
 303         spinlock_t *rsv_lock = &EXT3_SB(inode->i_sb)->s_rsv_window_lock;
 304
 305 +       ext3_mb_discard_inode_preallocations(inode);
 306 +
 307         if (!rsv_is_empty(&rsv->rsv_window)) {
 308                 spin_lock(rsv_lock);
 309                 if (!rsv_is_empty(&rsv->rsv_window))
 310 @@ -451,21 +453,25 @@ error_return:
 311         return;
 312  }
 313
 314 -/* Free given blocks, update quota and i_blocks field */
 315 -void ext3_free_blocks(handle_t *handle, struct inode *inode,
 316 -                       unsigned long block, unsigned long count)
 317 +void ext3_free_blocks(handle_t *handle, struct inode * inode,
 318 +               unsigned long block, unsigned long count, int metadata)
 319  {
 320 -       struct super_block * sb;
 321 -       int dquot_freed_blocks;
 322 +       struct super_block *sb;
 323 +       int freed;
 324 +
 325 +       /* this isn't the right place to decide whether block is metadata
 326 +        * inode.c/extents.c knows better, but for safety ... */
 327 +       if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
 328 +                       ext3_should_journal_data(inode))
 329 +               metadata = 1;
 330
 331         sb = inode->i_sb;
 332 -       if (!sb) {
 333 -               printk ("ext3_free_blocks: nonexistent device");
 334 -               return;
 335 -       }
 336 -       ext3_free_blocks_sb(handle, sb, block, count, &dquot_freed_blocks);
 337 -       if (dquot_freed_blocks)
 338 -               DQUOT_FREE_BLOCK(inode, dquot_freed_blocks);
 339 +       if (!test_opt(sb, MBALLOC) || !EXT3_SB(sb)->s_group_info)
 340 +               ext3_free_blocks_sb(handle, sb, block, count, &freed);
 341 +       else
 342 +               ext3_mb_free_blocks(handle, inode, block, count, metadata, &freed);
 343 +       if (freed)
 344 +               DQUOT_FREE_BLOCK(inode, freed);
 345         return;
 346  }
 347
 348 @@ -1131,7 +1137,7 @@ int ext3_should_retry_alloc(struct super
 349   * bitmap, and then for any free bit if that fails.
 350   * This function also updates quota and i_blocks field.
 351   */
 352 -int ext3_new_block(handle_t *handle, struct inode *inode,
 353 +int ext3_new_block_old(handle_t *handle, struct inode *inode,
 354                         unsigned long goal, int *errp)
 355  {
 356         struct buffer_head *bitmap_bh = NULL;
 357 Index: linux-2.6.9-full/fs/ext3/inode.c
 358 ===================================================================
 359 --- linux-2.6.9-full.orig/fs/ext3/inode.c       2007-03-28 01:29:39.000000000 +0400
 360 +++ linux-2.6.9-full/fs/ext3/inode.c    2007-03-28 15:45:41.000000000 +0400
 361 @@ -572,7 +572,7 @@ static int ext3_alloc_branch(handle_t *h
 362                 ext3_journal_forget(handle, branch[i].bh);
 363         }
 364         for (i = 0; i < keys; i++)
 365 -               ext3_free_blocks(handle, inode, le32_to_cpu(branch[i].key), 1);
 366 +               ext3_free_blocks(handle, inode, le32_to_cpu(branch[i].key), 1, 0);
 367         return err;
 368  }
 369
 370 @@ -673,7 +673,7 @@ err_out:
 371         if (err == -EAGAIN)
 372                 for (i = 0; i < num; i++)
 373                         ext3_free_blocks(handle, inode,
 374 -                                        le32_to_cpu(where[i].key), 1);
 375 +                                        le32_to_cpu(where[i].key), 1, 0);
 376         return err;
 377  }
 378
 379 @@ -1834,7 +1834,7 @@ ext3_clear_blocks(handle_t *handle, stru
 380                 }
 381         }
 382
 383 -       ext3_free_blocks(handle, inode, block_to_free, count);
 384 +       ext3_free_blocks(handle, inode, block_to_free, count, 0);
 385  }
 386
 387  /**
 388 @@ -2007,7 +2007,7 @@ static void ext3_free_branches(handle_t
 389                                 ext3_journal_test_restart(handle, inode);
 390                         }
 391
 392 -                       ext3_free_blocks(handle, inode, nr, 1);
 393 +                       ext3_free_blocks(handle, inode, nr, 1, 1);
 394
 395                         if (parent_bh) {
 396                                 /*