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

   1 Index: linux-2.6.16.27-0.9-full/include/linux/ext3_fs_i.h
   2 ===================================================================
   3 --- linux-2.6.16.27-0.9-full.orig/include/linux/ext3_fs_i.h     2007-03-28 05:12:50.000000000 +0400
   4 +++ linux-2.6.16.27-0.9-full/include/linux/ext3_fs_i.h  2007-03-28 16:03:20.000000000 +0400
   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-2.6.16.27-0.9-full/include/linux/ext3_fs_sb.h
  17 ===================================================================
  18 --- linux-2.6.16.27-0.9-full.orig/include/linux/ext3_fs_sb.h    2007-03-28 16:03:19.000000000 +0400
  19 +++ linux-2.6.16.27-0.9-full/include/linux/ext3_fs_sb.h 2007-03-28 16:03:20.000000000 +0400
  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-2.6.16.27-0.9-full/include/linux/ext3_fs.h
  37 ===================================================================
  38 --- linux-2.6.16.27-0.9-full.orig/include/linux/ext3_fs.h       2007-03-28 16:03:19.000000000 +0400
  39 +++ linux-2.6.16.27-0.9-full/include/linux/ext3_fs.h    2007-03-28 16:03:20.000000000 +0400
  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 @@ -767,8 +768,9 @@ 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-2.6.16.27-0.9-full/fs/ext3/super.c
  60 ===================================================================
  61 --- linux-2.6.16.27-0.9-full.orig/fs/ext3/super.c       2007-03-28 16:03:19.000000000 +0400
  62 +++ linux-2.6.16.27-0.9-full/fs/ext3/super.c    2007-03-28 16:03:20.000000000 +0400
  63 @@ -688,6 +688,7 @@ enum {
  64         Opt_ignore, Opt_barrier, Opt_err, Opt_resize, Opt_usrquota,
  65         Opt_iopen, Opt_noiopen, Opt_iopen_nopriv,
  66         Opt_extents, Opt_noextents, Opt_extdebug,
  67 +       Opt_mballoc, Opt_nomballoc, Opt_stripe,
  68         Opt_grpquota
  69  };
  70
  71 @@ -743,6 +744,9 @@ static match_table_t tokens = {
  72         {Opt_noextents, "noextents"},
  73         {Opt_extdebug, "extdebug"},
  74         {Opt_barrier, "barrier=%u"},
  75 +       {Opt_mballoc, "mballoc"},
  76 +       {Opt_nomballoc, "nomballoc"},
  77 +       {Opt_stripe, "stripe=%u"},
  78         {Opt_err, NULL},
  79         {Opt_resize, "resize"},
  80  };
  81 @@ -1092,6 +1096,19 @@ 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 @@ -1819,6 +1836,7 @@ static int ext3_fill_super (struct super
 102                 ext3_count_dirs(sb));
 103
 104         ext3_ext_init(sb);
 105 +       ext3_mb_init(sb, needs_recovery);
 106         lock_kernel();
 107         return 0;
 108
 109 Index: linux-2.6.16.27-0.9-full/fs/ext3/extents.c
 110 ===================================================================
 111 --- linux-2.6.16.27-0.9-full.orig/fs/ext3/extents.c     2007-03-28 05:13:39.000000000 +0400
 112 +++ linux-2.6.16.27-0.9-full/fs/ext3/extents.c  2007-03-28 16:03:20.000000000 +0400
 113 @@ -779,7 +779,7 @@ cleanup:
 114                 for (i = 0; i < depth; i++) {
 115                         if (!ablocks[i])
 116                                 continue;
 117 -                       ext3_free_blocks(handle, tree->inode, ablocks[i], 1);
 118 +                       ext3_free_blocks(handle, tree->inode, ablocks[i], 1, 1);
 119                 }
 120         }
 121         kfree(ablocks);
 122 @@ -1586,7 +1586,7 @@ int ext3_ext_rm_idx(handle_t *handle, st
 123                   path->p_idx->ei_leaf);
 124         bh = sb_find_get_block(tree->inode->i_sb, path->p_idx->ei_leaf);
 125         ext3_forget(handle, 1, tree->inode, bh, path->p_idx->ei_leaf);
 126 -       ext3_free_blocks(handle, tree->inode, path->p_idx->ei_leaf, 1);
 127 +       ext3_free_blocks(handle, tree->inode, path->p_idx->ei_leaf, 1, 1);
 128         return err;
 129  }
 130
 131 @@ -2071,10 +2071,12 @@ ext3_remove_blocks(struct ext3_extents_t
 132         int needed = ext3_remove_blocks_credits(tree, ex, from, to);
 133         handle_t *handle = ext3_journal_start(tree->inode, needed);
 134         struct buffer_head *bh;
 135 -       int i;
 136 +       int i, metadata = 0;
 137
 138         if (IS_ERR(handle))
 139                 return PTR_ERR(handle);
 140 +       if (S_ISDIR(tree->inode->i_mode) || S_ISLNK(tree->inode->i_mode))
 141 +               metadata = 1;
 142         if (from >= ex->ee_block && to == ex->ee_block + ex->ee_len - 1) {
 143                 /* tail removal */
 144                 unsigned long num, start;
 145 @@ -2086,7 +2088,7 @@ ext3_remove_blocks(struct ext3_extents_t
 146                         bh = sb_find_get_block(tree->inode->i_sb, start + i);
 147                         ext3_forget(handle, 0, tree->inode, bh, start + i);
 148                 }
 149 -               ext3_free_blocks(handle, tree->inode, start, num);
 150 +               ext3_free_blocks(handle, tree->inode, start, num, metadata);
 151         } else if (from == ex->ee_block && to <= ex->ee_block + ex->ee_len - 1) {
 152                 printk("strange request: removal %lu-%lu from %u:%u\n",
 153                        from, to, ex->ee_block, ex->ee_len);
 154 @@ -2177,11 +2179,8 @@ int ext3_ext_get_block(handle_t *handle,
 155         struct ext3_extent *ex;
 156         int goal, newblock, err = 0, depth;
 157         struct ext3_extents_tree tree;
 158 -       unsigned long next;
 159 -       int allocated = 0;
 160 -
 161 -       /* until we have multiblock allocation */
 162 -       max_blocks = 1;
 163 +       unsigned long allocated = 0;
 164 +       struct ext3_allocation_request ar;
 165
 166         clear_buffer_new(bh_result);
 167         ext3_init_tree_desc(&tree, inode);
 168 @@ -2253,18 +2252,33 @@ int ext3_ext_get_block(handle_t *handle,
 169                 goto out2;
 170         }
 171
 172 +       /* find neighbour allocated blocks */
 173 +       ar.lleft = iblock;
 174 +       err = ext3_ext_search_left(&tree, path, &ar.lleft, &ar.pleft);
 175 +       if (err)
 176 +               goto out2;
 177 +       ar.lright = iblock;
 178 +       err = ext3_ext_search_right(&tree, path, &ar.lright, &ar.pright);
 179 +       if (err)
 180 +               goto out2;
 181 +
 182         /* find next allocated block so that we know how many
 183          * blocks we can allocate without ovelapping next extent */
 184 -       EXT_ASSERT(iblock >= ex->ee_block + ex->ee_len);
 185 -       next = ext3_ext_next_allocated_block(path);
 186 -       EXT_ASSERT(next > iblock);
 187 -       allocated = next - iblock;
 188 +       EXT_ASSERT(ar.pright == 0 || ar.lright > iblock);
 189 +       if (ar.pright == 0)
 190 +               allocated = EXT_MAX_BLOCK - iblock;
 191 +       else
 192 +               allocated = ar.lright - iblock;
 193         if (allocated > max_blocks)
 194                 allocated = max_blocks;
 195
 196         /* allocate new block */
 197 -       goal = ext3_ext_find_goal(inode, path, iblock);
 198 -       newblock = ext3_new_block(handle, inode, goal, &err);
 199 +       ar.inode = inode;
 200 +       ar.goal = ext3_ext_find_goal(inode, path, iblock);
 201 +       ar.logical = iblock;
 202 +       ar.len = allocated;
 203 +       ar.flags = EXT3_MB_HINT_DATA;
 204 +       newblock = ext3_mb_new_blocks(handle, &ar, &err);
 205         if (!newblock)
 206                 goto out2;
 207         ext_debug(&tree, "allocate new block: goal %d, found %d\n",
 208 @@ -2274,11 +2288,14 @@ int ext3_ext_get_block(handle_t *handle,
 209         newex.ee_block = iblock;
 210         newex.ee_start = newblock;
 211         newex.ee_start_hi = 0;
 212 -       newex.ee_len = 1;
 213 +       newex.ee_len = ar.len;
 214         err = ext3_ext_insert_extent(handle, &tree, path, &newex);
 215         if (err) {
 216                 /* free data blocks we just allocated */
 217 -               ext3_free_blocks(handle, inode, newex.ee_start, newex.ee_len);
 218 +               /* not a good idea to call discard here directly,
 219 +                * but otherwise we'd need to call it every free() */
 220 +               ext3_mb_discard_inode_preallocations(inode);
 221 +               ext3_free_blocks(handle, inode, newex.ee_start, newex.ee_len, 0);
 222                 goto out2;
 223         }
 224
 225 @@ -2287,6 +2304,7 @@ int ext3_ext_get_block(handle_t *handle,
 226
 227         /* previous routine could use block we allocated */
 228         newblock = newex.ee_start;
 229 +       allocated = newex.ee_len;
 230         set_buffer_new(bh_result);
 231
 232         ext3_ext_put_in_cache(&tree, newex.ee_block, newex.ee_len,
 233 @@ -2339,6 +2357,9 @@ void ext3_ext_truncate(struct inode * in
 234         down(&EXT3_I(inode)->truncate_sem);
 235         ext3_ext_invalidate_cache(&tree);
 236
 237 +       /* it's important to discard preallocations under truncate_sem */
 238 +       ext3_mb_discard_inode_preallocations(inode);
 239 +
 240         /*
 241          * TODO: optimization is possible here
 242          * probably we need not scaning at all,
 243 Index: linux-2.6.16.27-0.9-full/fs/ext3/Makefile
 244 ===================================================================
 245 --- linux-2.6.16.27-0.9-full.orig/fs/ext3/Makefile      2007-03-28 05:12:50.000000000 +0400
 246 +++ linux-2.6.16.27-0.9-full/fs/ext3/Makefile   2007-03-28 16:03:20.000000000 +0400
 247 @@ -6,7 +6,7 @@ obj-$(CONFIG_EXT3_FS) += ext3.o
 248
 249  ext3-y := balloc.o bitmap.o dir.o file.o fsync.o ialloc.o inode.o iopen.o \
 250            ioctl.o namei.o super.o symlink.o hash.o resize.o \
 251 -          extents.o
 252 +          extents.o mballoc.o
 253
 254  ext3-$(CONFIG_EXT3_FS_XATTR)    += xattr.o xattr_user.o xattr_trusted.o
 255  ext3-$(CONFIG_EXT3_FS_POSIX_ACL) += acl.o
 256 Index: linux-2.6.16.27-0.9-full/fs/ext3/xattr.c
 257 ===================================================================
 258 --- linux-2.6.16.27-0.9-full.orig/fs/ext3/xattr.c       2007-03-13 02:56:52.000000000 +0300
 259 +++ linux-2.6.16.27-0.9-full/fs/ext3/xattr.c    2007-03-28 16:03:20.000000000 +0400
 260 @@ -484,7 +484,7 @@ ext3_xattr_release_block(handle_t *handl
 261                 ea_bdebug(bh, "refcount now=0; freeing");
 262                 if (ce)
 263                         mb_cache_entry_free(ce);
 264 -               ext3_free_blocks(handle, inode, bh->b_blocknr, 1);
 265 +               ext3_free_blocks(handle, inode, bh->b_blocknr, 1, 1);
 266                 get_bh(bh);
 267                 ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
 268         } else {
 269 @@ -804,7 +804,7 @@ inserted:
 270                         new_bh = sb_getblk(sb, block);
 271                         if (!new_bh) {
 272  getblk_failed:
 273 -                               ext3_free_blocks(handle, inode, block, 1);
 274 +                               ext3_free_blocks(handle, inode, block, 1, 1);
 275                                 error = -EIO;
 276                                 goto cleanup;
 277                         }
 278 Index: linux-2.6.16.27-0.9-full/fs/ext3/balloc.c
 279 ===================================================================
 280 --- linux-2.6.16.27-0.9-full.orig/fs/ext3/balloc.c      2007-03-13 02:56:52.000000000 +0300
 281 +++ linux-2.6.16.27-0.9-full/fs/ext3/balloc.c   2007-03-28 16:03:20.000000000 +0400
 282 @@ -80,7 +80,7 @@ struct ext3_group_desc * ext3_get_group_
 283   *
 284   * Return buffer_head on success or NULL in case of failure.
 285   */
 286 -static struct buffer_head *
 287 +struct buffer_head *
 288  read_block_bitmap(struct super_block *sb, unsigned int block_group)
 289  {
 290         struct ext3_group_desc * desc;
 291 @@ -296,6 +296,8 @@ void ext3_discard_reservation(struct ino
 292         struct ext3_reserve_window_node *rsv;
 293         spinlock_t *rsv_lock = &EXT3_SB(inode->i_sb)->s_rsv_window_lock;
 294
 295 +       ext3_mb_discard_inode_preallocations(inode);
 296 +
 297         if (!block_i)
 298                 return;
 299
 300 @@ -491,21 +493,25 @@ error_return:
 301         return;
 302  }
 303
 304 -/* Free given blocks, update quota and i_blocks field */
 305 -void ext3_free_blocks(handle_t *handle, struct inode *inode,
 306 -                       unsigned long block, unsigned long count)
 307 +void ext3_free_blocks(handle_t *handle, struct inode * inode,
 308 +               unsigned long block, unsigned long count, int metadata)
 309  {
 310 -       struct super_block * sb;
 311 -       int dquot_freed_blocks;
 312 +       struct super_block *sb;
 313 +       int freed;
 314 +
 315 +       /* this isn't the right place to decide whether block is metadata
 316 +        * inode.c/extents.c knows better, but for safety ... */
 317 +       if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
 318 +                       ext3_should_journal_data(inode))
 319 +               metadata = 1;
 320
 321         sb = inode->i_sb;
 322 -       if (!sb) {
 323 -               printk ("ext3_free_blocks: nonexistent device");
 324 -               return;
 325 -       }
 326 -       ext3_free_blocks_sb(handle, sb, block, count, &dquot_freed_blocks);
 327 -       if (dquot_freed_blocks)
 328 -               DQUOT_FREE_BLOCK(inode, dquot_freed_blocks);
 329 +       if (!test_opt(sb, MBALLOC) || !EXT3_SB(sb)->s_group_info)
 330 +               ext3_free_blocks_sb(handle, sb, block, count, &freed);
 331 +       else
 332 +               ext3_mb_free_blocks(handle, inode, block, count, metadata, &freed);
 333 +       if (freed)
 334 +               DQUOT_FREE_BLOCK(inode, freed);
 335         return;
 336  }
 337
 338 @@ -1154,7 +1160,7 @@ int ext3_should_retry_alloc(struct super
 339   * bitmap, and then for any free bit if that fails.
 340   * This function also updates quota and i_blocks field.
 341   */
 342 -int ext3_new_block(handle_t *handle, struct inode *inode,
 343 +int ext3_new_block_old(handle_t *handle, struct inode *inode,
 344                         unsigned long goal, int *errp)
 345  {
 346         struct buffer_head *bitmap_bh = NULL;
 347 Index: linux-2.6.16.27-0.9-full/fs/ext3/inode.c
 348 ===================================================================
 349 --- linux-2.6.16.27-0.9-full.orig/fs/ext3/inode.c       2007-03-28 05:13:38.000000000 +0400
 350 +++ linux-2.6.16.27-0.9-full/fs/ext3/inode.c    2007-03-28 16:03:20.000000000 +0400
 351 @@ -568,7 +568,7 @@ static int ext3_alloc_branch(handle_t *h
 352                 ext3_journal_forget(handle, branch[i].bh);
 353         }
 354         for (i = 0; i < keys; i++)
 355 -               ext3_free_blocks(handle, inode, le32_to_cpu(branch[i].key), 1);
 356 +               ext3_free_blocks(handle, inode, le32_to_cpu(branch[i].key), 1, 0);
 357         return err;
 358  }
 359
 360 @@ -1865,7 +1865,7 @@ ext3_clear_blocks(handle_t *handle, stru
 361                 }
 362         }
 363
 364 -       ext3_free_blocks(handle, inode, block_to_free, count);
 365 +       ext3_free_blocks(handle, inode, block_to_free, count, 0);
 366  }
 367
 368  /**
 369 @@ -2038,7 +2038,7 @@ static void ext3_free_branches(handle_t
 370                                 ext3_journal_test_restart(handle, inode);
 371                         }
 372
 373 -                       ext3_free_blocks(handle, inode, nr, 1);
 374 +                       ext3_free_blocks(handle, inode, nr, 1, 1);
 375
 376                         if (parent_bh) {
 377                                 /*