lustre/kernel_patches/patches/ext3-nanosecond-2.6-rhel4.patch

   1 Index: linux-2.6.9-rhel4/fs/ext3/ialloc.c
   2 ===================================================================
   3 --- linux-2.6.9-rhel4.orig/fs/ext3/ialloc.c
   4 +++ linux-2.6.9-rhel4/fs/ext3/ialloc.c
   5 @@ -612,7 +612,8 @@ got:
   6         /* This is the optimal IO size (for stat), not the fs block size */
   7         inode->i_blksize = PAGE_SIZE;
   8         inode->i_blocks = 0;
   9 -       inode->i_mtime = inode->i_atime = inode->i_ctime = CURRENT_TIME_SEC;
  10 +       inode->i_mtime = inode->i_atime = inode->i_ctime = ei->i_crtime =
  11 +                                                      ext3_current_time(inode);
  12
  13         memset(ei->i_data, 0, sizeof(ei->i_data));
  14         ei->i_next_alloc_block = 0;
  15 @@ -650,9 +651,8 @@ got:
  16         spin_unlock(&sbi->s_next_gen_lock);
  17
  18         ei->i_state = EXT3_STATE_NEW;
  19 -       ei->i_extra_isize =
  20 -               (EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) ?
  21 -               sizeof(struct ext3_inode) - EXT3_GOOD_OLD_INODE_SIZE : 0;
  22 +
  23 +       ei->i_extra_isize = EXT3_SB(sb)->s_want_extra_isize;
  24
  25         ret = inode;
  26         if(DQUOT_ALLOC_INODE(inode)) {
  27 Index: linux-2.6.9-rhel4/fs/ext3/inode.c
  28 ===================================================================
  29 --- linux-2.6.9-rhel4.orig/fs/ext3/inode.c
  30 +++ linux-2.6.9-rhel4/fs/ext3/inode.c
  31 @@ -627,7 +627,7 @@ static int ext3_splice_branch(handle_t *
  32
  33         /* We are done with atomic stuff, now do the rest of housekeeping */
  34
  35 -       inode->i_ctime = CURRENT_TIME_SEC;
  36 +       inode->i_ctime = ext3_current_time(inode);
  37         ext3_mark_inode_dirty(handle, inode);
  38
  39         /* had we spliced it onto indirect block? */
  40 @@ -2227,7 +2227,7 @@ do_indirects:
  41                         ;
  42         }
  43         up(&ei->truncate_sem);
  44 -       inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
  45 +       inode->i_mtime = inode->i_ctime = ext3_current_time(inode);
  46         ext3_mark_inode_dirty(handle, inode);
  47
  48         /* In a multi-transaction truncate, we only make the final
  49 @@ -2454,10 +2454,6 @@ void ext3_read_inode(struct inode * inod
  50         }
  51         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
  52         inode->i_size = le32_to_cpu(raw_inode->i_size);
  53 -       inode->i_atime.tv_sec = le32_to_cpu(raw_inode->i_atime);
  54 -       inode->i_ctime.tv_sec = le32_to_cpu(raw_inode->i_ctime);
  55 -       inode->i_mtime.tv_sec = le32_to_cpu(raw_inode->i_mtime);
  56 -       inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
  57
  58         ei->i_state = 0;
  59         ei->i_next_alloc_block = 0;
  60 @@ -2518,6 +2514,11 @@ void ext3_read_inode(struct inode * inod
  61         else
  62                 ei->i_extra_isize = 0;
  63
  64 +       EXT3_INODE_GET_XTIME(i_ctime, inode, raw_inode);
  65 +       EXT3_INODE_GET_XTIME(i_mtime, inode, raw_inode);
  66 +       EXT3_INODE_GET_XTIME(i_atime, inode, raw_inode);
  67 +       EXT3_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
  68 +
  69         if (S_ISREG(inode->i_mode)) {
  70                 inode->i_op = &ext3_file_inode_operations;
  71                 inode->i_fop = &ext3_file_operations;
  72 @@ -2598,9 +2599,12 @@ static int ext3_do_update_inode(handle_t
  73         }
  74         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
  75         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
  76 -       raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
  77 -       raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
  78 -       raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
  79 +
  80 +       EXT3_INODE_SET_XTIME(i_ctime, inode, raw_inode);
  81 +       EXT3_INODE_SET_XTIME(i_mtime, inode, raw_inode);
  82 +       EXT3_INODE_SET_XTIME(i_atime, inode, raw_inode);
  83 +       EXT3_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
  84 +
  85         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
  86         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
  87         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
  88 Index: linux-2.6.9-rhel4/fs/ext3/ioctl.c
  89 ===================================================================
  90 --- linux-2.6.9-rhel4.orig/fs/ext3/ioctl.c
  91 +++ linux-2.6.9-rhel4/fs/ext3/ioctl.c
  92 @@ -112,7 +112,7 @@ int ext3_ioctl (struct inode * inode, st
  93                 ei->i_flags = flags;
  94
  95                 ext3_set_inode_flags(inode);
  96 -               inode->i_ctime = CURRENT_TIME_SEC;
  97 +               inode->i_ctime = ext3_current_time(inode);
  98
  99                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
 100  flags_err:
 101 @@ -150,7 +150,7 @@ flags_err:
 102                         return PTR_ERR(handle);
 103                 err = ext3_reserve_inode_write(handle, inode, &iloc);
 104                 if (err == 0) {
 105 -                       inode->i_ctime = CURRENT_TIME_SEC;
 106 +                       inode->i_ctime = ext3_current_time(inode);
 107                         inode->i_generation = generation;
 108                         err = ext3_mark_iloc_dirty(handle, inode, &iloc);
 109                 }
 110 Index: linux-2.6.9-rhel4/fs/ext3/namei.c
 111 ===================================================================
 112 --- linux-2.6.9-rhel4.orig/fs/ext3/namei.c
 113 +++ linux-2.6.9-rhel4/fs/ext3/namei.c
 114 @@ -1302,7 +1302,7 @@ static int add_dirent_to_buf(handle_t *h
 115          * happen is that the times are slightly out of date
 116          * and/or different from the directory change time.
 117          */
 118 -       dir->i_mtime = dir->i_ctime = CURRENT_TIME_SEC;
 119 +       dir->i_mtime = dir->i_ctime = ext3_current_time(dir);
 120         ext3_update_dx_flag(dir);
 121         dir->i_version++;
 122         ext3_mark_inode_dirty(handle, dir);
 123 @@ -2098,7 +2098,7 @@ static int ext3_rmdir (struct inode * di
 124         inode->i_version++;
 125         inode->i_nlink = 0;
 126         ext3_orphan_add(handle, inode);
 127 -       inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME_SEC;
 128 +       inode->i_ctime = dir->i_ctime = dir->i_mtime = ext3_current_time(inode);
 129         ext3_mark_inode_dirty(handle, inode);
 130         ext3_dec_count(handle, dir);
 131         ext3_update_dx_flag(dir);
 132 @@ -2148,13 +2148,13 @@ static int ext3_unlink(struct inode * di
 133         retval = ext3_delete_entry(handle, dir, de, bh);
 134         if (retval)
 135                 goto end_unlink;
 136 -       dir->i_ctime = dir->i_mtime = CURRENT_TIME_SEC;
 137 +       dir->i_ctime = dir->i_mtime = ext3_current_time(dir);
 138         ext3_update_dx_flag(dir);
 139         ext3_mark_inode_dirty(handle, dir);
 140         ext3_dec_count(handle, inode);
 141         if (!inode->i_nlink)
 142                 ext3_orphan_add(handle, inode);
 143 -       inode->i_ctime = dir->i_ctime;
 144 +       inode->i_ctime = ext3_current_time(inode);
 145         ext3_mark_inode_dirty(handle, inode);
 146         retval = 0;
 147
 148 @@ -2255,7 +2255,7 @@ retry:
 149         if (IS_DIRSYNC(dir))
 150                 handle->h_sync = 1;
 151
 152 -       inode->i_ctime = CURRENT_TIME_SEC;
 153 +       inode->i_ctime = ext3_current_time(inode);
 154         ext3_inc_count(handle, inode);
 155         atomic_inc(&inode->i_count);
 156
 157 @@ -2357,7 +2357,7 @@ static int ext3_rename (struct inode * o
 158          * Like most other Unix systems, set the ctime for inodes on a
 159          * rename.
 160          */
 161 -       old_inode->i_ctime = CURRENT_TIME_SEC;
 162 +       old_inode->i_ctime = ext3_current_time(old_inode);
 163         ext3_mark_inode_dirty(handle, old_inode);
 164
 165         /*
 166 @@ -2390,9 +2390,9 @@ static int ext3_rename (struct inode * o
 167
 168         if (new_inode) {
 169                 ext3_dec_count(handle, new_inode);
 170 -               new_inode->i_ctime = CURRENT_TIME_SEC;
 171 +               new_inode->i_ctime = ext3_current_time(new_inode);
 172         }
 173 -       old_dir->i_ctime = old_dir->i_mtime = CURRENT_TIME_SEC;
 174 +       old_dir->i_ctime = old_dir->i_mtime = ext3_current_time(old_dir);
 175         ext3_update_dx_flag(old_dir);
 176         if (dir_bh) {
 177                 BUFFER_TRACE(dir_bh, "get_write_access");
 178 Index: linux-2.6.9-rhel4/fs/ext3/super.c
 179 ===================================================================
 180 --- linux-2.6.9-rhel4.orig/fs/ext3/super.c
 181 +++ linux-2.6.9-rhel4/fs/ext3/super.c
 182 @@ -1468,6 +1468,8 @@ static int ext3_fill_super (struct super
 183                                 sbi->s_inode_size);
 184                         goto failed_mount;
 185                 }
 186 +               if (sbi->s_inode_size > EXT3_GOOD_OLD_INODE_SIZE)
 187 +                       sb->s_time_gran = 1 << (EXT3_EPOCH_BITS - 2);
 188         }
 189         sbi->s_frag_size = EXT3_MIN_FRAG_SIZE <<
 190                                    le32_to_cpu(es->s_log_frag_size);
 191 @@ -1654,6 +1656,32 @@ static int ext3_fill_super (struct super
 192         }
 193
 194         ext3_setup_super (sb, es, sb->s_flags & MS_RDONLY);
 195 +
 196 +       /* determine the minimum size of new large inodes, if present */
 197 +       if (sbi->s_inode_size > EXT3_GOOD_OLD_INODE_SIZE) {
 198 +               sbi->s_want_extra_isize = sizeof(struct ext3_inode) -
 199 +                                                    EXT3_GOOD_OLD_INODE_SIZE;
 200 +               if (EXT3_HAS_RO_COMPAT_FEATURE(sb,
 201 +                                      EXT3_FEATURE_RO_COMPAT_EXTRA_ISIZE)) {
 202 +                       if (sbi->s_want_extra_isize <
 203 +                           le16_to_cpu(es->s_want_extra_isize))
 204 +                               sbi->s_want_extra_isize =
 205 +                                       le16_to_cpu(es->s_want_extra_isize);
 206 +                       if (sbi->s_want_extra_isize <
 207 +                           le16_to_cpu(es->s_min_extra_isize))
 208 +                               sbi->s_want_extra_isize =
 209 +                                       le16_to_cpu(es->s_min_extra_isize);
 210 +               }
 211 +       }
 212 +       /* Check if enough inode space is available */
 213 +       if (EXT3_GOOD_OLD_INODE_SIZE + sbi->s_want_extra_isize >
 214 +                                                       sbi->s_inode_size) {
 215 +               sbi->s_want_extra_isize = sizeof(struct ext3_inode) -
 216 +                                                      EXT3_GOOD_OLD_INODE_SIZE;
 217 +               printk(KERN_INFO "EXT3-fs: required extra inode space not"
 218 +                       "available.\n");
 219 +       }
 220 +
 221         /*
 222          * akpm: core read_super() calls in here with the superblock locked.
 223          * That deadlocks, because orphan cleanup needs to lock the superblock
 224 Index: linux-2.6.9-rhel4/fs/ext3/xattr.c
 225 ===================================================================
 226 --- linux-2.6.9-rhel4.orig/fs/ext3/xattr.c
 227 +++ linux-2.6.9-rhel4/fs/ext3/xattr.c
 228 @@ -1305,7 +1305,7 @@ getblk_failed:
 229
 230         /* Update the inode. */
 231         EXT3_I(inode)->i_file_acl = new_bh ? new_bh->b_blocknr : 0;
 232 -       inode->i_ctime = CURRENT_TIME_SEC;
 233 +       inode->i_ctime = ext3_current_time(inode);
 234         ext3_mark_inode_dirty(handle, inode);
 235         if (IS_SYNC(inode))
 236                 handle->h_sync = 1;
 237 Index: linux-2.6.9-rhel4/include/linux/ext3_fs.h
 238 ===================================================================
 239 --- linux-2.6.9-rhel4.orig/include/linux/ext3_fs.h
 240 +++ linux-2.6.9-rhel4/include/linux/ext3_fs.h
 241 @@ -258,7 +258,7 @@ struct ext3_inode {
 242         __le16  i_uid;          /* Low 16 bits of Owner Uid */
 243         __le32  i_size;         /* Size in bytes */
 244         __le32  i_atime;        /* Access time */
 245 -       __le32  i_ctime;        /* Creation time */
 246 +       __le32  i_ctime;        /* Inode Change time */
 247         __le32  i_mtime;        /* Modification time */
 248         __le32  i_dtime;        /* Deletion Time */
 249         __le16  i_gid;          /* Low 16 bits of Group Id */
 250 @@ -307,10 +307,73 @@ struct ext3_inode {
 251         } osd2;                         /* OS dependent 2 */
 252         __u16   i_extra_isize;
 253         __u16   i_pad1;
 254 +       __le32  i_ctime_extra;  /* extra Change time      (nsec << 2 | epoch) */
 255 +       __le32  i_mtime_extra;  /* extra Modification time(nsec << 2 | epoch) */
 256 +       __le32  i_atime_extra;  /* extra Access time      (nsec << 2 | epoch) */
 257 +       __le32  i_crtime;       /* File Creation time */
 258 +       __le32  i_crtime_extra; /* extra File Creation time (nsec << 2 | epoch) */
 259  };
 260
 261  #define i_size_high    i_dir_acl
 262
 263 +#define EXT3_EPOCH_BITS 2
 264 +#define EXT3_EPOCH_MASK ((1 << EXT3_EPOCH_BITS) - 1)
 265 +#define EXT3_NSEC_MASK  (~0UL << EXT3_EPOCH_BITS)
 266 +
 267 +#define EXT3_FITS_IN_INODE(ext3_inode, einode, field)   \
 268 +       ((offsetof(typeof(*ext3_inode), field) +        \
 269 +        sizeof((ext3_inode)->field))                   \
 270 +        <= (EXT3_GOOD_OLD_INODE_SIZE +                 \
 271 +            (einode)->i_extra_isize))                  \
 272 +
 273 +static inline __le32 ext3_encode_extra_time(struct timespec *time)
 274 +{
 275 +       return cpu_to_le32((sizeof(time->tv_sec) > 4 ?
 276 +                          time->tv_sec >> 32 : 0) |
 277 +                          ((time->tv_nsec << 2) & EXT3_NSEC_MASK));
 278 +}
 279 +
 280 +static inline void ext3_decode_extra_time(struct timespec *time, __le32 extra) {
 281 +       if (sizeof(time->tv_sec) > 4)
 282 +              time->tv_sec |= (__u64)(le32_to_cpu(extra) & EXT3_EPOCH_MASK)
 283 +                                      << 32;
 284 +       time->tv_nsec = (le32_to_cpu(extra) & EXT3_NSEC_MASK) >> 2;
 285 +}
 286 +
 287 +#define EXT3_INODE_SET_XTIME(xtime, inode, raw_inode)                   \
 288 +do {                                                                    \
 289 +       (raw_inode)->xtime = cpu_to_le32((inode)->xtime.tv_sec);                 \
 290 +       if (EXT3_FITS_IN_INODE(raw_inode, EXT3_I(inode), xtime ## _extra))\
 291 +              (raw_inode)->xtime ## _extra =                            \
 292 +                              ext3_encode_extra_time(&(inode)->xtime);  \
 293 +} while (0)
 294 +
 295 +#define EXT3_EINODE_SET_XTIME(xtime, einode, raw_inode)                         \
 296 +do {                                                                    \
 297 +       if (EXT3_FITS_IN_INODE(raw_inode, einode, xtime))                \
 298 +              (raw_inode)->xtime = cpu_to_le32((einode)->xtime.tv_sec); \
 299 +       if (EXT3_FITS_IN_INODE(raw_inode, einode, xtime ## _extra))      \
 300 +              (raw_inode)->xtime ## _extra =                            \
 301 +                              ext3_encode_extra_time(&(einode)->xtime); \
 302 +} while (0)
 303 +
 304 +#define EXT3_INODE_GET_XTIME(xtime, inode, raw_inode)                   \
 305 +do {                                                                    \
 306 +       (inode)->xtime.tv_sec = le32_to_cpu((raw_inode)->xtime);                 \
 307 +       if (EXT3_FITS_IN_INODE(raw_inode, EXT3_I(inode), xtime ## _extra))\
 308 +              ext3_decode_extra_time(&(inode)->xtime,                   \
 309 +                                     raw_inode->xtime ## _extra);       \
 310 +} while (0)
 311 +
 312 +#define EXT3_EINODE_GET_XTIME(xtime, einode, raw_inode)                         \
 313 +do {                                                                    \
 314 +       if (EXT3_FITS_IN_INODE(raw_inode, einode, xtime))                \
 315 +              (einode)->xtime.tv_sec = le32_to_cpu((raw_inode)->xtime); \
 316 +       if (EXT3_FITS_IN_INODE(raw_inode, einode, xtime ## _extra))      \
 317 +              ext3_decode_extra_time(&(einode)->xtime,                  \
 318 +                                     raw_inode->xtime ## _extra);       \
 319 +} while (0)
 320 +
 321  #if defined(__KERNEL__) || defined(__linux__)
 322  #define i_reserved1    osd1.linux1.l_i_reserved1
 323  #define i_frag         osd2.linux2.l_i_frag
 324 @@ -490,7 +553,9 @@ struct ext3_super_block {
 325         __u16   s_reserved_word_pad;
 326         __le32  s_default_mount_opts;
 327         __le32  s_first_meta_bg;        /* First metablock block group */
 328 -       __u32   s_reserved[190];        /* Padding to the end of the block */
 329 +       __le16  s_min_extra_isize;      /* All inodes have at least # bytes */
 330 +       __le16  s_want_extra_isize;     /* New inodes should reserve # bytes */
 331 +       __u32   s_reserved[189];        /* Padding to the end of the block */
 332  };
 333
 334  #ifdef __KERNEL__
 335 @@ -565,6 +630,8 @@ static inline struct ext3_inode_info *EX
 336  #define EXT3_FEATURE_RO_COMPAT_LARGE_FILE      0x0002
 337  #define EXT3_FEATURE_RO_COMPAT_BTREE_DIR       0x0004
 338  #define EXT3_FEATURE_RO_COMPAT_DIR_NLINK       0x0020
 339 +#define EXT3_FEATURE_RO_COMPAT_EXTRA_ISIZE     0x0040
 340 +
 341
 342  #define EXT3_FEATURE_INCOMPAT_COMPRESSION      0x0001
 343  #define EXT3_FEATURE_INCOMPAT_FILETYPE         0x0002
 344 @@ -581,6 +648,7 @@ static inline struct ext3_inode_info *EX
 345  #define EXT3_FEATURE_RO_COMPAT_SUPP    (EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER| \
 346                                          EXT3_FEATURE_RO_COMPAT_LARGE_FILE| \
 347                                          EXT3_FEATURE_RO_COMPAT_DIR_NLINK| \
 348 +                                        EXT3_FEATURE_RO_COMPAT_EXTRA_ISIZE| \
 349                                          EXT3_FEATURE_RO_COMPAT_BTREE_DIR)
 350
 351  /*
 352 @@ -708,6 +776,12 @@ static inline struct ext3_inode *ext3_ra
 353         return (struct ext3_inode *) (iloc->bh->b_data + iloc->offset);
 354  }
 355
 356 +static inline struct timespec ext3_current_time(struct inode *inode)
 357 +{
 358 +       return (inode->i_sb->s_time_gran < NSEC_PER_SEC) ?
 359 +               current_fs_time(inode->i_sb) : CURRENT_TIME_SEC;
 360 +}
 361 +
 362  /*
 363   * This structure is stuffed into the struct file's private_data field
 364   * for directories.  It is where we put information so that we can do
 365 Index: linux-2.6.9-rhel4/include/linux/ext3_fs_i.h
 366 ===================================================================
 367 --- linux-2.6.9-rhel4.orig/include/linux/ext3_fs_i.h
 368 +++ linux-2.6.9-rhel4/include/linux/ext3_fs_i.h
 369 @@ -130,6 +130,7 @@ struct ext3_inode_info {
 370         struct inode vfs_inode;
 371
 372         __u32 i_cached_extent[4];
 373 +       struct timespec i_crtime;
 374  };
 375
 376  #endif /* _LINUX_EXT3_FS_I */
 377 Index: linux-2.6.9-rhel4/include/linux/ext3_fs_sb.h
 378 ===================================================================
 379 --- linux-2.6.9-rhel4.orig/include/linux/ext3_fs_sb.h
 380 +++ linux-2.6.9-rhel4/include/linux/ext3_fs_sb.h
 381 @@ -122,6 +122,8 @@ struct ext3_sb_info {
 382         spinlock_t s_bal_lock;
 383         unsigned long s_mb_buddies_generated;
 384         unsigned long long s_mb_generation_time;
 385 +
 386 +       unsigned int s_want_extra_isize; /* New inodes should reserve # bytes */
 387  };
 388
 389  #define EXT3_GROUP_INFO(sb, group)                                        \