lustre/lvfs/fsfilt_ext3.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  lustre/lib/fsfilt_ext3.c
   5  *  Lustre filesystem abstraction routines
   6  *
   7  *  Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   8  *   Author: Andreas Dilger <adilger@clusterfs.com>
   9  *
  10  *   This file is part of Lustre, http://www.lustre.org.
  11  *
  12  *   Lustre is free software; you can redistribute it and/or
  13  *   modify it under the terms of version 2 of the GNU General Public
  14  *   License as published by the Free Software Foundation.
  15  *
  16  *   Lustre is distributed in the hope that it will be useful,
  17  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19  *   GNU General Public License for more details.
  20  *
  21  *   You should have received a copy of the GNU General Public License
  22  *   along with Lustre; if not, write to the Free Software
  23  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  24  */
  25
  26 #define DEBUG_SUBSYSTEM S_FILTER
  27
  28 #include <linux/init.h>
  29 #include <linux/module.h>
  30 #include <linux/fs.h>
  31 #include <linux/jbd.h>
  32 #include <linux/slab.h>
  33 #include <linux/pagemap.h>
  34 #include <linux/quotaops.h>
  35 #include <linux/ext3_fs.h>
  36 #include <linux/ext3_jbd.h>
  37 #include <linux/ext3_extents.h>
  38 #include <linux/version.h>
  39 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
  40 #include <linux/ext3_xattr.h>
  41 #else
  42 #include <ext3/xattr.h>
  43 #endif
  44
  45 #include <libcfs/kp30.h>
  46 #include <linux/lustre_fsfilt.h>
  47 #include <linux/obd.h>
  48 #include <linux/obd_class.h>
  49 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
  50 #include <linux/module.h>
  51 #include <linux/iobuf.h>
  52 #endif
  53
  54 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,7))
  55 # define lock_24kernel() lock_kernel()
  56 # define unlock_24kernel() unlock_kernel()
  57 #else
  58 # define lock_24kernel() do {} while (0)
  59 # define unlock_24kernel() do {} while (0)
  60 #endif
  61
  62 static kmem_cache_t *fcb_cache;
  63 static atomic_t fcb_cache_count = ATOMIC_INIT(0);
  64
  65 struct fsfilt_cb_data {
  66         struct journal_callback cb_jcb; /* jbd private data - MUST BE FIRST */
  67         fsfilt_cb_t cb_func;            /* MDS/OBD completion function */
  68         struct obd_device *cb_obd;      /* MDS/OBD completion device */
  69         __u64 cb_last_num;              /* MDS/OST last committed operation */
  70         void *cb_data;                  /* MDS/OST completion function data */
  71 };
  72
  73 #ifndef EXT3_XATTR_INDEX_TRUSTED        /* temporary until we hit l28 kernel */
  74 #define EXT3_XATTR_INDEX_TRUSTED        4
  75 #endif
  76
  77 /*
  78  * We don't currently need any additional blocks for rmdir and
  79  * unlink transactions because we are storing the OST oa_id inside
  80  * the inode (which we will be changing anyways as part of this
  81  * transaction).
  82  */
  83 static void *fsfilt_ext3_start(struct inode *inode, int op, void *desc_private,
  84                                int logs)
  85 {
  86         /* For updates to the last recieved file */
  87         int nblocks = EXT3_SINGLEDATA_TRANS_BLOCKS;
  88         journal_t *journal;
  89         void *handle;
  90
  91         if (current->journal_info) {
  92                 CDEBUG(D_INODE, "increasing refcount on %p\n",
  93                        current->journal_info);
  94                 goto journal_start;
  95         }
  96
  97         if (logs)
  98                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
  99                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
 100
 101         switch(op) {
 102         case FSFILT_OP_RMDIR:
 103         case FSFILT_OP_UNLINK:
 104                 /* delete one file + create/update logs for each stripe */
 105                 nblocks += EXT3_DELETE_TRANS_BLOCKS;
 106                 /*nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 107                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;*/
 108                 break;
 109         case FSFILT_OP_RENAME:
 110                 /* modify additional directory */
 111                 nblocks += EXT3_SINGLEDATA_TRANS_BLOCKS;
 112                 /* no break */
 113         case FSFILT_OP_SYMLINK:
 114                 /* additional block + block bitmap + GDT for long symlink */
 115                 nblocks += 3;
 116                 /* no break */
 117         case FSFILT_OP_CREATE:
 118                 /* create/update logs for each stripe */
 119                 /*nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 120                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;*/
 121                 /* no break */
 122         case FSFILT_OP_MKDIR:
 123         case FSFILT_OP_MKNOD:
 124                 /* modify one inode + block bitmap + GDT */
 125                 nblocks += 3;
 126                 /* no break */
 127         case FSFILT_OP_LINK:
 128                 /* modify parent directory */
 129                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 130                         EXT3_DATA_TRANS_BLOCKS;
 131                 break;
 132         case FSFILT_OP_SETATTR:
 133                 /* Setattr on inode */
 134                 nblocks += 1;
 135                 break;
 136         case FSFILT_OP_CANCEL_UNLINK:
 137                 /* blocks for log header bitmap update OR
 138                  * blocks for catalog header bitmap update + unlink of logs */
 139                 nblocks = (LLOG_CHUNK_SIZE >> inode->i_blkbits) +
 140                         EXT3_DELETE_TRANS_BLOCKS * logs;
 141                 break;
 142         case FSFILT_OP_NOOP:
 143                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
 144                 break;
 145         default: CERROR("unknown transaction start op %d\n", op);
 146                 LBUG();
 147         }
 148
 149         LASSERT(current->journal_info == desc_private);
 150         journal = EXT3_SB(inode->i_sb)->s_journal;
 151         if (nblocks > journal->j_max_transaction_buffers) {
 152                 CERROR("too many credits %d for op %ux%u using %d instead\n",
 153                        nblocks, op, logs, journal->j_max_transaction_buffers);
 154                 nblocks = journal->j_max_transaction_buffers;
 155         }
 156
 157  journal_start:
 158         LASSERTF(nblocks > 0, "can't start %d credit transaction\n", nblocks);
 159         lock_24kernel();
 160         handle = journal_start(EXT3_JOURNAL(inode), nblocks);
 161         unlock_24kernel();
 162
 163         if (!IS_ERR(handle))
 164                 LASSERT(current->journal_info == handle);
 165         else
 166                 CERROR("error starting handle for op %u (%u credits): rc %ld\n",
 167                        op, nblocks, PTR_ERR(handle));
 168         return handle;
 169 }
 170
 171 /*
 172  * Calculate the number of buffer credits needed to write multiple pages in
 173  * a single ext3 transaction.  No, this shouldn't be here, but as yet ext3
 174  * doesn't have a nice API for calculating this sort of thing in advance.
 175  *
 176  * See comment above ext3_writepage_trans_blocks for details.  We assume
 177  * no data journaling is being done, but it does allow for all of the pages
 178  * being non-contiguous.  If we are guaranteed contiguous pages we could
 179  * reduce the number of (d)indirect blocks a lot.
 180  *
 181  * With N blocks per page and P pages, for each inode we have at most:
 182  * N*P indirect
 183  * min(N*P, blocksize/4 + 1) dindirect blocks
 184  * niocount tindirect
 185  *
 186  * For the entire filesystem, we have at most:
 187  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 188  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 189  * objcount inode blocks
 190  * 1 superblock
 191  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 192  *
 193  * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
 194  */
 195 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso,
 196                                       int niocount, struct niobuf_local *nb)
 197 {
 198         struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
 199         __u64 next_indir;
 200         const int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 201         int nbitmaps = 0, ngdblocks;
 202         int needed = objcount + 1; /* inodes + superblock */
 203         int i, j;
 204
 205         for (i = 0, j = 0; i < objcount; i++, fso++) {
 206                 /* two or more dindirect blocks in case we cross boundary */
 207                 int ndind = (long)((nb[j + fso->fso_bufcnt - 1].offset -
 208                                     nb[j].offset) >>
 209                                    sb->s_blocksize_bits) /
 210                         (EXT3_ADDR_PER_BLOCK(sb) * EXT3_ADDR_PER_BLOCK(sb));
 211                 nbitmaps += min(fso->fso_bufcnt, ndind > 0 ? ndind : 2);
 212
 213                 /* leaf, indirect, tindirect blocks for first block */
 214                 nbitmaps += blockpp + 2;
 215
 216                 j += fso->fso_bufcnt;
 217         }
 218
 219         next_indir = nb[0].offset +
 220                 (EXT3_ADDR_PER_BLOCK(sb) << sb->s_blocksize_bits);
 221         for (i = 1; i < niocount; i++) {
 222                 if (nb[i].offset >= next_indir) {
 223                         nbitmaps++;     /* additional indirect */
 224                         next_indir = nb[i].offset +
 225                                 (EXT3_ADDR_PER_BLOCK(sb)<<sb->s_blocksize_bits);
 226                 } else if (nb[i].offset != nb[i - 1].offset + sb->s_blocksize) {
 227                         nbitmaps++;     /* additional indirect */
 228                 }
 229                 nbitmaps += blockpp;    /* each leaf in different group? */
 230         }
 231
 232         ngdblocks = nbitmaps;
 233         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 234                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 235         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 236                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 237
 238         needed += nbitmaps + ngdblocks;
 239
 240         /* last_rcvd update */
 241         needed += EXT3_DATA_TRANS_BLOCKS;
 242
 243 #if defined(CONFIG_QUOTA) && !defined(__x86_64__) /* XXX */
 244         /* We assume that there will be 1 bit set in s_dquot.flags for each
 245          * quota file that is active.  This is at least true for now.
 246          */
 247         needed += hweight32(sb_any_quota_enabled(sb)) *
 248                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 249 #endif
 250
 251         return needed;
 252 }
 253
 254 /* We have to start a huge journal transaction here to hold all of the
 255  * metadata for the pages being written here.  This is necessitated by
 256  * the fact that we do lots of prepare_write operations before we do
 257  * any of the matching commit_write operations, so even if we split
 258  * up to use "smaller" transactions none of them could complete until
 259  * all of them were opened.  By having a single journal transaction,
 260  * we eliminate duplicate reservations for common blocks like the
 261  * superblock and group descriptors or bitmaps.
 262  *
 263  * We will start the transaction here, but each prepare_write will
 264  * add a refcount to the transaction, and each commit_write will
 265  * remove a refcount.  The transaction will be closed when all of
 266  * the pages have been written.
 267  */
 268 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
 269                                    int niocount, struct niobuf_local *nb,
 270                                    void *desc_private, int logs)
 271 {
 272         journal_t *journal;
 273         handle_t *handle;
 274         int needed;
 275         ENTRY;
 276
 277         LASSERT(current->journal_info == desc_private);
 278         journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
 279         needed = fsfilt_ext3_credits_needed(objcount, fso, niocount, nb);
 280
 281         /* The number of blocks we could _possibly_ dirty can very large.
 282          * We reduce our request if it is absurd (and we couldn't get that
 283          * many credits for a single handle anyways).
 284          *
 285          * At some point we have to limit the size of I/Os sent at one time,
 286          * increase the size of the journal, or we have to calculate the
 287          * actual journal requirements more carefully by checking all of
 288          * the blocks instead of being maximally pessimistic.  It remains to
 289          * be seen if this is a real problem or not.
 290          */
 291         if (needed > journal->j_max_transaction_buffers) {
 292                 CERROR("want too many journal credits (%d) using %d instead\n",
 293                        needed, journal->j_max_transaction_buffers);
 294                 needed = journal->j_max_transaction_buffers;
 295         }
 296
 297         LASSERTF(needed > 0, "can't start %d credit transaction\n", needed);
 298         lock_24kernel();
 299         handle = journal_start(journal, needed);
 300         unlock_24kernel();
 301         if (IS_ERR(handle)) {
 302                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 303                        PTR_ERR(handle));
 304         } else {
 305                 LASSERT(handle->h_buffer_credits >= needed);
 306                 LASSERT(current->journal_info == handle);
 307         }
 308
 309         RETURN(handle);
 310 }
 311
 312 static int fsfilt_ext3_commit(struct super_block *sb, struct inode *inode,
 313                               void *h, int force_sync)
 314 {
 315         int rc;
 316         handle_t *handle = h;
 317
 318         LASSERT(current->journal_info == handle);
 319         if (force_sync)
 320                 handle->h_sync = 1; /* recovery likes this */
 321
 322         lock_24kernel();
 323         rc = journal_stop(handle);
 324         unlock_24kernel();
 325
 326         return rc;
 327 }
 328
 329 static int fsfilt_ext3_commit_async(struct inode *inode, void *h,
 330                                     void **wait_handle)
 331 {
 332         unsigned long tid;
 333         transaction_t *transaction;
 334 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 335         unsigned long rtid;
 336 #endif
 337         handle_t *handle = h;
 338         journal_t *journal;
 339         int rc;
 340
 341         LASSERT(current->journal_info == handle);
 342
 343         lock_24kernel();
 344         transaction = handle->h_transaction;
 345         journal = transaction->t_journal;
 346         tid = transaction->t_tid;
 347         /* we don't want to be blocked */
 348         handle->h_sync = 0;
 349         rc = journal_stop(handle);
 350         if (rc) {
 351                 CERROR("error while stopping transaction: %d\n", rc);
 352                 unlock_kernel();
 353                 return rc;
 354         }
 355 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 356         rtid = log_start_commit(journal, transaction);
 357         if (rtid != tid)
 358                 CERROR("strange race: %lu != %lu\n",
 359                        (unsigned long) tid, (unsigned long) rtid);
 360 #else
 361         log_start_commit(journal, tid);
 362 #endif
 363         unlock_24kernel();
 364
 365         *wait_handle = (void *) tid;
 366         CDEBUG(D_INODE, "commit async: %lu\n", (unsigned long) tid);
 367         return 0;
 368 }
 369
 370 static int fsfilt_ext3_commit_wait(struct inode *inode, void *h)
 371 {
 372         tid_t tid = (tid_t)(long)h;
 373
 374         CDEBUG(D_INODE, "commit wait: %lu\n", (unsigned long) tid);
 375         if (is_journal_aborted(EXT3_JOURNAL(inode)))
 376                 return -EIO;
 377
 378         log_wait_commit(EXT3_JOURNAL(inode), tid);
 379
 380         return 0;
 381 }
 382
 383 static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
 384                                struct iattr *iattr, int do_trunc)
 385 {
 386         struct inode *inode = dentry->d_inode;
 387         int rc;
 388
 389         lock_24kernel();
 390
 391         /* A _really_ horrible hack to avoid removing the data stored
 392          * in the block pointers; this is really the "small" stripe MD data.
 393          * We can avoid further hackery by virtue of the MDS file size being
 394          * zero all the time (which doesn't invoke block truncate at unlink
 395          * time), so we assert we never change the MDS file size from zero. */
 396         if (iattr->ia_valid & ATTR_SIZE && !do_trunc) {
 397                 /* ATTR_SIZE would invoke truncate: clear it */
 398                 iattr->ia_valid &= ~ATTR_SIZE;
 399                 EXT3_I(inode)->i_disksize = inode->i_size = iattr->ia_size;
 400
 401                 /* make sure _something_ gets set - so new inode
 402                  * goes to disk (probably won't work over XFS */
 403                 if (!(iattr->ia_valid & (ATTR_MODE | ATTR_MTIME | ATTR_CTIME))){
 404                         iattr->ia_valid |= ATTR_MODE;
 405                         iattr->ia_mode = inode->i_mode;
 406                 }
 407         }
 408
 409         /* Don't allow setattr to change file type */
 410         iattr->ia_mode = (inode->i_mode & S_IFMT)|(iattr->ia_mode & ~S_IFMT);
 411
 412         /* We set these flags on the client, but have already checked perms
 413          * so don't confuse inode_change_ok. */
 414         iattr->ia_valid &= ~(ATTR_MTIME_SET | ATTR_ATIME_SET);
 415
 416         if (inode->i_op->setattr) {
 417                 rc = inode->i_op->setattr(dentry, iattr);
 418         } else {
 419                 rc = inode_change_ok(inode, iattr);
 420                 if (!rc)
 421                         rc = inode_setattr(inode, iattr);
 422         }
 423
 424         unlock_24kernel();
 425
 426         return rc;
 427 }
 428
 429 static int fsfilt_ext3_iocontrol(struct inode * inode, struct file *file,
 430                                  unsigned int cmd, unsigned long arg)
 431 {
 432         int rc = 0;
 433         ENTRY;
 434
 435         if (inode->i_fop->ioctl)
 436                 rc = inode->i_fop->ioctl(inode, file, cmd, arg);
 437         else
 438                 RETURN(-ENOTTY);
 439
 440         RETURN(rc);
 441 }
 442
 443 static int fsfilt_ext3_set_xattr(struct inode * inode, void *handle, char *name,
 444                                  void *buffer, int buffer_size)
 445 {
 446         int rc = 0;
 447
 448         lock_24kernel();
 449
 450         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
 451                                    name, buffer, buffer_size, 0);
 452         unlock_24kernel();
 453         if (rc)
 454                 CERROR("set xattr %s from inode %lu: rc %d\n",
 455                        name,  inode->i_ino, rc);
 456         return rc;
 457 }
 458
 459 static int fsfilt_ext3_get_xattr(struct inode *inode, char *name,
 460                                  void *buffer, int buffer_size)
 461 {
 462         int rc = 0;
 463
 464         lock_24kernel();
 465
 466         rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
 467                             name, buffer, buffer_size);
 468         unlock_24kernel();
 469
 470         if (buffer == NULL)
 471                 return (rc == -ENODATA) ? 0 : rc;
 472         if (rc < 0) {
 473                 CDEBUG(D_INFO, "error getting EA %s from inode %lu: rc %d\n",
 474                        name,  inode->i_ino, rc);
 475                 memset(buffer, 0, buffer_size);
 476                 return (rc == -ENODATA) ? 0 : rc;
 477         }
 478
 479         return rc;
 480 }
 481
 482 static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
 483                               void *lmm, int lmm_size,
 484                               enum ea_type type)
 485 {
 486         int rc;
 487
 488         switch(type) {
 489         case EA_LOV:
 490                 rc = fsfilt_ext3_set_xattr(inode, handle,
 491                                            XATTR_LUSTRE_MDS_LOV_EA,
 492                                            lmm, lmm_size);
 493                 break;
 494         case EA_MEA:
 495                 rc = fsfilt_ext3_set_xattr(inode, handle,
 496                                            XATTR_LUSTRE_MDS_MEA_EA,
 497                                            lmm, lmm_size);
 498                 break;
 499         case EA_SID:
 500                 rc = fsfilt_ext3_set_xattr(inode, handle,
 501                                            XATTR_LUSTRE_MDS_SID_EA,
 502                                            lmm, lmm_size);
 503                 break;
 504         case EA_PID:
 505                  rc = fsfilt_ext3_set_xattr(inode, handle,
 506                                             XATTR_LUSTRE_MDS_PID_EA,
 507                                             lmm, lmm_size);
 508                 break;
 509         case EA_KEY:
 510                 rc = fsfilt_ext3_set_xattr(inode, handle,
 511                                            XATTR_LUSTRE_MDS_KEY_EA,
 512                                            lmm, lmm_size);
 513                 break;
 514         default:
 515                 return -EINVAL;
 516         }
 517
 518         return rc;
 519 }
 520
 521 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm,
 522                               int lmm_size, enum ea_type type)
 523 {
 524         int rc;
 525
 526         switch (type) {
 527         case EA_LOV:
 528                 rc = fsfilt_ext3_get_xattr(inode,
 529                                            XATTR_LUSTRE_MDS_LOV_EA,
 530                                            lmm, lmm_size);
 531                 break;
 532         case EA_MEA:
 533                 rc = fsfilt_ext3_get_xattr(inode,
 534                                            XATTR_LUSTRE_MDS_MEA_EA,
 535                                            lmm, lmm_size);
 536                 break;
 537         case EA_SID:
 538                 rc = fsfilt_ext3_get_xattr(inode,
 539                                            XATTR_LUSTRE_MDS_SID_EA,
 540                                            lmm, lmm_size);
 541                 break;
 542         case EA_PID:
 543                 rc = fsfilt_ext3_get_xattr(inode,
 544                                            XATTR_LUSTRE_MDS_PID_EA,
 545                                            lmm, lmm_size);
 546                 break;
 547         case EA_KEY:
 548                 rc = fsfilt_ext3_get_xattr(inode,
 549                                            XATTR_LUSTRE_MDS_KEY_EA,
 550                                            lmm, lmm_size);
 551                 break;
 552         default:
 553                 return -EINVAL;
 554         }
 555
 556         return rc;
 557 }
 558
 559 static int fsfilt_ext3_send_bio(int rw, struct inode *inode, void *bio)
 560 {
 561         int rc = 0;
 562 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 563         submit_bio(rw, (struct bio *)bio);
 564 #else
 565         struct bio *b = (struct kiobuf *)bio;
 566         int blocks_per_page;
 567
 568         rc = brw_kiovec(rw, 1, &b, inode->i_dev,
 569                         b->blocks, 1 << inode->i_blkbits);
 570
 571         blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 572
 573         if (rc != (1 << inode->i_blkbits) * b->nr_pages * blocks_per_page) {
 574                 CERROR("short write?  expected %d, wrote %d\n",
 575                        (1 << inode->i_blkbits) * b->nr_pages *
 576                        blocks_per_page, rc);
 577         }
 578 #endif
 579         return rc;
 580 }
 581
 582 static struct page *fsfilt_ext3_getpage(struct inode *inode, long int index)
 583 {
 584         int rc;
 585         struct page *page;
 586
 587         page = grab_cache_page(inode->i_mapping, index);
 588         if (page == NULL)
 589                 return ERR_PTR(-ENOMEM);
 590
 591         if (PageUptodate(page)) {
 592                 unlock_page(page);
 593                 return page;
 594         }
 595
 596         rc = inode->i_mapping->a_ops->readpage(NULL, page);
 597         if (rc < 0) {
 598                 page_cache_release(page);
 599                 return ERR_PTR(rc);
 600         }
 601
 602         return page;
 603 }
 604
 605 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
 606                                     loff_t *off)
 607 {
 608         struct inode *inode = file->f_dentry->d_inode;
 609         int rc = 0;
 610
 611         if (S_ISREG(inode->i_mode))
 612                 rc = file->f_op->read(file, buf, count, off);
 613         else {
 614                 const int blkbits = inode->i_sb->s_blocksize_bits;
 615                 const int blksize = inode->i_sb->s_blocksize;
 616
 617                 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
 618                        count, inode->i_ino, *off);
 619                 while (count > 0) {
 620                         struct buffer_head *bh;
 621
 622                         bh = NULL;
 623                         if (*off < inode->i_size) {
 624                                 int err = 0;
 625
 626                                 bh = ext3_bread(NULL, inode, *off >> blkbits,
 627                                                 0, &err);
 628
 629                                 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
 630
 631                                 if (bh) {
 632                                         memcpy(buf, bh->b_data, blksize);
 633                                         brelse(bh);
 634                                 } else if (err) {
 635                                         /* XXX in theory we should just fake
 636                                          * this buffer and continue like ext3,
 637                                          * especially if this is a partial read
 638                                          */
 639                                         CERROR("error read dir %lu+%llu: %d\n",
 640                                                inode->i_ino, *off, err);
 641                                         RETURN(err);
 642                                 }
 643                         }
 644                         if (!bh) {
 645                                 struct ext3_dir_entry_2 *fake = (void *)buf;
 646
 647                                 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
 648                                 memset(fake, 0, sizeof(*fake));
 649                                 fake->rec_len = cpu_to_le32(blksize);
 650                         }
 651                         count -= blksize;
 652                         buf += blksize;
 653                         *off += blksize;
 654                         rc += blksize;
 655                 }
 656         }
 657
 658         return rc;
 659 }
 660
 661 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
 662 {
 663         struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
 664
 665         fcb->cb_func(fcb->cb_obd, fcb->cb_last_num, fcb->cb_data, error);
 666
 667         OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
 668         atomic_dec(&fcb_cache_count);
 669 }
 670
 671 static int fsfilt_ext3_add_journal_cb(struct obd_device *obd,
 672                                       struct super_block *sb,
 673                                       __u64 last_num, void *handle,
 674                                       fsfilt_cb_t cb_func,
 675                                       void *cb_data)
 676 {
 677         struct fsfilt_cb_data *fcb;
 678
 679         OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
 680         if (fcb == NULL)
 681                 RETURN(-ENOMEM);
 682
 683         atomic_inc(&fcb_cache_count);
 684         fcb->cb_func = cb_func;
 685         fcb->cb_obd = obd;
 686         fcb->cb_last_num = last_num;
 687         fcb->cb_data = cb_data;
 688
 689         CDEBUG(D_EXT2, "set callback for last_num: "LPD64"\n", last_num);
 690         lock_24kernel();
 691         journal_callback_set(handle, fsfilt_ext3_cb_func,
 692                              (struct journal_callback *)fcb);
 693         unlock_24kernel();
 694         return 0;
 695 }
 696
 697 /*
 698  * We need to hack the return value for the free inode counts because
 699  * the current EA code requires one filesystem block per inode with EAs,
 700  * so it is possible to run out of blocks before we run out of inodes.
 701  *
 702  * This can be removed when the ext3 EA code is fixed.
 703  */
 704 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
 705 {
 706         struct kstatfs sfs;
 707         int rc;
 708
 709         memset(&sfs, 0, sizeof(sfs));
 710
 711         rc = sb->s_op->statfs(sb, &sfs);
 712
 713         if (!rc && sfs.f_bfree < sfs.f_ffree) {
 714                 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
 715                 sfs.f_ffree = sfs.f_bfree;
 716         }
 717
 718         statfs_pack(osfs, &sfs);
 719         return rc;
 720 }
 721
 722 static int fsfilt_ext3_sync(struct super_block *sb)
 723 {
 724         return ext3_force_commit(sb);
 725 }
 726
 727 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
 728 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 729 #define ext3_up_truncate_sem(inode)  up_write(&EXT3_I(inode)->truncate_sem);
 730 #define ext3_down_truncate_sem(inode)  down_write(&EXT3_I(inode)->truncate_sem);
 731 #else
 732 #define ext3_up_truncate_sem(inode)  up(&EXT3_I(inode)->truncate_sem);
 733 #define ext3_down_truncate_sem(inode)  down(&EXT3_I(inode)->truncate_sem);
 734 #endif
 735
 736 #include <linux/lustre_version.h>
 737 #if EXT3_EXT_MAGIC == 0xf301
 738 #define ee_start e_start
 739 #define ee_block e_block
 740 #define ee_len   e_num
 741 #endif
 742 #ifndef EXT3_BB_MAX_BLOCKS
 743 #define ext3_mb_new_blocks(handle, inode, goal, count, aflags, err) \
 744         ext3_new_blocks(handle, inode, count, goal, err)
 745 #endif
 746
 747 struct bpointers {
 748         unsigned long *blocks;
 749         int *created;
 750         unsigned long start;
 751         int num;
 752         int init_num;
 753         int create;
 754 };
 755 static int ext3_ext_find_goal(struct inode *inode, struct ext3_ext_path *path,
 756                                 unsigned long block, int *aflags)
 757 {
 758         struct ext3_inode_info *ei = EXT3_I(inode);
 759         unsigned long bg_start;
 760         unsigned long colour;
 761         int depth;
 762
 763         if (path) {
 764                 struct ext3_extent *ex;
 765                 depth = path->p_depth;
 766
 767                 /* try to predict block placement */
 768                 if ((ex = path[depth].p_ext)) {
 769 #if 0
 770                         /* This prefers to eat into a contiguous extent
 771                          * rather than find an extent that the whole
 772                          * request will fit into.  This can fragment data
 773                          * block allocation and prevents our lovely 1M I/Os
 774                          * from reaching the disk intact. */
 775
 776                         if (ex->ee_block + ex->ee_len == block)
 777                                 *aflags |= 1;
 778 #endif
 779                         return ex->ee_start + (block - ex->ee_block);
 780                 }
 781
 782                 /* it looks index is empty
 783                  * try to find starting from index itself */
 784                 if (path[depth].p_bh)
 785                         return path[depth].p_bh->b_blocknr;
 786         }
 787
 788         /* OK. use inode's group */
 789         bg_start = (ei->i_block_group * EXT3_BLOCKS_PER_GROUP(inode->i_sb)) +
 790                 le32_to_cpu(EXT3_SB(inode->i_sb)->s_es->s_first_data_block);
 791         colour = (current->pid % 16) *
 792                         (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 793         return bg_start + colour + block;
 794 }
 795
 796 static int ext3_ext_new_extent_cb(struct ext3_extents_tree *tree,
 797                                   struct ext3_ext_path *path,
 798                                   struct ext3_extent *newex, int exist)
 799 {
 800         struct inode *inode = tree->inode;
 801         struct bpointers *bp = tree->private;
 802         int count, err, goal;
 803         unsigned long pblock;
 804         unsigned long tgen;
 805         loff_t new_i_size;
 806         handle_t *handle;
 807         int i, aflags = 0;
 808
 809         i = EXT_DEPTH(tree);
 810         EXT_ASSERT(i == path->p_depth);
 811         EXT_ASSERT(path[i].p_hdr);
 812
 813         if (exist) {
 814                 err = EXT_CONTINUE;
 815                 goto map;
 816         }
 817
 818         if (bp->create == 0) {
 819                 i = 0;
 820                 if (newex->ee_block < bp->start)
 821                         i = bp->start - newex->ee_block;
 822                 if (i >= newex->ee_len)
 823                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 824                                         i, newex->ee_len);
 825                 for (; i < newex->ee_len && bp->num; i++) {
 826                         *(bp->created) = 0;
 827                         bp->created++;
 828                         *(bp->blocks) = 0;
 829                         bp->blocks++;
 830                         bp->num--;
 831                         bp->start++;
 832                 }
 833
 834                 return EXT_CONTINUE;
 835         }
 836         tgen = EXT_GENERATION(tree);
 837         count = ext3_ext_calc_credits_for_insert(tree, path);
 838         ext3_up_truncate_sem(inode);
 839         lock_24kernel();
 840         handle = journal_start(EXT3_JOURNAL(inode), count + EXT3_ALLOC_NEEDED + 1);
 841         unlock_24kernel();
 842         if (IS_ERR(handle)) {
 843                 ext3_down_truncate_sem(inode);
 844                 return PTR_ERR(handle);
 845         }
 846
 847         if (tgen != EXT_GENERATION(tree)) {
 848                 /* the tree has changed. so path can be invalid at moment */
 849                 lock_24kernel();
 850                 journal_stop(handle);
 851                 unlock_24kernel();
 852                 ext3_down_truncate_sem(inode);
 853                 return EXT_REPEAT;
 854         }
 855         ext3_down_truncate_sem(inode);
 856         count = newex->ee_len;
 857         goal = ext3_ext_find_goal(inode, path, newex->ee_block, &aflags);
 858         aflags |= 2; /* block have been already reserved */
 859         pblock = ext3_mb_new_blocks(handle, inode, goal, &count, aflags, &err);
 860         if (!pblock)
 861                 goto out;
 862         EXT_ASSERT(count <= newex->ee_len);
 863
 864         /* insert new extent */
 865         newex->ee_start = pblock;
 866         newex->ee_len = count;
 867         err = ext3_ext_insert_extent(handle, tree, path, newex);
 868         if (err)
 869                 goto out;
 870
 871         /* correct on-disk inode size */
 872         if (newex->ee_len > 0) {
 873                 new_i_size = (loff_t) newex->ee_block + newex->ee_len;
 874                 new_i_size = new_i_size << inode->i_blkbits;
 875                 if (new_i_size > EXT3_I(inode)->i_disksize) {
 876                         EXT3_I(inode)->i_disksize = new_i_size;
 877                         err = ext3_mark_inode_dirty(handle, inode);
 878                 }
 879         }
 880 out:
 881         lock_24kernel();
 882         journal_stop(handle);
 883         unlock_24kernel();
 884 map:
 885         if (err >= 0) {
 886                 /* map blocks */
 887                 if (bp->num == 0) {
 888                         CERROR("hmm. why do we find this extent?\n");
 889                         CERROR("initial space: %lu:%u\n",
 890                                 bp->start, bp->init_num);
 891                         CERROR("current extent: %u/%u/%u %d\n",
 892                                 newex->ee_block, newex->ee_len,
 893                                 newex->ee_start, exist);
 894                 }
 895                 i = 0;
 896                 if (newex->ee_block < bp->start)
 897                         i = bp->start - newex->ee_block;
 898                 if (i >= newex->ee_len)
 899                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 900                                         i, newex->ee_len);
 901                 for (; i < newex->ee_len && bp->num; i++) {
 902                         *(bp->created) = (exist == 0 ? 1 : 0);
 903                         bp->created++;
 904                         *(bp->blocks) = newex->ee_start + i;
 905                         bp->blocks++;
 906                         bp->num--;
 907                         bp->start++;
 908                 }
 909         }
 910         return err;
 911 }
 912
 913 int fsfilt_map_nblocks(struct inode *inode, unsigned long block,
 914                        unsigned long num, unsigned long *blocks,
 915                        int *created, int create)
 916 {
 917         struct ext3_extents_tree tree;
 918         struct bpointers bp;
 919         int err, i;
 920
 921         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
 922                 block, block + num, (unsigned) inode->i_ino);
 923
 924         ext3_init_tree_desc(&tree, inode);
 925         tree.private = &bp;
 926         bp.blocks = blocks;
 927         bp.created = created;
 928         bp.start = block;
 929         bp.init_num = bp.num = num;
 930         bp.create = create;
 931
 932         ext3_down_truncate_sem(inode);
 933         err = ext3_ext_walk_space(&tree, block, num, ext3_ext_new_extent_cb);
 934         ext3_ext_invalidate_cache(&tree);
 935         ext3_up_truncate_sem(inode);
 936
 937         /* unmap underlying pages/buffers from blockdevice mapping */
 938         if (create) {
 939                 struct block_device *bdev = inode->i_sb->s_bdev;
 940                 for (i = 0; i < num; i++) {
 941                         if (created[i] == 0)
 942                                 continue;
 943                         unmap_underlying_metadata(bdev, blocks[i]);
 944                 }
 945         }
 946         return err;
 947 }
 948
 949 int fsfilt_ext3_map_ext_inode_pages(struct inode *inode, struct page **page,
 950                                     int pages, unsigned long *blocks,
 951                                     int *created, int create)
 952 {
 953         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 954         int rc = 0, i = 0;
 955         struct page *fp = NULL;
 956         int clen = 0;
 957
 958         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 959                 inode->i_ino, pages, (*page)->index);
 960
 961         /* pages are sorted already. so, we just have to find
 962          * contig. space and process them properly */
 963         while (i < pages) {
 964                 if (fp == NULL) {
 965                         /* start new extent */
 966                         fp = *page++;
 967                         clen = 1;
 968                         i++;
 969                         continue;
 970                 } else if (fp->index + clen == (*page)->index) {
 971                         /* continue the extent */
 972                         page++;
 973                         clen++;
 974                         i++;
 975                         continue;
 976                 }
 977
 978                 /* process found extent */
 979                 rc = fsfilt_map_nblocks(inode, fp->index * blocks_per_page,
 980                                         clen * blocks_per_page, blocks,
 981                                         created, create);
 982                 if (rc)
 983                         GOTO(cleanup, rc);
 984
 985                 /* look for next extent */
 986                 fp = NULL;
 987                 blocks += blocks_per_page * clen;
 988                 created += blocks_per_page * clen;
 989         }
 990
 991         if (fp)
 992                 rc = fsfilt_map_nblocks(inode, fp->index * blocks_per_page,
 993                                         clen * blocks_per_page, blocks,
 994                                         created, create);
 995 cleanup:
 996         return rc;
 997 }
 998 #endif
 999
1000 extern int ext3_map_inode_page(struct inode *inode, struct page *page,
1001                                unsigned long *blocks, int *created, int create);
1002 int fsfilt_ext3_map_bm_inode_pages(struct inode *inode, struct page **page,
1003                                    int pages, unsigned long *blocks,
1004                                    int *created, int create)
1005 {
1006         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1007         unsigned long *b;
1008         int rc = 0, i, *cr;
1009
1010         for (i = 0, cr = created, b = blocks; i < pages; i++, page++) {
1011                 rc = ext3_map_inode_page(inode, *page, b, cr, create);
1012                 if (rc) {
1013                         CERROR("ino %lu, blk %lu cr %u create %d: rc %d\n",
1014                                inode->i_ino, *b, *cr, create, rc);
1015                         break;
1016                 }
1017
1018                 b += blocks_per_page;
1019                 cr += blocks_per_page;
1020         }
1021         return rc;
1022 }
1023
1024 int fsfilt_ext3_map_inode_pages(struct inode *inode, struct page **page,
1025                                 int pages, unsigned long *blocks,
1026                                 int *created, int create,
1027                                 struct semaphore *optional_sem)
1028 {
1029         int rc;
1030 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
1031         if (EXT3_I(inode)->i_flags & EXT3_EXTENTS_FL) {
1032                 rc = fsfilt_ext3_map_ext_inode_pages(inode, page, pages,
1033                                                      blocks, created, create);
1034                 return rc;
1035         }
1036 #endif
1037         if (optional_sem != NULL)
1038                 down(optional_sem);
1039         rc = fsfilt_ext3_map_bm_inode_pages(inode, page, pages, blocks,
1040                                             created, create);
1041         if (optional_sem != NULL)
1042                 up(optional_sem);
1043
1044         return rc;
1045 }
1046
1047 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
1048                                int nblocks, loff_t newsize);
1049 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
1050                                       int nblocks, loff_t newsize)
1051 {
1052         return ext3_prep_san_write(inode, blocks, nblocks, newsize);
1053 }
1054
1055 static int fsfilt_ext3_read_record(struct file * file, void *buf,
1056                                    int size, loff_t *offs)
1057 {
1058         struct inode *inode = file->f_dentry->d_inode;
1059         unsigned long block;
1060         struct buffer_head *bh;
1061         int err, blocksize, csize, boffs;
1062
1063         /* prevent reading after eof */
1064         lock_kernel();
1065         if (inode->i_size < *offs + size) {
1066                 size = inode->i_size - *offs;
1067                 unlock_kernel();
1068                 if (size < 0) {
1069                         CERROR("size %llu is too short for read %u@%llu\n",
1070                                inode->i_size, size, *offs);
1071                         return -EIO;
1072                 } else if (size == 0) {
1073                         return 0;
1074                 }
1075         } else {
1076                 unlock_kernel();
1077         }
1078
1079         blocksize = 1 << inode->i_blkbits;
1080
1081         while (size > 0) {
1082                 block = *offs >> inode->i_blkbits;
1083                 boffs = *offs & (blocksize - 1);
1084                 csize = min(blocksize - boffs, size);
1085                 bh = ext3_bread(NULL, inode, block, 0, &err);
1086                 if (!bh) {
1087                         CERROR("can't read block: %d\n", err);
1088                         return err;
1089                 }
1090
1091                 memcpy(buf, bh->b_data + boffs, csize);
1092                 brelse(bh);
1093
1094                 *offs += csize;
1095                 buf += csize;
1096                 size -= csize;
1097         }
1098         return 0;
1099 }
1100
1101 static int fsfilt_ext3_write_record(struct file *file, void *buf, int bufsize,
1102                                     loff_t *offs, int force_sync)
1103 {
1104         struct buffer_head *bh = NULL;
1105         unsigned long block;
1106         struct inode *inode = file->f_dentry->d_inode;
1107         loff_t old_size = inode->i_size, offset = *offs;
1108         loff_t new_size = inode->i_size;
1109         journal_t *journal;
1110         handle_t *handle;
1111         int err = 0, block_count = 0, blocksize, size, boffs;
1112
1113         /* Determine how many transaction credits are needed */
1114         blocksize = 1 << inode->i_blkbits;
1115         block_count = (*offs & (blocksize - 1)) + bufsize;
1116         block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
1117
1118         journal = EXT3_SB(inode->i_sb)->s_journal;
1119         lock_24kernel();
1120         handle = journal_start(journal,
1121                                block_count * EXT3_DATA_TRANS_BLOCKS + 2);
1122         unlock_24kernel();
1123         if (IS_ERR(handle)) {
1124                 CERROR("can't start transaction\n");
1125                 return PTR_ERR(handle);
1126         }
1127
1128         while (bufsize > 0) {
1129                 if (bh != NULL)
1130                         brelse(bh);
1131
1132                 block = offset >> inode->i_blkbits;
1133                 boffs = offset & (blocksize - 1);
1134                 size = min(blocksize - boffs, bufsize);
1135                 bh = ext3_bread(handle, inode, block, 1, &err);
1136                 if (!bh) {
1137                         CERROR("can't read/create block: %d\n", err);
1138                         goto out;
1139                 }
1140
1141                 err = ext3_journal_get_write_access(handle, bh);
1142                 if (err) {
1143                         CERROR("journal_get_write_access() returned error %d\n",
1144                                err);
1145                         goto out;
1146                 }
1147                 LASSERT(bh->b_data + boffs + size <= bh->b_data + bh->b_size);
1148                 memcpy(bh->b_data + boffs, buf, size);
1149                 err = ext3_journal_dirty_metadata(handle, bh);
1150                 if (err) {
1151                         CERROR("journal_dirty_metadata() returned error %d\n",
1152                                err);
1153                         goto out;
1154                 }
1155                 if (offset + size > new_size)
1156                         new_size = offset + size;
1157                 offset += size;
1158                 bufsize -= size;
1159                 buf += size;
1160         }
1161
1162         if (force_sync)
1163                 handle->h_sync = 1; /* recovery likes this */
1164 out:
1165         if (bh)
1166                 brelse(bh);
1167
1168         /* correct in-core and on-disk sizes */
1169         if (new_size > inode->i_size) {
1170                 lock_kernel();
1171                 if (new_size > inode->i_size)
1172                         inode->i_size = new_size;
1173                 if (inode->i_size > EXT3_I(inode)->i_disksize)
1174                         EXT3_I(inode)->i_disksize = inode->i_size;
1175                 if (inode->i_size > old_size)
1176                         mark_inode_dirty(inode);
1177                 unlock_kernel();
1178         }
1179
1180         lock_24kernel();
1181         journal_stop(handle);
1182         unlock_24kernel();
1183
1184         if (err == 0)
1185                 *offs = offset;
1186         return err;
1187 }
1188
1189 static int fsfilt_ext3_setup(struct obd_device *obd, struct super_block *sb)
1190 {
1191 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1192         struct mds_obd *mds = &obd->u.mds;
1193 #endif
1194 #if 0
1195         EXT3_SB(sb)->dx_lock = fsfilt_ext3_dx_lock;
1196         EXT3_SB(sb)->dx_unlock = fsfilt_ext3_dx_unlock;
1197 #endif
1198 #ifdef S_PDIROPS
1199         CWARN("Enabling PDIROPS\n");
1200         set_opt(EXT3_SB(sb)->s_mount_opt, PDIROPS);
1201         sb->s_flags |= S_PDIROPS;
1202 #endif
1203         /* setup mdsnum in underlying fs */
1204 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1205         if (mds->mds_md_obd) {
1206                 struct ext3_sb_info *sbi = EXT3_SB(sb);
1207                 struct ext3_super_block *es = sbi->s_es;
1208                 handle_t *handle;
1209                 int err;
1210
1211                 if (!EXT3_HAS_INCOMPAT_FEATURE(sb, EXT3_FEATURE_INCOMPAT_MDSNUM)) {
1212                         CWARN("%s: set mdsnum %d in ext3\n",
1213                               obd->obd_name, mds->mds_num);
1214                         lock_24kernel();
1215                         handle = journal_start(sbi->s_journal, 1);
1216                         unlock_24kernel();
1217                         LASSERT(!IS_ERR(handle));
1218                         err = ext3_journal_get_write_access(handle, sbi->s_sbh);
1219                         LASSERT(err == 0);
1220                         EXT3_SET_INCOMPAT_FEATURE(sb,
1221                                                 EXT3_FEATURE_INCOMPAT_MDSNUM);
1222                         es->s_mdsnum = mds->mds_num;
1223                         err = ext3_journal_dirty_metadata(handle, sbi->s_sbh);
1224                         LASSERT(err == 0);
1225                         lock_24kernel();
1226                         journal_stop(handle);
1227                         unlock_24kernel();
1228                 } else {
1229                         CWARN("%s: mdsnum initialized to %u in ext3fs\n",
1230                                 obd->obd_name, es->s_mdsnum);
1231                 }
1232                 sbi->s_mdsnum = es->s_mdsnum;
1233         }
1234 #endif
1235         return 0;
1236 }
1237
1238 extern int ext3_add_dir_entry(struct dentry *dentry);
1239 extern int ext3_del_dir_entry(struct dentry *dentry);
1240
1241 static int fsfilt_ext3_add_dir_entry(struct obd_device *obd,
1242                                      struct dentry *parent,
1243                                      char *name, int namelen,
1244                                      unsigned long ino,
1245                                      unsigned long generation,
1246                                      unsigned long mds,
1247                                      unsigned long fid)
1248 {
1249 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1250         struct dentry *dentry;
1251         int err;
1252         LASSERT(ino != 0);
1253         LASSERT(namelen != 0);
1254         dentry = ll_lookup_one_len(name, parent, namelen);
1255         if (IS_ERR(dentry)) {
1256                 CERROR("can't lookup %*s in %lu/%lu: %d\n", dentry->d_name.len,
1257                        dentry->d_name.name, dentry->d_inode->i_ino,
1258                        (unsigned long) dentry->d_inode->i_generation,
1259                        (int) PTR_ERR(dentry));
1260                 RETURN(PTR_ERR(dentry));
1261         }
1262         if (dentry->d_inode != NULL || dentry->d_flags & DCACHE_CROSS_REF) {
1263                 CERROR("dentry %*s(0x%p) found\n", dentry->d_name.len,
1264                        dentry->d_name.name, dentry);
1265                 l_dput(dentry);
1266                 RETURN(-EEXIST);
1267         }
1268
1269         /* mds_reint_rename() may use this method to add dir entry
1270          * that points onto local inode. and we don't want to find
1271          * it cross-ref by subsequent lookups */
1272         d_drop(dentry);
1273
1274         dentry->d_flags |= DCACHE_CROSS_REF;
1275         dentry->d_inum = ino;
1276         dentry->d_mdsnum = mds;
1277         dentry->d_generation = generation;
1278         dentry->d_fid = fid;
1279         lock_24kernel();
1280         err = ext3_add_dir_entry(dentry);
1281         unlock_24kernel();
1282
1283         l_dput(dentry);
1284
1285         RETURN(err);
1286 #else
1287 #error "rebuild kernel and lustre with ext3-mds-num patch!"
1288         LASSERT(0);
1289 #endif
1290 }
1291
1292 static int fsfilt_ext3_del_dir_entry(struct obd_device *obd,
1293                                  struct dentry *dentry)
1294 {
1295 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1296         int err;
1297         lock_24kernel();
1298         err = ext3_del_dir_entry(dentry);
1299         unlock_24kernel();
1300         if (err == 0)
1301                 d_drop(dentry);
1302         return err;
1303 #else
1304 #error "rebuild kernel and lustre with ext3-mds-num patch!"
1305         LASSERT(0);
1306 #endif
1307 }
1308
1309 /* If fso is NULL, op is FSFILT operation, otherwise op is number of fso
1310    objects. Logs is number of logfiles to update */
1311 static int fsfilt_ext3_get_op_len(int op, struct fsfilt_objinfo *fso, int logs)
1312 {
1313         if ( !fso ) {
1314                 switch(op) {
1315                 case FSFILT_OP_CREATE:
1316                                  /* directory leaf, index & indirect & EA*/
1317                         return 4 + 3 * logs;
1318                 case FSFILT_OP_UNLINK:
1319                         return 3 * logs;
1320                 }
1321         } else {
1322                 int i;
1323                 int needed = 0;
1324                 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
1325                 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
1326                 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
1327                 for (i = 0; i < op; i++, fso++) {
1328                         int nblocks = fso->fso_bufcnt * blockpp;
1329                         int ndindirect = min(nblocks, addrpp + 1);
1330                         int nindir = nblocks + ndindirect + 1;
1331
1332                         needed += nindir;
1333                 }
1334                 return needed + 3 * logs;
1335         }
1336
1337         return 0;
1338 }
1339
1340
1341 #define EXTENTS_EA "write_extents"
1342 #define EXTENTS_EA_SIZE 64
1343
1344 int ext3_ext_in_ea_alloc_space(struct inode *, int, const char *, unsigned long, unsigned long);
1345 int ext3_ext_in_ea_remove_space(struct inode *, int, const char *, unsigned long, unsigned long);
1346 int ext3_ext_in_ea_get_extents(struct inode *, int, const char *, char **, int *);
1347 int ext3_ext_in_ea_get_extents_num(struct inode *, int, const char *, int *);
1348
1349 static int fsfilt_ext3_insert_extents_ea(struct inode *inode,
1350                                       unsigned long from,
1351                                       unsigned long num)
1352 {
1353         int rc = 0;
1354
1355         rc = ext3_ext_in_ea_alloc_space(inode, EXT3_XATTR_INDEX_TRUSTED,
1356                                         EXTENTS_EA, from, num);
1357         return rc;
1358 }
1359
1360 static int fsfilt_ext3_remove_extents_ea(struct inode *inode,
1361                                          unsigned long from,
1362                                          unsigned long num)
1363 {
1364         int rc = 0;
1365
1366         rc = ext3_ext_in_ea_remove_space(inode, EXT3_XATTR_INDEX_TRUSTED,
1367                                          EXTENTS_EA, from, num);
1368         return rc;
1369 }
1370
1371 extern int ext3_init_tree_in_ea(struct inode *inode, int name_index,
1372                                 const char *eaname, int size);
1373
1374 static int fsfilt_ext3_init_extents_ea(struct inode *inode)
1375 {
1376         int rc = 0;
1377
1378         rc = ext3_init_tree_in_ea(inode, EXT3_XATTR_INDEX_TRUSTED,
1379                                   EXTENTS_EA, 64);
1380         return rc;
1381 }
1382
1383 static int fsfilt_ext3_get_inode_write_extents(struct inode *inode,
1384                                          char **pbuf, int *size)
1385 {
1386         int rc = 0;
1387
1388         rc = ext3_ext_in_ea_get_extents(inode, EXT3_XATTR_INDEX_TRUSTED,
1389                                         EXTENTS_EA,  pbuf, size);
1390         return rc;
1391 }
1392
1393 static int fsfilt_ext3_get_write_extents_num(struct inode *inode, int *size)
1394 {
1395         int rc = 0;
1396
1397         rc = ext3_ext_in_ea_get_extents_num(inode, EXT3_XATTR_INDEX_TRUSTED,
1398                                             EXTENTS_EA, size);
1399         return rc;
1400 }
1401
1402 static struct fsfilt_operations fsfilt_ext3_ops = {
1403         .fs_type                    = "ext3",
1404         .fs_owner                   = THIS_MODULE,
1405         .fs_start                   = fsfilt_ext3_start,
1406         .fs_brw_start               = fsfilt_ext3_brw_start,
1407         .fs_commit                  = fsfilt_ext3_commit,
1408         .fs_commit_async            = fsfilt_ext3_commit_async,
1409         .fs_commit_wait             = fsfilt_ext3_commit_wait,
1410         .fs_setattr                 = fsfilt_ext3_setattr,
1411         .fs_iocontrol               = fsfilt_ext3_iocontrol,
1412         .fs_set_md                  = fsfilt_ext3_set_md,
1413         .fs_get_md                  = fsfilt_ext3_get_md,
1414         .fs_readpage                = fsfilt_ext3_readpage,
1415         .fs_add_journal_cb          = fsfilt_ext3_add_journal_cb,
1416         .fs_statfs                  = fsfilt_ext3_statfs,
1417         .fs_sync                    = fsfilt_ext3_sync,
1418         .fs_map_inode_pages         = fsfilt_ext3_map_inode_pages,
1419         .fs_prep_san_write          = fsfilt_ext3_prep_san_write,
1420         .fs_write_record            = fsfilt_ext3_write_record,
1421         .fs_read_record             = fsfilt_ext3_read_record,
1422         .fs_setup                   = fsfilt_ext3_setup,
1423         .fs_getpage                 = fsfilt_ext3_getpage,
1424         .fs_send_bio                = fsfilt_ext3_send_bio,
1425         .fs_set_xattr               = fsfilt_ext3_set_xattr,
1426         .fs_get_xattr               = fsfilt_ext3_get_xattr,
1427         .fs_get_op_len              = fsfilt_ext3_get_op_len,
1428         .fs_add_dir_entry           = fsfilt_ext3_add_dir_entry,
1429         .fs_del_dir_entry           = fsfilt_ext3_del_dir_entry,
1430         .fs_init_extents_ea         = fsfilt_ext3_init_extents_ea,
1431         .fs_insert_extents_ea       = fsfilt_ext3_insert_extents_ea,
1432         .fs_remove_extents_ea       = fsfilt_ext3_remove_extents_ea,
1433         .fs_get_inode_write_extents = fsfilt_ext3_get_inode_write_extents,
1434         .fs_get_write_extents_num   = fsfilt_ext3_get_write_extents_num,
1435 };
1436
1437 static int __init fsfilt_ext3_init(void)
1438 {
1439         int rc;
1440
1441         fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
1442                                       sizeof(struct fsfilt_cb_data), 0,
1443                                       0, NULL, NULL);
1444         if (!fcb_cache) {
1445                 CERROR("error allocating fsfilt journal callback cache\n");
1446                 GOTO(out, rc = -ENOMEM);
1447         }
1448
1449         rc = fsfilt_register_ops(&fsfilt_ext3_ops);
1450
1451         if (rc)
1452                 kmem_cache_destroy(fcb_cache);
1453 out:
1454         return rc;
1455 }
1456
1457 static void __exit fsfilt_ext3_exit(void)
1458 {
1459         fsfilt_unregister_ops(&fsfilt_ext3_ops);
1460         LASSERTF(kmem_cache_destroy(fcb_cache) == 0,
1461                  "can't free fsfilt callback cache: count %d\n",
1462                  atomic_read(&fcb_cache_count));
1463 }
1464
1465 module_init(fsfilt_ext3_init);
1466 module_exit(fsfilt_ext3_exit);
1467
1468 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1469 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
1470 MODULE_LICENSE("GPL");