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 <linux/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
  55 static kmem_cache_t *fcb_cache;
  56 static atomic_t fcb_cache_count = ATOMIC_INIT(0);
  57
  58 struct fsfilt_cb_data {
  59         struct journal_callback cb_jcb; /* jbd private data - MUST BE FIRST */
  60         fsfilt_cb_t cb_func;            /* MDS/OBD completion function */
  61         struct obd_device *cb_obd;      /* MDS/OBD completion device */
  62         __u64 cb_last_rcvd;             /* MDS/OST last committed operation */
  63         void *cb_data;                  /* MDS/OST completion function data */
  64 };
  65
  66 #ifndef EXT3_XATTR_INDEX_TRUSTED        /* temporary until we hit l28 kernel */
  67 #define EXT3_XATTR_INDEX_TRUSTED        4
  68 #endif
  69 #define XATTR_LUSTRE_MDS_LOV_EA         "lov"
  70
  71 /*
  72  * We don't currently need any additional blocks for rmdir and
  73  * unlink transactions because we are storing the OST oa_id inside
  74  * the inode (which we will be changing anyways as part of this
  75  * transaction).
  76  */
  77 static void *fsfilt_ext3_start(struct inode *inode, int op, void *desc_private,
  78                                int logs)
  79 {
  80         /* For updates to the last recieved file */
  81         int nblocks = EXT3_SINGLEDATA_TRANS_BLOCKS;
  82         journal_t *journal;
  83         void *handle;
  84
  85         if (current->journal_info) {
  86                 CDEBUG(D_INODE, "increasing refcount on %p\n",
  87                        current->journal_info);
  88                 goto journal_start;
  89         }
  90
  91         /* XXX BUG 3188 -- must return to one set of opcodes */
  92         /* FIXME - cache hook */
  93         if (op & 0x20) {
  94                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
  95                 op = op & ~0x20;
  96         }
  97
  98         /* FIXME - kml */
  99         if (op & 0x10) {
 100                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
 101                 op = op & ~0x10;
 102         }
 103
 104         switch(op) {
 105         case FSFILT_OP_RMDIR:
 106         case FSFILT_OP_UNLINK:
 107                 /* delete one file + create/update logs for each stripe */
 108                 nblocks += EXT3_DELETE_TRANS_BLOCKS;
 109                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 110                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
 111                 break;
 112         case FSFILT_OP_RENAME:
 113                 /* modify additional directory */
 114                 nblocks += EXT3_SINGLEDATA_TRANS_BLOCKS;
 115                 /* no break */
 116         case FSFILT_OP_SYMLINK:
 117                 /* additional block + block bitmap + GDT for long symlink */
 118                 nblocks += 3;
 119                 /* no break */
 120         case FSFILT_OP_CREATE:
 121                 /* create/update logs for each stripe */
 122                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 123                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
 124                 /* no break */
 125         case FSFILT_OP_MKDIR:
 126         case FSFILT_OP_MKNOD:
 127                 /* modify one inode + block bitmap + GDT */
 128                 nblocks += 3;
 129                 /* no break */
 130         case FSFILT_OP_LINK:
 131                 /* modify parent directory */
 132                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 133                         EXT3_DATA_TRANS_BLOCKS;
 134                 break;
 135         case FSFILT_OP_SETATTR:
 136                 /* Setattr on inode */
 137                 nblocks += 1;
 138                 break;
 139         case FSFILT_OP_CANCEL_UNLINK:
 140                 /* blocks for log header bitmap update OR
 141                  * blocks for catalog header bitmap update + unlink of logs */
 142                 nblocks = (LLOG_CHUNK_SIZE >> inode->i_blkbits) +
 143                         EXT3_DELETE_TRANS_BLOCKS * logs;
 144                 break;
 145         case FSFILT_OP_NOOP:
 146                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
 147                 break;
 148         default: CERROR("unknown transaction start op %d\n", op);
 149                 LBUG();
 150         }
 151
 152         LASSERT(current->journal_info == desc_private);
 153         journal = EXT3_SB(inode->i_sb)->s_journal;
 154         if (nblocks > journal->j_max_transaction_buffers) {
 155                 CERROR("too many credits %d for op %ux%u using %d instead\n",
 156                        nblocks, op, logs, journal->j_max_transaction_buffers);
 157                 nblocks = journal->j_max_transaction_buffers;
 158         }
 159
 160  journal_start:
 161         lock_kernel();
 162         handle = journal_start(EXT3_JOURNAL(inode), nblocks);
 163         unlock_kernel();
 164
 165         if (!IS_ERR(handle))
 166                 LASSERT(current->journal_info == handle);
 167         else
 168                 CERROR("error starting handle for op %u (%u credits): rc %ld\n",
 169                        op, nblocks, PTR_ERR(handle));
 170         return handle;
 171 }
 172
 173 /*
 174  * Calculate the number of buffer credits needed to write multiple pages in
 175  * a single ext3 transaction.  No, this shouldn't be here, but as yet ext3
 176  * doesn't have a nice API for calculating this sort of thing in advance.
 177  *
 178  * See comment above ext3_writepage_trans_blocks for details.  We assume
 179  * no data journaling is being done, but it does allow for all of the pages
 180  * being non-contiguous.  If we are guaranteed contiguous pages we could
 181  * reduce the number of (d)indirect blocks a lot.
 182  *
 183  * With N blocks per page and P pages, for each inode we have at most:
 184  * N*P indirect
 185  * min(N*P, blocksize/4 + 1) dindirect blocks
 186  * niocount tindirect
 187  *
 188  * For the entire filesystem, we have at most:
 189  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 190  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 191  * objcount inode blocks
 192  * 1 superblock
 193  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 194  *
 195  * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
 196  */
 197 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso,
 198                                       int niocount, struct niobuf_local *nb)
 199 {
 200         struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
 201         __u64 next_indir;
 202         const int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 203         int nbitmaps = 0, ngdblocks;
 204         int needed = objcount + 1; /* inodes + superblock */
 205         int i, j;
 206
 207         for (i = 0, j = 0; i < objcount; i++, fso++) {
 208                 /* two or more dindirect blocks in case we cross boundary */
 209                 int ndind = (long)((nb[j + fso->fso_bufcnt - 1].offset -
 210                                     nb[j].offset) >>
 211                                    sb->s_blocksize_bits) /
 212                         (EXT3_ADDR_PER_BLOCK(sb) * EXT3_ADDR_PER_BLOCK(sb));
 213                 nbitmaps += min(fso->fso_bufcnt, ndind > 0 ? ndind : 2);
 214
 215                 /* leaf, indirect, tindirect blocks for first block */
 216                 nbitmaps += blockpp + 2;
 217
 218                 j += fso->fso_bufcnt;
 219         }
 220
 221         next_indir = nb[0].offset +
 222                 (EXT3_ADDR_PER_BLOCK(sb) << sb->s_blocksize_bits);
 223         for (i = 1; i < niocount; i++) {
 224                 if (nb[i].offset >= next_indir) {
 225                         nbitmaps++;     /* additional indirect */
 226                         next_indir = nb[i].offset +
 227                                 (EXT3_ADDR_PER_BLOCK(sb)<<sb->s_blocksize_bits);
 228                 } else if (nb[i].offset != nb[i - 1].offset + sb->s_blocksize) {
 229                         nbitmaps++;     /* additional indirect */
 230                 }
 231                 nbitmaps += blockpp;    /* each leaf in different group? */
 232         }
 233
 234         ngdblocks = nbitmaps;
 235         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 236                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 237         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 238                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 239
 240         needed += nbitmaps + ngdblocks;
 241
 242         /* last_rcvd update */
 243         needed += EXT3_DATA_TRANS_BLOCKS;
 244
 245 #if defined(CONFIG_QUOTA) && !defined(__x86_64__) /* XXX */
 246         /* We assume that there will be 1 bit set in s_dquot.flags for each
 247          * quota file that is active.  This is at least true for now.
 248          */
 249         needed += hweight32(sb_any_quota_enabled(sb)) *
 250                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 251 #endif
 252
 253         return needed;
 254 }
 255
 256 /* We have to start a huge journal transaction here to hold all of the
 257  * metadata for the pages being written here.  This is necessitated by
 258  * the fact that we do lots of prepare_write operations before we do
 259  * any of the matching commit_write operations, so even if we split
 260  * up to use "smaller" transactions none of them could complete until
 261  * all of them were opened.  By having a single journal transaction,
 262  * we eliminate duplicate reservations for common blocks like the
 263  * superblock and group descriptors or bitmaps.
 264  *
 265  * We will start the transaction here, but each prepare_write will
 266  * add a refcount to the transaction, and each commit_write will
 267  * remove a refcount.  The transaction will be closed when all of
 268  * the pages have been written.
 269  */
 270 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
 271                                    int niocount, struct niobuf_local *nb,
 272                                    void *desc_private, int logs)
 273 {
 274         journal_t *journal;
 275         handle_t *handle;
 276         int needed;
 277         ENTRY;
 278
 279         LASSERT(current->journal_info == desc_private);
 280         journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
 281         needed = fsfilt_ext3_credits_needed(objcount, fso, niocount, nb);
 282
 283         /* The number of blocks we could _possibly_ dirty can very large.
 284          * We reduce our request if it is absurd (and we couldn't get that
 285          * many credits for a single handle anyways).
 286          *
 287          * At some point we have to limit the size of I/Os sent at one time,
 288          * increase the size of the journal, or we have to calculate the
 289          * actual journal requirements more carefully by checking all of
 290          * the blocks instead of being maximally pessimistic.  It remains to
 291          * be seen if this is a real problem or not.
 292          */
 293         if (needed > journal->j_max_transaction_buffers) {
 294                 CERROR("want too many journal credits (%d) using %d instead\n",
 295                        needed, journal->j_max_transaction_buffers);
 296                 needed = journal->j_max_transaction_buffers;
 297         }
 298
 299         lock_kernel();
 300         handle = journal_start(journal, needed);
 301         unlock_kernel();
 302         if (IS_ERR(handle)) {
 303                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 304                        PTR_ERR(handle));
 305         } else {
 306                 LASSERT(handle->h_buffer_credits >= needed);
 307                 LASSERT(current->journal_info == handle);
 308         }
 309
 310         RETURN(handle);
 311 }
 312
 313 static int fsfilt_ext3_commit(struct super_block *sb, struct inode *inode,
 314                               void *h, int force_sync)
 315 {
 316         int rc;
 317         handle_t *handle = h;
 318
 319         LASSERT(current->journal_info == handle);
 320         if (force_sync)
 321                 handle->h_sync = 1; /* recovery likes this */
 322
 323         lock_kernel();
 324         rc = journal_stop(handle);
 325         unlock_kernel();
 326
 327         return rc;
 328 }
 329
 330 static int fsfilt_ext3_commit_async(struct inode *inode, void *h,
 331                                     void **wait_handle)
 332 {
 333         unsigned long tid;
 334         transaction_t *transaction;
 335 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 336         unsigned long rtid;
 337 #endif
 338         handle_t *handle = h;
 339         journal_t *journal;
 340         int rc;
 341
 342         LASSERT(current->journal_info == handle);
 343
 344         lock_kernel();
 345         transaction = handle->h_transaction;
 346         journal = transaction->t_journal;
 347         tid = transaction->t_tid;
 348         /* we don't want to be blocked */
 349         handle->h_sync = 0;
 350         rc = journal_stop(handle);
 351         if (rc) {
 352                 CERROR("error while stopping transaction: %d\n", rc);
 353                 unlock_kernel();
 354                 return rc;
 355         }
 356 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 357         rtid = log_start_commit(journal, transaction);
 358         if (rtid != tid)
 359                 CERROR("strange race: %lu != %lu\n",
 360                        (unsigned long) tid, (unsigned long) rtid);
 361 #else
 362         log_start_commit(journal, transaction->t_tid);
 363 #endif
 364         unlock_kernel();
 365
 366         *wait_handle = (void *) tid;
 367         CDEBUG(D_INODE, "commit async: %lu\n", (unsigned long) tid);
 368         return 0;
 369 }
 370
 371 static int fsfilt_ext3_commit_wait(struct inode *inode, void *h)
 372 {
 373         tid_t tid = (tid_t)(long)h;
 374
 375         CDEBUG(D_INODE, "commit wait: %lu\n", (unsigned long) tid);
 376         if (is_journal_aborted(EXT3_JOURNAL(inode)))
 377                 return -EIO;
 378
 379         log_wait_commit(EXT3_JOURNAL(inode), tid);
 380
 381         return 0;
 382 }
 383
 384 static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
 385                                struct iattr *iattr, int do_trunc)
 386 {
 387         struct inode *inode = dentry->d_inode;
 388         int rc;
 389
 390         lock_kernel();
 391
 392         /* A _really_ horrible hack to avoid removing the data stored
 393          * in the block pointers; this is really the "small" stripe MD data.
 394          * We can avoid further hackery by virtue of the MDS file size being
 395          * zero all the time (which doesn't invoke block truncate at unlink
 396          * time), so we assert we never change the MDS file size from zero. */
 397         if (iattr->ia_valid & ATTR_SIZE && !do_trunc) {
 398                 /* ATTR_SIZE would invoke truncate: clear it */
 399                 iattr->ia_valid &= ~ATTR_SIZE;
 400                 EXT3_I(inode)->i_disksize = inode->i_size = iattr->ia_size;
 401
 402                 /* make sure _something_ gets set - so new inode
 403                  * goes to disk (probably won't work over XFS */
 404                 if (!(iattr->ia_valid & (ATTR_MODE | ATTR_MTIME | ATTR_CTIME))){
 405                         iattr->ia_valid |= ATTR_MODE;
 406                         iattr->ia_mode = inode->i_mode;
 407                 }
 408         }
 409
 410         /* Don't allow setattr to change file type */
 411         iattr->ia_mode = (inode->i_mode & S_IFMT)|(iattr->ia_mode & ~S_IFMT);
 412
 413         /* We set these flags on the client, but have already checked perms
 414          * so don't confuse inode_change_ok. */
 415         iattr->ia_valid &= ~(ATTR_MTIME_SET | ATTR_ATIME_SET);
 416
 417         if (inode->i_op->setattr) {
 418                 rc = inode->i_op->setattr(dentry, iattr);
 419         } else {
 420                 rc = inode_change_ok(inode, iattr);
 421                 if (!rc)
 422                         rc = inode_setattr(inode, iattr);
 423         }
 424
 425         unlock_kernel();
 426
 427         return rc;
 428 }
 429
 430 static int fsfilt_ext3_iocontrol(struct inode * inode, struct file *file,
 431                                  unsigned int cmd, unsigned long arg)
 432 {
 433         int rc = 0;
 434         ENTRY;
 435
 436         if (inode->i_fop->ioctl)
 437                 rc = inode->i_fop->ioctl(inode, file, cmd, arg);
 438         else
 439                 RETURN(-ENOTTY);
 440
 441         RETURN(rc);
 442 }
 443
 444 static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
 445                               void *lmm, int lmm_size)
 446 {
 447         int rc;
 448
 449         LASSERT(down_trylock(&inode->i_sem) != 0);
 450
 451         /* keep this when we get rid of OLD_EA (too noisy during conversion) */
 452         if (EXT3_I(inode)->i_file_acl /* || large inode EA flag */)
 453                 CWARN("setting EA on %lu/%u again... interesting\n",
 454                        inode->i_ino, inode->i_generation);
 455
 456         lock_kernel();
 457         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
 458                                    XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size, 0);
 459
 460         unlock_kernel();
 461
 462         if (rc)
 463                 CERROR("error adding MD data to inode %lu: rc = %d\n",
 464                        inode->i_ino, rc);
 465         return rc;
 466 }
 467
 468 /* Must be called with i_sem held */
 469 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm, int lmm_size)
 470 {
 471         int rc;
 472
 473         LASSERT(down_trylock(&inode->i_sem) != 0);
 474         lock_kernel();
 475
 476         rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
 477                             XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size);
 478         unlock_kernel();
 479
 480         /* This gives us the MD size */
 481         if (lmm == NULL)
 482                 return (rc == -ENODATA) ? 0 : rc;
 483
 484         if (rc < 0) {
 485                 CDEBUG(D_INFO, "error getting EA %d/%s from inode %lu: rc %d\n",
 486                        EXT3_XATTR_INDEX_TRUSTED, XATTR_LUSTRE_MDS_LOV_EA,
 487                        inode->i_ino, rc);
 488                 memset(lmm, 0, lmm_size);
 489                 return (rc == -ENODATA) ? 0 : rc;
 490         }
 491
 492         return rc;
 493 }
 494
 495 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 496 static int fsfilt_ext3_send_bio(int rw, struct inode *inode, struct bio *bio)
 497 {
 498         submit_bio(rw, bio);
 499         return 0;
 500 }
 501 #else
 502 static int fsfilt_ext3_send_bio(int rw, struct inode *inode, struct kiobuf *bio)
 503 {
 504         int rc, blocks_per_page;
 505
 506         rc = brw_kiovec(rw, 1, &bio, inode->i_dev,
 507                         bio->blocks, 1 << inode->i_blkbits);
 508
 509         blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 510
 511         if (rc != (1 << inode->i_blkbits) * bio->nr_pages * blocks_per_page) {
 512                 CERROR("short write?  expected %d, wrote %d\n",
 513                        (1 << inode->i_blkbits) * bio->nr_pages *
 514                        blocks_per_page, rc);
 515         }
 516
 517         return rc;
 518 }
 519 #endif
 520
 521 /* FIXME-UMKA: This should be used in 2.6.x io code later. */
 522 static struct page *fsfilt_ext3_getpage(struct inode *inode, long int index)
 523 {
 524         int rc;
 525         struct page *page;
 526
 527         page = grab_cache_page(inode->i_mapping, index);
 528         if (page == NULL)
 529                 return ERR_PTR(-ENOMEM);
 530
 531         if (PageUptodate(page)) {
 532                 unlock_page(page);
 533                 return page;
 534         }
 535
 536         rc = inode->i_mapping->a_ops->readpage(NULL, page);
 537         if (rc < 0) {
 538                 page_cache_release(page);
 539                 return ERR_PTR(rc);
 540         }
 541
 542         return page;
 543 }
 544
 545 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
 546                                     loff_t *off)
 547 {
 548         struct inode *inode = file->f_dentry->d_inode;
 549         int rc = 0;
 550
 551         if (S_ISREG(inode->i_mode))
 552                 rc = file->f_op->read(file, buf, count, off);
 553         else {
 554                 const int blkbits = inode->i_sb->s_blocksize_bits;
 555                 const int blksize = inode->i_sb->s_blocksize;
 556
 557                 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
 558                        count, inode->i_ino, *off);
 559                 while (count > 0) {
 560                         struct buffer_head *bh;
 561
 562                         bh = NULL;
 563                         if (*off < inode->i_size) {
 564                                 int err = 0;
 565
 566                                 bh = ext3_bread(NULL, inode, *off >> blkbits,
 567                                                 0, &err);
 568
 569                                 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
 570
 571                                 if (bh) {
 572                                         memcpy(buf, bh->b_data, blksize);
 573                                         brelse(bh);
 574                                 } else if (err) {
 575                                         /* XXX in theory we should just fake
 576                                          * this buffer and continue like ext3,
 577                                          * especially if this is a partial read
 578                                          */
 579                                         CERROR("error read dir %lu+%llu: %d\n",
 580                                                inode->i_ino, *off, err);
 581                                         RETURN(err);
 582                                 }
 583                         }
 584                         if (!bh) {
 585                                 struct ext3_dir_entry_2 *fake = (void *)buf;
 586
 587                                 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
 588                                 memset(fake, 0, sizeof(*fake));
 589                                 fake->rec_len = cpu_to_le32(blksize);
 590                         }
 591                         count -= blksize;
 592                         buf += blksize;
 593                         *off += blksize;
 594                         rc += blksize;
 595                 }
 596         }
 597
 598         return rc;
 599 }
 600
 601 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
 602 {
 603         struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
 604
 605         fcb->cb_func(fcb->cb_obd, fcb->cb_last_rcvd, fcb->cb_data, error);
 606
 607         OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
 608         atomic_dec(&fcb_cache_count);
 609 }
 610
 611 static int fsfilt_ext3_add_journal_cb(struct obd_device *obd,
 612                                       struct super_block *sb,
 613                                       __u64 last_rcvd,
 614                                       void *handle, fsfilt_cb_t cb_func,
 615                                       void *cb_data)
 616 {
 617         struct fsfilt_cb_data *fcb;
 618
 619         OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
 620         if (fcb == NULL)
 621                 RETURN(-ENOMEM);
 622
 623         atomic_inc(&fcb_cache_count);
 624         fcb->cb_func = cb_func;
 625         fcb->cb_obd = obd;
 626         fcb->cb_last_rcvd = last_rcvd;
 627         fcb->cb_data = cb_data;
 628
 629         CDEBUG(D_EXT2, "set callback for last_rcvd: "LPD64"\n", last_rcvd);
 630         lock_kernel();
 631         journal_callback_set(handle, fsfilt_ext3_cb_func,
 632                              (struct journal_callback *)fcb);
 633         unlock_kernel();
 634
 635         return 0;
 636 }
 637
 638 /*
 639  * We need to hack the return value for the free inode counts because
 640  * the current EA code requires one filesystem block per inode with EAs,
 641  * so it is possible to run out of blocks before we run out of inodes.
 642  *
 643  * This can be removed when the ext3 EA code is fixed.
 644  */
 645 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
 646 {
 647         struct kstatfs sfs;
 648         int rc;
 649
 650         memset(&sfs, 0, sizeof(sfs));
 651
 652         rc = sb->s_op->statfs(sb, &sfs);
 653
 654         if (!rc && sfs.f_bfree < sfs.f_ffree) {
 655                 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
 656                 sfs.f_ffree = sfs.f_bfree;
 657         }
 658
 659         statfs_pack(osfs, &sfs);
 660         return rc;
 661 }
 662
 663 static int fsfilt_ext3_sync(struct super_block *sb)
 664 {
 665         return ext3_force_commit(sb);
 666 }
 667
 668 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
 669 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 670 #define ext3_up_truncate_sem(inode)  up_write(&EXT3_I(inode)->truncate_sem);
 671 #define ext3_down_truncate_sem(inode)  down_write(&EXT3_I(inode)->truncate_sem);
 672 #else
 673 #define ext3_up_truncate_sem(inode)  up(&EXT3_I(inode)->truncate_sem);
 674 #define ext3_down_truncate_sem(inode)  down(&EXT3_I(inode)->truncate_sem);
 675 #endif
 676
 677 #include <linux/lustre_version.h>
 678 #if EXT3_EXT_MAGIC == 0xf301
 679 #define ee_start e_start
 680 #define ee_block e_block
 681 #define ee_len   e_num
 682 #endif
 683 #ifndef EXT3_BB_MAX_BLOCKS
 684 #define ext3_mb_new_blocks(handle, inode, goal, count, aflags, err) \
 685         ext3_new_blocks(handle, inode, count, goal, err)
 686 #endif
 687
 688 struct bpointers {
 689         unsigned long *blocks;
 690         int *created;
 691         unsigned long start;
 692         int num;
 693         int init_num;
 694         int create;
 695 };
 696 static int ext3_ext_find_goal(struct inode *inode, struct ext3_ext_path *path,
 697                                 unsigned long block, int *aflags)
 698 {
 699         struct ext3_inode_info *ei = EXT3_I(inode);
 700         unsigned long bg_start;
 701         unsigned long colour;
 702         int depth;
 703
 704         if (path) {
 705                 struct ext3_extent *ex;
 706                 depth = path->p_depth;
 707
 708                 /* try to predict block placement */
 709                 if ((ex = path[depth].p_ext)) {
 710                         if (ex->ee_block + ex->ee_len == block)
 711                                 *aflags |= 1;
 712                         return ex->ee_start + (block - ex->ee_block);
 713                 }
 714
 715                 /* it looks index is empty
 716                  * try to find starting from index itself */
 717                 if (path[depth].p_bh)
 718                         return path[depth].p_bh->b_blocknr;
 719         }
 720
 721         /* OK. use inode's group */
 722         bg_start = (ei->i_block_group * EXT3_BLOCKS_PER_GROUP(inode->i_sb)) +
 723                 le32_to_cpu(EXT3_SB(inode->i_sb)->s_es->s_first_data_block);
 724         colour = (current->pid % 16) *
 725                         (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 726         return bg_start + colour + block;
 727 }
 728
 729 static int ext3_ext_new_extent_cb(struct ext3_extents_tree *tree,
 730                                   struct ext3_ext_path *path,
 731                                   struct ext3_extent *newex, int exist)
 732 {
 733         struct inode *inode = tree->inode;
 734         struct bpointers *bp = tree->private;
 735         int count, err, goal;
 736         unsigned long pblock;
 737         unsigned long tgen;
 738         loff_t new_i_size;
 739         handle_t *handle;
 740         int i, aflags = 0;
 741
 742         i = EXT_DEPTH(tree);
 743         EXT_ASSERT(i == path->p_depth);
 744         EXT_ASSERT(path[i].p_hdr);
 745
 746         if (exist) {
 747                 err = EXT_CONTINUE;
 748                 goto map;
 749         }
 750
 751         if (bp->create == 0) {
 752                 i = 0;
 753                 if (newex->ee_block < bp->start)
 754                         i = bp->start - newex->ee_block;
 755                 if (i >= newex->ee_len)
 756                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 757                                         i, newex->ee_len);
 758                 for (; i < newex->ee_len && bp->num; i++) {
 759                         *(bp->created) = 0;
 760                         bp->created++;
 761                         *(bp->blocks) = 0;
 762                         bp->blocks++;
 763                         bp->num--;
 764                         bp->start++;
 765                 }
 766
 767                 return EXT_CONTINUE;
 768         }
 769         tgen = EXT_GENERATION(tree);
 770         count = ext3_ext_calc_credits_for_insert(tree, path);
 771         ext3_up_truncate_sem(inode);
 772
 773         lock_kernel();
 774         handle = journal_start(EXT3_JOURNAL(inode), count + EXT3_ALLOC_NEEDED + 1);
 775         unlock_kernel();
 776         if (IS_ERR(handle)) {
 777                 ext3_down_truncate_sem(inode);
 778                 return PTR_ERR(handle);
 779         }
 780
 781         if (tgen != EXT_GENERATION(tree)) {
 782                 /* the tree has changed. so path can be invalid at moment */
 783                 lock_kernel();
 784                 journal_stop(handle);
 785                 unlock_kernel();
 786                 ext3_down_truncate_sem(inode);
 787                 return EXT_REPEAT;
 788         }
 789
 790         ext3_down_truncate_sem(inode);
 791         count = newex->ee_len;
 792         goal = ext3_ext_find_goal(inode, path, newex->ee_block, &aflags);
 793         aflags |= 2; /* block have been already reserved */
 794         pblock = ext3_mb_new_blocks(handle, inode, goal, &count, aflags, &err);
 795         if (!pblock)
 796                 goto out;
 797         EXT_ASSERT(count <= newex->ee_len);
 798
 799         /* insert new extent */
 800         newex->ee_start = pblock;
 801         newex->ee_len = count;
 802         err = ext3_ext_insert_extent(handle, tree, path, newex);
 803         if (err)
 804                 goto out;
 805
 806         /* correct on-disk inode size */
 807         if (newex->ee_len > 0) {
 808                 new_i_size = (loff_t) newex->ee_block + newex->ee_len;
 809                 new_i_size = new_i_size << inode->i_blkbits;
 810                 if (new_i_size > EXT3_I(inode)->i_disksize) {
 811                         EXT3_I(inode)->i_disksize = new_i_size;
 812                         err = ext3_mark_inode_dirty(handle, inode);
 813                 }
 814         }
 815 out:
 816         lock_kernel();
 817         journal_stop(handle);
 818         unlock_kernel();
 819 map:
 820         if (err >= 0) {
 821                 /* map blocks */
 822                 if (bp->num == 0) {
 823                         CERROR("hmm. why do we find this extent?\n");
 824                         CERROR("initial space: %lu:%u\n",
 825                                 bp->start, bp->init_num);
 826                         CERROR("current extent: %u/%u/%u %d\n",
 827                                 newex->ee_block, newex->ee_len,
 828                                 newex->ee_start, exist);
 829                 }
 830                 i = 0;
 831                 if (newex->ee_block < bp->start)
 832                         i = bp->start - newex->ee_block;
 833                 if (i >= newex->ee_len)
 834                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 835                                         i, newex->ee_len);
 836                 for (; i < newex->ee_len && bp->num; i++) {
 837                         *(bp->created) = (exist == 0 ? 1 : 0);
 838                         bp->created++;
 839                         *(bp->blocks) = newex->ee_start + i;
 840                         bp->blocks++;
 841                         bp->num--;
 842                         bp->start++;
 843                 }
 844         }
 845         return err;
 846 }
 847
 848 int fsfilt_map_nblocks(struct inode *inode, unsigned long block,
 849                        unsigned long num, unsigned long *blocks,
 850                        int *created, int create)
 851 {
 852         struct ext3_extents_tree tree;
 853         struct bpointers bp;
 854         int err;
 855
 856         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
 857                 block, block + num, (unsigned) inode->i_ino);
 858
 859         ext3_init_tree_desc(&tree, inode);
 860         tree.private = &bp;
 861         bp.blocks = blocks;
 862         bp.created = created;
 863         bp.start = block;
 864         bp.init_num = bp.num = num;
 865         bp.create = create;
 866
 867         ext3_down_truncate_sem(inode);
 868         err = ext3_ext_walk_space(&tree, block, num, ext3_ext_new_extent_cb);
 869         ext3_ext_invalidate_cache(&tree);
 870         ext3_up_truncate_sem(inode);
 871
 872         return err;
 873 }
 874
 875 int fsfilt_ext3_map_ext_inode_pages(struct inode *inode, struct page **page,
 876                                     int pages, unsigned long *blocks,
 877                                     int *created, int create)
 878 {
 879         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 880         int rc = 0, i = 0;
 881         struct page *fp = NULL;
 882         int clen = 0;
 883
 884         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 885                 inode->i_ino, pages, (*page)->index);
 886
 887         /* pages are sorted already. so, we just have to find
 888          * contig. space and process them properly */
 889         while (i < pages) {
 890                 if (fp == NULL) {
 891                         /* start new extent */
 892                         fp = *page++;
 893                         clen = 1;
 894                         i++;
 895                         continue;
 896                 } else if (fp->index + clen == (*page)->index) {
 897                         /* continue the extent */
 898                         page++;
 899                         clen++;
 900                         i++;
 901                         continue;
 902                 }
 903
 904                 /* process found extent */
 905                 rc = fsfilt_map_nblocks(inode, fp->index * blocks_per_page,
 906                                         clen * blocks_per_page, blocks,
 907                                         created, create);
 908                 if (rc)
 909                         GOTO(cleanup, rc);
 910
 911                 /* look for next extent */
 912                 fp = NULL;
 913                 blocks += blocks_per_page * clen;
 914                 created += blocks_per_page * clen;
 915         }
 916
 917         if (fp)
 918                 rc = fsfilt_map_nblocks(inode, fp->index * blocks_per_page,
 919                                         clen * blocks_per_page, blocks,
 920                                         created, create);
 921 cleanup:
 922         return rc;
 923 }
 924 #endif
 925
 926 extern int ext3_map_inode_page(struct inode *inode, struct page *page,
 927                                unsigned long *blocks, int *created, int create);
 928 int fsfilt_ext3_map_bm_inode_pages(struct inode *inode, struct page **page,
 929                                    int pages, unsigned long *blocks,
 930                                    int *created, int create)
 931 {
 932         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 933         unsigned long *b;
 934         int rc = 0, i, *cr;
 935
 936         for (i = 0, cr = created, b = blocks; i < pages; i++, page++) {
 937                 rc = ext3_map_inode_page(inode, *page, b, cr, create);
 938                 if (rc) {
 939                         CERROR("ino %lu, blk %lu cr %u create %d: rc %d\n",
 940                                inode->i_ino, *b, *cr, create, rc);
 941                         break;
 942                 }
 943
 944                 b += blocks_per_page;
 945                 cr += blocks_per_page;
 946         }
 947         return rc;
 948 }
 949
 950 int fsfilt_ext3_map_inode_pages(struct inode *inode, struct page **page,
 951                                 int pages, unsigned long *blocks,
 952                                 int *created, int create,
 953                                 struct semaphore *optional_sem)
 954 {
 955         int rc;
 956 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
 957         if (EXT3_I(inode)->i_flags & EXT3_EXTENTS_FL) {
 958                 rc = fsfilt_ext3_map_ext_inode_pages(inode, page, pages,
 959                                                      blocks, created, create);
 960                 return rc;
 961         }
 962 #endif
 963         if (optional_sem != NULL)
 964                 down(optional_sem);
 965         rc = fsfilt_ext3_map_bm_inode_pages(inode, page, pages, blocks,
 966                                             created, create);
 967         if (optional_sem != NULL)
 968                 up(optional_sem);
 969
 970         return rc;
 971 }
 972
 973 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
 974                                int nblocks, loff_t newsize);
 975 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
 976                                       int nblocks, loff_t newsize)
 977 {
 978         return ext3_prep_san_write(inode, blocks, nblocks, newsize);
 979 }
 980
 981 static int fsfilt_ext3_read_record(struct file * file, void *buf,
 982                                    int size, loff_t *offs)
 983 {
 984         struct inode *inode = file->f_dentry->d_inode;
 985         unsigned long block;
 986         struct buffer_head *bh;
 987         int err, blocksize, csize, boffs;
 988
 989         /* prevent reading after eof */
 990         lock_kernel();
 991         if (inode->i_size < *offs + size) {
 992                 size = inode->i_size - *offs;
 993                 unlock_kernel();
 994                 if (size < 0) {
 995                         CERROR("size %llu is too short for read %u@%llu\n",
 996                                inode->i_size, size, *offs);
 997                         return -EIO;
 998                 } else if (size == 0) {
 999                         return 0;
1000                 }
1001         } else {
1002                 unlock_kernel();
1003         }
1004
1005         blocksize = 1 << inode->i_blkbits;
1006
1007         while (size > 0) {
1008                 block = *offs >> inode->i_blkbits;
1009                 boffs = *offs & (blocksize - 1);
1010                 csize = min(blocksize - boffs, size);
1011                 bh = ext3_bread(NULL, inode, block, 0, &err);
1012                 if (!bh) {
1013                         CERROR("can't read block: %d\n", err);
1014                         return err;
1015                 }
1016
1017                 memcpy(buf, bh->b_data + boffs, csize);
1018                 brelse(bh);
1019
1020                 *offs += csize;
1021                 buf += csize;
1022                 size -= csize;
1023         }
1024         return 0;
1025 }
1026
1027 static int fsfilt_ext3_write_record(struct file *file, void *buf, int bufsize,
1028                                     loff_t *offs, int force_sync)
1029 {
1030         struct buffer_head *bh = NULL;
1031         unsigned long block;
1032         struct inode *inode = file->f_dentry->d_inode;
1033         loff_t old_size = inode->i_size, offset = *offs;
1034         loff_t new_size = inode->i_size;
1035         journal_t *journal;
1036         handle_t *handle;
1037         int err, block_count = 0, blocksize, size, boffs;
1038
1039         /* Determine how many transaction credits are needed */
1040         blocksize = 1 << inode->i_blkbits;
1041         block_count = (*offs & (blocksize - 1)) + bufsize;
1042         block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
1043
1044         journal = EXT3_SB(inode->i_sb)->s_journal;
1045         lock_kernel();
1046         handle = journal_start(journal,
1047                                block_count * EXT3_DATA_TRANS_BLOCKS + 2);
1048         unlock_kernel();
1049         if (IS_ERR(handle)) {
1050                 CERROR("can't start transaction\n");
1051                 return PTR_ERR(handle);
1052         }
1053
1054         while (bufsize > 0) {
1055                 if (bh != NULL)
1056                         brelse(bh);
1057
1058                 block = offset >> inode->i_blkbits;
1059                 boffs = offset & (blocksize - 1);
1060                 size = min(blocksize - boffs, bufsize);
1061                 bh = ext3_bread(handle, inode, block, 1, &err);
1062                 if (!bh) {
1063                         CERROR("can't read/create block: %d\n", err);
1064                         goto out;
1065                 }
1066
1067                 err = ext3_journal_get_write_access(handle, bh);
1068                 if (err) {
1069                         CERROR("journal_get_write_access() returned error %d\n",
1070                                err);
1071                         goto out;
1072                 }
1073                 LASSERT(bh->b_data + boffs + size <= bh->b_data + bh->b_size);
1074                 memcpy(bh->b_data + boffs, buf, size);
1075                 err = ext3_journal_dirty_metadata(handle, bh);
1076                 if (err) {
1077                         CERROR("journal_dirty_metadata() returned error %d\n",
1078                                err);
1079                         goto out;
1080                 }
1081                 if (offset + size > new_size)
1082                         new_size = offset + size;
1083                 offset += size;
1084                 bufsize -= size;
1085                 buf += size;
1086         }
1087
1088         if (force_sync)
1089                 handle->h_sync = 1; /* recovery likes this */
1090 out:
1091         if (bh)
1092                 brelse(bh);
1093
1094         /* correct in-core and on-disk sizes */
1095         if (new_size > inode->i_size) {
1096                 lock_kernel();
1097                 if (new_size > inode->i_size)
1098                         inode->i_size = new_size;
1099                 if (inode->i_size > EXT3_I(inode)->i_disksize)
1100                         EXT3_I(inode)->i_disksize = inode->i_size;
1101                 if (inode->i_size > old_size)
1102                         mark_inode_dirty(inode);
1103                 unlock_kernel();
1104         }
1105
1106         lock_kernel();
1107         journal_stop(handle);
1108         unlock_kernel();
1109
1110         if (err == 0)
1111                 *offs = offset;
1112         return err;
1113 }
1114
1115 static int fsfilt_ext3_setup(struct obd_device *obd, struct super_block *sb)
1116 {
1117 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1118         struct mds_obd *mds = &obd->u.mds;
1119 #endif
1120 #if 0
1121         EXT3_SB(sb)->dx_lock = fsfilt_ext3_dx_lock;
1122         EXT3_SB(sb)->dx_unlock = fsfilt_ext3_dx_unlock;
1123 #endif
1124 #ifdef S_PDIROPS
1125         CWARN("Enabling PDIROPS\n");
1126         set_opt(EXT3_SB(sb)->s_mount_opt, PDIROPS);
1127         sb->s_flags |= S_PDIROPS;
1128 #endif
1129         /* setup mdsnum in underlying fs */
1130 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1131         if (mds->mds_lmv_obd) {
1132                 struct ext3_sb_info *sbi = EXT3_SB(sb);
1133                 struct ext3_super_block *es = sbi->s_es;
1134                 handle_t *handle;
1135                 int err;
1136
1137                 if (!EXT3_HAS_INCOMPAT_FEATURE(sb, EXT3_FEATURE_INCOMPAT_MDSNUM)) {
1138                         CWARN("%s: set mdsnum %d in ext3fs\n",
1139                                         obd->obd_name, mds->mds_num);
1140                         lock_kernel();
1141                         handle = journal_start(sbi->s_journal, 1);
1142                         unlock_kernel();
1143                         LASSERT(!IS_ERR(handle));
1144                         err = ext3_journal_get_write_access(handle, sbi->s_sbh);
1145                         LASSERT(err == 0);
1146                         EXT3_SET_INCOMPAT_FEATURE(sb,
1147                                                 EXT3_FEATURE_INCOMPAT_MDSNUM);
1148                         es->s_mdsnum = mds->mds_num;
1149                         err = ext3_journal_dirty_metadata(handle, sbi->s_sbh);
1150                         LASSERT(err == 0);
1151                         lock_kernel();
1152                         journal_stop(handle);
1153                         unlock_kernel();
1154                 } else {
1155                         CWARN("%s: mdsnum initialized to %u in ext3fs\n",
1156                                 obd->obd_name, es->s_mdsnum);
1157                 }
1158                 sbi->s_mdsnum = es->s_mdsnum;
1159         }
1160 #endif
1161         return 0;
1162 }
1163
1164 extern int ext3_add_dir_entry(struct dentry *dentry);
1165 extern int ext3_del_dir_entry(struct dentry *dentry);
1166
1167 static int fsfilt_ext3_add_dir_entry(struct obd_device *obd,
1168                                  struct dentry *parent,
1169                                  char *name, int namelen,
1170                                  unsigned long ino,
1171                                  unsigned long generation,
1172                                  unsigned mds)
1173 {
1174 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1175         struct dentry *dentry;
1176         int err;
1177         LASSERT(ino != 0);
1178         LASSERT(namelen != 0);
1179         dentry = ll_lookup_one_len(name, parent, namelen);
1180         if (IS_ERR(dentry)) {
1181                 CERROR("can't lookup %*s in %lu/%lu: %d\n", dentry->d_name.len,
1182                        dentry->d_name.name, dentry->d_inode->i_ino,
1183                        (unsigned long) dentry->d_inode->i_generation,
1184                        (int) PTR_ERR(dentry));
1185                 RETURN(PTR_ERR(dentry));
1186         }
1187         if (dentry->d_inode != NULL) {
1188                 CERROR("dentry %*s(0x%p) found\n", dentry->d_name.len,
1189                        dentry->d_name.name, dentry);
1190                 l_dput(dentry);
1191                 RETURN(-EEXIST);
1192         }
1193
1194         /* mds_reint_rename() may use this method to add dir entry
1195          * that points onto local inode. and we don't want to find
1196          * it cross-ref by subsequent lookups */
1197         d_drop(dentry);
1198
1199         dentry->d_flags |= DCACHE_CROSS_REF;
1200         dentry->d_inum = ino;
1201         dentry->d_mdsnum = mds;
1202         dentry->d_generation = generation;
1203         lock_kernel();
1204         err = ext3_add_dir_entry(dentry);
1205         unlock_kernel();
1206
1207         l_dput(dentry);
1208
1209         RETURN(err);
1210 #else
1211 #error "rebuild kernel and lustre with ext3-mds-num patch!"
1212         LASSERT(0);
1213 #endif
1214 }
1215
1216 static int fsfilt_ext3_del_dir_entry(struct obd_device *obd,
1217                                  struct dentry *dentry)
1218 {
1219 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1220         int err;
1221         lock_kernel();
1222         err = ext3_del_dir_entry(dentry);
1223         unlock_kernel();
1224         if (err == 0)
1225                 d_drop(dentry);
1226         return err;
1227 #else
1228 #error "rebuild kernel and lustre with ext3-mds-num patch!"
1229         LASSERT(0);
1230 #endif
1231 }
1232
1233 static int fsfilt_ext3_set_xattr(struct inode * inode, void *handle, char *name,
1234                                  void *buffer, int buffer_size)
1235 {
1236         int rc = 0;
1237
1238         lock_kernel();
1239
1240         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
1241                                    name, buffer, buffer_size, 0);
1242         unlock_kernel();
1243         if (rc)
1244                 CERROR("set xattr %s from inode %lu: rc %d\n",
1245                        name,  inode->i_ino, rc);
1246         return rc;
1247 }
1248
1249 static int fsfilt_ext3_get_xattr(struct inode *inode, char *name,
1250                                  void *buffer, int buffer_size)
1251 {
1252         int rc = 0;
1253
1254         lock_kernel();
1255
1256         rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
1257                             name, buffer, buffer_size);
1258         unlock_kernel();
1259
1260         if (buffer == NULL)
1261                 return (rc == -ENODATA) ? 0 : rc;
1262         if (rc < 0) {
1263                 CDEBUG(D_INFO, "error getting EA %s from inode %lu: rc %d\n",
1264                        name,  inode->i_ino, rc);
1265                 memset(buffer, 0, buffer_size);
1266                 return (rc == -ENODATA) ? 0 : rc;
1267         }
1268
1269         return rc;
1270 }
1271
1272 /* If fso is NULL, op is FSFILT operation, otherwise op is number of fso
1273    objects. Logs is number of logfiles to update */
1274 static int fsfilt_ext3_get_op_len(int op, struct fsfilt_objinfo *fso, int logs)
1275 {
1276         if ( !fso ) {
1277                 switch(op) {
1278                 case FSFILT_OP_CREATE:
1279                                  /* directory leaf, index & indirect & EA*/
1280                         return 4 + 3 * logs;
1281                 case FSFILT_OP_UNLINK:
1282                         return 3 * logs;
1283                 }
1284         } else {
1285                 int i;
1286                 int needed = 0;
1287                 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
1288                 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
1289                 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
1290                 for (i = 0; i < op; i++, fso++) {
1291                         int nblocks = fso->fso_bufcnt * blockpp;
1292                         int ndindirect = min(nblocks, addrpp + 1);
1293                         int nindir = nblocks + ndindirect + 1;
1294
1295                         needed += nindir;
1296                 }
1297                 return needed + 3 * logs;
1298         }
1299
1300         return 0;
1301 }
1302
1303
1304 #define EXTENTS_EA "write_extents"
1305 #define EXTENTS_EA_SIZE 64
1306
1307 int ext3_ext_in_ea_alloc_space(struct inode *, int, const char *, unsigned long, unsigned long);
1308 int ext3_ext_in_ea_remove_space(struct inode *, int, const char *, unsigned long, unsigned long);
1309 int ext3_ext_in_ea_get_extents(struct inode *, int, const char *, char **, int *);
1310 int ext3_ext_in_ea_get_extents_num(struct inode *, int, const char *, int *);
1311
1312 static int fsfilt_ext3_insert_extents_ea(struct inode *inode,
1313                                       unsigned long from,
1314                                       unsigned long num)
1315 {
1316         int rc = 0;
1317
1318         rc = ext3_ext_in_ea_alloc_space(inode, EXT3_XATTR_INDEX_TRUSTED,
1319                                         EXTENTS_EA, from, num);
1320         return rc;
1321 }
1322
1323 static int fsfilt_ext3_remove_extents_ea(struct inode *inode,
1324                                          unsigned long from,
1325                                          unsigned long num)
1326 {
1327         int rc = 0;
1328
1329         rc = ext3_ext_in_ea_remove_space(inode, EXT3_XATTR_INDEX_TRUSTED,
1330                                          EXTENTS_EA, from, num);
1331         return rc;
1332 }
1333
1334 extern int ext3_init_tree_in_ea(struct inode *inode, int name_index,
1335                                 const char *eaname, int size);
1336
1337 static int fsfilt_ext3_init_extents_ea(struct inode *inode)
1338 {
1339         int rc = 0;
1340
1341         rc = ext3_init_tree_in_ea(inode, EXT3_XATTR_INDEX_TRUSTED,
1342                                   EXTENTS_EA, 64);
1343         return rc;
1344 }
1345
1346 static int fsfilt_ext3_get_inode_write_extents(struct inode *inode,
1347                                          char **pbuf, int *size)
1348 {
1349         int rc = 0;
1350
1351         rc = ext3_ext_in_ea_get_extents(inode, EXT3_XATTR_INDEX_TRUSTED,
1352                                         EXTENTS_EA,  pbuf, size);
1353         return rc;
1354 }
1355
1356 static int fsfilt_ext3_get_write_extents_num(struct inode *inode, int *size)
1357 {
1358         int rc = 0;
1359
1360         rc = ext3_ext_in_ea_get_extents_num(inode, EXT3_XATTR_INDEX_TRUSTED,
1361                                             EXTENTS_EA, size);
1362         return rc;
1363 }
1364
1365 static struct fsfilt_operations fsfilt_ext3_ops = {
1366         .fs_type                = "ext3",
1367         .fs_owner               = THIS_MODULE,
1368         .fs_start               = fsfilt_ext3_start,
1369         .fs_brw_start           = fsfilt_ext3_brw_start,
1370         .fs_commit              = fsfilt_ext3_commit,
1371         .fs_commit_async        = fsfilt_ext3_commit_async,
1372         .fs_commit_wait         = fsfilt_ext3_commit_wait,
1373         .fs_setattr             = fsfilt_ext3_setattr,
1374         .fs_iocontrol           = fsfilt_ext3_iocontrol,
1375         .fs_set_md              = fsfilt_ext3_set_md,
1376         .fs_get_md              = fsfilt_ext3_get_md,
1377         .fs_readpage            = fsfilt_ext3_readpage,
1378         .fs_add_journal_cb      = fsfilt_ext3_add_journal_cb,
1379         .fs_statfs              = fsfilt_ext3_statfs,
1380         .fs_sync                = fsfilt_ext3_sync,
1381         .fs_map_inode_pages     = fsfilt_ext3_map_inode_pages,
1382         .fs_prep_san_write      = fsfilt_ext3_prep_san_write,
1383         .fs_write_record        = fsfilt_ext3_write_record,
1384         .fs_read_record         = fsfilt_ext3_read_record,
1385         .fs_setup               = fsfilt_ext3_setup,
1386         .fs_getpage             = fsfilt_ext3_getpage,
1387         .fs_send_bio            = fsfilt_ext3_send_bio,
1388         .fs_set_xattr           = fsfilt_ext3_set_xattr,
1389         .fs_get_xattr           = fsfilt_ext3_get_xattr,
1390         .fs_get_op_len          = fsfilt_ext3_get_op_len,
1391         .fs_add_dir_entry       = fsfilt_ext3_add_dir_entry,
1392         .fs_del_dir_entry       = fsfilt_ext3_del_dir_entry,
1393         .fs_init_extents_ea         = fsfilt_ext3_init_extents_ea,
1394         .fs_insert_extents_ea       = fsfilt_ext3_insert_extents_ea,
1395         .fs_remove_extents_ea       = fsfilt_ext3_remove_extents_ea,
1396         .fs_get_inode_write_extents = fsfilt_ext3_get_inode_write_extents,
1397         .fs_get_write_extents_num   = fsfilt_ext3_get_write_extents_num,
1398 };
1399
1400 static int __init fsfilt_ext3_init(void)
1401 {
1402         int rc;
1403
1404         fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
1405                                       sizeof(struct fsfilt_cb_data), 0,
1406                                       0, NULL, NULL);
1407         if (!fcb_cache) {
1408                 CERROR("error allocating fsfilt journal callback cache\n");
1409                 GOTO(out, rc = -ENOMEM);
1410         }
1411
1412         rc = fsfilt_register_ops(&fsfilt_ext3_ops);
1413
1414         if (rc)
1415                 kmem_cache_destroy(fcb_cache);
1416 out:
1417         return rc;
1418 }
1419
1420 static void __exit fsfilt_ext3_exit(void)
1421 {
1422         fsfilt_unregister_ops(&fsfilt_ext3_ops);
1423         LASSERTF(kmem_cache_destroy(fcb_cache) == 0,
1424                  "can't free fsfilt callback cache: count %d\n",
1425                  atomic_read(&fcb_cache_count));
1426 }
1427
1428 module_init(fsfilt_ext3_init);
1429 module_exit(fsfilt_ext3_exit);
1430
1431 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1432 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
1433 MODULE_LICENSE("GPL");