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/fs.h>
  29 #include <linux/jbd.h>
  30 #include <linux/slab.h>
  31 #include <linux/pagemap.h>
  32 #include <linux/quotaops.h>
  33 #include <linux/ext3_fs.h>
  34 #include <linux/ext3_jbd.h>
  35 #include <linux/version.h>
  36 /* XXX ugh */
  37 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
  38  #include <linux/ext3_xattr.h>
  39 #else
  40  #include <linux/../../fs/ext3/xattr.h>
  41 #endif
  42 #include <linux/kp30.h>
  43 #include <linux/lustre_fsfilt.h>
  44 #include <linux/obd.h>
  45 #include <linux/obd_class.h>
  46 #include <linux/module.h>
  47
  48 static kmem_cache_t *fcb_cache;
  49 static atomic_t fcb_cache_count = ATOMIC_INIT(0);
  50
  51 struct fsfilt_cb_data {
  52         struct journal_callback cb_jcb; /* jbd private data - MUST BE FIRST */
  53         fsfilt_cb_t cb_func;            /* MDS/OBD completion function */
  54         struct obd_device *cb_obd;      /* MDS/OBD completion device */
  55         __u64 cb_last_rcvd;             /* MDS/OST last committed operation */
  56         void *cb_data;                  /* MDS/OST completion function data */
  57 };
  58
  59 #ifndef EXT3_XATTR_INDEX_TRUSTED        /* temporary until we hit l28 kernel */
  60 #define EXT3_XATTR_INDEX_TRUSTED        4
  61 #endif
  62 #define XATTR_LUSTRE_MDS_LOV_EA         "lov"
  63
  64 #define EXT3_XATTR_INDEX_LUSTRE         5                         /* old */
  65 #define XATTR_LUSTRE_MDS_OBJID          "system.lustre_mds_objid" /* old */
  66
  67 /*
  68  * We don't currently need any additional blocks for rmdir and
  69  * unlink transactions because we are storing the OST oa_id inside
  70  * the inode (which we will be changing anyways as part of this
  71  * transaction).
  72  */
  73 static void *fsfilt_ext3_start(struct inode *inode, int op, void *desc_private)
  74 {
  75         /* For updates to the last recieved file */
  76         int nblocks = EXT3_DATA_TRANS_BLOCKS;
  77         int blocksize, block_count = 0;
  78         void *handle;
  79
  80         if (current->journal_info) {
  81                 CDEBUG(D_INODE, "increasing refcount on %p\n", current->journal_info);
  82                 goto journal_start;
  83         }
  84
  85         switch(op) {
  86         case FSFILT_OP_CREATE_LOG:
  87                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
  88                 op = FSFILT_OP_CREATE;
  89                 break;
  90         case FSFILT_OP_UNLINK_LOG:
  91                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
  92                 op = FSFILT_OP_UNLINK;
  93                 break;
  94         }
  95
  96         switch(op) {
  97         case FSFILT_OP_RMDIR:
  98         case FSFILT_OP_UNLINK:
  99                 nblocks += EXT3_DELETE_TRANS_BLOCKS;
 100                 break;
 101         case FSFILT_OP_RENAME:
 102                 /* modify additional directory */
 103                 nblocks += EXT3_DATA_TRANS_BLOCKS;
 104                 /* no break */
 105         case FSFILT_OP_SYMLINK:
 106                 /* additional block + block bitmap + GDT for long symlink */
 107                 nblocks += 3;
 108                 /* no break */
 109         case FSFILT_OP_CREATE:
 110         case FSFILT_OP_MKDIR:
 111         case FSFILT_OP_MKNOD:
 112                 /* modify one inode + block bitmap + GDT */
 113                 nblocks += 3;
 114                 /* no break */
 115         case FSFILT_OP_LINK:
 116                 /* modify parent directory */
 117                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
 118                 break;
 119         case FSFILT_OP_SETATTR:
 120                 /* Setattr on inode */
 121                 nblocks += 1;
 122                 break;
 123         case FSFILT_OP_CANCEL_UNLINK_LOG:
 124                 blocksize = 1 << inode->i_blkbits;
 125                 block_count = (blocksize - 1) + LLOG_CHUNK_SIZE;
 126                 block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
 127                 block_count = block_count * EXT3_DATA_TRANS_BLOCKS + 2;
 128                 nblocks = 2 * 2 * block_count;
 129                 break;
 130         default: CERROR("unknown transaction start op %d\n", op);
 131                  LBUG();
 132         }
 133
 134         LASSERT(current->journal_info == desc_private);
 135
 136  journal_start:
 137         lock_kernel();
 138         handle = journal_start(EXT3_JOURNAL(inode), nblocks);
 139         unlock_kernel();
 140
 141         if (!IS_ERR(handle))
 142                 LASSERT(current->journal_info == handle);
 143         return handle;
 144 }
 145
 146 /*
 147  * Calculate the number of buffer credits needed to write multiple pages in
 148  * a single ext3 transaction.  No, this shouldn't be here, but as yet ext3
 149  * doesn't have a nice API for calculating this sort of thing in advance.
 150  *
 151  * See comment above ext3_writepage_trans_blocks for details.  We assume
 152  * no data journaling is being done, but it does allow for all of the pages
 153  * being non-contiguous.  If we are guaranteed contiguous pages we could
 154  * reduce the number of (d)indirect blocks a lot.
 155  *
 156  * With N blocks per page and P pages, for each inode we have at most:
 157  * N*P indirect
 158  * min(N*P, blocksize/4 + 1) dindirect blocks
 159  * niocount tindirect
 160  *
 161  * For the entire filesystem, we have at most:
 162  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 163  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 164  * objcount inode blocks
 165  * 1 superblock
 166  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 167  *
 168  * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
 169  */
 170 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso,
 171                                       int niocount, struct niobuf_local *nb)
 172 {
 173         struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
 174         __u64 next_indir;
 175         const int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 176         int nbitmaps = 0, ngdblocks;
 177         int needed = objcount + 1; /* inodes + superblock */
 178         int i, j;
 179
 180         for (i = 0, j = 0; i < objcount; i++, fso++) {
 181                 /* two or more dindirect blocks in case we cross boundary */
 182                 int ndind = (long)((nb[j + fso->fso_bufcnt - 1].offset -
 183                                     nb[j].offset) >>
 184                                    sb->s_blocksize_bits) /
 185                         (EXT3_ADDR_PER_BLOCK(sb) * EXT3_ADDR_PER_BLOCK(sb));
 186                 nbitmaps += min(fso->fso_bufcnt, ndind > 0 ? ndind : 2);
 187
 188                 /* leaf, indirect, tindirect blocks for first block */
 189                 nbitmaps += blockpp + 2;
 190
 191                 j += fso->fso_bufcnt;
 192         }
 193
 194         next_indir = nb[0].offset +
 195                 (EXT3_ADDR_PER_BLOCK(sb) << sb->s_blocksize_bits);
 196         for (i = 1; i < niocount; i++) {
 197                 if (nb[i].offset >= next_indir) {
 198                         nbitmaps++;     /* additional indirect */
 199                         next_indir = nb[i].offset +
 200                                 (EXT3_ADDR_PER_BLOCK(sb)<<sb->s_blocksize_bits);
 201                 } else if (nb[i].offset != nb[i - 1].offset + sb->s_blocksize) {
 202                         nbitmaps++;     /* additional indirect */
 203                 }
 204                 nbitmaps += blockpp;    /* each leaf in different group? */
 205         }
 206
 207         ngdblocks = nbitmaps;
 208         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 209                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 210         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 211                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 212
 213         needed += nbitmaps + ngdblocks;
 214
 215         /* last_rcvd update */
 216         needed += EXT3_DATA_TRANS_BLOCKS;
 217
 218 #if defined(CONFIG_QUOTA) && !defined(__x86_64__) /* XXX */
 219         /* We assume that there will be 1 bit set in s_dquot.flags for each
 220          * quota file that is active.  This is at least true for now.
 221          */
 222         needed += hweight32(sb_any_quota_enabled(sb)) *
 223                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 224 #endif
 225
 226         return needed;
 227 }
 228
 229 /* We have to start a huge journal transaction here to hold all of the
 230  * metadata for the pages being written here.  This is necessitated by
 231  * the fact that we do lots of prepare_write operations before we do
 232  * any of the matching commit_write operations, so even if we split
 233  * up to use "smaller" transactions none of them could complete until
 234  * all of them were opened.  By having a single journal transaction,
 235  * we eliminate duplicate reservations for common blocks like the
 236  * superblock and group descriptors or bitmaps.
 237  *
 238  * We will start the transaction here, but each prepare_write will
 239  * add a refcount to the transaction, and each commit_write will
 240  * remove a refcount.  The transaction will be closed when all of
 241  * the pages have been written.
 242  */
 243 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
 244                                    int niocount, struct niobuf_local *nb,
 245                                    void *desc_private)
 246 {
 247         journal_t *journal;
 248         handle_t *handle;
 249         int needed;
 250         ENTRY;
 251
 252         LASSERT(current->journal_info == desc_private);
 253         journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
 254         needed = fsfilt_ext3_credits_needed(objcount, fso, niocount, nb);
 255
 256         /* The number of blocks we could _possibly_ dirty can very large.
 257          * We reduce our request if it is absurd (and we couldn't get that
 258          * many credits for a single handle anyways).
 259          *
 260          * At some point we have to limit the size of I/Os sent at one time,
 261          * increase the size of the journal, or we have to calculate the
 262          * actual journal requirements more carefully by checking all of
 263          * the blocks instead of being maximally pessimistic.  It remains to
 264          * be seen if this is a real problem or not.
 265          */
 266         if (needed > journal->j_max_transaction_buffers) {
 267                 CERROR("want too many journal credits (%d) using %d instead\n",
 268                        needed, journal->j_max_transaction_buffers);
 269                 needed = journal->j_max_transaction_buffers;
 270         }
 271
 272         lock_kernel();
 273         handle = journal_start(journal, needed);
 274         unlock_kernel();
 275         if (IS_ERR(handle)) {
 276                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 277                        PTR_ERR(handle));
 278         } else {
 279                 LASSERT(handle->h_buffer_credits >= needed);
 280                 LASSERT(current->journal_info == handle);
 281         }
 282
 283         RETURN(handle);
 284 }
 285
 286 static int fsfilt_ext3_commit(struct inode *inode, void *h, int force_sync)
 287 {
 288         int rc;
 289         handle_t *handle = h;
 290
 291         LASSERT(current->journal_info == handle);
 292         if (force_sync)
 293                 handle->h_sync = 1; /* recovery likes this */
 294
 295         lock_kernel();
 296         rc = journal_stop(handle);
 297         unlock_kernel();
 298
 299         // LASSERT(current->journal_info == NULL);
 300         return rc;
 301 }
 302
 303 static int fsfilt_ext3_commit_async(struct inode *inode, void *h,
 304                                         void **wait_handle)
 305 {
 306         transaction_t *transaction;
 307         unsigned long tid, rtid;
 308         handle_t *handle = h;
 309         journal_t *journal;
 310         int rc;
 311
 312         LASSERT(current->journal_info == handle);
 313
 314         lock_kernel();
 315         transaction = handle->h_transaction;
 316         journal = transaction->t_journal;
 317         tid = transaction->t_tid;
 318         /* we don't want to be blocked */
 319         handle->h_sync = 0;
 320         rc = journal_stop(handle);
 321         if (rc) {
 322                 CERROR("error while stopping transaction: %d\n", rc);
 323                 unlock_kernel();
 324                 return rc;
 325         }
 326 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 327         rtid = log_start_commit(journal, transaction);
 328         if (rtid != tid)
 329                 CERROR("strange race: %lu != %lu\n",
 330                        (unsigned long) tid, (unsigned long) rtid);
 331 #else
 332         log_start_commit(journal, transaction->t_tid);
 333 #endif
 334         unlock_kernel();
 335
 336         *wait_handle = (void *) tid;
 337         CDEBUG(D_INODE, "commit async: %lu\n", (unsigned long) tid);
 338         return 0;
 339 }
 340
 341 static int fsfilt_ext3_commit_wait(struct inode *inode, void *h)
 342 {
 343         tid_t tid = (tid_t)(long)h;
 344
 345         CDEBUG(D_INODE, "commit wait: %lu\n", (unsigned long) tid);
 346         if (is_journal_aborted(EXT3_JOURNAL(inode)))
 347                 return -EIO;
 348
 349         log_wait_commit(EXT3_JOURNAL(inode), tid);
 350
 351         return 0;
 352 }
 353
 354 static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
 355                                struct iattr *iattr, int do_trunc)
 356 {
 357         struct inode *inode = dentry->d_inode;
 358         int rc;
 359
 360         lock_kernel();
 361
 362         /* A _really_ horrible hack to avoid removing the data stored
 363          * in the block pointers; this is really the "small" stripe MD data.
 364          * We can avoid further hackery by virtue of the MDS file size being
 365          * zero all the time (which doesn't invoke block truncate at unlink
 366          * time), so we assert we never change the MDS file size from zero. */
 367         if (iattr->ia_valid & ATTR_SIZE && !do_trunc) {
 368                 /* ATTR_SIZE would invoke truncate: clear it */
 369                 iattr->ia_valid &= ~ATTR_SIZE;
 370                 EXT3_I(inode)->i_disksize = inode->i_size = iattr->ia_size;
 371
 372                 /* make sure _something_ gets set - so new inode
 373                  * goes to disk (probably won't work over XFS */
 374                 if (!(iattr->ia_valid & (ATTR_MODE | ATTR_MTIME | ATTR_CTIME))){
 375                         iattr->ia_valid |= ATTR_MODE;
 376                         iattr->ia_mode = inode->i_mode;
 377                 }
 378         }
 379
 380         /* Don't allow setattr to change file type */
 381         iattr->ia_mode = (inode->i_mode & S_IFMT)|(iattr->ia_mode & ~S_IFMT);
 382
 383         /* We set these flags on the client, but have already checked perms
 384          * so don't confuse inode_change_ok. */
 385         iattr->ia_valid &= ~(ATTR_MTIME_SET | ATTR_ATIME_SET);
 386
 387         if (inode->i_op->setattr) {
 388                 rc = inode->i_op->setattr(dentry, iattr);
 389         } else {
 390                 rc = inode_change_ok(inode, iattr);
 391                 if (!rc)
 392                         rc = inode_setattr(inode, iattr);
 393         }
 394
 395         unlock_kernel();
 396
 397         return rc;
 398 }
 399
 400 static int fsfilt_ext3_iocontrol(struct inode * inode, struct file *file,
 401                                  unsigned int cmd, unsigned long arg)
 402 {
 403         int rc = 0;
 404         ENTRY;
 405
 406         if (inode->i_fop->ioctl)
 407                 rc = inode->i_fop->ioctl(inode, file, cmd, arg);
 408         else
 409                 RETURN(-ENOTTY);
 410
 411         RETURN(rc);
 412 }
 413
 414 #undef INLINE_EA
 415 #undef OLD_EA
 416 static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
 417                               void *lmm, int lmm_size)
 418 {
 419         int rc, old_ea = 0;
 420
 421 #ifdef INLINE_EA  /* can go away before 1.0 - just for testing bug 2097 now */
 422         /* Nasty hack city - store stripe MD data in the block pointers if
 423          * it will fit, because putting it in an EA currently kills the MDS
 424          * performance.  We'll fix this with "fast EAs" in the future.
 425          */
 426         if (inode->i_blocks == 0 && lmm_size <= sizeof(EXT3_I(inode)->i_data) -
 427                                             sizeof(EXT3_I(inode)->i_data[0])) {
 428                 unsigned old_size = EXT3_I(inode)->i_data[0];
 429                 if (old_size != 0) {
 430                         LASSERT(old_size < sizeof(EXT3_I(inode)->i_data));
 431                         CERROR("setting EA on %lu/%u again... interesting\n",
 432                                inode->i_ino, inode->i_generation);
 433                 }
 434
 435                 EXT3_I(inode)->i_data[0] = cpu_to_le32(lmm_size);
 436                 memcpy(&EXT3_I(inode)->i_data[1], lmm, lmm_size);
 437                 mark_inode_dirty(inode);
 438                 return 0;
 439         }
 440 #endif
 441 #ifdef OLD_EA
 442         /* keep this when we get rid of OLD_EA (too noisy during conversion) */
 443         if (EXT3_I(inode)->i_file_acl /* || large inode EA flag */) {
 444                 CWARN("setting EA on %lu/%u again... interesting\n",
 445                        inode->i_ino, inode->i_generation);
 446                 old_ea = 1;
 447         }
 448
 449         lock_kernel();
 450         /* this can go away before 1.0.  For bug 2097 testing only. */
 451         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_LUSTRE,
 452                                    XATTR_LUSTRE_MDS_OBJID, lmm, lmm_size, 0);
 453 #else
 454         lock_kernel();
 455         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
 456                                    XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size, 0);
 457
 458         /* This tries to delete the old-format LOV EA, but only as long as we
 459          * have successfully saved the new-format LOV EA (we can always try
 460          * the conversion again the next time the file is accessed).  It is
 461          * possible (although unlikely) that the new-format LOV EA couldn't be
 462          * saved because it ran out of space but we would need a file striped
 463          * over least 123 OSTs before the two EAs filled a 4kB block.
 464          *
 465          * This can be removed when all filesystems have converted to the
 466          * new EA format, but otherwise adds little if any overhead.  If we
 467          * wanted backward compatibility for existing files, we could keep
 468          * the old EA around for a while but we'd have to clean it up later. */
 469         if (rc >= 0 && old_ea) {
 470                 int err = ext3_xattr_set_handle(handle, inode,
 471                                                 EXT3_XATTR_INDEX_LUSTRE,
 472                                                 XATTR_LUSTRE_MDS_OBJID,
 473                                                 NULL, 0, 0);
 474                 if (err)
 475                         CERROR("error deleting old LOV EA on %lu/%u: rc %d\n",
 476                                inode->i_ino, inode->i_generation, err);
 477         }
 478 #endif
 479         unlock_kernel();
 480
 481         if (rc)
 482                 CERROR("error adding MD data to inode %lu: rc = %d\n",
 483                        inode->i_ino, rc);
 484         return rc;
 485 }
 486
 487 /* Must be called with i_sem held */
 488 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm, int lmm_size)
 489 {
 490         int rc;
 491
 492         LASSERT(down_trylock(&inode->i_sem) != 0);
 493         lock_kernel();
 494         /* Keep support for reading "inline EAs" until we convert
 495          * users over to new format entirely.  See bug 841/2097. */
 496         if (inode->i_blocks == 0 && EXT3_I(inode)->i_data[0]) {
 497                 unsigned size = le32_to_cpu(EXT3_I(inode)->i_data[0]);
 498                 void *handle;
 499
 500                 LASSERT(size < sizeof(EXT3_I(inode)->i_data));
 501                 if (lmm) {
 502                         if (size > lmm_size) {
 503                                 CERROR("inline EA on %lu/%u bad size %u > %u\n",
 504                                        inode->i_ino, inode->i_generation,
 505                                        size, lmm_size);
 506                                 return -ERANGE;
 507                         }
 508                         memcpy(lmm, &EXT3_I(inode)->i_data[1], size);
 509                 }
 510
 511 #ifndef INLINE_EA
 512                 /* migrate LOV EA data to external block - keep same format */
 513                 CWARN("DEBUG: migrate inline EA for inode %lu/%u to block\n",
 514                       inode->i_ino, inode->i_generation);
 515
 516                 handle = journal_start(EXT3_JOURNAL(inode),
 517                                        EXT3_XATTR_TRANS_BLOCKS);
 518                 if (!IS_ERR(handle)) {
 519                         int err;
 520                         rc = fsfilt_ext3_set_md(inode, handle,
 521                                                 &EXT3_I(inode)->i_data[1],size);
 522                         if (rc == 0) {
 523                                 memset(EXT3_I(inode)->i_data, 0,
 524                                        sizeof(EXT3_I(inode)->i_data));
 525                                 mark_inode_dirty(inode);
 526                         }
 527                         err = journal_stop(handle);
 528                         if (err && rc == 0)
 529                                 rc = err;
 530                 } else {
 531                         rc = PTR_ERR(handle);
 532                 }
 533 #endif
 534                 unlock_kernel();
 535                 return size;
 536         }
 537
 538         rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
 539                             XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size);
 540         /* try old EA type if new one failed - MDS will convert it for us */
 541         if (rc == -ENODATA) {
 542                 CDEBUG(D_INFO,"failed new LOV EA %d/%s from inode %lu: rc %d\n",
 543                        EXT3_XATTR_INDEX_TRUSTED, XATTR_LUSTRE_MDS_LOV_EA,
 544                        inode->i_ino, rc);
 545
 546                 rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_LUSTRE,
 547                                     XATTR_LUSTRE_MDS_OBJID, lmm, lmm_size);
 548         }
 549         unlock_kernel();
 550
 551         /* This gives us the MD size */
 552         if (lmm == NULL)
 553                 return (rc == -ENODATA) ? 0 : rc;
 554
 555         if (rc < 0) {
 556                 CDEBUG(D_INFO, "error getting EA %d/%s from inode %lu: rc %d\n",
 557                        EXT3_XATTR_INDEX_LUSTRE, XATTR_LUSTRE_MDS_OBJID,
 558                        inode->i_ino, rc);
 559                 memset(lmm, 0, lmm_size);
 560                 return (rc == -ENODATA) ? 0 : rc;
 561         }
 562
 563         return rc;
 564 }
 565
 566 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
 567                                     loff_t *off)
 568 {
 569         struct inode *inode = file->f_dentry->d_inode;
 570         int rc = 0;
 571
 572         if (S_ISREG(inode->i_mode))
 573                 rc = file->f_op->read(file, buf, count, off);
 574         else {
 575                 const int blkbits = inode->i_sb->s_blocksize_bits;
 576                 const int blksize = inode->i_sb->s_blocksize;
 577
 578                 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
 579                        count, inode->i_ino, *off);
 580                 while (count > 0) {
 581                         struct buffer_head *bh;
 582
 583                         bh = NULL;
 584                         if (*off < inode->i_size) {
 585                                 int err = 0;
 586
 587                                 bh = ext3_bread(NULL, inode, *off >> blkbits,
 588                                                 0, &err);
 589
 590                                 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
 591
 592                                 if (bh) {
 593                                         memcpy(buf, bh->b_data, blksize);
 594                                         brelse(bh);
 595                                 } else if (err) {
 596                                         /* XXX in theory we should just fake
 597                                          * this buffer and continue like ext3,
 598                                          * especially if this is a partial read
 599                                          */
 600                                         CERROR("error read dir %lu+%llu: %d\n",
 601                                                inode->i_ino, *off, err);
 602                                         RETURN(err);
 603                                 }
 604                         }
 605                         if (!bh) {
 606                                 struct ext3_dir_entry_2 *fake = (void *)buf;
 607
 608                                 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
 609                                 memset(fake, 0, sizeof(*fake));
 610                                 fake->rec_len = cpu_to_le32(blksize);
 611                         }
 612                         count -= blksize;
 613                         buf += blksize;
 614                         *off += blksize;
 615                         rc += blksize;
 616                 }
 617         }
 618
 619         return rc;
 620 }
 621
 622 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
 623 {
 624         struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
 625
 626         fcb->cb_func(fcb->cb_obd, fcb->cb_last_rcvd, fcb->cb_data, error);
 627
 628         OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
 629         atomic_dec(&fcb_cache_count);
 630 }
 631
 632 static int fsfilt_ext3_add_journal_cb(struct obd_device *obd, __u64 last_rcvd,
 633                                       void *handle, fsfilt_cb_t cb_func,
 634                                       void *cb_data)
 635 {
 636         struct fsfilt_cb_data *fcb;
 637
 638         OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
 639         if (fcb == NULL)
 640                 RETURN(-ENOMEM);
 641
 642         atomic_inc(&fcb_cache_count);
 643         fcb->cb_func = cb_func;
 644         fcb->cb_obd = obd;
 645         fcb->cb_last_rcvd = last_rcvd;
 646         fcb->cb_data = cb_data;
 647
 648         CDEBUG(D_EXT2, "set callback for last_rcvd: "LPD64"\n", last_rcvd);
 649         lock_kernel();
 650         journal_callback_set(handle, fsfilt_ext3_cb_func,
 651                              (struct journal_callback *)fcb);
 652         unlock_kernel();
 653
 654         return 0;
 655 }
 656
 657 /*
 658  * We need to hack the return value for the free inode counts because
 659  * the current EA code requires one filesystem block per inode with EAs,
 660  * so it is possible to run out of blocks before we run out of inodes.
 661  *
 662  * This can be removed when the ext3 EA code is fixed.
 663  */
 664 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
 665 {
 666         struct kstatfs sfs;
 667         int rc = vfs_statfs(sb, &sfs);
 668
 669         if (!rc && sfs.f_bfree < sfs.f_ffree) {
 670                 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
 671                 sfs.f_ffree = sfs.f_bfree;
 672         }
 673
 674         statfs_pack(osfs, &sfs);
 675         return rc;
 676 }
 677
 678 static int fsfilt_ext3_sync(struct super_block *sb)
 679 {
 680         return ext3_force_commit(sb);
 681 }
 682
 683 extern int ext3_map_inode_page(struct inode *inode, struct page *page,
 684                                unsigned long *blocks, int *created, int create);
 685 int fsfilt_ext3_map_inode_page(struct inode *inode, struct page *page,
 686                                unsigned long *blocks, int *created, int create)
 687 {
 688         return ext3_map_inode_page(inode, page, blocks, created, create);
 689 }
 690
 691 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
 692                                int nblocks, loff_t newsize);
 693 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
 694                                       int nblocks, loff_t newsize)
 695 {
 696         return ext3_prep_san_write(inode, blocks, nblocks, newsize);
 697 }
 698
 699 static int fsfilt_ext3_read_record(struct file * file, void *buf,
 700                                    int size, loff_t *offs)
 701 {
 702         struct inode *inode = file->f_dentry->d_inode;
 703         unsigned long block;
 704         struct buffer_head *bh;
 705         int err, blocksize, csize, boffs;
 706
 707         /* prevent reading after eof */
 708         lock_kernel();
 709         if (inode->i_size < *offs + size) {
 710                 size = inode->i_size - *offs;
 711                 unlock_kernel();
 712                 if (size < 0) {
 713                         CERROR("size %llu is too short for read %u@%llu\n",
 714                                inode->i_size, size, *offs);
 715                         return -EIO;
 716                 } else if (size == 0) {
 717                         return 0;
 718                 }
 719         } else {
 720                 unlock_kernel();
 721         }
 722
 723         blocksize = 1 << inode->i_blkbits;
 724
 725         while (size > 0) {
 726                 block = *offs >> inode->i_blkbits;
 727                 boffs = *offs & (blocksize - 1);
 728                 csize = min(blocksize - boffs, size);
 729                 bh = ext3_bread(NULL, inode, block, 0, &err);
 730                 if (!bh) {
 731                         CERROR("can't read block: %d\n", err);
 732                         return err;
 733                 }
 734
 735                 memcpy(buf, bh->b_data + boffs, csize);
 736                 brelse(bh);
 737
 738                 *offs += csize;
 739                 buf += csize;
 740                 size -= csize;
 741         }
 742         return 0;
 743 }
 744
 745 static int fsfilt_ext3_write_record(struct file *file, void *buf, int bufsize,
 746                                     loff_t *offs, int force_sync)
 747 {
 748         struct buffer_head *bh = NULL;
 749         unsigned long block;
 750         struct inode *inode = file->f_dentry->d_inode;
 751         loff_t old_size = inode->i_size, offset = *offs;
 752         loff_t new_size = inode->i_size;
 753         journal_t *journal;
 754         handle_t *handle;
 755         int err, block_count = 0, blocksize, size, boffs;
 756
 757         /* Determine how many transaction credits are needed */
 758         blocksize = 1 << inode->i_blkbits;
 759         block_count = (*offs & (blocksize - 1)) + bufsize;
 760         block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
 761
 762         journal = EXT3_SB(inode->i_sb)->s_journal;
 763         lock_kernel();
 764         handle = journal_start(journal,
 765                                block_count * EXT3_DATA_TRANS_BLOCKS + 2);
 766         unlock_kernel();
 767         if (IS_ERR(handle)) {
 768                 CERROR("can't start transaction\n");
 769                 return PTR_ERR(handle);
 770         }
 771
 772         while (bufsize > 0) {
 773                 if (bh != NULL)
 774                         brelse(bh);
 775
 776                 block = offset >> inode->i_blkbits;
 777                 boffs = offset & (blocksize - 1);
 778                 size = min(blocksize - boffs, bufsize);
 779                 bh = ext3_bread(handle, inode, block, 1, &err);
 780                 if (!bh) {
 781                         CERROR("can't read/create block: %d\n", err);
 782                         goto out;
 783                 }
 784
 785                 err = ext3_journal_get_write_access(handle, bh);
 786                 if (err) {
 787                         CERROR("journal_get_write_access() returned error %d\n",
 788                                err);
 789                         goto out;
 790                 }
 791                 LASSERT(bh->b_data + boffs + size <= bh->b_data + bh->b_size);
 792                 memcpy(bh->b_data + boffs, buf, size);
 793                 err = ext3_journal_dirty_metadata(handle, bh);
 794                 if (err) {
 795                         CERROR("journal_dirty_metadata() returned error %d\n",
 796                                err);
 797                         goto out;
 798                 }
 799                 if (offset + size > new_size)
 800                         new_size = offset + size;
 801                 offset += size;
 802                 bufsize -= size;
 803                 buf += size;
 804         }
 805
 806         if (force_sync)
 807                 handle->h_sync = 1; /* recovery likes this */
 808 out:
 809         if (bh)
 810                 brelse(bh);
 811
 812         /* correct in-core and on-disk sizes */
 813         if (new_size > inode->i_size) {
 814                 lock_kernel();
 815                 if (new_size > inode->i_size)
 816                         inode->i_size = new_size;
 817                 if (inode->i_size > EXT3_I(inode)->i_disksize)
 818                         EXT3_I(inode)->i_disksize = inode->i_size;
 819                 if (inode->i_size > old_size)
 820                         mark_inode_dirty(inode);
 821                 unlock_kernel();
 822         }
 823
 824         lock_kernel();
 825         journal_stop(handle);
 826         unlock_kernel();
 827
 828         if (err == 0)
 829                 *offs = offset;
 830         return err;
 831 }
 832
 833 static int fsfilt_ext3_setup(struct super_block *sb)
 834 {
 835 #if 0
 836         EXT3_SB(sb)->dx_lock = fsfilt_ext3_dx_lock;
 837         EXT3_SB(sb)->dx_unlock = fsfilt_ext3_dx_unlock;
 838 #endif
 839 #ifdef S_PDIROPS
 840         CWARN("Enabling PDIROPS\n");
 841         set_opt(EXT3_SB(sb)->s_mount_opt, PDIROPS);
 842         sb->s_flags |= S_PDIROPS;
 843 #endif
 844         return 0;
 845 }
 846
 847 static struct fsfilt_operations fsfilt_ext3_ops = {
 848         fs_type:                "ext3",
 849         fs_owner:               THIS_MODULE,
 850         fs_start:               fsfilt_ext3_start,
 851         fs_brw_start:           fsfilt_ext3_brw_start,
 852         fs_commit:              fsfilt_ext3_commit,
 853         fs_commit_async:        fsfilt_ext3_commit_async,
 854         fs_commit_wait:         fsfilt_ext3_commit_wait,
 855         fs_setattr:             fsfilt_ext3_setattr,
 856         fs_iocontrol:           fsfilt_ext3_iocontrol,
 857         fs_set_md:              fsfilt_ext3_set_md,
 858         fs_get_md:              fsfilt_ext3_get_md,
 859         fs_readpage:            fsfilt_ext3_readpage,
 860         fs_add_journal_cb:      fsfilt_ext3_add_journal_cb,
 861         fs_statfs:              fsfilt_ext3_statfs,
 862         fs_sync:                fsfilt_ext3_sync,
 863         fs_map_inode_page:      fsfilt_ext3_map_inode_page,
 864         fs_prep_san_write:      fsfilt_ext3_prep_san_write,
 865         fs_write_record:        fsfilt_ext3_write_record,
 866         fs_read_record:         fsfilt_ext3_read_record,
 867         fs_setup:               fsfilt_ext3_setup,
 868 };
 869
 870 static int __init fsfilt_ext3_init(void)
 871 {
 872         int rc;
 873
 874         //rc = ext3_xattr_register();
 875         fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
 876                                       sizeof(struct fsfilt_cb_data), 0,
 877                                       0, NULL, NULL);
 878         if (!fcb_cache) {
 879                 CERROR("error allocating fsfilt journal callback cache\n");
 880                 GOTO(out, rc = -ENOMEM);
 881         }
 882
 883         rc = fsfilt_register_ops(&fsfilt_ext3_ops);
 884
 885         if (rc)
 886                 kmem_cache_destroy(fcb_cache);
 887 out:
 888         return rc;
 889 }
 890
 891 static void __exit fsfilt_ext3_exit(void)
 892 {
 893         int rc;
 894
 895         fsfilt_unregister_ops(&fsfilt_ext3_ops);
 896         rc = kmem_cache_destroy(fcb_cache);
 897
 898         if (rc || atomic_read(&fcb_cache_count)) {
 899                 CERROR("can't free fsfilt callback cache: count %d, rc = %d\n",
 900                        atomic_read(&fcb_cache_count), rc);
 901         }
 902
 903         //rc = ext3_xattr_unregister();
 904 }
 905
 906 module_init(fsfilt_ext3_init);
 907 module_exit(fsfilt_ext3_exit);
 908
 909 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
 910 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
 911 MODULE_LICENSE("GPL");