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