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