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                                int logs)
  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", current->journal_info);
  82                 goto journal_start;
  83         }
  84
  85         switch(op) {
  86         case FSFILT_OP_RMDIR:
  87         case FSFILT_OP_UNLINK:
  88                 nblocks += EXT3_DELETE_TRANS_BLOCKS;
  89                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
  90                             EXT3_DATA_TRANS_BLOCKS) * logs;
  91                 break;
  92         case FSFILT_OP_RENAME:
  93                 /* modify additional directory */
  94                 nblocks += EXT3_DATA_TRANS_BLOCKS;
  95                 /* no break */
  96         case FSFILT_OP_SYMLINK:
  97                 /* additional block + block bitmap + GDT for long symlink */
  98                 nblocks += 3;
  99                 /* no break */
 100         case FSFILT_OP_CREATE:
 101                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 102                             EXT3_DATA_TRANS_BLOCKS) * logs;
 103         case FSFILT_OP_MKDIR:
 104         case FSFILT_OP_MKNOD:
 105                 /* modify one inode + block bitmap + GDT */
 106                 nblocks += 3;
 107                 /* no break */
 108         case FSFILT_OP_LINK:
 109                 /* modify parent directory */
 110                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
 111                 break;
 112         case FSFILT_OP_SETATTR:
 113                 /* Setattr on inode */
 114                 nblocks += 1;
 115                 break;
 116         case FSFILT_OP_CANCEL_UNLINK:
 117                 nblocks = (LLOG_CHUNK_SIZE >> inode->i_blkbits) +
 118                         EXT3_DELETE_TRANS_BLOCKS * logs;
 119                 break;
 120         default: CERROR("unknown transaction start op %d\n", op);
 121                  LBUG();
 122         }
 123
 124         LASSERT(current->journal_info == desc_private);
 125
 126  journal_start:
 127         lock_kernel();
 128         handle = journal_start(EXT3_JOURNAL(inode), nblocks);
 129         unlock_kernel();
 130
 131         if (!IS_ERR(handle))
 132                 LASSERT(current->journal_info == handle);
 133         return handle;
 134 }
 135
 136 /*
 137  * Calculate the number of buffer credits needed to write multiple pages in
 138  * a single ext3 transaction.  No, this shouldn't be here, but as yet ext3
 139  * doesn't have a nice API for calculating this sort of thing in advance.
 140  *
 141  * See comment above ext3_writepage_trans_blocks for details.  We assume
 142  * no data journaling is being done, but it does allow for all of the pages
 143  * being non-contiguous.  If we are guaranteed contiguous pages we could
 144  * reduce the number of (d)indirect blocks a lot.
 145  *
 146  * With N blocks per page and P pages, for each inode we have at most:
 147  * N*P indirect
 148  * min(N*P, blocksize/4 + 1) dindirect blocks
 149  * niocount tindirect
 150  *
 151  * For the entire filesystem, we have at most:
 152  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 153  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 154  * objcount inode blocks
 155  * 1 superblock
 156  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 157  *
 158  * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
 159  */
 160 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso,
 161                                       int niocount, struct niobuf_local *nb)
 162 {
 163         struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
 164         __u64 next_indir;
 165         const int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 166         int nbitmaps = 0, ngdblocks;
 167         int needed = objcount + 1; /* inodes + superblock */
 168         int i, j;
 169
 170         for (i = 0, j = 0; i < objcount; i++, fso++) {
 171                 /* two or more dindirect blocks in case we cross boundary */
 172                 int ndind = (long)((nb[j + fso->fso_bufcnt - 1].offset -
 173                                     nb[j].offset) >>
 174                                    sb->s_blocksize_bits) /
 175                         (EXT3_ADDR_PER_BLOCK(sb) * EXT3_ADDR_PER_BLOCK(sb));
 176                 nbitmaps += min(fso->fso_bufcnt, ndind > 0 ? ndind : 2);
 177
 178                 /* leaf, indirect, tindirect blocks for first block */
 179                 nbitmaps += blockpp + 2;
 180
 181                 j += fso->fso_bufcnt;
 182         }
 183
 184         next_indir = nb[0].offset +
 185                 (EXT3_ADDR_PER_BLOCK(sb) << sb->s_blocksize_bits);
 186         for (i = 1; i < niocount; i++) {
 187                 if (nb[i].offset >= next_indir) {
 188                         nbitmaps++;     /* additional indirect */
 189                         next_indir = nb[i].offset +
 190                                 (EXT3_ADDR_PER_BLOCK(sb)<<sb->s_blocksize_bits);
 191                 } else if (nb[i].offset != nb[i - 1].offset + sb->s_blocksize) {
 192                         nbitmaps++;     /* additional indirect */
 193                 }
 194                 nbitmaps += blockpp;    /* each leaf in different group? */
 195         }
 196
 197         ngdblocks = nbitmaps;
 198         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 199                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 200         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 201                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 202
 203         needed += nbitmaps + ngdblocks;
 204
 205         /* last_rcvd update */
 206         needed += EXT3_DATA_TRANS_BLOCKS;
 207
 208 #if defined(CONFIG_QUOTA) && !defined(__x86_64__) /* XXX */
 209         /* We assume that there will be 1 bit set in s_dquot.flags for each
 210          * quota file that is active.  This is at least true for now.
 211          */
 212         needed += hweight32(sb_any_quota_enabled(sb)) *
 213                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 214 #endif
 215
 216         return needed;
 217 }
 218
 219 /* We have to start a huge journal transaction here to hold all of the
 220  * metadata for the pages being written here.  This is necessitated by
 221  * the fact that we do lots of prepare_write operations before we do
 222  * any of the matching commit_write operations, so even if we split
 223  * up to use "smaller" transactions none of them could complete until
 224  * all of them were opened.  By having a single journal transaction,
 225  * we eliminate duplicate reservations for common blocks like the
 226  * superblock and group descriptors or bitmaps.
 227  *
 228  * We will start the transaction here, but each prepare_write will
 229  * add a refcount to the transaction, and each commit_write will
 230  * remove a refcount.  The transaction will be closed when all of
 231  * the pages have been written.
 232  */
 233 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
 234                                    int niocount, struct niobuf_local *nb,
 235                                    void *desc_private, int logs)
 236 {
 237         journal_t *journal;
 238         handle_t *handle;
 239         int needed;
 240         ENTRY;
 241
 242         LASSERT(current->journal_info == desc_private);
 243         journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
 244         needed = fsfilt_ext3_credits_needed(objcount, fso, niocount, nb);
 245
 246         /* The number of blocks we could _possibly_ dirty can very large.
 247          * We reduce our request if it is absurd (and we couldn't get that
 248          * many credits for a single handle anyways).
 249          *
 250          * At some point we have to limit the size of I/Os sent at one time,
 251          * increase the size of the journal, or we have to calculate the
 252          * actual journal requirements more carefully by checking all of
 253          * the blocks instead of being maximally pessimistic.  It remains to
 254          * be seen if this is a real problem or not.
 255          */
 256         if (needed > journal->j_max_transaction_buffers) {
 257                 CERROR("want too many journal credits (%d) using %d instead\n",
 258                        needed, journal->j_max_transaction_buffers);
 259                 needed = journal->j_max_transaction_buffers;
 260         }
 261
 262         lock_kernel();
 263         handle = journal_start(journal, needed);
 264         unlock_kernel();
 265         if (IS_ERR(handle)) {
 266                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 267                        PTR_ERR(handle));
 268         } else {
 269                 LASSERT(handle->h_buffer_credits >= needed);
 270                 LASSERT(current->journal_info == handle);
 271         }
 272
 273         RETURN(handle);
 274 }
 275
 276 static int fsfilt_ext3_commit(struct inode *inode, void *h, int force_sync)
 277 {
 278         int rc;
 279         handle_t *handle = h;
 280
 281         LASSERT(current->journal_info == handle);
 282         if (force_sync)
 283                 handle->h_sync = 1; /* recovery likes this */
 284
 285         lock_kernel();
 286         rc = journal_stop(handle);
 287         unlock_kernel();
 288
 289         // LASSERT(current->journal_info == NULL);
 290         return rc;
 291 }
 292
 293 static int fsfilt_ext3_commit_async(struct inode *inode, void *h,
 294                                         void **wait_handle)
 295 {
 296         transaction_t *transaction;
 297         unsigned long tid, rtid;
 298         handle_t *handle = h;
 299         journal_t *journal;
 300         int rc;
 301
 302         LASSERT(current->journal_info == handle);
 303
 304         lock_kernel();
 305         transaction = handle->h_transaction;
 306         journal = transaction->t_journal;
 307         tid = transaction->t_tid;
 308         /* we don't want to be blocked */
 309         handle->h_sync = 0;
 310         rc = journal_stop(handle);
 311         if (rc) {
 312                 CERROR("error while stopping transaction: %d\n", rc);
 313                 unlock_kernel();
 314                 return rc;
 315         }
 316 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 317         rtid = log_start_commit(journal, transaction);
 318         if (rtid != tid)
 319                 CERROR("strange race: %lu != %lu\n",
 320                        (unsigned long) tid, (unsigned long) rtid);
 321 #else
 322         log_start_commit(journal, transaction->t_tid);
 323 #endif
 324         unlock_kernel();
 325
 326         *wait_handle = (void *) tid;
 327         CDEBUG(D_INODE, "commit async: %lu\n", (unsigned long) tid);
 328         return 0;
 329 }
 330
 331 static int fsfilt_ext3_commit_wait(struct inode *inode, void *h)
 332 {
 333         tid_t tid = (tid_t)(long)h;
 334
 335         CDEBUG(D_INODE, "commit wait: %lu\n", (unsigned long) tid);
 336         if (is_journal_aborted(EXT3_JOURNAL(inode)))
 337                 return -EIO;
 338
 339         log_wait_commit(EXT3_JOURNAL(inode), tid);
 340
 341         return 0;
 342 }
 343
 344 static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
 345                                struct iattr *iattr, int do_trunc)
 346 {
 347         struct inode *inode = dentry->d_inode;
 348         int rc;
 349
 350         lock_kernel();
 351
 352         /* A _really_ horrible hack to avoid removing the data stored
 353          * in the block pointers; this is really the "small" stripe MD data.
 354          * We can avoid further hackery by virtue of the MDS file size being
 355          * zero all the time (which doesn't invoke block truncate at unlink
 356          * time), so we assert we never change the MDS file size from zero. */
 357         if (iattr->ia_valid & ATTR_SIZE && !do_trunc) {
 358                 /* ATTR_SIZE would invoke truncate: clear it */
 359                 iattr->ia_valid &= ~ATTR_SIZE;
 360                 EXT3_I(inode)->i_disksize = inode->i_size = iattr->ia_size;
 361
 362                 /* make sure _something_ gets set - so new inode
 363                  * goes to disk (probably won't work over XFS */
 364                 if (!(iattr->ia_valid & (ATTR_MODE | ATTR_MTIME | ATTR_CTIME))){
 365                         iattr->ia_valid |= ATTR_MODE;
 366                         iattr->ia_mode = inode->i_mode;
 367                 }
 368         }
 369
 370         /* Don't allow setattr to change file type */
 371         iattr->ia_mode = (inode->i_mode & S_IFMT)|(iattr->ia_mode & ~S_IFMT);
 372
 373         /* We set these flags on the client, but have already checked perms
 374          * so don't confuse inode_change_ok. */
 375         iattr->ia_valid &= ~(ATTR_MTIME_SET | ATTR_ATIME_SET);
 376
 377         if (inode->i_op->setattr) {
 378                 rc = inode->i_op->setattr(dentry, iattr);
 379         } else {
 380                 rc = inode_change_ok(inode, iattr);
 381                 if (!rc)
 382                         rc = inode_setattr(inode, iattr);
 383         }
 384
 385         unlock_kernel();
 386
 387         return rc;
 388 }
 389
 390 static int fsfilt_ext3_iocontrol(struct inode * inode, struct file *file,
 391                                  unsigned int cmd, unsigned long arg)
 392 {
 393         int rc = 0;
 394         ENTRY;
 395
 396         if (inode->i_fop->ioctl)
 397                 rc = inode->i_fop->ioctl(inode, file, cmd, arg);
 398         else
 399                 RETURN(-ENOTTY);
 400
 401         RETURN(rc);
 402 }
 403
 404 #undef INLINE_EA
 405 #undef OLD_EA
 406 static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
 407                               void *lmm, int lmm_size)
 408 {
 409         int rc, old_ea = 0;
 410
 411 #ifdef INLINE_EA  /* can go away before 1.0 - just for testing bug 2097 now */
 412         /* Nasty hack city - store stripe MD data in the block pointers if
 413          * it will fit, because putting it in an EA currently kills the MDS
 414          * performance.  We'll fix this with "fast EAs" in the future.
 415          */
 416         if (inode->i_blocks == 0 && lmm_size <= sizeof(EXT3_I(inode)->i_data) -
 417                                             sizeof(EXT3_I(inode)->i_data[0])) {
 418                 unsigned old_size = EXT3_I(inode)->i_data[0];
 419                 if (old_size != 0) {
 420                         LASSERT(old_size < sizeof(EXT3_I(inode)->i_data));
 421                         CERROR("setting EA on %lu/%u again... interesting\n",
 422                                inode->i_ino, inode->i_generation);
 423                 }
 424
 425                 EXT3_I(inode)->i_data[0] = cpu_to_le32(lmm_size);
 426                 memcpy(&EXT3_I(inode)->i_data[1], lmm, lmm_size);
 427                 mark_inode_dirty(inode);
 428                 return 0;
 429         }
 430 #endif
 431 #ifdef OLD_EA
 432         /* keep this when we get rid of OLD_EA (too noisy during conversion) */
 433         if (EXT3_I(inode)->i_file_acl /* || large inode EA flag */) {
 434                 CWARN("setting EA on %lu/%u again... interesting\n",
 435                        inode->i_ino, inode->i_generation);
 436                 old_ea = 1;
 437         }
 438
 439         lock_kernel();
 440         /* this can go away before 1.0.  For bug 2097 testing only. */
 441         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_LUSTRE,
 442                                    XATTR_LUSTRE_MDS_OBJID, lmm, lmm_size, 0);
 443 #else
 444         lock_kernel();
 445         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
 446                                    XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size, 0);
 447
 448         /* This tries to delete the old-format LOV EA, but only as long as we
 449          * have successfully saved the new-format LOV EA (we can always try
 450          * the conversion again the next time the file is accessed).  It is
 451          * possible (although unlikely) that the new-format LOV EA couldn't be
 452          * saved because it ran out of space but we would need a file striped
 453          * over least 123 OSTs before the two EAs filled a 4kB block.
 454          *
 455          * This can be removed when all filesystems have converted to the
 456          * new EA format, but otherwise adds little if any overhead.  If we
 457          * wanted backward compatibility for existing files, we could keep
 458          * the old EA around for a while but we'd have to clean it up later. */
 459         if (rc >= 0 && old_ea) {
 460                 int err = ext3_xattr_set_handle(handle, inode,
 461                                                 EXT3_XATTR_INDEX_LUSTRE,
 462                                                 XATTR_LUSTRE_MDS_OBJID,
 463                                                 NULL, 0, 0);
 464                 if (err)
 465                         CERROR("error deleting old LOV EA on %lu/%u: rc %d\n",
 466                                inode->i_ino, inode->i_generation, err);
 467         }
 468 #endif
 469         unlock_kernel();
 470
 471         if (rc)
 472                 CERROR("error adding MD data to inode %lu: rc = %d\n",
 473                        inode->i_ino, rc);
 474         return rc;
 475 }
 476
 477 /* Must be called with i_sem held */
 478 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm, int lmm_size)
 479 {
 480         int rc;
 481
 482         LASSERT(down_trylock(&inode->i_sem) != 0);
 483         lock_kernel();
 484         /* Keep support for reading "inline EAs" until we convert
 485          * users over to new format entirely.  See bug 841/2097. */
 486         if (inode->i_blocks == 0 && EXT3_I(inode)->i_data[0]) {
 487                 unsigned size = le32_to_cpu(EXT3_I(inode)->i_data[0]);
 488                 void *handle;
 489
 490                 LASSERT(size < sizeof(EXT3_I(inode)->i_data));
 491                 if (lmm) {
 492                         if (size > lmm_size) {
 493                                 CERROR("inline EA on %lu/%u bad size %u > %u\n",
 494                                        inode->i_ino, inode->i_generation,
 495                                        size, lmm_size);
 496                                 return -ERANGE;
 497                         }
 498                         memcpy(lmm, &EXT3_I(inode)->i_data[1], size);
 499                 }
 500
 501 #ifndef INLINE_EA
 502                 /* migrate LOV EA data to external block - keep same format */
 503                 CWARN("DEBUG: migrate inline EA for inode %lu/%u to block\n",
 504                       inode->i_ino, inode->i_generation);
 505
 506                 handle = journal_start(EXT3_JOURNAL(inode),
 507                                        EXT3_XATTR_TRANS_BLOCKS);
 508                 if (!IS_ERR(handle)) {
 509                         int err;
 510                         rc = fsfilt_ext3_set_md(inode, handle,
 511                                                 &EXT3_I(inode)->i_data[1],size);
 512                         if (rc == 0) {
 513                                 memset(EXT3_I(inode)->i_data, 0,
 514                                        sizeof(EXT3_I(inode)->i_data));
 515                                 mark_inode_dirty(inode);
 516                         }
 517                         err = journal_stop(handle);
 518                         if (err && rc == 0)
 519                                 rc = err;
 520                 } else {
 521                         rc = PTR_ERR(handle);
 522                 }
 523 #endif
 524                 unlock_kernel();
 525                 return size;
 526         }
 527
 528         rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
 529                             XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size);
 530         /* try old EA type if new one failed - MDS will convert it for us */
 531         if (rc == -ENODATA) {
 532                 CDEBUG(D_INFO,"failed new LOV EA %d/%s from inode %lu: rc %d\n",
 533                        EXT3_XATTR_INDEX_TRUSTED, XATTR_LUSTRE_MDS_LOV_EA,
 534                        inode->i_ino, rc);
 535
 536                 rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_LUSTRE,
 537                                     XATTR_LUSTRE_MDS_OBJID, lmm, lmm_size);
 538         }
 539         unlock_kernel();
 540
 541         /* This gives us the MD size */
 542         if (lmm == NULL)
 543                 return (rc == -ENODATA) ? 0 : rc;
 544
 545         if (rc < 0) {
 546                 CDEBUG(D_INFO, "error getting EA %d/%s from inode %lu: rc %d\n",
 547                        EXT3_XATTR_INDEX_LUSTRE, XATTR_LUSTRE_MDS_OBJID,
 548                        inode->i_ino, rc);
 549                 memset(lmm, 0, lmm_size);
 550                 return (rc == -ENODATA) ? 0 : rc;
 551         }
 552
 553         return rc;
 554 }
 555
 556 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
 557                                     loff_t *off)
 558 {
 559         struct inode *inode = file->f_dentry->d_inode;
 560         int rc = 0;
 561
 562         if (S_ISREG(inode->i_mode))
 563                 rc = file->f_op->read(file, buf, count, off);
 564         else {
 565                 const int blkbits = inode->i_sb->s_blocksize_bits;
 566                 const int blksize = inode->i_sb->s_blocksize;
 567
 568                 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
 569                        count, inode->i_ino, *off);
 570                 while (count > 0) {
 571                         struct buffer_head *bh;
 572
 573                         bh = NULL;
 574                         if (*off < inode->i_size) {
 575                                 int err = 0;
 576
 577                                 bh = ext3_bread(NULL, inode, *off >> blkbits,
 578                                                 0, &err);
 579
 580                                 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
 581
 582                                 if (bh) {
 583                                         memcpy(buf, bh->b_data, blksize);
 584                                         brelse(bh);
 585                                 } else if (err) {
 586                                         /* XXX in theory we should just fake
 587                                          * this buffer and continue like ext3,
 588                                          * especially if this is a partial read
 589                                          */
 590                                         CERROR("error read dir %lu+%llu: %d\n",
 591                                                inode->i_ino, *off, err);
 592                                         RETURN(err);
 593                                 }
 594                         }
 595                         if (!bh) {
 596                                 struct ext3_dir_entry_2 *fake = (void *)buf;
 597
 598                                 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
 599                                 memset(fake, 0, sizeof(*fake));
 600                                 fake->rec_len = cpu_to_le32(blksize);
 601                         }
 602                         count -= blksize;
 603                         buf += blksize;
 604                         *off += blksize;
 605                         rc += blksize;
 606                 }
 607         }
 608
 609         return rc;
 610 }
 611
 612 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
 613 {
 614         struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
 615
 616         fcb->cb_func(fcb->cb_obd, fcb->cb_last_rcvd, fcb->cb_data, error);
 617
 618         OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
 619         atomic_dec(&fcb_cache_count);
 620 }
 621
 622 static int fsfilt_ext3_add_journal_cb(struct obd_device *obd, __u64 last_rcvd,
 623                                       void *handle, fsfilt_cb_t cb_func,
 624                                       void *cb_data)
 625 {
 626         struct fsfilt_cb_data *fcb;
 627
 628         OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
 629         if (fcb == NULL)
 630                 RETURN(-ENOMEM);
 631
 632         atomic_inc(&fcb_cache_count);
 633         fcb->cb_func = cb_func;
 634         fcb->cb_obd = obd;
 635         fcb->cb_last_rcvd = last_rcvd;
 636         fcb->cb_data = cb_data;
 637
 638         CDEBUG(D_EXT2, "set callback for last_rcvd: "LPD64"\n", last_rcvd);
 639         lock_kernel();
 640         journal_callback_set(handle, fsfilt_ext3_cb_func,
 641                              (struct journal_callback *)fcb);
 642         unlock_kernel();
 643
 644         return 0;
 645 }
 646
 647 /*
 648  * We need to hack the return value for the free inode counts because
 649  * the current EA code requires one filesystem block per inode with EAs,
 650  * so it is possible to run out of blocks before we run out of inodes.
 651  *
 652  * This can be removed when the ext3 EA code is fixed.
 653  */
 654 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
 655 {
 656         struct kstatfs sfs;
 657         int rc;
 658
 659         memset(&sfs, 0, sizeof(sfs));
 660
 661         rc = sb->s_op->statfs(sb, &sfs);
 662
 663         if (!rc && sfs.f_bfree < sfs.f_ffree) {
 664                 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
 665                 sfs.f_ffree = sfs.f_bfree;
 666         }
 667
 668         statfs_pack(osfs, &sfs);
 669         return rc;
 670 }
 671
 672 static int fsfilt_ext3_sync(struct super_block *sb)
 673 {
 674         return ext3_force_commit(sb);
 675 }
 676
 677 extern int ext3_map_inode_page(struct inode *inode, struct page *page,
 678                                unsigned long *blocks, int *created, int create);
 679 int fsfilt_ext3_map_inode_page(struct inode *inode, struct page *page,
 680                                unsigned long *blocks, int *created, int create)
 681 {
 682         return ext3_map_inode_page(inode, page, blocks, created, create);
 683 }
 684
 685 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
 686                                int nblocks, loff_t newsize);
 687 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
 688                                       int nblocks, loff_t newsize)
 689 {
 690         return ext3_prep_san_write(inode, blocks, nblocks, newsize);
 691 }
 692
 693 static int fsfilt_ext3_read_record(struct file * file, void *buf,
 694                                    int size, loff_t *offs)
 695 {
 696         struct inode *inode = file->f_dentry->d_inode;
 697         unsigned long block;
 698         struct buffer_head *bh;
 699         int err, blocksize, csize, boffs;
 700
 701         /* prevent reading after eof */
 702         lock_kernel();
 703         if (inode->i_size < *offs + size) {
 704                 size = inode->i_size - *offs;
 705                 unlock_kernel();
 706                 if (size < 0) {
 707                         CERROR("size %llu is too short for read %u@%llu\n",
 708                                inode->i_size, size, *offs);
 709                         return -EIO;
 710                 } else if (size == 0) {
 711                         return 0;
 712                 }
 713         } else {
 714                 unlock_kernel();
 715         }
 716
 717         blocksize = 1 << inode->i_blkbits;
 718
 719         while (size > 0) {
 720                 block = *offs >> inode->i_blkbits;
 721                 boffs = *offs & (blocksize - 1);
 722                 csize = min(blocksize - boffs, size);
 723                 bh = ext3_bread(NULL, inode, block, 0, &err);
 724                 if (!bh) {
 725                         CERROR("can't read block: %d\n", err);
 726                         return err;
 727                 }
 728
 729                 memcpy(buf, bh->b_data + boffs, csize);
 730                 brelse(bh);
 731
 732                 *offs += csize;
 733                 buf += csize;
 734                 size -= csize;
 735         }
 736         return 0;
 737 }
 738
 739 static int fsfilt_ext3_write_record(struct file *file, void *buf, int bufsize,
 740                                     loff_t *offs, int force_sync)
 741 {
 742         struct buffer_head *bh = NULL;
 743         unsigned long block;
 744         struct inode *inode = file->f_dentry->d_inode;
 745         loff_t old_size = inode->i_size, offset = *offs;
 746         loff_t new_size = inode->i_size;
 747         journal_t *journal;
 748         handle_t *handle;
 749         int err, block_count = 0, blocksize, size, boffs;
 750
 751         /* Determine how many transaction credits are needed */
 752         blocksize = 1 << inode->i_blkbits;
 753         block_count = (*offs & (blocksize - 1)) + bufsize;
 754         block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
 755
 756         journal = EXT3_SB(inode->i_sb)->s_journal;
 757         lock_kernel();
 758         handle = journal_start(journal,
 759                                block_count * EXT3_DATA_TRANS_BLOCKS + 2);
 760         unlock_kernel();
 761         if (IS_ERR(handle)) {
 762                 CERROR("can't start transaction\n");
 763                 return PTR_ERR(handle);
 764         }
 765
 766         while (bufsize > 0) {
 767                 if (bh != NULL)
 768                         brelse(bh);
 769
 770                 block = offset >> inode->i_blkbits;
 771                 boffs = offset & (blocksize - 1);
 772                 size = min(blocksize - boffs, bufsize);
 773                 bh = ext3_bread(handle, inode, block, 1, &err);
 774                 if (!bh) {
 775                         CERROR("can't read/create block: %d\n", err);
 776                         goto out;
 777                 }
 778
 779                 err = ext3_journal_get_write_access(handle, bh);
 780                 if (err) {
 781                         CERROR("journal_get_write_access() returned error %d\n",
 782                                err);
 783                         goto out;
 784                 }
 785                 LASSERT(bh->b_data + boffs + size <= bh->b_data + bh->b_size);
 786                 memcpy(bh->b_data + boffs, buf, size);
 787                 err = ext3_journal_dirty_metadata(handle, bh);
 788                 if (err) {
 789                         CERROR("journal_dirty_metadata() returned error %d\n",
 790                                err);
 791                         goto out;
 792                 }
 793                 if (offset + size > new_size)
 794                         new_size = offset + size;
 795                 offset += size;
 796                 bufsize -= size;
 797                 buf += size;
 798         }
 799
 800         if (force_sync)
 801                 handle->h_sync = 1; /* recovery likes this */
 802 out:
 803         if (bh)
 804                 brelse(bh);
 805
 806         /* correct in-core and on-disk sizes */
 807         if (new_size > inode->i_size) {
 808                 lock_kernel();
 809                 if (new_size > inode->i_size)
 810                         inode->i_size = new_size;
 811                 if (inode->i_size > EXT3_I(inode)->i_disksize)
 812                         EXT3_I(inode)->i_disksize = inode->i_size;
 813                 if (inode->i_size > old_size)
 814                         mark_inode_dirty(inode);
 815                 unlock_kernel();
 816         }
 817
 818         lock_kernel();
 819         journal_stop(handle);
 820         unlock_kernel();
 821
 822         if (err == 0)
 823                 *offs = offset;
 824         return err;
 825 }
 826
 827 static int fsfilt_ext3_setup(struct super_block *sb)
 828 {
 829 #if 0
 830         EXT3_SB(sb)->dx_lock = fsfilt_ext3_dx_lock;
 831         EXT3_SB(sb)->dx_unlock = fsfilt_ext3_dx_unlock;
 832 #endif
 833 #ifdef S_PDIROPS
 834         CWARN("Enabling PDIROPS\n");
 835         set_opt(EXT3_SB(sb)->s_mount_opt, PDIROPS);
 836         sb->s_flags |= S_PDIROPS;
 837 #endif
 838         return 0;
 839 }
 840
 841 /* If fso is NULL, op is FSFILT operation, otherwise op is number of fso
 842    objects. Logs is number of logfiles to update */
 843 static int fsfilt_ext3_get_op_len(int op, struct fsfilt_objinfo *fso, int logs)
 844 {
 845         if ( !fso ) {
 846                 switch(op) {
 847                 case FSFILT_OP_CREATE:
 848                                  /* directory leaf, index & indirect & EA*/
 849                         return 4 + 3 * logs;
 850                 case FSFILT_OP_UNLINK:
 851                         return 3 * logs;
 852                 }
 853
 854         } else {
 855                 int i;
 856                 int needed = 0;
 857                 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
 858                 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 859                 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
 860                 for (i = 0; i < op; i++, fso++) {
 861                         int nblocks = fso->fso_bufcnt * blockpp;
 862                         int ndindirect = min(nblocks, addrpp + 1);
 863                         int nindir = nblocks + ndindirect + 1;
 864
 865                         needed += nindir;
 866                 }
 867                 return needed + 3 * logs;
 868         }
 869
 870         return 0;
 871 }
 872
 873 static struct fsfilt_operations fsfilt_ext3_ops = {
 874         fs_type:                "ext3",
 875         fs_owner:               THIS_MODULE,
 876         fs_start:               fsfilt_ext3_start,
 877         fs_brw_start:           fsfilt_ext3_brw_start,
 878         fs_commit:              fsfilt_ext3_commit,
 879         fs_commit_async:        fsfilt_ext3_commit_async,
 880         fs_commit_wait:         fsfilt_ext3_commit_wait,
 881         fs_setattr:             fsfilt_ext3_setattr,
 882         fs_iocontrol:           fsfilt_ext3_iocontrol,
 883         fs_set_md:              fsfilt_ext3_set_md,
 884         fs_get_md:              fsfilt_ext3_get_md,
 885         fs_readpage:            fsfilt_ext3_readpage,
 886         fs_add_journal_cb:      fsfilt_ext3_add_journal_cb,
 887         fs_statfs:              fsfilt_ext3_statfs,
 888         fs_sync:                fsfilt_ext3_sync,
 889         fs_map_inode_page:      fsfilt_ext3_map_inode_page,
 890         fs_prep_san_write:      fsfilt_ext3_prep_san_write,
 891         fs_write_record:        fsfilt_ext3_write_record,
 892         fs_read_record:         fsfilt_ext3_read_record,
 893         fs_setup:               fsfilt_ext3_setup,
 894         fs_get_op_len:          fsfilt_ext3_get_op_len,
 895 };
 896
 897 static int __init fsfilt_ext3_init(void)
 898 {
 899         int rc;
 900
 901         //rc = ext3_xattr_register();
 902         fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
 903                                       sizeof(struct fsfilt_cb_data), 0,
 904                                       0, NULL, NULL);
 905         if (!fcb_cache) {
 906                 CERROR("error allocating fsfilt journal callback cache\n");
 907                 GOTO(out, rc = -ENOMEM);
 908         }
 909
 910         rc = fsfilt_register_ops(&fsfilt_ext3_ops);
 911
 912         if (rc)
 913                 kmem_cache_destroy(fcb_cache);
 914 out:
 915         return rc;
 916 }
 917
 918 static void __exit fsfilt_ext3_exit(void)
 919 {
 920         int rc;
 921
 922         fsfilt_unregister_ops(&fsfilt_ext3_ops);
 923         rc = kmem_cache_destroy(fcb_cache);
 924
 925         if (rc || atomic_read(&fcb_cache_count)) {
 926                 CERROR("can't free fsfilt callback cache: count %d, rc = %d\n",
 927                        atomic_read(&fcb_cache_count), rc);
 928         }
 929
 930         //rc = ext3_xattr_unregister();
 931 }
 932
 933 module_init(fsfilt_ext3_init);
 934 module_exit(fsfilt_ext3_exit);
 935
 936 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
 937 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
 938 MODULE_LICENSE("GPL");