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 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
  39 #include <linux/ext3_xattr.h>
  40 #else
  41 #include <ext3/xattr.h>
  42 #endif
  43
  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                                int logs)
  76 {
  77         /* For updates to the last recieved file */
  78         int nblocks = EXT3_SINGLEDATA_TRANS_BLOCKS;
  79         journal_t *journal;
  80         void *handle;
  81
  82         if (current->journal_info) {
  83                 CDEBUG(D_INODE, "increasing refcount on %p\n",
  84                        current->journal_info);
  85                 goto journal_start;
  86         }
  87
  88         switch(op) {
  89         case FSFILT_OP_RMDIR:
  90         case FSFILT_OP_UNLINK:
  91                 /* delete one file + create/update logs for each stripe */
  92                 nblocks += EXT3_DELETE_TRANS_BLOCKS;
  93                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
  94                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
  95                 break;
  96         case FSFILT_OP_RENAME:
  97                 /* modify additional directory */
  98                 nblocks += EXT3_SINGLEDATA_TRANS_BLOCKS;
  99                 /* no break */
 100         case FSFILT_OP_SYMLINK:
 101                 /* additional block + block bitmap + GDT for long symlink */
 102                 nblocks += 3;
 103                 /* no break */
 104         case FSFILT_OP_CREATE:
 105                 /* create/update logs for each stripe */
 106                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 107                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
 108                 /* no break */
 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 +
 117                         EXT3_DATA_TRANS_BLOCKS;
 118                 break;
 119         case FSFILT_OP_SETATTR:
 120                 /* Setattr on inode */
 121                 nblocks += 1;
 122                 break;
 123         case FSFILT_OP_CANCEL_UNLINK:
 124                 /* blocks for log header bitmap update OR
 125                  * blocks for catalog header bitmap update + unlink of logs */
 126                 nblocks = (LLOG_CHUNK_SIZE >> inode->i_blkbits) +
 127                         EXT3_DELETE_TRANS_BLOCKS * logs;
 128                 break;
 129         default: CERROR("unknown transaction start op %d\n", op);
 130                  LBUG();
 131         }
 132
 133         LASSERT(current->journal_info == desc_private);
 134         journal = EXT3_SB(inode->i_sb)->s_journal;
 135         if (nblocks > journal->j_max_transaction_buffers) {
 136                 CERROR("too many credits %d for op %ux%u using %d instead\n",
 137                        nblocks, op, logs, journal->j_max_transaction_buffers);
 138                 nblocks = journal->j_max_transaction_buffers;
 139         }
 140
 141  journal_start:
 142         lock_kernel();
 143         handle = journal_start(EXT3_JOURNAL(inode), nblocks);
 144         unlock_kernel();
 145
 146         if (!IS_ERR(handle))
 147                 LASSERT(current->journal_info == handle);
 148         else
 149                 CERROR("error starting handle for op %u (%u credits): rc %ld\n",
 150                        op, nblocks, PTR_ERR(handle));
 151         return handle;
 152 }
 153
 154 /*
 155  * Calculate the number of buffer credits needed to write multiple pages in
 156  * a single ext3 transaction.  No, this shouldn't be here, but as yet ext3
 157  * doesn't have a nice API for calculating this sort of thing in advance.
 158  *
 159  * See comment above ext3_writepage_trans_blocks for details.  We assume
 160  * no data journaling is being done, but it does allow for all of the pages
 161  * being non-contiguous.  If we are guaranteed contiguous pages we could
 162  * reduce the number of (d)indirect blocks a lot.
 163  *
 164  * With N blocks per page and P pages, for each inode we have at most:
 165  * N*P indirect
 166  * min(N*P, blocksize/4 + 1) dindirect blocks
 167  * niocount tindirect
 168  *
 169  * For the entire filesystem, we have at most:
 170  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 171  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 172  * objcount inode blocks
 173  * 1 superblock
 174  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 175  *
 176  * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
 177  */
 178 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso,
 179                                       int niocount, struct niobuf_local *nb)
 180 {
 181         struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
 182         __u64 next_indir;
 183         const int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 184         int nbitmaps = 0, ngdblocks;
 185         int needed = objcount + 1; /* inodes + superblock */
 186         int i, j;
 187
 188         for (i = 0, j = 0; i < objcount; i++, fso++) {
 189                 /* two or more dindirect blocks in case we cross boundary */
 190                 int ndind = (long)((nb[j + fso->fso_bufcnt - 1].offset -
 191                                     nb[j].offset) >>
 192                                    sb->s_blocksize_bits) /
 193                         (EXT3_ADDR_PER_BLOCK(sb) * EXT3_ADDR_PER_BLOCK(sb));
 194                 nbitmaps += min(fso->fso_bufcnt, ndind > 0 ? ndind : 2);
 195
 196                 /* leaf, indirect, tindirect blocks for first block */
 197                 nbitmaps += blockpp + 2;
 198
 199                 j += fso->fso_bufcnt;
 200         }
 201
 202         next_indir = nb[0].offset +
 203                 (EXT3_ADDR_PER_BLOCK(sb) << sb->s_blocksize_bits);
 204         for (i = 1; i < niocount; i++) {
 205                 if (nb[i].offset >= next_indir) {
 206                         nbitmaps++;     /* additional indirect */
 207                         next_indir = nb[i].offset +
 208                                 (EXT3_ADDR_PER_BLOCK(sb)<<sb->s_blocksize_bits);
 209                 } else if (nb[i].offset != nb[i - 1].offset + sb->s_blocksize) {
 210                         nbitmaps++;     /* additional indirect */
 211                 }
 212                 nbitmaps += blockpp;    /* each leaf in different group? */
 213         }
 214
 215         ngdblocks = nbitmaps;
 216         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 217                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 218         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 219                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 220
 221         needed += nbitmaps + ngdblocks;
 222
 223         /* last_rcvd update */
 224         needed += EXT3_DATA_TRANS_BLOCKS;
 225
 226 #if defined(CONFIG_QUOTA) && !defined(__x86_64__) /* XXX */
 227         /* We assume that there will be 1 bit set in s_dquot.flags for each
 228          * quota file that is active.  This is at least true for now.
 229          */
 230         needed += hweight32(sb_any_quota_enabled(sb)) *
 231                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 232 #endif
 233
 234         return needed;
 235 }
 236
 237 /* We have to start a huge journal transaction here to hold all of the
 238  * metadata for the pages being written here.  This is necessitated by
 239  * the fact that we do lots of prepare_write operations before we do
 240  * any of the matching commit_write operations, so even if we split
 241  * up to use "smaller" transactions none of them could complete until
 242  * all of them were opened.  By having a single journal transaction,
 243  * we eliminate duplicate reservations for common blocks like the
 244  * superblock and group descriptors or bitmaps.
 245  *
 246  * We will start the transaction here, but each prepare_write will
 247  * add a refcount to the transaction, and each commit_write will
 248  * remove a refcount.  The transaction will be closed when all of
 249  * the pages have been written.
 250  */
 251 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
 252                                    int niocount, struct niobuf_local *nb,
 253                                    void *desc_private, int logs)
 254 {
 255         journal_t *journal;
 256         handle_t *handle;
 257         int needed;
 258         ENTRY;
 259
 260         LASSERT(current->journal_info == desc_private);
 261         journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
 262         needed = fsfilt_ext3_credits_needed(objcount, fso, niocount, nb);
 263
 264         /* The number of blocks we could _possibly_ dirty can very large.
 265          * We reduce our request if it is absurd (and we couldn't get that
 266          * many credits for a single handle anyways).
 267          *
 268          * At some point we have to limit the size of I/Os sent at one time,
 269          * increase the size of the journal, or we have to calculate the
 270          * actual journal requirements more carefully by checking all of
 271          * the blocks instead of being maximally pessimistic.  It remains to
 272          * be seen if this is a real problem or not.
 273          */
 274         if (needed > journal->j_max_transaction_buffers) {
 275                 CERROR("want too many journal credits (%d) using %d instead\n",
 276                        needed, journal->j_max_transaction_buffers);
 277                 needed = journal->j_max_transaction_buffers;
 278         }
 279
 280         lock_kernel();
 281         handle = journal_start(journal, needed);
 282         unlock_kernel();
 283         if (IS_ERR(handle)) {
 284                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 285                        PTR_ERR(handle));
 286         } else {
 287                 LASSERT(handle->h_buffer_credits >= needed);
 288                 LASSERT(current->journal_info == handle);
 289         }
 290
 291         RETURN(handle);
 292 }
 293
 294 static int fsfilt_ext3_commit(struct inode *inode, void *h, int force_sync)
 295 {
 296         int rc;
 297         handle_t *handle = h;
 298
 299         LASSERT(current->journal_info == handle);
 300         if (force_sync)
 301                 handle->h_sync = 1; /* recovery likes this */
 302
 303         lock_kernel();
 304         rc = journal_stop(handle);
 305         unlock_kernel();
 306
 307         return rc;
 308 }
 309
 310 static int fsfilt_ext3_commit_async(struct inode *inode, void *h,
 311                                     void **wait_handle)
 312 {
 313         unsigned long tid;
 314         transaction_t *transaction;
 315 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 316         unsigned long rtid;
 317 #endif
 318         handle_t *handle = h;
 319         journal_t *journal;
 320         int rc;
 321
 322         LASSERT(current->journal_info == handle);
 323
 324         lock_kernel();
 325         transaction = handle->h_transaction;
 326         journal = transaction->t_journal;
 327         tid = transaction->t_tid;
 328         /* we don't want to be blocked */
 329         handle->h_sync = 0;
 330         rc = journal_stop(handle);
 331         if (rc) {
 332                 CERROR("error while stopping transaction: %d\n", rc);
 333                 unlock_kernel();
 334                 return rc;
 335         }
 336 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 337         rtid = log_start_commit(journal, transaction);
 338         if (rtid != tid)
 339                 CERROR("strange race: %lu != %lu\n",
 340                        (unsigned long) tid, (unsigned long) rtid);
 341 #else
 342         log_start_commit(journal, transaction->t_tid);
 343 #endif
 344         unlock_kernel();
 345
 346         *wait_handle = (void *) tid;
 347         CDEBUG(D_INODE, "commit async: %lu\n", (unsigned long) tid);
 348         return 0;
 349 }
 350
 351 static int fsfilt_ext3_commit_wait(struct inode *inode, void *h)
 352 {
 353         tid_t tid = (tid_t)(long)h;
 354
 355         CDEBUG(D_INODE, "commit wait: %lu\n", (unsigned long) tid);
 356         if (is_journal_aborted(EXT3_JOURNAL(inode)))
 357                 return -EIO;
 358
 359         log_wait_commit(EXT3_JOURNAL(inode), tid);
 360
 361         return 0;
 362 }
 363
 364 static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
 365                                struct iattr *iattr, int do_trunc)
 366 {
 367         struct inode *inode = dentry->d_inode;
 368         int rc;
 369
 370         lock_kernel();
 371
 372         /* A _really_ horrible hack to avoid removing the data stored
 373          * in the block pointers; this is really the "small" stripe MD data.
 374          * We can avoid further hackery by virtue of the MDS file size being
 375          * zero all the time (which doesn't invoke block truncate at unlink
 376          * time), so we assert we never change the MDS file size from zero. */
 377         if (iattr->ia_valid & ATTR_SIZE && !do_trunc) {
 378                 /* ATTR_SIZE would invoke truncate: clear it */
 379                 iattr->ia_valid &= ~ATTR_SIZE;
 380                 EXT3_I(inode)->i_disksize = inode->i_size = iattr->ia_size;
 381
 382                 /* make sure _something_ gets set - so new inode
 383                  * goes to disk (probably won't work over XFS */
 384                 if (!(iattr->ia_valid & (ATTR_MODE | ATTR_MTIME | ATTR_CTIME))){
 385                         iattr->ia_valid |= ATTR_MODE;
 386                         iattr->ia_mode = inode->i_mode;
 387                 }
 388         }
 389
 390         /* Don't allow setattr to change file type */
 391         iattr->ia_mode = (inode->i_mode & S_IFMT)|(iattr->ia_mode & ~S_IFMT);
 392
 393         /* We set these flags on the client, but have already checked perms
 394          * so don't confuse inode_change_ok. */
 395         iattr->ia_valid &= ~(ATTR_MTIME_SET | ATTR_ATIME_SET);
 396
 397         if (inode->i_op->setattr) {
 398                 rc = inode->i_op->setattr(dentry, iattr);
 399         } else {
 400                 rc = inode_change_ok(inode, iattr);
 401                 if (!rc)
 402                         rc = inode_setattr(inode, iattr);
 403         }
 404
 405         unlock_kernel();
 406
 407         return rc;
 408 }
 409
 410 static int fsfilt_ext3_iocontrol(struct inode * inode, struct file *file,
 411                                  unsigned int cmd, unsigned long arg)
 412 {
 413         int rc = 0;
 414         ENTRY;
 415
 416         if (inode->i_fop->ioctl)
 417                 rc = inode->i_fop->ioctl(inode, file, cmd, arg);
 418         else
 419                 RETURN(-ENOTTY);
 420
 421         RETURN(rc);
 422 }
 423
 424 #undef INLINE_EA
 425 #undef OLD_EA
 426 static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
 427                               void *lmm, int lmm_size)
 428 {
 429         int rc, old_ea = 0;
 430
 431 #ifdef INLINE_EA  /* can go away before 1.0 - just for testing bug 2097 now */
 432         /* Nasty hack city - store stripe MD data in the block pointers if
 433          * it will fit, because putting it in an EA currently kills the MDS
 434          * performance.  We'll fix this with "fast EAs" in the future.
 435          */
 436         if (inode->i_blocks == 0 && lmm_size <= sizeof(EXT3_I(inode)->i_data) -
 437                                             sizeof(EXT3_I(inode)->i_data[0])) {
 438                 unsigned old_size = EXT3_I(inode)->i_data[0];
 439                 if (old_size != 0) {
 440                         LASSERT(old_size < sizeof(EXT3_I(inode)->i_data));
 441                         CERROR("setting EA on %lu/%u again... interesting\n",
 442                                inode->i_ino, inode->i_generation);
 443                 }
 444
 445                 EXT3_I(inode)->i_data[0] = cpu_to_le32(lmm_size);
 446                 memcpy(&EXT3_I(inode)->i_data[1], lmm, lmm_size);
 447                 mark_inode_dirty(inode);
 448                 return 0;
 449         }
 450 #endif
 451 #ifdef OLD_EA
 452         /* keep this when we get rid of OLD_EA (too noisy during conversion) */
 453         if (EXT3_I(inode)->i_file_acl /* || large inode EA flag */) {
 454                 CWARN("setting EA on %lu/%u again... interesting\n",
 455                        inode->i_ino, inode->i_generation);
 456                 old_ea = 1;
 457         }
 458
 459         lock_kernel();
 460         /* this can go away before 1.0.  For bug 2097 testing only. */
 461         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_LUSTRE,
 462                                    XATTR_LUSTRE_MDS_OBJID, lmm, lmm_size, 0);
 463 #else
 464         lock_kernel();
 465         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
 466                                    XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size, 0);
 467
 468         /* This tries to delete the old-format LOV EA, but only as long as we
 469          * have successfully saved the new-format LOV EA (we can always try
 470          * the conversion again the next time the file is accessed).  It is
 471          * possible (although unlikely) that the new-format LOV EA couldn't be
 472          * saved because it ran out of space but we would need a file striped
 473          * over least 123 OSTs before the two EAs filled a 4kB block.
 474          *
 475          * This can be removed when all filesystems have converted to the
 476          * new EA format, but otherwise adds little if any overhead.  If we
 477          * wanted backward compatibility for existing files, we could keep
 478          * the old EA around for a while but we'd have to clean it up later. */
 479         if (rc >= 0 && old_ea) {
 480                 int err = ext3_xattr_set_handle(handle, inode,
 481                                                 EXT3_XATTR_INDEX_LUSTRE,
 482                                                 XATTR_LUSTRE_MDS_OBJID,
 483                                                 NULL, 0, 0);
 484                 if (err)
 485                         CERROR("error deleting old LOV EA on %lu/%u: rc %d\n",
 486                                inode->i_ino, inode->i_generation, err);
 487         }
 488 #endif
 489         unlock_kernel();
 490
 491         if (rc)
 492                 CERROR("error adding MD data to inode %lu: rc = %d\n",
 493                        inode->i_ino, rc);
 494         return rc;
 495 }
 496
 497 /* Must be called with i_sem held */
 498 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm, int lmm_size)
 499 {
 500         int rc;
 501
 502         LASSERT(down_trylock(&inode->i_sem) != 0);
 503         lock_kernel();
 504         /* Keep support for reading "inline EAs" until we convert
 505          * users over to new format entirely.  See bug 841/2097. */
 506         if (inode->i_blocks == 0 && EXT3_I(inode)->i_data[0]) {
 507                 unsigned size = le32_to_cpu(EXT3_I(inode)->i_data[0]);
 508                 void *handle;
 509
 510                 LASSERT(size < sizeof(EXT3_I(inode)->i_data));
 511                 if (lmm) {
 512                         if (size > lmm_size) {
 513                                 CERROR("inline EA on %lu/%u bad size %u > %u\n",
 514                                        inode->i_ino, inode->i_generation,
 515                                        size, lmm_size);
 516                                 return -ERANGE;
 517                         }
 518                         memcpy(lmm, &EXT3_I(inode)->i_data[1], size);
 519                 }
 520
 521 #ifndef INLINE_EA
 522                 /* migrate LOV EA data to external block - keep same format */
 523                 CWARN("DEBUG: migrate inline EA for inode %lu/%u to block\n",
 524                       inode->i_ino, inode->i_generation);
 525
 526                 handle = journal_start(EXT3_JOURNAL(inode),
 527                                        EXT3_XATTR_TRANS_BLOCKS);
 528                 if (!IS_ERR(handle)) {
 529                         int err;
 530                         rc = fsfilt_ext3_set_md(inode, handle,
 531                                                 &EXT3_I(inode)->i_data[1],size);
 532                         if (rc == 0) {
 533                                 memset(EXT3_I(inode)->i_data, 0,
 534                                        sizeof(EXT3_I(inode)->i_data));
 535                                 mark_inode_dirty(inode);
 536                         }
 537                         err = journal_stop(handle);
 538                         if (err && rc == 0)
 539                                 rc = err;
 540                 } else {
 541                         rc = PTR_ERR(handle);
 542                 }
 543 #endif
 544                 unlock_kernel();
 545                 return size;
 546         }
 547
 548         rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
 549                             XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size);
 550         /* try old EA type if new one failed - MDS will convert it for us */
 551         if (rc == -ENODATA) {
 552                 CDEBUG(D_INFO,"failed new LOV EA %d/%s from inode %lu: rc %d\n",
 553                        EXT3_XATTR_INDEX_TRUSTED, XATTR_LUSTRE_MDS_LOV_EA,
 554                        inode->i_ino, rc);
 555
 556                 rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_LUSTRE,
 557                                     XATTR_LUSTRE_MDS_OBJID, lmm, lmm_size);
 558         }
 559         unlock_kernel();
 560
 561         /* This gives us the MD size */
 562         if (lmm == NULL)
 563                 return (rc == -ENODATA) ? 0 : rc;
 564
 565         if (rc < 0) {
 566                 CDEBUG(D_INFO, "error getting EA %d/%s from inode %lu: rc %d\n",
 567                        EXT3_XATTR_INDEX_LUSTRE, XATTR_LUSTRE_MDS_OBJID,
 568                        inode->i_ino, rc);
 569                 memset(lmm, 0, lmm_size);
 570                 return (rc == -ENODATA) ? 0 : rc;
 571         }
 572
 573         return rc;
 574 }
 575
 576 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
 577                                     loff_t *off)
 578 {
 579         struct inode *inode = file->f_dentry->d_inode;
 580         int rc = 0;
 581
 582         if (S_ISREG(inode->i_mode))
 583                 rc = file->f_op->read(file, buf, count, off);
 584         else {
 585                 const int blkbits = inode->i_sb->s_blocksize_bits;
 586                 const int blksize = inode->i_sb->s_blocksize;
 587
 588                 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
 589                        count, inode->i_ino, *off);
 590                 while (count > 0) {
 591                         struct buffer_head *bh;
 592
 593                         bh = NULL;
 594                         if (*off < inode->i_size) {
 595                                 int err = 0;
 596
 597                                 bh = ext3_bread(NULL, inode, *off >> blkbits,
 598                                                 0, &err);
 599
 600                                 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
 601
 602                                 if (bh) {
 603                                         memcpy(buf, bh->b_data, blksize);
 604                                         brelse(bh);
 605                                 } else if (err) {
 606                                         /* XXX in theory we should just fake
 607                                          * this buffer and continue like ext3,
 608                                          * especially if this is a partial read
 609                                          */
 610                                         CERROR("error read dir %lu+%llu: %d\n",
 611                                                inode->i_ino, *off, err);
 612                                         RETURN(err);
 613                                 }
 614                         }
 615                         if (!bh) {
 616                                 struct ext3_dir_entry_2 *fake = (void *)buf;
 617
 618                                 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
 619                                 memset(fake, 0, sizeof(*fake));
 620                                 fake->rec_len = cpu_to_le32(blksize);
 621                         }
 622                         count -= blksize;
 623                         buf += blksize;
 624                         *off += blksize;
 625                         rc += blksize;
 626                 }
 627         }
 628
 629         return rc;
 630 }
 631
 632 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
 633 {
 634         struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
 635
 636         fcb->cb_func(fcb->cb_obd, fcb->cb_last_rcvd, fcb->cb_data, error);
 637
 638         OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
 639         atomic_dec(&fcb_cache_count);
 640 }
 641
 642 static int fsfilt_ext3_add_journal_cb(struct obd_device *obd, __u64 last_rcvd,
 643                                       void *handle, fsfilt_cb_t cb_func,
 644                                       void *cb_data)
 645 {
 646         struct fsfilt_cb_data *fcb;
 647
 648         OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
 649         if (fcb == NULL)
 650                 RETURN(-ENOMEM);
 651
 652         atomic_inc(&fcb_cache_count);
 653         fcb->cb_func = cb_func;
 654         fcb->cb_obd = obd;
 655         fcb->cb_last_rcvd = last_rcvd;
 656         fcb->cb_data = cb_data;
 657
 658         CDEBUG(D_EXT2, "set callback for last_rcvd: "LPD64"\n", last_rcvd);
 659         lock_kernel();
 660         journal_callback_set(handle, fsfilt_ext3_cb_func,
 661                              (struct journal_callback *)fcb);
 662         unlock_kernel();
 663
 664         return 0;
 665 }
 666
 667 /*
 668  * We need to hack the return value for the free inode counts because
 669  * the current EA code requires one filesystem block per inode with EAs,
 670  * so it is possible to run out of blocks before we run out of inodes.
 671  *
 672  * This can be removed when the ext3 EA code is fixed.
 673  */
 674 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
 675 {
 676         struct kstatfs sfs;
 677         int rc;
 678
 679         memset(&sfs, 0, sizeof(sfs));
 680
 681         rc = sb->s_op->statfs(sb, &sfs);
 682
 683         if (!rc && sfs.f_bfree < sfs.f_ffree) {
 684                 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
 685                 sfs.f_ffree = sfs.f_bfree;
 686         }
 687
 688         statfs_pack(osfs, &sfs);
 689         return rc;
 690 }
 691
 692 static int fsfilt_ext3_sync(struct super_block *sb)
 693 {
 694         return ext3_force_commit(sb);
 695 }
 696
 697 extern int ext3_map_inode_page(struct inode *inode, struct page *page,
 698                                unsigned long *blocks, int *created, int create);
 699 int fsfilt_ext3_map_inode_page(struct inode *inode, struct page *page,
 700                                unsigned long *blocks, int *created, int create)
 701 {
 702         return ext3_map_inode_page(inode, page, blocks, created, create);
 703 }
 704
 705 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
 706                                int nblocks, loff_t newsize);
 707 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
 708                                       int nblocks, loff_t newsize)
 709 {
 710         return ext3_prep_san_write(inode, blocks, nblocks, newsize);
 711 }
 712
 713 static int fsfilt_ext3_read_record(struct file * file, void *buf,
 714                                    int size, loff_t *offs)
 715 {
 716         struct inode *inode = file->f_dentry->d_inode;
 717         unsigned long block;
 718         struct buffer_head *bh;
 719         int err, blocksize, csize, boffs;
 720
 721         /* prevent reading after eof */
 722         lock_kernel();
 723         if (inode->i_size < *offs + size) {
 724                 size = inode->i_size - *offs;
 725                 unlock_kernel();
 726                 if (size < 0) {
 727                         CERROR("size %llu is too short for read %u@%llu\n",
 728                                inode->i_size, size, *offs);
 729                         return -EIO;
 730                 } else if (size == 0) {
 731                         return 0;
 732                 }
 733         } else {
 734                 unlock_kernel();
 735         }
 736
 737         blocksize = 1 << inode->i_blkbits;
 738
 739         while (size > 0) {
 740                 block = *offs >> inode->i_blkbits;
 741                 boffs = *offs & (blocksize - 1);
 742                 csize = min(blocksize - boffs, size);
 743                 bh = ext3_bread(NULL, inode, block, 0, &err);
 744                 if (!bh) {
 745                         CERROR("can't read block: %d\n", err);
 746                         return err;
 747                 }
 748
 749                 memcpy(buf, bh->b_data + boffs, csize);
 750                 brelse(bh);
 751
 752                 *offs += csize;
 753                 buf += csize;
 754                 size -= csize;
 755         }
 756         return 0;
 757 }
 758
 759 static int fsfilt_ext3_write_record(struct file *file, void *buf, int bufsize,
 760                                     loff_t *offs, int force_sync)
 761 {
 762         struct buffer_head *bh = NULL;
 763         unsigned long block;
 764         struct inode *inode = file->f_dentry->d_inode;
 765         loff_t old_size = inode->i_size, offset = *offs;
 766         loff_t new_size = inode->i_size;
 767         journal_t *journal;
 768         handle_t *handle;
 769         int err, block_count = 0, blocksize, size, boffs;
 770
 771         /* Determine how many transaction credits are needed */
 772         blocksize = 1 << inode->i_blkbits;
 773         block_count = (*offs & (blocksize - 1)) + bufsize;
 774         block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
 775
 776         journal = EXT3_SB(inode->i_sb)->s_journal;
 777         lock_kernel();
 778         handle = journal_start(journal,
 779                                block_count * EXT3_DATA_TRANS_BLOCKS + 2);
 780         unlock_kernel();
 781         if (IS_ERR(handle)) {
 782                 CERROR("can't start transaction\n");
 783                 return PTR_ERR(handle);
 784         }
 785
 786         while (bufsize > 0) {
 787                 if (bh != NULL)
 788                         brelse(bh);
 789
 790                 block = offset >> inode->i_blkbits;
 791                 boffs = offset & (blocksize - 1);
 792                 size = min(blocksize - boffs, bufsize);
 793                 bh = ext3_bread(handle, inode, block, 1, &err);
 794                 if (!bh) {
 795                         CERROR("can't read/create block: %d\n", err);
 796                         goto out;
 797                 }
 798
 799                 err = ext3_journal_get_write_access(handle, bh);
 800                 if (err) {
 801                         CERROR("journal_get_write_access() returned error %d\n",
 802                                err);
 803                         goto out;
 804                 }
 805                 LASSERT(bh->b_data + boffs + size <= bh->b_data + bh->b_size);
 806                 memcpy(bh->b_data + boffs, buf, size);
 807                 err = ext3_journal_dirty_metadata(handle, bh);
 808                 if (err) {
 809                         CERROR("journal_dirty_metadata() returned error %d\n",
 810                                err);
 811                         goto out;
 812                 }
 813                 if (offset + size > new_size)
 814                         new_size = offset + size;
 815                 offset += size;
 816                 bufsize -= size;
 817                 buf += size;
 818         }
 819
 820         if (force_sync)
 821                 handle->h_sync = 1; /* recovery likes this */
 822 out:
 823         if (bh)
 824                 brelse(bh);
 825
 826         /* correct in-core and on-disk sizes */
 827         if (new_size > inode->i_size) {
 828                 lock_kernel();
 829                 if (new_size > inode->i_size)
 830                         inode->i_size = new_size;
 831                 if (inode->i_size > EXT3_I(inode)->i_disksize)
 832                         EXT3_I(inode)->i_disksize = inode->i_size;
 833                 if (inode->i_size > old_size)
 834                         mark_inode_dirty(inode);
 835                 unlock_kernel();
 836         }
 837
 838         lock_kernel();
 839         journal_stop(handle);
 840         unlock_kernel();
 841
 842         if (err == 0)
 843                 *offs = offset;
 844         return err;
 845 }
 846
 847 static int fsfilt_ext3_setup(struct super_block *sb)
 848 {
 849 #if 0
 850         EXT3_SB(sb)->dx_lock = fsfilt_ext3_dx_lock;
 851         EXT3_SB(sb)->dx_unlock = fsfilt_ext3_dx_unlock;
 852 #endif
 853 #ifdef S_PDIROPS
 854         CWARN("Enabling PDIROPS\n");
 855         set_opt(EXT3_SB(sb)->s_mount_opt, PDIROPS);
 856         sb->s_flags |= S_PDIROPS;
 857 #endif
 858         return 0;
 859 }
 860
 861 /* If fso is NULL, op is FSFILT operation, otherwise op is number of fso
 862    objects. Logs is number of logfiles to update */
 863 static int fsfilt_ext3_get_op_len(int op, struct fsfilt_objinfo *fso, int logs)
 864 {
 865         if ( !fso ) {
 866                 switch(op) {
 867                 case FSFILT_OP_CREATE:
 868                                  /* directory leaf, index & indirect & EA*/
 869                         return 4 + 3 * logs;
 870                 case FSFILT_OP_UNLINK:
 871                         return 3 * logs;
 872                 }
 873         } else {
 874                 int i;
 875                 int needed = 0;
 876                 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
 877                 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 878                 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
 879                 for (i = 0; i < op; i++, fso++) {
 880                         int nblocks = fso->fso_bufcnt * blockpp;
 881                         int ndindirect = min(nblocks, addrpp + 1);
 882                         int nindir = nblocks + ndindirect + 1;
 883
 884                         needed += nindir;
 885                 }
 886                 return needed + 3 * logs;
 887         }
 888
 889         return 0;
 890 }
 891
 892 static struct fsfilt_operations fsfilt_ext3_ops = {
 893         .fs_type                = "ext3",
 894         .fs_owner               = THIS_MODULE,
 895         .fs_start               = fsfilt_ext3_start,
 896         .fs_brw_start           = fsfilt_ext3_brw_start,
 897         .fs_commit              = fsfilt_ext3_commit,
 898         .fs_commit_async        = fsfilt_ext3_commit_async,
 899         .fs_commit_wait         = fsfilt_ext3_commit_wait,
 900         .fs_setattr             = fsfilt_ext3_setattr,
 901         .fs_iocontrol           = fsfilt_ext3_iocontrol,
 902         .fs_set_md              = fsfilt_ext3_set_md,
 903         .fs_get_md              = fsfilt_ext3_get_md,
 904         .fs_readpage            = fsfilt_ext3_readpage,
 905         .fs_add_journal_cb      = fsfilt_ext3_add_journal_cb,
 906         .fs_statfs              = fsfilt_ext3_statfs,
 907         .fs_sync                = fsfilt_ext3_sync,
 908         .fs_map_inode_page      = fsfilt_ext3_map_inode_page,
 909         .fs_prep_san_write      = fsfilt_ext3_prep_san_write,
 910         .fs_write_record        = fsfilt_ext3_write_record,
 911         .fs_read_record         = fsfilt_ext3_read_record,
 912         .fs_setup               = fsfilt_ext3_setup,
 913         .fs_get_op_len          = fsfilt_ext3_get_op_len,
 914 };
 915
 916 static int __init fsfilt_ext3_init(void)
 917 {
 918         int rc;
 919
 920         //rc = ext3_xattr_register();
 921         fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
 922                                       sizeof(struct fsfilt_cb_data), 0,
 923                                       0, NULL, NULL);
 924         if (!fcb_cache) {
 925                 CERROR("error allocating fsfilt journal callback cache\n");
 926                 GOTO(out, rc = -ENOMEM);
 927         }
 928
 929         rc = fsfilt_register_ops(&fsfilt_ext3_ops);
 930
 931         if (rc)
 932                 kmem_cache_destroy(fcb_cache);
 933 out:
 934         return rc;
 935 }
 936
 937 static void __exit fsfilt_ext3_exit(void)
 938 {
 939         fsfilt_unregister_ops(&fsfilt_ext3_ops);
 940         LASSERTF(kmem_cache_destroy(fcb_cache) == 0,
 941                  "can't free fsfilt callback cache: count %d\n",
 942                  atomic_read(&fcb_cache_count));
 943
 944         //rc = ext3_xattr_unregister();
 945 }
 946
 947 module_init(fsfilt_ext3_init);
 948 module_exit(fsfilt_ext3_exit);
 949
 950 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
 951 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
 952 MODULE_LICENSE("GPL");