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 /*
  66  * We don't currently need any additional blocks for rmdir and
  67  * unlink transactions because we are storing the OST oa_id inside
  68  * the inode (which we will be changing anyways as part of this
  69  * transaction).
  70  */
  71 static void *fsfilt_ext3_start(struct inode *inode, int op, void *desc_private,
  72                                int logs)
  73 {
  74         /* For updates to the last recieved file */
  75         int nblocks = EXT3_SINGLEDATA_TRANS_BLOCKS;
  76         journal_t *journal;
  77         void *handle;
  78
  79         if (current->journal_info) {
  80                 CDEBUG(D_INODE, "increasing refcount on %p\n",
  81                        current->journal_info);
  82                 goto journal_start;
  83         }
  84
  85         switch(op) {
  86         case FSFILT_OP_RMDIR:
  87         case FSFILT_OP_UNLINK:
  88                 /* delete one file + create/update logs for each stripe */
  89                 nblocks += EXT3_DELETE_TRANS_BLOCKS;
  90                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
  91                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
  92                 break;
  93         case FSFILT_OP_RENAME:
  94                 /* modify additional directory */
  95                 nblocks += EXT3_SINGLEDATA_TRANS_BLOCKS;
  96                 /* no break */
  97         case FSFILT_OP_SYMLINK:
  98                 /* additional block + block bitmap + GDT for long symlink */
  99                 nblocks += 3;
 100                 /* no break */
 101         case FSFILT_OP_CREATE:
 102                 /* create/update logs for each stripe */
 103                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 104                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
 105                 /* no break */
 106         case FSFILT_OP_MKDIR:
 107         case FSFILT_OP_MKNOD:
 108                 /* modify one inode + block bitmap + GDT */
 109                 nblocks += 3;
 110                 /* no break */
 111         case FSFILT_OP_LINK:
 112                 /* modify parent directory */
 113                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 114                         EXT3_DATA_TRANS_BLOCKS;
 115                 break;
 116         case FSFILT_OP_SETATTR:
 117                 /* Setattr on inode */
 118                 nblocks += 1;
 119                 break;
 120         case FSFILT_OP_CANCEL_UNLINK:
 121                 /* blocks for log header bitmap update OR
 122                  * blocks for catalog header bitmap update + unlink of logs */
 123                 nblocks = (LLOG_CHUNK_SIZE >> inode->i_blkbits) +
 124                         EXT3_DELETE_TRANS_BLOCKS * logs;
 125                 break;
 126         default: CERROR("unknown transaction start op %d\n", op);
 127                  LBUG();
 128         }
 129
 130         LASSERT(current->journal_info == desc_private);
 131         journal = EXT3_SB(inode->i_sb)->s_journal;
 132         if (nblocks > journal->j_max_transaction_buffers) {
 133                 CERROR("too many credits %d for op %ux%u using %d instead\n",
 134                        nblocks, op, logs, journal->j_max_transaction_buffers);
 135                 nblocks = journal->j_max_transaction_buffers;
 136         }
 137
 138  journal_start:
 139         lock_kernel();
 140         handle = journal_start(EXT3_JOURNAL(inode), nblocks);
 141         unlock_kernel();
 142
 143         if (!IS_ERR(handle))
 144                 LASSERT(current->journal_info == handle);
 145         else
 146                 CERROR("error starting handle for op %u (%u credits): rc %ld\n",
 147                        op, nblocks, PTR_ERR(handle));
 148         return handle;
 149 }
 150
 151 /*
 152  * Calculate the number of buffer credits needed to write multiple pages in
 153  * a single ext3 transaction.  No, this shouldn't be here, but as yet ext3
 154  * doesn't have a nice API for calculating this sort of thing in advance.
 155  *
 156  * See comment above ext3_writepage_trans_blocks for details.  We assume
 157  * no data journaling is being done, but it does allow for all of the pages
 158  * being non-contiguous.  If we are guaranteed contiguous pages we could
 159  * reduce the number of (d)indirect blocks a lot.
 160  *
 161  * With N blocks per page and P pages, for each inode we have at most:
 162  * N*P indirect
 163  * min(N*P, blocksize/4 + 1) dindirect blocks
 164  * niocount tindirect
 165  *
 166  * For the entire filesystem, we have at most:
 167  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 168  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 169  * objcount inode blocks
 170  * 1 superblock
 171  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 172  *
 173  * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
 174  */
 175 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso,
 176                                       int niocount, struct niobuf_local *nb)
 177 {
 178         struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
 179         __u64 next_indir;
 180         const int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 181         int nbitmaps = 0, ngdblocks;
 182         int needed = objcount + 1; /* inodes + superblock */
 183         int i, j;
 184
 185         for (i = 0, j = 0; i < objcount; i++, fso++) {
 186                 /* two or more dindirect blocks in case we cross boundary */
 187                 int ndind = (long)((nb[j + fso->fso_bufcnt - 1].offset -
 188                                     nb[j].offset) >>
 189                                    sb->s_blocksize_bits) /
 190                         (EXT3_ADDR_PER_BLOCK(sb) * EXT3_ADDR_PER_BLOCK(sb));
 191                 nbitmaps += min(fso->fso_bufcnt, ndind > 0 ? ndind : 2);
 192
 193                 /* leaf, indirect, tindirect blocks for first block */
 194                 nbitmaps += blockpp + 2;
 195
 196                 j += fso->fso_bufcnt;
 197         }
 198
 199         next_indir = nb[0].offset +
 200                 (EXT3_ADDR_PER_BLOCK(sb) << sb->s_blocksize_bits);
 201         for (i = 1; i < niocount; i++) {
 202                 if (nb[i].offset >= next_indir) {
 203                         nbitmaps++;     /* additional indirect */
 204                         next_indir = nb[i].offset +
 205                                 (EXT3_ADDR_PER_BLOCK(sb)<<sb->s_blocksize_bits);
 206                 } else if (nb[i].offset != nb[i - 1].offset + sb->s_blocksize) {
 207                         nbitmaps++;     /* additional indirect */
 208                 }
 209                 nbitmaps += blockpp;    /* each leaf in different group? */
 210         }
 211
 212         ngdblocks = nbitmaps;
 213         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 214                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 215         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 216                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 217
 218         needed += nbitmaps + ngdblocks;
 219
 220         /* last_rcvd update */
 221         needed += EXT3_DATA_TRANS_BLOCKS;
 222
 223 #if defined(CONFIG_QUOTA) && !defined(__x86_64__) /* XXX */
 224         /* We assume that there will be 1 bit set in s_dquot.flags for each
 225          * quota file that is active.  This is at least true for now.
 226          */
 227         needed += hweight32(sb_any_quota_enabled(sb)) *
 228                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 229 #endif
 230
 231         return needed;
 232 }
 233
 234 /* We have to start a huge journal transaction here to hold all of the
 235  * metadata for the pages being written here.  This is necessitated by
 236  * the fact that we do lots of prepare_write operations before we do
 237  * any of the matching commit_write operations, so even if we split
 238  * up to use "smaller" transactions none of them could complete until
 239  * all of them were opened.  By having a single journal transaction,
 240  * we eliminate duplicate reservations for common blocks like the
 241  * superblock and group descriptors or bitmaps.
 242  *
 243  * We will start the transaction here, but each prepare_write will
 244  * add a refcount to the transaction, and each commit_write will
 245  * remove a refcount.  The transaction will be closed when all of
 246  * the pages have been written.
 247  */
 248 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
 249                                    int niocount, struct niobuf_local *nb,
 250                                    void *desc_private, int logs)
 251 {
 252         journal_t *journal;
 253         handle_t *handle;
 254         int needed;
 255         ENTRY;
 256
 257         LASSERT(current->journal_info == desc_private);
 258         journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
 259         needed = fsfilt_ext3_credits_needed(objcount, fso, niocount, nb);
 260
 261         /* The number of blocks we could _possibly_ dirty can very large.
 262          * We reduce our request if it is absurd (and we couldn't get that
 263          * many credits for a single handle anyways).
 264          *
 265          * At some point we have to limit the size of I/Os sent at one time,
 266          * increase the size of the journal, or we have to calculate the
 267          * actual journal requirements more carefully by checking all of
 268          * the blocks instead of being maximally pessimistic.  It remains to
 269          * be seen if this is a real problem or not.
 270          */
 271         if (needed > journal->j_max_transaction_buffers) {
 272                 CERROR("want too many journal credits (%d) using %d instead\n",
 273                        needed, journal->j_max_transaction_buffers);
 274                 needed = journal->j_max_transaction_buffers;
 275         }
 276
 277         lock_kernel();
 278         handle = journal_start(journal, needed);
 279         unlock_kernel();
 280         if (IS_ERR(handle)) {
 281                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 282                        PTR_ERR(handle));
 283         } else {
 284                 LASSERT(handle->h_buffer_credits >= needed);
 285                 LASSERT(current->journal_info == handle);
 286         }
 287
 288         RETURN(handle);
 289 }
 290
 291 static int fsfilt_ext3_commit(struct inode *inode, void *h, int force_sync)
 292 {
 293         int rc;
 294         handle_t *handle = h;
 295
 296         LASSERT(current->journal_info == handle);
 297         if (force_sync)
 298                 handle->h_sync = 1; /* recovery likes this */
 299
 300         lock_kernel();
 301         rc = journal_stop(handle);
 302         unlock_kernel();
 303
 304         return rc;
 305 }
 306
 307 static int fsfilt_ext3_commit_async(struct inode *inode, void *h,
 308                                     void **wait_handle)
 309 {
 310         unsigned long tid;
 311         transaction_t *transaction;
 312 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 313         unsigned long rtid;
 314 #endif
 315         handle_t *handle = h;
 316         journal_t *journal;
 317         int rc;
 318
 319         LASSERT(current->journal_info == handle);
 320
 321         lock_kernel();
 322         transaction = handle->h_transaction;
 323         journal = transaction->t_journal;
 324         tid = transaction->t_tid;
 325         /* we don't want to be blocked */
 326         handle->h_sync = 0;
 327         rc = journal_stop(handle);
 328         if (rc) {
 329                 CERROR("error while stopping transaction: %d\n", rc);
 330                 unlock_kernel();
 331                 return rc;
 332         }
 333 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 334         rtid = log_start_commit(journal, transaction);
 335         if (rtid != tid)
 336                 CERROR("strange race: %lu != %lu\n",
 337                        (unsigned long) tid, (unsigned long) rtid);
 338 #else
 339         log_start_commit(journal, transaction->t_tid);
 340 #endif
 341         unlock_kernel();
 342
 343         *wait_handle = (void *) tid;
 344         CDEBUG(D_INODE, "commit async: %lu\n", (unsigned long) tid);
 345         return 0;
 346 }
 347
 348 static int fsfilt_ext3_commit_wait(struct inode *inode, void *h)
 349 {
 350         tid_t tid = (tid_t)(long)h;
 351
 352         CDEBUG(D_INODE, "commit wait: %lu\n", (unsigned long) tid);
 353         if (is_journal_aborted(EXT3_JOURNAL(inode)))
 354                 return -EIO;
 355
 356         log_wait_commit(EXT3_JOURNAL(inode), tid);
 357
 358         return 0;
 359 }
 360
 361 static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
 362                                struct iattr *iattr, int do_trunc)
 363 {
 364         struct inode *inode = dentry->d_inode;
 365         int rc;
 366
 367         lock_kernel();
 368
 369         /* A _really_ horrible hack to avoid removing the data stored
 370          * in the block pointers; this is really the "small" stripe MD data.
 371          * We can avoid further hackery by virtue of the MDS file size being
 372          * zero all the time (which doesn't invoke block truncate at unlink
 373          * time), so we assert we never change the MDS file size from zero. */
 374         if (iattr->ia_valid & ATTR_SIZE && !do_trunc) {
 375                 /* ATTR_SIZE would invoke truncate: clear it */
 376                 iattr->ia_valid &= ~ATTR_SIZE;
 377                 EXT3_I(inode)->i_disksize = inode->i_size = iattr->ia_size;
 378
 379                 /* make sure _something_ gets set - so new inode
 380                  * goes to disk (probably won't work over XFS */
 381                 if (!(iattr->ia_valid & (ATTR_MODE | ATTR_MTIME | ATTR_CTIME))){
 382                         iattr->ia_valid |= ATTR_MODE;
 383                         iattr->ia_mode = inode->i_mode;
 384                 }
 385         }
 386
 387         /* Don't allow setattr to change file type */
 388         iattr->ia_mode = (inode->i_mode & S_IFMT)|(iattr->ia_mode & ~S_IFMT);
 389
 390         /* We set these flags on the client, but have already checked perms
 391          * so don't confuse inode_change_ok. */
 392         iattr->ia_valid &= ~(ATTR_MTIME_SET | ATTR_ATIME_SET);
 393
 394         if (inode->i_op->setattr) {
 395                 rc = inode->i_op->setattr(dentry, iattr);
 396         } else {
 397                 rc = inode_change_ok(inode, iattr);
 398                 if (!rc)
 399                         rc = inode_setattr(inode, iattr);
 400         }
 401
 402         unlock_kernel();
 403
 404         return rc;
 405 }
 406
 407 static int fsfilt_ext3_iocontrol(struct inode * inode, struct file *file,
 408                                  unsigned int cmd, unsigned long arg)
 409 {
 410         int rc = 0;
 411         ENTRY;
 412
 413         if (inode->i_fop->ioctl)
 414                 rc = inode->i_fop->ioctl(inode, file, cmd, arg);
 415         else
 416                 RETURN(-ENOTTY);
 417
 418         RETURN(rc);
 419 }
 420
 421 static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
 422                               void *lmm, int lmm_size)
 423 {
 424         int rc;
 425
 426         /* keep this when we get rid of OLD_EA (too noisy during conversion) */
 427         if (EXT3_I(inode)->i_file_acl /* || large inode EA flag */)
 428                 CWARN("setting EA on %lu/%u again... interesting\n",
 429                        inode->i_ino, inode->i_generation);
 430
 431         lock_kernel();
 432         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
 433                                    XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size, 0);
 434
 435         unlock_kernel();
 436
 437         if (rc)
 438                 CERROR("error adding MD data to inode %lu: rc = %d\n",
 439                        inode->i_ino, rc);
 440         return rc;
 441 }
 442
 443 /* Must be called with i_sem held */
 444 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm, int lmm_size)
 445 {
 446         int rc;
 447
 448         LASSERT(down_trylock(&inode->i_sem) != 0);
 449         lock_kernel();
 450
 451         rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
 452                             XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size);
 453         unlock_kernel();
 454
 455         /* This gives us the MD size */
 456         if (lmm == NULL)
 457                 return (rc == -ENODATA) ? 0 : rc;
 458
 459         if (rc < 0) {
 460                 CDEBUG(D_INFO, "error getting EA %d/%s from inode %lu: rc %d\n",
 461                        EXT3_XATTR_INDEX_TRUSTED, XATTR_LUSTRE_MDS_LOV_EA,
 462                        inode->i_ino, rc);
 463                 memset(lmm, 0, lmm_size);
 464                 return (rc == -ENODATA) ? 0 : rc;
 465         }
 466
 467         return rc;
 468 }
 469
 470 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
 471                                     loff_t *off)
 472 {
 473         struct inode *inode = file->f_dentry->d_inode;
 474         int rc = 0;
 475
 476         if (S_ISREG(inode->i_mode))
 477                 rc = file->f_op->read(file, buf, count, off);
 478         else {
 479                 const int blkbits = inode->i_sb->s_blocksize_bits;
 480                 const int blksize = inode->i_sb->s_blocksize;
 481
 482                 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
 483                        count, inode->i_ino, *off);
 484                 while (count > 0) {
 485                         struct buffer_head *bh;
 486
 487                         bh = NULL;
 488                         if (*off < inode->i_size) {
 489                                 int err = 0;
 490
 491                                 bh = ext3_bread(NULL, inode, *off >> blkbits,
 492                                                 0, &err);
 493
 494                                 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
 495
 496                                 if (bh) {
 497                                         memcpy(buf, bh->b_data, blksize);
 498                                         brelse(bh);
 499                                 } else if (err) {
 500                                         /* XXX in theory we should just fake
 501                                          * this buffer and continue like ext3,
 502                                          * especially if this is a partial read
 503                                          */
 504                                         CERROR("error read dir %lu+%llu: %d\n",
 505                                                inode->i_ino, *off, err);
 506                                         RETURN(err);
 507                                 }
 508                         }
 509                         if (!bh) {
 510                                 struct ext3_dir_entry_2 *fake = (void *)buf;
 511
 512                                 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
 513                                 memset(fake, 0, sizeof(*fake));
 514                                 fake->rec_len = cpu_to_le32(blksize);
 515                         }
 516                         count -= blksize;
 517                         buf += blksize;
 518                         *off += blksize;
 519                         rc += blksize;
 520                 }
 521         }
 522
 523         return rc;
 524 }
 525
 526 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
 527 {
 528         struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
 529
 530         fcb->cb_func(fcb->cb_obd, fcb->cb_last_rcvd, fcb->cb_data, error);
 531
 532         OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
 533         atomic_dec(&fcb_cache_count);
 534 }
 535
 536 static int fsfilt_ext3_add_journal_cb(struct obd_device *obd, __u64 last_rcvd,
 537                                       void *handle, fsfilt_cb_t cb_func,
 538                                       void *cb_data)
 539 {
 540         struct fsfilt_cb_data *fcb;
 541
 542         OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
 543         if (fcb == NULL)
 544                 RETURN(-ENOMEM);
 545
 546         atomic_inc(&fcb_cache_count);
 547         fcb->cb_func = cb_func;
 548         fcb->cb_obd = obd;
 549         fcb->cb_last_rcvd = last_rcvd;
 550         fcb->cb_data = cb_data;
 551
 552         CDEBUG(D_EXT2, "set callback for last_rcvd: "LPD64"\n", last_rcvd);
 553         lock_kernel();
 554         journal_callback_set(handle, fsfilt_ext3_cb_func,
 555                              (struct journal_callback *)fcb);
 556         unlock_kernel();
 557
 558         return 0;
 559 }
 560
 561 /*
 562  * We need to hack the return value for the free inode counts because
 563  * the current EA code requires one filesystem block per inode with EAs,
 564  * so it is possible to run out of blocks before we run out of inodes.
 565  *
 566  * This can be removed when the ext3 EA code is fixed.
 567  */
 568 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
 569 {
 570         struct kstatfs sfs;
 571         int rc;
 572
 573         memset(&sfs, 0, sizeof(sfs));
 574
 575         rc = sb->s_op->statfs(sb, &sfs);
 576
 577         if (!rc && sfs.f_bfree < sfs.f_ffree) {
 578                 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
 579                 sfs.f_ffree = sfs.f_bfree;
 580         }
 581
 582         statfs_pack(osfs, &sfs);
 583         return rc;
 584 }
 585
 586 static int fsfilt_ext3_sync(struct super_block *sb)
 587 {
 588         return ext3_force_commit(sb);
 589 }
 590
 591 extern int ext3_map_inode_page(struct inode *inode, struct page *page,
 592                                unsigned long *blocks, int *created, int create);
 593 int fsfilt_ext3_map_inode_page(struct inode *inode, struct page *page,
 594                                unsigned long *blocks, int *created, int create)
 595 {
 596         return ext3_map_inode_page(inode, page, blocks, created, create);
 597 }
 598
 599 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
 600                                int nblocks, loff_t newsize);
 601 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
 602                                       int nblocks, loff_t newsize)
 603 {
 604         return ext3_prep_san_write(inode, blocks, nblocks, newsize);
 605 }
 606
 607 static int fsfilt_ext3_read_record(struct file * file, void *buf,
 608                                    int size, loff_t *offs)
 609 {
 610         struct inode *inode = file->f_dentry->d_inode;
 611         unsigned long block;
 612         struct buffer_head *bh;
 613         int err, blocksize, csize, boffs;
 614
 615         /* prevent reading after eof */
 616         lock_kernel();
 617         if (inode->i_size < *offs + size) {
 618                 size = inode->i_size - *offs;
 619                 unlock_kernel();
 620                 if (size < 0) {
 621                         CERROR("size %llu is too short for read %u@%llu\n",
 622                                inode->i_size, size, *offs);
 623                         return -EIO;
 624                 } else if (size == 0) {
 625                         return 0;
 626                 }
 627         } else {
 628                 unlock_kernel();
 629         }
 630
 631         blocksize = 1 << inode->i_blkbits;
 632
 633         while (size > 0) {
 634                 block = *offs >> inode->i_blkbits;
 635                 boffs = *offs & (blocksize - 1);
 636                 csize = min(blocksize - boffs, size);
 637                 bh = ext3_bread(NULL, inode, block, 0, &err);
 638                 if (!bh) {
 639                         CERROR("can't read block: %d\n", err);
 640                         return err;
 641                 }
 642
 643                 memcpy(buf, bh->b_data + boffs, csize);
 644                 brelse(bh);
 645
 646                 *offs += csize;
 647                 buf += csize;
 648                 size -= csize;
 649         }
 650         return 0;
 651 }
 652
 653 static int fsfilt_ext3_write_record(struct file *file, void *buf, int bufsize,
 654                                     loff_t *offs, int force_sync)
 655 {
 656         struct buffer_head *bh = NULL;
 657         unsigned long block;
 658         struct inode *inode = file->f_dentry->d_inode;
 659         loff_t old_size = inode->i_size, offset = *offs;
 660         loff_t new_size = inode->i_size;
 661         journal_t *journal;
 662         handle_t *handle;
 663         int err, block_count = 0, blocksize, size, boffs;
 664
 665         /* Determine how many transaction credits are needed */
 666         blocksize = 1 << inode->i_blkbits;
 667         block_count = (*offs & (blocksize - 1)) + bufsize;
 668         block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
 669
 670         journal = EXT3_SB(inode->i_sb)->s_journal;
 671         lock_kernel();
 672         handle = journal_start(journal,
 673                                block_count * EXT3_DATA_TRANS_BLOCKS + 2);
 674         unlock_kernel();
 675         if (IS_ERR(handle)) {
 676                 CERROR("can't start transaction\n");
 677                 return PTR_ERR(handle);
 678         }
 679
 680         while (bufsize > 0) {
 681                 if (bh != NULL)
 682                         brelse(bh);
 683
 684                 block = offset >> inode->i_blkbits;
 685                 boffs = offset & (blocksize - 1);
 686                 size = min(blocksize - boffs, bufsize);
 687                 bh = ext3_bread(handle, inode, block, 1, &err);
 688                 if (!bh) {
 689                         CERROR("can't read/create block: %d\n", err);
 690                         goto out;
 691                 }
 692
 693                 err = ext3_journal_get_write_access(handle, bh);
 694                 if (err) {
 695                         CERROR("journal_get_write_access() returned error %d\n",
 696                                err);
 697                         goto out;
 698                 }
 699                 LASSERT(bh->b_data + boffs + size <= bh->b_data + bh->b_size);
 700                 memcpy(bh->b_data + boffs, buf, size);
 701                 err = ext3_journal_dirty_metadata(handle, bh);
 702                 if (err) {
 703                         CERROR("journal_dirty_metadata() returned error %d\n",
 704                                err);
 705                         goto out;
 706                 }
 707                 if (offset + size > new_size)
 708                         new_size = offset + size;
 709                 offset += size;
 710                 bufsize -= size;
 711                 buf += size;
 712         }
 713
 714         if (force_sync)
 715                 handle->h_sync = 1; /* recovery likes this */
 716 out:
 717         if (bh)
 718                 brelse(bh);
 719
 720         /* correct in-core and on-disk sizes */
 721         if (new_size > inode->i_size) {
 722                 lock_kernel();
 723                 if (new_size > inode->i_size)
 724                         inode->i_size = new_size;
 725                 if (inode->i_size > EXT3_I(inode)->i_disksize)
 726                         EXT3_I(inode)->i_disksize = inode->i_size;
 727                 if (inode->i_size > old_size)
 728                         mark_inode_dirty(inode);
 729                 unlock_kernel();
 730         }
 731
 732         lock_kernel();
 733         journal_stop(handle);
 734         unlock_kernel();
 735
 736         if (err == 0)
 737                 *offs = offset;
 738         return err;
 739 }
 740
 741 static int fsfilt_ext3_setup(struct super_block *sb)
 742 {
 743 #if 0
 744         EXT3_SB(sb)->dx_lock = fsfilt_ext3_dx_lock;
 745         EXT3_SB(sb)->dx_unlock = fsfilt_ext3_dx_unlock;
 746 #endif
 747 #ifdef S_PDIROPS
 748         CWARN("Enabling PDIROPS\n");
 749         set_opt(EXT3_SB(sb)->s_mount_opt, PDIROPS);
 750         sb->s_flags |= S_PDIROPS;
 751 #endif
 752         return 0;
 753 }
 754
 755 /* If fso is NULL, op is FSFILT operation, otherwise op is number of fso
 756    objects. Logs is number of logfiles to update */
 757 static int fsfilt_ext3_get_op_len(int op, struct fsfilt_objinfo *fso, int logs)
 758 {
 759         if ( !fso ) {
 760                 switch(op) {
 761                 case FSFILT_OP_CREATE:
 762                                  /* directory leaf, index & indirect & EA*/
 763                         return 4 + 3 * logs;
 764                 case FSFILT_OP_UNLINK:
 765                         return 3 * logs;
 766                 }
 767         } else {
 768                 int i;
 769                 int needed = 0;
 770                 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
 771                 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 772                 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
 773                 for (i = 0; i < op; i++, fso++) {
 774                         int nblocks = fso->fso_bufcnt * blockpp;
 775                         int ndindirect = min(nblocks, addrpp + 1);
 776                         int nindir = nblocks + ndindirect + 1;
 777
 778                         needed += nindir;
 779                 }
 780                 return needed + 3 * logs;
 781         }
 782
 783         return 0;
 784 }
 785
 786 static struct fsfilt_operations fsfilt_ext3_ops = {
 787         .fs_type                = "ext3",
 788         .fs_owner               = THIS_MODULE,
 789         .fs_start               = fsfilt_ext3_start,
 790         .fs_brw_start           = fsfilt_ext3_brw_start,
 791         .fs_commit              = fsfilt_ext3_commit,
 792         .fs_commit_async        = fsfilt_ext3_commit_async,
 793         .fs_commit_wait         = fsfilt_ext3_commit_wait,
 794         .fs_setattr             = fsfilt_ext3_setattr,
 795         .fs_iocontrol           = fsfilt_ext3_iocontrol,
 796         .fs_set_md              = fsfilt_ext3_set_md,
 797         .fs_get_md              = fsfilt_ext3_get_md,
 798         .fs_readpage            = fsfilt_ext3_readpage,
 799         .fs_add_journal_cb      = fsfilt_ext3_add_journal_cb,
 800         .fs_statfs              = fsfilt_ext3_statfs,
 801         .fs_sync                = fsfilt_ext3_sync,
 802         .fs_map_inode_page      = fsfilt_ext3_map_inode_page,
 803         .fs_prep_san_write      = fsfilt_ext3_prep_san_write,
 804         .fs_write_record        = fsfilt_ext3_write_record,
 805         .fs_read_record         = fsfilt_ext3_read_record,
 806         .fs_setup               = fsfilt_ext3_setup,
 807         .fs_get_op_len          = fsfilt_ext3_get_op_len,
 808 };
 809
 810 static int __init fsfilt_ext3_init(void)
 811 {
 812         int rc;
 813
 814         //rc = ext3_xattr_register();
 815         fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
 816                                       sizeof(struct fsfilt_cb_data), 0,
 817                                       0, NULL, NULL);
 818         if (!fcb_cache) {
 819                 CERROR("error allocating fsfilt journal callback cache\n");
 820                 GOTO(out, rc = -ENOMEM);
 821         }
 822
 823         rc = fsfilt_register_ops(&fsfilt_ext3_ops);
 824
 825         if (rc)
 826                 kmem_cache_destroy(fcb_cache);
 827 out:
 828         return rc;
 829 }
 830
 831 static void __exit fsfilt_ext3_exit(void)
 832 {
 833         int rc;
 834
 835         fsfilt_unregister_ops(&fsfilt_ext3_ops);
 836         rc = kmem_cache_destroy(fcb_cache);
 837
 838         if (rc || atomic_read(&fcb_cache_count)) {
 839                 CERROR("can't free fsfilt callback cache: count %d, rc = %d\n",
 840                        atomic_read(&fcb_cache_count), rc);
 841         }
 842
 843         //rc = ext3_xattr_unregister();
 844 }
 845
 846 module_init(fsfilt_ext3_init);
 847 module_exit(fsfilt_ext3_exit);
 848
 849 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
 850 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
 851 MODULE_LICENSE("GPL");