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 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
  49 #include <linux/module.h>
  50 #include <linux/iobuf.h>
  51 #endif
  52
  53 static kmem_cache_t *fcb_cache;
  54 static atomic_t fcb_cache_count = ATOMIC_INIT(0);
  55
  56 struct fsfilt_cb_data {
  57         struct journal_callback cb_jcb; /* jbd private data - MUST BE FIRST */
  58         fsfilt_cb_t cb_func;            /* MDS/OBD completion function */
  59         struct obd_device *cb_obd;      /* MDS/OBD completion device */
  60         __u64 cb_last_rcvd;             /* MDS/OST last committed operation */
  61         void *cb_data;                  /* MDS/OST completion function data */
  62 };
  63
  64 #ifndef EXT3_XATTR_INDEX_TRUSTED        /* temporary until we hit l28 kernel */
  65 #define EXT3_XATTR_INDEX_TRUSTED        4
  66 #endif
  67 #define XATTR_LUSTRE_MDS_LOV_EA         "lov"
  68
  69 /*
  70  * We don't currently need any additional blocks for rmdir and
  71  * unlink transactions because we are storing the OST oa_id inside
  72  * the inode (which we will be changing anyways as part of this
  73  * transaction).
  74  */
  75 static void *fsfilt_ext3_start(struct inode *inode, int op, void *desc_private,
  76                                int logs)
  77 {
  78         /* For updates to the last recieved file */
  79         int nblocks = EXT3_SINGLEDATA_TRANS_BLOCKS;
  80         journal_t *journal;
  81         void *handle;
  82
  83         if (current->journal_info) {
  84                 CDEBUG(D_INODE, "increasing refcount on %p\n",
  85                        current->journal_info);
  86                 goto journal_start;
  87         }
  88
  89         /* XXX BUG 3188 -- must return to one set of opcodes */
  90         /* FIXME - cache hook */
  91         if (op & 0x20) {
  92                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
  93                 op = op & ~0x20;
  94         }
  95
  96         /* FIXME - kml */
  97         if (op & 0x10) {
  98                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
  99                 op = op & ~0x10;
 100         }
 101
 102         switch(op) {
 103         case FSFILT_OP_RMDIR:
 104         case FSFILT_OP_UNLINK:
 105                 /* delete one file + create/update logs for each stripe */
 106                 nblocks += EXT3_DELETE_TRANS_BLOCKS;
 107                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 108                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
 109                 break;
 110         case FSFILT_OP_RENAME:
 111                 /* modify additional directory */
 112                 nblocks += EXT3_SINGLEDATA_TRANS_BLOCKS;
 113                 /* no break */
 114         case FSFILT_OP_SYMLINK:
 115                 /* additional block + block bitmap + GDT for long symlink */
 116                 nblocks += 3;
 117                 /* no break */
 118         case FSFILT_OP_CREATE:
 119                 /* create/update logs for each stripe */
 120                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 121                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
 122                 /* no break */
 123         case FSFILT_OP_MKDIR:
 124         case FSFILT_OP_MKNOD:
 125                 /* modify one inode + block bitmap + GDT */
 126                 nblocks += 3;
 127                 /* no break */
 128         case FSFILT_OP_LINK:
 129                 /* modify parent directory */
 130                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 131                         EXT3_DATA_TRANS_BLOCKS;
 132                 break;
 133         case FSFILT_OP_SETATTR:
 134                 /* Setattr on inode */
 135                 nblocks += 1;
 136                 break;
 137         case FSFILT_OP_CANCEL_UNLINK:
 138                 /* blocks for log header bitmap update OR
 139                  * blocks for catalog header bitmap update + unlink of logs */
 140                 nblocks = (LLOG_CHUNK_SIZE >> inode->i_blkbits) +
 141                         EXT3_DELETE_TRANS_BLOCKS * logs;
 142                 break;
 143         case FSFILT_OP_NOOP:
 144                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
 145                 break;
 146         default: CERROR("unknown transaction start op %d\n", op);
 147                 LBUG();
 148         }
 149
 150         LASSERT(current->journal_info == desc_private);
 151         journal = EXT3_SB(inode->i_sb)->s_journal;
 152         if (nblocks > journal->j_max_transaction_buffers) {
 153                 CERROR("too many credits %d for op %ux%u using %d instead\n",
 154                        nblocks, op, logs, journal->j_max_transaction_buffers);
 155                 nblocks = journal->j_max_transaction_buffers;
 156         }
 157
 158  journal_start:
 159         lock_kernel();
 160         handle = journal_start(EXT3_JOURNAL(inode), nblocks);
 161         unlock_kernel();
 162
 163         if (!IS_ERR(handle))
 164                 LASSERT(current->journal_info == handle);
 165         else
 166                 CERROR("error starting handle for op %u (%u credits): rc %ld\n",
 167                        op, nblocks, PTR_ERR(handle));
 168         return handle;
 169 }
 170
 171 /*
 172  * Calculate the number of buffer credits needed to write multiple pages in
 173  * a single ext3 transaction.  No, this shouldn't be here, but as yet ext3
 174  * doesn't have a nice API for calculating this sort of thing in advance.
 175  *
 176  * See comment above ext3_writepage_trans_blocks for details.  We assume
 177  * no data journaling is being done, but it does allow for all of the pages
 178  * being non-contiguous.  If we are guaranteed contiguous pages we could
 179  * reduce the number of (d)indirect blocks a lot.
 180  *
 181  * With N blocks per page and P pages, for each inode we have at most:
 182  * N*P indirect
 183  * min(N*P, blocksize/4 + 1) dindirect blocks
 184  * niocount tindirect
 185  *
 186  * For the entire filesystem, we have at most:
 187  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 188  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 189  * objcount inode blocks
 190  * 1 superblock
 191  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 192  *
 193  * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
 194  */
 195 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso,
 196                                       int niocount, struct niobuf_local *nb)
 197 {
 198         struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
 199         __u64 next_indir;
 200         const int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 201         int nbitmaps = 0, ngdblocks;
 202         int needed = objcount + 1; /* inodes + superblock */
 203         int i, j;
 204
 205         for (i = 0, j = 0; i < objcount; i++, fso++) {
 206                 /* two or more dindirect blocks in case we cross boundary */
 207                 int ndind = (long)((nb[j + fso->fso_bufcnt - 1].offset -
 208                                     nb[j].offset) >>
 209                                    sb->s_blocksize_bits) /
 210                         (EXT3_ADDR_PER_BLOCK(sb) * EXT3_ADDR_PER_BLOCK(sb));
 211                 nbitmaps += min(fso->fso_bufcnt, ndind > 0 ? ndind : 2);
 212
 213                 /* leaf, indirect, tindirect blocks for first block */
 214                 nbitmaps += blockpp + 2;
 215
 216                 j += fso->fso_bufcnt;
 217         }
 218
 219         next_indir = nb[0].offset +
 220                 (EXT3_ADDR_PER_BLOCK(sb) << sb->s_blocksize_bits);
 221         for (i = 1; i < niocount; i++) {
 222                 if (nb[i].offset >= next_indir) {
 223                         nbitmaps++;     /* additional indirect */
 224                         next_indir = nb[i].offset +
 225                                 (EXT3_ADDR_PER_BLOCK(sb)<<sb->s_blocksize_bits);
 226                 } else if (nb[i].offset != nb[i - 1].offset + sb->s_blocksize) {
 227                         nbitmaps++;     /* additional indirect */
 228                 }
 229                 nbitmaps += blockpp;    /* each leaf in different group? */
 230         }
 231
 232         ngdblocks = nbitmaps;
 233         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 234                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 235         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 236                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 237
 238         needed += nbitmaps + ngdblocks;
 239
 240         /* last_rcvd update */
 241         needed += EXT3_DATA_TRANS_BLOCKS;
 242
 243 #if defined(CONFIG_QUOTA) && !defined(__x86_64__) /* XXX */
 244         /* We assume that there will be 1 bit set in s_dquot.flags for each
 245          * quota file that is active.  This is at least true for now.
 246          */
 247         needed += hweight32(sb_any_quota_enabled(sb)) *
 248                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 249 #endif
 250
 251         return needed;
 252 }
 253
 254 /* We have to start a huge journal transaction here to hold all of the
 255  * metadata for the pages being written here.  This is necessitated by
 256  * the fact that we do lots of prepare_write operations before we do
 257  * any of the matching commit_write operations, so even if we split
 258  * up to use "smaller" transactions none of them could complete until
 259  * all of them were opened.  By having a single journal transaction,
 260  * we eliminate duplicate reservations for common blocks like the
 261  * superblock and group descriptors or bitmaps.
 262  *
 263  * We will start the transaction here, but each prepare_write will
 264  * add a refcount to the transaction, and each commit_write will
 265  * remove a refcount.  The transaction will be closed when all of
 266  * the pages have been written.
 267  */
 268 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
 269                                    int niocount, struct niobuf_local *nb,
 270                                    void *desc_private, int logs)
 271 {
 272         journal_t *journal;
 273         handle_t *handle;
 274         int needed;
 275         ENTRY;
 276
 277         LASSERT(current->journal_info == desc_private);
 278         journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
 279         needed = fsfilt_ext3_credits_needed(objcount, fso, niocount, nb);
 280
 281         /* The number of blocks we could _possibly_ dirty can very large.
 282          * We reduce our request if it is absurd (and we couldn't get that
 283          * many credits for a single handle anyways).
 284          *
 285          * At some point we have to limit the size of I/Os sent at one time,
 286          * increase the size of the journal, or we have to calculate the
 287          * actual journal requirements more carefully by checking all of
 288          * the blocks instead of being maximally pessimistic.  It remains to
 289          * be seen if this is a real problem or not.
 290          */
 291         if (needed > journal->j_max_transaction_buffers) {
 292                 CERROR("want too many journal credits (%d) using %d instead\n",
 293                        needed, journal->j_max_transaction_buffers);
 294                 needed = journal->j_max_transaction_buffers;
 295         }
 296
 297         lock_kernel();
 298         handle = journal_start(journal, needed);
 299         unlock_kernel();
 300         if (IS_ERR(handle)) {
 301                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 302                        PTR_ERR(handle));
 303         } else {
 304                 LASSERT(handle->h_buffer_credits >= needed);
 305                 LASSERT(current->journal_info == handle);
 306         }
 307
 308         RETURN(handle);
 309 }
 310
 311 static int fsfilt_ext3_commit(struct inode *inode, void *h, int force_sync)
 312 {
 313         int rc;
 314         handle_t *handle = h;
 315
 316         LASSERT(current->journal_info == handle);
 317         if (force_sync)
 318                 handle->h_sync = 1; /* recovery likes this */
 319
 320         lock_kernel();
 321         rc = journal_stop(handle);
 322         unlock_kernel();
 323
 324         return rc;
 325 }
 326
 327 static int fsfilt_ext3_commit_async(struct inode *inode, void *h,
 328                                     void **wait_handle)
 329 {
 330         unsigned long tid;
 331         transaction_t *transaction;
 332 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 333         unsigned long rtid;
 334 #endif
 335         handle_t *handle = h;
 336         journal_t *journal;
 337         int rc;
 338
 339         LASSERT(current->journal_info == handle);
 340
 341         lock_kernel();
 342         transaction = handle->h_transaction;
 343         journal = transaction->t_journal;
 344         tid = transaction->t_tid;
 345         /* we don't want to be blocked */
 346         handle->h_sync = 0;
 347         rc = journal_stop(handle);
 348         if (rc) {
 349                 CERROR("error while stopping transaction: %d\n", rc);
 350                 unlock_kernel();
 351                 return rc;
 352         }
 353 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 354         rtid = log_start_commit(journal, transaction);
 355         if (rtid != tid)
 356                 CERROR("strange race: %lu != %lu\n",
 357                        (unsigned long) tid, (unsigned long) rtid);
 358 #else
 359         log_start_commit(journal, transaction->t_tid);
 360 #endif
 361         unlock_kernel();
 362
 363         *wait_handle = (void *) tid;
 364         CDEBUG(D_INODE, "commit async: %lu\n", (unsigned long) tid);
 365         return 0;
 366 }
 367
 368 static int fsfilt_ext3_commit_wait(struct inode *inode, void *h)
 369 {
 370         tid_t tid = (tid_t)(long)h;
 371
 372         CDEBUG(D_INODE, "commit wait: %lu\n", (unsigned long) tid);
 373         if (is_journal_aborted(EXT3_JOURNAL(inode)))
 374                 return -EIO;
 375
 376         log_wait_commit(EXT3_JOURNAL(inode), tid);
 377
 378         return 0;
 379 }
 380
 381 static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
 382                                struct iattr *iattr, int do_trunc)
 383 {
 384         struct inode *inode = dentry->d_inode;
 385         int rc;
 386
 387         lock_kernel();
 388
 389         /* A _really_ horrible hack to avoid removing the data stored
 390          * in the block pointers; this is really the "small" stripe MD data.
 391          * We can avoid further hackery by virtue of the MDS file size being
 392          * zero all the time (which doesn't invoke block truncate at unlink
 393          * time), so we assert we never change the MDS file size from zero. */
 394         if (iattr->ia_valid & ATTR_SIZE && !do_trunc) {
 395                 /* ATTR_SIZE would invoke truncate: clear it */
 396                 iattr->ia_valid &= ~ATTR_SIZE;
 397                 EXT3_I(inode)->i_disksize = inode->i_size = iattr->ia_size;
 398
 399                 /* make sure _something_ gets set - so new inode
 400                  * goes to disk (probably won't work over XFS */
 401                 if (!(iattr->ia_valid & (ATTR_MODE | ATTR_MTIME | ATTR_CTIME))){
 402                         iattr->ia_valid |= ATTR_MODE;
 403                         iattr->ia_mode = inode->i_mode;
 404                 }
 405         }
 406
 407         /* Don't allow setattr to change file type */
 408         iattr->ia_mode = (inode->i_mode & S_IFMT)|(iattr->ia_mode & ~S_IFMT);
 409
 410         /* We set these flags on the client, but have already checked perms
 411          * so don't confuse inode_change_ok. */
 412         iattr->ia_valid &= ~(ATTR_MTIME_SET | ATTR_ATIME_SET);
 413
 414         if (inode->i_op->setattr) {
 415                 rc = inode->i_op->setattr(dentry, iattr);
 416         } else {
 417                 rc = inode_change_ok(inode, iattr);
 418                 if (!rc)
 419                         rc = inode_setattr(inode, iattr);
 420         }
 421
 422         unlock_kernel();
 423
 424         return rc;
 425 }
 426
 427 static int fsfilt_ext3_iocontrol(struct inode * inode, struct file *file,
 428                                  unsigned int cmd, unsigned long arg)
 429 {
 430         int rc = 0;
 431         ENTRY;
 432
 433         if (inode->i_fop->ioctl)
 434                 rc = inode->i_fop->ioctl(inode, file, cmd, arg);
 435         else
 436                 RETURN(-ENOTTY);
 437
 438         RETURN(rc);
 439 }
 440
 441 static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
 442                               void *lmm, int lmm_size)
 443 {
 444         int rc;
 445
 446         /* keep this when we get rid of OLD_EA (too noisy during conversion) */
 447         if (EXT3_I(inode)->i_file_acl /* || large inode EA flag */)
 448                 CWARN("setting EA on %lu/%u again... interesting\n",
 449                        inode->i_ino, inode->i_generation);
 450
 451         lock_kernel();
 452         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
 453                                    XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size, 0);
 454
 455         unlock_kernel();
 456
 457         if (rc)
 458                 CERROR("error adding MD data to inode %lu: rc = %d\n",
 459                        inode->i_ino, rc);
 460         return rc;
 461 }
 462
 463 /* Must be called with i_sem held */
 464 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm, int lmm_size)
 465 {
 466         int rc;
 467
 468         LASSERT(down_trylock(&inode->i_sem) != 0);
 469         lock_kernel();
 470
 471         rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
 472                             XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size);
 473         unlock_kernel();
 474
 475         /* This gives us the MD size */
 476         if (lmm == NULL)
 477                 return (rc == -ENODATA) ? 0 : rc;
 478
 479         if (rc < 0) {
 480                 CDEBUG(D_INFO, "error getting EA %d/%s from inode %lu: rc %d\n",
 481                        EXT3_XATTR_INDEX_TRUSTED, XATTR_LUSTRE_MDS_LOV_EA,
 482                        inode->i_ino, rc);
 483                 memset(lmm, 0, lmm_size);
 484                 return (rc == -ENODATA) ? 0 : rc;
 485         }
 486
 487         return rc;
 488 }
 489
 490 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 491 static int fsfilt_ext3_send_bio(struct inode *inode, struct bio *bio)
 492 {
 493         submit_bio(WRITE, bio);
 494         return 0;
 495 }
 496 #else
 497 static int fsfilt_ext3_send_bio(struct inode *inode, struct kiobuf *bio)
 498 {
 499         int rc, blocks_per_page;
 500
 501         rc = brw_kiovec(WRITE, 1, &bio, inode->i_dev,
 502                         bio->blocks, 1 << inode->i_blkbits);
 503
 504         blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 505
 506         if (rc != (1 << inode->i_blkbits) * bio->nr_pages *
 507             blocks_per_page) {
 508                 CERROR("short write?  expected %d, wrote %d\n",
 509                        (1 << inode->i_blkbits) * bio->nr_pages *
 510                        blocks_per_page, rc);
 511         }
 512
 513         return rc;
 514 }
 515 #endif
 516
 517 /* FIXME-UMKA: This should be used in 2.6.x io code later. */
 518 static struct page *fsfilt_ext3_getpage(struct inode *inode, long int index)
 519 {
 520         int rc;
 521         struct page *page;
 522
 523         page = grab_cache_page(inode->i_mapping, index);
 524         if (page == NULL)
 525                 return ERR_PTR(-ENOMEM);
 526
 527         if (PageUptodate(page)) {
 528                 unlock_page(page);
 529                 return page;
 530         }
 531
 532         rc = inode->i_mapping->a_ops->readpage(NULL, page);
 533         if (rc < 0) {
 534                 page_cache_release(page);
 535                 return ERR_PTR(rc);
 536         }
 537
 538         return page;
 539 }
 540
 541 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
 542                                     loff_t *off)
 543 {
 544         struct inode *inode = file->f_dentry->d_inode;
 545         int rc = 0;
 546
 547         if (S_ISREG(inode->i_mode))
 548                 rc = file->f_op->read(file, buf, count, off);
 549         else {
 550                 const int blkbits = inode->i_sb->s_blocksize_bits;
 551                 const int blksize = inode->i_sb->s_blocksize;
 552
 553                 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
 554                        count, inode->i_ino, *off);
 555                 while (count > 0) {
 556                         struct buffer_head *bh;
 557
 558                         bh = NULL;
 559                         if (*off < inode->i_size) {
 560                                 int err = 0;
 561
 562                                 bh = ext3_bread(NULL, inode, *off >> blkbits,
 563                                                 0, &err);
 564
 565                                 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
 566
 567                                 if (bh) {
 568                                         memcpy(buf, bh->b_data, blksize);
 569                                         brelse(bh);
 570                                 } else if (err) {
 571                                         /* XXX in theory we should just fake
 572                                          * this buffer and continue like ext3,
 573                                          * especially if this is a partial read
 574                                          */
 575                                         CERROR("error read dir %lu+%llu: %d\n",
 576                                                inode->i_ino, *off, err);
 577                                         RETURN(err);
 578                                 }
 579                         }
 580                         if (!bh) {
 581                                 struct ext3_dir_entry_2 *fake = (void *)buf;
 582
 583                                 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
 584                                 memset(fake, 0, sizeof(*fake));
 585                                 fake->rec_len = cpu_to_le32(blksize);
 586                         }
 587                         count -= blksize;
 588                         buf += blksize;
 589                         *off += blksize;
 590                         rc += blksize;
 591                 }
 592         }
 593
 594         return rc;
 595 }
 596
 597 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
 598 {
 599         struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
 600
 601         fcb->cb_func(fcb->cb_obd, fcb->cb_last_rcvd, fcb->cb_data, error);
 602
 603         OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
 604         atomic_dec(&fcb_cache_count);
 605 }
 606
 607 static int fsfilt_ext3_add_journal_cb(struct obd_device *obd,
 608                                       struct super_block *sb,
 609                                       __u64 last_rcvd,
 610                                       void *handle, fsfilt_cb_t cb_func,
 611                                       void *cb_data)
 612 {
 613         struct fsfilt_cb_data *fcb;
 614
 615         OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
 616         if (fcb == NULL)
 617                 RETURN(-ENOMEM);
 618
 619         atomic_inc(&fcb_cache_count);
 620         fcb->cb_func = cb_func;
 621         fcb->cb_obd = obd;
 622         fcb->cb_last_rcvd = last_rcvd;
 623         fcb->cb_data = cb_data;
 624
 625         CDEBUG(D_EXT2, "set callback for last_rcvd: "LPD64"\n", last_rcvd);
 626         lock_kernel();
 627         journal_callback_set(handle, fsfilt_ext3_cb_func,
 628                              (struct journal_callback *)fcb);
 629         unlock_kernel();
 630
 631         return 0;
 632 }
 633
 634 /*
 635  * We need to hack the return value for the free inode counts because
 636  * the current EA code requires one filesystem block per inode with EAs,
 637  * so it is possible to run out of blocks before we run out of inodes.
 638  *
 639  * This can be removed when the ext3 EA code is fixed.
 640  */
 641 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
 642 {
 643         struct kstatfs sfs;
 644         int rc;
 645
 646         memset(&sfs, 0, sizeof(sfs));
 647
 648         rc = sb->s_op->statfs(sb, &sfs);
 649
 650         if (!rc && sfs.f_bfree < sfs.f_ffree) {
 651                 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
 652                 sfs.f_ffree = sfs.f_bfree;
 653         }
 654
 655         statfs_pack(osfs, &sfs);
 656         return rc;
 657 }
 658
 659 static int fsfilt_ext3_sync(struct super_block *sb)
 660 {
 661         return ext3_force_commit(sb);
 662 }
 663
 664 extern int ext3_map_inode_page(struct inode *inode, struct page *page,
 665                                unsigned long *blocks, int *created, int create);
 666 int fsfilt_ext3_map_inode_page(struct inode *inode, struct page *page,
 667                                unsigned long *blocks, int *created, int create)
 668 {
 669         return ext3_map_inode_page(inode, page, blocks, created, create);
 670 }
 671
 672 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
 673                                int nblocks, loff_t newsize);
 674 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
 675                                       int nblocks, loff_t newsize)
 676 {
 677         return ext3_prep_san_write(inode, blocks, nblocks, newsize);
 678 }
 679
 680 static int fsfilt_ext3_read_record(struct file * file, void *buf,
 681                                    int size, loff_t *offs)
 682 {
 683         struct inode *inode = file->f_dentry->d_inode;
 684         unsigned long block;
 685         struct buffer_head *bh;
 686         int err, blocksize, csize, boffs;
 687
 688         /* prevent reading after eof */
 689         lock_kernel();
 690         if (inode->i_size < *offs + size) {
 691                 size = inode->i_size - *offs;
 692                 unlock_kernel();
 693                 if (size < 0) {
 694                         CERROR("size %llu is too short for read %u@%llu\n",
 695                                inode->i_size, size, *offs);
 696                         return -EIO;
 697                 } else if (size == 0) {
 698                         return 0;
 699                 }
 700         } else {
 701                 unlock_kernel();
 702         }
 703
 704         blocksize = 1 << inode->i_blkbits;
 705
 706         while (size > 0) {
 707                 block = *offs >> inode->i_blkbits;
 708                 boffs = *offs & (blocksize - 1);
 709                 csize = min(blocksize - boffs, size);
 710                 bh = ext3_bread(NULL, inode, block, 0, &err);
 711                 if (!bh) {
 712                         CERROR("can't read block: %d\n", err);
 713                         return err;
 714                 }
 715
 716                 memcpy(buf, bh->b_data + boffs, csize);
 717                 brelse(bh);
 718
 719                 *offs += csize;
 720                 buf += csize;
 721                 size -= csize;
 722         }
 723         return 0;
 724 }
 725
 726 static int fsfilt_ext3_write_record(struct file *file, void *buf, int bufsize,
 727                                     loff_t *offs, int force_sync)
 728 {
 729         struct buffer_head *bh = NULL;
 730         unsigned long block;
 731         struct inode *inode = file->f_dentry->d_inode;
 732         loff_t old_size = inode->i_size, offset = *offs;
 733         loff_t new_size = inode->i_size;
 734         journal_t *journal;
 735         handle_t *handle;
 736         int err, block_count = 0, blocksize, size, boffs;
 737
 738         /* Determine how many transaction credits are needed */
 739         blocksize = 1 << inode->i_blkbits;
 740         block_count = (*offs & (blocksize - 1)) + bufsize;
 741         block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
 742
 743         journal = EXT3_SB(inode->i_sb)->s_journal;
 744         lock_kernel();
 745         handle = journal_start(journal,
 746                                block_count * EXT3_DATA_TRANS_BLOCKS + 2);
 747         unlock_kernel();
 748         if (IS_ERR(handle)) {
 749                 CERROR("can't start transaction\n");
 750                 return PTR_ERR(handle);
 751         }
 752
 753         while (bufsize > 0) {
 754                 if (bh != NULL)
 755                         brelse(bh);
 756
 757                 block = offset >> inode->i_blkbits;
 758                 boffs = offset & (blocksize - 1);
 759                 size = min(blocksize - boffs, bufsize);
 760                 bh = ext3_bread(handle, inode, block, 1, &err);
 761                 if (!bh) {
 762                         CERROR("can't read/create block: %d\n", err);
 763                         goto out;
 764                 }
 765
 766                 err = ext3_journal_get_write_access(handle, bh);
 767                 if (err) {
 768                         CERROR("journal_get_write_access() returned error %d\n",
 769                                err);
 770                         goto out;
 771                 }
 772                 LASSERT(bh->b_data + boffs + size <= bh->b_data + bh->b_size);
 773                 memcpy(bh->b_data + boffs, buf, size);
 774                 err = ext3_journal_dirty_metadata(handle, bh);
 775                 if (err) {
 776                         CERROR("journal_dirty_metadata() returned error %d\n",
 777                                err);
 778                         goto out;
 779                 }
 780                 if (offset + size > new_size)
 781                         new_size = offset + size;
 782                 offset += size;
 783                 bufsize -= size;
 784                 buf += size;
 785         }
 786
 787         if (force_sync)
 788                 handle->h_sync = 1; /* recovery likes this */
 789 out:
 790         if (bh)
 791                 brelse(bh);
 792
 793         /* correct in-core and on-disk sizes */
 794         if (new_size > inode->i_size) {
 795                 lock_kernel();
 796                 if (new_size > inode->i_size)
 797                         inode->i_size = new_size;
 798                 if (inode->i_size > EXT3_I(inode)->i_disksize)
 799                         EXT3_I(inode)->i_disksize = inode->i_size;
 800                 if (inode->i_size > old_size)
 801                         mark_inode_dirty(inode);
 802                 unlock_kernel();
 803         }
 804
 805         lock_kernel();
 806         journal_stop(handle);
 807         unlock_kernel();
 808
 809         if (err == 0)
 810                 *offs = offset;
 811         return err;
 812 }
 813
 814 static int fsfilt_ext3_setup(struct super_block *sb)
 815 {
 816 #if 0
 817         EXT3_SB(sb)->dx_lock = fsfilt_ext3_dx_lock;
 818         EXT3_SB(sb)->dx_unlock = fsfilt_ext3_dx_unlock;
 819 #endif
 820 #ifdef S_PDIROPS
 821         CWARN("Enabling PDIROPS\n");
 822         set_opt(EXT3_SB(sb)->s_mount_opt, PDIROPS);
 823         sb->s_flags |= S_PDIROPS;
 824 #endif
 825         return 0;
 826 }
 827
 828 static int fsfilt_ext3_set_xattr(struct inode * inode, void *handle, char *name,
 829                                  void *buffer, int buffer_size)
 830 {
 831         int rc = 0;
 832
 833         lock_kernel();
 834
 835         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
 836                                    name, buffer, buffer_size, 0);
 837         unlock_kernel();
 838         if (rc)
 839                 CERROR("set xattr %s from inode %lu: rc %d\n",
 840                        name,  inode->i_ino, rc);
 841         return rc;
 842 }
 843
 844 static int fsfilt_ext3_get_xattr(struct inode *inode, char *name,
 845                                  void *buffer, int buffer_size)
 846 {
 847         int rc = 0;
 848         lock_kernel();
 849
 850         rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
 851                             name, buffer, buffer_size);
 852         unlock_kernel();
 853
 854         if (buffer == NULL)
 855                 return (rc == -ENODATA) ? 0 : rc;
 856         if (rc < 0) {
 857                 CDEBUG(D_INFO, "error getting EA %s from inode %lu: rc %d\n",
 858                        name,  inode->i_ino, rc);
 859                 memset(buffer, 0, buffer_size);
 860                 return (rc == -ENODATA) ? 0 : rc;
 861         }
 862
 863         return rc;
 864 }
 865
 866 /* If fso is NULL, op is FSFILT operation, otherwise op is number of fso
 867    objects. Logs is number of logfiles to update */
 868 static int fsfilt_ext3_get_op_len(int op, struct fsfilt_objinfo *fso, int logs)
 869 {
 870         if ( !fso ) {
 871                 switch(op) {
 872                 case FSFILT_OP_CREATE:
 873                                  /* directory leaf, index & indirect & EA*/
 874                         return 4 + 3 * logs;
 875                 case FSFILT_OP_UNLINK:
 876                         return 3 * logs;
 877                 }
 878         } else {
 879                 int i;
 880                 int needed = 0;
 881                 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
 882                 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 883                 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
 884                 for (i = 0; i < op; i++, fso++) {
 885                         int nblocks = fso->fso_bufcnt * blockpp;
 886                         int ndindirect = min(nblocks, addrpp + 1);
 887                         int nindir = nblocks + ndindirect + 1;
 888
 889                         needed += nindir;
 890                 }
 891                 return needed + 3 * logs;
 892         }
 893
 894         return 0;
 895 }
 896
 897 static struct fsfilt_operations fsfilt_ext3_ops = {
 898         .fs_type                = "ext3",
 899         .fs_owner               = THIS_MODULE,
 900         .fs_start               = fsfilt_ext3_start,
 901         .fs_brw_start           = fsfilt_ext3_brw_start,
 902         .fs_commit              = fsfilt_ext3_commit,
 903         .fs_commit_async        = fsfilt_ext3_commit_async,
 904         .fs_commit_wait         = fsfilt_ext3_commit_wait,
 905         .fs_setattr             = fsfilt_ext3_setattr,
 906         .fs_iocontrol           = fsfilt_ext3_iocontrol,
 907         .fs_set_md              = fsfilt_ext3_set_md,
 908         .fs_get_md              = fsfilt_ext3_get_md,
 909         .fs_readpage            = fsfilt_ext3_readpage,
 910         .fs_add_journal_cb      = fsfilt_ext3_add_journal_cb,
 911         .fs_statfs              = fsfilt_ext3_statfs,
 912         .fs_sync                = fsfilt_ext3_sync,
 913         .fs_map_inode_page      = fsfilt_ext3_map_inode_page,
 914         .fs_prep_san_write      = fsfilt_ext3_prep_san_write,
 915         .fs_write_record        = fsfilt_ext3_write_record,
 916         .fs_read_record         = fsfilt_ext3_read_record,
 917         .fs_setup               = fsfilt_ext3_setup,
 918         .fs_getpage             = fsfilt_ext3_getpage,
 919         .fs_send_bio            = fsfilt_ext3_send_bio,
 920         .fs_set_xattr           = fsfilt_ext3_set_xattr,
 921         .fs_get_xattr           = fsfilt_ext3_get_xattr,
 922         .fs_get_op_len          = fsfilt_ext3_get_op_len,
 923 };
 924
 925 static int __init fsfilt_ext3_init(void)
 926 {
 927         int rc;
 928
 929         fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
 930                                       sizeof(struct fsfilt_cb_data), 0,
 931                                       0, NULL, NULL);
 932         if (!fcb_cache) {
 933                 CERROR("error allocating fsfilt journal callback cache\n");
 934                 GOTO(out, rc = -ENOMEM);
 935         }
 936
 937         rc = fsfilt_register_ops(&fsfilt_ext3_ops);
 938
 939         if (rc)
 940                 kmem_cache_destroy(fcb_cache);
 941 out:
 942         return rc;
 943 }
 944
 945 static void __exit fsfilt_ext3_exit(void)
 946 {
 947         int rc;
 948
 949         fsfilt_unregister_ops(&fsfilt_ext3_ops);
 950         rc = kmem_cache_destroy(fcb_cache);
 951
 952         if (rc || atomic_read(&fcb_cache_count)) {
 953                 CERROR("can't free fsfilt callback cache: count %d, rc = %d\n",
 954                        atomic_read(&fcb_cache_count), rc);
 955         }
 956 }
 957
 958 module_init(fsfilt_ext3_init);
 959 module_exit(fsfilt_ext3_exit);
 960
 961 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
 962 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
 963 MODULE_LICENSE("GPL");