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 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
  54 #include <linux/ext3_extents.h>
  55 #endif
  56
  57 static kmem_cache_t *fcb_cache;
  58 static atomic_t fcb_cache_count = ATOMIC_INIT(0);
  59
  60 struct fsfilt_cb_data {
  61         struct journal_callback cb_jcb; /* jbd private data - MUST BE FIRST */
  62         fsfilt_cb_t cb_func;            /* MDS/OBD completion function */
  63         struct obd_device *cb_obd;      /* MDS/OBD completion device */
  64         __u64 cb_last_rcvd;             /* MDS/OST last committed operation */
  65         void *cb_data;                  /* MDS/OST completion function data */
  66 };
  67
  68 #ifndef EXT3_XATTR_INDEX_TRUSTED        /* temporary until we hit l28 kernel */
  69 #define EXT3_XATTR_INDEX_TRUSTED        4
  70 #endif
  71 #define XATTR_LUSTRE_MDS_LOV_EA         "lov"
  72
  73 /*
  74  * We don't currently need any additional blocks for rmdir and
  75  * unlink transactions because we are storing the OST oa_id inside
  76  * the inode (which we will be changing anyways as part of this
  77  * transaction).
  78  */
  79 static void *fsfilt_ext3_start(struct inode *inode, int op, void *desc_private,
  80                                int logs)
  81 {
  82         /* For updates to the last recieved file */
  83         int nblocks = EXT3_SINGLEDATA_TRANS_BLOCKS;
  84         journal_t *journal;
  85         void *handle;
  86
  87         if (current->journal_info) {
  88                 CDEBUG(D_INODE, "increasing refcount on %p\n",
  89                        current->journal_info);
  90                 goto journal_start;
  91         }
  92
  93         /* XXX BUG 3188 -- must return to one set of opcodes */
  94         /* FIXME - cache hook */
  95         if (op & 0x20) {
  96                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
  97                 op = op & ~0x20;
  98         }
  99
 100         /* FIXME - kml */
 101         if (op & 0x10) {
 102                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
 103                 op = op & ~0x10;
 104         }
 105
 106         switch(op) {
 107         case FSFILT_OP_RMDIR:
 108         case FSFILT_OP_UNLINK:
 109                 /* delete one file + create/update logs for each stripe */
 110                 nblocks += EXT3_DELETE_TRANS_BLOCKS;
 111                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 112                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
 113                 break;
 114         case FSFILT_OP_RENAME:
 115                 /* modify additional directory */
 116                 nblocks += EXT3_SINGLEDATA_TRANS_BLOCKS;
 117                 /* no break */
 118         case FSFILT_OP_SYMLINK:
 119                 /* additional block + block bitmap + GDT for long symlink */
 120                 nblocks += 3;
 121                 /* no break */
 122         case FSFILT_OP_CREATE:
 123                 /* create/update logs for each stripe */
 124                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 125                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
 126                 /* no break */
 127         case FSFILT_OP_MKDIR:
 128         case FSFILT_OP_MKNOD:
 129                 /* modify one inode + block bitmap + GDT */
 130                 nblocks += 3;
 131                 /* no break */
 132         case FSFILT_OP_LINK:
 133                 /* modify parent directory */
 134                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 135                         EXT3_DATA_TRANS_BLOCKS;
 136                 break;
 137         case FSFILT_OP_SETATTR:
 138                 /* Setattr on inode */
 139                 nblocks += 1;
 140                 break;
 141         case FSFILT_OP_CANCEL_UNLINK:
 142                 /* blocks for log header bitmap update OR
 143                  * blocks for catalog header bitmap update + unlink of logs */
 144                 nblocks = (LLOG_CHUNK_SIZE >> inode->i_blkbits) +
 145                         EXT3_DELETE_TRANS_BLOCKS * logs;
 146                 break;
 147         case FSFILT_OP_NOOP:
 148                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
 149                 break;
 150         default: CERROR("unknown transaction start op %d\n", op);
 151                 LBUG();
 152         }
 153
 154         LASSERT(current->journal_info == desc_private);
 155         journal = EXT3_SB(inode->i_sb)->s_journal;
 156         if (nblocks > journal->j_max_transaction_buffers) {
 157                 CERROR("too many credits %d for op %ux%u using %d instead\n",
 158                        nblocks, op, logs, journal->j_max_transaction_buffers);
 159                 nblocks = journal->j_max_transaction_buffers;
 160         }
 161
 162  journal_start:
 163         lock_kernel();
 164         handle = journal_start(EXT3_JOURNAL(inode), nblocks);
 165         unlock_kernel();
 166
 167         if (!IS_ERR(handle))
 168                 LASSERT(current->journal_info == handle);
 169         else
 170                 CERROR("error starting handle for op %u (%u credits): rc %ld\n",
 171                        op, nblocks, PTR_ERR(handle));
 172         return handle;
 173 }
 174
 175 /*
 176  * Calculate the number of buffer credits needed to write multiple pages in
 177  * a single ext3 transaction.  No, this shouldn't be here, but as yet ext3
 178  * doesn't have a nice API for calculating this sort of thing in advance.
 179  *
 180  * See comment above ext3_writepage_trans_blocks for details.  We assume
 181  * no data journaling is being done, but it does allow for all of the pages
 182  * being non-contiguous.  If we are guaranteed contiguous pages we could
 183  * reduce the number of (d)indirect blocks a lot.
 184  *
 185  * With N blocks per page and P pages, for each inode we have at most:
 186  * N*P indirect
 187  * min(N*P, blocksize/4 + 1) dindirect blocks
 188  * niocount tindirect
 189  *
 190  * For the entire filesystem, we have at most:
 191  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 192  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 193  * objcount inode blocks
 194  * 1 superblock
 195  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 196  *
 197  * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
 198  */
 199 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso,
 200                                       int niocount, struct niobuf_local *nb)
 201 {
 202         struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
 203         __u64 next_indir;
 204         const int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 205         int nbitmaps = 0, ngdblocks;
 206         int needed = objcount + 1; /* inodes + superblock */
 207         int i, j;
 208
 209         for (i = 0, j = 0; i < objcount; i++, fso++) {
 210                 /* two or more dindirect blocks in case we cross boundary */
 211                 int ndind = (long)((nb[j + fso->fso_bufcnt - 1].offset -
 212                                     nb[j].offset) >>
 213                                    sb->s_blocksize_bits) /
 214                         (EXT3_ADDR_PER_BLOCK(sb) * EXT3_ADDR_PER_BLOCK(sb));
 215                 nbitmaps += min(fso->fso_bufcnt, ndind > 0 ? ndind : 2);
 216
 217                 /* leaf, indirect, tindirect blocks for first block */
 218                 nbitmaps += blockpp + 2;
 219
 220                 j += fso->fso_bufcnt;
 221         }
 222
 223         next_indir = nb[0].offset +
 224                 (EXT3_ADDR_PER_BLOCK(sb) << sb->s_blocksize_bits);
 225         for (i = 1; i < niocount; i++) {
 226                 if (nb[i].offset >= next_indir) {
 227                         nbitmaps++;     /* additional indirect */
 228                         next_indir = nb[i].offset +
 229                                 (EXT3_ADDR_PER_BLOCK(sb)<<sb->s_blocksize_bits);
 230                 } else if (nb[i].offset != nb[i - 1].offset + sb->s_blocksize) {
 231                         nbitmaps++;     /* additional indirect */
 232                 }
 233                 nbitmaps += blockpp;    /* each leaf in different group? */
 234         }
 235
 236         ngdblocks = nbitmaps;
 237         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 238                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 239         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 240                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 241
 242         needed += nbitmaps + ngdblocks;
 243
 244         /* last_rcvd update */
 245         needed += EXT3_DATA_TRANS_BLOCKS;
 246
 247 #if defined(CONFIG_QUOTA) && !defined(__x86_64__) /* XXX */
 248         /* We assume that there will be 1 bit set in s_dquot.flags for each
 249          * quota file that is active.  This is at least true for now.
 250          */
 251         needed += hweight32(sb_any_quota_enabled(sb)) *
 252                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 253 #endif
 254
 255         return needed;
 256 }
 257
 258 /* We have to start a huge journal transaction here to hold all of the
 259  * metadata for the pages being written here.  This is necessitated by
 260  * the fact that we do lots of prepare_write operations before we do
 261  * any of the matching commit_write operations, so even if we split
 262  * up to use "smaller" transactions none of them could complete until
 263  * all of them were opened.  By having a single journal transaction,
 264  * we eliminate duplicate reservations for common blocks like the
 265  * superblock and group descriptors or bitmaps.
 266  *
 267  * We will start the transaction here, but each prepare_write will
 268  * add a refcount to the transaction, and each commit_write will
 269  * remove a refcount.  The transaction will be closed when all of
 270  * the pages have been written.
 271  */
 272 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
 273                                    int niocount, struct niobuf_local *nb,
 274                                    void *desc_private, int logs)
 275 {
 276         journal_t *journal;
 277         handle_t *handle;
 278         int needed;
 279         ENTRY;
 280
 281         LASSERT(current->journal_info == desc_private);
 282         journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
 283         needed = fsfilt_ext3_credits_needed(objcount, fso, niocount, nb);
 284
 285         /* The number of blocks we could _possibly_ dirty can very large.
 286          * We reduce our request if it is absurd (and we couldn't get that
 287          * many credits for a single handle anyways).
 288          *
 289          * At some point we have to limit the size of I/Os sent at one time,
 290          * increase the size of the journal, or we have to calculate the
 291          * actual journal requirements more carefully by checking all of
 292          * the blocks instead of being maximally pessimistic.  It remains to
 293          * be seen if this is a real problem or not.
 294          */
 295         if (needed > journal->j_max_transaction_buffers) {
 296                 CERROR("want too many journal credits (%d) using %d instead\n",
 297                        needed, journal->j_max_transaction_buffers);
 298                 needed = journal->j_max_transaction_buffers;
 299         }
 300
 301         lock_kernel();
 302         handle = journal_start(journal, needed);
 303         unlock_kernel();
 304         if (IS_ERR(handle)) {
 305                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 306                        PTR_ERR(handle));
 307         } else {
 308                 LASSERT(handle->h_buffer_credits >= needed);
 309                 LASSERT(current->journal_info == handle);
 310         }
 311
 312         RETURN(handle);
 313 }
 314
 315 static int fsfilt_ext3_commit(struct inode *inode, void *h, int force_sync)
 316 {
 317         int rc;
 318         handle_t *handle = h;
 319
 320         LASSERT(current->journal_info == handle);
 321         if (force_sync)
 322                 handle->h_sync = 1; /* recovery likes this */
 323
 324         lock_kernel();
 325         rc = journal_stop(handle);
 326         unlock_kernel();
 327
 328         return rc;
 329 }
 330
 331 static int fsfilt_ext3_commit_async(struct inode *inode, void *h,
 332                                     void **wait_handle)
 333 {
 334         unsigned long tid;
 335         transaction_t *transaction;
 336 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 337         unsigned long rtid;
 338 #endif
 339         handle_t *handle = h;
 340         journal_t *journal;
 341         int rc;
 342
 343         LASSERT(current->journal_info == handle);
 344
 345         lock_kernel();
 346         transaction = handle->h_transaction;
 347         journal = transaction->t_journal;
 348         tid = transaction->t_tid;
 349         /* we don't want to be blocked */
 350         handle->h_sync = 0;
 351         rc = journal_stop(handle);
 352         if (rc) {
 353                 CERROR("error while stopping transaction: %d\n", rc);
 354                 unlock_kernel();
 355                 return rc;
 356         }
 357 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 358         rtid = log_start_commit(journal, transaction);
 359         if (rtid != tid)
 360                 CERROR("strange race: %lu != %lu\n",
 361                        (unsigned long) tid, (unsigned long) rtid);
 362 #else
 363         log_start_commit(journal, transaction->t_tid);
 364 #endif
 365         unlock_kernel();
 366
 367         *wait_handle = (void *) tid;
 368         CDEBUG(D_INODE, "commit async: %lu\n", (unsigned long) tid);
 369         return 0;
 370 }
 371
 372 static int fsfilt_ext3_commit_wait(struct inode *inode, void *h)
 373 {
 374         tid_t tid = (tid_t)(long)h;
 375
 376         CDEBUG(D_INODE, "commit wait: %lu\n", (unsigned long) tid);
 377         if (is_journal_aborted(EXT3_JOURNAL(inode)))
 378                 return -EIO;
 379
 380         log_wait_commit(EXT3_JOURNAL(inode), tid);
 381
 382         return 0;
 383 }
 384
 385 static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
 386                                struct iattr *iattr, int do_trunc)
 387 {
 388         struct inode *inode = dentry->d_inode;
 389         int rc;
 390
 391         lock_kernel();
 392
 393         /* A _really_ horrible hack to avoid removing the data stored
 394          * in the block pointers; this is really the "small" stripe MD data.
 395          * We can avoid further hackery by virtue of the MDS file size being
 396          * zero all the time (which doesn't invoke block truncate at unlink
 397          * time), so we assert we never change the MDS file size from zero. */
 398         if (iattr->ia_valid & ATTR_SIZE && !do_trunc) {
 399                 /* ATTR_SIZE would invoke truncate: clear it */
 400                 iattr->ia_valid &= ~ATTR_SIZE;
 401                 EXT3_I(inode)->i_disksize = inode->i_size = iattr->ia_size;
 402
 403                 /* make sure _something_ gets set - so new inode
 404                  * goes to disk (probably won't work over XFS */
 405                 if (!(iattr->ia_valid & (ATTR_MODE | ATTR_MTIME | ATTR_CTIME))){
 406                         iattr->ia_valid |= ATTR_MODE;
 407                         iattr->ia_mode = inode->i_mode;
 408                 }
 409         }
 410
 411         /* Don't allow setattr to change file type */
 412         iattr->ia_mode = (inode->i_mode & S_IFMT)|(iattr->ia_mode & ~S_IFMT);
 413
 414         /* We set these flags on the client, but have already checked perms
 415          * so don't confuse inode_change_ok. */
 416         iattr->ia_valid &= ~(ATTR_MTIME_SET | ATTR_ATIME_SET);
 417
 418         if (inode->i_op->setattr) {
 419                 rc = inode->i_op->setattr(dentry, iattr);
 420         } else {
 421                 rc = inode_change_ok(inode, iattr);
 422                 if (!rc)
 423                         rc = inode_setattr(inode, iattr);
 424         }
 425
 426         unlock_kernel();
 427
 428         return rc;
 429 }
 430
 431 static int fsfilt_ext3_iocontrol(struct inode * inode, struct file *file,
 432                                  unsigned int cmd, unsigned long arg)
 433 {
 434         int rc = 0;
 435         ENTRY;
 436
 437         if (inode->i_fop->ioctl)
 438                 rc = inode->i_fop->ioctl(inode, file, cmd, arg);
 439         else
 440                 RETURN(-ENOTTY);
 441
 442         RETURN(rc);
 443 }
 444
 445 static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
 446                               void *lmm, int lmm_size)
 447 {
 448         int rc;
 449
 450         /* keep this when we get rid of OLD_EA (too noisy during conversion) */
 451         if (EXT3_I(inode)->i_file_acl /* || large inode EA flag */)
 452                 CWARN("setting EA on %lu/%u again... interesting\n",
 453                        inode->i_ino, inode->i_generation);
 454
 455         lock_kernel();
 456         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
 457                                    XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size, 0);
 458
 459         unlock_kernel();
 460
 461         if (rc)
 462                 CERROR("error adding MD data to inode %lu: rc = %d\n",
 463                        inode->i_ino, rc);
 464         return rc;
 465 }
 466
 467 /* Must be called with i_sem held */
 468 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm, int lmm_size)
 469 {
 470         int rc;
 471
 472         LASSERT(down_trylock(&inode->i_sem) != 0);
 473         lock_kernel();
 474
 475         rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
 476                             XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size);
 477         unlock_kernel();
 478
 479         /* This gives us the MD size */
 480         if (lmm == NULL)
 481                 return (rc == -ENODATA) ? 0 : rc;
 482
 483         if (rc < 0) {
 484                 CDEBUG(D_INFO, "error getting EA %d/%s from inode %lu: rc %d\n",
 485                        EXT3_XATTR_INDEX_TRUSTED, XATTR_LUSTRE_MDS_LOV_EA,
 486                        inode->i_ino, rc);
 487                 memset(lmm, 0, lmm_size);
 488                 return (rc == -ENODATA) ? 0 : rc;
 489         }
 490
 491         return rc;
 492 }
 493
 494 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 495 static int fsfilt_ext3_send_bio(struct inode *inode, struct bio *bio)
 496 {
 497         submit_bio(WRITE, bio);
 498         return 0;
 499 }
 500 #else
 501 static int fsfilt_ext3_send_bio(struct inode *inode, struct kiobuf *bio)
 502 {
 503         int rc, blocks_per_page;
 504
 505         rc = brw_kiovec(WRITE, 1, &bio, inode->i_dev,
 506                         bio->blocks, 1 << inode->i_blkbits);
 507
 508         blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 509
 510         if (rc != (1 << inode->i_blkbits) * bio->nr_pages *
 511             blocks_per_page) {
 512                 CERROR("short write?  expected %d, wrote %d\n",
 513                        (1 << inode->i_blkbits) * bio->nr_pages *
 514                        blocks_per_page, rc);
 515         }
 516
 517         return rc;
 518 }
 519 #endif
 520
 521 /* FIXME-UMKA: This should be used in 2.6.x io code later. */
 522 static struct page *fsfilt_ext3_getpage(struct inode *inode, long int index)
 523 {
 524         int rc;
 525         struct page *page;
 526
 527         page = grab_cache_page(inode->i_mapping, index);
 528         if (page == NULL)
 529                 return ERR_PTR(-ENOMEM);
 530
 531         if (PageUptodate(page)) {
 532                 unlock_page(page);
 533                 return page;
 534         }
 535
 536         rc = inode->i_mapping->a_ops->readpage(NULL, page);
 537         if (rc < 0) {
 538                 page_cache_release(page);
 539                 return ERR_PTR(rc);
 540         }
 541
 542         return page;
 543 }
 544
 545 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
 546                                     loff_t *off)
 547 {
 548         struct inode *inode = file->f_dentry->d_inode;
 549         int rc = 0;
 550
 551         if (S_ISREG(inode->i_mode))
 552                 rc = file->f_op->read(file, buf, count, off);
 553         else {
 554                 const int blkbits = inode->i_sb->s_blocksize_bits;
 555                 const int blksize = inode->i_sb->s_blocksize;
 556
 557                 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
 558                        count, inode->i_ino, *off);
 559                 while (count > 0) {
 560                         struct buffer_head *bh;
 561
 562                         bh = NULL;
 563                         if (*off < inode->i_size) {
 564                                 int err = 0;
 565
 566                                 bh = ext3_bread(NULL, inode, *off >> blkbits,
 567                                                 0, &err);
 568
 569                                 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
 570
 571                                 if (bh) {
 572                                         memcpy(buf, bh->b_data, blksize);
 573                                         brelse(bh);
 574                                 } else if (err) {
 575                                         /* XXX in theory we should just fake
 576                                          * this buffer and continue like ext3,
 577                                          * especially if this is a partial read
 578                                          */
 579                                         CERROR("error read dir %lu+%llu: %d\n",
 580                                                inode->i_ino, *off, err);
 581                                         RETURN(err);
 582                                 }
 583                         }
 584                         if (!bh) {
 585                                 struct ext3_dir_entry_2 *fake = (void *)buf;
 586
 587                                 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
 588                                 memset(fake, 0, sizeof(*fake));
 589                                 fake->rec_len = cpu_to_le32(blksize);
 590                         }
 591                         count -= blksize;
 592                         buf += blksize;
 593                         *off += blksize;
 594                         rc += blksize;
 595                 }
 596         }
 597
 598         return rc;
 599 }
 600
 601 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
 602 {
 603         struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
 604
 605         fcb->cb_func(fcb->cb_obd, fcb->cb_last_rcvd, fcb->cb_data, error);
 606
 607         OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
 608         atomic_dec(&fcb_cache_count);
 609 }
 610
 611 static int fsfilt_ext3_add_journal_cb(struct obd_device *obd,
 612                                       struct super_block *sb,
 613                                       __u64 last_rcvd,
 614                                       void *handle, fsfilt_cb_t cb_func,
 615                                       void *cb_data)
 616 {
 617         struct fsfilt_cb_data *fcb;
 618
 619         OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
 620         if (fcb == NULL)
 621                 RETURN(-ENOMEM);
 622
 623         atomic_inc(&fcb_cache_count);
 624         fcb->cb_func = cb_func;
 625         fcb->cb_obd = obd;
 626         fcb->cb_last_rcvd = last_rcvd;
 627         fcb->cb_data = cb_data;
 628
 629         CDEBUG(D_EXT2, "set callback for last_rcvd: "LPD64"\n", last_rcvd);
 630         lock_kernel();
 631         journal_callback_set(handle, fsfilt_ext3_cb_func,
 632                              (struct journal_callback *)fcb);
 633         unlock_kernel();
 634
 635         return 0;
 636 }
 637
 638 /*
 639  * We need to hack the return value for the free inode counts because
 640  * the current EA code requires one filesystem block per inode with EAs,
 641  * so it is possible to run out of blocks before we run out of inodes.
 642  *
 643  * This can be removed when the ext3 EA code is fixed.
 644  */
 645 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
 646 {
 647         struct kstatfs sfs;
 648         int rc;
 649
 650         memset(&sfs, 0, sizeof(sfs));
 651
 652         rc = sb->s_op->statfs(sb, &sfs);
 653
 654         if (!rc && sfs.f_bfree < sfs.f_ffree) {
 655                 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
 656                 sfs.f_ffree = sfs.f_bfree;
 657         }
 658
 659         statfs_pack(osfs, &sfs);
 660         return rc;
 661 }
 662
 663 static int fsfilt_ext3_sync(struct super_block *sb)
 664 {
 665         return ext3_force_commit(sb);
 666 }
 667
 668 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
 669 struct bpointers {
 670         unsigned long *blocks;
 671         int *created;
 672         unsigned long start;
 673         int num;
 674         int init_num;
 675 };
 676
 677 static int ext3_ext_new_extent_cb(struct ext3_extents_tree *tree,
 678                                   struct ext3_ext_path *path,
 679                                   struct ext3_extent *newex, int exist)
 680 {
 681         struct inode *inode = tree->inode;
 682         struct bpointers *bp = tree->private;
 683         int count, err, goal;
 684         unsigned long pblock;
 685         unsigned long tgen;
 686         loff_t new_i_size;
 687         handle_t *handle;
 688         int i;
 689
 690         i = EXT_DEPTH(tree);
 691         EXT_ASSERT(i == path->p_depth);
 692         EXT_ASSERT(path[i].p_hdr);
 693
 694         if (exist) {
 695                 err = EXT_CONTINUE;
 696                 goto map;
 697         }
 698
 699         tgen = EXT_GENERATION(tree);
 700         count = ext3_ext_calc_credits_for_insert(tree, path);
 701         up_write(&EXT3_I(inode)->truncate_sem);
 702
 703         handle = ext3_journal_start(inode, count + EXT3_ALLOC_NEEDED + 1);
 704         if (IS_ERR(handle)) {
 705                 down_write(&EXT3_I(inode)->truncate_sem);
 706                 return PTR_ERR(handle);
 707         }
 708
 709         if (tgen != EXT_GENERATION(tree)) {
 710                 /* the tree has changed. so path can be invalid at moment */
 711                 ext3_journal_stop(handle, inode);
 712                 down_write(&EXT3_I(inode)->truncate_sem);
 713                 return EXT_REPEAT;
 714         }
 715
 716         down_write(&EXT3_I(inode)->truncate_sem);
 717         goal = ext3_ext_find_goal(inode, path, newex->e_block);
 718         count = newex->e_num;
 719         pblock = ext3_new_blocks(handle, inode, &count, goal, &err);
 720         if (!pblock)
 721                 goto out;
 722         EXT_ASSERT(count <= newex->e_num);
 723
 724         /* insert new extent */
 725         newex->e_start = pblock;
 726         newex->e_num = count;
 727         err = ext3_ext_insert_extent(handle, tree, path, newex);
 728         if (err)
 729                 goto out;
 730
 731         /* correct on-disk inode size */
 732         if (newex->e_num > 0) {
 733                 new_i_size = (loff_t) newex->e_block + newex->e_num;
 734                 new_i_size = new_i_size << inode->i_blkbits;
 735                 if (new_i_size > EXT3_I(inode)->i_disksize) {
 736                         EXT3_I(inode)->i_disksize = new_i_size;
 737                         err = ext3_mark_inode_dirty(handle, inode);
 738                 }
 739         }
 740
 741 out:
 742         ext3_journal_stop(handle, inode);
 743 map:
 744         if (err >= 0) {
 745                 /* map blocks */
 746                 if (bp->num == 0) {
 747                         CERROR("hmm. why do we find this extent?\n");
 748                         CERROR("initial space: %lu:%u\n",
 749                                 bp->start, bp->init_num);
 750                         CERROR("current extent: %u/%u/%u %d\n",
 751                                 newex->e_block, newex->e_num,
 752                                 newex->e_start, exist);
 753                 }
 754                 i = 0;
 755                 if (newex->e_block < bp->start)
 756                         i = bp->start - newex->e_block;
 757                 if (i >= newex->e_num)
 758                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 759                                         i, newex->e_num);
 760                 for (; i < newex->e_num && bp->num; i++) {
 761                         *(bp->created) = (exist == 0 ? 1 : 0);
 762                         bp->created++;
 763                         *(bp->blocks) = newex->e_start + i;
 764                         bp->blocks++;
 765                         bp->num--;
 766                 }
 767         }
 768         return err;
 769 }
 770
 771 int fsfilt_map_nblocks(struct inode *inode, unsigned long block,
 772                        unsigned long num, unsigned long *blocks,
 773                        int *created, int create)
 774 {
 775         struct ext3_extents_tree tree;
 776         struct bpointers bp;
 777         int err;
 778
 779         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
 780                 block, block + num, (unsigned) inode->i_ino);
 781
 782         ext3_init_tree_desc(&tree, inode);
 783         tree.private = &bp;
 784         bp.blocks = blocks;
 785         bp.created = created;
 786         bp.start = block;
 787         bp.init_num = bp.num = num;
 788
 789         down_write(&EXT3_I(inode)->truncate_sem);
 790         err = ext3_ext_walk_space(&tree, block, num, ext3_ext_new_extent_cb);
 791         ext3_ext_invalidate_cache(&tree);
 792         up_write(&EXT3_I(inode)->truncate_sem);
 793
 794         return err;
 795 }
 796
 797 int fsfilt_ext3_map_ext_inode_pages(struct inode *inode, struct page **page,
 798                                     int pages, unsigned long *blocks,
 799                                     int *created, int create)
 800 {
 801         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 802         int rc = 0, i = 0;
 803         struct page *fp = NULL;
 804         int clen = 0;
 805
 806         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 807                 inode->i_ino, pages, (*page)->index);
 808
 809         /* pages are sorted already. so, we just have to find
 810          * contig. space and process them properly */
 811         while (i < pages) {
 812                 if (fp == NULL) {
 813                         /* start new extent */
 814                         fp = *page++;
 815                         clen = 1;
 816                         i++;
 817                         continue;
 818                 } else if (fp->index + clen == (*page)->index) {
 819                         /* continue the extent */
 820                         page++;
 821                         clen++;
 822                         i++;
 823                         continue;
 824                 }
 825
 826                 /* process found extent */
 827                 rc = fsfilt_map_nblocks(inode, fp->index * blocks_per_page,
 828                                         clen * blocks_per_page, blocks,
 829                                         created, create);
 830                 if (rc)
 831                         GOTO(cleanup, rc);
 832
 833                 /* look for next extent */
 834                 fp = NULL;
 835                 blocks += blocks_per_page * clen;
 836                 created += blocks_per_page * clen;
 837         }
 838
 839         if (fp)
 840                 rc = fsfilt_map_nblocks(inode, fp->index * blocks_per_page,
 841                                         clen * blocks_per_page, blocks,
 842                                         created, create);
 843 cleanup:
 844         return rc;
 845 }
 846 #endif
 847
 848 extern int ext3_map_inode_page(struct inode *inode, struct page *page,
 849                                unsigned long *blocks, int *created, int create);
 850 int fsfilt_ext3_map_bm_inode_pages(struct inode *inode, struct page **page,
 851                                    int pages, unsigned long *blocks,
 852                                    int *created, int create)
 853 {
 854         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 855         unsigned long *b;
 856         int rc = 0, i, *cr;
 857
 858         for (i = 0, cr = created, b = blocks; i < pages; i++, page++) {
 859                 rc = ext3_map_inode_page(inode, *page, b, cr, create);
 860                 if (rc) {
 861                         CERROR("ino %lu, blk %lu cr %u create %d: rc %d\n",
 862                                inode->i_ino, *b, *cr, create, rc);
 863                         break;
 864                 }
 865
 866                 b += blocks_per_page;
 867                 cr += blocks_per_page;
 868         }
 869         return rc;
 870 }
 871
 872 int fsfilt_ext3_map_inode_pages(struct inode *inode, struct page **page,
 873                                 int pages, unsigned long *blocks,
 874                                 int *created, int create,
 875                                 struct semaphore *optional_sem)
 876 {
 877         int rc;
 878 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
 879         if (EXT3_I(inode)->i_flags & EXT3_EXTENTS_FL) {
 880                 rc = fsfilt_ext3_map_ext_inode_pages(inode, page, pages,
 881                                                      blocks, created, create);
 882                 return rc;
 883         }
 884 #endif
 885         if (optional_sem != NULL)
 886                 down(optional_sem);
 887         rc = fsfilt_ext3_map_bm_inode_pages(inode, page, pages, blocks,
 888                                             created, create);
 889         if (optional_sem != NULL)
 890                 up(optional_sem);
 891
 892         return rc;
 893 }
 894
 895 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
 896                                int nblocks, loff_t newsize);
 897 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
 898                                       int nblocks, loff_t newsize)
 899 {
 900         return ext3_prep_san_write(inode, blocks, nblocks, newsize);
 901 }
 902
 903 static int fsfilt_ext3_read_record(struct file * file, void *buf,
 904                                    int size, loff_t *offs)
 905 {
 906         struct inode *inode = file->f_dentry->d_inode;
 907         unsigned long block;
 908         struct buffer_head *bh;
 909         int err, blocksize, csize, boffs;
 910
 911         /* prevent reading after eof */
 912         lock_kernel();
 913         if (inode->i_size < *offs + size) {
 914                 size = inode->i_size - *offs;
 915                 unlock_kernel();
 916                 if (size < 0) {
 917                         CERROR("size %llu is too short for read %u@%llu\n",
 918                                inode->i_size, size, *offs);
 919                         return -EIO;
 920                 } else if (size == 0) {
 921                         return 0;
 922                 }
 923         } else {
 924                 unlock_kernel();
 925         }
 926
 927         blocksize = 1 << inode->i_blkbits;
 928
 929         while (size > 0) {
 930                 block = *offs >> inode->i_blkbits;
 931                 boffs = *offs & (blocksize - 1);
 932                 csize = min(blocksize - boffs, size);
 933                 bh = ext3_bread(NULL, inode, block, 0, &err);
 934                 if (!bh) {
 935                         CERROR("can't read block: %d\n", err);
 936                         return err;
 937                 }
 938
 939                 memcpy(buf, bh->b_data + boffs, csize);
 940                 brelse(bh);
 941
 942                 *offs += csize;
 943                 buf += csize;
 944                 size -= csize;
 945         }
 946         return 0;
 947 }
 948
 949 static int fsfilt_ext3_write_record(struct file *file, void *buf, int bufsize,
 950                                     loff_t *offs, int force_sync)
 951 {
 952         struct buffer_head *bh = NULL;
 953         unsigned long block;
 954         struct inode *inode = file->f_dentry->d_inode;
 955         loff_t old_size = inode->i_size, offset = *offs;
 956         loff_t new_size = inode->i_size;
 957         journal_t *journal;
 958         handle_t *handle;
 959         int err, block_count = 0, blocksize, size, boffs;
 960
 961         /* Determine how many transaction credits are needed */
 962         blocksize = 1 << inode->i_blkbits;
 963         block_count = (*offs & (blocksize - 1)) + bufsize;
 964         block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
 965
 966         journal = EXT3_SB(inode->i_sb)->s_journal;
 967         lock_kernel();
 968         handle = journal_start(journal,
 969                                block_count * EXT3_DATA_TRANS_BLOCKS + 2);
 970         unlock_kernel();
 971         if (IS_ERR(handle)) {
 972                 CERROR("can't start transaction\n");
 973                 return PTR_ERR(handle);
 974         }
 975
 976         while (bufsize > 0) {
 977                 if (bh != NULL)
 978                         brelse(bh);
 979
 980                 block = offset >> inode->i_blkbits;
 981                 boffs = offset & (blocksize - 1);
 982                 size = min(blocksize - boffs, bufsize);
 983                 bh = ext3_bread(handle, inode, block, 1, &err);
 984                 if (!bh) {
 985                         CERROR("can't read/create block: %d\n", err);
 986                         goto out;
 987                 }
 988
 989                 err = ext3_journal_get_write_access(handle, bh);
 990                 if (err) {
 991                         CERROR("journal_get_write_access() returned error %d\n",
 992                                err);
 993                         goto out;
 994                 }
 995                 LASSERT(bh->b_data + boffs + size <= bh->b_data + bh->b_size);
 996                 memcpy(bh->b_data + boffs, buf, size);
 997                 err = ext3_journal_dirty_metadata(handle, bh);
 998                 if (err) {
 999                         CERROR("journal_dirty_metadata() returned error %d\n",
1000                                err);
1001                         goto out;
1002                 }
1003                 if (offset + size > new_size)
1004                         new_size = offset + size;
1005                 offset += size;
1006                 bufsize -= size;
1007                 buf += size;
1008         }
1009
1010         if (force_sync)
1011                 handle->h_sync = 1; /* recovery likes this */
1012 out:
1013         if (bh)
1014                 brelse(bh);
1015
1016         /* correct in-core and on-disk sizes */
1017         if (new_size > inode->i_size) {
1018                 lock_kernel();
1019                 if (new_size > inode->i_size)
1020                         inode->i_size = new_size;
1021                 if (inode->i_size > EXT3_I(inode)->i_disksize)
1022                         EXT3_I(inode)->i_disksize = inode->i_size;
1023                 if (inode->i_size > old_size)
1024                         mark_inode_dirty(inode);
1025                 unlock_kernel();
1026         }
1027
1028         lock_kernel();
1029         journal_stop(handle);
1030         unlock_kernel();
1031
1032         if (err == 0)
1033                 *offs = offset;
1034         return err;
1035 }
1036
1037 static int fsfilt_ext3_setup(struct super_block *sb)
1038 {
1039 #if 0
1040         EXT3_SB(sb)->dx_lock = fsfilt_ext3_dx_lock;
1041         EXT3_SB(sb)->dx_unlock = fsfilt_ext3_dx_unlock;
1042 #endif
1043 #ifdef S_PDIROPS
1044         CWARN("Enabling PDIROPS\n");
1045         set_opt(EXT3_SB(sb)->s_mount_opt, PDIROPS);
1046         sb->s_flags |= S_PDIROPS;
1047 #endif
1048         return 0;
1049 }
1050
1051 static int fsfilt_ext3_set_xattr(struct inode * inode, void *handle, char *name,
1052                                  void *buffer, int buffer_size)
1053 {
1054         int rc = 0;
1055
1056         lock_kernel();
1057
1058         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
1059                                    name, buffer, buffer_size, 0);
1060         unlock_kernel();
1061         if (rc)
1062                 CERROR("set xattr %s from inode %lu: rc %d\n",
1063                        name,  inode->i_ino, rc);
1064         return rc;
1065 }
1066
1067 static int fsfilt_ext3_get_xattr(struct inode *inode, char *name,
1068                                  void *buffer, int buffer_size)
1069 {
1070         int rc = 0;
1071         lock_kernel();
1072
1073         rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
1074                             name, buffer, buffer_size);
1075         unlock_kernel();
1076
1077         if (buffer == NULL)
1078                 return (rc == -ENODATA) ? 0 : rc;
1079         if (rc < 0) {
1080                 CDEBUG(D_INFO, "error getting EA %s from inode %lu: rc %d\n",
1081                        name,  inode->i_ino, rc);
1082                 memset(buffer, 0, buffer_size);
1083                 return (rc == -ENODATA) ? 0 : rc;
1084         }
1085
1086         return rc;
1087 }
1088
1089 /* If fso is NULL, op is FSFILT operation, otherwise op is number of fso
1090    objects. Logs is number of logfiles to update */
1091 static int fsfilt_ext3_get_op_len(int op, struct fsfilt_objinfo *fso, int logs)
1092 {
1093         if ( !fso ) {
1094                 switch(op) {
1095                 case FSFILT_OP_CREATE:
1096                                  /* directory leaf, index & indirect & EA*/
1097                         return 4 + 3 * logs;
1098                 case FSFILT_OP_UNLINK:
1099                         return 3 * logs;
1100                 }
1101         } else {
1102                 int i;
1103                 int needed = 0;
1104                 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
1105                 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
1106                 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
1107                 for (i = 0; i < op; i++, fso++) {
1108                         int nblocks = fso->fso_bufcnt * blockpp;
1109                         int ndindirect = min(nblocks, addrpp + 1);
1110                         int nindir = nblocks + ndindirect + 1;
1111
1112                         needed += nindir;
1113                 }
1114                 return needed + 3 * logs;
1115         }
1116
1117         return 0;
1118 }
1119
1120 static struct fsfilt_operations fsfilt_ext3_ops = {
1121         .fs_type                = "ext3",
1122         .fs_owner               = THIS_MODULE,
1123         .fs_start               = fsfilt_ext3_start,
1124         .fs_brw_start           = fsfilt_ext3_brw_start,
1125         .fs_commit              = fsfilt_ext3_commit,
1126         .fs_commit_async        = fsfilt_ext3_commit_async,
1127         .fs_commit_wait         = fsfilt_ext3_commit_wait,
1128         .fs_setattr             = fsfilt_ext3_setattr,
1129         .fs_iocontrol           = fsfilt_ext3_iocontrol,
1130         .fs_set_md              = fsfilt_ext3_set_md,
1131         .fs_get_md              = fsfilt_ext3_get_md,
1132         .fs_readpage            = fsfilt_ext3_readpage,
1133         .fs_add_journal_cb      = fsfilt_ext3_add_journal_cb,
1134         .fs_statfs              = fsfilt_ext3_statfs,
1135         .fs_sync                = fsfilt_ext3_sync,
1136         .fs_map_inode_pages     = fsfilt_ext3_map_inode_pages,
1137         .fs_prep_san_write      = fsfilt_ext3_prep_san_write,
1138         .fs_write_record        = fsfilt_ext3_write_record,
1139         .fs_read_record         = fsfilt_ext3_read_record,
1140         .fs_setup               = fsfilt_ext3_setup,
1141         .fs_getpage             = fsfilt_ext3_getpage,
1142         .fs_send_bio            = fsfilt_ext3_send_bio,
1143         .fs_set_xattr           = fsfilt_ext3_set_xattr,
1144         .fs_get_xattr           = fsfilt_ext3_get_xattr,
1145         .fs_get_op_len          = fsfilt_ext3_get_op_len,
1146 };
1147
1148 static int __init fsfilt_ext3_init(void)
1149 {
1150         int rc;
1151
1152         fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
1153                                       sizeof(struct fsfilt_cb_data), 0,
1154                                       0, NULL, NULL);
1155         if (!fcb_cache) {
1156                 CERROR("error allocating fsfilt journal callback cache\n");
1157                 GOTO(out, rc = -ENOMEM);
1158         }
1159
1160         rc = fsfilt_register_ops(&fsfilt_ext3_ops);
1161
1162         if (rc)
1163                 kmem_cache_destroy(fcb_cache);
1164 out:
1165         return rc;
1166 }
1167
1168 static void __exit fsfilt_ext3_exit(void)
1169 {
1170         int rc;
1171
1172         fsfilt_unregister_ops(&fsfilt_ext3_ops);
1173         rc = kmem_cache_destroy(fcb_cache);
1174
1175         if (rc || atomic_read(&fcb_cache_count)) {
1176                 CERROR("can't free fsfilt callback cache: count %d, rc = %d\n",
1177                        atomic_read(&fcb_cache_count), rc);
1178         }
1179 }
1180
1181 module_init(fsfilt_ext3_init);
1182 module_exit(fsfilt_ext3_exit);
1183
1184 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1185 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
1186 MODULE_LICENSE("GPL");