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 #include <linux/bitops.h>
  39 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
  40 #include <linux/ext3_xattr.h>
  41 #else
  42 #include <ext3/xattr.h>
  43 #endif
  44
  45 #include <linux/kp30.h>
  46 #include <linux/lustre_fsfilt.h>
  47 #include <linux/obd.h>
  48 #include <linux/obd_class.h>
  49 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
  50 #include <linux/iobuf.h>
  51 #endif
  52
  53 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
  54 #include <linux/ext3_extents.h>
  55 #endif
  56
  57 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,7))
  58 # define lock_24kernel() lock_kernel()
  59 # define unlock_24kernel() unlock_kernel()
  60 #else
  61 # define lock_24kernel() do {} while (0)
  62 # define unlock_24kernel() do {} while (0)
  63 #endif
  64
  65 static kmem_cache_t *fcb_cache;
  66 static atomic_t fcb_cache_count = ATOMIC_INIT(0);
  67
  68 struct fsfilt_cb_data {
  69         struct journal_callback cb_jcb; /* jbd private data - MUST BE FIRST */
  70         fsfilt_cb_t cb_func;            /* MDS/OBD completion function */
  71         struct obd_device *cb_obd;      /* MDS/OBD completion device */
  72         __u64 cb_last_rcvd;             /* MDS/OST last committed operation */
  73         void *cb_data;                  /* MDS/OST completion function data */
  74 };
  75
  76 #ifndef EXT3_XATTR_INDEX_TRUSTED        /* temporary until we hit l28 kernel */
  77 #define EXT3_XATTR_INDEX_TRUSTED        4
  78 #endif
  79 #define XATTR_LUSTRE_MDS_LOV_EA         "lov"
  80
  81 /*
  82  * We don't currently need any additional blocks for rmdir and
  83  * unlink transactions because we are storing the OST oa_id inside
  84  * the inode (which we will be changing anyways as part of this
  85  * transaction).
  86  */
  87 static void *fsfilt_ext3_start(struct inode *inode, int op, void *desc_private,
  88                                int logs)
  89 {
  90         /* For updates to the last recieved file */
  91         int nblocks = EXT3_SINGLEDATA_TRANS_BLOCKS;
  92         journal_t *journal;
  93         void *handle;
  94
  95         if (current->journal_info) {
  96                 CDEBUG(D_INODE, "increasing refcount on %p\n",
  97                        current->journal_info);
  98                 goto journal_start;
  99         }
 100
 101         switch(op) {
 102         case FSFILT_OP_RMDIR:
 103         case FSFILT_OP_UNLINK:
 104                 /* delete one file + create/update logs for each stripe */
 105                 nblocks += EXT3_DELETE_TRANS_BLOCKS;
 106                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 107                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
 108                 break;
 109         case FSFILT_OP_RENAME:
 110                 /* modify additional directory */
 111                 nblocks += EXT3_SINGLEDATA_TRANS_BLOCKS;
 112                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 113                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
 114                 /* no break */
 115         case FSFILT_OP_SYMLINK:
 116                 /* additional block + block bitmap + GDT for long symlink */
 117                 nblocks += 3;
 118                 /* no break */
 119         case FSFILT_OP_CREATE:
 120                 /* create/update logs for each stripe */
 121                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 122                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
 123                 /* no break */
 124         case FSFILT_OP_MKDIR:
 125         case FSFILT_OP_MKNOD:
 126                 /* modify one inode + block bitmap + GDT */
 127                 nblocks += 3;
 128                 /* no break */
 129         case FSFILT_OP_LINK:
 130                 /* modify parent directory */
 131                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 132                         EXT3_DATA_TRANS_BLOCKS;
 133                 break;
 134         case FSFILT_OP_SETATTR:
 135                 /* Setattr on inode */
 136                 nblocks += 1;
 137                 break;
 138         case FSFILT_OP_CANCEL_UNLINK:
 139                 /* blocks for log header bitmap update OR
 140                  * blocks for catalog header bitmap update + unlink of logs */
 141                 nblocks = (LLOG_CHUNK_SIZE >> inode->i_blkbits) +
 142                         EXT3_DELETE_TRANS_BLOCKS * logs;
 143                 break;
 144         default: CERROR("unknown transaction start op %d\n", op);
 145                 LBUG();
 146         }
 147
 148         LASSERT(current->journal_info == desc_private);
 149         journal = EXT3_SB(inode->i_sb)->s_journal;
 150         if (nblocks > journal->j_max_transaction_buffers) {
 151                 CERROR("too many credits %d for op %ux%u using %d instead\n",
 152                        nblocks, op, logs, journal->j_max_transaction_buffers);
 153                 nblocks = journal->j_max_transaction_buffers;
 154         }
 155
 156  journal_start:
 157         LASSERTF(nblocks > 0, "can't start %d credit transaction\n", nblocks);
 158         lock_24kernel();
 159         handle = journal_start(EXT3_JOURNAL(inode), nblocks);
 160         unlock_24kernel();
 161
 162         if (!IS_ERR(handle))
 163                 LASSERT(current->journal_info == handle);
 164         else
 165                 CERROR("error starting handle for op %u (%u credits): rc %ld\n",
 166                        op, nblocks, PTR_ERR(handle));
 167         return handle;
 168 }
 169
 170 /*
 171  * Calculate the number of buffer credits needed to write multiple pages in
 172  * a single ext3 transaction.  No, this shouldn't be here, but as yet ext3
 173  * doesn't have a nice API for calculating this sort of thing in advance.
 174  *
 175  * See comment above ext3_writepage_trans_blocks for details.  We assume
 176  * no data journaling is being done, but it does allow for all of the pages
 177  * being non-contiguous.  If we are guaranteed contiguous pages we could
 178  * reduce the number of (d)indirect blocks a lot.
 179  *
 180  * With N blocks per page and P pages, for each inode we have at most:
 181  * N*P indirect
 182  * min(N*P, blocksize/4 + 1) dindirect blocks
 183  * niocount tindirect
 184  *
 185  * For the entire filesystem, we have at most:
 186  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 187  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 188  * objcount inode blocks
 189  * 1 superblock
 190  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 191  *
 192  * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
 193  */
 194 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso,
 195                                       int niocount, struct niobuf_local *nb)
 196 {
 197         struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
 198         __u64 next_indir;
 199         const int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 200         int nbitmaps = 0, ngdblocks;
 201         int needed = objcount + 1; /* inodes + superblock */
 202         int i, j;
 203
 204         for (i = 0, j = 0; i < objcount; i++, fso++) {
 205                 /* two or more dindirect blocks in case we cross boundary */
 206                 int ndind = (long)((nb[j + fso->fso_bufcnt - 1].offset -
 207                                     nb[j].offset) >>
 208                                    sb->s_blocksize_bits) /
 209                         (EXT3_ADDR_PER_BLOCK(sb) * EXT3_ADDR_PER_BLOCK(sb));
 210                 nbitmaps += min(fso->fso_bufcnt, ndind > 0 ? ndind : 2);
 211
 212                 /* leaf, indirect, tindirect blocks for first block */
 213                 nbitmaps += blockpp + 2;
 214
 215                 j += fso->fso_bufcnt;
 216         }
 217
 218         next_indir = nb[0].offset +
 219                 (EXT3_ADDR_PER_BLOCK(sb) << sb->s_blocksize_bits);
 220         for (i = 1; i < niocount; i++) {
 221                 if (nb[i].offset >= next_indir) {
 222                         nbitmaps++;     /* additional indirect */
 223                         next_indir = nb[i].offset +
 224                                 (EXT3_ADDR_PER_BLOCK(sb)<<sb->s_blocksize_bits);
 225                 } else if (nb[i].offset != nb[i - 1].offset + sb->s_blocksize) {
 226                         nbitmaps++;     /* additional indirect */
 227                 }
 228                 nbitmaps += blockpp;    /* each leaf in different group? */
 229         }
 230
 231         ngdblocks = nbitmaps;
 232         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 233                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 234         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 235                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 236
 237         needed += nbitmaps + ngdblocks;
 238
 239         /* last_rcvd update */
 240         needed += EXT3_DATA_TRANS_BLOCKS;
 241
 242 #if defined(CONFIG_QUOTA)
 243         /* We assume that there will be 1 bit set in s_dquot.flags for each
 244          * quota file that is active.  This is at least true for now.
 245          */
 246         needed += hweight32(sb_any_quota_enabled(sb)) *
 247                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 248 #endif
 249
 250         return needed;
 251 }
 252
 253 /* We have to start a huge journal transaction here to hold all of the
 254  * metadata for the pages being written here.  This is necessitated by
 255  * the fact that we do lots of prepare_write operations before we do
 256  * any of the matching commit_write operations, so even if we split
 257  * up to use "smaller" transactions none of them could complete until
 258  * all of them were opened.  By having a single journal transaction,
 259  * we eliminate duplicate reservations for common blocks like the
 260  * superblock and group descriptors or bitmaps.
 261  *
 262  * We will start the transaction here, but each prepare_write will
 263  * add a refcount to the transaction, and each commit_write will
 264  * remove a refcount.  The transaction will be closed when all of
 265  * the pages have been written.
 266  */
 267 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
 268                                    int niocount, struct niobuf_local *nb,
 269                                    void *desc_private, int logs)
 270 {
 271         journal_t *journal;
 272         handle_t *handle;
 273         int needed;
 274         ENTRY;
 275
 276         LASSERT(current->journal_info == desc_private);
 277         journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
 278         needed = fsfilt_ext3_credits_needed(objcount, fso, niocount, nb);
 279
 280         /* The number of blocks we could _possibly_ dirty can very large.
 281          * We reduce our request if it is absurd (and we couldn't get that
 282          * many credits for a single handle anyways).
 283          *
 284          * At some point we have to limit the size of I/Os sent at one time,
 285          * increase the size of the journal, or we have to calculate the
 286          * actual journal requirements more carefully by checking all of
 287          * the blocks instead of being maximally pessimistic.  It remains to
 288          * be seen if this is a real problem or not.
 289          */
 290         if (needed > journal->j_max_transaction_buffers) {
 291                 CERROR("want too many journal credits (%d) using %d instead\n",
 292                        needed, journal->j_max_transaction_buffers);
 293                 needed = journal->j_max_transaction_buffers;
 294         }
 295
 296         LASSERTF(needed > 0, "can't start %d credit transaction\n", needed);
 297         lock_24kernel();
 298         handle = journal_start(journal, needed);
 299         unlock_24kernel();
 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_24kernel();
 321         rc = journal_stop(handle);
 322         unlock_24kernel();
 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         LASSERT(down_trylock(&inode->i_sem) != 0);
 447
 448         if (EXT3_I(inode)->i_file_acl /* || large inode EA flag */)
 449                 CWARN("setting EA on %lu/%u again... interesting\n",
 450                        inode->i_ino, inode->i_generation);
 451
 452         lock_kernel();
 453         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
 454                                    XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size, 0);
 455
 456         unlock_kernel();
 457
 458         if (rc)
 459                 CERROR("error adding MD data to inode %lu: rc = %d\n",
 460                        inode->i_ino, rc);
 461         return rc;
 462 }
 463
 464 /* Must be called with i_sem held */
 465 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm, int lmm_size)
 466 {
 467         int rc;
 468
 469         LASSERT(down_trylock(&inode->i_sem) != 0);
 470         lock_kernel();
 471
 472         rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
 473                             XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size);
 474         unlock_kernel();
 475
 476         /* This gives us the MD size */
 477         if (lmm == NULL)
 478                 return (rc == -ENODATA) ? 0 : rc;
 479
 480         if (rc < 0) {
 481                 CDEBUG(D_INFO, "error getting EA %d/%s from inode %lu: rc %d\n",
 482                        EXT3_XATTR_INDEX_TRUSTED, XATTR_LUSTRE_MDS_LOV_EA,
 483                        inode->i_ino, rc);
 484                 memset(lmm, 0, lmm_size);
 485                 return (rc == -ENODATA) ? 0 : rc;
 486         }
 487
 488         return rc;
 489 }
 490
 491 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 492 static int fsfilt_ext3_send_bio(int rw, struct inode *inode, struct bio *bio)
 493 {
 494         submit_bio(rw, bio);
 495         return 0;
 496 }
 497 #else
 498 static int fsfilt_ext3_send_bio(int rw, struct inode *inode, struct kiobuf *bio)
 499 {
 500         int rc, blocks_per_page;
 501
 502         rc = brw_kiovec(rw, 1, &bio, inode->i_dev,
 503                         bio->blocks, 1 << inode->i_blkbits);
 504
 505         blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 506
 507         if (rc != (1 << inode->i_blkbits) * bio->nr_pages * 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 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
 518                                     loff_t *off)
 519 {
 520         struct inode *inode = file->f_dentry->d_inode;
 521         int rc = 0;
 522
 523         if (S_ISREG(inode->i_mode))
 524                 rc = file->f_op->read(file, buf, count, off);
 525         else {
 526                 const int blkbits = inode->i_sb->s_blocksize_bits;
 527                 const int blksize = inode->i_sb->s_blocksize;
 528
 529                 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
 530                        count, inode->i_ino, *off);
 531                 while (count > 0) {
 532                         struct buffer_head *bh;
 533
 534                         bh = NULL;
 535                         if (*off < inode->i_size) {
 536                                 int err = 0;
 537
 538                                 bh = ext3_bread(NULL, inode, *off >> blkbits,
 539                                                 0, &err);
 540
 541                                 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
 542
 543                                 if (bh) {
 544                                         memcpy(buf, bh->b_data, blksize);
 545                                         brelse(bh);
 546                                 } else if (err) {
 547                                         /* XXX in theory we should just fake
 548                                          * this buffer and continue like ext3,
 549                                          * especially if this is a partial read
 550                                          */
 551                                         CERROR("error read dir %lu+%llu: %d\n",
 552                                                inode->i_ino, *off, err);
 553                                         RETURN(err);
 554                                 }
 555                         }
 556                         if (!bh) {
 557                                 struct ext3_dir_entry_2 *fake = (void *)buf;
 558
 559                                 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
 560                                 memset(fake, 0, sizeof(*fake));
 561                                 fake->rec_len = cpu_to_le32(blksize);
 562                         }
 563                         count -= blksize;
 564                         buf += blksize;
 565                         *off += blksize;
 566                         rc += blksize;
 567                 }
 568         }
 569
 570         return rc;
 571 }
 572
 573 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
 574 {
 575         struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
 576
 577         fcb->cb_func(fcb->cb_obd, fcb->cb_last_rcvd, fcb->cb_data, error);
 578
 579         OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
 580         atomic_dec(&fcb_cache_count);
 581 }
 582
 583 static int fsfilt_ext3_add_journal_cb(struct obd_device *obd, __u64 last_rcvd,
 584                                       void *handle, fsfilt_cb_t cb_func,
 585                                       void *cb_data)
 586 {
 587         struct fsfilt_cb_data *fcb;
 588
 589         OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
 590         if (fcb == NULL)
 591                 RETURN(-ENOMEM);
 592
 593         atomic_inc(&fcb_cache_count);
 594         fcb->cb_func = cb_func;
 595         fcb->cb_obd = obd;
 596         fcb->cb_last_rcvd = last_rcvd;
 597         fcb->cb_data = cb_data;
 598
 599         CDEBUG(D_EXT2, "set callback for last_rcvd: "LPD64"\n", last_rcvd);
 600         lock_24kernel();
 601         journal_callback_set(handle, fsfilt_ext3_cb_func,
 602                              (struct journal_callback *)fcb);
 603         unlock_24kernel();
 604
 605         return 0;
 606 }
 607
 608 /*
 609  * We need to hack the return value for the free inode counts because
 610  * the current EA code requires one filesystem block per inode with EAs,
 611  * so it is possible to run out of blocks before we run out of inodes.
 612  *
 613  * This can be removed when the ext3 EA code is fixed.
 614  */
 615 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
 616 {
 617         struct kstatfs sfs;
 618         int rc;
 619
 620         memset(&sfs, 0, sizeof(sfs));
 621
 622         rc = sb->s_op->statfs(sb, &sfs);
 623
 624         if (!rc && sfs.f_bfree < sfs.f_ffree) {
 625                 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
 626                 sfs.f_ffree = sfs.f_bfree;
 627         }
 628
 629         statfs_pack(osfs, &sfs);
 630         return rc;
 631 }
 632
 633 static int fsfilt_ext3_sync(struct super_block *sb)
 634 {
 635         return ext3_force_commit(sb);
 636 }
 637
 638 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
 639 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 640 #define ext3_up_truncate_sem(inode)  up_write(&EXT3_I(inode)->truncate_sem);
 641 #define ext3_down_truncate_sem(inode)  down_write(&EXT3_I(inode)->truncate_sem);
 642 #else
 643 #define ext3_up_truncate_sem(inode)  up(&EXT3_I(inode)->truncate_sem);
 644 #define ext3_down_truncate_sem(inode)  down(&EXT3_I(inode)->truncate_sem);
 645 #endif
 646
 647 #include <linux/lustre_version.h>
 648 #if EXT3_EXT_MAGIC == 0xf301
 649 #define ee_start e_start
 650 #define ee_block e_block
 651 #define ee_len   e_num
 652 #endif
 653 #ifndef EXT3_BB_MAX_BLOCKS
 654 #define ext3_mb_new_blocks(handle, inode, goal, count, aflags, err) \
 655         ext3_new_blocks(handle, inode, count, goal, err)
 656 #endif
 657
 658 struct bpointers {
 659         unsigned long *blocks;
 660         int *created;
 661         unsigned long start;
 662         int num;
 663         int init_num;
 664         int create;
 665 };
 666
 667 static int ext3_ext_find_goal(struct inode *inode, struct ext3_ext_path *path,
 668                               unsigned long block, int *aflags)
 669 {
 670         struct ext3_inode_info *ei = EXT3_I(inode);
 671         unsigned long bg_start;
 672         unsigned long colour;
 673         int depth;
 674
 675         if (path) {
 676                 struct ext3_extent *ex;
 677                 depth = path->p_depth;
 678
 679                 /* try to predict block placement */
 680                 if ((ex = path[depth].p_ext)) {
 681 #if 0
 682                         /* This prefers to eat into a contiguous extent
 683                          * rather than find an extent that the whole
 684                          * request will fit into.  This can fragment data
 685                          * block allocation and prevents our lovely 1M I/Os
 686                          * from reaching the disk intact. */
 687                         if (ex->ee_block + ex->ee_len == block)
 688                                 *aflags |= 1;
 689 #endif
 690                         return ex->ee_start + (block - ex->ee_block);
 691                 }
 692
 693                 /* it looks index is empty
 694                  * try to find starting from index itself */
 695                 if (path[depth].p_bh)
 696                         return path[depth].p_bh->b_blocknr;
 697         }
 698
 699         /* OK. use inode's group */
 700         bg_start = (ei->i_block_group * EXT3_BLOCKS_PER_GROUP(inode->i_sb)) +
 701                 le32_to_cpu(EXT3_SB(inode->i_sb)->s_es->s_first_data_block);
 702         colour = (current->pid % 16) *
 703                 (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 704         return bg_start + colour + block;
 705 }
 706
 707 static int ext3_ext_new_extent_cb(struct ext3_extents_tree *tree,
 708                                   struct ext3_ext_path *path,
 709                                   struct ext3_extent *newex, int exist)
 710 {
 711         struct inode *inode = tree->inode;
 712         struct bpointers *bp = tree->private;
 713         int count, err, goal;
 714         unsigned long pblock;
 715         unsigned long tgen;
 716         loff_t new_i_size;
 717         handle_t *handle;
 718         int i, aflags = 0;
 719
 720         i = EXT_DEPTH(tree);
 721         EXT_ASSERT(i == path->p_depth);
 722         EXT_ASSERT(path[i].p_hdr);
 723
 724         if (exist) {
 725                 err = EXT_CONTINUE;
 726                 goto map;
 727         }
 728
 729         if (bp->create == 0) {
 730                 i = 0;
 731                 if (newex->ee_block < bp->start)
 732                         i = bp->start - newex->ee_block;
 733                 if (i >= newex->ee_len)
 734                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 735                                         i, newex->ee_len);
 736                 for (; i < newex->ee_len && bp->num; i++) {
 737                         *(bp->created) = 0;
 738                         bp->created++;
 739                         *(bp->blocks) = 0;
 740                         bp->blocks++;
 741                         bp->num--;
 742                         bp->start++;
 743                 }
 744
 745                 return EXT_CONTINUE;
 746         }
 747
 748         tgen = EXT_GENERATION(tree);
 749         count = ext3_ext_calc_credits_for_insert(tree, path);
 750         ext3_up_truncate_sem(inode);
 751
 752         lock_24kernel();
 753         handle = journal_start(EXT3_JOURNAL(inode), count + EXT3_ALLOC_NEEDED + 1);
 754         unlock_24kernel();
 755         if (IS_ERR(handle)) {
 756                 ext3_down_truncate_sem(inode);
 757                 return PTR_ERR(handle);
 758         }
 759
 760         if (tgen != EXT_GENERATION(tree)) {
 761                 /* the tree has changed. so path can be invalid at moment */
 762                 lock_24kernel();
 763                 journal_stop(handle);
 764                 unlock_24kernel();
 765                 ext3_down_truncate_sem(inode);
 766                 return EXT_REPEAT;
 767         }
 768
 769         ext3_down_truncate_sem(inode);
 770         count = newex->ee_len;
 771         goal = ext3_ext_find_goal(inode, path, newex->ee_block, &aflags);
 772         aflags |= 2; /* block have been already reserved */
 773         pblock = ext3_mb_new_blocks(handle, inode, goal, &count, aflags, &err);
 774         if (!pblock)
 775                 goto out;
 776         EXT_ASSERT(count <= newex->ee_len);
 777
 778         /* insert new extent */
 779         newex->ee_start = pblock;
 780         newex->ee_len = count;
 781         err = ext3_ext_insert_extent(handle, tree, path, newex);
 782         if (err)
 783                 goto out;
 784
 785         /* correct on-disk inode size */
 786         if (newex->ee_len > 0) {
 787                 new_i_size = (loff_t) newex->ee_block + newex->ee_len;
 788                 new_i_size = new_i_size << inode->i_blkbits;
 789                 if (new_i_size > EXT3_I(inode)->i_disksize) {
 790                         EXT3_I(inode)->i_disksize = new_i_size;
 791                         err = ext3_mark_inode_dirty(handle, inode);
 792                 }
 793         }
 794
 795 out:
 796         lock_24kernel();
 797         journal_stop(handle);
 798         unlock_24kernel();
 799 map:
 800         if (err >= 0) {
 801                 /* map blocks */
 802                 if (bp->num == 0) {
 803                         CERROR("hmm. why do we find this extent?\n");
 804                         CERROR("initial space: %lu:%u\n",
 805                                 bp->start, bp->init_num);
 806                         CERROR("current extent: %u/%u/%u %d\n",
 807                                 newex->ee_block, newex->ee_len,
 808                                 newex->ee_start, exist);
 809                 }
 810                 i = 0;
 811                 if (newex->ee_block < bp->start)
 812                         i = bp->start - newex->ee_block;
 813                 if (i >= newex->ee_len)
 814                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 815                                         i, newex->ee_len);
 816                 for (; i < newex->ee_len && bp->num; i++) {
 817                         *(bp->created) = (exist == 0 ? 1 : 0);
 818                         bp->created++;
 819                         *(bp->blocks) = newex->ee_start + i;
 820                         bp->blocks++;
 821                         bp->num--;
 822                         bp->start++;
 823                 }
 824         }
 825         return err;
 826 }
 827
 828 int fsfilt_map_nblocks(struct inode *inode, unsigned long block,
 829                        unsigned long num, unsigned long *blocks,
 830                        int *created, int create)
 831 {
 832         struct ext3_extents_tree tree;
 833         struct bpointers bp;
 834         int err;
 835
 836         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
 837                 block, block + num, (unsigned) inode->i_ino);
 838
 839         ext3_init_tree_desc(&tree, inode);
 840         tree.private = &bp;
 841         bp.blocks = blocks;
 842         bp.created = created;
 843         bp.start = block;
 844         bp.init_num = bp.num = num;
 845         bp.create = create;
 846
 847         ext3_down_truncate_sem(inode);
 848         err = ext3_ext_walk_space(&tree, block, num, ext3_ext_new_extent_cb);
 849         ext3_ext_invalidate_cache(&tree);
 850         ext3_up_truncate_sem(inode);
 851
 852         return err;
 853 }
 854
 855 int fsfilt_ext3_map_ext_inode_pages(struct inode *inode, struct page **page,
 856                                     int pages, unsigned long *blocks,
 857                                     int *created, int create)
 858 {
 859         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 860         int rc = 0, i = 0;
 861         struct page *fp = NULL;
 862         int clen = 0;
 863
 864         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 865                 inode->i_ino, pages, (*page)->index);
 866
 867         /* pages are sorted already. so, we just have to find
 868          * contig. space and process them properly */
 869         while (i < pages) {
 870                 if (fp == NULL) {
 871                         /* start new extent */
 872                         fp = *page++;
 873                         clen = 1;
 874                         i++;
 875                         continue;
 876                 } else if (fp->index + clen == (*page)->index) {
 877                         /* continue the extent */
 878                         page++;
 879                         clen++;
 880                         i++;
 881                         continue;
 882                 }
 883
 884                 /* process found extent */
 885                 rc = fsfilt_map_nblocks(inode, fp->index * blocks_per_page,
 886                                         clen * blocks_per_page, blocks,
 887                                         created, create);
 888                 if (rc)
 889                         GOTO(cleanup, rc);
 890
 891                 /* look for next extent */
 892                 fp = NULL;
 893                 blocks += blocks_per_page * clen;
 894                 created += blocks_per_page * clen;
 895         }
 896
 897         if (fp)
 898                 rc = fsfilt_map_nblocks(inode, fp->index * blocks_per_page,
 899                                         clen * blocks_per_page, blocks,
 900                                         created, create);
 901 cleanup:
 902         return rc;
 903 }
 904 #endif
 905
 906 extern int ext3_map_inode_page(struct inode *inode, struct page *page,
 907                                unsigned long *blocks, int *created, int create);
 908 int fsfilt_ext3_map_bm_inode_pages(struct inode *inode, struct page **page,
 909                                    int pages, unsigned long *blocks,
 910                                    int *created, int create)
 911 {
 912         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 913         unsigned long *b;
 914         int rc = 0, i, *cr;
 915
 916         for (i = 0, cr = created, b = blocks; i < pages; i++, page++) {
 917                 rc = ext3_map_inode_page(inode, *page, b, cr, create);
 918                 if (rc) {
 919                         CERROR("ino %lu, blk %lu cr %u create %d: rc %d\n",
 920                                inode->i_ino, *b, *cr, create, rc);
 921                         break;
 922                 }
 923
 924                 b += blocks_per_page;
 925                 cr += blocks_per_page;
 926         }
 927         return rc;
 928 }
 929
 930 int fsfilt_ext3_map_inode_pages(struct inode *inode, struct page **page,
 931                                 int pages, unsigned long *blocks,
 932                                 int *created, int create,
 933                                 struct semaphore *optional_sem)
 934 {
 935         int rc;
 936 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
 937         if (EXT3_I(inode)->i_flags & EXT3_EXTENTS_FL) {
 938                 rc = fsfilt_ext3_map_ext_inode_pages(inode, page, pages,
 939                                                      blocks, created, create);
 940                 return rc;
 941         }
 942 #endif
 943         if (optional_sem != NULL)
 944                 down(optional_sem);
 945         rc = fsfilt_ext3_map_bm_inode_pages(inode, page, pages, blocks,
 946                                             created, create);
 947         if (optional_sem != NULL)
 948                 up(optional_sem);
 949
 950         return rc;
 951 }
 952
 953 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
 954                                int nblocks, loff_t newsize);
 955 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
 956                                       int nblocks, loff_t newsize)
 957 {
 958         return ext3_prep_san_write(inode, blocks, nblocks, newsize);
 959 }
 960
 961 static int fsfilt_ext3_read_record(struct file * file, void *buf,
 962                                    int size, loff_t *offs)
 963 {
 964         struct inode *inode = file->f_dentry->d_inode;
 965         unsigned long block;
 966         struct buffer_head *bh;
 967         int err, blocksize, csize, boffs;
 968
 969         /* prevent reading after eof */
 970         lock_kernel();
 971         if (inode->i_size < *offs + size) {
 972                 size = inode->i_size - *offs;
 973                 unlock_kernel();
 974                 if (size < 0) {
 975                         CERROR("size %llu is too short for read %u@%llu\n",
 976                                inode->i_size, size, *offs);
 977                         return -EIO;
 978                 } else if (size == 0) {
 979                         return 0;
 980                 }
 981         } else {
 982                 unlock_kernel();
 983         }
 984
 985         blocksize = 1 << inode->i_blkbits;
 986
 987         while (size > 0) {
 988                 block = *offs >> inode->i_blkbits;
 989                 boffs = *offs & (blocksize - 1);
 990                 csize = min(blocksize - boffs, size);
 991                 bh = ext3_bread(NULL, inode, block, 0, &err);
 992                 if (!bh) {
 993                         CERROR("can't read block: %d\n", err);
 994                         return err;
 995                 }
 996
 997                 memcpy(buf, bh->b_data + boffs, csize);
 998                 brelse(bh);
 999
1000                 *offs += csize;
1001                 buf += csize;
1002                 size -= csize;
1003         }
1004         return 0;
1005 }
1006
1007 static int fsfilt_ext3_write_record(struct file *file, void *buf, int bufsize,
1008                                     loff_t *offs, int force_sync)
1009 {
1010         struct buffer_head *bh = NULL;
1011         unsigned long block;
1012         struct inode *inode = file->f_dentry->d_inode;
1013         loff_t old_size = inode->i_size, offset = *offs;
1014         loff_t new_size = inode->i_size;
1015         journal_t *journal;
1016         handle_t *handle;
1017         int err, block_count = 0, blocksize, size, boffs;
1018
1019         /* Determine how many transaction credits are needed */
1020         blocksize = 1 << inode->i_blkbits;
1021         block_count = (*offs & (blocksize - 1)) + bufsize;
1022         block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
1023
1024         journal = EXT3_SB(inode->i_sb)->s_journal;
1025         lock_24kernel();
1026         handle = journal_start(journal,
1027                                block_count * EXT3_DATA_TRANS_BLOCKS + 2);
1028         unlock_24kernel();
1029         if (IS_ERR(handle)) {
1030                 CERROR("can't start transaction\n");
1031                 return PTR_ERR(handle);
1032         }
1033
1034         while (bufsize > 0) {
1035                 if (bh != NULL)
1036                         brelse(bh);
1037
1038                 block = offset >> inode->i_blkbits;
1039                 boffs = offset & (blocksize - 1);
1040                 size = min(blocksize - boffs, bufsize);
1041                 bh = ext3_bread(handle, inode, block, 1, &err);
1042                 if (!bh) {
1043                         CERROR("can't read/create block: %d\n", err);
1044                         goto out;
1045                 }
1046
1047                 err = ext3_journal_get_write_access(handle, bh);
1048                 if (err) {
1049                         CERROR("journal_get_write_access() returned error %d\n",
1050                                err);
1051                         goto out;
1052                 }
1053                 LASSERT(bh->b_data + boffs + size <= bh->b_data + bh->b_size);
1054                 memcpy(bh->b_data + boffs, buf, size);
1055                 err = ext3_journal_dirty_metadata(handle, bh);
1056                 if (err) {
1057                         CERROR("journal_dirty_metadata() returned error %d\n",
1058                                err);
1059                         goto out;
1060                 }
1061                 if (offset + size > new_size)
1062                         new_size = offset + size;
1063                 offset += size;
1064                 bufsize -= size;
1065                 buf += size;
1066         }
1067
1068         if (force_sync)
1069                 handle->h_sync = 1; /* recovery likes this */
1070 out:
1071         if (bh)
1072                 brelse(bh);
1073
1074         /* correct in-core and on-disk sizes */
1075         if (new_size > inode->i_size) {
1076                 lock_kernel();
1077                 if (new_size > inode->i_size)
1078                         inode->i_size = new_size;
1079                 if (inode->i_size > EXT3_I(inode)->i_disksize)
1080                         EXT3_I(inode)->i_disksize = inode->i_size;
1081                 if (inode->i_size > old_size)
1082                         mark_inode_dirty(inode);
1083                 unlock_kernel();
1084         }
1085
1086         lock_24kernel();
1087         journal_stop(handle);
1088         unlock_24kernel();
1089
1090         if (err == 0)
1091                 *offs = offset;
1092         return err;
1093 }
1094
1095 static int fsfilt_ext3_setup(struct super_block *sb)
1096 {
1097 #if 0
1098         EXT3_SB(sb)->dx_lock = fsfilt_ext3_dx_lock;
1099         EXT3_SB(sb)->dx_unlock = fsfilt_ext3_dx_unlock;
1100 #endif
1101 #ifdef S_PDIROPS
1102         CWARN("Enabling PDIROPS\n");
1103         set_opt(EXT3_SB(sb)->s_mount_opt, PDIROPS);
1104         sb->s_flags |= S_PDIROPS;
1105 #endif
1106         return 0;
1107 }
1108
1109 /* If fso is NULL, op is FSFILT operation, otherwise op is number of fso
1110    objects. Logs is number of logfiles to update */
1111 static int fsfilt_ext3_get_op_len(int op, struct fsfilt_objinfo *fso, int logs)
1112 {
1113         if ( !fso ) {
1114                 switch(op) {
1115                 case FSFILT_OP_CREATE:
1116                                  /* directory leaf, index & indirect & EA*/
1117                         return 4 + 3 * logs;
1118                 case FSFILT_OP_UNLINK:
1119                         return 3 * logs;
1120                 }
1121         } else {
1122                 int i;
1123                 int needed = 0;
1124                 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
1125                 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
1126                 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
1127                 for (i = 0; i < op; i++, fso++) {
1128                         int nblocks = fso->fso_bufcnt * blockpp;
1129                         int ndindirect = min(nblocks, addrpp + 1);
1130                         int nindir = nblocks + ndindirect + 1;
1131
1132                         needed += nindir;
1133                 }
1134                 return needed + 3 * logs;
1135         }
1136
1137         return 0;
1138 }
1139
1140 static struct fsfilt_operations fsfilt_ext3_ops = {
1141         .fs_type                = "ext3",
1142         .fs_owner               = THIS_MODULE,
1143         .fs_start               = fsfilt_ext3_start,
1144         .fs_brw_start           = fsfilt_ext3_brw_start,
1145         .fs_commit              = fsfilt_ext3_commit,
1146         .fs_commit_async        = fsfilt_ext3_commit_async,
1147         .fs_commit_wait         = fsfilt_ext3_commit_wait,
1148         .fs_setattr             = fsfilt_ext3_setattr,
1149         .fs_iocontrol           = fsfilt_ext3_iocontrol,
1150         .fs_set_md              = fsfilt_ext3_set_md,
1151         .fs_get_md              = fsfilt_ext3_get_md,
1152         .fs_readpage            = fsfilt_ext3_readpage,
1153         .fs_add_journal_cb      = fsfilt_ext3_add_journal_cb,
1154         .fs_statfs              = fsfilt_ext3_statfs,
1155         .fs_sync                = fsfilt_ext3_sync,
1156         .fs_map_inode_pages     = fsfilt_ext3_map_inode_pages,
1157         .fs_prep_san_write      = fsfilt_ext3_prep_san_write,
1158         .fs_write_record        = fsfilt_ext3_write_record,
1159         .fs_read_record         = fsfilt_ext3_read_record,
1160         .fs_setup               = fsfilt_ext3_setup,
1161         .fs_send_bio            = fsfilt_ext3_send_bio,
1162         .fs_get_op_len          = fsfilt_ext3_get_op_len,
1163 };
1164
1165 static int __init fsfilt_ext3_init(void)
1166 {
1167         int rc;
1168
1169         fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
1170                                       sizeof(struct fsfilt_cb_data), 0,
1171                                       0, NULL, NULL);
1172         if (!fcb_cache) {
1173                 CERROR("error allocating fsfilt journal callback cache\n");
1174                 GOTO(out, rc = -ENOMEM);
1175         }
1176
1177         rc = fsfilt_register_ops(&fsfilt_ext3_ops);
1178
1179         if (rc)
1180                 kmem_cache_destroy(fcb_cache);
1181 out:
1182         return rc;
1183 }
1184
1185 static void __exit fsfilt_ext3_exit(void)
1186 {
1187         fsfilt_unregister_ops(&fsfilt_ext3_ops);
1188         LASSERTF(kmem_cache_destroy(fcb_cache) == 0,
1189                  "can't free fsfilt callback cache: count %d\n",
1190                  atomic_read(&fcb_cache_count));
1191 }
1192
1193 module_init(fsfilt_ext3_init);
1194 module_exit(fsfilt_ext3_exit);
1195
1196 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1197 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
1198 MODULE_LICENSE("GPL");