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/ext3_extents.h>
  38 #include <linux/version.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 <libcfs/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/module.h>
  51 #include <linux/iobuf.h>
  52 #endif
  53
  54 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,7))
  55 # define lock_24kernel() lock_kernel()
  56 # define unlock_24kernel() unlock_kernel()
  57 #else
  58 # define lock_24kernel() do {} while (0)
  59 # define unlock_24kernel() do {} while (0)
  60 #endif
  61
  62 static kmem_cache_t *fcb_cache;
  63 static atomic_t fcb_cache_count = ATOMIC_INIT(0);
  64
  65 struct fsfilt_cb_data {
  66         struct journal_callback cb_jcb; /* jbd private data - MUST BE FIRST */
  67         fsfilt_cb_t cb_func;            /* MDS/OBD completion function */
  68         struct obd_device *cb_obd;      /* MDS/OBD completion device */
  69         __u64 cb_last_num;              /* MDS/OST last committed operation */
  70         void *cb_data;                  /* MDS/OST completion function data */
  71 };
  72
  73 #ifndef EXT3_XATTR_INDEX_TRUSTED        /* temporary until we hit l28 kernel */
  74 #define EXT3_XATTR_INDEX_TRUSTED        4
  75 #endif
  76
  77 /*
  78  * We don't currently need any additional blocks for rmdir and
  79  * unlink transactions because we are storing the OST oa_id inside
  80  * the inode (which we will be changing anyways as part of this
  81  * transaction).
  82  */
  83 static void *fsfilt_ext3_start(struct inode *inode, int op, void *desc_private,
  84                                int logs)
  85 {
  86         /* For updates to the last recieved file */
  87         int nblocks = EXT3_SINGLEDATA_TRANS_BLOCKS;
  88         journal_t *journal;
  89         void *handle;
  90
  91         if (current->journal_info) {
  92                 CDEBUG(D_INODE, "increasing refcount on %p\n",
  93                        current->journal_info);
  94                 goto journal_start;
  95         }
  96
  97         if (logs)
  98                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
  99                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
 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                 /* no break */
 113         case FSFILT_OP_SYMLINK:
 114                 /* additional block + block bitmap + GDT for long symlink */
 115                 nblocks += 3;
 116                 /* no break */
 117         case FSFILT_OP_CREATE:
 118                 /* create/update logs for each stripe */
 119                 /*nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 120                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;*/
 121                 /* no break */
 122         case FSFILT_OP_MKDIR:
 123         case FSFILT_OP_MKNOD:
 124                 /* modify one inode + block bitmap + GDT */
 125                 nblocks += 3;
 126                 /* no break */
 127         case FSFILT_OP_LINK:
 128                 /* modify parent directory */
 129                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 130                         EXT3_DATA_TRANS_BLOCKS;
 131                 break;
 132         case FSFILT_OP_SETATTR:
 133                 /* Setattr on inode */
 134                 nblocks += 1;
 135                 break;
 136         case FSFILT_OP_CANCEL_UNLINK:
 137                 /* blocks for log header bitmap update OR
 138                  * blocks for catalog header bitmap update + unlink of logs */
 139                 nblocks = (LLOG_CHUNK_SIZE >> inode->i_blkbits) +
 140                         EXT3_DELETE_TRANS_BLOCKS * logs;
 141                 break;
 142         case FSFILT_OP_NOOP:
 143                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
 144                 break;
 145         default: CERROR("unknown transaction start op %d\n", op);
 146                 LBUG();
 147         }
 148
 149         LASSERT(current->journal_info == desc_private);
 150         journal = EXT3_SB(inode->i_sb)->s_journal;
 151         if (nblocks > journal->j_max_transaction_buffers) {
 152                 CERROR("too many credits %d for op %ux%u using %d instead\n",
 153                        nblocks, op, logs, journal->j_max_transaction_buffers);
 154                 nblocks = journal->j_max_transaction_buffers;
 155         }
 156
 157  journal_start:
 158         LASSERTF(nblocks > 0, "can't start %d credit transaction\n", nblocks);
 159         lock_24kernel();
 160         handle = journal_start(EXT3_JOURNAL(inode), nblocks);
 161         unlock_24kernel();
 162
 163         if (!IS_ERR(handle))
 164                 LASSERT(current->journal_info == handle);
 165         else
 166                 CERROR("error starting handle for op %u (%u credits): rc %ld\n",
 167                        op, nblocks, PTR_ERR(handle));
 168         return handle;
 169 }
 170
 171 /*
 172  * Calculate the number of buffer credits needed to write multiple pages in
 173  * a single ext3 transaction.  No, this shouldn't be here, but as yet ext3
 174  * doesn't have a nice API for calculating this sort of thing in advance.
 175  *
 176  * See comment above ext3_writepage_trans_blocks for details.  We assume
 177  * no data journaling is being done, but it does allow for all of the pages
 178  * being non-contiguous.  If we are guaranteed contiguous pages we could
 179  * reduce the number of (d)indirect blocks a lot.
 180  *
 181  * With N blocks per page and P pages, for each inode we have at most:
 182  * N*P indirect
 183  * min(N*P, blocksize/4 + 1) dindirect blocks
 184  * niocount tindirect
 185  *
 186  * For the entire filesystem, we have at most:
 187  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 188  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 189  * objcount inode blocks
 190  * 1 superblock
 191  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 192  *
 193  * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
 194  */
 195 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso,
 196                                       int niocount, struct niobuf_local *nb)
 197 {
 198         struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
 199         __u64 next_indir;
 200         const int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 201         int nbitmaps = 0, ngdblocks;
 202         int needed = objcount + 1; /* inodes + superblock */
 203         int i, j;
 204
 205         for (i = 0, j = 0; i < objcount; i++, fso++) {
 206                 /* two or more dindirect blocks in case we cross boundary */
 207                 int ndind = (long)((nb[j + fso->fso_bufcnt - 1].offset -
 208                                     nb[j].offset) >>
 209                                    sb->s_blocksize_bits) /
 210                         (EXT3_ADDR_PER_BLOCK(sb) * EXT3_ADDR_PER_BLOCK(sb));
 211                 nbitmaps += min(fso->fso_bufcnt, ndind > 0 ? ndind : 2);
 212
 213                 /* leaf, indirect, tindirect blocks for first block */
 214                 nbitmaps += blockpp + 2;
 215
 216                 j += fso->fso_bufcnt;
 217         }
 218
 219         next_indir = nb[0].offset +
 220                 (EXT3_ADDR_PER_BLOCK(sb) << sb->s_blocksize_bits);
 221         for (i = 1; i < niocount; i++) {
 222                 if (nb[i].offset >= next_indir) {
 223                         nbitmaps++;     /* additional indirect */
 224                         next_indir = nb[i].offset +
 225                                 (EXT3_ADDR_PER_BLOCK(sb)<<sb->s_blocksize_bits);
 226                 } else if (nb[i].offset != nb[i - 1].offset + sb->s_blocksize) {
 227                         nbitmaps++;     /* additional indirect */
 228                 }
 229                 nbitmaps += blockpp;    /* each leaf in different group? */
 230         }
 231
 232         ngdblocks = nbitmaps;
 233         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 234                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 235         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 236                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 237
 238         needed += nbitmaps + ngdblocks;
 239
 240         /* last_rcvd update */
 241         needed += EXT3_DATA_TRANS_BLOCKS;
 242
 243 #if defined(CONFIG_QUOTA) && !defined(__x86_64__) /* XXX */
 244         /* We assume that there will be 1 bit set in s_dquot.flags for each
 245          * quota file that is active.  This is at least true for now.
 246          */
 247         needed += hweight32(sb_any_quota_enabled(sb)) *
 248                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 249 #endif
 250
 251         return needed;
 252 }
 253
 254 /* We have to start a huge journal transaction here to hold all of the
 255  * metadata for the pages being written here.  This is necessitated by
 256  * the fact that we do lots of prepare_write operations before we do
 257  * any of the matching commit_write operations, so even if we split
 258  * up to use "smaller" transactions none of them could complete until
 259  * all of them were opened.  By having a single journal transaction,
 260  * we eliminate duplicate reservations for common blocks like the
 261  * superblock and group descriptors or bitmaps.
 262  *
 263  * We will start the transaction here, but each prepare_write will
 264  * add a refcount to the transaction, and each commit_write will
 265  * remove a refcount.  The transaction will be closed when all of
 266  * the pages have been written.
 267  */
 268 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
 269                                    int niocount, struct niobuf_local *nb,
 270                                    void *desc_private, int logs)
 271 {
 272         journal_t *journal;
 273         handle_t *handle;
 274         int needed;
 275         ENTRY;
 276
 277         LASSERT(current->journal_info == desc_private);
 278         journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
 279         needed = fsfilt_ext3_credits_needed(objcount, fso, niocount, nb);
 280
 281         /* The number of blocks we could _possibly_ dirty can very large.
 282          * We reduce our request if it is absurd (and we couldn't get that
 283          * many credits for a single handle anyways).
 284          *
 285          * At some point we have to limit the size of I/Os sent at one time,
 286          * increase the size of the journal, or we have to calculate the
 287          * actual journal requirements more carefully by checking all of
 288          * the blocks instead of being maximally pessimistic.  It remains to
 289          * be seen if this is a real problem or not.
 290          */
 291         if (needed > journal->j_max_transaction_buffers) {
 292                 CERROR("want too many journal credits (%d) using %d instead\n",
 293                        needed, journal->j_max_transaction_buffers);
 294                 needed = journal->j_max_transaction_buffers;
 295         }
 296
 297         LASSERTF(needed > 0, "can't start %d credit transaction\n", needed);
 298         lock_24kernel();
 299         handle = journal_start(journal, needed);
 300         unlock_24kernel();
 301         if (IS_ERR(handle)) {
 302                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 303                        PTR_ERR(handle));
 304         } else {
 305                 LASSERT(handle->h_buffer_credits >= needed);
 306                 LASSERT(current->journal_info == handle);
 307         }
 308
 309         RETURN(handle);
 310 }
 311
 312 static int fsfilt_ext3_commit(struct super_block *sb, struct inode *inode,
 313                               void *h, int force_sync)
 314 {
 315         int rc;
 316         handle_t *handle = h;
 317
 318         LASSERT(current->journal_info == handle);
 319         if (force_sync)
 320                 handle->h_sync = 1; /* recovery likes this */
 321
 322         lock_24kernel();
 323         rc = journal_stop(handle);
 324         unlock_24kernel();
 325
 326         return rc;
 327 }
 328
 329 static int fsfilt_ext3_commit_async(struct inode *inode, void *h,
 330                                     void **wait_handle)
 331 {
 332         unsigned long tid;
 333         transaction_t *transaction;
 334 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 335         unsigned long rtid;
 336 #endif
 337         handle_t *handle = h;
 338         journal_t *journal;
 339         int rc;
 340
 341         LASSERT(current->journal_info == handle);
 342
 343         lock_24kernel();
 344         transaction = handle->h_transaction;
 345         journal = transaction->t_journal;
 346         tid = transaction->t_tid;
 347         /* we don't want to be blocked */
 348         handle->h_sync = 0;
 349         rc = journal_stop(handle);
 350         if (rc) {
 351                 CERROR("error while stopping transaction: %d\n", rc);
 352                 unlock_kernel();
 353                 return rc;
 354         }
 355 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 356         rtid = log_start_commit(journal, transaction);
 357         if (rtid != tid)
 358                 CERROR("strange race: %lu != %lu\n",
 359                        (unsigned long) tid, (unsigned long) rtid);
 360 #else
 361         log_start_commit(journal, tid);
 362 #endif
 363         unlock_24kernel();
 364
 365         *wait_handle = (void *) tid;
 366         CDEBUG(D_INODE, "commit async: %lu\n", (unsigned long) tid);
 367         return 0;
 368 }
 369
 370 static int fsfilt_ext3_commit_wait(struct inode *inode, void *h)
 371 {
 372         journal_t *journal = EXT3_JOURNAL(inode);
 373         tid_t tid = (tid_t)(long)h;
 374
 375         CDEBUG(D_INODE, "commit wait: %lu\n", (unsigned long) tid);
 376         if (unlikely(is_journal_aborted(journal)))
 377                 return -EIO;
 378
 379         log_wait_commit(EXT3_JOURNAL(inode), tid);
 380
 381         if (unlikely(is_journal_aborted(journal)))
 382                 return -EIO;
 383         return 0;
 384 }
 385
 386 static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
 387                                struct iattr *iattr, int do_trunc)
 388 {
 389         struct inode *inode = dentry->d_inode;
 390         int rc;
 391
 392         lock_24kernel();
 393
 394         /* A _really_ horrible hack to avoid removing the data stored
 395          * in the block pointers; this is really the "small" stripe MD data.
 396          * We can avoid further hackery by virtue of the MDS file size being
 397          * zero all the time (which doesn't invoke block truncate at unlink
 398          * time), so we assert we never change the MDS file size from zero. */
 399         if (iattr->ia_valid & ATTR_SIZE && !do_trunc) {
 400                 /* ATTR_SIZE would invoke truncate: clear it */
 401                 iattr->ia_valid &= ~ATTR_SIZE;
 402                 EXT3_I(inode)->i_disksize = inode->i_size = iattr->ia_size;
 403
 404                 /* make sure _something_ gets set - so new inode
 405                  * goes to disk (probably won't work over XFS */
 406                 if (!(iattr->ia_valid & (ATTR_MODE | ATTR_MTIME | ATTR_CTIME))){
 407                         iattr->ia_valid |= ATTR_MODE;
 408                         iattr->ia_mode = inode->i_mode;
 409                 }
 410         }
 411
 412         /* Don't allow setattr to change file type */
 413         iattr->ia_mode = (inode->i_mode & S_IFMT)|(iattr->ia_mode & ~S_IFMT);
 414
 415         /* We set these flags on the client, but have already checked perms
 416          * so don't confuse inode_change_ok. */
 417         iattr->ia_valid &= ~(ATTR_MTIME_SET | ATTR_ATIME_SET);
 418
 419         if (inode->i_op->setattr) {
 420                 rc = inode->i_op->setattr(dentry, iattr);
 421         } else {
 422                 rc = inode_change_ok(inode, iattr);
 423                 if (!rc)
 424                         rc = inode_setattr(inode, iattr);
 425         }
 426
 427         unlock_24kernel();
 428
 429         return rc;
 430 }
 431
 432 static int fsfilt_ext3_iocontrol(struct inode * inode, struct file *file,
 433                                  unsigned int cmd, unsigned long arg)
 434 {
 435         int rc = 0;
 436         ENTRY;
 437
 438         if (inode->i_fop->ioctl)
 439                 rc = inode->i_fop->ioctl(inode, file, cmd, arg);
 440         else
 441                 RETURN(-ENOTTY);
 442
 443         RETURN(rc);
 444 }
 445
 446 static int fsfilt_ext3_set_xattr(struct inode * inode, void *handle, char *name,
 447                                  void *buffer, int buffer_size)
 448 {
 449         int rc = 0;
 450
 451         lock_24kernel();
 452
 453         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
 454                                    name, buffer, buffer_size, 0);
 455         unlock_24kernel();
 456         if (rc)
 457                 CERROR("set xattr %s from inode %lu: rc %d\n",
 458                        name,  inode->i_ino, rc);
 459         return rc;
 460 }
 461
 462 static int fsfilt_ext3_get_xattr(struct inode *inode, char *name,
 463                                  void *buffer, int buffer_size)
 464 {
 465         int rc = 0;
 466
 467         lock_24kernel();
 468
 469         rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
 470                             name, buffer, buffer_size);
 471         unlock_24kernel();
 472
 473         if (buffer == NULL)
 474                 return (rc == -ENODATA) ? 0 : rc;
 475         if (rc < 0) {
 476                 CDEBUG(D_INFO, "error getting EA %s from inode %lu: rc %d\n",
 477                        name,  inode->i_ino, rc);
 478                 memset(buffer, 0, buffer_size);
 479                 return (rc == -ENODATA) ? 0 : rc;
 480         }
 481
 482         return rc;
 483 }
 484
 485 static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
 486                               void *lmm, int lmm_size,
 487                               enum ea_type type)
 488 {
 489         int rc;
 490
 491         switch(type) {
 492         case EA_LOV:
 493                 rc = fsfilt_ext3_set_xattr(inode, handle,
 494                                            XATTR_LUSTRE_MDS_LOV_EA,
 495                                            lmm, lmm_size);
 496                 break;
 497         case EA_MEA:
 498                 rc = fsfilt_ext3_set_xattr(inode, handle,
 499                                            XATTR_LUSTRE_MDS_MEA_EA,
 500                                            lmm, lmm_size);
 501                 break;
 502         case EA_SID:
 503                 rc = fsfilt_ext3_set_xattr(inode, handle,
 504                                            XATTR_LUSTRE_MDS_SID_EA,
 505                                            lmm, lmm_size);
 506                 break;
 507         case EA_PID:
 508                  rc = fsfilt_ext3_set_xattr(inode, handle,
 509                                             XATTR_LUSTRE_MDS_PID_EA,
 510                                             lmm, lmm_size);
 511                 break;
 512         case EA_KEY:
 513                 rc = fsfilt_ext3_set_xattr(inode, handle,
 514                                            XATTR_LUSTRE_MDS_KEY_EA,
 515                                            lmm, lmm_size);
 516                 break;
 517         default:
 518                 return -EINVAL;
 519         }
 520
 521         return rc;
 522 }
 523
 524 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm,
 525                               int lmm_size, enum ea_type type)
 526 {
 527         int rc;
 528
 529         switch (type) {
 530         case EA_LOV:
 531                 rc = fsfilt_ext3_get_xattr(inode,
 532                                            XATTR_LUSTRE_MDS_LOV_EA,
 533                                            lmm, lmm_size);
 534                 break;
 535         case EA_MEA:
 536                 rc = fsfilt_ext3_get_xattr(inode,
 537                                            XATTR_LUSTRE_MDS_MEA_EA,
 538                                            lmm, lmm_size);
 539                 break;
 540         case EA_SID:
 541                 rc = fsfilt_ext3_get_xattr(inode,
 542                                            XATTR_LUSTRE_MDS_SID_EA,
 543                                            lmm, lmm_size);
 544                 break;
 545         case EA_PID:
 546                 rc = fsfilt_ext3_get_xattr(inode,
 547                                            XATTR_LUSTRE_MDS_PID_EA,
 548                                            lmm, lmm_size);
 549                 break;
 550         case EA_KEY:
 551                 rc = fsfilt_ext3_get_xattr(inode,
 552                                            XATTR_LUSTRE_MDS_KEY_EA,
 553                                            lmm, lmm_size);
 554                 break;
 555         default:
 556                 return -EINVAL;
 557         }
 558
 559         return rc;
 560 }
 561
 562 static int fsfilt_ext3_send_bio(int rw, struct inode *inode, void *bio)
 563 {
 564         int rc = 0;
 565 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 566         submit_bio(rw, (struct bio *)bio);
 567 #else
 568         struct bio *b = (struct kiobuf *)bio;
 569         int blocks_per_page;
 570
 571         rc = brw_kiovec(rw, 1, &b, inode->i_dev,
 572                         b->blocks, 1 << inode->i_blkbits);
 573
 574         blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 575
 576         if (rc != (1 << inode->i_blkbits) * b->nr_pages * blocks_per_page) {
 577                 CERROR("short write?  expected %d, wrote %d\n",
 578                        (1 << inode->i_blkbits) * b->nr_pages *
 579                        blocks_per_page, rc);
 580         }
 581 #endif
 582         return rc;
 583 }
 584
 585 static struct page *fsfilt_ext3_getpage(struct inode *inode, long int index)
 586 {
 587         int rc;
 588         struct page *page;
 589
 590         page = grab_cache_page(inode->i_mapping, index);
 591         if (page == NULL)
 592                 return ERR_PTR(-ENOMEM);
 593
 594         if (PageUptodate(page)) {
 595                 unlock_page(page);
 596                 return page;
 597         }
 598
 599         rc = inode->i_mapping->a_ops->readpage(NULL, page);
 600         if (rc < 0) {
 601                 page_cache_release(page);
 602                 return ERR_PTR(rc);
 603         }
 604
 605         return page;
 606 }
 607
 608 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
 609                                     loff_t *off)
 610 {
 611         struct inode *inode = file->f_dentry->d_inode;
 612         int rc = 0;
 613
 614         if (S_ISREG(inode->i_mode))
 615                 rc = file->f_op->read(file, buf, count, off);
 616         else {
 617                 const int blkbits = inode->i_sb->s_blocksize_bits;
 618                 const int blksize = inode->i_sb->s_blocksize;
 619
 620                 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
 621                        count, inode->i_ino, *off);
 622                 while (count > 0) {
 623                         struct buffer_head *bh;
 624
 625                         bh = NULL;
 626                         if (*off < inode->i_size) {
 627                                 int err = 0;
 628
 629                                 bh = ext3_bread(NULL, inode, *off >> blkbits,
 630                                                 0, &err);
 631
 632                                 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
 633
 634                                 if (bh) {
 635                                         memcpy(buf, bh->b_data, blksize);
 636                                         brelse(bh);
 637                                 } else if (err) {
 638                                         /* XXX in theory we should just fake
 639                                          * this buffer and continue like ext3,
 640                                          * especially if this is a partial read
 641                                          */
 642                                         CERROR("error read dir %lu+%llu: %d\n",
 643                                                inode->i_ino, *off, err);
 644                                         RETURN(err);
 645                                 }
 646                         }
 647                         if (!bh) {
 648                                 struct ext3_dir_entry_2 *fake = (void *)buf;
 649
 650                                 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
 651                                 memset(fake, 0, sizeof(*fake));
 652                                 fake->rec_len = cpu_to_le32(blksize);
 653                         }
 654                         count -= blksize;
 655                         buf += blksize;
 656                         *off += blksize;
 657                         rc += blksize;
 658                 }
 659         }
 660
 661         return rc;
 662 }
 663
 664 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
 665 {
 666         struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
 667
 668         fcb->cb_func(fcb->cb_obd, fcb->cb_last_num, fcb->cb_data, error);
 669
 670         OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
 671         atomic_dec(&fcb_cache_count);
 672 }
 673
 674 static int fsfilt_ext3_add_journal_cb(struct obd_device *obd,
 675                                       struct super_block *sb,
 676                                       __u64 last_num, void *handle,
 677                                       fsfilt_cb_t cb_func,
 678                                       void *cb_data)
 679 {
 680         struct fsfilt_cb_data *fcb;
 681
 682         OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
 683         if (fcb == NULL)
 684                 RETURN(-ENOMEM);
 685
 686         atomic_inc(&fcb_cache_count);
 687         fcb->cb_func = cb_func;
 688         fcb->cb_obd = obd;
 689         fcb->cb_last_num = last_num;
 690         fcb->cb_data = cb_data;
 691
 692         CDEBUG(D_EXT2, "set callback for last_num: "LPD64"\n", last_num);
 693         lock_24kernel();
 694         journal_callback_set(handle, fsfilt_ext3_cb_func,
 695                              (struct journal_callback *)fcb);
 696         unlock_24kernel();
 697         return 0;
 698 }
 699
 700 /*
 701  * We need to hack the return value for the free inode counts because
 702  * the current EA code requires one filesystem block per inode with EAs,
 703  * so it is possible to run out of blocks before we run out of inodes.
 704  *
 705  * This can be removed when the ext3 EA code is fixed.
 706  */
 707 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
 708 {
 709         struct kstatfs sfs;
 710         int rc;
 711
 712         memset(&sfs, 0, sizeof(sfs));
 713
 714         rc = sb->s_op->statfs(sb, &sfs);
 715
 716         if (!rc && sfs.f_bfree < sfs.f_ffree) {
 717                 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
 718                 sfs.f_ffree = sfs.f_bfree;
 719         }
 720
 721         statfs_pack(osfs, &sfs);
 722         return rc;
 723 }
 724
 725 static int fsfilt_ext3_sync(struct super_block *sb)
 726 {
 727         return ext3_force_commit(sb);
 728 }
 729
 730 #if defined(EXT3_MULTIBLOCK_ALLOCATOR) && (!defined(EXT3_EXT_CACHE_NO) || defined(EXT_CACHE_MARK))
 731 #warning "kernel code has old extents/mballoc patch, disabling"
 732 #undef EXT3_MULTIBLOCK_ALLOCATOR
 733 #endif
 734 #ifndef EXT3_EXTENTS_FL
 735 #define EXT3_EXTENTS_FL                 0x00080000 /* Inode uses extents */
 736 #endif
 737
 738 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
 739 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 740 #define ext3_up_truncate_sem(inode)  up_write(&EXT3_I(inode)->truncate_sem);
 741 #define ext3_down_truncate_sem(inode)  down_write(&EXT3_I(inode)->truncate_sem);
 742 #else
 743 #define ext3_up_truncate_sem(inode)  up(&EXT3_I(inode)->truncate_sem);
 744 #define ext3_down_truncate_sem(inode)  down(&EXT3_I(inode)->truncate_sem);
 745 #endif
 746
 747 #include <linux/lustre_version.h>
 748 #if EXT3_EXT_MAGIC == 0xf301
 749 #define ee_start e_start
 750 #define ee_block e_block
 751 #define ee_len   e_num
 752 #endif
 753 #ifndef EXT3_BB_MAX_BLOCKS
 754 #define ext3_mb_new_blocks(handle, inode, goal, count, aflags, err) \
 755         ext3_new_blocks(handle, inode, count, goal, err)
 756 #endif
 757
 758 struct bpointers {
 759         unsigned long *blocks;
 760         int *created;
 761         unsigned long start;
 762         int num;
 763         int init_num;
 764         int create;
 765 };
 766
 767 static int ext3_ext_find_goal(struct inode *inode, struct ext3_ext_path *path,
 768                               unsigned long block, int *aflags)
 769 {
 770         struct ext3_inode_info *ei = EXT3_I(inode);
 771         unsigned long bg_start;
 772         unsigned long colour;
 773         int depth;
 774
 775         if (path) {
 776                 struct ext3_extent *ex;
 777                 depth = path->p_depth;
 778
 779                 /* try to predict block placement */
 780                 if ((ex = path[depth].p_ext)) {
 781 #if 0
 782                         /* This prefers to eat into a contiguous extent
 783                          * rather than find an extent that the whole
 784                          * request will fit into.  This can fragment data
 785                          * block allocation and prevents our lovely 1M I/Os
 786                          * from reaching the disk intact. */
 787                         if (ex->ee_block + ex->ee_len == block)
 788                                 *aflags |= 1;
 789 #endif
 790                         return ex->ee_start + (block - ex->ee_block);
 791                 }
 792
 793                 /* it looks index is empty
 794                  * try to find starting from index itself */
 795                 if (path[depth].p_bh)
 796                         return path[depth].p_bh->b_blocknr;
 797         }
 798
 799         /* OK. use inode's group */
 800         bg_start = (ei->i_block_group * EXT3_BLOCKS_PER_GROUP(inode->i_sb)) +
 801                 le32_to_cpu(EXT3_SB(inode->i_sb)->s_es->s_first_data_block);
 802         colour = (current->pid % 16) *
 803                 (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 804         return bg_start + colour + block;
 805 }
 806
 807 static int ext3_ext_new_extent_cb(struct ext3_extents_tree *tree,
 808                                   struct ext3_ext_path *path,
 809                                   struct ext3_ext_cache *cex)
 810 {
 811         struct inode *inode = tree->inode;
 812         struct bpointers *bp = tree->private;
 813         struct ext3_extent nex;
 814         int count, err, goal;
 815         unsigned long pblock;
 816         loff_t new_i_size;
 817         handle_t *handle;
 818         int i, aflags = 0;
 819
 820         i = EXT_DEPTH(tree);
 821         EXT_ASSERT(i == path->p_depth);
 822         EXT_ASSERT(path[i].p_hdr);
 823
 824         if (cex->ec_type == EXT3_EXT_CACHE_EXTENT) {
 825                 err = EXT_CONTINUE;
 826                 goto map;
 827         }
 828
 829         if (bp->create == 0) {
 830                 i = 0;
 831                 if (cex->ec_block < bp->start)
 832                         i = bp->start - cex->ec_block;
 833                 if (i >= cex->ec_len)
 834                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 835                                         i, cex->ec_len);
 836                 for (; i < cex->ec_len && bp->num; i++) {
 837                         *(bp->created) = 0;
 838                         bp->created++;
 839                         *(bp->blocks) = 0;
 840                         bp->blocks++;
 841                         bp->num--;
 842                         bp->start++;
 843                 }
 844
 845                 return EXT_CONTINUE;
 846         }
 847
 848         count = ext3_ext_calc_credits_for_insert(tree, path);
 849
 850         lock_24kernel();
 851         handle = journal_start(EXT3_JOURNAL(inode), count+EXT3_ALLOC_NEEDED+1);
 852         unlock_24kernel();
 853         if (IS_ERR(handle))
 854                 return PTR_ERR(handle);
 855
 856         count = cex->ec_len;
 857         goal = ext3_ext_find_goal(inode, path, cex->ec_block, &aflags);
 858         aflags |= 2; /* block have been already reserved */
 859         lock_24kernel();
 860         pblock = ext3_mb_new_blocks(handle, inode, goal, &count, aflags, &err);
 861         unlock_24kernel();
 862         if (!pblock)
 863                 goto out;
 864         EXT_ASSERT(count <= cex->ec_len);
 865         if (count < cex->ec_len)
 866                 CDEBUG(D_ERROR, "%d < %lu\n", count, cex->ec_len);
 867
 868         /* insert new extent */
 869         nex.ee_block = cex->ec_block;
 870         nex.ee_start = pblock;
 871         nex.ee_len = count;
 872         err = ext3_ext_insert_extent(handle, tree, path, &nex);
 873         if (err)
 874                 goto out;
 875
 876         /*
 877          * Putting len of the actual extent we just inserted,
 878          * we are asking ext3_ext_walk_space() to continue
 879          * scaning after that block
 880          */
 881         cex->ec_len = nex.ee_len;
 882         cex->ec_start = nex.ee_start;
 883         BUG_ON(nex.ee_len == 0);
 884         BUG_ON(nex.ee_block != cex->ec_block);
 885
 886         /* correct on-disk inode size */
 887         if (nex.ee_len > 0) {
 888                 new_i_size = (loff_t) nex.ee_block + nex.ee_len;
 889                 new_i_size = new_i_size << inode->i_blkbits;
 890                 if (new_i_size > EXT3_I(inode)->i_disksize) {
 891                         EXT3_I(inode)->i_disksize = new_i_size;
 892                         err = ext3_mark_inode_dirty(handle, inode);
 893                 }
 894         }
 895
 896 out:
 897         lock_24kernel();
 898         journal_stop(handle);
 899         unlock_24kernel();
 900 map:
 901         if (err >= 0) {
 902                 struct block_device *bdev = inode->i_sb->s_bdev;
 903
 904                 /* map blocks */
 905                 if (bp->num == 0) {
 906                         CERROR("hmm. why do we find this extent?\n");
 907                         CERROR("initial space: %lu:%u\n",
 908                                 bp->start, bp->init_num);
 909                         CERROR("current extent: %u/%u/%u %d\n",
 910                                 cex->ec_block, cex->ec_len,
 911                                 cex->ec_start, cex->ec_type);
 912                 }
 913                 i = 0;
 914                 if (cex->ec_block < bp->start)
 915                         i = bp->start - cex->ec_block;
 916                 if (i >= cex->ec_len)
 917                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 918                                         i, cex->ec_len);
 919                 for (; i < cex->ec_len && bp->num; i++) {
 920                         *(bp->blocks) = cex->ec_start + i;
 921                         if (cex->ec_type == EXT3_EXT_CACHE_EXTENT) {
 922                                 *(bp->created) = 0;
 923                         } else {
 924                                 *(bp->created) = 1;
 925                                 /* unmap any possible underlying metadata from
 926                                  * the block device mapping.  bug 6998.
 927                                  * This only compiles on 2.6, but there are
 928                                  * no users of mballoc on 2.4. */
 929                                 unmap_underlying_metadata(bdev, *(bp->blocks));
 930                         }
 931                         bp->created++;
 932                         bp->blocks++;
 933                         bp->num--;
 934                         bp->start++;
 935                 }
 936         }
 937         return err;
 938 }
 939
 940 int fsfilt_map_nblocks(struct inode *inode, unsigned long block,
 941                        unsigned long num, unsigned long *blocks,
 942                        int *created, int create)
 943 {
 944         struct ext3_extents_tree tree;
 945         struct bpointers bp;
 946         int err;
 947
 948         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
 949                 block, block + num, (unsigned) inode->i_ino);
 950
 951         ext3_init_tree_desc(&tree, inode);
 952         tree.private = &bp;
 953         bp.blocks = blocks;
 954         bp.created = created;
 955         bp.start = block;
 956         bp.init_num = bp.num = num;
 957         bp.create = create;
 958
 959         err = ext3_ext_walk_space(&tree, block, num, ext3_ext_new_extent_cb);
 960         ext3_ext_invalidate_cache(&tree);
 961
 962         return err;
 963 }
 964
 965 int fsfilt_ext3_map_ext_inode_pages(struct inode *inode, struct page **page,
 966                                     int pages, unsigned long *blocks,
 967                                     int *created, int create)
 968 {
 969         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 970         int rc = 0, i = 0;
 971         struct page *fp = NULL;
 972         int clen = 0;
 973
 974         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 975                 inode->i_ino, pages, (*page)->index);
 976
 977         /* pages are sorted already. so, we just have to find
 978          * contig. space and process them properly */
 979         while (i < pages) {
 980                 if (fp == NULL) {
 981                         /* start new extent */
 982                         fp = *page++;
 983                         clen = 1;
 984                         i++;
 985                         continue;
 986                 } else if (fp->index + clen == (*page)->index) {
 987                         /* continue the extent */
 988                         page++;
 989                         clen++;
 990                         i++;
 991                         continue;
 992                 }
 993
 994                 /* process found extent */
 995                 rc = fsfilt_map_nblocks(inode, fp->index * blocks_per_page,
 996                                         clen * blocks_per_page, blocks,
 997                                         created, create);
 998                 if (rc)
 999                         GOTO(cleanup, rc);
1000
1001                 /* look for next extent */
1002                 fp = NULL;
1003                 blocks += blocks_per_page * clen;
1004                 created += blocks_per_page * clen;
1005         }
1006
1007         if (fp)
1008                 rc = fsfilt_map_nblocks(inode, fp->index * blocks_per_page,
1009                                         clen * blocks_per_page, blocks,
1010                                         created, create);
1011 cleanup:
1012         return rc;
1013 }
1014 #endif
1015
1016 extern int ext3_map_inode_page(struct inode *inode, struct page *page,
1017                                unsigned long *blocks, int *created, int create);
1018 int fsfilt_ext3_map_bm_inode_pages(struct inode *inode, struct page **page,
1019                                    int pages, unsigned long *blocks,
1020                                    int *created, int create)
1021 {
1022         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1023         unsigned long *b;
1024         int rc = 0, i, *cr;
1025
1026         for (i = 0, cr = created, b = blocks; i < pages; i++, page++) {
1027                 rc = ext3_map_inode_page(inode, *page, b, cr, create);
1028                 if (rc) {
1029                         CERROR("ino %lu, blk %lu cr %u create %d: rc %d\n",
1030                                inode->i_ino, *b, *cr, create, rc);
1031                         break;
1032                 }
1033
1034                 b += blocks_per_page;
1035                 cr += blocks_per_page;
1036         }
1037         return rc;
1038 }
1039
1040 int fsfilt_ext3_map_inode_pages(struct inode *inode, struct page **page,
1041                                 int pages, unsigned long *blocks,
1042                                 int *created, int create,
1043                                 struct semaphore *optional_sem)
1044 {
1045         int rc = 0;
1046         if (optional_sem != NULL)
1047                 down(optional_sem);
1048 #ifdef EXT3_MULTIBLOCK_ALLOCATOR
1049         if (EXT3_I(inode)->i_flags & EXT3_EXTENTS_FL)
1050                 rc = fsfilt_ext3_map_ext_inode_pages(inode, page, pages,
1051                                                      blocks, created, create);
1052         else
1053 #endif
1054                 rc = fsfilt_ext3_map_bm_inode_pages(inode, page, pages, blocks,
1055                                                     created, create);
1056         if (optional_sem != NULL)
1057                 up(optional_sem);
1058
1059         return rc;
1060 }
1061
1062 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
1063                                int nblocks, loff_t newsize);
1064 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
1065                                       int nblocks, loff_t newsize)
1066 {
1067         return ext3_prep_san_write(inode, blocks, nblocks, newsize);
1068 }
1069
1070 static int fsfilt_ext3_read_record(struct file * file, void *buf,
1071                                    int size, loff_t *offs)
1072 {
1073         struct inode *inode = file->f_dentry->d_inode;
1074         unsigned long block;
1075         struct buffer_head *bh;
1076         int err, blocksize, csize, boffs;
1077
1078         /* prevent reading after eof */
1079         lock_kernel();
1080         if (inode->i_size < *offs + size) {
1081                 size = inode->i_size - *offs;
1082                 unlock_kernel();
1083                 if (size < 0) {
1084                         CERROR("size %llu is too short for read %u@%llu\n",
1085                                inode->i_size, size, *offs);
1086                         return -EIO;
1087                 } else if (size == 0) {
1088                         return 0;
1089                 }
1090         } else {
1091                 unlock_kernel();
1092         }
1093
1094         blocksize = 1 << inode->i_blkbits;
1095
1096         while (size > 0) {
1097                 block = *offs >> inode->i_blkbits;
1098                 boffs = *offs & (blocksize - 1);
1099                 csize = min(blocksize - boffs, size);
1100                 bh = ext3_bread(NULL, inode, block, 0, &err);
1101                 if (!bh) {
1102                         CERROR("can't read block: %d\n", err);
1103                         return err;
1104                 }
1105
1106                 memcpy(buf, bh->b_data + boffs, csize);
1107                 brelse(bh);
1108
1109                 *offs += csize;
1110                 buf += csize;
1111                 size -= csize;
1112         }
1113         return 0;
1114 }
1115
1116 static int fsfilt_ext3_write_record(struct file *file, void *buf, int bufsize,
1117                                     loff_t *offs, int force_sync)
1118 {
1119         struct buffer_head *bh = NULL;
1120         unsigned long block;
1121         struct inode *inode = file->f_dentry->d_inode;
1122         loff_t old_size = inode->i_size, offset = *offs;
1123         loff_t new_size = inode->i_size;
1124         journal_t *journal;
1125         handle_t *handle;
1126         int err, block_count = 0, blocksize, size, boffs;
1127
1128         /* Determine how many transaction credits are needed */
1129         blocksize = 1 << inode->i_blkbits;
1130         block_count = (*offs & (blocksize - 1)) + bufsize;
1131         block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
1132
1133         journal = EXT3_SB(inode->i_sb)->s_journal;
1134         lock_24kernel();
1135         handle = journal_start(journal,
1136                                block_count * EXT3_DATA_TRANS_BLOCKS + 2);
1137         unlock_24kernel();
1138         if (IS_ERR(handle)) {
1139                 CERROR("can't start transaction\n");
1140                 return PTR_ERR(handle);
1141         }
1142
1143         while (bufsize > 0) {
1144                 if (bh != NULL)
1145                         brelse(bh);
1146
1147                 block = offset >> inode->i_blkbits;
1148                 boffs = offset & (blocksize - 1);
1149                 size = min(blocksize - boffs, bufsize);
1150                 bh = ext3_bread(handle, inode, block, 1, &err);
1151                 if (!bh) {
1152                         CERROR("can't read/create block: %d\n", err);
1153                         goto out;
1154                 }
1155
1156                 err = ext3_journal_get_write_access(handle, bh);
1157                 if (err) {
1158                         CERROR("journal_get_write_access() returned error %d\n",
1159                                err);
1160                         goto out;
1161                 }
1162                 LASSERT(bh->b_data + boffs + size <= bh->b_data + bh->b_size);
1163                 memcpy(bh->b_data + boffs, buf, size);
1164                 err = ext3_journal_dirty_metadata(handle, bh);
1165                 if (err) {
1166                         CERROR("journal_dirty_metadata() returned error %d\n",
1167                                err);
1168                         goto out;
1169                 }
1170                 if (offset + size > new_size)
1171                         new_size = offset + size;
1172                 offset += size;
1173                 bufsize -= size;
1174                 buf += size;
1175         }
1176
1177         if (force_sync)
1178                 handle->h_sync = 1; /* recovery likes this */
1179 out:
1180         if (bh)
1181                 brelse(bh);
1182
1183         /* correct in-core and on-disk sizes */
1184         if (new_size > inode->i_size) {
1185                 lock_kernel();
1186                 if (new_size > inode->i_size)
1187                         inode->i_size = new_size;
1188                 if (inode->i_size > EXT3_I(inode)->i_disksize)
1189                         EXT3_I(inode)->i_disksize = inode->i_size;
1190                 if (inode->i_size > old_size)
1191                         mark_inode_dirty(inode);
1192                 unlock_kernel();
1193         }
1194
1195         lock_24kernel();
1196         journal_stop(handle);
1197         unlock_24kernel();
1198
1199         if (err == 0)
1200                 *offs = offset;
1201         return err;
1202 }
1203
1204 static int fsfilt_ext3_setup(struct obd_device *obd, struct super_block *sb)
1205 {
1206 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1207         struct mds_obd *mds = &obd->u.mds;
1208 #endif
1209 #if 0
1210         EXT3_SB(sb)->dx_lock = fsfilt_ext3_dx_lock;
1211         EXT3_SB(sb)->dx_unlock = fsfilt_ext3_dx_unlock;
1212 #endif
1213 #ifdef S_PDIROPS
1214         CWARN("Enabling PDIROPS\n");
1215         set_opt(EXT3_SB(sb)->s_mount_opt, PDIROPS);
1216         sb->s_flags |= S_PDIROPS;
1217 #endif
1218         /* setup mdsnum in underlying fs */
1219 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1220         if (mds->mds_md_obd) {
1221                 struct ext3_sb_info *sbi = EXT3_SB(sb);
1222                 struct ext3_super_block *es = sbi->s_es;
1223                 handle_t *handle;
1224                 int err;
1225
1226                 if (!EXT3_HAS_INCOMPAT_FEATURE(sb, EXT3_FEATURE_INCOMPAT_MDSNUM)) {
1227                         CWARN("%s: set mdsnum %d in ext3\n",
1228                               obd->obd_name, mds->mds_num);
1229                         lock_24kernel();
1230                         handle = journal_start(sbi->s_journal, 1);
1231                         unlock_24kernel();
1232                         LASSERT(!IS_ERR(handle));
1233                         err = ext3_journal_get_write_access(handle, sbi->s_sbh);
1234                         LASSERT(err == 0);
1235                         EXT3_SET_INCOMPAT_FEATURE(sb,
1236                                                 EXT3_FEATURE_INCOMPAT_MDSNUM);
1237                         es->s_mdsnum = mds->mds_num;
1238                         err = ext3_journal_dirty_metadata(handle, sbi->s_sbh);
1239                         LASSERT(err == 0);
1240                         lock_24kernel();
1241                         journal_stop(handle);
1242                         unlock_24kernel();
1243                 } else {
1244                         CWARN("%s: mdsnum initialized to %u in ext3fs\n",
1245                                 obd->obd_name, es->s_mdsnum);
1246                 }
1247                 sbi->s_mdsnum = es->s_mdsnum;
1248         }
1249 #endif
1250         return 0;
1251 }
1252
1253 extern int ext3_add_dir_entry(struct dentry *dentry);
1254 extern int ext3_del_dir_entry(struct dentry *dentry);
1255
1256 static int fsfilt_ext3_add_dir_entry(struct obd_device *obd,
1257                                      struct dentry *parent,
1258                                      char *name, int namelen,
1259                                      unsigned long ino,
1260                                      unsigned long generation,
1261                                      unsigned long mds,
1262                                      unsigned long fid)
1263 {
1264 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1265         struct dentry *dentry;
1266         int err;
1267         LASSERT(ino != 0);
1268         LASSERT(namelen != 0);
1269         dentry = ll_lookup_one_len(name, parent, namelen);
1270         if (IS_ERR(dentry)) {
1271                 CERROR("can't lookup %*s in %lu/%lu: %d\n", dentry->d_name.len,
1272                        dentry->d_name.name, dentry->d_inode->i_ino,
1273                        (unsigned long) dentry->d_inode->i_generation,
1274                        (int) PTR_ERR(dentry));
1275                 RETURN(PTR_ERR(dentry));
1276         }
1277         if (dentry->d_inode != NULL || dentry->d_flags & DCACHE_CROSS_REF) {
1278                 CERROR("dentry %*s(0x%p) found\n", dentry->d_name.len,
1279                        dentry->d_name.name, dentry);
1280                 l_dput(dentry);
1281                 RETURN(-EEXIST);
1282         }
1283
1284         /* mds_reint_rename() may use this method to add dir entry
1285          * that points onto local inode. and we don't want to find
1286          * it cross-ref by subsequent lookups */
1287         d_drop(dentry);
1288
1289         dentry->d_flags |= DCACHE_CROSS_REF;
1290         dentry->d_inum = ino;
1291         dentry->d_mdsnum = mds;
1292         dentry->d_generation = generation;
1293         dentry->d_fid = fid;
1294         lock_24kernel();
1295         err = ext3_add_dir_entry(dentry);
1296         unlock_24kernel();
1297
1298         l_dput(dentry);
1299
1300         RETURN(err);
1301 #else
1302 #error "rebuild kernel and lustre with ext3-mds-num patch!"
1303         LASSERT(0);
1304 #endif
1305 }
1306
1307 static int fsfilt_ext3_del_dir_entry(struct obd_device *obd,
1308                                  struct dentry *dentry)
1309 {
1310 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
1311         int err;
1312         lock_24kernel();
1313         err = ext3_del_dir_entry(dentry);
1314         unlock_24kernel();
1315         if (err == 0)
1316                 d_drop(dentry);
1317         return err;
1318 #else
1319 #error "rebuild kernel and lustre with ext3-mds-num patch!"
1320         LASSERT(0);
1321 #endif
1322 }
1323
1324 /* If fso is NULL, op is FSFILT operation, otherwise op is number of fso
1325    objects. Logs is number of logfiles to update */
1326 static int fsfilt_ext3_get_op_len(int op, struct fsfilt_objinfo *fso, int logs)
1327 {
1328         if ( !fso ) {
1329                 switch(op) {
1330                 case FSFILT_OP_CREATE:
1331                                  /* directory leaf, index & indirect & EA*/
1332                         return 4 + 3 * logs;
1333                 case FSFILT_OP_UNLINK:
1334                         return 3 * logs;
1335                 }
1336         } else {
1337                 int i;
1338                 int needed = 0;
1339                 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
1340                 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
1341                 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
1342                 for (i = 0; i < op; i++, fso++) {
1343                         int nblocks = fso->fso_bufcnt * blockpp;
1344                         int ndindirect = min(nblocks, addrpp + 1);
1345                         int nindir = nblocks + ndindirect + 1;
1346
1347                         needed += nindir;
1348                 }
1349                 return needed + 3 * logs;
1350         }
1351
1352         return 0;
1353 }
1354
1355
1356 #if 0
1357 #define EXTENTS_EA "write_extents"
1358 #define EXTENTS_EA_SIZE 64
1359
1360 int ext3_ext_in_ea_alloc_space(struct inode *, int, const char *, unsigned long, unsigned long);
1361 int ext3_ext_in_ea_remove_space(struct inode *, int, const char *, unsigned long, unsigned long);
1362 int ext3_ext_in_ea_get_extents(struct inode *, int, const char *, char **, int *);
1363 int ext3_ext_in_ea_get_extents_num(struct inode *, int, const char *, int *);
1364
1365 static int fsfilt_ext3_insert_extents_ea(struct inode *inode,
1366                                       unsigned long from,
1367                                       unsigned long num)
1368 {
1369         int rc = 0;
1370
1371         rc = ext3_ext_in_ea_alloc_space(inode, EXT3_XATTR_INDEX_TRUSTED,
1372                                         EXTENTS_EA, from, num);
1373         return rc;
1374 }
1375
1376 static int fsfilt_ext3_remove_extents_ea(struct inode *inode,
1377                                          unsigned long from,
1378                                          unsigned long num)
1379 {
1380         int rc = 0;
1381
1382         rc = ext3_ext_in_ea_remove_space(inode, EXT3_XATTR_INDEX_TRUSTED,
1383                                          EXTENTS_EA, from, num);
1384         return rc;
1385 }
1386
1387 extern int ext3_init_tree_in_ea(struct inode *inode, int name_index,
1388                                 const char *eaname, int size);
1389
1390 static int fsfilt_ext3_init_extents_ea(struct inode *inode)
1391 {
1392         int rc = 0;
1393
1394         rc = ext3_init_tree_in_ea(inode, EXT3_XATTR_INDEX_TRUSTED,
1395                                   EXTENTS_EA, 64);
1396         return rc;
1397 }
1398
1399 static int fsfilt_ext3_get_inode_write_extents(struct inode *inode,
1400                                          char **pbuf, int *size)
1401 {
1402         int rc = 0;
1403
1404         rc = ext3_ext_in_ea_get_extents(inode, EXT3_XATTR_INDEX_TRUSTED,
1405                                         EXTENTS_EA,  pbuf, size);
1406         return rc;
1407 }
1408
1409 static int fsfilt_ext3_get_write_extents_num(struct inode *inode, int *size)
1410 {
1411         int rc = 0;
1412
1413         rc = ext3_ext_in_ea_get_extents_num(inode, EXT3_XATTR_INDEX_TRUSTED,
1414                                             EXTENTS_EA, size);
1415         return rc;
1416 }
1417 #endif
1418
1419 static struct fsfilt_operations fsfilt_ext3_ops = {
1420         .fs_type                    = "ext3",
1421         .fs_owner                   = THIS_MODULE,
1422         .fs_start                   = fsfilt_ext3_start,
1423         .fs_brw_start               = fsfilt_ext3_brw_start,
1424         .fs_commit                  = fsfilt_ext3_commit,
1425         .fs_commit_async            = fsfilt_ext3_commit_async,
1426         .fs_commit_wait             = fsfilt_ext3_commit_wait,
1427         .fs_setattr                 = fsfilt_ext3_setattr,
1428         .fs_iocontrol               = fsfilt_ext3_iocontrol,
1429         .fs_set_md                  = fsfilt_ext3_set_md,
1430         .fs_get_md                  = fsfilt_ext3_get_md,
1431         .fs_readpage                = fsfilt_ext3_readpage,
1432         .fs_add_journal_cb          = fsfilt_ext3_add_journal_cb,
1433         .fs_statfs                  = fsfilt_ext3_statfs,
1434         .fs_sync                    = fsfilt_ext3_sync,
1435         .fs_map_inode_pages         = fsfilt_ext3_map_inode_pages,
1436         .fs_prep_san_write          = fsfilt_ext3_prep_san_write,
1437         .fs_write_record            = fsfilt_ext3_write_record,
1438         .fs_read_record             = fsfilt_ext3_read_record,
1439         .fs_setup                   = fsfilt_ext3_setup,
1440         .fs_getpage                 = fsfilt_ext3_getpage,
1441         .fs_send_bio                = fsfilt_ext3_send_bio,
1442         .fs_set_xattr               = fsfilt_ext3_set_xattr,
1443         .fs_get_xattr               = fsfilt_ext3_get_xattr,
1444         .fs_get_op_len              = fsfilt_ext3_get_op_len,
1445         .fs_add_dir_entry           = fsfilt_ext3_add_dir_entry,
1446         .fs_del_dir_entry           = fsfilt_ext3_del_dir_entry,
1447 #if 0
1448         .fs_init_extents_ea         = fsfilt_ext3_init_extents_ea,
1449         .fs_insert_extents_ea       = fsfilt_ext3_insert_extents_ea,
1450         .fs_remove_extents_ea       = fsfilt_ext3_remove_extents_ea,
1451         .fs_get_inode_write_extents = fsfilt_ext3_get_inode_write_extents,
1452         .fs_get_write_extents_num   = fsfilt_ext3_get_write_extents_num,
1453 #endif
1454 };
1455
1456 static int __init fsfilt_ext3_init(void)
1457 {
1458         int rc;
1459
1460         fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
1461                                       sizeof(struct fsfilt_cb_data), 0,
1462                                       0, NULL, NULL);
1463         if (!fcb_cache) {
1464                 CERROR("error allocating fsfilt journal callback cache\n");
1465                 GOTO(out, rc = -ENOMEM);
1466         }
1467
1468         rc = fsfilt_register_ops(&fsfilt_ext3_ops);
1469
1470         if (rc)
1471                 kmem_cache_destroy(fcb_cache);
1472 out:
1473         return rc;
1474 }
1475
1476 static void __exit fsfilt_ext3_exit(void)
1477 {
1478         fsfilt_unregister_ops(&fsfilt_ext3_ops);
1479         LASSERTF(kmem_cache_destroy(fcb_cache) == 0,
1480                  "can't free fsfilt callback cache: count %d\n",
1481                  atomic_read(&fcb_cache_count));
1482 }
1483
1484 module_init(fsfilt_ext3_init);
1485 module_exit(fsfilt_ext3_exit);
1486
1487 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1488 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
1489 MODULE_LICENSE("GPL");