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/fs.h>
  29 #include <linux/jbd.h>
  30 #include <linux/slab.h>
  31 #include <linux/pagemap.h>
  32 #include <linux/quotaops.h>
  33 #include <linux/ext3_fs.h>
  34 #include <linux/ext3_jbd.h>
  35 #include <linux/version.h>
  36 /* XXX ugh */
  37 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
  38  #include <linux/ext3_xattr.h>
  39 #else
  40  #include <linux/../../fs/ext3/xattr.h>
  41 #endif
  42 #include <linux/kp30.h>
  43 #include <linux/lustre_fsfilt.h>
  44 #include <linux/obd.h>
  45 #include <linux/obd_class.h>
  46 #include <linux/module.h>
  47
  48 static kmem_cache_t *fcb_cache;
  49 static atomic_t fcb_cache_count = ATOMIC_INIT(0);
  50
  51 struct fsfilt_cb_data {
  52         struct journal_callback cb_jcb; /* jbd private data - MUST BE FIRST */
  53         fsfilt_cb_t cb_func;            /* MDS/OBD completion function */
  54         struct obd_device *cb_obd;      /* MDS/OBD completion device */
  55         __u64 cb_last_rcvd;             /* MDS/OST last committed operation */
  56         void *cb_data;                  /* MDS/OST completion function data */
  57 };
  58
  59 #ifndef EXT3_XATTR_INDEX_TRUSTED        /* temporary until we hit l28 kernel */
  60 #define EXT3_XATTR_INDEX_TRUSTED        4
  61 #endif
  62 #define XATTR_LUSTRE_MDS_LOV_EA         "lov"
  63
  64 #define EXT3_XATTR_INDEX_LUSTRE         5                         /* old */
  65 #define XATTR_LUSTRE_MDS_OBJID          "system.lustre_mds_objid" /* old */
  66
  67 /*
  68  * We don't currently need any additional blocks for rmdir and
  69  * unlink transactions because we are storing the OST oa_id inside
  70  * the inode (which we will be changing anyways as part of this
  71  * transaction).
  72  */
  73 static void *fsfilt_ext3_start(struct inode *inode, int op, void *desc_private)
  74 {
  75         /* For updates to the last recieved file */
  76         int nblocks = EXT3_DATA_TRANS_BLOCKS;
  77         void *handle;
  78
  79         if (current->journal_info) {
  80                 CDEBUG(D_INODE, "increasing refcount on %p\n",
  81                        current->journal_info);
  82                 goto journal_start;
  83         }
  84
  85         switch(op) {
  86         case FSFILT_OP_CREATE_LOG:
  87                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
  88                 op = FSFILT_OP_CREATE;
  89                 break;
  90         case FSFILT_OP_UNLINK_LOG:
  91                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
  92                 op = FSFILT_OP_UNLINK;
  93                 break;
  94         }
  95
  96         switch(op) {
  97         case FSFILT_OP_RMDIR:
  98         case FSFILT_OP_UNLINK:
  99                 nblocks += EXT3_DELETE_TRANS_BLOCKS;
 100                 break;
 101         case FSFILT_OP_RENAME:
 102                 /* modify additional directory */
 103                 nblocks += EXT3_DATA_TRANS_BLOCKS;
 104                 /* no break */
 105         case FSFILT_OP_SYMLINK:
 106                 /* additional block + block bitmap + GDT for long symlink */
 107                 nblocks += 3;
 108                 /* no break */
 109         case FSFILT_OP_CREATE:
 110         case FSFILT_OP_MKDIR:
 111         case FSFILT_OP_MKNOD:
 112                 /* modify one inode + block bitmap + GDT */
 113                 nblocks += 3;
 114                 /* no break */
 115         case FSFILT_OP_LINK:
 116                 /* modify parent directory */
 117                 nblocks += EXT3_INDEX_EXTRA_TRANS_BLOCKS+EXT3_DATA_TRANS_BLOCKS;
 118                 break;
 119         case FSFILT_OP_SETATTR:
 120                 /* Setattr on inode */
 121                 nblocks += 1;
 122                 break;
 123         case FSFILT_OP_CANCEL_UNLINK_LOG:
 124                 nblocks = (LLOG_CHUNK_SIZE >> inode->i_blkbits) +
 125                         EXT3_DELETE_TRANS_BLOCKS;
 126                 break;
 127         default: CERROR("unknown transaction start op %d\n", op);
 128                  LBUG();
 129         }
 130
 131         LASSERT(current->journal_info == desc_private);
 132
 133  journal_start:
 134         lock_kernel();
 135         handle = journal_start(EXT3_JOURNAL(inode), nblocks);
 136         unlock_kernel();
 137
 138         if (!IS_ERR(handle))
 139                 LASSERT(current->journal_info == handle);
 140         return handle;
 141 }
 142
 143 /*
 144  * Calculate the number of buffer credits needed to write multiple pages in
 145  * a single ext3 transaction.  No, this shouldn't be here, but as yet ext3
 146  * doesn't have a nice API for calculating this sort of thing in advance.
 147  *
 148  * See comment above ext3_writepage_trans_blocks for details.  We assume
 149  * no data journaling is being done, but it does allow for all of the pages
 150  * being non-contiguous.  If we are guaranteed contiguous pages we could
 151  * reduce the number of (d)indirect blocks a lot.
 152  *
 153  * With N blocks per page and P pages, for each inode we have at most:
 154  * N*P indirect
 155  * min(N*P, blocksize/4 + 1) dindirect blocks
 156  * niocount tindirect
 157  *
 158  * For the entire filesystem, we have at most:
 159  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 160  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 161  * objcount inode blocks
 162  * 1 superblock
 163  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 164  *
 165  * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
 166  */
 167 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso,
 168                                       int niocount, struct niobuf_local *nb)
 169 {
 170         struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
 171         __u64 next_indir;
 172         const int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 173         int nbitmaps = 0, ngdblocks;
 174         int needed = objcount + 1; /* inodes + superblock */
 175         int i, j;
 176
 177         for (i = 0, j = 0; i < objcount; i++, fso++) {
 178                 /* two or more dindirect blocks in case we cross boundary */
 179                 int ndind = (long)((nb[j + fso->fso_bufcnt - 1].offset -
 180                                     nb[j].offset) >>
 181                                    sb->s_blocksize_bits) /
 182                         (EXT3_ADDR_PER_BLOCK(sb) * EXT3_ADDR_PER_BLOCK(sb));
 183                 nbitmaps += min(fso->fso_bufcnt, ndind > 0 ? ndind : 2);
 184
 185                 /* leaf, indirect, tindirect blocks for first block */
 186                 nbitmaps += blockpp + 2;
 187
 188                 j += fso->fso_bufcnt;
 189         }
 190
 191         next_indir = nb[0].offset +
 192                 (EXT3_ADDR_PER_BLOCK(sb) << sb->s_blocksize_bits);
 193         for (i = 1; i < niocount; i++) {
 194                 if (nb[i].offset >= next_indir) {
 195                         nbitmaps++;     /* additional indirect */
 196                         next_indir = nb[i].offset +
 197                                 (EXT3_ADDR_PER_BLOCK(sb)<<sb->s_blocksize_bits);
 198                 } else if (nb[i].offset != nb[i - 1].offset + sb->s_blocksize) {
 199                         nbitmaps++;     /* additional indirect */
 200                 }
 201                 nbitmaps += blockpp;    /* each leaf in different group? */
 202         }
 203
 204         ngdblocks = nbitmaps;
 205         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 206                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 207         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 208                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 209
 210         needed += nbitmaps + ngdblocks;
 211
 212         /* last_rcvd update */
 213         needed += EXT3_DATA_TRANS_BLOCKS;
 214
 215 #if defined(CONFIG_QUOTA) && !defined(__x86_64__) /* XXX */
 216         /* We assume that there will be 1 bit set in s_dquot.flags for each
 217          * quota file that is active.  This is at least true for now.
 218          */
 219         needed += hweight32(sb_any_quota_enabled(sb)) *
 220                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 221 #endif
 222
 223         return needed;
 224 }
 225
 226 /* We have to start a huge journal transaction here to hold all of the
 227  * metadata for the pages being written here.  This is necessitated by
 228  * the fact that we do lots of prepare_write operations before we do
 229  * any of the matching commit_write operations, so even if we split
 230  * up to use "smaller" transactions none of them could complete until
 231  * all of them were opened.  By having a single journal transaction,
 232  * we eliminate duplicate reservations for common blocks like the
 233  * superblock and group descriptors or bitmaps.
 234  *
 235  * We will start the transaction here, but each prepare_write will
 236  * add a refcount to the transaction, and each commit_write will
 237  * remove a refcount.  The transaction will be closed when all of
 238  * the pages have been written.
 239  */
 240 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
 241                                    int niocount, struct niobuf_local *nb,
 242                                    void *desc_private)
 243 {
 244         journal_t *journal;
 245         handle_t *handle;
 246         int needed;
 247         ENTRY;
 248
 249         LASSERT(current->journal_info == desc_private);
 250         journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
 251         needed = fsfilt_ext3_credits_needed(objcount, fso, niocount, nb);
 252
 253         /* The number of blocks we could _possibly_ dirty can very large.
 254          * We reduce our request if it is absurd (and we couldn't get that
 255          * many credits for a single handle anyways).
 256          *
 257          * At some point we have to limit the size of I/Os sent at one time,
 258          * increase the size of the journal, or we have to calculate the
 259          * actual journal requirements more carefully by checking all of
 260          * the blocks instead of being maximally pessimistic.  It remains to
 261          * be seen if this is a real problem or not.
 262          */
 263         if (needed > journal->j_max_transaction_buffers) {
 264                 CERROR("want too many journal credits (%d) using %d instead\n",
 265                        needed, journal->j_max_transaction_buffers);
 266                 needed = journal->j_max_transaction_buffers;
 267         }
 268
 269         lock_kernel();
 270         handle = journal_start(journal, needed);
 271         unlock_kernel();
 272         if (IS_ERR(handle)) {
 273                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 274                        PTR_ERR(handle));
 275         } else {
 276                 LASSERT(handle->h_buffer_credits >= needed);
 277                 LASSERT(current->journal_info == handle);
 278         }
 279
 280         RETURN(handle);
 281 }
 282
 283 static int fsfilt_ext3_commit(struct inode *inode, void *h, int force_sync)
 284 {
 285         int rc;
 286         handle_t *handle = h;
 287
 288         LASSERT(current->journal_info == handle);
 289         if (force_sync)
 290                 handle->h_sync = 1; /* recovery likes this */
 291
 292         lock_kernel();
 293         rc = journal_stop(handle);
 294         unlock_kernel();
 295
 296         // LASSERT(current->journal_info == NULL);
 297         return rc;
 298 }
 299
 300 static int fsfilt_ext3_commit_async(struct inode *inode, void *h,
 301                                         void **wait_handle)
 302 {
 303         transaction_t *transaction;
 304         unsigned long tid, rtid;
 305         handle_t *handle = h;
 306         journal_t *journal;
 307         int rc;
 308
 309         LASSERT(current->journal_info == handle);
 310
 311         lock_kernel();
 312         transaction = handle->h_transaction;
 313         journal = transaction->t_journal;
 314         tid = transaction->t_tid;
 315         /* we don't want to be blocked */
 316         handle->h_sync = 0;
 317         rc = journal_stop(handle);
 318         if (rc) {
 319                 CERROR("error while stopping transaction: %d\n", rc);
 320                 unlock_kernel();
 321                 return rc;
 322         }
 323 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 324         rtid = log_start_commit(journal, transaction);
 325         if (rtid != tid)
 326                 CERROR("strange race: %lu != %lu\n",
 327                        (unsigned long) tid, (unsigned long) rtid);
 328 #else
 329         log_start_commit(journal, transaction->t_tid);
 330 #endif
 331         unlock_kernel();
 332
 333         *wait_handle = (void *) tid;
 334         CDEBUG(D_INODE, "commit async: %lu\n", (unsigned long) tid);
 335         return 0;
 336 }
 337
 338 static int fsfilt_ext3_commit_wait(struct inode *inode, void *h)
 339 {
 340         tid_t tid = (tid_t)(long)h;
 341
 342         CDEBUG(D_INODE, "commit wait: %lu\n", (unsigned long) tid);
 343         if (is_journal_aborted(EXT3_JOURNAL(inode)))
 344                 return -EIO;
 345
 346         log_wait_commit(EXT3_JOURNAL(inode), tid);
 347
 348         return 0;
 349 }
 350
 351 static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
 352                                struct iattr *iattr, int do_trunc)
 353 {
 354         struct inode *inode = dentry->d_inode;
 355         int rc;
 356
 357         lock_kernel();
 358
 359         /* A _really_ horrible hack to avoid removing the data stored
 360          * in the block pointers; this is really the "small" stripe MD data.
 361          * We can avoid further hackery by virtue of the MDS file size being
 362          * zero all the time (which doesn't invoke block truncate at unlink
 363          * time), so we assert we never change the MDS file size from zero. */
 364         if (iattr->ia_valid & ATTR_SIZE && !do_trunc) {
 365                 /* ATTR_SIZE would invoke truncate: clear it */
 366                 iattr->ia_valid &= ~ATTR_SIZE;
 367                 EXT3_I(inode)->i_disksize = inode->i_size = iattr->ia_size;
 368
 369                 /* make sure _something_ gets set - so new inode
 370                  * goes to disk (probably won't work over XFS */
 371                 if (!(iattr->ia_valid & (ATTR_MODE | ATTR_MTIME | ATTR_CTIME))){
 372                         iattr->ia_valid |= ATTR_MODE;
 373                         iattr->ia_mode = inode->i_mode;
 374                 }
 375         }
 376
 377         /* Don't allow setattr to change file type */
 378         iattr->ia_mode = (inode->i_mode & S_IFMT)|(iattr->ia_mode & ~S_IFMT);
 379
 380         /* We set these flags on the client, but have already checked perms
 381          * so don't confuse inode_change_ok. */
 382         iattr->ia_valid &= ~(ATTR_MTIME_SET | ATTR_ATIME_SET);
 383
 384         if (inode->i_op->setattr) {
 385                 rc = inode->i_op->setattr(dentry, iattr);
 386         } else {
 387                 rc = inode_change_ok(inode, iattr);
 388                 if (!rc)
 389                         rc = inode_setattr(inode, iattr);
 390         }
 391
 392         unlock_kernel();
 393
 394         return rc;
 395 }
 396
 397 static int fsfilt_ext3_iocontrol(struct inode * inode, struct file *file,
 398                                  unsigned int cmd, unsigned long arg)
 399 {
 400         int rc = 0;
 401         ENTRY;
 402
 403         if (inode->i_fop->ioctl)
 404                 rc = inode->i_fop->ioctl(inode, file, cmd, arg);
 405         else
 406                 RETURN(-ENOTTY);
 407
 408         RETURN(rc);
 409 }
 410
 411 #undef INLINE_EA
 412 #undef OLD_EA
 413 static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
 414                               void *lmm, int lmm_size)
 415 {
 416         int rc, old_ea = 0;
 417
 418 #ifdef INLINE_EA  /* can go away before 1.0 - just for testing bug 2097 now */
 419         /* Nasty hack city - store stripe MD data in the block pointers if
 420          * it will fit, because putting it in an EA currently kills the MDS
 421          * performance.  We'll fix this with "fast EAs" in the future.
 422          */
 423         if (inode->i_blocks == 0 && lmm_size <= sizeof(EXT3_I(inode)->i_data) -
 424                                             sizeof(EXT3_I(inode)->i_data[0])) {
 425                 unsigned old_size = EXT3_I(inode)->i_data[0];
 426                 if (old_size != 0) {
 427                         LASSERT(old_size < sizeof(EXT3_I(inode)->i_data));
 428                         CERROR("setting EA on %lu/%u again... interesting\n",
 429                                inode->i_ino, inode->i_generation);
 430                 }
 431
 432                 EXT3_I(inode)->i_data[0] = cpu_to_le32(lmm_size);
 433                 memcpy(&EXT3_I(inode)->i_data[1], lmm, lmm_size);
 434                 mark_inode_dirty(inode);
 435                 return 0;
 436         }
 437 #endif
 438 #ifdef OLD_EA
 439         /* keep this when we get rid of OLD_EA (too noisy during conversion) */
 440         if (EXT3_I(inode)->i_file_acl /* || large inode EA flag */) {
 441                 CWARN("setting EA on %lu/%u again... interesting\n",
 442                        inode->i_ino, inode->i_generation);
 443                 old_ea = 1;
 444         }
 445
 446         lock_kernel();
 447         /* this can go away before 1.0.  For bug 2097 testing only. */
 448         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_LUSTRE,
 449                                    XATTR_LUSTRE_MDS_OBJID, lmm, lmm_size, 0);
 450 #else
 451         lock_kernel();
 452         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
 453                                    XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size, 0);
 454
 455         /* This tries to delete the old-format LOV EA, but only as long as we
 456          * have successfully saved the new-format LOV EA (we can always try
 457          * the conversion again the next time the file is accessed).  It is
 458          * possible (although unlikely) that the new-format LOV EA couldn't be
 459          * saved because it ran out of space but we would need a file striped
 460          * over least 123 OSTs before the two EAs filled a 4kB block.
 461          *
 462          * This can be removed when all filesystems have converted to the
 463          * new EA format, but otherwise adds little if any overhead.  If we
 464          * wanted backward compatibility for existing files, we could keep
 465          * the old EA around for a while but we'd have to clean it up later. */
 466         if (rc >= 0 && old_ea) {
 467                 int err = ext3_xattr_set_handle(handle, inode,
 468                                                 EXT3_XATTR_INDEX_LUSTRE,
 469                                                 XATTR_LUSTRE_MDS_OBJID,
 470                                                 NULL, 0, 0);
 471                 if (err)
 472                         CERROR("error deleting old LOV EA on %lu/%u: rc %d\n",
 473                                inode->i_ino, inode->i_generation, err);
 474         }
 475 #endif
 476         unlock_kernel();
 477
 478         if (rc)
 479                 CERROR("error adding MD data to inode %lu: rc = %d\n",
 480                        inode->i_ino, rc);
 481         return rc;
 482 }
 483
 484 /* Must be called with i_sem held */
 485 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm, int lmm_size)
 486 {
 487         int rc;
 488
 489         LASSERT(down_trylock(&inode->i_sem) != 0);
 490         lock_kernel();
 491         /* Keep support for reading "inline EAs" until we convert
 492          * users over to new format entirely.  See bug 841/2097. */
 493         if (inode->i_blocks == 0 && EXT3_I(inode)->i_data[0]) {
 494                 unsigned size = le32_to_cpu(EXT3_I(inode)->i_data[0]);
 495                 void *handle;
 496
 497                 LASSERT(size < sizeof(EXT3_I(inode)->i_data));
 498                 if (lmm) {
 499                         if (size > lmm_size) {
 500                                 CERROR("inline EA on %lu/%u bad size %u > %u\n",
 501                                        inode->i_ino, inode->i_generation,
 502                                        size, lmm_size);
 503                                 return -ERANGE;
 504                         }
 505                         memcpy(lmm, &EXT3_I(inode)->i_data[1], size);
 506                 }
 507
 508 #ifndef INLINE_EA
 509                 /* migrate LOV EA data to external block - keep same format */
 510                 CWARN("DEBUG: migrate inline EA for inode %lu/%u to block\n",
 511                       inode->i_ino, inode->i_generation);
 512
 513                 handle = journal_start(EXT3_JOURNAL(inode),
 514                                        EXT3_XATTR_TRANS_BLOCKS);
 515                 if (!IS_ERR(handle)) {
 516                         int err;
 517                         rc = fsfilt_ext3_set_md(inode, handle,
 518                                                 &EXT3_I(inode)->i_data[1],size);
 519                         if (rc == 0) {
 520                                 memset(EXT3_I(inode)->i_data, 0,
 521                                        sizeof(EXT3_I(inode)->i_data));
 522                                 mark_inode_dirty(inode);
 523                         }
 524                         err = journal_stop(handle);
 525                         if (err && rc == 0)
 526                                 rc = err;
 527                 } else {
 528                         rc = PTR_ERR(handle);
 529                 }
 530 #endif
 531                 unlock_kernel();
 532                 return size;
 533         }
 534
 535         rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
 536                             XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size);
 537         /* try old EA type if new one failed - MDS will convert it for us */
 538         if (rc == -ENODATA) {
 539                 CDEBUG(D_INFO,"failed new LOV EA %d/%s from inode %lu: rc %d\n",
 540                        EXT3_XATTR_INDEX_TRUSTED, XATTR_LUSTRE_MDS_LOV_EA,
 541                        inode->i_ino, rc);
 542
 543                 rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_LUSTRE,
 544                                     XATTR_LUSTRE_MDS_OBJID, lmm, lmm_size);
 545         }
 546         unlock_kernel();
 547
 548         /* This gives us the MD size */
 549         if (lmm == NULL)
 550                 return (rc == -ENODATA) ? 0 : rc;
 551
 552         if (rc < 0) {
 553                 CDEBUG(D_INFO, "error getting EA %d/%s from inode %lu: rc %d\n",
 554                        EXT3_XATTR_INDEX_LUSTRE, XATTR_LUSTRE_MDS_OBJID,
 555                        inode->i_ino, rc);
 556                 memset(lmm, 0, lmm_size);
 557                 return (rc == -ENODATA) ? 0 : rc;
 558         }
 559
 560         return rc;
 561 }
 562
 563 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
 564                                     loff_t *off)
 565 {
 566         struct inode *inode = file->f_dentry->d_inode;
 567         int rc = 0;
 568
 569         if (S_ISREG(inode->i_mode))
 570                 rc = file->f_op->read(file, buf, count, off);
 571         else {
 572                 const int blkbits = inode->i_sb->s_blocksize_bits;
 573                 const int blksize = inode->i_sb->s_blocksize;
 574
 575                 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
 576                        count, inode->i_ino, *off);
 577                 while (count > 0) {
 578                         struct buffer_head *bh;
 579
 580                         bh = NULL;
 581                         if (*off < inode->i_size) {
 582                                 int err = 0;
 583
 584                                 bh = ext3_bread(NULL, inode, *off >> blkbits,
 585                                                 0, &err);
 586
 587                                 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
 588
 589                                 if (bh) {
 590                                         memcpy(buf, bh->b_data, blksize);
 591                                         brelse(bh);
 592                                 } else if (err) {
 593                                         /* XXX in theory we should just fake
 594                                          * this buffer and continue like ext3,
 595                                          * especially if this is a partial read
 596                                          */
 597                                         CERROR("error read dir %lu+%llu: %d\n",
 598                                                inode->i_ino, *off, err);
 599                                         RETURN(err);
 600                                 }
 601                         }
 602                         if (!bh) {
 603                                 struct ext3_dir_entry_2 *fake = (void *)buf;
 604
 605                                 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
 606                                 memset(fake, 0, sizeof(*fake));
 607                                 fake->rec_len = cpu_to_le32(blksize);
 608                         }
 609                         count -= blksize;
 610                         buf += blksize;
 611                         *off += blksize;
 612                         rc += blksize;
 613                 }
 614         }
 615
 616         return rc;
 617 }
 618
 619 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
 620 {
 621         struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
 622
 623         fcb->cb_func(fcb->cb_obd, fcb->cb_last_rcvd, fcb->cb_data, error);
 624
 625         OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
 626         atomic_dec(&fcb_cache_count);
 627 }
 628
 629 static int fsfilt_ext3_add_journal_cb(struct obd_device *obd, __u64 last_rcvd,
 630                                       void *handle, fsfilt_cb_t cb_func,
 631                                       void *cb_data)
 632 {
 633         struct fsfilt_cb_data *fcb;
 634
 635         OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
 636         if (fcb == NULL)
 637                 RETURN(-ENOMEM);
 638
 639         atomic_inc(&fcb_cache_count);
 640         fcb->cb_func = cb_func;
 641         fcb->cb_obd = obd;
 642         fcb->cb_last_rcvd = last_rcvd;
 643         fcb->cb_data = cb_data;
 644
 645         CDEBUG(D_EXT2, "set callback for last_rcvd: "LPD64"\n", last_rcvd);
 646         lock_kernel();
 647         journal_callback_set(handle, fsfilt_ext3_cb_func,
 648                              (struct journal_callback *)fcb);
 649         unlock_kernel();
 650
 651         return 0;
 652 }
 653
 654 /*
 655  * We need to hack the return value for the free inode counts because
 656  * the current EA code requires one filesystem block per inode with EAs,
 657  * so it is possible to run out of blocks before we run out of inodes.
 658  *
 659  * This can be removed when the ext3 EA code is fixed.
 660  */
 661 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
 662 {
 663         struct kstatfs sfs;
 664         int rc;
 665
 666         memset(&sfs, 0, sizeof(sfs));
 667
 668         rc = sb->s_op->statfs(sb, &sfs);
 669
 670         if (!rc && sfs.f_bfree < sfs.f_ffree) {
 671                 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
 672                 sfs.f_ffree = sfs.f_bfree;
 673         }
 674
 675         statfs_pack(osfs, &sfs);
 676         return rc;
 677 }
 678
 679 static int fsfilt_ext3_sync(struct super_block *sb)
 680 {
 681         return ext3_force_commit(sb);
 682 }
 683
 684 extern int ext3_map_inode_page(struct inode *inode, struct page *page,
 685                                unsigned long *blocks, int *created, int create);
 686 int fsfilt_ext3_map_inode_page(struct inode *inode, struct page *page,
 687                                unsigned long *blocks, int *created, int create)
 688 {
 689         return ext3_map_inode_page(inode, page, blocks, created, create);
 690 }
 691
 692 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
 693                                int nblocks, loff_t newsize);
 694 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
 695                                       int nblocks, loff_t newsize)
 696 {
 697         return ext3_prep_san_write(inode, blocks, nblocks, newsize);
 698 }
 699
 700 static int fsfilt_ext3_read_record(struct file * file, void *buf,
 701                                    int size, loff_t *offs)
 702 {
 703         struct inode *inode = file->f_dentry->d_inode;
 704         unsigned long block;
 705         struct buffer_head *bh;
 706         int err, blocksize, csize, boffs;
 707
 708         /* prevent reading after eof */
 709         lock_kernel();
 710         if (inode->i_size < *offs + size) {
 711                 size = inode->i_size - *offs;
 712                 unlock_kernel();
 713                 if (size < 0) {
 714                         CERROR("size %llu is too short for read %u@%llu\n",
 715                                inode->i_size, size, *offs);
 716                         return -EIO;
 717                 } else if (size == 0) {
 718                         return 0;
 719                 }
 720         } else {
 721                 unlock_kernel();
 722         }
 723
 724         blocksize = 1 << inode->i_blkbits;
 725
 726         while (size > 0) {
 727                 block = *offs >> inode->i_blkbits;
 728                 boffs = *offs & (blocksize - 1);
 729                 csize = min(blocksize - boffs, size);
 730                 bh = ext3_bread(NULL, inode, block, 0, &err);
 731                 if (!bh) {
 732                         CERROR("can't read block: %d\n", err);
 733                         return err;
 734                 }
 735
 736                 memcpy(buf, bh->b_data + boffs, csize);
 737                 brelse(bh);
 738
 739                 *offs += csize;
 740                 buf += csize;
 741                 size -= csize;
 742         }
 743         return 0;
 744 }
 745
 746 static int fsfilt_ext3_write_record(struct file *file, void *buf, int bufsize,
 747                                     loff_t *offs, int force_sync)
 748 {
 749         struct buffer_head *bh = NULL;
 750         unsigned long block;
 751         struct inode *inode = file->f_dentry->d_inode;
 752         loff_t old_size = inode->i_size, offset = *offs;
 753         loff_t new_size = inode->i_size;
 754         journal_t *journal;
 755         handle_t *handle;
 756         int err, block_count = 0, blocksize, size, boffs;
 757
 758         /* Determine how many transaction credits are needed */
 759         blocksize = 1 << inode->i_blkbits;
 760         block_count = (*offs & (blocksize - 1)) + bufsize;
 761         block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
 762
 763         journal = EXT3_SB(inode->i_sb)->s_journal;
 764         lock_kernel();
 765         handle = journal_start(journal,
 766                                block_count * EXT3_DATA_TRANS_BLOCKS + 2);
 767         unlock_kernel();
 768         if (IS_ERR(handle)) {
 769                 CERROR("can't start transaction\n");
 770                 return PTR_ERR(handle);
 771         }
 772
 773         while (bufsize > 0) {
 774                 if (bh != NULL)
 775                         brelse(bh);
 776
 777                 block = offset >> inode->i_blkbits;
 778                 boffs = offset & (blocksize - 1);
 779                 size = min(blocksize - boffs, bufsize);
 780                 bh = ext3_bread(handle, inode, block, 1, &err);
 781                 if (!bh) {
 782                         CERROR("can't read/create block: %d\n", err);
 783                         goto out;
 784                 }
 785
 786                 err = ext3_journal_get_write_access(handle, bh);
 787                 if (err) {
 788                         CERROR("journal_get_write_access() returned error %d\n",
 789                                err);
 790                         goto out;
 791                 }
 792                 LASSERT(bh->b_data + boffs + size <= bh->b_data + bh->b_size);
 793                 memcpy(bh->b_data + boffs, buf, size);
 794                 err = ext3_journal_dirty_metadata(handle, bh);
 795                 if (err) {
 796                         CERROR("journal_dirty_metadata() returned error %d\n",
 797                                err);
 798                         goto out;
 799                 }
 800                 if (offset + size > new_size)
 801                         new_size = offset + size;
 802                 offset += size;
 803                 bufsize -= size;
 804                 buf += size;
 805         }
 806
 807         if (force_sync)
 808                 handle->h_sync = 1; /* recovery likes this */
 809 out:
 810         if (bh)
 811                 brelse(bh);
 812
 813         /* correct in-core and on-disk sizes */
 814         if (new_size > inode->i_size) {
 815                 lock_kernel();
 816                 if (new_size > inode->i_size)
 817                         inode->i_size = new_size;
 818                 if (inode->i_size > EXT3_I(inode)->i_disksize)
 819                         EXT3_I(inode)->i_disksize = inode->i_size;
 820                 if (inode->i_size > old_size)
 821                         mark_inode_dirty(inode);
 822                 unlock_kernel();
 823         }
 824
 825         lock_kernel();
 826         journal_stop(handle);
 827         unlock_kernel();
 828
 829         if (err == 0)
 830                 *offs = offset;
 831         return err;
 832 }
 833
 834 static int fsfilt_ext3_setup(struct super_block *sb)
 835 {
 836 #if 0
 837         EXT3_SB(sb)->dx_lock = fsfilt_ext3_dx_lock;
 838         EXT3_SB(sb)->dx_unlock = fsfilt_ext3_dx_unlock;
 839 #endif
 840 #ifdef S_PDIROPS
 841         CWARN("Enabling PDIROPS\n");
 842         set_opt(EXT3_SB(sb)->s_mount_opt, PDIROPS);
 843         sb->s_flags |= S_PDIROPS;
 844 #endif
 845         return 0;
 846 }
 847
 848 static struct fsfilt_operations fsfilt_ext3_ops = {
 849         fs_type:                "ext3",
 850         fs_owner:               THIS_MODULE,
 851         fs_start:               fsfilt_ext3_start,
 852         fs_brw_start:           fsfilt_ext3_brw_start,
 853         fs_commit:              fsfilt_ext3_commit,
 854         fs_commit_async:        fsfilt_ext3_commit_async,
 855         fs_commit_wait:         fsfilt_ext3_commit_wait,
 856         fs_setattr:             fsfilt_ext3_setattr,
 857         fs_iocontrol:           fsfilt_ext3_iocontrol,
 858         fs_set_md:              fsfilt_ext3_set_md,
 859         fs_get_md:              fsfilt_ext3_get_md,
 860         fs_readpage:            fsfilt_ext3_readpage,
 861         fs_add_journal_cb:      fsfilt_ext3_add_journal_cb,
 862         fs_statfs:              fsfilt_ext3_statfs,
 863         fs_sync:                fsfilt_ext3_sync,
 864         fs_map_inode_page:      fsfilt_ext3_map_inode_page,
 865         fs_prep_san_write:      fsfilt_ext3_prep_san_write,
 866         fs_write_record:        fsfilt_ext3_write_record,
 867         fs_read_record:         fsfilt_ext3_read_record,
 868         fs_setup:               fsfilt_ext3_setup,
 869 };
 870
 871 static int __init fsfilt_ext3_init(void)
 872 {
 873         int rc;
 874
 875         //rc = ext3_xattr_register();
 876         fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
 877                                       sizeof(struct fsfilt_cb_data), 0,
 878                                       0, NULL, NULL);
 879         if (!fcb_cache) {
 880                 CERROR("error allocating fsfilt journal callback cache\n");
 881                 GOTO(out, rc = -ENOMEM);
 882         }
 883
 884         rc = fsfilt_register_ops(&fsfilt_ext3_ops);
 885
 886         if (rc)
 887                 kmem_cache_destroy(fcb_cache);
 888 out:
 889         return rc;
 890 }
 891
 892 static void __exit fsfilt_ext3_exit(void)
 893 {
 894         int rc;
 895
 896         fsfilt_unregister_ops(&fsfilt_ext3_ops);
 897         rc = kmem_cache_destroy(fcb_cache);
 898
 899         if (rc || atomic_read(&fcb_cache_count)) {
 900                 CERROR("can't free fsfilt callback cache: count %d, rc = %d\n",
 901                        atomic_read(&fcb_cache_count), rc);
 902         }
 903
 904         //rc = ext3_xattr_unregister();
 905 }
 906
 907 module_init(fsfilt_ext3_init);
 908 module_exit(fsfilt_ext3_exit);
 909
 910 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
 911 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
 912 MODULE_LICENSE("GPL");