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