lustre/lvfs/fsfilt_ext3.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  lustre/lib/fsfilt_ext3.c
   5  *  Lustre filesystem abstraction routines
   6  *
   7  *  Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   8  *   Author: Andreas Dilger <adilger@clusterfs.com>
   9  *
  10  *   This file is part of Lustre, http://www.lustre.org.
  11  *
  12  *   Lustre is free software; you can redistribute it and/or
  13  *   modify it under the terms of version 2 of the GNU General Public
  14  *   License as published by the Free Software Foundation.
  15  *
  16  *   Lustre is distributed in the hope that it will be useful,
  17  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19  *   GNU General Public License for more details.
  20  *
  21  *   You should have received a copy of the GNU General Public License
  22  *   along with Lustre; if not, write to the Free Software
  23  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  24  */
  25
  26 #define DEBUG_SUBSYSTEM S_FILTER
  27
  28 #include <linux/init.h>
  29 #include <linux/module.h>
  30 #include <linux/fs.h>
  31 #include <linux/jbd.h>
  32 #include <linux/slab.h>
  33 #include <linux/pagemap.h>
  34 #include <linux/quotaops.h>
  35 #include <linux/ext3_fs.h>
  36 #include <linux/ext3_jbd.h>
  37 #include <linux/version.h>
  38 /* XXX ugh */
  39 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
  40  #include <linux/ext3_xattr.h>
  41 #else
  42  #include <linux/../../fs/ext3/xattr.h>
  43 #endif
  44 #include <linux/kp30.h>
  45 #include <linux/lustre_fsfilt.h>
  46 #include <linux/obd.h>
  47 #include <linux/obd_class.h>
  48
  49 static kmem_cache_t *fcb_cache;
  50 static atomic_t fcb_cache_count = ATOMIC_INIT(0);
  51
  52 struct fsfilt_cb_data {
  53         struct journal_callback cb_jcb; /* jbd private data - MUST BE FIRST */
  54         fsfilt_cb_t cb_func;            /* MDS/OBD completion function */
  55         struct obd_device *cb_obd;      /* MDS/OBD completion device */
  56         __u64 cb_last_rcvd;             /* MDS/OST last committed operation */
  57         void *cb_data;                  /* MDS/OST completion function data */
  58 };
  59
  60 #ifndef EXT3_XATTR_INDEX_TRUSTED        /* temporary until we hit l28 kernel */
  61 #define EXT3_XATTR_INDEX_TRUSTED        4
  62 #endif
  63 #define XATTR_LUSTRE_MDS_LOV_EA         "lov"
  64
  65 #define EXT3_XATTR_INDEX_LUSTRE         5                         /* old */
  66 #define XATTR_LUSTRE_MDS_OBJID          "system.lustre_mds_objid" /* old */
  67
  68 /*
  69  * We don't currently need any additional blocks for rmdir and
  70  * unlink transactions because we are storing the OST oa_id inside
  71  * the inode (which we will be changing anyways as part of this
  72  * transaction).
  73  */
  74 static void *fsfilt_ext3_start(struct inode *inode, int op, void *desc_private,
  75                                int logs)
  76 {
  77         /* For updates to the last recieved file */
  78         int nblocks = EXT3_SINGLEDATA_TRANS_BLOCKS;
  79         journal_t *journal;
  80         void *handle;
  81
  82         if (current->journal_info) {
  83                 CDEBUG(D_INODE, "increasing refcount on %p\n",
  84                        current->journal_info);
  85                 goto journal_start;
  86         }
  87
  88         switch(op) {
  89         case FSFILT_OP_RMDIR:
  90         case FSFILT_OP_UNLINK:
  91                 /* delete one file + create/update logs for each stripe */
  92                 nblocks += EXT3_DELETE_TRANS_BLOCKS;
  93                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
  94                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
  95                 break;
  96         case FSFILT_OP_RENAME:
  97                 /* modify additional directory */
  98                 nblocks += EXT3_SINGLEDATA_TRANS_BLOCKS;
  99                 /* no break */
 100         case FSFILT_OP_SYMLINK:
 101                 /* additional block + block bitmap + GDT for long symlink */
 102                 nblocks += 3;
 103                 /* no break */
 104         case FSFILT_OP_CREATE:
 105                 /* create/update logs for each stripe */
 106                 nblocks += (EXT3_INDEX_EXTRA_TRANS_BLOCKS +
 107                             EXT3_SINGLEDATA_TRANS_BLOCKS) * logs;
 108                 /* no break */
 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 +
 117                         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:
 124                 /* blocks for log header bitmap update OR
 125                  * blocks for catalog header bitmap update + unlink of logs */
 126                 nblocks = (LLOG_CHUNK_SIZE >> inode->i_blkbits) +
 127                         EXT3_DELETE_TRANS_BLOCKS * logs;
 128                 break;
 129         default: CERROR("unknown transaction start op %d\n", op);
 130                  LBUG();
 131         }
 132
 133         LASSERT(current->journal_info == desc_private);
 134         journal = EXT3_SB(inode->i_sb)->s_journal;
 135         if (nblocks > journal->j_max_transaction_buffers) {
 136                 CERROR("too many credits %d for op %ux%u using %d instead\n",
 137                        nblocks, op, logs, journal->j_max_transaction_buffers);
 138                 nblocks = journal->j_max_transaction_buffers;
 139         }
 140
 141  journal_start:
 142         lock_kernel();
 143         handle = journal_start(EXT3_JOURNAL(inode), nblocks);
 144         unlock_kernel();
 145
 146         if (!IS_ERR(handle))
 147                 LASSERT(current->journal_info == handle);
 148         else
 149                 CERROR("error starting handle for op %u (%u credits): rc %ld\n",
 150                        op, nblocks, PTR_ERR(handle));
 151         return handle;
 152 }
 153
 154 /*
 155  * Calculate the number of buffer credits needed to write multiple pages in
 156  * a single ext3 transaction.  No, this shouldn't be here, but as yet ext3
 157  * doesn't have a nice API for calculating this sort of thing in advance.
 158  *
 159  * See comment above ext3_writepage_trans_blocks for details.  We assume
 160  * no data journaling is being done, but it does allow for all of the pages
 161  * being non-contiguous.  If we are guaranteed contiguous pages we could
 162  * reduce the number of (d)indirect blocks a lot.
 163  *
 164  * With N blocks per page and P pages, for each inode we have at most:
 165  * N*P indirect
 166  * min(N*P, blocksize/4 + 1) dindirect blocks
 167  * niocount tindirect
 168  *
 169  * For the entire filesystem, we have at most:
 170  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 171  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 172  * objcount inode blocks
 173  * 1 superblock
 174  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 175  *
 176  * 1 EXT3_DATA_TRANS_BLOCKS for the last_rcvd update.
 177  */
 178 static int fsfilt_ext3_credits_needed(int objcount, struct fsfilt_objinfo *fso,
 179                                       int niocount, struct niobuf_local *nb)
 180 {
 181         struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
 182         __u64 next_indir;
 183         const int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 184         int nbitmaps = 0, ngdblocks;
 185         int needed = objcount + 1; /* inodes + superblock */
 186         int i, j;
 187
 188         for (i = 0, j = 0; i < objcount; i++, fso++) {
 189                 /* two or more dindirect blocks in case we cross boundary */
 190                 int ndind = (long)((nb[j + fso->fso_bufcnt - 1].offset -
 191                                     nb[j].offset) >>
 192                                    sb->s_blocksize_bits) /
 193                         (EXT3_ADDR_PER_BLOCK(sb) * EXT3_ADDR_PER_BLOCK(sb));
 194                 nbitmaps += min(fso->fso_bufcnt, ndind > 0 ? ndind : 2);
 195
 196                 /* leaf, indirect, tindirect blocks for first block */
 197                 nbitmaps += blockpp + 2;
 198
 199                 j += fso->fso_bufcnt;
 200         }
 201
 202         next_indir = nb[0].offset +
 203                 (EXT3_ADDR_PER_BLOCK(sb) << sb->s_blocksize_bits);
 204         for (i = 1; i < niocount; i++) {
 205                 if (nb[i].offset >= next_indir) {
 206                         nbitmaps++;     /* additional indirect */
 207                         next_indir = nb[i].offset +
 208                                 (EXT3_ADDR_PER_BLOCK(sb)<<sb->s_blocksize_bits);
 209                 } else if (nb[i].offset != nb[i - 1].offset + sb->s_blocksize) {
 210                         nbitmaps++;     /* additional indirect */
 211                 }
 212                 nbitmaps += blockpp;    /* each leaf in different group? */
 213         }
 214
 215         ngdblocks = nbitmaps;
 216         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 217                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 218         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 219                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 220
 221         needed += nbitmaps + ngdblocks;
 222
 223         /* last_rcvd update */
 224         needed += EXT3_DATA_TRANS_BLOCKS;
 225
 226 #if defined(CONFIG_QUOTA) && !defined(__x86_64__) /* XXX */
 227         /* We assume that there will be 1 bit set in s_dquot.flags for each
 228          * quota file that is active.  This is at least true for now.
 229          */
 230         needed += hweight32(sb_any_quota_enabled(sb)) *
 231                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 232 #endif
 233
 234         return needed;
 235 }
 236
 237 /* We have to start a huge journal transaction here to hold all of the
 238  * metadata for the pages being written here.  This is necessitated by
 239  * the fact that we do lots of prepare_write operations before we do
 240  * any of the matching commit_write operations, so even if we split
 241  * up to use "smaller" transactions none of them could complete until
 242  * all of them were opened.  By having a single journal transaction,
 243  * we eliminate duplicate reservations for common blocks like the
 244  * superblock and group descriptors or bitmaps.
 245  *
 246  * We will start the transaction here, but each prepare_write will
 247  * add a refcount to the transaction, and each commit_write will
 248  * remove a refcount.  The transaction will be closed when all of
 249  * the pages have been written.
 250  */
 251 static void *fsfilt_ext3_brw_start(int objcount, struct fsfilt_objinfo *fso,
 252                                    int niocount, struct niobuf_local *nb,
 253                                    void *desc_private, int logs)
 254 {
 255         journal_t *journal;
 256         handle_t *handle;
 257         int needed;
 258         ENTRY;
 259
 260         LASSERT(current->journal_info == desc_private);
 261         journal = EXT3_SB(fso->fso_dentry->d_inode->i_sb)->s_journal;
 262         needed = fsfilt_ext3_credits_needed(objcount, fso, niocount, nb);
 263
 264         /* The number of blocks we could _possibly_ dirty can very large.
 265          * We reduce our request if it is absurd (and we couldn't get that
 266          * many credits for a single handle anyways).
 267          *
 268          * At some point we have to limit the size of I/Os sent at one time,
 269          * increase the size of the journal, or we have to calculate the
 270          * actual journal requirements more carefully by checking all of
 271          * the blocks instead of being maximally pessimistic.  It remains to
 272          * be seen if this is a real problem or not.
 273          */
 274         if (needed > journal->j_max_transaction_buffers) {
 275                 CERROR("want too many journal credits (%d) using %d instead\n",
 276                        needed, journal->j_max_transaction_buffers);
 277                 needed = journal->j_max_transaction_buffers;
 278         }
 279
 280         lock_kernel();
 281         handle = journal_start(journal, needed);
 282         unlock_kernel();
 283         if (IS_ERR(handle)) {
 284                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 285                        PTR_ERR(handle));
 286         } else {
 287                 LASSERT(handle->h_buffer_credits >= needed);
 288                 LASSERT(current->journal_info == handle);
 289         }
 290
 291         RETURN(handle);
 292 }
 293
 294 static int fsfilt_ext3_commit(struct inode *inode, void *h, int force_sync)
 295 {
 296         int rc;
 297         handle_t *handle = h;
 298
 299         LASSERT(current->journal_info == handle);
 300         if (force_sync)
 301                 handle->h_sync = 1; /* recovery likes this */
 302
 303         lock_kernel();
 304         rc = journal_stop(handle);
 305         unlock_kernel();
 306
 307         // LASSERT(current->journal_info == NULL);
 308         return rc;
 309 }
 310
 311 static int fsfilt_ext3_commit_async(struct inode *inode, void *h,
 312                                         void **wait_handle)
 313 {
 314         transaction_t *transaction;
 315         unsigned long tid, rtid;
 316         handle_t *handle = h;
 317         journal_t *journal;
 318         int rc;
 319
 320         LASSERT(current->journal_info == handle);
 321
 322         lock_kernel();
 323         transaction = handle->h_transaction;
 324         journal = transaction->t_journal;
 325         tid = transaction->t_tid;
 326         /* we don't want to be blocked */
 327         handle->h_sync = 0;
 328         rc = journal_stop(handle);
 329         if (rc) {
 330                 CERROR("error while stopping transaction: %d\n", rc);
 331                 unlock_kernel();
 332                 return rc;
 333         }
 334 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)
 335         rtid = log_start_commit(journal, transaction);
 336         if (rtid != tid)
 337                 CERROR("strange race: %lu != %lu\n",
 338                        (unsigned long) tid, (unsigned long) rtid);
 339 #else
 340         log_start_commit(journal, transaction->t_tid);
 341 #endif
 342         unlock_kernel();
 343
 344         *wait_handle = (void *) tid;
 345         CDEBUG(D_INODE, "commit async: %lu\n", (unsigned long) tid);
 346         return 0;
 347 }
 348
 349 static int fsfilt_ext3_commit_wait(struct inode *inode, void *h)
 350 {
 351         tid_t tid = (tid_t)(long)h;
 352
 353         CDEBUG(D_INODE, "commit wait: %lu\n", (unsigned long) tid);
 354         if (is_journal_aborted(EXT3_JOURNAL(inode)))
 355                 return -EIO;
 356
 357         log_wait_commit(EXT3_JOURNAL(inode), tid);
 358
 359         return 0;
 360 }
 361
 362 static int fsfilt_ext3_setattr(struct dentry *dentry, void *handle,
 363                                struct iattr *iattr, int do_trunc)
 364 {
 365         struct inode *inode = dentry->d_inode;
 366         int rc;
 367
 368         lock_kernel();
 369
 370         /* A _really_ horrible hack to avoid removing the data stored
 371          * in the block pointers; this is really the "small" stripe MD data.
 372          * We can avoid further hackery by virtue of the MDS file size being
 373          * zero all the time (which doesn't invoke block truncate at unlink
 374          * time), so we assert we never change the MDS file size from zero. */
 375         if (iattr->ia_valid & ATTR_SIZE && !do_trunc) {
 376                 /* ATTR_SIZE would invoke truncate: clear it */
 377                 iattr->ia_valid &= ~ATTR_SIZE;
 378                 EXT3_I(inode)->i_disksize = inode->i_size = iattr->ia_size;
 379
 380                 /* make sure _something_ gets set - so new inode
 381                  * goes to disk (probably won't work over XFS */
 382                 if (!(iattr->ia_valid & (ATTR_MODE | ATTR_MTIME | ATTR_CTIME))){
 383                         iattr->ia_valid |= ATTR_MODE;
 384                         iattr->ia_mode = inode->i_mode;
 385                 }
 386         }
 387
 388         /* Don't allow setattr to change file type */
 389         iattr->ia_mode = (inode->i_mode & S_IFMT)|(iattr->ia_mode & ~S_IFMT);
 390
 391         /* We set these flags on the client, but have already checked perms
 392          * so don't confuse inode_change_ok. */
 393         iattr->ia_valid &= ~(ATTR_MTIME_SET | ATTR_ATIME_SET);
 394
 395         if (inode->i_op->setattr) {
 396                 rc = inode->i_op->setattr(dentry, iattr);
 397         } else {
 398                 rc = inode_change_ok(inode, iattr);
 399                 if (!rc)
 400                         rc = inode_setattr(inode, iattr);
 401         }
 402
 403         unlock_kernel();
 404
 405         return rc;
 406 }
 407
 408 static int fsfilt_ext3_iocontrol(struct inode * inode, struct file *file,
 409                                  unsigned int cmd, unsigned long arg)
 410 {
 411         int rc = 0;
 412         ENTRY;
 413
 414         if (inode->i_fop->ioctl)
 415                 rc = inode->i_fop->ioctl(inode, file, cmd, arg);
 416         else
 417                 RETURN(-ENOTTY);
 418
 419         RETURN(rc);
 420 }
 421
 422 #undef INLINE_EA
 423 #undef OLD_EA
 424 static int fsfilt_ext3_set_md(struct inode *inode, void *handle,
 425                               void *lmm, int lmm_size)
 426 {
 427         int rc, old_ea = 0;
 428
 429 #ifdef INLINE_EA  /* can go away before 1.0 - just for testing bug 2097 now */
 430         /* Nasty hack city - store stripe MD data in the block pointers if
 431          * it will fit, because putting it in an EA currently kills the MDS
 432          * performance.  We'll fix this with "fast EAs" in the future.
 433          */
 434         if (inode->i_blocks == 0 && lmm_size <= sizeof(EXT3_I(inode)->i_data) -
 435                                             sizeof(EXT3_I(inode)->i_data[0])) {
 436                 unsigned old_size = EXT3_I(inode)->i_data[0];
 437                 if (old_size != 0) {
 438                         LASSERT(old_size < sizeof(EXT3_I(inode)->i_data));
 439                         CERROR("setting EA on %lu/%u again... interesting\n",
 440                                inode->i_ino, inode->i_generation);
 441                 }
 442
 443                 EXT3_I(inode)->i_data[0] = cpu_to_le32(lmm_size);
 444                 memcpy(&EXT3_I(inode)->i_data[1], lmm, lmm_size);
 445                 mark_inode_dirty(inode);
 446                 return 0;
 447         }
 448 #endif
 449 #ifdef OLD_EA
 450         /* keep this when we get rid of OLD_EA (too noisy during conversion) */
 451         if (EXT3_I(inode)->i_file_acl /* || large inode EA flag */) {
 452                 CWARN("setting EA on %lu/%u again... interesting\n",
 453                        inode->i_ino, inode->i_generation);
 454                 old_ea = 1;
 455         }
 456
 457         lock_kernel();
 458         /* this can go away before 1.0.  For bug 2097 testing only. */
 459         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_LUSTRE,
 460                                    XATTR_LUSTRE_MDS_OBJID, lmm, lmm_size, 0);
 461 #else
 462         lock_kernel();
 463         rc = ext3_xattr_set_handle(handle, inode, EXT3_XATTR_INDEX_TRUSTED,
 464                                    XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size, 0);
 465
 466         /* This tries to delete the old-format LOV EA, but only as long as we
 467          * have successfully saved the new-format LOV EA (we can always try
 468          * the conversion again the next time the file is accessed).  It is
 469          * possible (although unlikely) that the new-format LOV EA couldn't be
 470          * saved because it ran out of space but we would need a file striped
 471          * over least 123 OSTs before the two EAs filled a 4kB block.
 472          *
 473          * This can be removed when all filesystems have converted to the
 474          * new EA format, but otherwise adds little if any overhead.  If we
 475          * wanted backward compatibility for existing files, we could keep
 476          * the old EA around for a while but we'd have to clean it up later. */
 477         if (rc >= 0 && old_ea) {
 478                 int err = ext3_xattr_set_handle(handle, inode,
 479                                                 EXT3_XATTR_INDEX_LUSTRE,
 480                                                 XATTR_LUSTRE_MDS_OBJID,
 481                                                 NULL, 0, 0);
 482                 if (err)
 483                         CERROR("error deleting old LOV EA on %lu/%u: rc %d\n",
 484                                inode->i_ino, inode->i_generation, err);
 485         }
 486 #endif
 487         unlock_kernel();
 488
 489         if (rc)
 490                 CERROR("error adding MD data to inode %lu: rc = %d\n",
 491                        inode->i_ino, rc);
 492         return rc;
 493 }
 494
 495 /* Must be called with i_sem held */
 496 static int fsfilt_ext3_get_md(struct inode *inode, void *lmm, int lmm_size)
 497 {
 498         int rc;
 499
 500         LASSERT(down_trylock(&inode->i_sem) != 0);
 501         lock_kernel();
 502         /* Keep support for reading "inline EAs" until we convert
 503          * users over to new format entirely.  See bug 841/2097. */
 504         if (inode->i_blocks == 0 && EXT3_I(inode)->i_data[0]) {
 505                 unsigned size = le32_to_cpu(EXT3_I(inode)->i_data[0]);
 506                 void *handle;
 507
 508                 LASSERT(size < sizeof(EXT3_I(inode)->i_data));
 509                 if (lmm) {
 510                         if (size > lmm_size) {
 511                                 CERROR("inline EA on %lu/%u bad size %u > %u\n",
 512                                        inode->i_ino, inode->i_generation,
 513                                        size, lmm_size);
 514                                 return -ERANGE;
 515                         }
 516                         memcpy(lmm, &EXT3_I(inode)->i_data[1], size);
 517                 }
 518
 519 #ifndef INLINE_EA
 520                 /* migrate LOV EA data to external block - keep same format */
 521                 CWARN("DEBUG: migrate inline EA for inode %lu/%u to block\n",
 522                       inode->i_ino, inode->i_generation);
 523
 524                 handle = journal_start(EXT3_JOURNAL(inode),
 525                                        EXT3_XATTR_TRANS_BLOCKS);
 526                 if (!IS_ERR(handle)) {
 527                         int err;
 528                         rc = fsfilt_ext3_set_md(inode, handle,
 529                                                 &EXT3_I(inode)->i_data[1],size);
 530                         if (rc == 0) {
 531                                 memset(EXT3_I(inode)->i_data, 0,
 532                                        sizeof(EXT3_I(inode)->i_data));
 533                                 mark_inode_dirty(inode);
 534                         }
 535                         err = journal_stop(handle);
 536                         if (err && rc == 0)
 537                                 rc = err;
 538                 } else {
 539                         rc = PTR_ERR(handle);
 540                 }
 541 #endif
 542                 unlock_kernel();
 543                 return size;
 544         }
 545
 546         rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_TRUSTED,
 547                             XATTR_LUSTRE_MDS_LOV_EA, lmm, lmm_size);
 548         /* try old EA type if new one failed - MDS will convert it for us */
 549         if (rc == -ENODATA) {
 550                 CDEBUG(D_INFO,"failed new LOV EA %d/%s from inode %lu: rc %d\n",
 551                        EXT3_XATTR_INDEX_TRUSTED, XATTR_LUSTRE_MDS_LOV_EA,
 552                        inode->i_ino, rc);
 553
 554                 rc = ext3_xattr_get(inode, EXT3_XATTR_INDEX_LUSTRE,
 555                                     XATTR_LUSTRE_MDS_OBJID, lmm, lmm_size);
 556         }
 557         unlock_kernel();
 558
 559         /* This gives us the MD size */
 560         if (lmm == NULL)
 561                 return (rc == -ENODATA) ? 0 : rc;
 562
 563         if (rc < 0) {
 564                 CDEBUG(D_INFO, "error getting EA %d/%s from inode %lu: rc %d\n",
 565                        EXT3_XATTR_INDEX_LUSTRE, XATTR_LUSTRE_MDS_OBJID,
 566                        inode->i_ino, rc);
 567                 memset(lmm, 0, lmm_size);
 568                 return (rc == -ENODATA) ? 0 : rc;
 569         }
 570
 571         return rc;
 572 }
 573
 574 static ssize_t fsfilt_ext3_readpage(struct file *file, char *buf, size_t count,
 575                                     loff_t *off)
 576 {
 577         struct inode *inode = file->f_dentry->d_inode;
 578         int rc = 0;
 579
 580         if (S_ISREG(inode->i_mode))
 581                 rc = file->f_op->read(file, buf, count, off);
 582         else {
 583                 const int blkbits = inode->i_sb->s_blocksize_bits;
 584                 const int blksize = inode->i_sb->s_blocksize;
 585
 586                 CDEBUG(D_EXT2, "reading "LPSZ" at dir %lu+%llu\n",
 587                        count, inode->i_ino, *off);
 588                 while (count > 0) {
 589                         struct buffer_head *bh;
 590
 591                         bh = NULL;
 592                         if (*off < inode->i_size) {
 593                                 int err = 0;
 594
 595                                 bh = ext3_bread(NULL, inode, *off >> blkbits,
 596                                                 0, &err);
 597
 598                                 CDEBUG(D_EXT2, "read %u@%llu\n", blksize, *off);
 599
 600                                 if (bh) {
 601                                         memcpy(buf, bh->b_data, blksize);
 602                                         brelse(bh);
 603                                 } else if (err) {
 604                                         /* XXX in theory we should just fake
 605                                          * this buffer and continue like ext3,
 606                                          * especially if this is a partial read
 607                                          */
 608                                         CERROR("error read dir %lu+%llu: %d\n",
 609                                                inode->i_ino, *off, err);
 610                                         RETURN(err);
 611                                 }
 612                         }
 613                         if (!bh) {
 614                                 struct ext3_dir_entry_2 *fake = (void *)buf;
 615
 616                                 CDEBUG(D_EXT2, "fake %u@%llu\n", blksize, *off);
 617                                 memset(fake, 0, sizeof(*fake));
 618                                 fake->rec_len = cpu_to_le32(blksize);
 619                         }
 620                         count -= blksize;
 621                         buf += blksize;
 622                         *off += blksize;
 623                         rc += blksize;
 624                 }
 625         }
 626
 627         return rc;
 628 }
 629
 630 static void fsfilt_ext3_cb_func(struct journal_callback *jcb, int error)
 631 {
 632         struct fsfilt_cb_data *fcb = (struct fsfilt_cb_data *)jcb;
 633
 634         fcb->cb_func(fcb->cb_obd, fcb->cb_last_rcvd, fcb->cb_data, error);
 635
 636         OBD_SLAB_FREE(fcb, fcb_cache, sizeof *fcb);
 637         atomic_dec(&fcb_cache_count);
 638 }
 639
 640 static int fsfilt_ext3_add_journal_cb(struct obd_device *obd, __u64 last_rcvd,
 641                                       void *handle, fsfilt_cb_t cb_func,
 642                                       void *cb_data)
 643 {
 644         struct fsfilt_cb_data *fcb;
 645
 646         OBD_SLAB_ALLOC(fcb, fcb_cache, GFP_NOFS, sizeof *fcb);
 647         if (fcb == NULL)
 648                 RETURN(-ENOMEM);
 649
 650         atomic_inc(&fcb_cache_count);
 651         fcb->cb_func = cb_func;
 652         fcb->cb_obd = obd;
 653         fcb->cb_last_rcvd = last_rcvd;
 654         fcb->cb_data = cb_data;
 655
 656         CDEBUG(D_EXT2, "set callback for last_rcvd: "LPD64"\n", last_rcvd);
 657         lock_kernel();
 658         journal_callback_set(handle, fsfilt_ext3_cb_func,
 659                              (struct journal_callback *)fcb);
 660         unlock_kernel();
 661
 662         return 0;
 663 }
 664
 665 /*
 666  * We need to hack the return value for the free inode counts because
 667  * the current EA code requires one filesystem block per inode with EAs,
 668  * so it is possible to run out of blocks before we run out of inodes.
 669  *
 670  * This can be removed when the ext3 EA code is fixed.
 671  */
 672 static int fsfilt_ext3_statfs(struct super_block *sb, struct obd_statfs *osfs)
 673 {
 674         struct kstatfs sfs;
 675         int rc;
 676
 677         memset(&sfs, 0, sizeof(sfs));
 678
 679         rc = sb->s_op->statfs(sb, &sfs);
 680
 681         if (!rc && sfs.f_bfree < sfs.f_ffree) {
 682                 sfs.f_files = (sfs.f_files - sfs.f_ffree) + sfs.f_bfree;
 683                 sfs.f_ffree = sfs.f_bfree;
 684         }
 685
 686         statfs_pack(osfs, &sfs);
 687         return rc;
 688 }
 689
 690 static int fsfilt_ext3_sync(struct super_block *sb)
 691 {
 692         return ext3_force_commit(sb);
 693 }
 694
 695 extern int ext3_map_inode_page(struct inode *inode, struct page *page,
 696                                unsigned long *blocks, int *created, int create);
 697 int fsfilt_ext3_map_inode_page(struct inode *inode, struct page *page,
 698                                unsigned long *blocks, int *created, int create)
 699 {
 700         return ext3_map_inode_page(inode, page, blocks, created, create);
 701 }
 702
 703 extern int ext3_prep_san_write(struct inode *inode, long *blocks,
 704                                int nblocks, loff_t newsize);
 705 static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
 706                                       int nblocks, loff_t newsize)
 707 {
 708         return ext3_prep_san_write(inode, blocks, nblocks, newsize);
 709 }
 710
 711 static int fsfilt_ext3_read_record(struct file * file, void *buf,
 712                                    int size, loff_t *offs)
 713 {
 714         struct inode *inode = file->f_dentry->d_inode;
 715         unsigned long block;
 716         struct buffer_head *bh;
 717         int err, blocksize, csize, boffs;
 718
 719         /* prevent reading after eof */
 720         lock_kernel();
 721         if (inode->i_size < *offs + size) {
 722                 size = inode->i_size - *offs;
 723                 unlock_kernel();
 724                 if (size < 0) {
 725                         CERROR("size %llu is too short for read %u@%llu\n",
 726                                inode->i_size, size, *offs);
 727                         return -EIO;
 728                 } else if (size == 0) {
 729                         return 0;
 730                 }
 731         } else {
 732                 unlock_kernel();
 733         }
 734
 735         blocksize = 1 << inode->i_blkbits;
 736
 737         while (size > 0) {
 738                 block = *offs >> inode->i_blkbits;
 739                 boffs = *offs & (blocksize - 1);
 740                 csize = min(blocksize - boffs, size);
 741                 bh = ext3_bread(NULL, inode, block, 0, &err);
 742                 if (!bh) {
 743                         CERROR("can't read block: %d\n", err);
 744                         return err;
 745                 }
 746
 747                 memcpy(buf, bh->b_data + boffs, csize);
 748                 brelse(bh);
 749
 750                 *offs += csize;
 751                 buf += csize;
 752                 size -= csize;
 753         }
 754         return 0;
 755 }
 756
 757 static int fsfilt_ext3_write_record(struct file *file, void *buf, int bufsize,
 758                                     loff_t *offs, int force_sync)
 759 {
 760         struct buffer_head *bh = NULL;
 761         unsigned long block;
 762         struct inode *inode = file->f_dentry->d_inode;
 763         loff_t old_size = inode->i_size, offset = *offs;
 764         loff_t new_size = inode->i_size;
 765         journal_t *journal;
 766         handle_t *handle;
 767         int err, block_count = 0, blocksize, size, boffs;
 768
 769         /* Determine how many transaction credits are needed */
 770         blocksize = 1 << inode->i_blkbits;
 771         block_count = (*offs & (blocksize - 1)) + bufsize;
 772         block_count = (block_count + blocksize - 1) >> inode->i_blkbits;
 773
 774         journal = EXT3_SB(inode->i_sb)->s_journal;
 775         lock_kernel();
 776         handle = journal_start(journal,
 777                                block_count * EXT3_DATA_TRANS_BLOCKS + 2);
 778         unlock_kernel();
 779         if (IS_ERR(handle)) {
 780                 CERROR("can't start transaction\n");
 781                 return PTR_ERR(handle);
 782         }
 783
 784         while (bufsize > 0) {
 785                 if (bh != NULL)
 786                         brelse(bh);
 787
 788                 block = offset >> inode->i_blkbits;
 789                 boffs = offset & (blocksize - 1);
 790                 size = min(blocksize - boffs, bufsize);
 791                 bh = ext3_bread(handle, inode, block, 1, &err);
 792                 if (!bh) {
 793                         CERROR("can't read/create block: %d\n", err);
 794                         goto out;
 795                 }
 796
 797                 err = ext3_journal_get_write_access(handle, bh);
 798                 if (err) {
 799                         CERROR("journal_get_write_access() returned error %d\n",
 800                                err);
 801                         goto out;
 802                 }
 803                 LASSERT(bh->b_data + boffs + size <= bh->b_data + bh->b_size);
 804                 memcpy(bh->b_data + boffs, buf, size);
 805                 err = ext3_journal_dirty_metadata(handle, bh);
 806                 if (err) {
 807                         CERROR("journal_dirty_metadata() returned error %d\n",
 808                                err);
 809                         goto out;
 810                 }
 811                 if (offset + size > new_size)
 812                         new_size = offset + size;
 813                 offset += size;
 814                 bufsize -= size;
 815                 buf += size;
 816         }
 817
 818         if (force_sync)
 819                 handle->h_sync = 1; /* recovery likes this */
 820 out:
 821         if (bh)
 822                 brelse(bh);
 823
 824         /* correct in-core and on-disk sizes */
 825         if (new_size > inode->i_size) {
 826                 lock_kernel();
 827                 if (new_size > inode->i_size)
 828                         inode->i_size = new_size;
 829                 if (inode->i_size > EXT3_I(inode)->i_disksize)
 830                         EXT3_I(inode)->i_disksize = inode->i_size;
 831                 if (inode->i_size > old_size)
 832                         mark_inode_dirty(inode);
 833                 unlock_kernel();
 834         }
 835
 836         lock_kernel();
 837         journal_stop(handle);
 838         unlock_kernel();
 839
 840         if (err == 0)
 841                 *offs = offset;
 842         return err;
 843 }
 844
 845 static int fsfilt_ext3_setup(struct obd_device *obd, struct super_block *sb)
 846 {
 847         struct mds_obd *mds = &obd->u.mds;
 848 #if 0
 849         EXT3_SB(sb)->dx_lock = fsfilt_ext3_dx_lock;
 850         EXT3_SB(sb)->dx_unlock = fsfilt_ext3_dx_unlock;
 851 #endif
 852 #ifdef S_PDIROPS
 853         CWARN("Enabling PDIROPS\n");
 854         set_opt(EXT3_SB(sb)->s_mount_opt, PDIROPS);
 855         sb->s_flags |= S_PDIROPS;
 856 #endif
 857         /* setup mdsnum in underlying fs */
 858 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
 859         if (mds->mds_lmv_obd) {
 860                 struct ext3_sb_info *sbi = EXT3_SB(sb);
 861                 struct ext3_super_block *es = sbi->s_es;
 862                 handle_t *handle;
 863                 int err;
 864
 865                 if (!EXT3_HAS_INCOMPAT_FEATURE(sb, EXT3_FEATURE_INCOMPAT_MDSNUM)) {
 866                         CWARN("%s: set mdsnum %d in ext3fs\n",
 867                                         obd->obd_name, mds->mds_num);
 868                         lock_kernel();
 869                         handle = journal_start(sbi->s_journal, 1);
 870                         unlock_kernel();
 871                         LASSERT(!IS_ERR(handle));
 872                         err = ext3_journal_get_write_access(handle, sbi->s_sbh);
 873                         LASSERT(err == 0);
 874                         EXT3_SET_INCOMPAT_FEATURE(sb,
 875                                                 EXT3_FEATURE_INCOMPAT_MDSNUM);
 876                         es->s_mdsnum = mds->mds_num;
 877                         err = ext3_journal_dirty_metadata(handle, sbi->s_sbh);
 878                         LASSERT(err == 0);
 879                         lock_kernel();
 880                         journal_stop(handle);
 881                         unlock_kernel();
 882                 } else {
 883                         CWARN("%s: mdsnum initialized to %u in ext3fs\n",
 884                                 obd->obd_name, es->s_mdsnum);
 885                 }
 886                 sbi->s_mdsnum = es->s_mdsnum;
 887         }
 888 #endif
 889         return 0;
 890 }
 891
 892 extern int ext3_add_dir_entry(struct dentry *dentry);
 893 extern int ext3_del_dir_entry(struct dentry *dentry);
 894
 895 static int fsfilt_ext3_add_dir_entry(struct obd_device *obd,
 896                                  struct dentry *parent,
 897                                  char *name, int namelen,
 898                                  unsigned long ino,
 899                                  unsigned long generation,
 900                                  unsigned mds)
 901 {
 902 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
 903         struct dentry *dentry;
 904         int err;
 905         dentry = ll_lookup_one_len(name, parent, namelen);
 906         if (IS_ERR(dentry))
 907                 RETURN(PTR_ERR(dentry));
 908         if (dentry->d_inode != NULL) {
 909                 CERROR("dentry %*s(0x%p) found\n", dentry->d_name.len,
 910                        dentry->d_name.name, dentry);
 911                 l_dput(dentry);
 912                 RETURN(-EEXIST);
 913         }
 914
 915         /* mds_reint_rename() may use this method to add dir entry
 916          * that points onto local inode. and we don't want to find
 917          * it cross-ref by subsequent lookups */
 918         d_drop(dentry);
 919
 920         dentry->d_flags |= DCACHE_CROSS_REF;
 921         dentry->d_inum = ino;
 922         dentry->d_mdsnum = mds;
 923         dentry->d_generation = generation;
 924         lock_kernel();
 925         err = ext3_add_dir_entry(dentry);
 926         unlock_kernel();
 927
 928         l_dput(dentry);
 929
 930         RETURN(err);
 931 #else
 932 #error "rebuild kernel and lustre with ext3-mds-num patch!"
 933         LASSERT(0);
 934 #endif
 935 }
 936
 937 static int fsfilt_ext3_del_dir_entry(struct obd_device *obd,
 938                                  struct dentry *dentry)
 939 {
 940 #ifdef EXT3_FEATURE_INCOMPAT_MDSNUM
 941         int err;
 942         lock_kernel();
 943         err = ext3_del_dir_entry(dentry);
 944         unlock_kernel();
 945         if (err == 0)
 946                 d_drop(dentry);
 947         return err;
 948 #else
 949 #error "rebuild kernel and lustre with ext3-mds-num patch!"
 950         LASSERT(0);
 951 #endif
 952 }
 953
 954 /* If fso is NULL, op is FSFILT operation, otherwise op is number of fso
 955    objects. Logs is number of logfiles to update */
 956 static int fsfilt_ext3_get_op_len(int op, struct fsfilt_objinfo *fso, int logs)
 957 {
 958         if ( !fso ) {
 959                 switch(op) {
 960                 case FSFILT_OP_CREATE:
 961                                  /* directory leaf, index & indirect & EA*/
 962                         return 4 + 3 * logs;
 963                 case FSFILT_OP_UNLINK:
 964                         return 3 * logs;
 965                 }
 966         } else {
 967                 int i;
 968                 int needed = 0;
 969                 struct super_block *sb = fso->fso_dentry->d_inode->i_sb;
 970                 int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 971                 int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
 972                 for (i = 0; i < op; i++, fso++) {
 973                         int nblocks = fso->fso_bufcnt * blockpp;
 974                         int ndindirect = min(nblocks, addrpp + 1);
 975                         int nindir = nblocks + ndindirect + 1;
 976
 977                         needed += nindir;
 978                 }
 979                 return needed + 3 * logs;
 980         }
 981
 982         return 0;
 983 }
 984
 985 static struct fsfilt_operations fsfilt_ext3_ops = {
 986         fs_type:                "ext3",
 987         fs_owner:               THIS_MODULE,
 988         fs_start:               fsfilt_ext3_start,
 989         fs_brw_start:           fsfilt_ext3_brw_start,
 990         fs_commit:              fsfilt_ext3_commit,
 991         fs_commit_async:        fsfilt_ext3_commit_async,
 992         fs_commit_wait:         fsfilt_ext3_commit_wait,
 993         fs_setattr:             fsfilt_ext3_setattr,
 994         fs_iocontrol:           fsfilt_ext3_iocontrol,
 995         fs_set_md:              fsfilt_ext3_set_md,
 996         fs_get_md:              fsfilt_ext3_get_md,
 997         fs_readpage:            fsfilt_ext3_readpage,
 998         fs_add_journal_cb:      fsfilt_ext3_add_journal_cb,
 999         fs_statfs:              fsfilt_ext3_statfs,
1000         fs_sync:                fsfilt_ext3_sync,
1001         fs_map_inode_page:      fsfilt_ext3_map_inode_page,
1002         fs_prep_san_write:      fsfilt_ext3_prep_san_write,
1003         fs_write_record:        fsfilt_ext3_write_record,
1004         fs_read_record:         fsfilt_ext3_read_record,
1005         fs_setup:               fsfilt_ext3_setup,
1006         fs_get_op_len:          fsfilt_ext3_get_op_len,
1007         fs_add_dir_entry:       fsfilt_ext3_add_dir_entry,
1008         fs_del_dir_entry:       fsfilt_ext3_del_dir_entry,
1009 };
1010
1011 static int __init fsfilt_ext3_init(void)
1012 {
1013         int rc;
1014
1015         //rc = ext3_xattr_register();
1016         fcb_cache = kmem_cache_create("fsfilt_ext3_fcb",
1017                                       sizeof(struct fsfilt_cb_data), 0,
1018                                       0, NULL, NULL);
1019         if (!fcb_cache) {
1020                 CERROR("error allocating fsfilt journal callback cache\n");
1021                 GOTO(out, rc = -ENOMEM);
1022         }
1023
1024         rc = fsfilt_register_ops(&fsfilt_ext3_ops);
1025
1026         if (rc)
1027                 kmem_cache_destroy(fcb_cache);
1028 out:
1029         return rc;
1030 }
1031
1032 static void __exit fsfilt_ext3_exit(void)
1033 {
1034         int rc;
1035
1036         fsfilt_unregister_ops(&fsfilt_ext3_ops);
1037         rc = kmem_cache_destroy(fcb_cache);
1038
1039         if (rc || atomic_read(&fcb_cache_count)) {
1040                 CERROR("can't free fsfilt callback cache: count %d, rc = %d\n",
1041                        atomic_read(&fcb_cache_count), rc);
1042         }
1043
1044         //rc = ext3_xattr_unregister();
1045 }
1046
1047 module_init(fsfilt_ext3_init);
1048 module_exit(fsfilt_ext3_exit);
1049
1050 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1051 MODULE_DESCRIPTION("Lustre ext3 Filesystem Helper v0.1");
1052 MODULE_LICENSE("GPL");