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