git.oblomov.eu Git - linux-2.6/blob - fs/ocfs2/extent_map.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * extent_map.c
   5  *
   6  * In-memory extent map for OCFS2.  Man, this code was prettier in
   7  * the library.
   8  *
   9  * Copyright (C) 2004 Oracle.  All rights reserved.
  10  *
  11  * This program is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU General Public
  13  * License, version 2,  as published by the Free Software Foundation.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  */
  25
  26 #include <linux/fs.h>
  27 #include <linux/init.h>
  28 #include <linux/types.h>
  29 #include <linux/slab.h>
  30 #include <linux/rbtree.h>
  31
  32 #define MLOG_MASK_PREFIX ML_EXTENT_MAP
  33 #include <cluster/masklog.h>
  34
  35 #include "ocfs2.h"
  36
  37 #include "extent_map.h"
  38 #include "inode.h"
  39 #include "super.h"
  40
  41 #include "buffer_head_io.h"
  42
  43
  44 /*
  45  * SUCK SUCK SUCK
  46  * Our headers are so bad that struct ocfs2_extent_map is in ocfs.h
  47  */
  48
  49 struct ocfs2_extent_map_entry {
  50         struct rb_node e_node;
  51         int e_tree_depth;
  52         struct ocfs2_extent_rec e_rec;
  53 };
  54
  55 struct ocfs2_em_insert_context {
  56         int need_left;
  57         int need_right;
  58         struct ocfs2_extent_map_entry *new_ent;
  59         struct ocfs2_extent_map_entry *old_ent;
  60         struct ocfs2_extent_map_entry *left_ent;
  61         struct ocfs2_extent_map_entry *right_ent;
  62 };
  63
  64 static kmem_cache_t *ocfs2_em_ent_cachep = NULL;
  65
  66
  67 static struct ocfs2_extent_map_entry *
  68 ocfs2_extent_map_lookup(struct ocfs2_extent_map *em,
  69                         u32 cpos, u32 clusters,
  70                         struct rb_node ***ret_p,
  71                         struct rb_node **ret_parent);
  72 static int ocfs2_extent_map_insert(struct inode *inode,
  73                                    struct ocfs2_extent_rec *rec,
  74                                    int tree_depth);
  75 static int ocfs2_extent_map_insert_entry(struct ocfs2_extent_map *em,
  76                                          struct ocfs2_extent_map_entry *ent);
  77 static int ocfs2_extent_map_find_leaf(struct inode *inode,
  78                                       u32 cpos, u32 clusters,
  79                                       struct ocfs2_extent_list *el);
  80 static int ocfs2_extent_map_lookup_read(struct inode *inode,
  81                                         u32 cpos, u32 clusters,
  82                                         struct ocfs2_extent_map_entry **ret_ent);
  83 static int ocfs2_extent_map_try_insert(struct inode *inode,
  84                                        struct ocfs2_extent_rec *rec,
  85                                        int tree_depth,
  86                                        struct ocfs2_em_insert_context *ctxt);
  87
  88 /* returns 1 only if the rec contains all the given clusters -- that is that
  89  * rec's cpos is <= the cluster cpos and that the rec endpoint (cpos +
  90  * clusters) is >= the argument's endpoint */
  91 static int ocfs2_extent_rec_contains_clusters(struct ocfs2_extent_rec *rec,
  92                                               u32 cpos, u32 clusters)
  93 {
  94         if (le32_to_cpu(rec->e_cpos) > cpos)
  95                 return 0;
  96         if (cpos + clusters > le32_to_cpu(rec->e_cpos) +
  97                               le32_to_cpu(rec->e_clusters))
  98                 return 0;
  99         return 1;
 100 }
 101
 102
 103 /*
 104  * Find an entry in the tree that intersects the region passed in.
 105  * Note that this will find straddled intervals, it is up to the
 106  * callers to enforce any boundary conditions.
 107  *
 108  * Callers must hold ip_lock.  This lookup is not guaranteed to return
 109  * a tree_depth 0 match, and as such can race inserts if the lock
 110  * were not held.
 111  *
 112  * The rb_node garbage lets insertion share the search.  Trivial
 113  * callers pass NULL.
 114  */
 115 static struct ocfs2_extent_map_entry *
 116 ocfs2_extent_map_lookup(struct ocfs2_extent_map *em,
 117                         u32 cpos, u32 clusters,
 118                         struct rb_node ***ret_p,
 119                         struct rb_node **ret_parent)
 120 {
 121         struct rb_node **p = &em->em_extents.rb_node;
 122         struct rb_node *parent = NULL;
 123         struct ocfs2_extent_map_entry *ent = NULL;
 124
 125         while (*p)
 126         {
 127                 parent = *p;
 128                 ent = rb_entry(parent, struct ocfs2_extent_map_entry,
 129                                e_node);
 130                 if ((cpos + clusters) <= le32_to_cpu(ent->e_rec.e_cpos)) {
 131                         p = &(*p)->rb_left;
 132                         ent = NULL;
 133                 } else if (cpos >= (le32_to_cpu(ent->e_rec.e_cpos) +
 134                                     le32_to_cpu(ent->e_rec.e_clusters))) {
 135                         p = &(*p)->rb_right;
 136                         ent = NULL;
 137                 } else
 138                         break;
 139         }
 140
 141         if (ret_p != NULL)
 142                 *ret_p = p;
 143         if (ret_parent != NULL)
 144                 *ret_parent = parent;
 145         return ent;
 146 }
 147
 148 /*
 149  * Find the leaf containing the interval we want.  While we're on our
 150  * way down the tree, fill in every record we see at any depth, because
 151  * we might want it later.
 152  *
 153  * Note that this code is run without ip_lock.  That's because it
 154  * sleeps while reading.  If someone is also filling the extent list at
 155  * the same time we are, we might have to restart.
 156  */
 157 static int ocfs2_extent_map_find_leaf(struct inode *inode,
 158                                       u32 cpos, u32 clusters,
 159                                       struct ocfs2_extent_list *el)
 160 {
 161         int i, ret;
 162         struct buffer_head *eb_bh = NULL;
 163         u64 blkno;
 164         u32 rec_end;
 165         struct ocfs2_extent_block *eb;
 166         struct ocfs2_extent_rec *rec;
 167
 168         /*
 169          * The bh data containing the el cannot change here, because
 170          * we hold alloc_sem.  So we can do this without other
 171          * locks.
 172          */
 173         while (el->l_tree_depth)
 174         {
 175                 blkno = 0;
 176                 for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
 177                         rec = &el->l_recs[i];
 178                         rec_end = (le32_to_cpu(rec->e_cpos) +
 179                                    le32_to_cpu(rec->e_clusters));
 180
 181                         ret = -EBADR;
 182                         if (rec_end > OCFS2_I(inode)->ip_clusters) {
 183                                 mlog_errno(ret);
 184                                 goto out_free;
 185                         }
 186
 187                         if (rec_end <= cpos) {
 188                                 ret = ocfs2_extent_map_insert(inode, rec,
 189                                                 le16_to_cpu(el->l_tree_depth));
 190                                 if (ret && (ret != -EEXIST)) {
 191                                         mlog_errno(ret);
 192                                         goto out_free;
 193                                 }
 194                                 continue;
 195                         }
 196                         if ((cpos + clusters) <= le32_to_cpu(rec->e_cpos)) {
 197                                 ret = ocfs2_extent_map_insert(inode, rec,
 198                                                 le16_to_cpu(el->l_tree_depth));
 199                                 if (ret && (ret != -EEXIST)) {
 200                                         mlog_errno(ret);
 201                                         goto out_free;
 202                                 }
 203                                 continue;
 204                         }
 205
 206                         /*
 207                          * We've found a record that matches our
 208                          * interval.  We don't insert it because we're
 209                          * about to traverse it.
 210                          */
 211
 212                         /* Check to see if we're stradling */
 213                         ret = -ESRCH;
 214                         if (!ocfs2_extent_rec_contains_clusters(rec,
 215                                                                 cpos,
 216                                                                 clusters)) {
 217                                 mlog_errno(ret);
 218                                 goto out_free;
 219                         }
 220
 221                         /*
 222                          * If we've already found a record, the el has
 223                          * two records covering the same interval.
 224                          * EEEK!
 225                          */
 226                         ret = -EBADR;
 227                         if (blkno) {
 228                                 mlog_errno(ret);
 229                                 goto out_free;
 230                         }
 231
 232                         blkno = le64_to_cpu(rec->e_blkno);
 233                 }
 234
 235                 /*
 236                  * We don't support holes, and we're still up
 237                  * in the branches, so we'd better have found someone
 238                  */
 239                 ret = -EBADR;
 240                 if (!blkno) {
 241                         mlog_errno(ret);
 242                         goto out_free;
 243                 }
 244
 245                 if (eb_bh) {
 246                         brelse(eb_bh);
 247                         eb_bh = NULL;
 248                 }
 249                 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
 250                                        blkno, &eb_bh, OCFS2_BH_CACHED,
 251                                        inode);
 252                 if (ret) {
 253                         mlog_errno(ret);
 254                         goto out_free;
 255                 }
 256                 eb = (struct ocfs2_extent_block *)eb_bh->b_data;
 257                 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
 258                         OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
 259                         ret = -EIO;
 260                         goto out_free;
 261                 }
 262                 el = &eb->h_list;
 263         }
 264
 265         BUG_ON(el->l_tree_depth);
 266
 267         for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
 268                 rec = &el->l_recs[i];
 269                 ret = ocfs2_extent_map_insert(inode, rec,
 270                                               le16_to_cpu(el->l_tree_depth));
 271                 if (ret) {
 272                         mlog_errno(ret);
 273                         goto out_free;
 274                 }
 275         }
 276
 277         ret = 0;
 278
 279 out_free:
 280         if (eb_bh)
 281                 brelse(eb_bh);
 282
 283         return ret;
 284 }
 285
 286 /*
 287  * This lookup actually will read from disk.  It has one invariant:
 288  * It will never re-traverse blocks.  This means that all inserts should
 289  * be new regions or more granular regions (both allowed by insert).
 290  */
 291 static int ocfs2_extent_map_lookup_read(struct inode *inode,
 292                                         u32 cpos,
 293                                         u32 clusters,
 294                                         struct ocfs2_extent_map_entry **ret_ent)
 295 {
 296         int ret;
 297         u64 blkno;
 298         struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
 299         struct ocfs2_extent_map_entry *ent;
 300         struct buffer_head *bh = NULL;
 301         struct ocfs2_extent_block *eb;
 302         struct ocfs2_dinode *di;
 303         struct ocfs2_extent_list *el;
 304
 305         spin_lock(&OCFS2_I(inode)->ip_lock);
 306         ent = ocfs2_extent_map_lookup(em, cpos, clusters, NULL, NULL);
 307         if (ent) {
 308                 if (!ent->e_tree_depth) {
 309                         spin_unlock(&OCFS2_I(inode)->ip_lock);
 310                         *ret_ent = ent;
 311                         return 0;
 312                 }
 313                 blkno = le64_to_cpu(ent->e_rec.e_blkno);
 314                 spin_unlock(&OCFS2_I(inode)->ip_lock);
 315
 316                 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb), blkno, &bh,
 317                                        OCFS2_BH_CACHED, inode);
 318                 if (ret) {
 319                         mlog_errno(ret);
 320                         if (bh)
 321                                 brelse(bh);
 322                         return ret;
 323                 }
 324                 eb = (struct ocfs2_extent_block *)bh->b_data;
 325                 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
 326                         OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
 327                         brelse(bh);
 328                         return -EIO;
 329                 }
 330                 el = &eb->h_list;
 331         } else {
 332                 spin_unlock(&OCFS2_I(inode)->ip_lock);
 333
 334                 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
 335                                        OCFS2_I(inode)->ip_blkno, &bh,
 336                                        OCFS2_BH_CACHED, inode);
 337                 if (ret) {
 338                         mlog_errno(ret);
 339                         if (bh)
 340                                 brelse(bh);
 341                         return ret;
 342                 }
 343                 di = (struct ocfs2_dinode *)bh->b_data;
 344                 if (!OCFS2_IS_VALID_DINODE(di)) {
 345                         brelse(bh);
 346                         OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, di);
 347                         return -EIO;
 348                 }
 349                 el = &di->id2.i_list;
 350         }
 351
 352         ret = ocfs2_extent_map_find_leaf(inode, cpos, clusters, el);
 353         brelse(bh);
 354         if (ret) {
 355                 mlog_errno(ret);
 356                 return ret;
 357         }
 358
 359         ent = ocfs2_extent_map_lookup(em, cpos, clusters, NULL, NULL);
 360         if (!ent) {
 361                 ret = -ESRCH;
 362                 mlog_errno(ret);
 363                 return ret;
 364         }
 365
 366         /* FIXME: Make sure this isn't a corruption */
 367         BUG_ON(ent->e_tree_depth);
 368
 369         *ret_ent = ent;
 370
 371         return 0;
 372 }
 373
 374 /*
 375  * Callers must hold ip_lock.  This can insert pieces of the tree,
 376  * thus racing lookup if the lock weren't held.
 377  */
 378 static int ocfs2_extent_map_insert_entry(struct ocfs2_extent_map *em,
 379                                          struct ocfs2_extent_map_entry *ent)
 380 {
 381         struct rb_node **p, *parent;
 382         struct ocfs2_extent_map_entry *old_ent;
 383
 384         old_ent = ocfs2_extent_map_lookup(em, le32_to_cpu(ent->e_rec.e_cpos),
 385                                           le32_to_cpu(ent->e_rec.e_clusters),
 386                                           &p, &parent);
 387         if (old_ent)
 388                 return -EEXIST;
 389
 390         rb_link_node(&ent->e_node, parent, p);
 391         rb_insert_color(&ent->e_node, &em->em_extents);
 392
 393         return 0;
 394 }
 395
 396
 397 /*
 398  * Simple rule: on any return code other than -EAGAIN, anything left
 399  * in the insert_context will be freed.
 400  */
 401 static int ocfs2_extent_map_try_insert(struct inode *inode,
 402                                        struct ocfs2_extent_rec *rec,
 403                                        int tree_depth,
 404                                        struct ocfs2_em_insert_context *ctxt)
 405 {
 406         int ret;
 407         struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
 408         struct ocfs2_extent_map_entry *old_ent;
 409
 410         ctxt->need_left = 0;
 411         ctxt->need_right = 0;
 412         ctxt->old_ent = NULL;
 413
 414         spin_lock(&OCFS2_I(inode)->ip_lock);
 415         ret = ocfs2_extent_map_insert_entry(em, ctxt->new_ent);
 416         if (!ret) {
 417                 ctxt->new_ent = NULL;
 418                 goto out_unlock;
 419         }
 420
 421         old_ent = ocfs2_extent_map_lookup(em, le32_to_cpu(rec->e_cpos),
 422                                           le32_to_cpu(rec->e_clusters), NULL,
 423                                           NULL);
 424
 425         BUG_ON(!old_ent);
 426
 427         ret = -EEXIST;
 428         if (old_ent->e_tree_depth < tree_depth)
 429                 goto out_unlock;
 430
 431         if (old_ent->e_tree_depth == tree_depth) {
 432                 if (!memcmp(rec, &old_ent->e_rec,
 433                             sizeof(struct ocfs2_extent_rec)))
 434                         ret = 0;
 435
 436                 /* FIXME: Should this be ESRCH/EBADR??? */
 437                 goto out_unlock;
 438         }
 439
 440         /*
 441          * We do it in this order specifically so that no actual tree
 442          * changes occur until we have all the pieces we need.  We
 443          * don't want malloc failures to leave an inconsistent tree.
 444          * Whenever we drop the lock, another process could be
 445          * inserting.  Also note that, if another process just beat us
 446          * to an insert, we might not need the same pieces we needed
 447          * the first go round.  In the end, the pieces we need will
 448          * be used, and the pieces we don't will be freed.
 449          */
 450         ctxt->need_left = !!(le32_to_cpu(rec->e_cpos) >
 451                              le32_to_cpu(old_ent->e_rec.e_cpos));
 452         ctxt->need_right = !!((le32_to_cpu(old_ent->e_rec.e_cpos) +
 453                                le32_to_cpu(old_ent->e_rec.e_clusters)) >
 454                               (le32_to_cpu(rec->e_cpos) + le32_to_cpu(rec->e_clusters)));
 455         ret = -EAGAIN;
 456         if (ctxt->need_left) {
 457                 if (!ctxt->left_ent)
 458                         goto out_unlock;
 459                 *(ctxt->left_ent) = *old_ent;
 460                 ctxt->left_ent->e_rec.e_clusters =
 461                         cpu_to_le32(le32_to_cpu(rec->e_cpos) -
 462                                     le32_to_cpu(ctxt->left_ent->e_rec.e_cpos));
 463         }
 464         if (ctxt->need_right) {
 465                 if (!ctxt->right_ent)
 466                         goto out_unlock;
 467                 *(ctxt->right_ent) = *old_ent;
 468                 ctxt->right_ent->e_rec.e_cpos =
 469                         cpu_to_le32(le32_to_cpu(rec->e_cpos) +
 470                                     le32_to_cpu(rec->e_clusters));
 471                 ctxt->right_ent->e_rec.e_clusters =
 472                         cpu_to_le32((le32_to_cpu(old_ent->e_rec.e_cpos) +
 473                                      le32_to_cpu(old_ent->e_rec.e_clusters)) -
 474                                     le32_to_cpu(ctxt->right_ent->e_rec.e_cpos));
 475         }
 476
 477         rb_erase(&old_ent->e_node, &em->em_extents);
 478         /* Now that he's erased, set him up for deletion */
 479         ctxt->old_ent = old_ent;
 480
 481         if (ctxt->need_left) {
 482                 ret = ocfs2_extent_map_insert_entry(em,
 483                                                     ctxt->left_ent);
 484                 if (ret)
 485                         goto out_unlock;
 486                 ctxt->left_ent = NULL;
 487         }
 488
 489         if (ctxt->need_right) {
 490                 ret = ocfs2_extent_map_insert_entry(em,
 491                                                     ctxt->right_ent);
 492                 if (ret)
 493                         goto out_unlock;
 494                 ctxt->right_ent = NULL;
 495         }
 496
 497         ret = ocfs2_extent_map_insert_entry(em, ctxt->new_ent);
 498
 499         if (!ret)
 500                 ctxt->new_ent = NULL;
 501
 502 out_unlock:
 503         spin_unlock(&OCFS2_I(inode)->ip_lock);
 504
 505         return ret;
 506 }
 507
 508
 509 static int ocfs2_extent_map_insert(struct inode *inode,
 510                                    struct ocfs2_extent_rec *rec,
 511                                    int tree_depth)
 512 {
 513         int ret;
 514         struct ocfs2_em_insert_context ctxt = {0, };
 515
 516         if ((le32_to_cpu(rec->e_cpos) + le32_to_cpu(rec->e_clusters)) >
 517             OCFS2_I(inode)->ip_map.em_clusters) {
 518                 ret = -EBADR;
 519                 mlog_errno(ret);
 520                 return ret;
 521         }
 522
 523         /* Zero e_clusters means a truncated tail record.  It better be EOF */
 524         if (!rec->e_clusters) {
 525                 if ((le32_to_cpu(rec->e_cpos) + le32_to_cpu(rec->e_clusters)) !=
 526                     OCFS2_I(inode)->ip_map.em_clusters) {
 527                         ret = -EBADR;
 528                         mlog_errno(ret);
 529                         return ret;
 530                 }
 531
 532                 /* Ignore the truncated tail */
 533                 return 0;
 534         }
 535
 536         ret = -ENOMEM;
 537         ctxt.new_ent = kmem_cache_alloc(ocfs2_em_ent_cachep,
 538                                         GFP_KERNEL);
 539         if (!ctxt.new_ent) {
 540                 mlog_errno(ret);
 541                 return ret;
 542         }
 543
 544         ctxt.new_ent->e_rec = *rec;
 545         ctxt.new_ent->e_tree_depth = tree_depth;
 546
 547         do {
 548                 ret = -ENOMEM;
 549                 if (ctxt.need_left && !ctxt.left_ent) {
 550                         ctxt.left_ent =
 551                                 kmem_cache_alloc(ocfs2_em_ent_cachep,
 552                                                  GFP_KERNEL);
 553                         if (!ctxt.left_ent)
 554                                 break;
 555                 }
 556                 if (ctxt.need_right && !ctxt.right_ent) {
 557                         ctxt.right_ent =
 558                                 kmem_cache_alloc(ocfs2_em_ent_cachep,
 559                                                  GFP_KERNEL);
 560                         if (!ctxt.right_ent)
 561                                 break;
 562                 }
 563
 564                 ret = ocfs2_extent_map_try_insert(inode, rec,
 565                                                   tree_depth, &ctxt);
 566         } while (ret == -EAGAIN);
 567
 568         if (ret < 0)
 569                 mlog_errno(ret);
 570
 571         if (ctxt.left_ent)
 572                 kmem_cache_free(ocfs2_em_ent_cachep, ctxt.left_ent);
 573         if (ctxt.right_ent)
 574                 kmem_cache_free(ocfs2_em_ent_cachep, ctxt.right_ent);
 575         if (ctxt.old_ent)
 576                 kmem_cache_free(ocfs2_em_ent_cachep, ctxt.old_ent);
 577         if (ctxt.new_ent)
 578                 kmem_cache_free(ocfs2_em_ent_cachep, ctxt.new_ent);
 579
 580         return ret;
 581 }
 582
 583 /*
 584  * Append this record to the tail of the extent map.  It must be
 585  * tree_depth 0.  The record might be an extension of an existing
 586  * record, and as such that needs to be handled.  eg:
 587  *
 588  * Existing record in the extent map:
 589  *
 590  *      cpos = 10, len = 10
 591  *      |---------|
 592  *
 593  * New Record:
 594  *
 595  *      cpos = 10, len = 20
 596  *      |------------------|
 597  *
 598  * The passed record is the new on-disk record.  The new_clusters value
 599  * is how many clusters were added to the file.  If the append is a
 600  * contiguous append, the new_clusters has been added to
 601  * rec->e_clusters.  If the append is an entirely new extent, then
 602  * rec->e_clusters is == new_clusters.
 603  */
 604 int ocfs2_extent_map_append(struct inode *inode,
 605                             struct ocfs2_extent_rec *rec,
 606                             u32 new_clusters)
 607 {
 608         int ret;
 609         struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
 610         struct ocfs2_extent_map_entry *ent;
 611         struct ocfs2_extent_rec *old;
 612
 613         BUG_ON(!new_clusters);
 614         BUG_ON(le32_to_cpu(rec->e_clusters) < new_clusters);
 615
 616         if (em->em_clusters < OCFS2_I(inode)->ip_clusters) {
 617                 /*
 618                  * Size changed underneath us on disk.  Drop any
 619                  * straddling records and update our idea of
 620                  * i_clusters
 621                  */
 622                 ocfs2_extent_map_drop(inode, em->em_clusters - 1);
 623                 em->em_clusters = OCFS2_I(inode)->ip_clusters;
 624         }
 625
 626         mlog_bug_on_msg((le32_to_cpu(rec->e_cpos) +
 627                          le32_to_cpu(rec->e_clusters)) !=
 628                         (em->em_clusters + new_clusters),
 629                         "Inode %"MLFu64":\n"
 630                         "rec->e_cpos = %u + rec->e_clusters = %u = %u\n"
 631                         "em->em_clusters = %u + new_clusters = %u = %u\n",
 632                         OCFS2_I(inode)->ip_blkno,
 633                         le32_to_cpu(rec->e_cpos), le32_to_cpu(rec->e_clusters),
 634                         le32_to_cpu(rec->e_cpos) + le32_to_cpu(rec->e_clusters),
 635                         em->em_clusters, new_clusters,
 636                         em->em_clusters + new_clusters);
 637
 638         em->em_clusters += new_clusters;
 639
 640         ret = -ENOENT;
 641         if (le32_to_cpu(rec->e_clusters) > new_clusters) {
 642                 /* This is a contiguous append */
 643                 ent = ocfs2_extent_map_lookup(em, le32_to_cpu(rec->e_cpos), 1,
 644                                               NULL, NULL);
 645                 if (ent) {
 646                         old = &ent->e_rec;
 647                         BUG_ON((le32_to_cpu(rec->e_cpos) +
 648                                 le32_to_cpu(rec->e_clusters)) !=
 649                                  (le32_to_cpu(old->e_cpos) +
 650                                   le32_to_cpu(old->e_clusters) +
 651                                   new_clusters));
 652                         if (ent->e_tree_depth == 0) {
 653                                 BUG_ON(le32_to_cpu(old->e_cpos) !=
 654                                        le32_to_cpu(rec->e_cpos));
 655                                 BUG_ON(le64_to_cpu(old->e_blkno) !=
 656                                        le64_to_cpu(rec->e_blkno));
 657                                 ret = 0;
 658                         }
 659                         /*
 660                          * Let non-leafs fall through as -ENOENT to
 661                          * force insertion of the new leaf.
 662                          */
 663                         le32_add_cpu(&old->e_clusters, new_clusters);
 664                 }
 665         }
 666
 667         if (ret == -ENOENT)
 668                 ret = ocfs2_extent_map_insert(inode, rec, 0);
 669         if (ret < 0)
 670                 mlog_errno(ret);
 671         return ret;
 672 }
 673
 674 #if 0
 675 /* Code here is included but defined out as it completes the extent
 676  * map api and may be used in the future. */
 677
 678 /*
 679  * Look up the record containing this cluster offset.  This record is
 680  * part of the extent map.  Do not free it.  Any changes you make to
 681  * it will reflect in the extent map.  So, if your last extent
 682  * is (cpos = 10, clusters = 10) and you truncate the file by 5
 683  * clusters, you can do:
 684  *
 685  * ret = ocfs2_extent_map_get_rec(em, orig_size - 5, &rec);
 686  * rec->e_clusters -= 5;
 687  *
 688  * The lookup does not read from disk.  If the map isn't filled in for
 689  * an entry, you won't find it.
 690  *
 691  * Also note that the returned record is valid until alloc_sem is
 692  * dropped.  After that, truncate and extend can happen.  Caveat Emptor.
 693  */
 694 int ocfs2_extent_map_get_rec(struct inode *inode, u32 cpos,
 695                              struct ocfs2_extent_rec **rec,
 696                              int *tree_depth)
 697 {
 698         int ret = -ENOENT;
 699         struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
 700         struct ocfs2_extent_map_entry *ent;
 701
 702         *rec = NULL;
 703
 704         if (cpos >= OCFS2_I(inode)->ip_clusters)
 705                 return -EINVAL;
 706
 707         if (cpos >= em->em_clusters) {
 708                 /*
 709                  * Size changed underneath us on disk.  Drop any
 710                  * straddling records and update our idea of
 711                  * i_clusters
 712                  */
 713                 ocfs2_extent_map_drop(inode, em->em_clusters - 1);
 714                 em->em_clusters = OCFS2_I(inode)->ip_clusters ;
 715         }
 716
 717         ent = ocfs2_extent_map_lookup(&OCFS2_I(inode)->ip_map, cpos, 1,
 718                                       NULL, NULL);
 719
 720         if (ent) {
 721                 *rec = &ent->e_rec;
 722                 if (tree_depth)
 723                         *tree_depth = ent->e_tree_depth;
 724                 ret = 0;
 725         }
 726
 727         return ret;
 728 }
 729
 730 int ocfs2_extent_map_get_clusters(struct inode *inode,
 731                                   u32 v_cpos, int count,
 732                                   u32 *p_cpos, int *ret_count)
 733 {
 734         int ret;
 735         u32 coff, ccount;
 736         struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
 737         struct ocfs2_extent_map_entry *ent = NULL;
 738
 739         *p_cpos = ccount = 0;
 740
 741         if ((v_cpos + count) > OCFS2_I(inode)->ip_clusters)
 742                 return -EINVAL;
 743
 744         if ((v_cpos + count) > em->em_clusters) {
 745                 /*
 746                  * Size changed underneath us on disk.  Drop any
 747                  * straddling records and update our idea of
 748                  * i_clusters
 749                  */
 750                 ocfs2_extent_map_drop(inode, em->em_clusters - 1);
 751                 em->em_clusters = OCFS2_I(inode)->ip_clusters;
 752         }
 753
 754
 755         ret = ocfs2_extent_map_lookup_read(inode, v_cpos, count, &ent);
 756         if (ret)
 757                 return ret;
 758
 759         if (ent) {
 760                 /* We should never find ourselves straddling an interval */
 761                 if (!ocfs2_extent_rec_contains_clusters(&ent->e_rec,
 762                                                         v_cpos,
 763                                                         count))
 764                         return -ESRCH;
 765
 766                 coff = v_cpos - le32_to_cpu(ent->e_rec.e_cpos);
 767                 *p_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 768                                 le64_to_cpu(ent->e_rec.e_blkno)) +
 769                           coff;
 770
 771                 if (ret_count)
 772                         *ret_count = le32_to_cpu(ent->e_rec.e_clusters) - coff;
 773
 774                 return 0;
 775         }
 776
 777
 778         return -ENOENT;
 779 }
 780
 781 #endif  /*  0  */
 782
 783 int ocfs2_extent_map_get_blocks(struct inode *inode,
 784                                 u64 v_blkno, int count,
 785                                 u64 *p_blkno, int *ret_count)
 786 {
 787         int ret;
 788         u64 boff;
 789         u32 cpos, clusters;
 790         int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1);
 791         struct ocfs2_extent_map_entry *ent = NULL;
 792         struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
 793         struct ocfs2_extent_rec *rec;
 794
 795         *p_blkno = 0;
 796
 797         cpos = ocfs2_blocks_to_clusters(inode->i_sb, v_blkno);
 798         clusters = ocfs2_blocks_to_clusters(inode->i_sb,
 799                                             (u64)count + bpc - 1);
 800         if ((cpos + clusters) > OCFS2_I(inode)->ip_clusters) {
 801                 ret = -EINVAL;
 802                 mlog_errno(ret);
 803                 return ret;
 804         }
 805
 806         if ((cpos + clusters) > em->em_clusters) {
 807                 /*
 808                  * Size changed underneath us on disk.  Drop any
 809                  * straddling records and update our idea of
 810                  * i_clusters
 811                  */
 812                 ocfs2_extent_map_drop(inode, em->em_clusters - 1);
 813                 em->em_clusters = OCFS2_I(inode)->ip_clusters;
 814         }
 815
 816         ret = ocfs2_extent_map_lookup_read(inode, cpos, clusters, &ent);
 817         if (ret) {
 818                 mlog_errno(ret);
 819                 return ret;
 820         }
 821
 822         if (ent)
 823         {
 824                 rec = &ent->e_rec;
 825
 826                 /* We should never find ourselves straddling an interval */
 827                 if (!ocfs2_extent_rec_contains_clusters(rec, cpos, clusters)) {
 828                         ret = -ESRCH;
 829                         mlog_errno(ret);
 830                         return ret;
 831                 }
 832
 833                 boff = ocfs2_clusters_to_blocks(inode->i_sb, cpos -
 834                                                 le32_to_cpu(rec->e_cpos));
 835                 boff += (v_blkno & (u64)(bpc - 1));
 836                 *p_blkno = le64_to_cpu(rec->e_blkno) + boff;
 837
 838                 if (ret_count) {
 839                         *ret_count = ocfs2_clusters_to_blocks(inode->i_sb,
 840                                         le32_to_cpu(rec->e_clusters)) - boff;
 841                 }
 842
 843                 return 0;
 844         }
 845
 846         return -ENOENT;
 847 }
 848
 849 int ocfs2_extent_map_init(struct inode *inode)
 850 {
 851         struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
 852
 853         em->em_extents = RB_ROOT;
 854         em->em_clusters = 0;
 855
 856         return 0;
 857 }
 858
 859 /* Needs the lock */
 860 static void __ocfs2_extent_map_drop(struct inode *inode,
 861                                     u32 new_clusters,
 862                                     struct rb_node **free_head,
 863                                     struct ocfs2_extent_map_entry **tail_ent)
 864 {
 865         struct rb_node *node, *next;
 866         struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
 867         struct ocfs2_extent_map_entry *ent;
 868
 869         *free_head = NULL;
 870
 871         ent = NULL;
 872         node = rb_last(&em->em_extents);
 873         while (node)
 874         {
 875                 next = rb_prev(node);
 876
 877                 ent = rb_entry(node, struct ocfs2_extent_map_entry,
 878                                e_node);
 879                 if (le32_to_cpu(ent->e_rec.e_cpos) < new_clusters)
 880                         break;
 881
 882                 rb_erase(&ent->e_node, &em->em_extents);
 883
 884                 node->rb_right = *free_head;
 885                 *free_head = node;
 886
 887                 ent = NULL;
 888                 node = next;
 889         }
 890
 891         /* Do we have an entry straddling new_clusters? */
 892         if (tail_ent) {
 893                 if (ent &&
 894                     ((le32_to_cpu(ent->e_rec.e_cpos) +
 895                       le32_to_cpu(ent->e_rec.e_clusters)) > new_clusters))
 896                         *tail_ent = ent;
 897                 else
 898                         *tail_ent = NULL;
 899         }
 900 }
 901
 902 static void __ocfs2_extent_map_drop_cleanup(struct rb_node *free_head)
 903 {
 904         struct rb_node *node;
 905         struct ocfs2_extent_map_entry *ent;
 906
 907         while (free_head) {
 908                 node = free_head;
 909                 free_head = node->rb_right;
 910
 911                 ent = rb_entry(node, struct ocfs2_extent_map_entry,
 912                                e_node);
 913                 kmem_cache_free(ocfs2_em_ent_cachep, ent);
 914         }
 915 }
 916
 917 /*
 918  * Remove all entries past new_clusters, inclusive of an entry that
 919  * contains new_clusters.  This is effectively a cache forget.
 920  *
 921  * If you want to also clip the last extent by some number of clusters,
 922  * you need to call ocfs2_extent_map_trunc().
 923  * This code does not check or modify ip_clusters.
 924  */
 925 int ocfs2_extent_map_drop(struct inode *inode, u32 new_clusters)
 926 {
 927         struct rb_node *free_head = NULL;
 928         struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
 929         struct ocfs2_extent_map_entry *ent;
 930
 931         spin_lock(&OCFS2_I(inode)->ip_lock);
 932
 933         __ocfs2_extent_map_drop(inode, new_clusters, &free_head, &ent);
 934
 935         if (ent) {
 936                 rb_erase(&ent->e_node, &em->em_extents);
 937                 ent->e_node.rb_right = free_head;
 938                 free_head = &ent->e_node;
 939         }
 940
 941         spin_unlock(&OCFS2_I(inode)->ip_lock);
 942
 943         if (free_head)
 944                 __ocfs2_extent_map_drop_cleanup(free_head);
 945
 946         return 0;
 947 }
 948
 949 /*
 950  * Remove all entries past new_clusters and also clip any extent
 951  * straddling new_clusters, if there is one.  This does not check
 952  * or modify ip_clusters
 953  */
 954 int ocfs2_extent_map_trunc(struct inode *inode, u32 new_clusters)
 955 {
 956         struct rb_node *free_head = NULL;
 957         struct ocfs2_extent_map_entry *ent = NULL;
 958
 959         spin_lock(&OCFS2_I(inode)->ip_lock);
 960
 961         __ocfs2_extent_map_drop(inode, new_clusters, &free_head, &ent);
 962
 963         if (ent)
 964                 ent->e_rec.e_clusters = cpu_to_le32(new_clusters -
 965                                                le32_to_cpu(ent->e_rec.e_cpos));
 966
 967         OCFS2_I(inode)->ip_map.em_clusters = new_clusters;
 968
 969         spin_unlock(&OCFS2_I(inode)->ip_lock);
 970
 971         if (free_head)
 972                 __ocfs2_extent_map_drop_cleanup(free_head);
 973
 974         return 0;
 975 }
 976
 977 int __init init_ocfs2_extent_maps(void)
 978 {
 979         ocfs2_em_ent_cachep =
 980                 kmem_cache_create("ocfs2_em_ent",
 981                                   sizeof(struct ocfs2_extent_map_entry),
 982                                   0, SLAB_HWCACHE_ALIGN, NULL, NULL);
 983         if (!ocfs2_em_ent_cachep)
 984                 return -ENOMEM;
 985
 986         return 0;
 987 }
 988
 989 void exit_ocfs2_extent_maps(void)
 990 {
 991         kmem_cache_destroy(ocfs2_em_ent_cachep);
 992 }