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  * Block/Cluster mapping functions
   7  *
   8  * Copyright (C) 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License, version 2,  as published by the Free Software Foundation.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public
  20  * License along with this program; if not, write to the
  21  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  22  * Boston, MA 021110-1307, USA.
  23  */
  24
  25 #include <linux/fs.h>
  26 #include <linux/init.h>
  27 #include <linux/types.h>
  28
  29 #define MLOG_MASK_PREFIX ML_EXTENT_MAP
  30 #include <cluster/masklog.h>
  31
  32 #include "ocfs2.h"
  33
  34 #include "alloc.h"
  35 #include "extent_map.h"
  36 #include "inode.h"
  37 #include "super.h"
  38
  39 #include "buffer_head_io.h"
  40
  41 /*
  42  * The extent caching implementation is intentionally trivial.
  43  *
  44  * We only cache a small number of extents stored directly on the
  45  * inode, so linear order operations are acceptable. If we ever want
  46  * to increase the size of the extent map, then these algorithms must
  47  * get smarter.
  48  */
  49
  50 void ocfs2_extent_map_init(struct inode *inode)
  51 {
  52         struct ocfs2_inode_info *oi = OCFS2_I(inode);
  53
  54         oi->ip_extent_map.em_num_items = 0;
  55         INIT_LIST_HEAD(&oi->ip_extent_map.em_list);
  56 }
  57
  58 static void __ocfs2_extent_map_lookup(struct ocfs2_extent_map *em,
  59                                       unsigned int cpos,
  60                                       struct ocfs2_extent_map_item **ret_emi)
  61 {
  62         unsigned int range;
  63         struct ocfs2_extent_map_item *emi;
  64
  65         *ret_emi = NULL;
  66
  67         list_for_each_entry(emi, &em->em_list, ei_list) {
  68                 range = emi->ei_cpos + emi->ei_clusters;
  69
  70                 if (cpos >= emi->ei_cpos && cpos < range) {
  71                         list_move(&emi->ei_list, &em->em_list);
  72
  73                         *ret_emi = emi;
  74                         break;
  75                 }
  76         }
  77 }
  78
  79 static int ocfs2_extent_map_lookup(struct inode *inode, unsigned int cpos,
  80                                    unsigned int *phys, unsigned int *len,
  81                                    unsigned int *flags)
  82 {
  83         unsigned int coff;
  84         struct ocfs2_inode_info *oi = OCFS2_I(inode);
  85         struct ocfs2_extent_map_item *emi;
  86
  87         spin_lock(&oi->ip_lock);
  88
  89         __ocfs2_extent_map_lookup(&oi->ip_extent_map, cpos, &emi);
  90         if (emi) {
  91                 coff = cpos - emi->ei_cpos;
  92                 *phys = emi->ei_phys + coff;
  93                 if (len)
  94                         *len = emi->ei_clusters - coff;
  95                 if (flags)
  96                         *flags = emi->ei_flags;
  97         }
  98
  99         spin_unlock(&oi->ip_lock);
 100
 101         if (emi == NULL)
 102                 return -ENOENT;
 103
 104         return 0;
 105 }
 106
 107 /*
 108  * Forget about all clusters equal to or greater than cpos.
 109  */
 110 void ocfs2_extent_map_trunc(struct inode *inode, unsigned int cpos)
 111 {
 112         struct list_head *p, *n;
 113         struct ocfs2_extent_map_item *emi;
 114         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 115         struct ocfs2_extent_map *em = &oi->ip_extent_map;
 116         LIST_HEAD(tmp_list);
 117         unsigned int range;
 118
 119         spin_lock(&oi->ip_lock);
 120         list_for_each_safe(p, n, &em->em_list) {
 121                 emi = list_entry(p, struct ocfs2_extent_map_item, ei_list);
 122
 123                 if (emi->ei_cpos >= cpos) {
 124                         /* Full truncate of this record. */
 125                         list_move(&emi->ei_list, &tmp_list);
 126                         BUG_ON(em->em_num_items == 0);
 127                         em->em_num_items--;
 128                         continue;
 129                 }
 130
 131                 range = emi->ei_cpos + emi->ei_clusters;
 132                 if (range > cpos) {
 133                         /* Partial truncate */
 134                         emi->ei_clusters = cpos - emi->ei_cpos;
 135                 }
 136         }
 137         spin_unlock(&oi->ip_lock);
 138
 139         list_for_each_safe(p, n, &tmp_list) {
 140                 emi = list_entry(p, struct ocfs2_extent_map_item, ei_list);
 141                 list_del(&emi->ei_list);
 142                 kfree(emi);
 143         }
 144 }
 145
 146 /*
 147  * Is any part of emi2 contained within emi1
 148  */
 149 static int ocfs2_ei_is_contained(struct ocfs2_extent_map_item *emi1,
 150                                  struct ocfs2_extent_map_item *emi2)
 151 {
 152         unsigned int range1, range2;
 153
 154         /*
 155          * Check if logical start of emi2 is inside emi1
 156          */
 157         range1 = emi1->ei_cpos + emi1->ei_clusters;
 158         if (emi2->ei_cpos >= emi1->ei_cpos && emi2->ei_cpos < range1)
 159                 return 1;
 160
 161         /*
 162          * Check if logical end of emi2 is inside emi1
 163          */
 164         range2 = emi2->ei_cpos + emi2->ei_clusters;
 165         if (range2 > emi1->ei_cpos && range2 <= range1)
 166                 return 1;
 167
 168         return 0;
 169 }
 170
 171 static void ocfs2_copy_emi_fields(struct ocfs2_extent_map_item *dest,
 172                                   struct ocfs2_extent_map_item *src)
 173 {
 174         dest->ei_cpos = src->ei_cpos;
 175         dest->ei_phys = src->ei_phys;
 176         dest->ei_clusters = src->ei_clusters;
 177         dest->ei_flags = src->ei_flags;
 178 }
 179
 180 /*
 181  * Try to merge emi with ins. Returns 1 if merge succeeds, zero
 182  * otherwise.
 183  */
 184 static int ocfs2_try_to_merge_extent_map(struct ocfs2_extent_map_item *emi,
 185                                          struct ocfs2_extent_map_item *ins)
 186 {
 187         /*
 188          * Handle contiguousness
 189          */
 190         if (ins->ei_phys == (emi->ei_phys + emi->ei_clusters) &&
 191             ins->ei_cpos == (emi->ei_cpos + emi->ei_clusters) &&
 192             ins->ei_flags == emi->ei_flags) {
 193                 emi->ei_clusters += ins->ei_clusters;
 194                 return 1;
 195         } else if ((ins->ei_phys + ins->ei_clusters) == emi->ei_phys &&
 196                    (ins->ei_cpos + ins->ei_clusters) == emi->ei_phys &&
 197                    ins->ei_flags == emi->ei_flags) {
 198                 emi->ei_phys = ins->ei_phys;
 199                 emi->ei_cpos = ins->ei_cpos;
 200                 emi->ei_clusters += ins->ei_clusters;
 201                 return 1;
 202         }
 203
 204         /*
 205          * Overlapping extents - this shouldn't happen unless we've
 206          * split an extent to change it's flags. That is exceedingly
 207          * rare, so there's no sense in trying to optimize it yet.
 208          */
 209         if (ocfs2_ei_is_contained(emi, ins) ||
 210             ocfs2_ei_is_contained(ins, emi)) {
 211                 ocfs2_copy_emi_fields(emi, ins);
 212                 return 1;
 213         }
 214
 215         /* No merge was possible. */
 216         return 0;
 217 }
 218
 219 /*
 220  * In order to reduce complexity on the caller, this insert function
 221  * is intentionally liberal in what it will accept.
 222  *
 223  * The only rule is that the truncate call *must* be used whenever
 224  * records have been deleted. This avoids inserting overlapping
 225  * records with different physical mappings.
 226  */
 227 void ocfs2_extent_map_insert_rec(struct inode *inode,
 228                                  struct ocfs2_extent_rec *rec)
 229 {
 230         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 231         struct ocfs2_extent_map *em = &oi->ip_extent_map;
 232         struct ocfs2_extent_map_item *emi, *new_emi = NULL;
 233         struct ocfs2_extent_map_item ins;
 234
 235         ins.ei_cpos = le32_to_cpu(rec->e_cpos);
 236         ins.ei_phys = ocfs2_blocks_to_clusters(inode->i_sb,
 237                                                le64_to_cpu(rec->e_blkno));
 238         ins.ei_clusters = le16_to_cpu(rec->e_leaf_clusters);
 239         ins.ei_flags = rec->e_flags;
 240
 241 search:
 242         spin_lock(&oi->ip_lock);
 243
 244         list_for_each_entry(emi, &em->em_list, ei_list) {
 245                 if (ocfs2_try_to_merge_extent_map(emi, &ins)) {
 246                         list_move(&emi->ei_list, &em->em_list);
 247                         spin_unlock(&oi->ip_lock);
 248                         goto out;
 249                 }
 250         }
 251
 252         /*
 253          * No item could be merged.
 254          *
 255          * Either allocate and add a new item, or overwrite the last recently
 256          * inserted.
 257          */
 258
 259         if (em->em_num_items < OCFS2_MAX_EXTENT_MAP_ITEMS) {
 260                 if (new_emi == NULL) {
 261                         spin_unlock(&oi->ip_lock);
 262
 263                         new_emi = kmalloc(sizeof(*new_emi), GFP_NOFS);
 264                         if (new_emi == NULL)
 265                                 goto out;
 266
 267                         goto search;
 268                 }
 269
 270                 ocfs2_copy_emi_fields(new_emi, &ins);
 271                 list_add(&new_emi->ei_list, &em->em_list);
 272                 em->em_num_items++;
 273                 new_emi = NULL;
 274         } else {
 275                 BUG_ON(list_empty(&em->em_list) || em->em_num_items == 0);
 276                 emi = list_entry(em->em_list.prev,
 277                                  struct ocfs2_extent_map_item, ei_list);
 278                 list_move(&emi->ei_list, &em->em_list);
 279                 ocfs2_copy_emi_fields(emi, &ins);
 280         }
 281
 282         spin_unlock(&oi->ip_lock);
 283
 284 out:
 285         if (new_emi)
 286                 kfree(new_emi);
 287 }
 288
 289 /*
 290  * Return the 1st index within el which contains an extent start
 291  * larger than v_cluster.
 292  */
 293 static int ocfs2_search_for_hole_index(struct ocfs2_extent_list *el,
 294                                        u32 v_cluster)
 295 {
 296         int i;
 297         struct ocfs2_extent_rec *rec;
 298
 299         for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
 300                 rec = &el->l_recs[i];
 301
 302                 if (v_cluster < le32_to_cpu(rec->e_cpos))
 303                         break;
 304         }
 305
 306         return i;
 307 }
 308
 309 /*
 310  * Figure out the size of a hole which starts at v_cluster within the given
 311  * extent list.
 312  *
 313  * If there is no more allocation past v_cluster, we return the maximum
 314  * cluster size minus v_cluster.
 315  *
 316  * If we have in-inode extents, then el points to the dinode list and
 317  * eb_bh is NULL. Otherwise, eb_bh should point to the extent block
 318  * containing el.
 319  */
 320 static int ocfs2_figure_hole_clusters(struct inode *inode,
 321                                       struct ocfs2_extent_list *el,
 322                                       struct buffer_head *eb_bh,
 323                                       u32 v_cluster,
 324                                       u32 *num_clusters)
 325 {
 326         int ret, i;
 327         struct buffer_head *next_eb_bh = NULL;
 328         struct ocfs2_extent_block *eb, *next_eb;
 329
 330         i = ocfs2_search_for_hole_index(el, v_cluster);
 331
 332         if (i == le16_to_cpu(el->l_next_free_rec) && eb_bh) {
 333                 eb = (struct ocfs2_extent_block *)eb_bh->b_data;
 334
 335                 /*
 336                  * Check the next leaf for any extents.
 337                  */
 338
 339                 if (le64_to_cpu(eb->h_next_leaf_blk) == 0ULL)
 340                         goto no_more_extents;
 341
 342                 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
 343                                        le64_to_cpu(eb->h_next_leaf_blk),
 344                                        &next_eb_bh, OCFS2_BH_CACHED, inode);
 345                 if (ret) {
 346                         mlog_errno(ret);
 347                         goto out;
 348                 }
 349                 next_eb = (struct ocfs2_extent_block *)next_eb_bh->b_data;
 350
 351                 if (!OCFS2_IS_VALID_EXTENT_BLOCK(next_eb)) {
 352                         ret = -EROFS;
 353                         OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, next_eb);
 354                         goto out;
 355                 }
 356
 357                 el = &next_eb->h_list;
 358
 359                 i = ocfs2_search_for_hole_index(el, v_cluster);
 360         }
 361
 362 no_more_extents:
 363         if (i == le16_to_cpu(el->l_next_free_rec)) {
 364                 /*
 365                  * We're at the end of our existing allocation. Just
 366                  * return the maximum number of clusters we could
 367                  * possibly allocate.
 368                  */
 369                 *num_clusters = UINT_MAX - v_cluster;
 370         } else {
 371                 *num_clusters = le32_to_cpu(el->l_recs[i].e_cpos) - v_cluster;
 372         }
 373
 374         ret = 0;
 375 out:
 376         brelse(next_eb_bh);
 377         return ret;
 378 }
 379
 380 /*
 381  * Return the index of the extent record which contains cluster #v_cluster.
 382  * -1 is returned if it was not found.
 383  *
 384  * Should work fine on interior and exterior nodes.
 385  */
 386 static int ocfs2_search_extent_list(struct ocfs2_extent_list *el,
 387                                     u32 v_cluster)
 388 {
 389         int ret = -1;
 390         int i;
 391         struct ocfs2_extent_rec *rec;
 392         u32 rec_end, rec_start, clusters;
 393
 394         for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
 395                 rec = &el->l_recs[i];
 396
 397                 rec_start = le32_to_cpu(rec->e_cpos);
 398                 clusters = ocfs2_rec_clusters(el, rec);
 399
 400                 rec_end = rec_start + clusters;
 401
 402                 if (v_cluster >= rec_start && v_cluster < rec_end) {
 403                         ret = i;
 404                         break;
 405                 }
 406         }
 407
 408         return ret;
 409 }
 410
 411 int ocfs2_get_clusters(struct inode *inode, u32 v_cluster,
 412                        u32 *p_cluster, u32 *num_clusters,
 413                        unsigned int *extent_flags)
 414 {
 415         int ret, i;
 416         unsigned int flags = 0;
 417         struct buffer_head *di_bh = NULL;
 418         struct buffer_head *eb_bh = NULL;
 419         struct ocfs2_dinode *di;
 420         struct ocfs2_extent_block *eb;
 421         struct ocfs2_extent_list *el;
 422         struct ocfs2_extent_rec *rec;
 423         u32 coff;
 424
 425         ret = ocfs2_extent_map_lookup(inode, v_cluster, p_cluster,
 426                                       num_clusters, extent_flags);
 427         if (ret == 0)
 428                 goto out;
 429
 430         ret = ocfs2_read_block(OCFS2_SB(inode->i_sb), OCFS2_I(inode)->ip_blkno,
 431                                &di_bh, OCFS2_BH_CACHED, inode);
 432         if (ret) {
 433                 mlog_errno(ret);
 434                 goto out;
 435         }
 436
 437         di = (struct ocfs2_dinode *) di_bh->b_data;
 438         el = &di->id2.i_list;
 439
 440         if (el->l_tree_depth) {
 441                 ret = ocfs2_find_leaf(inode, el, v_cluster, &eb_bh);
 442                 if (ret) {
 443                         mlog_errno(ret);
 444                         goto out;
 445                 }
 446
 447                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
 448                 el = &eb->h_list;
 449
 450                 if (el->l_tree_depth) {
 451                         ocfs2_error(inode->i_sb,
 452                                     "Inode %lu has non zero tree depth in "
 453                                     "leaf block %llu\n", inode->i_ino,
 454                                     (unsigned long long)eb_bh->b_blocknr);
 455                         ret = -EROFS;
 456                         goto out;
 457                 }
 458         }
 459
 460         i = ocfs2_search_extent_list(el, v_cluster);
 461         if (i == -1) {
 462                 /*
 463                  * A hole was found. Return some canned values that
 464                  * callers can key on. If asked for, num_clusters will
 465                  * be populated with the size of the hole.
 466                  */
 467                 *p_cluster = 0;
 468                 if (num_clusters) {
 469                         ret = ocfs2_figure_hole_clusters(inode, el, eb_bh,
 470                                                          v_cluster,
 471                                                          num_clusters);
 472                         if (ret) {
 473                                 mlog_errno(ret);
 474                                 goto out;
 475                         }
 476                 }
 477         } else {
 478                 rec = &el->l_recs[i];
 479
 480                 BUG_ON(v_cluster < le32_to_cpu(rec->e_cpos));
 481
 482                 if (!rec->e_blkno) {
 483                         ocfs2_error(inode->i_sb, "Inode %lu has bad extent "
 484                                     "record (%u, %u, 0)", inode->i_ino,
 485                                     le32_to_cpu(rec->e_cpos),
 486                                     ocfs2_rec_clusters(el, rec));
 487                         ret = -EROFS;
 488                         goto out;
 489                 }
 490
 491                 coff = v_cluster - le32_to_cpu(rec->e_cpos);
 492
 493                 *p_cluster = ocfs2_blocks_to_clusters(inode->i_sb,
 494                                                     le64_to_cpu(rec->e_blkno));
 495                 *p_cluster = *p_cluster + coff;
 496
 497                 if (num_clusters)
 498                         *num_clusters = ocfs2_rec_clusters(el, rec) - coff;
 499
 500                 flags = rec->e_flags;
 501
 502                 ocfs2_extent_map_insert_rec(inode, rec);
 503         }
 504
 505         if (extent_flags)
 506                 *extent_flags = flags;
 507
 508 out:
 509         brelse(di_bh);
 510         brelse(eb_bh);
 511         return ret;
 512 }
 513
 514 /*
 515  * This expects alloc_sem to be held. The allocation cannot change at
 516  * all while the map is in the process of being updated.
 517  */
 518 int ocfs2_extent_map_get_blocks(struct inode *inode, u64 v_blkno, u64 *p_blkno,
 519                                 u64 *ret_count, unsigned int *extent_flags)
 520 {
 521         int ret;
 522         int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1);
 523         u32 cpos, num_clusters, p_cluster;
 524         u64 boff = 0;
 525
 526         cpos = ocfs2_blocks_to_clusters(inode->i_sb, v_blkno);
 527
 528         ret = ocfs2_get_clusters(inode, cpos, &p_cluster, &num_clusters,
 529                                  extent_flags);
 530         if (ret) {
 531                 mlog_errno(ret);
 532                 goto out;
 533         }
 534
 535         /*
 536          * p_cluster == 0 indicates a hole.
 537          */
 538         if (p_cluster) {
 539                 boff = ocfs2_clusters_to_blocks(inode->i_sb, p_cluster);
 540                 boff += (v_blkno & (u64)(bpc - 1));
 541         }
 542
 543         *p_blkno = boff;
 544
 545         if (ret_count) {
 546                 *ret_count = ocfs2_clusters_to_blocks(inode->i_sb, num_clusters);
 547                 *ret_count -= v_blkno & (u64)(bpc - 1);
 548         }
 549
 550 out:
 551         return ret;
 552 }