fs/ubifs/tnc_commit.c

   1 /*
   2  * This file is part of UBIFS.
   3  *
   4  * Copyright (C) 2006-2008 Nokia Corporation.
   5  *
   6  * This program is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License version 2 as published by
   8  * the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  13  * more details.
  14  *
  15  * You should have received a copy of the GNU General Public License along with
  16  * this program; if not, write to the Free Software Foundation, Inc., 51
  17  * Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
  18  *
  19  * Authors: Adrian Hunter
  20  *          Artem Bityutskiy (Битюцкий Артём)
  21  */
  22
  23 /* This file implements TNC functions for committing */
  24
  25 #include <linux/random.h>
  26 #include "ubifs.h"
  27
  28 /**
  29  * make_idx_node - make an index node for fill-the-gaps method of TNC commit.
  30  * @c: UBIFS file-system description object
  31  * @idx: buffer in which to place new index node
  32  * @znode: znode from which to make new index node
  33  * @lnum: LEB number where new index node will be written
  34  * @offs: offset where new index node will be written
  35  * @len: length of new index node
  36  */
  37 static int make_idx_node(struct ubifs_info *c, struct ubifs_idx_node *idx,
  38                          struct ubifs_znode *znode, int lnum, int offs, int len)
  39 {
  40         struct ubifs_znode *zp;
  41         int i, err;
  42
  43         /* Make index node */
  44         idx->ch.node_type = UBIFS_IDX_NODE;
  45         idx->child_cnt = cpu_to_le16(znode->child_cnt);
  46         idx->level = cpu_to_le16(znode->level);
  47         for (i = 0; i < znode->child_cnt; i++) {
  48                 struct ubifs_branch *br = ubifs_idx_branch(c, idx, i);
  49                 struct ubifs_zbranch *zbr = &znode->zbranch[i];
  50
  51                 key_write_idx(c, &zbr->key, &br->key);
  52                 br->lnum = cpu_to_le32(zbr->lnum);
  53                 br->offs = cpu_to_le32(zbr->offs);
  54                 br->len = cpu_to_le32(zbr->len);
  55                 if (!zbr->lnum || !zbr->len) {
  56                         ubifs_err("bad ref in znode");
  57                         dbg_dump_znode(c, znode);
  58                         if (zbr->znode)
  59                                 dbg_dump_znode(c, zbr->znode);
  60                 }
  61         }
  62         ubifs_prepare_node(c, idx, len, 0);
  63
  64 #ifdef CONFIG_UBIFS_FS_DEBUG
  65         znode->lnum = lnum;
  66         znode->offs = offs;
  67         znode->len = len;
  68 #endif
  69
  70         err = insert_old_idx_znode(c, znode);
  71
  72         /* Update the parent */
  73         zp = znode->parent;
  74         if (zp) {
  75                 struct ubifs_zbranch *zbr;
  76
  77                 zbr = &zp->zbranch[znode->iip];
  78                 zbr->lnum = lnum;
  79                 zbr->offs = offs;
  80                 zbr->len = len;
  81         } else {
  82                 c->zroot.lnum = lnum;
  83                 c->zroot.offs = offs;
  84                 c->zroot.len = len;
  85         }
  86         c->calc_idx_sz += ALIGN(len, 8);
  87
  88         atomic_long_dec(&c->dirty_zn_cnt);
  89
  90         ubifs_assert(ubifs_zn_dirty(znode));
  91         ubifs_assert(ubifs_zn_cow(znode));
  92
  93         /*
  94          * Note, unlike 'write_index()' we do not add memory barriers here
  95          * because this function is called with @c->tnc_mutex locked.
  96          */
  97         __clear_bit(DIRTY_ZNODE, &znode->flags);
  98         __clear_bit(COW_ZNODE, &znode->flags);
  99
 100         return err;
 101 }
 102
 103 /**
 104  * fill_gap - make index nodes in gaps in dirty index LEBs.
 105  * @c: UBIFS file-system description object
 106  * @lnum: LEB number that gap appears in
 107  * @gap_start: offset of start of gap
 108  * @gap_end: offset of end of gap
 109  * @dirt: adds dirty space to this
 110  *
 111  * This function returns the number of index nodes written into the gap.
 112  */
 113 static int fill_gap(struct ubifs_info *c, int lnum, int gap_start, int gap_end,
 114                     int *dirt)
 115 {
 116         int len, gap_remains, gap_pos, written, pad_len;
 117
 118         ubifs_assert((gap_start & 7) == 0);
 119         ubifs_assert((gap_end & 7) == 0);
 120         ubifs_assert(gap_end >= gap_start);
 121
 122         gap_remains = gap_end - gap_start;
 123         if (!gap_remains)
 124                 return 0;
 125         gap_pos = gap_start;
 126         written = 0;
 127         while (c->enext) {
 128                 len = ubifs_idx_node_sz(c, c->enext->child_cnt);
 129                 if (len < gap_remains) {
 130                         struct ubifs_znode *znode = c->enext;
 131                         const int alen = ALIGN(len, 8);
 132                         int err;
 133
 134                         ubifs_assert(alen <= gap_remains);
 135                         err = make_idx_node(c, c->ileb_buf + gap_pos, znode,
 136                                             lnum, gap_pos, len);
 137                         if (err)
 138                                 return err;
 139                         gap_remains -= alen;
 140                         gap_pos += alen;
 141                         c->enext = znode->cnext;
 142                         if (c->enext == c->cnext)
 143                                 c->enext = NULL;
 144                         written += 1;
 145                 } else
 146                         break;
 147         }
 148         if (gap_end == c->leb_size) {
 149                 c->ileb_len = ALIGN(gap_pos, c->min_io_size);
 150                 /* Pad to end of min_io_size */
 151                 pad_len = c->ileb_len - gap_pos;
 152         } else
 153                 /* Pad to end of gap */
 154                 pad_len = gap_remains;
 155         dbg_gc("LEB %d:%d to %d len %d nodes written %d wasted bytes %d",
 156                lnum, gap_start, gap_end, gap_end - gap_start, written, pad_len);
 157         ubifs_pad(c, c->ileb_buf + gap_pos, pad_len);
 158         *dirt += pad_len;
 159         return written;
 160 }
 161
 162 /**
 163  * find_old_idx - find an index node obsoleted since the last commit start.
 164  * @c: UBIFS file-system description object
 165  * @lnum: LEB number of obsoleted index node
 166  * @offs: offset of obsoleted index node
 167  *
 168  * Returns %1 if found and %0 otherwise.
 169  */
 170 static int find_old_idx(struct ubifs_info *c, int lnum, int offs)
 171 {
 172         struct ubifs_old_idx *o;
 173         struct rb_node *p;
 174
 175         p = c->old_idx.rb_node;
 176         while (p) {
 177                 o = rb_entry(p, struct ubifs_old_idx, rb);
 178                 if (lnum < o->lnum)
 179                         p = p->rb_left;
 180                 else if (lnum > o->lnum)
 181                         p = p->rb_right;
 182                 else if (offs < o->offs)
 183                         p = p->rb_left;
 184                 else if (offs > o->offs)
 185                         p = p->rb_right;
 186                 else
 187                         return 1;
 188         }
 189         return 0;
 190 }
 191
 192 /**
 193  * is_idx_node_in_use - determine if an index node can be overwritten.
 194  * @c: UBIFS file-system description object
 195  * @key: key of index node
 196  * @level: index node level
 197  * @lnum: LEB number of index node
 198  * @offs: offset of index node
 199  *
 200  * If @key / @lnum / @offs identify an index node that was not part of the old
 201  * index, then this function returns %0 (obsolete).  Else if the index node was
 202  * part of the old index but is now dirty %1 is returned, else if it is clean %2
 203  * is returned. A negative error code is returned on failure.
 204  */
 205 static int is_idx_node_in_use(struct ubifs_info *c, union ubifs_key *key,
 206                               int level, int lnum, int offs)
 207 {
 208         int ret;
 209
 210         ret = is_idx_node_in_tnc(c, key, level, lnum, offs);
 211         if (ret < 0)
 212                 return ret; /* Error code */
 213         if (ret == 0)
 214                 if (find_old_idx(c, lnum, offs))
 215                         return 1;
 216         return ret;
 217 }
 218
 219 /**
 220  * layout_leb_in_gaps - layout index nodes using in-the-gaps method.
 221  * @c: UBIFS file-system description object
 222  * @p: return LEB number here
 223  *
 224  * This function lays out new index nodes for dirty znodes using in-the-gaps
 225  * method of TNC commit.
 226  * This function merely puts the next znode into the next gap, making no attempt
 227  * to try to maximise the number of znodes that fit.
 228  * This function returns the number of index nodes written into the gaps, or a
 229  * negative error code on failure.
 230  */
 231 static int layout_leb_in_gaps(struct ubifs_info *c, int *p)
 232 {
 233         struct ubifs_scan_leb *sleb;
 234         struct ubifs_scan_node *snod;
 235         int lnum, dirt = 0, gap_start, gap_end, err, written, tot_written;
 236
 237         tot_written = 0;
 238         /* Get an index LEB with lots of obsolete index nodes */
 239         lnum = ubifs_find_dirty_idx_leb(c);
 240         if (lnum < 0)
 241                 /*
 242                  * There also may be dirt in the index head that could be
 243                  * filled, however we do not check there at present.
 244                  */
 245                 return lnum; /* Error code */
 246         *p = lnum;
 247         dbg_gc("LEB %d", lnum);
 248         /*
 249          * Scan the index LEB.  We use the generic scan for this even though
 250          * it is more comprehensive and less efficient than is needed for this
 251          * purpose.
 252          */
 253         sleb = ubifs_scan(c, lnum, 0, c->ileb_buf, 0);
 254         c->ileb_len = 0;
 255         if (IS_ERR(sleb))
 256                 return PTR_ERR(sleb);
 257         gap_start = 0;
 258         list_for_each_entry(snod, &sleb->nodes, list) {
 259                 struct ubifs_idx_node *idx;
 260                 int in_use, level;
 261
 262                 ubifs_assert(snod->type == UBIFS_IDX_NODE);
 263                 idx = snod->node;
 264                 key_read(c, ubifs_idx_key(c, idx), &snod->key);
 265                 level = le16_to_cpu(idx->level);
 266                 /* Determine if the index node is in use (not obsolete) */
 267                 in_use = is_idx_node_in_use(c, &snod->key, level, lnum,
 268                                             snod->offs);
 269                 if (in_use < 0) {
 270                         ubifs_scan_destroy(sleb);
 271                         return in_use; /* Error code */
 272                 }
 273                 if (in_use) {
 274                         if (in_use == 1)
 275                                 dirt += ALIGN(snod->len, 8);
 276                         /*
 277                          * The obsolete index nodes form gaps that can be
 278                          * overwritten.  This gap has ended because we have
 279                          * found an index node that is still in use
 280                          * i.e. not obsolete
 281                          */
 282                         gap_end = snod->offs;
 283                         /* Try to fill gap */
 284                         written = fill_gap(c, lnum, gap_start, gap_end, &dirt);
 285                         if (written < 0) {
 286                                 ubifs_scan_destroy(sleb);
 287                                 return written; /* Error code */
 288                         }
 289                         tot_written += written;
 290                         gap_start = ALIGN(snod->offs + snod->len, 8);
 291                 }
 292         }
 293         ubifs_scan_destroy(sleb);
 294         c->ileb_len = c->leb_size;
 295         gap_end = c->leb_size;
 296         /* Try to fill gap */
 297         written = fill_gap(c, lnum, gap_start, gap_end, &dirt);
 298         if (written < 0)
 299                 return written; /* Error code */
 300         tot_written += written;
 301         if (tot_written == 0) {
 302                 struct ubifs_lprops lp;
 303
 304                 dbg_gc("LEB %d wrote %d index nodes", lnum, tot_written);
 305                 err = ubifs_read_one_lp(c, lnum, &lp);
 306                 if (err)
 307                         return err;
 308                 if (lp.free == c->leb_size) {
 309                         /*
 310                          * We must have snatched this LEB from the idx_gc list
 311                          * so we need to correct the free and dirty space.
 312                          */
 313                         err = ubifs_change_one_lp(c, lnum,
 314                                                   c->leb_size - c->ileb_len,
 315                                                   dirt, 0, 0, 0);
 316                         if (err)
 317                                 return err;
 318                 }
 319                 return 0;
 320         }
 321         err = ubifs_change_one_lp(c, lnum, c->leb_size - c->ileb_len, dirt,
 322                                   0, 0, 0);
 323         if (err)
 324                 return err;
 325         err = ubifs_leb_change(c, lnum, c->ileb_buf, c->ileb_len,
 326                                UBI_SHORTTERM);
 327         if (err)
 328                 return err;
 329         dbg_gc("LEB %d wrote %d index nodes", lnum, tot_written);
 330         return tot_written;
 331 }
 332
 333 /**
 334  * get_leb_cnt - calculate the number of empty LEBs needed to commit.
 335  * @c: UBIFS file-system description object
 336  * @cnt: number of znodes to commit
 337  *
 338  * This function returns the number of empty LEBs needed to commit @cnt znodes
 339  * to the current index head.  The number is not exact and may be more than
 340  * needed.
 341  */
 342 static int get_leb_cnt(struct ubifs_info *c, int cnt)
 343 {
 344         int d;
 345
 346         /* Assume maximum index node size (i.e. overestimate space needed) */
 347         cnt -= (c->leb_size - c->ihead_offs) / c->max_idx_node_sz;
 348         if (cnt < 0)
 349                 cnt = 0;
 350         d = c->leb_size / c->max_idx_node_sz;
 351         return DIV_ROUND_UP(cnt, d);
 352 }
 353
 354 /**
 355  * layout_in_gaps - in-the-gaps method of committing TNC.
 356  * @c: UBIFS file-system description object
 357  * @cnt: number of dirty znodes to commit.
 358  *
 359  * This function lays out new index nodes for dirty znodes using in-the-gaps
 360  * method of TNC commit.
 361  *
 362  * This function returns %0 on success and a negative error code on failure.
 363  */
 364 static int layout_in_gaps(struct ubifs_info *c, int cnt)
 365 {
 366         int err, leb_needed_cnt, written, *p;
 367
 368         dbg_gc("%d znodes to write", cnt);
 369
 370         c->gap_lebs = kmalloc(sizeof(int) * (c->lst.idx_lebs + 1), GFP_NOFS);
 371         if (!c->gap_lebs)
 372                 return -ENOMEM;
 373
 374         p = c->gap_lebs;
 375         do {
 376                 ubifs_assert(p < c->gap_lebs + sizeof(int) * c->lst.idx_lebs);
 377                 written = layout_leb_in_gaps(c, p);
 378                 if (written < 0) {
 379                         err = written;
 380                         if (err != -ENOSPC) {
 381                                 kfree(c->gap_lebs);
 382                                 c->gap_lebs = NULL;
 383                                 return err;
 384                         }
 385                         if (!dbg_is_chk_index(c)) {
 386                                 /*
 387                                  * Do not print scary warnings if the debugging
 388                                  * option which forces in-the-gaps is enabled.
 389                                  */
 390                                 ubifs_warn("out of space");
 391                                 dbg_dump_budg(c, &c->bi);
 392                                 dbg_dump_lprops(c);
 393                         }
 394                         /* Try to commit anyway */
 395                         err = 0;
 396                         break;
 397                 }
 398                 p++;
 399                 cnt -= written;
 400                 leb_needed_cnt = get_leb_cnt(c, cnt);
 401                 dbg_gc("%d znodes remaining, need %d LEBs, have %d", cnt,
 402                        leb_needed_cnt, c->ileb_cnt);
 403         } while (leb_needed_cnt > c->ileb_cnt);
 404
 405         *p = -1;
 406         return 0;
 407 }
 408
 409 /**
 410  * layout_in_empty_space - layout index nodes in empty space.
 411  * @c: UBIFS file-system description object
 412  *
 413  * This function lays out new index nodes for dirty znodes using empty LEBs.
 414  *
 415  * This function returns %0 on success and a negative error code on failure.
 416  */
 417 static int layout_in_empty_space(struct ubifs_info *c)
 418 {
 419         struct ubifs_znode *znode, *cnext, *zp;
 420         int lnum, offs, len, next_len, buf_len, buf_offs, used, avail;
 421         int wlen, blen, err;
 422
 423         cnext = c->enext;
 424         if (!cnext)
 425                 return 0;
 426
 427         lnum = c->ihead_lnum;
 428         buf_offs = c->ihead_offs;
 429
 430         buf_len = ubifs_idx_node_sz(c, c->fanout);
 431         buf_len = ALIGN(buf_len, c->min_io_size);
 432         used = 0;
 433         avail = buf_len;
 434
 435         /* Ensure there is enough room for first write */
 436         next_len = ubifs_idx_node_sz(c, cnext->child_cnt);
 437         if (buf_offs + next_len > c->leb_size)
 438                 lnum = -1;
 439
 440         while (1) {
 441                 znode = cnext;
 442
 443                 len = ubifs_idx_node_sz(c, znode->child_cnt);
 444
 445                 /* Determine the index node position */
 446                 if (lnum == -1) {
 447                         if (c->ileb_nxt >= c->ileb_cnt) {
 448                                 ubifs_err("out of space");
 449                                 return -ENOSPC;
 450                         }
 451                         lnum = c->ilebs[c->ileb_nxt++];
 452                         buf_offs = 0;
 453                         used = 0;
 454                         avail = buf_len;
 455                 }
 456
 457                 offs = buf_offs + used;
 458
 459 #ifdef CONFIG_UBIFS_FS_DEBUG
 460                 znode->lnum = lnum;
 461                 znode->offs = offs;
 462                 znode->len = len;
 463 #endif
 464
 465                 /* Update the parent */
 466                 zp = znode->parent;
 467                 if (zp) {
 468                         struct ubifs_zbranch *zbr;
 469                         int i;
 470
 471                         i = znode->iip;
 472                         zbr = &zp->zbranch[i];
 473                         zbr->lnum = lnum;
 474                         zbr->offs = offs;
 475                         zbr->len = len;
 476                 } else {
 477                         c->zroot.lnum = lnum;
 478                         c->zroot.offs = offs;
 479                         c->zroot.len = len;
 480                 }
 481                 c->calc_idx_sz += ALIGN(len, 8);
 482
 483                 /*
 484                  * Once lprops is updated, we can decrease the dirty znode count
 485                  * but it is easier to just do it here.
 486                  */
 487                 atomic_long_dec(&c->dirty_zn_cnt);
 488
 489                 /*
 490                  * Calculate the next index node length to see if there is
 491                  * enough room for it
 492                  */
 493                 cnext = znode->cnext;
 494                 if (cnext == c->cnext)
 495                         next_len = 0;
 496                 else
 497                         next_len = ubifs_idx_node_sz(c, cnext->child_cnt);
 498
 499                 /* Update buffer positions */
 500                 wlen = used + len;
 501                 used += ALIGN(len, 8);
 502                 avail -= ALIGN(len, 8);
 503
 504                 if (next_len != 0 &&
 505                     buf_offs + used + next_len <= c->leb_size &&
 506                     avail > 0)
 507                         continue;
 508
 509                 if (avail <= 0 && next_len &&
 510                     buf_offs + used + next_len <= c->leb_size)
 511                         blen = buf_len;
 512                 else
 513                         blen = ALIGN(wlen, c->min_io_size);
 514
 515                 /* The buffer is full or there are no more znodes to do */
 516                 buf_offs += blen;
 517                 if (next_len) {
 518                         if (buf_offs + next_len > c->leb_size) {
 519                                 err = ubifs_update_one_lp(c, lnum,
 520                                         c->leb_size - buf_offs, blen - used,
 521                                         0, 0);
 522                                 if (err)
 523                                         return err;
 524                                 lnum = -1;
 525                         }
 526                         used -= blen;
 527                         if (used < 0)
 528                                 used = 0;
 529                         avail = buf_len - used;
 530                         continue;
 531                 }
 532                 err = ubifs_update_one_lp(c, lnum, c->leb_size - buf_offs,
 533                                           blen - used, 0, 0);
 534                 if (err)
 535                         return err;
 536                 break;
 537         }
 538
 539 #ifdef CONFIG_UBIFS_FS_DEBUG
 540         c->dbg->new_ihead_lnum = lnum;
 541         c->dbg->new_ihead_offs = buf_offs;
 542 #endif
 543
 544         return 0;
 545 }
 546
 547 /**
 548  * layout_commit - determine positions of index nodes to commit.
 549  * @c: UBIFS file-system description object
 550  * @no_space: indicates that insufficient empty LEBs were allocated
 551  * @cnt: number of znodes to commit
 552  *
 553  * Calculate and update the positions of index nodes to commit.  If there were
 554  * an insufficient number of empty LEBs allocated, then index nodes are placed
 555  * into the gaps created by obsolete index nodes in non-empty index LEBs.  For
 556  * this purpose, an obsolete index node is one that was not in the index as at
 557  * the end of the last commit.  To write "in-the-gaps" requires that those index
 558  * LEBs are updated atomically in-place.
 559  */
 560 static int layout_commit(struct ubifs_info *c, int no_space, int cnt)
 561 {
 562         int err;
 563
 564         if (no_space) {
 565                 err = layout_in_gaps(c, cnt);
 566                 if (err)
 567                         return err;
 568         }
 569         err = layout_in_empty_space(c);
 570         return err;
 571 }
 572
 573 /**
 574  * find_first_dirty - find first dirty znode.
 575  * @znode: znode to begin searching from
 576  */
 577 static struct ubifs_znode *find_first_dirty(struct ubifs_znode *znode)
 578 {
 579         int i, cont;
 580
 581         if (!znode)
 582                 return NULL;
 583
 584         while (1) {
 585                 if (znode->level == 0) {
 586                         if (ubifs_zn_dirty(znode))
 587                                 return znode;
 588                         return NULL;
 589                 }
 590                 cont = 0;
 591                 for (i = 0; i < znode->child_cnt; i++) {
 592                         struct ubifs_zbranch *zbr = &znode->zbranch[i];
 593
 594                         if (zbr->znode && ubifs_zn_dirty(zbr->znode)) {
 595                                 znode = zbr->znode;
 596                                 cont = 1;
 597                                 break;
 598                         }
 599                 }
 600                 if (!cont) {
 601                         if (ubifs_zn_dirty(znode))
 602                                 return znode;
 603                         return NULL;
 604                 }
 605         }
 606 }
 607
 608 /**
 609  * find_next_dirty - find next dirty znode.
 610  * @znode: znode to begin searching from
 611  */
 612 static struct ubifs_znode *find_next_dirty(struct ubifs_znode *znode)
 613 {
 614         int n = znode->iip + 1;
 615
 616         znode = znode->parent;
 617         if (!znode)
 618                 return NULL;
 619         for (; n < znode->child_cnt; n++) {
 620                 struct ubifs_zbranch *zbr = &znode->zbranch[n];
 621
 622                 if (zbr->znode && ubifs_zn_dirty(zbr->znode))
 623                         return find_first_dirty(zbr->znode);
 624         }
 625         return znode;
 626 }
 627
 628 /**
 629  * get_znodes_to_commit - create list of dirty znodes to commit.
 630  * @c: UBIFS file-system description object
 631  *
 632  * This function returns the number of znodes to commit.
 633  */
 634 static int get_znodes_to_commit(struct ubifs_info *c)
 635 {
 636         struct ubifs_znode *znode, *cnext;
 637         int cnt = 0;
 638
 639         c->cnext = find_first_dirty(c->zroot.znode);
 640         znode = c->enext = c->cnext;
 641         if (!znode) {
 642                 dbg_cmt("no znodes to commit");
 643                 return 0;
 644         }
 645         cnt += 1;
 646         while (1) {
 647                 ubifs_assert(!ubifs_zn_cow(znode));
 648                 __set_bit(COW_ZNODE, &znode->flags);
 649                 znode->alt = 0;
 650                 cnext = find_next_dirty(znode);
 651                 if (!cnext) {
 652                         znode->cnext = c->cnext;
 653                         break;
 654                 }
 655                 znode->cnext = cnext;
 656                 znode = cnext;
 657                 cnt += 1;
 658         }
 659         dbg_cmt("committing %d znodes", cnt);
 660         ubifs_assert(cnt == atomic_long_read(&c->dirty_zn_cnt));
 661         return cnt;
 662 }
 663
 664 /**
 665  * alloc_idx_lebs - allocate empty LEBs to be used to commit.
 666  * @c: UBIFS file-system description object
 667  * @cnt: number of znodes to commit
 668  *
 669  * This function returns %-ENOSPC if it cannot allocate a sufficient number of
 670  * empty LEBs.  %0 is returned on success, otherwise a negative error code
 671  * is returned.
 672  */
 673 static int alloc_idx_lebs(struct ubifs_info *c, int cnt)
 674 {
 675         int i, leb_cnt, lnum;
 676
 677         c->ileb_cnt = 0;
 678         c->ileb_nxt = 0;
 679         leb_cnt = get_leb_cnt(c, cnt);
 680         dbg_cmt("need about %d empty LEBS for TNC commit", leb_cnt);
 681         if (!leb_cnt)
 682                 return 0;
 683         c->ilebs = kmalloc(leb_cnt * sizeof(int), GFP_NOFS);
 684         if (!c->ilebs)
 685                 return -ENOMEM;
 686         for (i = 0; i < leb_cnt; i++) {
 687                 lnum = ubifs_find_free_leb_for_idx(c);
 688                 if (lnum < 0)
 689                         return lnum;
 690                 c->ilebs[c->ileb_cnt++] = lnum;
 691                 dbg_cmt("LEB %d", lnum);
 692         }
 693         if (dbg_is_chk_index(c) && !(random32() & 7))
 694                 return -ENOSPC;
 695         return 0;
 696 }
 697
 698 /**
 699  * free_unused_idx_lebs - free unused LEBs that were allocated for the commit.
 700  * @c: UBIFS file-system description object
 701  *
 702  * It is possible that we allocate more empty LEBs for the commit than we need.
 703  * This functions frees the surplus.
 704  *
 705  * This function returns %0 on success and a negative error code on failure.
 706  */
 707 static int free_unused_idx_lebs(struct ubifs_info *c)
 708 {
 709         int i, err = 0, lnum, er;
 710
 711         for (i = c->ileb_nxt; i < c->ileb_cnt; i++) {
 712                 lnum = c->ilebs[i];
 713                 dbg_cmt("LEB %d", lnum);
 714                 er = ubifs_change_one_lp(c, lnum, LPROPS_NC, LPROPS_NC, 0,
 715                                          LPROPS_INDEX | LPROPS_TAKEN, 0);
 716                 if (!err)
 717                         err = er;
 718         }
 719         return err;
 720 }
 721
 722 /**
 723  * free_idx_lebs - free unused LEBs after commit end.
 724  * @c: UBIFS file-system description object
 725  *
 726  * This function returns %0 on success and a negative error code on failure.
 727  */
 728 static int free_idx_lebs(struct ubifs_info *c)
 729 {
 730         int err;
 731
 732         err = free_unused_idx_lebs(c);
 733         kfree(c->ilebs);
 734         c->ilebs = NULL;
 735         return err;
 736 }
 737
 738 /**
 739  * ubifs_tnc_start_commit - start TNC commit.
 740  * @c: UBIFS file-system description object
 741  * @zroot: new index root position is returned here
 742  *
 743  * This function prepares the list of indexing nodes to commit and lays out
 744  * their positions on flash. If there is not enough free space it uses the
 745  * in-gap commit method. Returns zero in case of success and a negative error
 746  * code in case of failure.
 747  */
 748 int ubifs_tnc_start_commit(struct ubifs_info *c, struct ubifs_zbranch *zroot)
 749 {
 750         int err = 0, cnt;
 751
 752         mutex_lock(&c->tnc_mutex);
 753         err = dbg_check_tnc(c, 1);
 754         if (err)
 755                 goto out;
 756         cnt = get_znodes_to_commit(c);
 757         if (cnt != 0) {
 758                 int no_space = 0;
 759
 760                 err = alloc_idx_lebs(c, cnt);
 761                 if (err == -ENOSPC)
 762                         no_space = 1;
 763                 else if (err)
 764                         goto out_free;
 765                 err = layout_commit(c, no_space, cnt);
 766                 if (err)
 767                         goto out_free;
 768                 ubifs_assert(atomic_long_read(&c->dirty_zn_cnt) == 0);
 769                 err = free_unused_idx_lebs(c);
 770                 if (err)
 771                         goto out;
 772         }
 773         destroy_old_idx(c);
 774         memcpy(zroot, &c->zroot, sizeof(struct ubifs_zbranch));
 775
 776         err = ubifs_save_dirty_idx_lnums(c);
 777         if (err)
 778                 goto out;
 779
 780         spin_lock(&c->space_lock);
 781         /*
 782          * Although we have not finished committing yet, update size of the
 783          * committed index ('c->bi.old_idx_sz') and zero out the index growth
 784          * budget. It is OK to do this now, because we've reserved all the
 785          * space which is needed to commit the index, and it is save for the
 786          * budgeting subsystem to assume the index is already committed,
 787          * even though it is not.
 788          */
 789         ubifs_assert(c->bi.min_idx_lebs == ubifs_calc_min_idx_lebs(c));
 790         c->bi.old_idx_sz = c->calc_idx_sz;
 791         c->bi.uncommitted_idx = 0;
 792         c->bi.min_idx_lebs = ubifs_calc_min_idx_lebs(c);
 793         spin_unlock(&c->space_lock);
 794         mutex_unlock(&c->tnc_mutex);
 795
 796         dbg_cmt("number of index LEBs %d", c->lst.idx_lebs);
 797         dbg_cmt("size of index %llu", c->calc_idx_sz);
 798         return err;
 799
 800 out_free:
 801         free_idx_lebs(c);
 802 out:
 803         mutex_unlock(&c->tnc_mutex);
 804         return err;
 805 }
 806
 807 /**
 808  * write_index - write index nodes.
 809  * @c: UBIFS file-system description object
 810  *
 811  * This function writes the index nodes whose positions were laid out in the
 812  * layout_in_empty_space function.
 813  */
 814 static int write_index(struct ubifs_info *c)
 815 {
 816         struct ubifs_idx_node *idx;
 817         struct ubifs_znode *znode, *cnext;
 818         int i, lnum, offs, len, next_len, buf_len, buf_offs, used;
 819         int avail, wlen, err, lnum_pos = 0, blen, nxt_offs;
 820
 821         cnext = c->enext;
 822         if (!cnext)
 823                 return 0;
 824
 825         /*
 826          * Always write index nodes to the index head so that index nodes and
 827          * other types of nodes are never mixed in the same erase block.
 828          */
 829         lnum = c->ihead_lnum;
 830         buf_offs = c->ihead_offs;
 831
 832         /* Allocate commit buffer */
 833         buf_len = ALIGN(c->max_idx_node_sz, c->min_io_size);
 834         used = 0;
 835         avail = buf_len;
 836
 837         /* Ensure there is enough room for first write */
 838         next_len = ubifs_idx_node_sz(c, cnext->child_cnt);
 839         if (buf_offs + next_len > c->leb_size) {
 840                 err = ubifs_update_one_lp(c, lnum, LPROPS_NC, 0, 0,
 841                                           LPROPS_TAKEN);
 842                 if (err)
 843                         return err;
 844                 lnum = -1;
 845         }
 846
 847         while (1) {
 848                 cond_resched();
 849
 850                 znode = cnext;
 851                 idx = c->cbuf + used;
 852
 853                 /* Make index node */
 854                 idx->ch.node_type = UBIFS_IDX_NODE;
 855                 idx->child_cnt = cpu_to_le16(znode->child_cnt);
 856                 idx->level = cpu_to_le16(znode->level);
 857                 for (i = 0; i < znode->child_cnt; i++) {
 858                         struct ubifs_branch *br = ubifs_idx_branch(c, idx, i);
 859                         struct ubifs_zbranch *zbr = &znode->zbranch[i];
 860
 861                         key_write_idx(c, &zbr->key, &br->key);
 862                         br->lnum = cpu_to_le32(zbr->lnum);
 863                         br->offs = cpu_to_le32(zbr->offs);
 864                         br->len = cpu_to_le32(zbr->len);
 865                         if (!zbr->lnum || !zbr->len) {
 866                                 ubifs_err("bad ref in znode");
 867                                 dbg_dump_znode(c, znode);
 868                                 if (zbr->znode)
 869                                         dbg_dump_znode(c, zbr->znode);
 870                         }
 871                 }
 872                 len = ubifs_idx_node_sz(c, znode->child_cnt);
 873                 ubifs_prepare_node(c, idx, len, 0);
 874
 875                 /* Determine the index node position */
 876                 if (lnum == -1) {
 877                         lnum = c->ilebs[lnum_pos++];
 878                         buf_offs = 0;
 879                         used = 0;
 880                         avail = buf_len;
 881                 }
 882                 offs = buf_offs + used;
 883
 884 #ifdef CONFIG_UBIFS_FS_DEBUG
 885                 if (lnum != znode->lnum || offs != znode->offs ||
 886                     len != znode->len) {
 887                         ubifs_err("inconsistent znode posn");
 888                         return -EINVAL;
 889                 }
 890 #endif
 891
 892                 /* Grab some stuff from znode while we still can */
 893                 cnext = znode->cnext;
 894
 895                 ubifs_assert(ubifs_zn_dirty(znode));
 896                 ubifs_assert(ubifs_zn_cow(znode));
 897
 898                 /*
 899                  * It is important that other threads should see %DIRTY_ZNODE
 900                  * flag cleared before %COW_ZNODE. Specifically, it matters in
 901                  * the 'dirty_cow_znode()' function. This is the reason for the
 902                  * first barrier. Also, we want the bit changes to be seen to
 903                  * other threads ASAP, to avoid unnecesarry copying, which is
 904                  * the reason for the second barrier.
 905                  */
 906                 clear_bit(DIRTY_ZNODE, &znode->flags);
 907                 smp_mb__before_clear_bit();
 908                 clear_bit(COW_ZNODE, &znode->flags);
 909                 smp_mb__after_clear_bit();
 910
 911                 /*
 912                  * We have marked the znode as clean but have not updated the
 913                  * @c->clean_zn_cnt counter. If this znode becomes dirty again
 914                  * before 'free_obsolete_znodes()' is called, then
 915                  * @c->clean_zn_cnt will be decremented before it gets
 916                  * incremented (resulting in 2 decrements for the same znode).
 917                  * This means that @c->clean_zn_cnt may become negative for a
 918                  * while.
 919                  *
 920                  * Q: why we cannot increment @c->clean_zn_cnt?
 921                  * A: because we do not have the @c->tnc_mutex locked, and the
 922                  *    following code would be racy and buggy:
 923                  *
 924                  *    if (!ubifs_zn_obsolete(znode)) {
 925                  *            atomic_long_inc(&c->clean_zn_cnt);
 926                  *            atomic_long_inc(&ubifs_clean_zn_cnt);
 927                  *    }
 928                  *
 929                  *    Thus, we just delay the @c->clean_zn_cnt update until we
 930                  *    have the mutex locked.
 931                  */
 932
 933                 /* Do not access znode from this point on */
 934
 935                 /* Update buffer positions */
 936                 wlen = used + len;
 937                 used += ALIGN(len, 8);
 938                 avail -= ALIGN(len, 8);
 939
 940                 /*
 941                  * Calculate the next index node length to see if there is
 942                  * enough room for it
 943                  */
 944                 if (cnext == c->cnext)
 945                         next_len = 0;
 946                 else
 947                         next_len = ubifs_idx_node_sz(c, cnext->child_cnt);
 948
 949                 nxt_offs = buf_offs + used + next_len;
 950                 if (next_len && nxt_offs <= c->leb_size) {
 951                         if (avail > 0)
 952                                 continue;
 953                         else
 954                                 blen = buf_len;
 955                 } else {
 956                         wlen = ALIGN(wlen, 8);
 957                         blen = ALIGN(wlen, c->min_io_size);
 958                         ubifs_pad(c, c->cbuf + wlen, blen - wlen);
 959                 }
 960
 961                 /* The buffer is full or there are no more znodes to do */
 962                 err = ubifs_leb_write(c, lnum, c->cbuf, buf_offs, blen,
 963                                       UBI_SHORTTERM);
 964                 if (err)
 965                         return err;
 966                 buf_offs += blen;
 967                 if (next_len) {
 968                         if (nxt_offs > c->leb_size) {
 969                                 err = ubifs_update_one_lp(c, lnum, LPROPS_NC, 0,
 970                                                           0, LPROPS_TAKEN);
 971                                 if (err)
 972                                         return err;
 973                                 lnum = -1;
 974                         }
 975                         used -= blen;
 976                         if (used < 0)
 977                                 used = 0;
 978                         avail = buf_len - used;
 979                         memmove(c->cbuf, c->cbuf + blen, used);
 980                         continue;
 981                 }
 982                 break;
 983         }
 984
 985 #ifdef CONFIG_UBIFS_FS_DEBUG
 986         if (lnum != c->dbg->new_ihead_lnum ||
 987             buf_offs != c->dbg->new_ihead_offs) {
 988                 ubifs_err("inconsistent ihead");
 989                 return -EINVAL;
 990         }
 991 #endif
 992
 993         c->ihead_lnum = lnum;
 994         c->ihead_offs = buf_offs;
 995
 996         return 0;
 997 }
 998
 999 /**
1000  * free_obsolete_znodes - free obsolete znodes.
1001  * @c: UBIFS file-system description object
1002  *
1003  * At the end of commit end, obsolete znodes are freed.
1004  */
1005 static void free_obsolete_znodes(struct ubifs_info *c)
1006 {
1007         struct ubifs_znode *znode, *cnext;
1008
1009         cnext = c->cnext;
1010         do {
1011                 znode = cnext;
1012                 cnext = znode->cnext;
1013                 if (ubifs_zn_obsolete(znode))
1014                         kfree(znode);
1015                 else {
1016                         znode->cnext = NULL;
1017                         atomic_long_inc(&c->clean_zn_cnt);
1018                         atomic_long_inc(&ubifs_clean_zn_cnt);
1019                 }
1020         } while (cnext != c->cnext);
1021 }
1022
1023 /**
1024  * return_gap_lebs - return LEBs used by the in-gap commit method.
1025  * @c: UBIFS file-system description object
1026  *
1027  * This function clears the "taken" flag for the LEBs which were used by the
1028  * "commit in-the-gaps" method.
1029  */
1030 static int return_gap_lebs(struct ubifs_info *c)
1031 {
1032         int *p, err;
1033
1034         if (!c->gap_lebs)
1035                 return 0;
1036
1037         dbg_cmt("");
1038         for (p = c->gap_lebs; *p != -1; p++) {
1039                 err = ubifs_change_one_lp(c, *p, LPROPS_NC, LPROPS_NC, 0,
1040                                           LPROPS_TAKEN, 0);
1041                 if (err)
1042                         return err;
1043         }
1044
1045         kfree(c->gap_lebs);
1046         c->gap_lebs = NULL;
1047         return 0;
1048 }
1049
1050 /**
1051  * ubifs_tnc_end_commit - update the TNC for commit end.
1052  * @c: UBIFS file-system description object
1053  *
1054  * Write the dirty znodes.
1055  */
1056 int ubifs_tnc_end_commit(struct ubifs_info *c)
1057 {
1058         int err;
1059
1060         if (!c->cnext)
1061                 return 0;
1062
1063         err = return_gap_lebs(c);
1064         if (err)
1065                 return err;
1066
1067         err = write_index(c);
1068         if (err)
1069                 return err;
1070
1071         mutex_lock(&c->tnc_mutex);
1072
1073         dbg_cmt("TNC height is %d", c->zroot.znode->level + 1);
1074
1075         free_obsolete_znodes(c);
1076
1077         c->cnext = NULL;
1078         kfree(c->ilebs);
1079         c->ilebs = NULL;
1080
1081         mutex_unlock(&c->tnc_mutex);
1082
1083         return 0;
1084 }