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