fs/ubifs/tnc_commit.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  * This file is part of UBIFS.
0004  *
0005  * Copyright (C) 2006-2008 Nokia Corporation.
0006  *
0007  * Authors: Adrian Hunter
0008  *          Artem Bityutskiy (Битюцкий Артём)
0009  */
0010
0011 /* This file implements TNC functions for committing */
0012
0013 #include <linux/random.h>
0014 #include "ubifs.h"
0015
0016 /**
0017  * make_idx_node - make an index node for fill-the-gaps method of TNC commit.
0018  * @c: UBIFS file-system description object
0019  * @idx: buffer in which to place new index node
0020  * @znode: znode from which to make new index node
0021  * @lnum: LEB number where new index node will be written
0022  * @offs: offset where new index node will be written
0023  * @len: length of new index node
0024  */
0025 static int make_idx_node(struct ubifs_info *c, struct ubifs_idx_node *idx,
0026              struct ubifs_znode *znode, int lnum, int offs, int len)
0027 {
0028     struct ubifs_znode *zp;
0029     u8 hash[UBIFS_HASH_ARR_SZ];
0030     int i, err;
0031
0032     /* Make index node */
0033     idx->ch.node_type = UBIFS_IDX_NODE;
0034     idx->child_cnt = cpu_to_le16(znode->child_cnt);
0035     idx->level = cpu_to_le16(znode->level);
0036     for (i = 0; i < znode->child_cnt; i++) {
0037         struct ubifs_branch *br = ubifs_idx_branch(c, idx, i);
0038         struct ubifs_zbranch *zbr = &znode->zbranch[i];
0039
0040         key_write_idx(c, &zbr->key, &br->key);
0041         br->lnum = cpu_to_le32(zbr->lnum);
0042         br->offs = cpu_to_le32(zbr->offs);
0043         br->len = cpu_to_le32(zbr->len);
0044         ubifs_copy_hash(c, zbr->hash, ubifs_branch_hash(c, br));
0045         if (!zbr->lnum || !zbr->len) {
0046             ubifs_err(c, "bad ref in znode");
0047             ubifs_dump_znode(c, znode);
0048             if (zbr->znode)
0049                 ubifs_dump_znode(c, zbr->znode);
0050
0051             return -EINVAL;
0052         }
0053     }
0054     ubifs_prepare_node(c, idx, len, 0);
0055     ubifs_node_calc_hash(c, idx, hash);
0056
0057     znode->lnum = lnum;
0058     znode->offs = offs;
0059     znode->len = len;
0060
0061     err = insert_old_idx_znode(c, znode);
0062
0063     /* Update the parent */
0064     zp = znode->parent;
0065     if (zp) {
0066         struct ubifs_zbranch *zbr;
0067
0068         zbr = &zp->zbranch[znode->iip];
0069         zbr->lnum = lnum;
0070         zbr->offs = offs;
0071         zbr->len = len;
0072         ubifs_copy_hash(c, hash, zbr->hash);
0073     } else {
0074         c->zroot.lnum = lnum;
0075         c->zroot.offs = offs;
0076         c->zroot.len = len;
0077         ubifs_copy_hash(c, hash, c->zroot.hash);
0078     }
0079     c->calc_idx_sz += ALIGN(len, 8);
0080
0081     atomic_long_dec(&c->dirty_zn_cnt);
0082
0083     ubifs_assert(c, ubifs_zn_dirty(znode));
0084     ubifs_assert(c, ubifs_zn_cow(znode));
0085
0086     /*
0087      * Note, unlike 'write_index()' we do not add memory barriers here
0088      * because this function is called with @c->tnc_mutex locked.
0089      */
0090     __clear_bit(DIRTY_ZNODE, &znode->flags);
0091     __clear_bit(COW_ZNODE, &znode->flags);
0092
0093     return err;
0094 }
0095
0096 /**
0097  * fill_gap - make index nodes in gaps in dirty index LEBs.
0098  * @c: UBIFS file-system description object
0099  * @lnum: LEB number that gap appears in
0100  * @gap_start: offset of start of gap
0101  * @gap_end: offset of end of gap
0102  * @dirt: adds dirty space to this
0103  *
0104  * This function returns the number of index nodes written into the gap.
0105  */
0106 static int fill_gap(struct ubifs_info *c, int lnum, int gap_start, int gap_end,
0107             int *dirt)
0108 {
0109     int len, gap_remains, gap_pos, written, pad_len;
0110
0111     ubifs_assert(c, (gap_start & 7) == 0);
0112     ubifs_assert(c, (gap_end & 7) == 0);
0113     ubifs_assert(c, gap_end >= gap_start);
0114
0115     gap_remains = gap_end - gap_start;
0116     if (!gap_remains)
0117         return 0;
0118     gap_pos = gap_start;
0119     written = 0;
0120     while (c->enext) {
0121         len = ubifs_idx_node_sz(c, c->enext->child_cnt);
0122         if (len < gap_remains) {
0123             struct ubifs_znode *znode = c->enext;
0124             const int alen = ALIGN(len, 8);
0125             int err;
0126
0127             ubifs_assert(c, alen <= gap_remains);
0128             err = make_idx_node(c, c->ileb_buf + gap_pos, znode,
0129                         lnum, gap_pos, len);
0130             if (err)
0131                 return err;
0132             gap_remains -= alen;
0133             gap_pos += alen;
0134             c->enext = znode->cnext;
0135             if (c->enext == c->cnext)
0136                 c->enext = NULL;
0137             written += 1;
0138         } else
0139             break;
0140     }
0141     if (gap_end == c->leb_size) {
0142         c->ileb_len = ALIGN(gap_pos, c->min_io_size);
0143         /* Pad to end of min_io_size */
0144         pad_len = c->ileb_len - gap_pos;
0145     } else
0146         /* Pad to end of gap */
0147         pad_len = gap_remains;
0148     dbg_gc("LEB %d:%d to %d len %d nodes written %d wasted bytes %d",
0149            lnum, gap_start, gap_end, gap_end - gap_start, written, pad_len);
0150     ubifs_pad(c, c->ileb_buf + gap_pos, pad_len);
0151     *dirt += pad_len;
0152     return written;
0153 }
0154
0155 /**
0156  * find_old_idx - find an index node obsoleted since the last commit start.
0157  * @c: UBIFS file-system description object
0158  * @lnum: LEB number of obsoleted index node
0159  * @offs: offset of obsoleted index node
0160  *
0161  * Returns %1 if found and %0 otherwise.
0162  */
0163 static int find_old_idx(struct ubifs_info *c, int lnum, int offs)
0164 {
0165     struct ubifs_old_idx *o;
0166     struct rb_node *p;
0167
0168     p = c->old_idx.rb_node;
0169     while (p) {
0170         o = rb_entry(p, struct ubifs_old_idx, rb);
0171         if (lnum < o->lnum)
0172             p = p->rb_left;
0173         else if (lnum > o->lnum)
0174             p = p->rb_right;
0175         else if (offs < o->offs)
0176             p = p->rb_left;
0177         else if (offs > o->offs)
0178             p = p->rb_right;
0179         else
0180             return 1;
0181     }
0182     return 0;
0183 }
0184
0185 /**
0186  * is_idx_node_in_use - determine if an index node can be overwritten.
0187  * @c: UBIFS file-system description object
0188  * @key: key of index node
0189  * @level: index node level
0190  * @lnum: LEB number of index node
0191  * @offs: offset of index node
0192  *
0193  * If @key / @lnum / @offs identify an index node that was not part of the old
0194  * index, then this function returns %0 (obsolete).  Else if the index node was
0195  * part of the old index but is now dirty %1 is returned, else if it is clean %2
0196  * is returned. A negative error code is returned on failure.
0197  */
0198 static int is_idx_node_in_use(struct ubifs_info *c, union ubifs_key *key,
0199                   int level, int lnum, int offs)
0200 {
0201     int ret;
0202
0203     ret = is_idx_node_in_tnc(c, key, level, lnum, offs);
0204     if (ret < 0)
0205         return ret; /* Error code */
0206     if (ret == 0)
0207         if (find_old_idx(c, lnum, offs))
0208             return 1;
0209     return ret;
0210 }
0211
0212 /**
0213  * layout_leb_in_gaps - layout index nodes using in-the-gaps method.
0214  * @c: UBIFS file-system description object
0215  * @p: return LEB number in @c->gap_lebs[p]
0216  *
0217  * This function lays out new index nodes for dirty znodes using in-the-gaps
0218  * method of TNC commit.
0219  * This function merely puts the next znode into the next gap, making no attempt
0220  * to try to maximise the number of znodes that fit.
0221  * This function returns the number of index nodes written into the gaps, or a
0222  * negative error code on failure.
0223  */
0224 static int layout_leb_in_gaps(struct ubifs_info *c, int p)
0225 {
0226     struct ubifs_scan_leb *sleb;
0227     struct ubifs_scan_node *snod;
0228     int lnum, dirt = 0, gap_start, gap_end, err, written, tot_written;
0229
0230     tot_written = 0;
0231     /* Get an index LEB with lots of obsolete index nodes */
0232     lnum = ubifs_find_dirty_idx_leb(c);
0233     if (lnum < 0)
0234         /*
0235          * There also may be dirt in the index head that could be
0236          * filled, however we do not check there at present.
0237          */
0238         return lnum; /* Error code */
0239     c->gap_lebs[p] = lnum;
0240     dbg_gc("LEB %d", lnum);
0241     /*
0242      * Scan the index LEB.  We use the generic scan for this even though
0243      * it is more comprehensive and less efficient than is needed for this
0244      * purpose.
0245      */
0246     sleb = ubifs_scan(c, lnum, 0, c->ileb_buf, 0);
0247     c->ileb_len = 0;
0248     if (IS_ERR(sleb))
0249         return PTR_ERR(sleb);
0250     gap_start = 0;
0251     list_for_each_entry(snod, &sleb->nodes, list) {
0252         struct ubifs_idx_node *idx;
0253         int in_use, level;
0254
0255         ubifs_assert(c, snod->type == UBIFS_IDX_NODE);
0256         idx = snod->node;
0257         key_read(c, ubifs_idx_key(c, idx), &snod->key);
0258         level = le16_to_cpu(idx->level);
0259         /* Determine if the index node is in use (not obsolete) */
0260         in_use = is_idx_node_in_use(c, &snod->key, level, lnum,
0261                         snod->offs);
0262         if (in_use < 0) {
0263             ubifs_scan_destroy(sleb);
0264             return in_use; /* Error code */
0265         }
0266         if (in_use) {
0267             if (in_use == 1)
0268                 dirt += ALIGN(snod->len, 8);
0269             /*
0270              * The obsolete index nodes form gaps that can be
0271              * overwritten.  This gap has ended because we have
0272              * found an index node that is still in use
0273              * i.e. not obsolete
0274              */
0275             gap_end = snod->offs;
0276             /* Try to fill gap */
0277             written = fill_gap(c, lnum, gap_start, gap_end, &dirt);
0278             if (written < 0) {
0279                 ubifs_scan_destroy(sleb);
0280                 return written; /* Error code */
0281             }
0282             tot_written += written;
0283             gap_start = ALIGN(snod->offs + snod->len, 8);
0284         }
0285     }
0286     ubifs_scan_destroy(sleb);
0287     c->ileb_len = c->leb_size;
0288     gap_end = c->leb_size;
0289     /* Try to fill gap */
0290     written = fill_gap(c, lnum, gap_start, gap_end, &dirt);
0291     if (written < 0)
0292         return written; /* Error code */
0293     tot_written += written;
0294     if (tot_written == 0) {
0295         struct ubifs_lprops lp;
0296
0297         dbg_gc("LEB %d wrote %d index nodes", lnum, tot_written);
0298         err = ubifs_read_one_lp(c, lnum, &lp);
0299         if (err)
0300             return err;
0301         if (lp.free == c->leb_size) {
0302             /*
0303              * We must have snatched this LEB from the idx_gc list
0304              * so we need to correct the free and dirty space.
0305              */
0306             err = ubifs_change_one_lp(c, lnum,
0307                           c->leb_size - c->ileb_len,
0308                           dirt, 0, 0, 0);
0309             if (err)
0310                 return err;
0311         }
0312         return 0;
0313     }
0314     err = ubifs_change_one_lp(c, lnum, c->leb_size - c->ileb_len, dirt,
0315                   0, 0, 0);
0316     if (err)
0317         return err;
0318     err = ubifs_leb_change(c, lnum, c->ileb_buf, c->ileb_len);
0319     if (err)
0320         return err;
0321     dbg_gc("LEB %d wrote %d index nodes", lnum, tot_written);
0322     return tot_written;
0323 }
0324
0325 /**
0326  * get_leb_cnt - calculate the number of empty LEBs needed to commit.
0327  * @c: UBIFS file-system description object
0328  * @cnt: number of znodes to commit
0329  *
0330  * This function returns the number of empty LEBs needed to commit @cnt znodes
0331  * to the current index head.  The number is not exact and may be more than
0332  * needed.
0333  */
0334 static int get_leb_cnt(struct ubifs_info *c, int cnt)
0335 {
0336     int d;
0337
0338     /* Assume maximum index node size (i.e. overestimate space needed) */
0339     cnt -= (c->leb_size - c->ihead_offs) / c->max_idx_node_sz;
0340     if (cnt < 0)
0341         cnt = 0;
0342     d = c->leb_size / c->max_idx_node_sz;
0343     return DIV_ROUND_UP(cnt, d);
0344 }
0345
0346 /**
0347  * layout_in_gaps - in-the-gaps method of committing TNC.
0348  * @c: UBIFS file-system description object
0349  * @cnt: number of dirty znodes to commit.
0350  *
0351  * This function lays out new index nodes for dirty znodes using in-the-gaps
0352  * method of TNC commit.
0353  *
0354  * This function returns %0 on success and a negative error code on failure.
0355  */
0356 static int layout_in_gaps(struct ubifs_info *c, int cnt)
0357 {
0358     int err, leb_needed_cnt, written, p = 0, old_idx_lebs, *gap_lebs;
0359
0360     dbg_gc("%d znodes to write", cnt);
0361
0362     c->gap_lebs = kmalloc_array(c->lst.idx_lebs + 1, sizeof(int),
0363                     GFP_NOFS);
0364     if (!c->gap_lebs)
0365         return -ENOMEM;
0366
0367     old_idx_lebs = c->lst.idx_lebs;
0368     do {
0369         ubifs_assert(c, p < c->lst.idx_lebs);
0370         written = layout_leb_in_gaps(c, p);
0371         if (written < 0) {
0372             err = written;
0373             if (err != -ENOSPC) {
0374                 kfree(c->gap_lebs);
0375                 c->gap_lebs = NULL;
0376                 return err;
0377             }
0378             if (!dbg_is_chk_index(c)) {
0379                 /*
0380                  * Do not print scary warnings if the debugging
0381                  * option which forces in-the-gaps is enabled.
0382                  */
0383                 ubifs_warn(c, "out of space");
0384                 ubifs_dump_budg(c, &c->bi);
0385                 ubifs_dump_lprops(c);
0386             }
0387             /* Try to commit anyway */
0388             break;
0389         }
0390         p++;
0391         cnt -= written;
0392         leb_needed_cnt = get_leb_cnt(c, cnt);
0393         dbg_gc("%d znodes remaining, need %d LEBs, have %d", cnt,
0394                leb_needed_cnt, c->ileb_cnt);
0395         /*
0396          * Dynamically change the size of @c->gap_lebs to prevent
0397          * oob, because @c->lst.idx_lebs could be increased by
0398          * function @get_idx_gc_leb (called by layout_leb_in_gaps->
0399          * ubifs_find_dirty_idx_leb) during loop. Only enlarge
0400          * @c->gap_lebs when needed.
0401          *
0402          */
0403         if (leb_needed_cnt > c->ileb_cnt && p >= old_idx_lebs &&
0404             old_idx_lebs < c->lst.idx_lebs) {
0405             old_idx_lebs = c->lst.idx_lebs;
0406             gap_lebs = krealloc(c->gap_lebs, sizeof(int) *
0407                            (old_idx_lebs + 1), GFP_NOFS);
0408             if (!gap_lebs) {
0409                 kfree(c->gap_lebs);
0410                 c->gap_lebs = NULL;
0411                 return -ENOMEM;
0412             }
0413             c->gap_lebs = gap_lebs;
0414         }
0415     } while (leb_needed_cnt > c->ileb_cnt);
0416
0417     c->gap_lebs[p] = -1;
0418     return 0;
0419 }
0420
0421 /**
0422  * layout_in_empty_space - layout index nodes in empty space.
0423  * @c: UBIFS file-system description object
0424  *
0425  * This function lays out new index nodes for dirty znodes using empty LEBs.
0426  *
0427  * This function returns %0 on success and a negative error code on failure.
0428  */
0429 static int layout_in_empty_space(struct ubifs_info *c)
0430 {
0431     struct ubifs_znode *znode, *cnext, *zp;
0432     int lnum, offs, len, next_len, buf_len, buf_offs, used, avail;
0433     int wlen, blen, err;
0434
0435     cnext = c->enext;
0436     if (!cnext)
0437         return 0;
0438
0439     lnum = c->ihead_lnum;
0440     buf_offs = c->ihead_offs;
0441
0442     buf_len = ubifs_idx_node_sz(c, c->fanout);
0443     buf_len = ALIGN(buf_len, c->min_io_size);
0444     used = 0;
0445     avail = buf_len;
0446
0447     /* Ensure there is enough room for first write */
0448     next_len = ubifs_idx_node_sz(c, cnext->child_cnt);
0449     if (buf_offs + next_len > c->leb_size)
0450         lnum = -1;
0451
0452     while (1) {
0453         znode = cnext;
0454
0455         len = ubifs_idx_node_sz(c, znode->child_cnt);
0456
0457         /* Determine the index node position */
0458         if (lnum == -1) {
0459             if (c->ileb_nxt >= c->ileb_cnt) {
0460                 ubifs_err(c, "out of space");
0461                 return -ENOSPC;
0462             }
0463             lnum = c->ilebs[c->ileb_nxt++];
0464             buf_offs = 0;
0465             used = 0;
0466             avail = buf_len;
0467         }
0468
0469         offs = buf_offs + used;
0470
0471         znode->lnum = lnum;
0472         znode->offs = offs;
0473         znode->len = len;
0474
0475         /* Update the parent */
0476         zp = znode->parent;
0477         if (zp) {
0478             struct ubifs_zbranch *zbr;
0479             int i;
0480
0481             i = znode->iip;
0482             zbr = &zp->zbranch[i];
0483             zbr->lnum = lnum;
0484             zbr->offs = offs;
0485             zbr->len = len;
0486         } else {
0487             c->zroot.lnum = lnum;
0488             c->zroot.offs = offs;
0489             c->zroot.len = len;
0490         }
0491         c->calc_idx_sz += ALIGN(len, 8);
0492
0493         /*
0494          * Once lprops is updated, we can decrease the dirty znode count
0495          * but it is easier to just do it here.
0496          */
0497         atomic_long_dec(&c->dirty_zn_cnt);
0498
0499         /*
0500          * Calculate the next index node length to see if there is
0501          * enough room for it
0502          */
0503         cnext = znode->cnext;
0504         if (cnext == c->cnext)
0505             next_len = 0;
0506         else
0507             next_len = ubifs_idx_node_sz(c, cnext->child_cnt);
0508
0509         /* Update buffer positions */
0510         wlen = used + len;
0511         used += ALIGN(len, 8);
0512         avail -= ALIGN(len, 8);
0513
0514         if (next_len != 0 &&
0515             buf_offs + used + next_len <= c->leb_size &&
0516             avail > 0)
0517             continue;
0518
0519         if (avail <= 0 && next_len &&
0520             buf_offs + used + next_len <= c->leb_size)
0521             blen = buf_len;
0522         else
0523             blen = ALIGN(wlen, c->min_io_size);
0524
0525         /* The buffer is full or there are no more znodes to do */
0526         buf_offs += blen;
0527         if (next_len) {
0528             if (buf_offs + next_len > c->leb_size) {
0529                 err = ubifs_update_one_lp(c, lnum,
0530                     c->leb_size - buf_offs, blen - used,
0531                     0, 0);
0532                 if (err)
0533                     return err;
0534                 lnum = -1;
0535             }
0536             used -= blen;
0537             if (used < 0)
0538                 used = 0;
0539             avail = buf_len - used;
0540             continue;
0541         }
0542         err = ubifs_update_one_lp(c, lnum, c->leb_size - buf_offs,
0543                       blen - used, 0, 0);
0544         if (err)
0545             return err;
0546         break;
0547     }
0548
0549     c->dbg->new_ihead_lnum = lnum;
0550     c->dbg->new_ihead_offs = buf_offs;
0551
0552     return 0;
0553 }
0554
0555 /**
0556  * layout_commit - determine positions of index nodes to commit.
0557  * @c: UBIFS file-system description object
0558  * @no_space: indicates that insufficient empty LEBs were allocated
0559  * @cnt: number of znodes to commit
0560  *
0561  * Calculate and update the positions of index nodes to commit.  If there were
0562  * an insufficient number of empty LEBs allocated, then index nodes are placed
0563  * into the gaps created by obsolete index nodes in non-empty index LEBs.  For
0564  * this purpose, an obsolete index node is one that was not in the index as at
0565  * the end of the last commit.  To write "in-the-gaps" requires that those index
0566  * LEBs are updated atomically in-place.
0567  */
0568 static int layout_commit(struct ubifs_info *c, int no_space, int cnt)
0569 {
0570     int err;
0571
0572     if (no_space) {
0573         err = layout_in_gaps(c, cnt);
0574         if (err)
0575             return err;
0576     }
0577     err = layout_in_empty_space(c);
0578     return err;
0579 }
0580
0581 /**
0582  * find_first_dirty - find first dirty znode.
0583  * @znode: znode to begin searching from
0584  */
0585 static struct ubifs_znode *find_first_dirty(struct ubifs_znode *znode)
0586 {
0587     int i, cont;
0588
0589     if (!znode)
0590         return NULL;
0591
0592     while (1) {
0593         if (znode->level == 0) {
0594             if (ubifs_zn_dirty(znode))
0595                 return znode;
0596             return NULL;
0597         }
0598         cont = 0;
0599         for (i = 0; i < znode->child_cnt; i++) {
0600             struct ubifs_zbranch *zbr = &znode->zbranch[i];
0601
0602             if (zbr->znode && ubifs_zn_dirty(zbr->znode)) {
0603                 znode = zbr->znode;
0604                 cont = 1;
0605                 break;
0606             }
0607         }
0608         if (!cont) {
0609             if (ubifs_zn_dirty(znode))
0610                 return znode;
0611             return NULL;
0612         }
0613     }
0614 }
0615
0616 /**
0617  * find_next_dirty - find next dirty znode.
0618  * @znode: znode to begin searching from
0619  */
0620 static struct ubifs_znode *find_next_dirty(struct ubifs_znode *znode)
0621 {
0622     int n = znode->iip + 1;
0623
0624     znode = znode->parent;
0625     if (!znode)
0626         return NULL;
0627     for (; n < znode->child_cnt; n++) {
0628         struct ubifs_zbranch *zbr = &znode->zbranch[n];
0629
0630         if (zbr->znode && ubifs_zn_dirty(zbr->znode))
0631             return find_first_dirty(zbr->znode);
0632     }
0633     return znode;
0634 }
0635
0636 /**
0637  * get_znodes_to_commit - create list of dirty znodes to commit.
0638  * @c: UBIFS file-system description object
0639  *
0640  * This function returns the number of znodes to commit.
0641  */
0642 static int get_znodes_to_commit(struct ubifs_info *c)
0643 {
0644     struct ubifs_znode *znode, *cnext;
0645     int cnt = 0;
0646
0647     c->cnext = find_first_dirty(c->zroot.znode);
0648     znode = c->enext = c->cnext;
0649     if (!znode) {
0650         dbg_cmt("no znodes to commit");
0651         return 0;
0652     }
0653     cnt += 1;
0654     while (1) {
0655         ubifs_assert(c, !ubifs_zn_cow(znode));
0656         __set_bit(COW_ZNODE, &znode->flags);
0657         znode->alt = 0;
0658         cnext = find_next_dirty(znode);
0659         if (!cnext) {
0660             znode->cnext = c->cnext;
0661             break;
0662         }
0663         znode->cparent = znode->parent;
0664         znode->ciip = znode->iip;
0665         znode->cnext = cnext;
0666         znode = cnext;
0667         cnt += 1;
0668     }
0669     dbg_cmt("committing %d znodes", cnt);
0670     ubifs_assert(c, cnt == atomic_long_read(&c->dirty_zn_cnt));
0671     return cnt;
0672 }
0673
0674 /**
0675  * alloc_idx_lebs - allocate empty LEBs to be used to commit.
0676  * @c: UBIFS file-system description object
0677  * @cnt: number of znodes to commit
0678  *
0679  * This function returns %-ENOSPC if it cannot allocate a sufficient number of
0680  * empty LEBs.  %0 is returned on success, otherwise a negative error code
0681  * is returned.
0682  */
0683 static int alloc_idx_lebs(struct ubifs_info *c, int cnt)
0684 {
0685     int i, leb_cnt, lnum;
0686
0687     c->ileb_cnt = 0;
0688     c->ileb_nxt = 0;
0689     leb_cnt = get_leb_cnt(c, cnt);
0690     dbg_cmt("need about %d empty LEBS for TNC commit", leb_cnt);
0691     if (!leb_cnt)
0692         return 0;
0693     c->ilebs = kmalloc_array(leb_cnt, sizeof(int), GFP_NOFS);
0694     if (!c->ilebs)
0695         return -ENOMEM;
0696     for (i = 0; i < leb_cnt; i++) {
0697         lnum = ubifs_find_free_leb_for_idx(c);
0698         if (lnum < 0)
0699             return lnum;
0700         c->ilebs[c->ileb_cnt++] = lnum;
0701         dbg_cmt("LEB %d", lnum);
0702     }
0703     if (dbg_is_chk_index(c) && !(prandom_u32() & 7))
0704         return -ENOSPC;
0705     return 0;
0706 }
0707
0708 /**
0709  * free_unused_idx_lebs - free unused LEBs that were allocated for the commit.
0710  * @c: UBIFS file-system description object
0711  *
0712  * It is possible that we allocate more empty LEBs for the commit than we need.
0713  * This functions frees the surplus.
0714  *
0715  * This function returns %0 on success and a negative error code on failure.
0716  */
0717 static int free_unused_idx_lebs(struct ubifs_info *c)
0718 {
0719     int i, err = 0, lnum, er;
0720
0721     for (i = c->ileb_nxt; i < c->ileb_cnt; i++) {
0722         lnum = c->ilebs[i];
0723         dbg_cmt("LEB %d", lnum);
0724         er = ubifs_change_one_lp(c, lnum, LPROPS_NC, LPROPS_NC, 0,
0725                      LPROPS_INDEX | LPROPS_TAKEN, 0);
0726         if (!err)
0727             err = er;
0728     }
0729     return err;
0730 }
0731
0732 /**
0733  * free_idx_lebs - free unused LEBs after commit end.
0734  * @c: UBIFS file-system description object
0735  *
0736  * This function returns %0 on success and a negative error code on failure.
0737  */
0738 static int free_idx_lebs(struct ubifs_info *c)
0739 {
0740     int err;
0741
0742     err = free_unused_idx_lebs(c);
0743     kfree(c->ilebs);
0744     c->ilebs = NULL;
0745     return err;
0746 }
0747
0748 /**
0749  * ubifs_tnc_start_commit - start TNC commit.
0750  * @c: UBIFS file-system description object
0751  * @zroot: new index root position is returned here
0752  *
0753  * This function prepares the list of indexing nodes to commit and lays out
0754  * their positions on flash. If there is not enough free space it uses the
0755  * in-gap commit method. Returns zero in case of success and a negative error
0756  * code in case of failure.
0757  */
0758 int ubifs_tnc_start_commit(struct ubifs_info *c, struct ubifs_zbranch *zroot)
0759 {
0760     int err = 0, cnt;
0761
0762     mutex_lock(&c->tnc_mutex);
0763     err = dbg_check_tnc(c, 1);
0764     if (err)
0765         goto out;
0766     cnt = get_znodes_to_commit(c);
0767     if (cnt != 0) {
0768         int no_space = 0;
0769
0770         err = alloc_idx_lebs(c, cnt);
0771         if (err == -ENOSPC)
0772             no_space = 1;
0773         else if (err)
0774             goto out_free;
0775         err = layout_commit(c, no_space, cnt);
0776         if (err)
0777             goto out_free;
0778         ubifs_assert(c, atomic_long_read(&c->dirty_zn_cnt) == 0);
0779         err = free_unused_idx_lebs(c);
0780         if (err)
0781             goto out;
0782     }
0783     destroy_old_idx(c);
0784     memcpy(zroot, &c->zroot, sizeof(struct ubifs_zbranch));
0785
0786     err = ubifs_save_dirty_idx_lnums(c);
0787     if (err)
0788         goto out;
0789
0790     spin_lock(&c->space_lock);
0791     /*
0792      * Although we have not finished committing yet, update size of the
0793      * committed index ('c->bi.old_idx_sz') and zero out the index growth
0794      * budget. It is OK to do this now, because we've reserved all the
0795      * space which is needed to commit the index, and it is save for the
0796      * budgeting subsystem to assume the index is already committed,
0797      * even though it is not.
0798      */
0799     ubifs_assert(c, c->bi.min_idx_lebs == ubifs_calc_min_idx_lebs(c));
0800     c->bi.old_idx_sz = c->calc_idx_sz;
0801     c->bi.uncommitted_idx = 0;
0802     c->bi.min_idx_lebs = ubifs_calc_min_idx_lebs(c);
0803     spin_unlock(&c->space_lock);
0804     mutex_unlock(&c->tnc_mutex);
0805
0806     dbg_cmt("number of index LEBs %d", c->lst.idx_lebs);
0807     dbg_cmt("size of index %llu", c->calc_idx_sz);
0808     return err;
0809
0810 out_free:
0811     free_idx_lebs(c);
0812 out:
0813     mutex_unlock(&c->tnc_mutex);
0814     return err;
0815 }
0816
0817 /**
0818  * write_index - write index nodes.
0819  * @c: UBIFS file-system description object
0820  *
0821  * This function writes the index nodes whose positions were laid out in the
0822  * layout_in_empty_space function.
0823  */
0824 static int write_index(struct ubifs_info *c)
0825 {
0826     struct ubifs_idx_node *idx;
0827     struct ubifs_znode *znode, *cnext;
0828     int i, lnum, offs, len, next_len, buf_len, buf_offs, used;
0829     int avail, wlen, err, lnum_pos = 0, blen, nxt_offs;
0830
0831     cnext = c->enext;
0832     if (!cnext)
0833         return 0;
0834
0835     /*
0836      * Always write index nodes to the index head so that index nodes and
0837      * other types of nodes are never mixed in the same erase block.
0838      */
0839     lnum = c->ihead_lnum;
0840     buf_offs = c->ihead_offs;
0841
0842     /* Allocate commit buffer */
0843     buf_len = ALIGN(c->max_idx_node_sz, c->min_io_size);
0844     used = 0;
0845     avail = buf_len;
0846
0847     /* Ensure there is enough room for first write */
0848     next_len = ubifs_idx_node_sz(c, cnext->child_cnt);
0849     if (buf_offs + next_len > c->leb_size) {
0850         err = ubifs_update_one_lp(c, lnum, LPROPS_NC, 0, 0,
0851                       LPROPS_TAKEN);
0852         if (err)
0853             return err;
0854         lnum = -1;
0855     }
0856
0857     while (1) {
0858         u8 hash[UBIFS_HASH_ARR_SZ];
0859
0860         cond_resched();
0861
0862         znode = cnext;
0863         idx = c->cbuf + used;
0864
0865         /* Make index node */
0866         idx->ch.node_type = UBIFS_IDX_NODE;
0867         idx->child_cnt = cpu_to_le16(znode->child_cnt);
0868         idx->level = cpu_to_le16(znode->level);
0869         for (i = 0; i < znode->child_cnt; i++) {
0870             struct ubifs_branch *br = ubifs_idx_branch(c, idx, i);
0871             struct ubifs_zbranch *zbr = &znode->zbranch[i];
0872
0873             key_write_idx(c, &zbr->key, &br->key);
0874             br->lnum = cpu_to_le32(zbr->lnum);
0875             br->offs = cpu_to_le32(zbr->offs);
0876             br->len = cpu_to_le32(zbr->len);
0877             ubifs_copy_hash(c, zbr->hash, ubifs_branch_hash(c, br));
0878             if (!zbr->lnum || !zbr->len) {
0879                 ubifs_err(c, "bad ref in znode");
0880                 ubifs_dump_znode(c, znode);
0881                 if (zbr->znode)
0882                     ubifs_dump_znode(c, zbr->znode);
0883
0884                 return -EINVAL;
0885             }
0886         }
0887         len = ubifs_idx_node_sz(c, znode->child_cnt);
0888         ubifs_prepare_node(c, idx, len, 0);
0889         ubifs_node_calc_hash(c, idx, hash);
0890
0891         mutex_lock(&c->tnc_mutex);
0892
0893         if (znode->cparent)
0894             ubifs_copy_hash(c, hash,
0895                     znode->cparent->zbranch[znode->ciip].hash);
0896
0897         if (znode->parent) {
0898             if (!ubifs_zn_obsolete(znode))
0899                 ubifs_copy_hash(c, hash,
0900                     znode->parent->zbranch[znode->iip].hash);
0901         } else {
0902             ubifs_copy_hash(c, hash, c->zroot.hash);
0903         }
0904
0905         mutex_unlock(&c->tnc_mutex);
0906
0907         /* Determine the index node position */
0908         if (lnum == -1) {
0909             lnum = c->ilebs[lnum_pos++];
0910             buf_offs = 0;
0911             used = 0;
0912             avail = buf_len;
0913         }
0914         offs = buf_offs + used;
0915
0916         if (lnum != znode->lnum || offs != znode->offs ||
0917             len != znode->len) {
0918             ubifs_err(c, "inconsistent znode posn");
0919             return -EINVAL;
0920         }
0921
0922         /* Grab some stuff from znode while we still can */
0923         cnext = znode->cnext;
0924
0925         ubifs_assert(c, ubifs_zn_dirty(znode));
0926         ubifs_assert(c, ubifs_zn_cow(znode));
0927
0928         /*
0929          * It is important that other threads should see %DIRTY_ZNODE
0930          * flag cleared before %COW_ZNODE. Specifically, it matters in
0931          * the 'dirty_cow_znode()' function. This is the reason for the
0932          * first barrier. Also, we want the bit changes to be seen to
0933          * other threads ASAP, to avoid unnecessary copying, which is
0934          * the reason for the second barrier.
0935          */
0936         clear_bit(DIRTY_ZNODE, &znode->flags);
0937         smp_mb__before_atomic();
0938         clear_bit(COW_ZNODE, &znode->flags);
0939         smp_mb__after_atomic();
0940
0941         /*
0942          * We have marked the znode as clean but have not updated the
0943          * @c->clean_zn_cnt counter. If this znode becomes dirty again
0944          * before 'free_obsolete_znodes()' is called, then
0945          * @c->clean_zn_cnt will be decremented before it gets
0946          * incremented (resulting in 2 decrements for the same znode).
0947          * This means that @c->clean_zn_cnt may become negative for a
0948          * while.
0949          *
0950          * Q: why we cannot increment @c->clean_zn_cnt?
0951          * A: because we do not have the @c->tnc_mutex locked, and the
0952          *    following code would be racy and buggy:
0953          *
0954          *    if (!ubifs_zn_obsolete(znode)) {
0955          *            atomic_long_inc(&c->clean_zn_cnt);
0956          *            atomic_long_inc(&ubifs_clean_zn_cnt);
0957          *    }
0958          *
0959          *    Thus, we just delay the @c->clean_zn_cnt update until we
0960          *    have the mutex locked.
0961          */
0962
0963         /* Do not access znode from this point on */
0964
0965         /* Update buffer positions */
0966         wlen = used + len;
0967         used += ALIGN(len, 8);
0968         avail -= ALIGN(len, 8);
0969
0970         /*
0971          * Calculate the next index node length to see if there is
0972          * enough room for it
0973          */
0974         if (cnext == c->cnext)
0975             next_len = 0;
0976         else
0977             next_len = ubifs_idx_node_sz(c, cnext->child_cnt);
0978
0979         nxt_offs = buf_offs + used + next_len;
0980         if (next_len && nxt_offs <= c->leb_size) {
0981             if (avail > 0)
0982                 continue;
0983             else
0984                 blen = buf_len;
0985         } else {
0986             wlen = ALIGN(wlen, 8);
0987             blen = ALIGN(wlen, c->min_io_size);
0988             ubifs_pad(c, c->cbuf + wlen, blen - wlen);
0989         }
0990
0991         /* The buffer is full or there are no more znodes to do */
0992         err = ubifs_leb_write(c, lnum, c->cbuf, buf_offs, blen);
0993         if (err)
0994             return err;
0995         buf_offs += blen;
0996         if (next_len) {
0997             if (nxt_offs > c->leb_size) {
0998                 err = ubifs_update_one_lp(c, lnum, LPROPS_NC, 0,
0999                               0, LPROPS_TAKEN);
1000                 if (err)
1001                     return err;
1002                 lnum = -1;
1003             }
1004             used -= blen;
1005             if (used < 0)
1006                 used = 0;
1007             avail = buf_len - used;
1008             memmove(c->cbuf, c->cbuf + blen, used);
1009             continue;
1010         }
1011         break;
1012     }
1013
1014     if (lnum != c->dbg->new_ihead_lnum ||
1015         buf_offs != c->dbg->new_ihead_offs) {
1016         ubifs_err(c, "inconsistent ihead");
1017         return -EINVAL;
1018     }
1019
1020     c->ihead_lnum = lnum;
1021     c->ihead_offs = buf_offs;
1022
1023     return 0;
1024 }
1025
1026 /**
1027  * free_obsolete_znodes - free obsolete znodes.
1028  * @c: UBIFS file-system description object
1029  *
1030  * At the end of commit end, obsolete znodes are freed.
1031  */
1032 static void free_obsolete_znodes(struct ubifs_info *c)
1033 {
1034     struct ubifs_znode *znode, *cnext;
1035
1036     cnext = c->cnext;
1037     do {
1038         znode = cnext;
1039         cnext = znode->cnext;
1040         if (ubifs_zn_obsolete(znode))
1041             kfree(znode);
1042         else {
1043             znode->cnext = NULL;
1044             atomic_long_inc(&c->clean_zn_cnt);
1045             atomic_long_inc(&ubifs_clean_zn_cnt);
1046         }
1047     } while (cnext != c->cnext);
1048 }
1049
1050 /**
1051  * return_gap_lebs - return LEBs used by the in-gap commit method.
1052  * @c: UBIFS file-system description object
1053  *
1054  * This function clears the "taken" flag for the LEBs which were used by the
1055  * "commit in-the-gaps" method.
1056  */
1057 static int return_gap_lebs(struct ubifs_info *c)
1058 {
1059     int *p, err;
1060
1061     if (!c->gap_lebs)
1062         return 0;
1063
1064     dbg_cmt("");
1065     for (p = c->gap_lebs; *p != -1; p++) {
1066         err = ubifs_change_one_lp(c, *p, LPROPS_NC, LPROPS_NC, 0,
1067                       LPROPS_TAKEN, 0);
1068         if (err)
1069             return err;
1070     }
1071
1072     kfree(c->gap_lebs);
1073     c->gap_lebs = NULL;
1074     return 0;
1075 }
1076
1077 /**
1078  * ubifs_tnc_end_commit - update the TNC for commit end.
1079  * @c: UBIFS file-system description object
1080  *
1081  * Write the dirty znodes.
1082  */
1083 int ubifs_tnc_end_commit(struct ubifs_info *c)
1084 {
1085     int err;
1086
1087     if (!c->cnext)
1088         return 0;
1089
1090     err = return_gap_lebs(c);
1091     if (err)
1092         return err;
1093
1094     err = write_index(c);
1095     if (err)
1096         return err;
1097
1098     mutex_lock(&c->tnc_mutex);
1099
1100     dbg_cmt("TNC height is %d", c->zroot.znode->level + 1);
1101
1102     free_obsolete_znodes(c);
1103
1104     c->cnext = NULL;
1105     kfree(c->ilebs);
1106     c->ilebs = NULL;
1107
1108     mutex_unlock(&c->tnc_mutex);
1109
1110     return 0;
1111 }