fs/reiserfs/journal.c

   1 /*
   2 ** Write ahead logging implementation copyright Chris Mason 2000
   3 **
   4 ** The background commits make this code very interrelated, and
   5 ** overly complex.  I need to rethink things a bit....The major players:
   6 **
   7 ** journal_begin -- call with the number of blocks you expect to log.
   8 **                  If the current transaction is too
   9 **                  old, it will block until the current transaction is
  10 **                  finished, and then start a new one.
  11 **                  Usually, your transaction will get joined in with
  12 **                  previous ones for speed.
  13 **
  14 ** journal_join  -- same as journal_begin, but won't block on the current
  15 **                  transaction regardless of age.  Don't ever call
  16 **                  this.  Ever.  There are only two places it should be
  17 **                  called from, and they are both inside this file.
  18 **
  19 ** journal_mark_dirty -- adds blocks into this transaction.  clears any flags
  20 **                       that might make them get sent to disk
  21 **                       and then marks them BH_JDirty.  Puts the buffer head
  22 **                       into the current transaction hash.
  23 **
  24 ** journal_end -- if the current transaction is batchable, it does nothing
  25 **                   otherwise, it could do an async/synchronous commit, or
  26 **                   a full flush of all log and real blocks in the
  27 **                   transaction.
  28 **
  29 ** flush_old_commits -- if the current transaction is too old, it is ended and
  30 **                      commit blocks are sent to disk.  Forces commit blocks
  31 **                      to disk for all backgrounded commits that have been
  32 **                      around too long.
  33 **                   -- Note, if you call this as an immediate flush from
  34 **                      from within kupdate, it will ignore the immediate flag
  35 */
  36
  37 #include <linux/time.h>
  38 #include <linux/semaphore.h>
  39 #include <linux/vmalloc.h>
  40 #include "reiserfs.h"
  41 #include <linux/kernel.h>
  42 #include <linux/errno.h>
  43 #include <linux/fcntl.h>
  44 #include <linux/stat.h>
  45 #include <linux/string.h>
  46 #include <linux/buffer_head.h>
  47 #include <linux/workqueue.h>
  48 #include <linux/writeback.h>
  49 #include <linux/blkdev.h>
  50 #include <linux/backing-dev.h>
  51 #include <linux/uaccess.h>
  52 #include <linux/slab.h>
  53
  54
  55 /* gets a struct reiserfs_journal_list * from a list head */
  56 #define JOURNAL_LIST_ENTRY(h) (list_entry((h), struct reiserfs_journal_list, \
  57                                j_list))
  58 #define JOURNAL_WORK_ENTRY(h) (list_entry((h), struct reiserfs_journal_list, \
  59                                j_working_list))
  60
  61 #define JOURNAL_TRANS_HALF 1018 /* must be correct to keep the desc and commit
  62                                    structs at 4k */
  63 #define BUFNR 64                /*read ahead */
  64
  65 /* cnode stat bits.  Move these into reiserfs_fs.h */
  66
  67 #define BLOCK_FREED 2           /* this block was freed, and can't be written.  */
  68 #define BLOCK_FREED_HOLDER 3    /* this block was freed during this transaction, and can't be written */
  69
  70 #define BLOCK_NEEDS_FLUSH 4     /* used in flush_journal_list */
  71 #define BLOCK_DIRTIED 5
  72
  73 /* journal list state bits */
  74 #define LIST_TOUCHED 1
  75 #define LIST_DIRTY   2
  76 #define LIST_COMMIT_PENDING  4  /* someone will commit this list */
  77
  78 /* flags for do_journal_end */
  79 #define FLUSH_ALL   1           /* flush commit and real blocks */
  80 #define COMMIT_NOW  2           /* end and commit this transaction */
  81 #define WAIT        4           /* wait for the log blocks to hit the disk */
  82
  83 static int do_journal_end(struct reiserfs_transaction_handle *,
  84                           struct super_block *, unsigned long nblocks,
  85                           int flags);
  86 static int flush_journal_list(struct super_block *s,
  87                               struct reiserfs_journal_list *jl, int flushall);
  88 static int flush_commit_list(struct super_block *s,
  89                              struct reiserfs_journal_list *jl, int flushall);
  90 static int can_dirty(struct reiserfs_journal_cnode *cn);
  91 static int journal_join(struct reiserfs_transaction_handle *th,
  92                         struct super_block *sb, unsigned long nblocks);
  93 static void release_journal_dev(struct super_block *super,
  94                                struct reiserfs_journal *journal);
  95 static int dirty_one_transaction(struct super_block *s,
  96                                  struct reiserfs_journal_list *jl);
  97 static void flush_async_commits(struct work_struct *work);
  98 static void queue_log_writer(struct super_block *s);
  99
 100 /* values for join in do_journal_begin_r */
 101 enum {
 102         JBEGIN_REG = 0,         /* regular journal begin */
 103         JBEGIN_JOIN = 1,        /* join the running transaction if at all possible */
 104         JBEGIN_ABORT = 2,       /* called from cleanup code, ignores aborted flag */
 105 };
 106
 107 static int do_journal_begin_r(struct reiserfs_transaction_handle *th,
 108                               struct super_block *sb,
 109                               unsigned long nblocks, int join);
 110
 111 static void init_journal_hash(struct super_block *sb)
 112 {
 113         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 114         memset(journal->j_hash_table, 0,
 115                JOURNAL_HASH_SIZE * sizeof(struct reiserfs_journal_cnode *));
 116 }
 117
 118 /*
 119 ** clears BH_Dirty and sticks the buffer on the clean list.  Called because I can't allow refile_buffer to
 120 ** make schedule happen after I've freed a block.  Look at remove_from_transaction and journal_mark_freed for
 121 ** more details.
 122 */
 123 static int reiserfs_clean_and_file_buffer(struct buffer_head *bh)
 124 {
 125         if (bh) {
 126                 clear_buffer_dirty(bh);
 127                 clear_buffer_journal_test(bh);
 128         }
 129         return 0;
 130 }
 131
 132 static struct reiserfs_bitmap_node *allocate_bitmap_node(struct super_block
 133                                                          *sb)
 134 {
 135         struct reiserfs_bitmap_node *bn;
 136         static int id;
 137
 138         bn = kmalloc(sizeof(struct reiserfs_bitmap_node), GFP_NOFS);
 139         if (!bn) {
 140                 return NULL;
 141         }
 142         bn->data = kzalloc(sb->s_blocksize, GFP_NOFS);
 143         if (!bn->data) {
 144                 kfree(bn);
 145                 return NULL;
 146         }
 147         bn->id = id++;
 148         INIT_LIST_HEAD(&bn->list);
 149         return bn;
 150 }
 151
 152 static struct reiserfs_bitmap_node *get_bitmap_node(struct super_block *sb)
 153 {
 154         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 155         struct reiserfs_bitmap_node *bn = NULL;
 156         struct list_head *entry = journal->j_bitmap_nodes.next;
 157
 158         journal->j_used_bitmap_nodes++;
 159       repeat:
 160
 161         if (entry != &journal->j_bitmap_nodes) {
 162                 bn = list_entry(entry, struct reiserfs_bitmap_node, list);
 163                 list_del(entry);
 164                 memset(bn->data, 0, sb->s_blocksize);
 165                 journal->j_free_bitmap_nodes--;
 166                 return bn;
 167         }
 168         bn = allocate_bitmap_node(sb);
 169         if (!bn) {
 170                 yield();
 171                 goto repeat;
 172         }
 173         return bn;
 174 }
 175 static inline void free_bitmap_node(struct super_block *sb,
 176                                     struct reiserfs_bitmap_node *bn)
 177 {
 178         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 179         journal->j_used_bitmap_nodes--;
 180         if (journal->j_free_bitmap_nodes > REISERFS_MAX_BITMAP_NODES) {
 181                 kfree(bn->data);
 182                 kfree(bn);
 183         } else {
 184                 list_add(&bn->list, &journal->j_bitmap_nodes);
 185                 journal->j_free_bitmap_nodes++;
 186         }
 187 }
 188
 189 static void allocate_bitmap_nodes(struct super_block *sb)
 190 {
 191         int i;
 192         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 193         struct reiserfs_bitmap_node *bn = NULL;
 194         for (i = 0; i < REISERFS_MIN_BITMAP_NODES; i++) {
 195                 bn = allocate_bitmap_node(sb);
 196                 if (bn) {
 197                         list_add(&bn->list, &journal->j_bitmap_nodes);
 198                         journal->j_free_bitmap_nodes++;
 199                 } else {
 200                         break;  /* this is ok, we'll try again when more are needed */
 201                 }
 202         }
 203 }
 204
 205 static int set_bit_in_list_bitmap(struct super_block *sb,
 206                                   b_blocknr_t block,
 207                                   struct reiserfs_list_bitmap *jb)
 208 {
 209         unsigned int bmap_nr = block / (sb->s_blocksize << 3);
 210         unsigned int bit_nr = block % (sb->s_blocksize << 3);
 211
 212         if (!jb->bitmaps[bmap_nr]) {
 213                 jb->bitmaps[bmap_nr] = get_bitmap_node(sb);
 214         }
 215         set_bit(bit_nr, (unsigned long *)jb->bitmaps[bmap_nr]->data);
 216         return 0;
 217 }
 218
 219 static void cleanup_bitmap_list(struct super_block *sb,
 220                                 struct reiserfs_list_bitmap *jb)
 221 {
 222         int i;
 223         if (jb->bitmaps == NULL)
 224                 return;
 225
 226         for (i = 0; i < reiserfs_bmap_count(sb); i++) {
 227                 if (jb->bitmaps[i]) {
 228                         free_bitmap_node(sb, jb->bitmaps[i]);
 229                         jb->bitmaps[i] = NULL;
 230                 }
 231         }
 232 }
 233
 234 /*
 235 ** only call this on FS unmount.
 236 */
 237 static int free_list_bitmaps(struct super_block *sb,
 238                              struct reiserfs_list_bitmap *jb_array)
 239 {
 240         int i;
 241         struct reiserfs_list_bitmap *jb;
 242         for (i = 0; i < JOURNAL_NUM_BITMAPS; i++) {
 243                 jb = jb_array + i;
 244                 jb->journal_list = NULL;
 245                 cleanup_bitmap_list(sb, jb);
 246                 vfree(jb->bitmaps);
 247                 jb->bitmaps = NULL;
 248         }
 249         return 0;
 250 }
 251
 252 static int free_bitmap_nodes(struct super_block *sb)
 253 {
 254         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 255         struct list_head *next = journal->j_bitmap_nodes.next;
 256         struct reiserfs_bitmap_node *bn;
 257
 258         while (next != &journal->j_bitmap_nodes) {
 259                 bn = list_entry(next, struct reiserfs_bitmap_node, list);
 260                 list_del(next);
 261                 kfree(bn->data);
 262                 kfree(bn);
 263                 next = journal->j_bitmap_nodes.next;
 264                 journal->j_free_bitmap_nodes--;
 265         }
 266
 267         return 0;
 268 }
 269
 270 /*
 271 ** get memory for JOURNAL_NUM_BITMAPS worth of bitmaps.
 272 ** jb_array is the array to be filled in.
 273 */
 274 int reiserfs_allocate_list_bitmaps(struct super_block *sb,
 275                                    struct reiserfs_list_bitmap *jb_array,
 276                                    unsigned int bmap_nr)
 277 {
 278         int i;
 279         int failed = 0;
 280         struct reiserfs_list_bitmap *jb;
 281         int mem = bmap_nr * sizeof(struct reiserfs_bitmap_node *);
 282
 283         for (i = 0; i < JOURNAL_NUM_BITMAPS; i++) {
 284                 jb = jb_array + i;
 285                 jb->journal_list = NULL;
 286                 jb->bitmaps = vzalloc(mem);
 287                 if (!jb->bitmaps) {
 288                         reiserfs_warning(sb, "clm-2000", "unable to "
 289                                          "allocate bitmaps for journal lists");
 290                         failed = 1;
 291                         break;
 292                 }
 293         }
 294         if (failed) {
 295                 free_list_bitmaps(sb, jb_array);
 296                 return -1;
 297         }
 298         return 0;
 299 }
 300
 301 /*
 302 ** find an available list bitmap.  If you can't find one, flush a commit list
 303 ** and try again
 304 */
 305 static struct reiserfs_list_bitmap *get_list_bitmap(struct super_block *sb,
 306                                                     struct reiserfs_journal_list
 307                                                     *jl)
 308 {
 309         int i, j;
 310         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 311         struct reiserfs_list_bitmap *jb = NULL;
 312
 313         for (j = 0; j < (JOURNAL_NUM_BITMAPS * 3); j++) {
 314                 i = journal->j_list_bitmap_index;
 315                 journal->j_list_bitmap_index = (i + 1) % JOURNAL_NUM_BITMAPS;
 316                 jb = journal->j_list_bitmap + i;
 317                 if (journal->j_list_bitmap[i].journal_list) {
 318                         flush_commit_list(sb,
 319                                           journal->j_list_bitmap[i].
 320                                           journal_list, 1);
 321                         if (!journal->j_list_bitmap[i].journal_list) {
 322                                 break;
 323                         }
 324                 } else {
 325                         break;
 326                 }
 327         }
 328         if (jb->journal_list) { /* double check to make sure if flushed correctly */
 329                 return NULL;
 330         }
 331         jb->journal_list = jl;
 332         return jb;
 333 }
 334
 335 /*
 336 ** allocates a new chunk of X nodes, and links them all together as a list.
 337 ** Uses the cnode->next and cnode->prev pointers
 338 ** returns NULL on failure
 339 */
 340 static struct reiserfs_journal_cnode *allocate_cnodes(int num_cnodes)
 341 {
 342         struct reiserfs_journal_cnode *head;
 343         int i;
 344         if (num_cnodes <= 0) {
 345                 return NULL;
 346         }
 347         head = vzalloc(num_cnodes * sizeof(struct reiserfs_journal_cnode));
 348         if (!head) {
 349                 return NULL;
 350         }
 351         head[0].prev = NULL;
 352         head[0].next = head + 1;
 353         for (i = 1; i < num_cnodes; i++) {
 354                 head[i].prev = head + (i - 1);
 355                 head[i].next = head + (i + 1);  /* if last one, overwrite it after the if */
 356         }
 357         head[num_cnodes - 1].next = NULL;
 358         return head;
 359 }
 360
 361 /*
 362 ** pulls a cnode off the free list, or returns NULL on failure
 363 */
 364 static struct reiserfs_journal_cnode *get_cnode(struct super_block *sb)
 365 {
 366         struct reiserfs_journal_cnode *cn;
 367         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 368
 369         reiserfs_check_lock_depth(sb, "get_cnode");
 370
 371         if (journal->j_cnode_free <= 0) {
 372                 return NULL;
 373         }
 374         journal->j_cnode_used++;
 375         journal->j_cnode_free--;
 376         cn = journal->j_cnode_free_list;
 377         if (!cn) {
 378                 return cn;
 379         }
 380         if (cn->next) {
 381                 cn->next->prev = NULL;
 382         }
 383         journal->j_cnode_free_list = cn->next;
 384         memset(cn, 0, sizeof(struct reiserfs_journal_cnode));
 385         return cn;
 386 }
 387
 388 /*
 389 ** returns a cnode to the free list
 390 */
 391 static void free_cnode(struct super_block *sb,
 392                        struct reiserfs_journal_cnode *cn)
 393 {
 394         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 395
 396         reiserfs_check_lock_depth(sb, "free_cnode");
 397
 398         journal->j_cnode_used--;
 399         journal->j_cnode_free++;
 400         /* memset(cn, 0, sizeof(struct reiserfs_journal_cnode)) ; */
 401         cn->next = journal->j_cnode_free_list;
 402         if (journal->j_cnode_free_list) {
 403                 journal->j_cnode_free_list->prev = cn;
 404         }
 405         cn->prev = NULL;        /* not needed with the memset, but I might kill the memset, and forget to do this */
 406         journal->j_cnode_free_list = cn;
 407 }
 408
 409 static void clear_prepared_bits(struct buffer_head *bh)
 410 {
 411         clear_buffer_journal_prepared(bh);
 412         clear_buffer_journal_restore_dirty(bh);
 413 }
 414
 415 /* return a cnode with same dev, block number and size in table, or null if not found */
 416 static inline struct reiserfs_journal_cnode *get_journal_hash_dev(struct
 417                                                                   super_block
 418                                                                   *sb,
 419                                                                   struct
 420                                                                   reiserfs_journal_cnode
 421                                                                   **table,
 422                                                                   long bl)
 423 {
 424         struct reiserfs_journal_cnode *cn;
 425         cn = journal_hash(table, sb, bl);
 426         while (cn) {
 427                 if (cn->blocknr == bl && cn->sb == sb)
 428                         return cn;
 429                 cn = cn->hnext;
 430         }
 431         return (struct reiserfs_journal_cnode *)0;
 432 }
 433
 434 /*
 435 ** this actually means 'can this block be reallocated yet?'.  If you set search_all, a block can only be allocated
 436 ** if it is not in the current transaction, was not freed by the current transaction, and has no chance of ever
 437 ** being overwritten by a replay after crashing.
 438 **
 439 ** If you don't set search_all, a block can only be allocated if it is not in the current transaction.  Since deleting
 440 ** a block removes it from the current transaction, this case should never happen.  If you don't set search_all, make
 441 ** sure you never write the block without logging it.
 442 **
 443 ** next_zero_bit is a suggestion about the next block to try for find_forward.
 444 ** when bl is rejected because it is set in a journal list bitmap, we search
 445 ** for the next zero bit in the bitmap that rejected bl.  Then, we return that
 446 ** through next_zero_bit for find_forward to try.
 447 **
 448 ** Just because we return something in next_zero_bit does not mean we won't
 449 ** reject it on the next call to reiserfs_in_journal
 450 **
 451 */
 452 int reiserfs_in_journal(struct super_block *sb,
 453                         unsigned int bmap_nr, int bit_nr, int search_all,
 454                         b_blocknr_t * next_zero_bit)
 455 {
 456         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 457         struct reiserfs_journal_cnode *cn;
 458         struct reiserfs_list_bitmap *jb;
 459         int i;
 460         unsigned long bl;
 461
 462         *next_zero_bit = 0;     /* always start this at zero. */
 463
 464         PROC_INFO_INC(sb, journal.in_journal);
 465         /* If we aren't doing a search_all, this is a metablock, and it will be logged before use.
 466          ** if we crash before the transaction that freed it commits,  this transaction won't
 467          ** have committed either, and the block will never be written
 468          */
 469         if (search_all) {
 470                 for (i = 0; i < JOURNAL_NUM_BITMAPS; i++) {
 471                         PROC_INFO_INC(sb, journal.in_journal_bitmap);
 472                         jb = journal->j_list_bitmap + i;
 473                         if (jb->journal_list && jb->bitmaps[bmap_nr] &&
 474                             test_bit(bit_nr,
 475                                      (unsigned long *)jb->bitmaps[bmap_nr]->
 476                                      data)) {
 477                                 *next_zero_bit =
 478                                     find_next_zero_bit((unsigned long *)
 479                                                        (jb->bitmaps[bmap_nr]->
 480                                                         data),
 481                                                        sb->s_blocksize << 3,
 482                                                        bit_nr + 1);
 483                                 return 1;
 484                         }
 485                 }
 486         }
 487
 488         bl = bmap_nr * (sb->s_blocksize << 3) + bit_nr;
 489         /* is it in any old transactions? */
 490         if (search_all
 491             && (cn =
 492                 get_journal_hash_dev(sb, journal->j_list_hash_table, bl))) {
 493                 return 1;
 494         }
 495
 496         /* is it in the current transaction.  This should never happen */
 497         if ((cn = get_journal_hash_dev(sb, journal->j_hash_table, bl))) {
 498                 BUG();
 499                 return 1;
 500         }
 501
 502         PROC_INFO_INC(sb, journal.in_journal_reusable);
 503         /* safe for reuse */
 504         return 0;
 505 }
 506
 507 /* insert cn into table
 508 */
 509 static inline void insert_journal_hash(struct reiserfs_journal_cnode **table,
 510                                        struct reiserfs_journal_cnode *cn)
 511 {
 512         struct reiserfs_journal_cnode *cn_orig;
 513
 514         cn_orig = journal_hash(table, cn->sb, cn->blocknr);
 515         cn->hnext = cn_orig;
 516         cn->hprev = NULL;
 517         if (cn_orig) {
 518                 cn_orig->hprev = cn;
 519         }
 520         journal_hash(table, cn->sb, cn->blocknr) = cn;
 521 }
 522
 523 /* lock the current transaction */
 524 static inline void lock_journal(struct super_block *sb)
 525 {
 526         PROC_INFO_INC(sb, journal.lock_journal);
 527
 528         reiserfs_mutex_lock_safe(&SB_JOURNAL(sb)->j_mutex, sb);
 529 }
 530
 531 /* unlock the current transaction */
 532 static inline void unlock_journal(struct super_block *sb)
 533 {
 534         mutex_unlock(&SB_JOURNAL(sb)->j_mutex);
 535 }
 536
 537 static inline void get_journal_list(struct reiserfs_journal_list *jl)
 538 {
 539         jl->j_refcount++;
 540 }
 541
 542 static inline void put_journal_list(struct super_block *s,
 543                                     struct reiserfs_journal_list *jl)
 544 {
 545         if (jl->j_refcount < 1) {
 546                 reiserfs_panic(s, "journal-2", "trans id %u, refcount at %d",
 547                                jl->j_trans_id, jl->j_refcount);
 548         }
 549         if (--jl->j_refcount == 0)
 550                 kfree(jl);
 551 }
 552
 553 /*
 554 ** this used to be much more involved, and I'm keeping it just in case things get ugly again.
 555 ** it gets called by flush_commit_list, and cleans up any data stored about blocks freed during a
 556 ** transaction.
 557 */
 558 static void cleanup_freed_for_journal_list(struct super_block *sb,
 559                                            struct reiserfs_journal_list *jl)
 560 {
 561
 562         struct reiserfs_list_bitmap *jb = jl->j_list_bitmap;
 563         if (jb) {
 564                 cleanup_bitmap_list(sb, jb);
 565         }
 566         jl->j_list_bitmap->journal_list = NULL;
 567         jl->j_list_bitmap = NULL;
 568 }
 569
 570 static int journal_list_still_alive(struct super_block *s,
 571                                     unsigned int trans_id)
 572 {
 573         struct reiserfs_journal *journal = SB_JOURNAL(s);
 574         struct list_head *entry = &journal->j_journal_list;
 575         struct reiserfs_journal_list *jl;
 576
 577         if (!list_empty(entry)) {
 578                 jl = JOURNAL_LIST_ENTRY(entry->next);
 579                 if (jl->j_trans_id <= trans_id) {
 580                         return 1;
 581                 }
 582         }
 583         return 0;
 584 }
 585
 586 /*
 587  * If page->mapping was null, we failed to truncate this page for
 588  * some reason.  Most likely because it was truncated after being
 589  * logged via data=journal.
 590  *
 591  * This does a check to see if the buffer belongs to one of these
 592  * lost pages before doing the final put_bh.  If page->mapping was
 593  * null, it tries to free buffers on the page, which should make the
 594  * final page_cache_release drop the page from the lru.
 595  */
 596 static void release_buffer_page(struct buffer_head *bh)
 597 {
 598         struct page *page = bh->b_page;
 599         if (!page->mapping && trylock_page(page)) {
 600                 page_cache_get(page);
 601                 put_bh(bh);
 602                 if (!page->mapping)
 603                         try_to_free_buffers(page);
 604                 unlock_page(page);
 605                 page_cache_release(page);
 606         } else {
 607                 put_bh(bh);
 608         }
 609 }
 610
 611 static void reiserfs_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
 612 {
 613         char b[BDEVNAME_SIZE];
 614
 615         if (buffer_journaled(bh)) {
 616                 reiserfs_warning(NULL, "clm-2084",
 617                                  "pinned buffer %lu:%s sent to disk",
 618                                  bh->b_blocknr, bdevname(bh->b_bdev, b));
 619         }
 620         if (uptodate)
 621                 set_buffer_uptodate(bh);
 622         else
 623                 clear_buffer_uptodate(bh);
 624
 625         unlock_buffer(bh);
 626         release_buffer_page(bh);
 627 }
 628
 629 static void reiserfs_end_ordered_io(struct buffer_head *bh, int uptodate)
 630 {
 631         if (uptodate)
 632                 set_buffer_uptodate(bh);
 633         else
 634                 clear_buffer_uptodate(bh);
 635         unlock_buffer(bh);
 636         put_bh(bh);
 637 }
 638
 639 static void submit_logged_buffer(struct buffer_head *bh)
 640 {
 641         get_bh(bh);
 642         bh->b_end_io = reiserfs_end_buffer_io_sync;
 643         clear_buffer_journal_new(bh);
 644         clear_buffer_dirty(bh);
 645         if (!test_clear_buffer_journal_test(bh))
 646                 BUG();
 647         if (!buffer_uptodate(bh))
 648                 BUG();
 649         submit_bh(WRITE, bh);
 650 }
 651
 652 static void submit_ordered_buffer(struct buffer_head *bh)
 653 {
 654         get_bh(bh);
 655         bh->b_end_io = reiserfs_end_ordered_io;
 656         clear_buffer_dirty(bh);
 657         if (!buffer_uptodate(bh))
 658                 BUG();
 659         submit_bh(WRITE, bh);
 660 }
 661
 662 #define CHUNK_SIZE 32
 663 struct buffer_chunk {
 664         struct buffer_head *bh[CHUNK_SIZE];
 665         int nr;
 666 };
 667
 668 static void write_chunk(struct buffer_chunk *chunk)
 669 {
 670         int i;
 671         for (i = 0; i < chunk->nr; i++) {
 672                 submit_logged_buffer(chunk->bh[i]);
 673         }
 674         chunk->nr = 0;
 675 }
 676
 677 static void write_ordered_chunk(struct buffer_chunk *chunk)
 678 {
 679         int i;
 680         for (i = 0; i < chunk->nr; i++) {
 681                 submit_ordered_buffer(chunk->bh[i]);
 682         }
 683         chunk->nr = 0;
 684 }
 685
 686 static int add_to_chunk(struct buffer_chunk *chunk, struct buffer_head *bh,
 687                         spinlock_t * lock, void (fn) (struct buffer_chunk *))
 688 {
 689         int ret = 0;
 690         BUG_ON(chunk->nr >= CHUNK_SIZE);
 691         chunk->bh[chunk->nr++] = bh;
 692         if (chunk->nr >= CHUNK_SIZE) {
 693                 ret = 1;
 694                 if (lock)
 695                         spin_unlock(lock);
 696                 fn(chunk);
 697                 if (lock)
 698                         spin_lock(lock);
 699         }
 700         return ret;
 701 }
 702
 703 static atomic_t nr_reiserfs_jh = ATOMIC_INIT(0);
 704 static struct reiserfs_jh *alloc_jh(void)
 705 {
 706         struct reiserfs_jh *jh;
 707         while (1) {
 708                 jh = kmalloc(sizeof(*jh), GFP_NOFS);
 709                 if (jh) {
 710                         atomic_inc(&nr_reiserfs_jh);
 711                         return jh;
 712                 }
 713                 yield();
 714         }
 715 }
 716
 717 /*
 718  * we want to free the jh when the buffer has been written
 719  * and waited on
 720  */
 721 void reiserfs_free_jh(struct buffer_head *bh)
 722 {
 723         struct reiserfs_jh *jh;
 724
 725         jh = bh->b_private;
 726         if (jh) {
 727                 bh->b_private = NULL;
 728                 jh->bh = NULL;
 729                 list_del_init(&jh->list);
 730                 kfree(jh);
 731                 if (atomic_read(&nr_reiserfs_jh) <= 0)
 732                         BUG();
 733                 atomic_dec(&nr_reiserfs_jh);
 734                 put_bh(bh);
 735         }
 736 }
 737
 738 static inline int __add_jh(struct reiserfs_journal *j, struct buffer_head *bh,
 739                            int tail)
 740 {
 741         struct reiserfs_jh *jh;
 742
 743         if (bh->b_private) {
 744                 spin_lock(&j->j_dirty_buffers_lock);
 745                 if (!bh->b_private) {
 746                         spin_unlock(&j->j_dirty_buffers_lock);
 747                         goto no_jh;
 748                 }
 749                 jh = bh->b_private;
 750                 list_del_init(&jh->list);
 751         } else {
 752               no_jh:
 753                 get_bh(bh);
 754                 jh = alloc_jh();
 755                 spin_lock(&j->j_dirty_buffers_lock);
 756                 /* buffer must be locked for __add_jh, should be able to have
 757                  * two adds at the same time
 758                  */
 759                 BUG_ON(bh->b_private);
 760                 jh->bh = bh;
 761                 bh->b_private = jh;
 762         }
 763         jh->jl = j->j_current_jl;
 764         if (tail)
 765                 list_add_tail(&jh->list, &jh->jl->j_tail_bh_list);
 766         else {
 767                 list_add_tail(&jh->list, &jh->jl->j_bh_list);
 768         }
 769         spin_unlock(&j->j_dirty_buffers_lock);
 770         return 0;
 771 }
 772
 773 int reiserfs_add_tail_list(struct inode *inode, struct buffer_head *bh)
 774 {
 775         return __add_jh(SB_JOURNAL(inode->i_sb), bh, 1);
 776 }
 777 int reiserfs_add_ordered_list(struct inode *inode, struct buffer_head *bh)
 778 {
 779         return __add_jh(SB_JOURNAL(inode->i_sb), bh, 0);
 780 }
 781
 782 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list)
 783 static int write_ordered_buffers(spinlock_t * lock,
 784                                  struct reiserfs_journal *j,
 785                                  struct reiserfs_journal_list *jl,
 786                                  struct list_head *list)
 787 {
 788         struct buffer_head *bh;
 789         struct reiserfs_jh *jh;
 790         int ret = j->j_errno;
 791         struct buffer_chunk chunk;
 792         struct list_head tmp;
 793         INIT_LIST_HEAD(&tmp);
 794
 795         chunk.nr = 0;
 796         spin_lock(lock);
 797         while (!list_empty(list)) {
 798                 jh = JH_ENTRY(list->next);
 799                 bh = jh->bh;
 800                 get_bh(bh);
 801                 if (!trylock_buffer(bh)) {
 802                         if (!buffer_dirty(bh)) {
 803                                 list_move(&jh->list, &tmp);
 804                                 goto loop_next;
 805                         }
 806                         spin_unlock(lock);
 807                         if (chunk.nr)
 808                                 write_ordered_chunk(&chunk);
 809                         wait_on_buffer(bh);
 810                         cond_resched();
 811                         spin_lock(lock);
 812                         goto loop_next;
 813                 }
 814                 /* in theory, dirty non-uptodate buffers should never get here,
 815                  * but the upper layer io error paths still have a few quirks.
 816                  * Handle them here as gracefully as we can
 817                  */
 818                 if (!buffer_uptodate(bh) && buffer_dirty(bh)) {
 819                         clear_buffer_dirty(bh);
 820                         ret = -EIO;
 821                 }
 822                 if (buffer_dirty(bh)) {
 823                         list_move(&jh->list, &tmp);
 824                         add_to_chunk(&chunk, bh, lock, write_ordered_chunk);
 825                 } else {
 826                         reiserfs_free_jh(bh);
 827                         unlock_buffer(bh);
 828                 }
 829               loop_next:
 830                 put_bh(bh);
 831                 cond_resched_lock(lock);
 832         }
 833         if (chunk.nr) {
 834                 spin_unlock(lock);
 835                 write_ordered_chunk(&chunk);
 836                 spin_lock(lock);
 837         }
 838         while (!list_empty(&tmp)) {
 839                 jh = JH_ENTRY(tmp.prev);
 840                 bh = jh->bh;
 841                 get_bh(bh);
 842                 reiserfs_free_jh(bh);
 843
 844                 if (buffer_locked(bh)) {
 845                         spin_unlock(lock);
 846                         wait_on_buffer(bh);
 847                         spin_lock(lock);
 848                 }
 849                 if (!buffer_uptodate(bh)) {
 850                         ret = -EIO;
 851                 }
 852                 /* ugly interaction with invalidatepage here.
 853                  * reiserfs_invalidate_page will pin any buffer that has a valid
 854                  * journal head from an older transaction.  If someone else sets
 855                  * our buffer dirty after we write it in the first loop, and
 856                  * then someone truncates the page away, nobody will ever write
 857                  * the buffer. We're safe if we write the page one last time
 858                  * after freeing the journal header.
 859                  */
 860                 if (buffer_dirty(bh) && unlikely(bh->b_page->mapping == NULL)) {
 861                         spin_unlock(lock);
 862                         ll_rw_block(WRITE, 1, &bh);
 863                         spin_lock(lock);
 864                 }
 865                 put_bh(bh);
 866                 cond_resched_lock(lock);
 867         }
 868         spin_unlock(lock);
 869         return ret;
 870 }
 871
 872 static int flush_older_commits(struct super_block *s,
 873                                struct reiserfs_journal_list *jl)
 874 {
 875         struct reiserfs_journal *journal = SB_JOURNAL(s);
 876         struct reiserfs_journal_list *other_jl;
 877         struct reiserfs_journal_list *first_jl;
 878         struct list_head *entry;
 879         unsigned int trans_id = jl->j_trans_id;
 880         unsigned int other_trans_id;
 881         unsigned int first_trans_id;
 882
 883       find_first:
 884         /*
 885          * first we walk backwards to find the oldest uncommitted transation
 886          */
 887         first_jl = jl;
 888         entry = jl->j_list.prev;
 889         while (1) {
 890                 other_jl = JOURNAL_LIST_ENTRY(entry);
 891                 if (entry == &journal->j_journal_list ||
 892                     atomic_read(&other_jl->j_older_commits_done))
 893                         break;
 894
 895                 first_jl = other_jl;
 896                 entry = other_jl->j_list.prev;
 897         }
 898
 899         /* if we didn't find any older uncommitted transactions, return now */
 900         if (first_jl == jl) {
 901                 return 0;
 902         }
 903
 904         first_trans_id = first_jl->j_trans_id;
 905
 906         entry = &first_jl->j_list;
 907         while (1) {
 908                 other_jl = JOURNAL_LIST_ENTRY(entry);
 909                 other_trans_id = other_jl->j_trans_id;
 910
 911                 if (other_trans_id < trans_id) {
 912                         if (atomic_read(&other_jl->j_commit_left) != 0) {
 913                                 flush_commit_list(s, other_jl, 0);
 914
 915                                 /* list we were called with is gone, return */
 916                                 if (!journal_list_still_alive(s, trans_id))
 917                                         return 1;
 918
 919                                 /* the one we just flushed is gone, this means all
 920                                  * older lists are also gone, so first_jl is no longer
 921                                  * valid either.  Go back to the beginning.
 922                                  */
 923                                 if (!journal_list_still_alive
 924                                     (s, other_trans_id)) {
 925                                         goto find_first;
 926                                 }
 927                         }
 928                         entry = entry->next;
 929                         if (entry == &journal->j_journal_list)
 930                                 return 0;
 931                 } else {
 932                         return 0;
 933                 }
 934         }
 935         return 0;
 936 }
 937
 938 static int reiserfs_async_progress_wait(struct super_block *s)
 939 {
 940         struct reiserfs_journal *j = SB_JOURNAL(s);
 941
 942         if (atomic_read(&j->j_async_throttle)) {
 943                 int depth;
 944
 945                 depth = reiserfs_write_unlock_nested(s);
 946                 congestion_wait(BLK_RW_ASYNC, HZ / 10);
 947                 reiserfs_write_lock_nested(s, depth);
 948         }
 949
 950         return 0;
 951 }
 952
 953 /*
 954 ** if this journal list still has commit blocks unflushed, send them to disk.
 955 **
 956 ** log areas must be flushed in order (transaction 2 can't commit before transaction 1)
 957 ** Before the commit block can by written, every other log block must be safely on disk
 958 **
 959 */
 960 static int flush_commit_list(struct super_block *s,
 961                              struct reiserfs_journal_list *jl, int flushall)
 962 {
 963         int i;
 964         b_blocknr_t bn;
 965         struct buffer_head *tbh = NULL;
 966         unsigned int trans_id = jl->j_trans_id;
 967         struct reiserfs_journal *journal = SB_JOURNAL(s);
 968         int retval = 0;
 969         int write_len;
 970         int depth;
 971
 972         reiserfs_check_lock_depth(s, "flush_commit_list");
 973
 974         if (atomic_read(&jl->j_older_commits_done)) {
 975                 return 0;
 976         }
 977
 978         /* before we can put our commit blocks on disk, we have to make sure everyone older than
 979          ** us is on disk too
 980          */
 981         BUG_ON(jl->j_len <= 0);
 982         BUG_ON(trans_id == journal->j_trans_id);
 983
 984         get_journal_list(jl);
 985         if (flushall) {
 986                 if (flush_older_commits(s, jl) == 1) {
 987                         /* list disappeared during flush_older_commits.  return */
 988                         goto put_jl;
 989                 }
 990         }
 991
 992         /* make sure nobody is trying to flush this one at the same time */
 993         reiserfs_mutex_lock_safe(&jl->j_commit_mutex, s);
 994
 995         if (!journal_list_still_alive(s, trans_id)) {
 996                 mutex_unlock(&jl->j_commit_mutex);
 997                 goto put_jl;
 998         }
 999         BUG_ON(jl->j_trans_id == 0);
1000
1001         /* this commit is done, exit */
1002         if (atomic_read(&(jl->j_commit_left)) <= 0) {
1003                 if (flushall) {
1004                         atomic_set(&(jl->j_older_commits_done), 1);
1005                 }
1006                 mutex_unlock(&jl->j_commit_mutex);
1007                 goto put_jl;
1008         }
1009
1010         if (!list_empty(&jl->j_bh_list)) {
1011                 int ret;
1012
1013                 /*
1014                  * We might sleep in numerous places inside
1015                  * write_ordered_buffers. Relax the write lock.
1016                  */
1017                 depth = reiserfs_write_unlock_nested(s);
1018                 ret = write_ordered_buffers(&journal->j_dirty_buffers_lock,
1019                                             journal, jl, &jl->j_bh_list);
1020                 if (ret < 0 && retval == 0)
1021                         retval = ret;
1022                 reiserfs_write_lock_nested(s, depth);
1023         }
1024         BUG_ON(!list_empty(&jl->j_bh_list));
1025         /*
1026          * for the description block and all the log blocks, submit any buffers
1027          * that haven't already reached the disk.  Try to write at least 256
1028          * log blocks. later on, we will only wait on blocks that correspond
1029          * to this transaction, but while we're unplugging we might as well
1030          * get a chunk of data on there.
1031          */
1032         atomic_inc(&journal->j_async_throttle);
1033         write_len = jl->j_len + 1;
1034         if (write_len < 256)
1035                 write_len = 256;
1036         for (i = 0 ; i < write_len ; i++) {
1037                 bn = SB_ONDISK_JOURNAL_1st_BLOCK(s) + (jl->j_start + i) %
1038                     SB_ONDISK_JOURNAL_SIZE(s);
1039                 tbh = journal_find_get_block(s, bn);
1040                 if (tbh) {
1041                         if (buffer_dirty(tbh)) {
1042                             depth = reiserfs_write_unlock_nested(s);
1043                             ll_rw_block(WRITE, 1, &tbh);
1044                             reiserfs_write_lock_nested(s, depth);
1045                         }
1046                         put_bh(tbh) ;
1047                 }
1048         }
1049         atomic_dec(&journal->j_async_throttle);
1050
1051         for (i = 0; i < (jl->j_len + 1); i++) {
1052                 bn = SB_ONDISK_JOURNAL_1st_BLOCK(s) +
1053                     (jl->j_start + i) % SB_ONDISK_JOURNAL_SIZE(s);
1054                 tbh = journal_find_get_block(s, bn);
1055
1056                 depth = reiserfs_write_unlock_nested(s);
1057                 __wait_on_buffer(tbh);
1058                 reiserfs_write_lock_nested(s, depth);
1059                 // since we're using ll_rw_blk above, it might have skipped over
1060                 // a locked buffer.  Double check here
1061                 //
1062                 /* redundant, sync_dirty_buffer() checks */
1063                 if (buffer_dirty(tbh)) {
1064                         depth = reiserfs_write_unlock_nested(s);
1065                         sync_dirty_buffer(tbh);
1066                         reiserfs_write_lock_nested(s, depth);
1067                 }
1068                 if (unlikely(!buffer_uptodate(tbh))) {
1069 #ifdef CONFIG_REISERFS_CHECK
1070                         reiserfs_warning(s, "journal-601",
1071                                          "buffer write failed");
1072 #endif
1073                         retval = -EIO;
1074                 }
1075                 put_bh(tbh);    /* once for journal_find_get_block */
1076                 put_bh(tbh);    /* once due to original getblk in do_journal_end */
1077                 atomic_dec(&(jl->j_commit_left));
1078         }
1079
1080         BUG_ON(atomic_read(&(jl->j_commit_left)) != 1);
1081
1082         /* If there was a write error in the journal - we can't commit
1083          * this transaction - it will be invalid and, if successful,
1084          * will just end up propagating the write error out to
1085          * the file system. */
1086         if (likely(!retval && !reiserfs_is_journal_aborted (journal))) {
1087                 if (buffer_dirty(jl->j_commit_bh))
1088                         BUG();
1089                 mark_buffer_dirty(jl->j_commit_bh) ;
1090                 depth = reiserfs_write_unlock_nested(s);
1091                 if (reiserfs_barrier_flush(s))
1092                         __sync_dirty_buffer(jl->j_commit_bh, WRITE_FLUSH_FUA);
1093                 else
1094                         sync_dirty_buffer(jl->j_commit_bh);
1095                 reiserfs_write_lock_nested(s, depth);
1096         }
1097
1098         /* If there was a write error in the journal - we can't commit this
1099          * transaction - it will be invalid and, if successful, will just end
1100          * up propagating the write error out to the filesystem. */
1101         if (unlikely(!buffer_uptodate(jl->j_commit_bh))) {
1102 #ifdef CONFIG_REISERFS_CHECK
1103                 reiserfs_warning(s, "journal-615", "buffer write failed");
1104 #endif
1105                 retval = -EIO;
1106         }
1107         bforget(jl->j_commit_bh);
1108         if (journal->j_last_commit_id != 0 &&
1109             (jl->j_trans_id - journal->j_last_commit_id) != 1) {
1110                 reiserfs_warning(s, "clm-2200", "last commit %lu, current %lu",
1111                                  journal->j_last_commit_id, jl->j_trans_id);
1112         }
1113         journal->j_last_commit_id = jl->j_trans_id;
1114
1115         /* now, every commit block is on the disk.  It is safe to allow blocks freed during this transaction to be reallocated */
1116         cleanup_freed_for_journal_list(s, jl);
1117
1118         retval = retval ? retval : journal->j_errno;
1119
1120         /* mark the metadata dirty */
1121         if (!retval)
1122                 dirty_one_transaction(s, jl);
1123         atomic_dec(&(jl->j_commit_left));
1124
1125         if (flushall) {
1126                 atomic_set(&(jl->j_older_commits_done), 1);
1127         }
1128         mutex_unlock(&jl->j_commit_mutex);
1129       put_jl:
1130         put_journal_list(s, jl);
1131
1132         if (retval)
1133                 reiserfs_abort(s, retval, "Journal write error in %s",
1134                                __func__);
1135         return retval;
1136 }
1137
1138 /*
1139 ** flush_journal_list frequently needs to find a newer transaction for a given block.  This does that, or
1140 ** returns NULL if it can't find anything
1141 */
1142 static struct reiserfs_journal_list *find_newer_jl_for_cn(struct
1143                                                           reiserfs_journal_cnode
1144                                                           *cn)
1145 {
1146         struct super_block *sb = cn->sb;
1147         b_blocknr_t blocknr = cn->blocknr;
1148
1149         cn = cn->hprev;
1150         while (cn) {
1151                 if (cn->sb == sb && cn->blocknr == blocknr && cn->jlist) {
1152                         return cn->jlist;
1153                 }
1154                 cn = cn->hprev;
1155         }
1156         return NULL;
1157 }
1158
1159 static void remove_journal_hash(struct super_block *,
1160                                 struct reiserfs_journal_cnode **,
1161                                 struct reiserfs_journal_list *, unsigned long,
1162                                 int);
1163
1164 /*
1165 ** once all the real blocks have been flushed, it is safe to remove them from the
1166 ** journal list for this transaction.  Aside from freeing the cnode, this also allows the
1167 ** block to be reallocated for data blocks if it had been deleted.
1168 */
1169 static void remove_all_from_journal_list(struct super_block *sb,
1170                                          struct reiserfs_journal_list *jl,
1171                                          int debug)
1172 {
1173         struct reiserfs_journal *journal = SB_JOURNAL(sb);
1174         struct reiserfs_journal_cnode *cn, *last;
1175         cn = jl->j_realblock;
1176
1177         /* which is better, to lock once around the whole loop, or
1178          ** to lock for each call to remove_journal_hash?
1179          */
1180         while (cn) {
1181                 if (cn->blocknr != 0) {
1182                         if (debug) {
1183                                 reiserfs_warning(sb, "reiserfs-2201",
1184                                                  "block %u, bh is %d, state %ld",
1185                                                  cn->blocknr, cn->bh ? 1 : 0,
1186                                                  cn->state);
1187                         }
1188                         cn->state = 0;
1189                         remove_journal_hash(sb, journal->j_list_hash_table,
1190                                             jl, cn->blocknr, 1);
1191                 }
1192                 last = cn;
1193                 cn = cn->next;
1194                 free_cnode(sb, last);
1195         }
1196         jl->j_realblock = NULL;
1197 }
1198
1199 /*
1200 ** if this timestamp is greater than the timestamp we wrote last to the header block, write it to the header block.
1201 ** once this is done, I can safely say the log area for this transaction won't ever be replayed, and I can start
1202 ** releasing blocks in this transaction for reuse as data blocks.
1203 ** called by flush_journal_list, before it calls remove_all_from_journal_list
1204 **
1205 */
1206 static int _update_journal_header_block(struct super_block *sb,
1207                                         unsigned long offset,
1208                                         unsigned int trans_id)
1209 {
1210         struct reiserfs_journal_header *jh;
1211         struct reiserfs_journal *journal = SB_JOURNAL(sb);
1212         int depth;
1213
1214         if (reiserfs_is_journal_aborted(journal))
1215                 return -EIO;
1216
1217         if (trans_id >= journal->j_last_flush_trans_id) {
1218                 if (buffer_locked((journal->j_header_bh))) {
1219                         depth = reiserfs_write_unlock_nested(sb);
1220                         __wait_on_buffer(journal->j_header_bh);
1221                         reiserfs_write_lock_nested(sb, depth);
1222                         if (unlikely(!buffer_uptodate(journal->j_header_bh))) {
1223 #ifdef CONFIG_REISERFS_CHECK
1224                                 reiserfs_warning(sb, "journal-699",
1225                                                  "buffer write failed");
1226 #endif
1227                                 return -EIO;
1228                         }
1229                 }
1230                 journal->j_last_flush_trans_id = trans_id;
1231                 journal->j_first_unflushed_offset = offset;
1232                 jh = (struct reiserfs_journal_header *)(journal->j_header_bh->
1233                                                         b_data);
1234                 jh->j_last_flush_trans_id = cpu_to_le32(trans_id);
1235                 jh->j_first_unflushed_offset = cpu_to_le32(offset);
1236                 jh->j_mount_id = cpu_to_le32(journal->j_mount_id);
1237
1238                 set_buffer_dirty(journal->j_header_bh);
1239                 depth = reiserfs_write_unlock_nested(sb);
1240
1241                 if (reiserfs_barrier_flush(sb))
1242                         __sync_dirty_buffer(journal->j_header_bh, WRITE_FLUSH_FUA);
1243                 else
1244                         sync_dirty_buffer(journal->j_header_bh);
1245
1246                 reiserfs_write_lock_nested(sb, depth);
1247                 if (!buffer_uptodate(journal->j_header_bh)) {
1248                         reiserfs_warning(sb, "journal-837",
1249                                          "IO error during journal replay");
1250                         return -EIO;
1251                 }
1252         }
1253         return 0;
1254 }
1255
1256 static int update_journal_header_block(struct super_block *sb,
1257                                        unsigned long offset,
1258                                        unsigned int trans_id)
1259 {
1260         return _update_journal_header_block(sb, offset, trans_id);
1261 }
1262
1263 /*
1264 ** flush any and all journal lists older than you are
1265 ** can only be called from flush_journal_list
1266 */
1267 static int flush_older_journal_lists(struct super_block *sb,
1268                                      struct reiserfs_journal_list *jl)
1269 {
1270         struct list_head *entry;
1271         struct reiserfs_journal_list *other_jl;
1272         struct reiserfs_journal *journal = SB_JOURNAL(sb);
1273         unsigned int trans_id = jl->j_trans_id;
1274
1275         /* we know we are the only ones flushing things, no extra race
1276          * protection is required.
1277          */
1278       restart:
1279         entry = journal->j_journal_list.next;
1280         /* Did we wrap? */
1281         if (entry == &journal->j_journal_list)
1282                 return 0;
1283         other_jl = JOURNAL_LIST_ENTRY(entry);
1284         if (other_jl->j_trans_id < trans_id) {
1285                 BUG_ON(other_jl->j_refcount <= 0);
1286                 /* do not flush all */
1287                 flush_journal_list(sb, other_jl, 0);
1288
1289                 /* other_jl is now deleted from the list */
1290                 goto restart;
1291         }
1292         return 0;
1293 }
1294
1295 static void del_from_work_list(struct super_block *s,
1296                                struct reiserfs_journal_list *jl)
1297 {
1298         struct reiserfs_journal *journal = SB_JOURNAL(s);
1299         if (!list_empty(&jl->j_working_list)) {
1300                 list_del_init(&jl->j_working_list);
1301                 journal->j_num_work_lists--;
1302         }
1303 }
1304
1305 /* flush a journal list, both commit and real blocks
1306 **
1307 ** always set flushall to 1, unless you are calling from inside
1308 ** flush_journal_list
1309 **
1310 ** IMPORTANT.  This can only be called while there are no journal writers,
1311 ** and the journal is locked.  That means it can only be called from
1312 ** do_journal_end, or by journal_release
1313 */
1314 static int flush_journal_list(struct super_block *s,
1315                               struct reiserfs_journal_list *jl, int flushall)
1316 {
1317         struct reiserfs_journal_list *pjl;
1318         struct reiserfs_journal_cnode *cn, *last;
1319         int count;
1320         int was_jwait = 0;
1321         int was_dirty = 0;
1322         struct buffer_head *saved_bh;
1323         unsigned long j_len_saved = jl->j_len;
1324         struct reiserfs_journal *journal = SB_JOURNAL(s);
1325         int err = 0;
1326         int depth;
1327
1328         BUG_ON(j_len_saved <= 0);
1329
1330         if (atomic_read(&journal->j_wcount) != 0) {
1331                 reiserfs_warning(s, "clm-2048", "called with wcount %d",
1332                                  atomic_read(&journal->j_wcount));
1333         }
1334
1335         /* if flushall == 0, the lock is already held */
1336         if (flushall) {
1337                 reiserfs_mutex_lock_safe(&journal->j_flush_mutex, s);
1338         } else if (mutex_trylock(&journal->j_flush_mutex)) {
1339                 BUG();
1340         }
1341
1342         count = 0;
1343         if (j_len_saved > journal->j_trans_max) {
1344                 reiserfs_panic(s, "journal-715", "length is %lu, trans id %lu",
1345                                j_len_saved, jl->j_trans_id);
1346                 return 0;
1347         }
1348
1349         /* if all the work is already done, get out of here */
1350         if (atomic_read(&(jl->j_nonzerolen)) <= 0 &&
1351             atomic_read(&(jl->j_commit_left)) <= 0) {
1352                 goto flush_older_and_return;
1353         }
1354
1355         /* start by putting the commit list on disk.  This will also flush
1356          ** the commit lists of any olders transactions
1357          */
1358         flush_commit_list(s, jl, 1);
1359
1360         if (!(jl->j_state & LIST_DIRTY)
1361             && !reiserfs_is_journal_aborted(journal))
1362                 BUG();
1363
1364         /* are we done now? */
1365         if (atomic_read(&(jl->j_nonzerolen)) <= 0 &&
1366             atomic_read(&(jl->j_commit_left)) <= 0) {
1367                 goto flush_older_and_return;
1368         }
1369
1370         /* loop through each cnode, see if we need to write it,
1371          ** or wait on a more recent transaction, or just ignore it
1372          */
1373         if (atomic_read(&(journal->j_wcount)) != 0) {
1374                 reiserfs_panic(s, "journal-844", "journal list is flushing, "
1375                                "wcount is not 0");
1376         }
1377         cn = jl->j_realblock;
1378         while (cn) {
1379                 was_jwait = 0;
1380                 was_dirty = 0;
1381                 saved_bh = NULL;
1382                 /* blocknr of 0 is no longer in the hash, ignore it */
1383                 if (cn->blocknr == 0) {
1384                         goto free_cnode;
1385                 }
1386
1387                 /* This transaction failed commit. Don't write out to the disk */
1388                 if (!(jl->j_state & LIST_DIRTY))
1389                         goto free_cnode;
1390
1391                 pjl = find_newer_jl_for_cn(cn);
1392                 /* the order is important here.  We check pjl to make sure we
1393                  ** don't clear BH_JDirty_wait if we aren't the one writing this
1394                  ** block to disk
1395                  */
1396                 if (!pjl && cn->bh) {
1397                         saved_bh = cn->bh;
1398
1399                         /* we do this to make sure nobody releases the buffer while
1400                          ** we are working with it
1401                          */
1402                         get_bh(saved_bh);
1403
1404                         if (buffer_journal_dirty(saved_bh)) {
1405                                 BUG_ON(!can_dirty(cn));
1406                                 was_jwait = 1;
1407                                 was_dirty = 1;
1408                         } else if (can_dirty(cn)) {
1409                                 /* everything with !pjl && jwait should be writable */
1410                                 BUG();
1411                         }
1412                 }
1413
1414                 /* if someone has this block in a newer transaction, just make
1415                  ** sure they are committed, and don't try writing it to disk
1416                  */
1417                 if (pjl) {
1418                         if (atomic_read(&pjl->j_commit_left))
1419                                 flush_commit_list(s, pjl, 1);
1420                         goto free_cnode;
1421                 }
1422
1423                 /* bh == NULL when the block got to disk on its own, OR,
1424                  ** the block got freed in a future transaction
1425                  */
1426                 if (saved_bh == NULL) {
1427                         goto free_cnode;
1428                 }
1429
1430                 /* this should never happen.  kupdate_one_transaction has this list
1431                  ** locked while it works, so we should never see a buffer here that
1432                  ** is not marked JDirty_wait
1433                  */
1434                 if ((!was_jwait) && !buffer_locked(saved_bh)) {
1435                         reiserfs_warning(s, "journal-813",
1436                                          "BAD! buffer %llu %cdirty %cjwait, "
1437                                          "not in a newer tranasction",
1438                                          (unsigned long long)saved_bh->
1439                                          b_blocknr, was_dirty ? ' ' : '!',
1440                                          was_jwait ? ' ' : '!');
1441                 }
1442                 if (was_dirty) {
1443                         /* we inc again because saved_bh gets decremented at free_cnode */
1444                         get_bh(saved_bh);
1445                         set_bit(BLOCK_NEEDS_FLUSH, &cn->state);
1446                         lock_buffer(saved_bh);
1447                         BUG_ON(cn->blocknr != saved_bh->b_blocknr);
1448                         if (buffer_dirty(saved_bh))
1449                                 submit_logged_buffer(saved_bh);
1450                         else
1451                                 unlock_buffer(saved_bh);
1452                         count++;
1453                 } else {
1454                         reiserfs_warning(s, "clm-2082",
1455                                          "Unable to flush buffer %llu in %s",
1456                                          (unsigned long long)saved_bh->
1457                                          b_blocknr, __func__);
1458                 }
1459               free_cnode:
1460                 last = cn;
1461                 cn = cn->next;
1462                 if (saved_bh) {
1463                         /* we incremented this to keep others from taking the buffer head away */
1464                         put_bh(saved_bh);
1465                         if (atomic_read(&(saved_bh->b_count)) < 0) {
1466                                 reiserfs_warning(s, "journal-945",
1467                                                  "saved_bh->b_count < 0");
1468                         }
1469                 }
1470         }
1471         if (count > 0) {
1472                 cn = jl->j_realblock;
1473                 while (cn) {
1474                         if (test_bit(BLOCK_NEEDS_FLUSH, &cn->state)) {
1475                                 if (!cn->bh) {
1476                                         reiserfs_panic(s, "journal-1011",
1477                                                        "cn->bh is NULL");
1478                                 }
1479
1480                                 depth = reiserfs_write_unlock_nested(s);
1481                                 __wait_on_buffer(cn->bh);
1482                                 reiserfs_write_lock_nested(s, depth);
1483
1484                                 if (!cn->bh) {
1485                                         reiserfs_panic(s, "journal-1012",
1486                                                        "cn->bh is NULL");
1487                                 }
1488                                 if (unlikely(!buffer_uptodate(cn->bh))) {
1489 #ifdef CONFIG_REISERFS_CHECK
1490                                         reiserfs_warning(s, "journal-949",
1491                                                          "buffer write failed");
1492 #endif
1493                                         err = -EIO;
1494                                 }
1495                                 /* note, we must clear the JDirty_wait bit after the up to date
1496                                  ** check, otherwise we race against our flushpage routine
1497                                  */
1498                                 BUG_ON(!test_clear_buffer_journal_dirty
1499                                        (cn->bh));
1500
1501                                 /* drop one ref for us */
1502                                 put_bh(cn->bh);
1503                                 /* drop one ref for journal_mark_dirty */
1504                                 release_buffer_page(cn->bh);
1505                         }
1506                         cn = cn->next;
1507                 }
1508         }
1509
1510         if (err)
1511                 reiserfs_abort(s, -EIO,
1512                                "Write error while pushing transaction to disk in %s",
1513                                __func__);
1514       flush_older_and_return:
1515
1516         /* before we can update the journal header block, we _must_ flush all
1517          ** real blocks from all older transactions to disk.  This is because
1518          ** once the header block is updated, this transaction will not be
1519          ** replayed after a crash
1520          */
1521         if (flushall) {
1522                 flush_older_journal_lists(s, jl);
1523         }
1524
1525         err = journal->j_errno;
1526         /* before we can remove everything from the hash tables for this
1527          ** transaction, we must make sure it can never be replayed
1528          **
1529          ** since we are only called from do_journal_end, we know for sure there
1530          ** are no allocations going on while we are flushing journal lists.  So,
1531          ** we only need to update the journal header block for the last list
1532          ** being flushed
1533          */
1534         if (!err && flushall) {
1535                 err =
1536                     update_journal_header_block(s,
1537                                                 (jl->j_start + jl->j_len +
1538                                                  2) % SB_ONDISK_JOURNAL_SIZE(s),
1539                                                 jl->j_trans_id);
1540                 if (err)
1541                         reiserfs_abort(s, -EIO,
1542                                        "Write error while updating journal header in %s",
1543                                        __func__);
1544         }
1545         remove_all_from_journal_list(s, jl, 0);
1546         list_del_init(&jl->j_list);
1547         journal->j_num_lists--;
1548         del_from_work_list(s, jl);
1549
1550         if (journal->j_last_flush_id != 0 &&
1551             (jl->j_trans_id - journal->j_last_flush_id) != 1) {
1552                 reiserfs_warning(s, "clm-2201", "last flush %lu, current %lu",
1553                                  journal->j_last_flush_id, jl->j_trans_id);
1554         }
1555         journal->j_last_flush_id = jl->j_trans_id;
1556
1557         /* not strictly required since we are freeing the list, but it should
1558          * help find code using dead lists later on
1559          */
1560         jl->j_len = 0;
1561         atomic_set(&(jl->j_nonzerolen), 0);
1562         jl->j_start = 0;
1563         jl->j_realblock = NULL;
1564         jl->j_commit_bh = NULL;
1565         jl->j_trans_id = 0;
1566         jl->j_state = 0;
1567         put_journal_list(s, jl);
1568         if (flushall)
1569                 mutex_unlock(&journal->j_flush_mutex);
1570         return err;
1571 }
1572
1573 static int write_one_transaction(struct super_block *s,
1574                                  struct reiserfs_journal_list *jl,
1575                                  struct buffer_chunk *chunk)
1576 {
1577         struct reiserfs_journal_cnode *cn;
1578         int ret = 0;
1579
1580         jl->j_state |= LIST_TOUCHED;
1581         del_from_work_list(s, jl);
1582         if (jl->j_len == 0 || atomic_read(&jl->j_nonzerolen) == 0) {
1583                 return 0;
1584         }
1585
1586         cn = jl->j_realblock;
1587         while (cn) {
1588                 /* if the blocknr == 0, this has been cleared from the hash,
1589                  ** skip it
1590                  */
1591                 if (cn->blocknr == 0) {
1592                         goto next;
1593                 }
1594                 if (cn->bh && can_dirty(cn) && buffer_dirty(cn->bh)) {
1595                         struct buffer_head *tmp_bh;
1596                         /* we can race against journal_mark_freed when we try
1597                          * to lock_buffer(cn->bh), so we have to inc the buffer
1598                          * count, and recheck things after locking
1599                          */
1600                         tmp_bh = cn->bh;
1601                         get_bh(tmp_bh);
1602                         lock_buffer(tmp_bh);
1603                         if (cn->bh && can_dirty(cn) && buffer_dirty(tmp_bh)) {
1604                                 if (!buffer_journal_dirty(tmp_bh) ||
1605                                     buffer_journal_prepared(tmp_bh))
1606                                         BUG();
1607                                 add_to_chunk(chunk, tmp_bh, NULL, write_chunk);
1608                                 ret++;
1609                         } else {
1610                                 /* note, cn->bh might be null now */
1611                                 unlock_buffer(tmp_bh);
1612                         }
1613                         put_bh(tmp_bh);
1614                 }
1615               next:
1616                 cn = cn->next;
1617                 cond_resched();
1618         }
1619         return ret;
1620 }
1621
1622 /* used by flush_commit_list */
1623 static int dirty_one_transaction(struct super_block *s,
1624                                  struct reiserfs_journal_list *jl)
1625 {
1626         struct reiserfs_journal_cnode *cn;
1627         struct reiserfs_journal_list *pjl;
1628         int ret = 0;
1629
1630         jl->j_state |= LIST_DIRTY;
1631         cn = jl->j_realblock;
1632         while (cn) {
1633                 /* look for a more recent transaction that logged this
1634                  ** buffer.  Only the most recent transaction with a buffer in
1635                  ** it is allowed to send that buffer to disk
1636                  */
1637                 pjl = find_newer_jl_for_cn(cn);
1638                 if (!pjl && cn->blocknr && cn->bh
1639                     && buffer_journal_dirty(cn->bh)) {
1640                         BUG_ON(!can_dirty(cn));
1641                         /* if the buffer is prepared, it will either be logged
1642                          * or restored.  If restored, we need to make sure
1643                          * it actually gets marked dirty
1644                          */
1645                         clear_buffer_journal_new(cn->bh);
1646                         if (buffer_journal_prepared(cn->bh)) {
1647                                 set_buffer_journal_restore_dirty(cn->bh);
1648                         } else {
1649                                 set_buffer_journal_test(cn->bh);
1650                                 mark_buffer_dirty(cn->bh);
1651                         }
1652                 }
1653                 cn = cn->next;
1654         }
1655         return ret;
1656 }
1657
1658 static int kupdate_transactions(struct super_block *s,
1659                                 struct reiserfs_journal_list *jl,
1660                                 struct reiserfs_journal_list **next_jl,
1661                                 unsigned int *next_trans_id,
1662                                 int num_blocks, int num_trans)
1663 {
1664         int ret = 0;
1665         int written = 0;
1666         int transactions_flushed = 0;
1667         unsigned int orig_trans_id = jl->j_trans_id;
1668         struct buffer_chunk chunk;
1669         struct list_head *entry;
1670         struct reiserfs_journal *journal = SB_JOURNAL(s);
1671         chunk.nr = 0;
1672
1673         reiserfs_mutex_lock_safe(&journal->j_flush_mutex, s);
1674         if (!journal_list_still_alive(s, orig_trans_id)) {
1675                 goto done;
1676         }
1677
1678         /* we've got j_flush_mutex held, nobody is going to delete any
1679          * of these lists out from underneath us
1680          */
1681         while ((num_trans && transactions_flushed < num_trans) ||
1682                (!num_trans && written < num_blocks)) {
1683
1684                 if (jl->j_len == 0 || (jl->j_state & LIST_TOUCHED) ||
1685                     atomic_read(&jl->j_commit_left)
1686                     || !(jl->j_state & LIST_DIRTY)) {
1687                         del_from_work_list(s, jl);
1688                         break;
1689                 }
1690                 ret = write_one_transaction(s, jl, &chunk);
1691
1692                 if (ret < 0)
1693                         goto done;
1694                 transactions_flushed++;
1695                 written += ret;
1696                 entry = jl->j_list.next;
1697
1698                 /* did we wrap? */
1699                 if (entry == &journal->j_journal_list) {
1700                         break;
1701                 }
1702                 jl = JOURNAL_LIST_ENTRY(entry);
1703
1704                 /* don't bother with older transactions */
1705                 if (jl->j_trans_id <= orig_trans_id)
1706                         break;
1707         }
1708         if (chunk.nr) {
1709                 write_chunk(&chunk);
1710         }
1711
1712       done:
1713         mutex_unlock(&journal->j_flush_mutex);
1714         return ret;
1715 }
1716
1717 /* for o_sync and fsync heavy applications, they tend to use
1718 ** all the journa list slots with tiny transactions.  These
1719 ** trigger lots and lots of calls to update the header block, which
1720 ** adds seeks and slows things down.
1721 **
1722 ** This function tries to clear out a large chunk of the journal lists
1723 ** at once, which makes everything faster since only the newest journal
1724 ** list updates the header block
1725 */
1726 static int flush_used_journal_lists(struct super_block *s,
1727                                     struct reiserfs_journal_list *jl)
1728 {
1729         unsigned long len = 0;
1730         unsigned long cur_len;
1731         int ret;
1732         int i;
1733         int limit = 256;
1734         struct reiserfs_journal_list *tjl;
1735         struct reiserfs_journal_list *flush_jl;
1736         unsigned int trans_id;
1737         struct reiserfs_journal *journal = SB_JOURNAL(s);
1738
1739         flush_jl = tjl = jl;
1740
1741         /* in data logging mode, try harder to flush a lot of blocks */
1742         if (reiserfs_data_log(s))
1743                 limit = 1024;
1744         /* flush for 256 transactions or limit blocks, whichever comes first */
1745         for (i = 0; i < 256 && len < limit; i++) {
1746                 if (atomic_read(&tjl->j_commit_left) ||
1747                     tjl->j_trans_id < jl->j_trans_id) {
1748                         break;
1749                 }
1750                 cur_len = atomic_read(&tjl->j_nonzerolen);
1751                 if (cur_len > 0) {
1752                         tjl->j_state &= ~LIST_TOUCHED;
1753                 }
1754                 len += cur_len;
1755                 flush_jl = tjl;
1756                 if (tjl->j_list.next == &journal->j_journal_list)
1757                         break;
1758                 tjl = JOURNAL_LIST_ENTRY(tjl->j_list.next);
1759         }
1760         get_journal_list(jl);
1761         get_journal_list(flush_jl);
1762         /* try to find a group of blocks we can flush across all the
1763          ** transactions, but only bother if we've actually spanned
1764          ** across multiple lists
1765          */
1766         if (flush_jl != jl) {
1767                 ret = kupdate_transactions(s, jl, &tjl, &trans_id, len, i);
1768         }
1769         flush_journal_list(s, flush_jl, 1);
1770         put_journal_list(s, flush_jl);
1771         put_journal_list(s, jl);
1772         return 0;
1773 }
1774
1775 /*
1776 ** removes any nodes in table with name block and dev as bh.
1777 ** only touchs the hnext and hprev pointers.
1778 */
1779 void remove_journal_hash(struct super_block *sb,
1780                          struct reiserfs_journal_cnode **table,
1781                          struct reiserfs_journal_list *jl,
1782                          unsigned long block, int remove_freed)
1783 {
1784         struct reiserfs_journal_cnode *cur;
1785         struct reiserfs_journal_cnode **head;
1786
1787         head = &(journal_hash(table, sb, block));
1788         if (!head) {
1789                 return;
1790         }
1791         cur = *head;
1792         while (cur) {
1793                 if (cur->blocknr == block && cur->sb == sb
1794                     && (jl == NULL || jl == cur->jlist)
1795                     && (!test_bit(BLOCK_FREED, &cur->state) || remove_freed)) {
1796                         if (cur->hnext) {
1797                                 cur->hnext->hprev = cur->hprev;
1798                         }
1799                         if (cur->hprev) {
1800                                 cur->hprev->hnext = cur->hnext;
1801                         } else {
1802                                 *head = cur->hnext;
1803                         }
1804                         cur->blocknr = 0;
1805                         cur->sb = NULL;
1806                         cur->state = 0;
1807                         if (cur->bh && cur->jlist)      /* anybody who clears the cur->bh will also dec the nonzerolen */
1808                                 atomic_dec(&(cur->jlist->j_nonzerolen));
1809                         cur->bh = NULL;
1810                         cur->jlist = NULL;
1811                 }
1812                 cur = cur->hnext;
1813         }
1814 }
1815
1816 static void free_journal_ram(struct super_block *sb)
1817 {
1818         struct reiserfs_journal *journal = SB_JOURNAL(sb);
1819         kfree(journal->j_current_jl);
1820         journal->j_num_lists--;
1821
1822         vfree(journal->j_cnode_free_orig);
1823         free_list_bitmaps(sb, journal->j_list_bitmap);
1824         free_bitmap_nodes(sb);  /* must be after free_list_bitmaps */
1825         if (journal->j_header_bh) {
1826                 brelse(journal->j_header_bh);
1827         }
1828         /* j_header_bh is on the journal dev, make sure not to release the journal
1829          * dev until we brelse j_header_bh
1830          */
1831         release_journal_dev(sb, journal);
1832         vfree(journal);
1833 }
1834
1835 /*
1836 ** call on unmount.  Only set error to 1 if you haven't made your way out
1837 ** of read_super() yet.  Any other caller must keep error at 0.
1838 */
1839 static int do_journal_release(struct reiserfs_transaction_handle *th,
1840                               struct super_block *sb, int error)
1841 {
1842         struct reiserfs_transaction_handle myth;
1843         int flushed = 0;
1844         struct reiserfs_journal *journal = SB_JOURNAL(sb);
1845
1846         /* we only want to flush out transactions if we were called with error == 0
1847          */
1848         if (!error && !(sb->s_flags & MS_RDONLY)) {
1849                 /* end the current trans */
1850                 BUG_ON(!th->t_trans_id);
1851                 do_journal_end(th, sb, 10, FLUSH_ALL);
1852
1853                 /* make sure something gets logged to force our way into the flush code */
1854                 if (!journal_join(&myth, sb, 1)) {
1855                         reiserfs_prepare_for_journal(sb,
1856                                                      SB_BUFFER_WITH_SB(sb),
1857                                                      1);
1858                         journal_mark_dirty(&myth, sb,
1859                                            SB_BUFFER_WITH_SB(sb));
1860                         do_journal_end(&myth, sb, 1, FLUSH_ALL);
1861                         flushed = 1;
1862                 }
1863         }
1864
1865         /* this also catches errors during the do_journal_end above */
1866         if (!error && reiserfs_is_journal_aborted(journal)) {
1867                 memset(&myth, 0, sizeof(myth));
1868                 if (!journal_join_abort(&myth, sb, 1)) {
1869                         reiserfs_prepare_for_journal(sb,
1870                                                      SB_BUFFER_WITH_SB(sb),
1871                                                      1);
1872                         journal_mark_dirty(&myth, sb,
1873                                            SB_BUFFER_WITH_SB(sb));
1874                         do_journal_end(&myth, sb, 1, FLUSH_ALL);
1875                 }
1876         }
1877
1878         /* wait for all commits to finish */
1879         cancel_delayed_work(&SB_JOURNAL(sb)->j_work);
1880
1881         /*
1882          * We must release the write lock here because
1883          * the workqueue job (flush_async_commit) needs this lock
1884          */
1885         reiserfs_write_unlock(sb);
1886
1887         cancel_delayed_work_sync(&REISERFS_SB(sb)->old_work);
1888         flush_workqueue(REISERFS_SB(sb)->commit_wq);
1889
1890         free_journal_ram(sb);
1891
1892         reiserfs_write_lock(sb);
1893
1894         return 0;
1895 }
1896
1897 /*
1898 ** call on unmount.  flush all journal trans, release all alloc'd ram
1899 */
1900 int journal_release(struct reiserfs_transaction_handle *th,
1901                     struct super_block *sb)
1902 {
1903         return do_journal_release(th, sb, 0);
1904 }
1905
1906 /*
1907 ** only call from an error condition inside reiserfs_read_super!
1908 */
1909 int journal_release_error(struct reiserfs_transaction_handle *th,
1910                           struct super_block *sb)
1911 {
1912         return do_journal_release(th, sb, 1);
1913 }
1914
1915 /* compares description block with commit block.  returns 1 if they differ, 0 if they are the same */
1916 static int journal_compare_desc_commit(struct super_block *sb,
1917                                        struct reiserfs_journal_desc *desc,
1918                                        struct reiserfs_journal_commit *commit)
1919 {
1920         if (get_commit_trans_id(commit) != get_desc_trans_id(desc) ||
1921             get_commit_trans_len(commit) != get_desc_trans_len(desc) ||
1922             get_commit_trans_len(commit) > SB_JOURNAL(sb)->j_trans_max ||
1923             get_commit_trans_len(commit) <= 0) {
1924                 return 1;
1925         }
1926         return 0;
1927 }
1928
1929 /* returns 0 if it did not find a description block
1930 ** returns -1 if it found a corrupt commit block
1931 ** returns 1 if both desc and commit were valid
1932 ** NOTE: only called during fs mount
1933 */
1934 static int journal_transaction_is_valid(struct super_block *sb,
1935                                         struct buffer_head *d_bh,
1936                                         unsigned int *oldest_invalid_trans_id,
1937                                         unsigned long *newest_mount_id)
1938 {
1939         struct reiserfs_journal_desc *desc;
1940         struct reiserfs_journal_commit *commit;
1941         struct buffer_head *c_bh;
1942         unsigned long offset;
1943
1944         if (!d_bh)
1945                 return 0;
1946
1947         desc = (struct reiserfs_journal_desc *)d_bh->b_data;
1948         if (get_desc_trans_len(desc) > 0
1949             && !memcmp(get_journal_desc_magic(d_bh), JOURNAL_DESC_MAGIC, 8)) {
1950                 if (oldest_invalid_trans_id && *oldest_invalid_trans_id
1951                     && get_desc_trans_id(desc) > *oldest_invalid_trans_id) {
1952                         reiserfs_debug(sb, REISERFS_DEBUG_CODE,
1953                                        "journal-986: transaction "
1954                                        "is valid returning because trans_id %d is greater than "
1955                                        "oldest_invalid %lu",
1956                                        get_desc_trans_id(desc),
1957                                        *oldest_invalid_trans_id);
1958                         return 0;
1959                 }
1960                 if (newest_mount_id
1961                     && *newest_mount_id > get_desc_mount_id(desc)) {
1962                         reiserfs_debug(sb, REISERFS_DEBUG_CODE,
1963                                        "journal-1087: transaction "
1964                                        "is valid returning because mount_id %d is less than "
1965                                        "newest_mount_id %lu",
1966                                        get_desc_mount_id(desc),
1967                                        *newest_mount_id);
1968                         return -1;
1969                 }
1970                 if (get_desc_trans_len(desc) > SB_JOURNAL(sb)->j_trans_max) {
1971                         reiserfs_warning(sb, "journal-2018",
1972                                          "Bad transaction length %d "
1973                                          "encountered, ignoring transaction",
1974                                          get_desc_trans_len(desc));
1975                         return -1;
1976                 }
1977                 offset = d_bh->b_blocknr - SB_ONDISK_JOURNAL_1st_BLOCK(sb);
1978
1979                 /* ok, we have a journal description block, lets see if the transaction was valid */
1980                 c_bh =
1981                     journal_bread(sb,
1982                                   SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
1983                                   ((offset + get_desc_trans_len(desc) +
1984                                     1) % SB_ONDISK_JOURNAL_SIZE(sb)));
1985                 if (!c_bh)
1986                         return 0;
1987                 commit = (struct reiserfs_journal_commit *)c_bh->b_data;
1988                 if (journal_compare_desc_commit(sb, desc, commit)) {
1989                         reiserfs_debug(sb, REISERFS_DEBUG_CODE,
1990                                        "journal_transaction_is_valid, commit offset %ld had bad "
1991                                        "time %d or length %d",
1992                                        c_bh->b_blocknr -
1993                                        SB_ONDISK_JOURNAL_1st_BLOCK(sb),
1994                                        get_commit_trans_id(commit),
1995                                        get_commit_trans_len(commit));
1996                         brelse(c_bh);
1997                         if (oldest_invalid_trans_id) {
1998                                 *oldest_invalid_trans_id =
1999                                     get_desc_trans_id(desc);
2000                                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2001                                                "journal-1004: "
2002                                                "transaction_is_valid setting oldest invalid trans_id "
2003                                                "to %d",
2004                                                get_desc_trans_id(desc));
2005                         }
2006                         return -1;
2007                 }
2008                 brelse(c_bh);
2009                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2010                                "journal-1006: found valid "
2011                                "transaction start offset %llu, len %d id %d",
2012                                d_bh->b_blocknr -
2013                                SB_ONDISK_JOURNAL_1st_BLOCK(sb),
2014                                get_desc_trans_len(desc),
2015                                get_desc_trans_id(desc));
2016                 return 1;
2017         } else {
2018                 return 0;
2019         }
2020 }
2021
2022 static void brelse_array(struct buffer_head **heads, int num)
2023 {
2024         int i;
2025         for (i = 0; i < num; i++) {
2026                 brelse(heads[i]);
2027         }
2028 }
2029
2030 /*
2031 ** given the start, and values for the oldest acceptable transactions,
2032 ** this either reads in a replays a transaction, or returns because the
2033 ** transaction is invalid, or too old.
2034 ** NOTE: only called during fs mount
2035 */
2036 static int journal_read_transaction(struct super_block *sb,
2037                                     unsigned long cur_dblock,
2038                                     unsigned long oldest_start,
2039                                     unsigned int oldest_trans_id,
2040                                     unsigned long newest_mount_id)
2041 {
2042         struct reiserfs_journal *journal = SB_JOURNAL(sb);
2043         struct reiserfs_journal_desc *desc;
2044         struct reiserfs_journal_commit *commit;
2045         unsigned int trans_id = 0;
2046         struct buffer_head *c_bh;
2047         struct buffer_head *d_bh;
2048         struct buffer_head **log_blocks = NULL;
2049         struct buffer_head **real_blocks = NULL;
2050         unsigned int trans_offset;
2051         int i;
2052         int trans_half;
2053
2054         d_bh = journal_bread(sb, cur_dblock);
2055         if (!d_bh)
2056                 return 1;
2057         desc = (struct reiserfs_journal_desc *)d_bh->b_data;
2058         trans_offset = d_bh->b_blocknr - SB_ONDISK_JOURNAL_1st_BLOCK(sb);
2059         reiserfs_debug(sb, REISERFS_DEBUG_CODE, "journal-1037: "
2060                        "journal_read_transaction, offset %llu, len %d mount_id %d",
2061                        d_bh->b_blocknr - SB_ONDISK_JOURNAL_1st_BLOCK(sb),
2062                        get_desc_trans_len(desc), get_desc_mount_id(desc));
2063         if (get_desc_trans_id(desc) < oldest_trans_id) {
2064                 reiserfs_debug(sb, REISERFS_DEBUG_CODE, "journal-1039: "
2065                                "journal_read_trans skipping because %lu is too old",
2066                                cur_dblock -
2067                                SB_ONDISK_JOURNAL_1st_BLOCK(sb));
2068                 brelse(d_bh);
2069                 return 1;
2070         }
2071         if (get_desc_mount_id(desc) != newest_mount_id) {
2072                 reiserfs_debug(sb, REISERFS_DEBUG_CODE, "journal-1146: "
2073                                "journal_read_trans skipping because %d is != "
2074                                "newest_mount_id %lu", get_desc_mount_id(desc),
2075                                newest_mount_id);
2076                 brelse(d_bh);
2077                 return 1;
2078         }
2079         c_bh = journal_bread(sb, SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
2080                              ((trans_offset + get_desc_trans_len(desc) + 1) %
2081                               SB_ONDISK_JOURNAL_SIZE(sb)));
2082         if (!c_bh) {
2083                 brelse(d_bh);
2084                 return 1;
2085         }
2086         commit = (struct reiserfs_journal_commit *)c_bh->b_data;
2087         if (journal_compare_desc_commit(sb, desc, commit)) {
2088                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2089                                "journal_read_transaction, "
2090                                "commit offset %llu had bad time %d or length %d",
2091                                c_bh->b_blocknr -
2092                                SB_ONDISK_JOURNAL_1st_BLOCK(sb),
2093                                get_commit_trans_id(commit),
2094                                get_commit_trans_len(commit));
2095                 brelse(c_bh);
2096                 brelse(d_bh);
2097                 return 1;
2098         }
2099
2100         if (bdev_read_only(sb->s_bdev)) {
2101                 reiserfs_warning(sb, "clm-2076",
2102                                  "device is readonly, unable to replay log");
2103                 brelse(c_bh);
2104                 brelse(d_bh);
2105                 return -EROFS;
2106         }
2107
2108         trans_id = get_desc_trans_id(desc);
2109         /* now we know we've got a good transaction, and it was inside the valid time ranges */
2110         log_blocks = kmalloc(get_desc_trans_len(desc) *
2111                              sizeof(struct buffer_head *), GFP_NOFS);
2112         real_blocks = kmalloc(get_desc_trans_len(desc) *
2113                               sizeof(struct buffer_head *), GFP_NOFS);
2114         if (!log_blocks || !real_blocks) {
2115                 brelse(c_bh);
2116                 brelse(d_bh);
2117                 kfree(log_blocks);
2118                 kfree(real_blocks);
2119                 reiserfs_warning(sb, "journal-1169",
2120                                  "kmalloc failed, unable to mount FS");
2121                 return -1;
2122         }
2123         /* get all the buffer heads */
2124         trans_half = journal_trans_half(sb->s_blocksize);
2125         for (i = 0; i < get_desc_trans_len(desc); i++) {
2126                 log_blocks[i] =
2127                     journal_getblk(sb,
2128                                    SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
2129                                    (trans_offset + 1 +
2130                                     i) % SB_ONDISK_JOURNAL_SIZE(sb));
2131                 if (i < trans_half) {
2132                         real_blocks[i] =
2133                             sb_getblk(sb,
2134                                       le32_to_cpu(desc->j_realblock[i]));
2135                 } else {
2136                         real_blocks[i] =
2137                             sb_getblk(sb,
2138                                       le32_to_cpu(commit->
2139                                                   j_realblock[i - trans_half]));
2140                 }
2141                 if (real_blocks[i]->b_blocknr > SB_BLOCK_COUNT(sb)) {
2142                         reiserfs_warning(sb, "journal-1207",
2143                                          "REPLAY FAILURE fsck required! "
2144                                          "Block to replay is outside of "
2145                                          "filesystem");
2146                         goto abort_replay;
2147                 }
2148                 /* make sure we don't try to replay onto log or reserved area */
2149                 if (is_block_in_log_or_reserved_area
2150                     (sb, real_blocks[i]->b_blocknr)) {
2151                         reiserfs_warning(sb, "journal-1204",
2152                                          "REPLAY FAILURE fsck required! "
2153                                          "Trying to replay onto a log block");
2154                       abort_replay:
2155                         brelse_array(log_blocks, i);
2156                         brelse_array(real_blocks, i);
2157                         brelse(c_bh);
2158                         brelse(d_bh);
2159                         kfree(log_blocks);
2160                         kfree(real_blocks);
2161                         return -1;
2162                 }
2163         }
2164         /* read in the log blocks, memcpy to the corresponding real block */
2165         ll_rw_block(READ, get_desc_trans_len(desc), log_blocks);
2166         for (i = 0; i < get_desc_trans_len(desc); i++) {
2167
2168                 wait_on_buffer(log_blocks[i]);
2169                 if (!buffer_uptodate(log_blocks[i])) {
2170                         reiserfs_warning(sb, "journal-1212",
2171                                          "REPLAY FAILURE fsck required! "
2172                                          "buffer write failed");
2173                         brelse_array(log_blocks + i,
2174                                      get_desc_trans_len(desc) - i);
2175                         brelse_array(real_blocks, get_desc_trans_len(desc));
2176                         brelse(c_bh);
2177                         brelse(d_bh);
2178                         kfree(log_blocks);
2179                         kfree(real_blocks);
2180                         return -1;
2181                 }
2182                 memcpy(real_blocks[i]->b_data, log_blocks[i]->b_data,
2183                        real_blocks[i]->b_size);
2184                 set_buffer_uptodate(real_blocks[i]);
2185                 brelse(log_blocks[i]);
2186         }
2187         /* flush out the real blocks */
2188         for (i = 0; i < get_desc_trans_len(desc); i++) {
2189                 set_buffer_dirty(real_blocks[i]);
2190                 write_dirty_buffer(real_blocks[i], WRITE);
2191         }
2192         for (i = 0; i < get_desc_trans_len(desc); i++) {
2193                 wait_on_buffer(real_blocks[i]);
2194                 if (!buffer_uptodate(real_blocks[i])) {
2195                         reiserfs_warning(sb, "journal-1226",
2196                                          "REPLAY FAILURE, fsck required! "
2197                                          "buffer write failed");
2198                         brelse_array(real_blocks + i,
2199                                      get_desc_trans_len(desc) - i);
2200                         brelse(c_bh);
2201                         brelse(d_bh);
2202                         kfree(log_blocks);
2203                         kfree(real_blocks);
2204                         return -1;
2205                 }
2206                 brelse(real_blocks[i]);
2207         }
2208         cur_dblock =
2209             SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
2210             ((trans_offset + get_desc_trans_len(desc) +
2211               2) % SB_ONDISK_JOURNAL_SIZE(sb));
2212         reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2213                        "journal-1095: setting journal " "start to offset %ld",
2214                        cur_dblock - SB_ONDISK_JOURNAL_1st_BLOCK(sb));
2215
2216         /* init starting values for the first transaction, in case this is the last transaction to be replayed. */
2217         journal->j_start = cur_dblock - SB_ONDISK_JOURNAL_1st_BLOCK(sb);
2218         journal->j_last_flush_trans_id = trans_id;
2219         journal->j_trans_id = trans_id + 1;
2220         /* check for trans_id overflow */
2221         if (journal->j_trans_id == 0)
2222                 journal->j_trans_id = 10;
2223         brelse(c_bh);
2224         brelse(d_bh);
2225         kfree(log_blocks);
2226         kfree(real_blocks);
2227         return 0;
2228 }
2229
2230 /* This function reads blocks starting from block and to max_block of bufsize
2231    size (but no more than BUFNR blocks at a time). This proved to improve
2232    mounting speed on self-rebuilding raid5 arrays at least.
2233    Right now it is only used from journal code. But later we might use it
2234    from other places.
2235    Note: Do not use journal_getblk/sb_getblk functions here! */
2236 static struct buffer_head *reiserfs_breada(struct block_device *dev,
2237                                            b_blocknr_t block, int bufsize,
2238                                            b_blocknr_t max_block)
2239 {
2240         struct buffer_head *bhlist[BUFNR];
2241         unsigned int blocks = BUFNR;
2242         struct buffer_head *bh;
2243         int i, j;
2244
2245         bh = __getblk(dev, block, bufsize);
2246         if (buffer_uptodate(bh))
2247                 return (bh);
2248
2249         if (block + BUFNR > max_block) {
2250                 blocks = max_block - block;
2251         }
2252         bhlist[0] = bh;
2253         j = 1;
2254         for (i = 1; i < blocks; i++) {
2255                 bh = __getblk(dev, block + i, bufsize);
2256                 if (buffer_uptodate(bh)) {
2257                         brelse(bh);
2258                         break;
2259                 } else
2260                         bhlist[j++] = bh;
2261         }
2262         ll_rw_block(READ, j, bhlist);
2263         for (i = 1; i < j; i++)
2264                 brelse(bhlist[i]);
2265         bh = bhlist[0];
2266         wait_on_buffer(bh);
2267         if (buffer_uptodate(bh))
2268                 return bh;
2269         brelse(bh);
2270         return NULL;
2271 }
2272
2273 /*
2274 ** read and replay the log
2275 ** on a clean unmount, the journal header's next unflushed pointer will
2276 ** be to an invalid transaction.  This tests that before finding all the
2277 ** transactions in the log, which makes normal mount times fast.
2278 ** After a crash, this starts with the next unflushed transaction, and
2279 ** replays until it finds one too old, or invalid.
2280 ** On exit, it sets things up so the first transaction will work correctly.
2281 ** NOTE: only called during fs mount
2282 */
2283 static int journal_read(struct super_block *sb)
2284 {
2285         struct reiserfs_journal *journal = SB_JOURNAL(sb);
2286         struct reiserfs_journal_desc *desc;
2287         unsigned int oldest_trans_id = 0;
2288         unsigned int oldest_invalid_trans_id = 0;
2289         time_t start;
2290         unsigned long oldest_start = 0;
2291         unsigned long cur_dblock = 0;
2292         unsigned long newest_mount_id = 9;
2293         struct buffer_head *d_bh;
2294         struct reiserfs_journal_header *jh;
2295         int valid_journal_header = 0;
2296         int replay_count = 0;
2297         int continue_replay = 1;
2298         int ret;
2299         char b[BDEVNAME_SIZE];
2300
2301         cur_dblock = SB_ONDISK_JOURNAL_1st_BLOCK(sb);
2302         reiserfs_info(sb, "checking transaction log (%s)\n",
2303                       bdevname(journal->j_dev_bd, b));
2304         start = get_seconds();
2305
2306         /* step 1, read in the journal header block.  Check the transaction it says
2307          ** is the first unflushed, and if that transaction is not valid,
2308          ** replay is done
2309          */
2310         journal->j_header_bh = journal_bread(sb,
2311                                              SB_ONDISK_JOURNAL_1st_BLOCK(sb)
2312                                              + SB_ONDISK_JOURNAL_SIZE(sb));
2313         if (!journal->j_header_bh) {
2314                 return 1;
2315         }
2316         jh = (struct reiserfs_journal_header *)(journal->j_header_bh->b_data);
2317         if (le32_to_cpu(jh->j_first_unflushed_offset) <
2318             SB_ONDISK_JOURNAL_SIZE(sb)
2319             && le32_to_cpu(jh->j_last_flush_trans_id) > 0) {
2320                 oldest_start =
2321                     SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
2322                     le32_to_cpu(jh->j_first_unflushed_offset);
2323                 oldest_trans_id = le32_to_cpu(jh->j_last_flush_trans_id) + 1;
2324                 newest_mount_id = le32_to_cpu(jh->j_mount_id);
2325                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2326                                "journal-1153: found in "
2327                                "header: first_unflushed_offset %d, last_flushed_trans_id "
2328                                "%lu", le32_to_cpu(jh->j_first_unflushed_offset),
2329                                le32_to_cpu(jh->j_last_flush_trans_id));
2330                 valid_journal_header = 1;
2331
2332                 /* now, we try to read the first unflushed offset.  If it is not valid,
2333                  ** there is nothing more we can do, and it makes no sense to read
2334                  ** through the whole log.
2335                  */
2336                 d_bh =
2337                     journal_bread(sb,
2338                                   SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
2339                                   le32_to_cpu(jh->j_first_unflushed_offset));
2340                 ret = journal_transaction_is_valid(sb, d_bh, NULL, NULL);
2341                 if (!ret) {
2342                         continue_replay = 0;
2343                 }
2344                 brelse(d_bh);
2345                 goto start_log_replay;
2346         }
2347
2348         /* ok, there are transactions that need to be replayed.  start with the first log block, find
2349          ** all the valid transactions, and pick out the oldest.
2350          */
2351         while (continue_replay
2352                && cur_dblock <
2353                (SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
2354                 SB_ONDISK_JOURNAL_SIZE(sb))) {
2355                 /* Note that it is required for blocksize of primary fs device and journal
2356                    device to be the same */
2357                 d_bh =
2358                     reiserfs_breada(journal->j_dev_bd, cur_dblock,
2359                                     sb->s_blocksize,
2360                                     SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
2361                                     SB_ONDISK_JOURNAL_SIZE(sb));
2362                 ret =
2363                     journal_transaction_is_valid(sb, d_bh,
2364                                                  &oldest_invalid_trans_id,
2365                                                  &newest_mount_id);
2366                 if (ret == 1) {
2367                         desc = (struct reiserfs_journal_desc *)d_bh->b_data;
2368                         if (oldest_start == 0) {        /* init all oldest_ values */
2369                                 oldest_trans_id = get_desc_trans_id(desc);
2370                                 oldest_start = d_bh->b_blocknr;
2371                                 newest_mount_id = get_desc_mount_id(desc);
2372                                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2373                                                "journal-1179: Setting "
2374                                                "oldest_start to offset %llu, trans_id %lu",
2375                                                oldest_start -
2376                                                SB_ONDISK_JOURNAL_1st_BLOCK
2377                                                (sb), oldest_trans_id);
2378                         } else if (oldest_trans_id > get_desc_trans_id(desc)) {
2379                                 /* one we just read was older */
2380                                 oldest_trans_id = get_desc_trans_id(desc);
2381                                 oldest_start = d_bh->b_blocknr;
2382                                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2383                                                "journal-1180: Resetting "
2384                                                "oldest_start to offset %lu, trans_id %lu",
2385                                                oldest_start -
2386                                                SB_ONDISK_JOURNAL_1st_BLOCK
2387                                                (sb), oldest_trans_id);
2388                         }
2389                         if (newest_mount_id < get_desc_mount_id(desc)) {
2390                                 newest_mount_id = get_desc_mount_id(desc);
2391                                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2392                                                "journal-1299: Setting "
2393                                                "newest_mount_id to %d",
2394                                                get_desc_mount_id(desc));
2395                         }
2396                         cur_dblock += get_desc_trans_len(desc) + 2;
2397                 } else {
2398                         cur_dblock++;
2399                 }
2400                 brelse(d_bh);
2401         }
2402
2403       start_log_replay:
2404         cur_dblock = oldest_start;
2405         if (oldest_trans_id) {
2406                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2407                                "journal-1206: Starting replay "
2408                                "from offset %llu, trans_id %lu",
2409                                cur_dblock - SB_ONDISK_JOURNAL_1st_BLOCK(sb),
2410                                oldest_trans_id);
2411
2412         }
2413         replay_count = 0;
2414         while (continue_replay && oldest_trans_id > 0) {
2415                 ret =
2416                     journal_read_transaction(sb, cur_dblock, oldest_start,
2417                                              oldest_trans_id, newest_mount_id);
2418                 if (ret < 0) {
2419                         return ret;
2420                 } else if (ret != 0) {
2421                         break;
2422                 }
2423                 cur_dblock =
2424                     SB_ONDISK_JOURNAL_1st_BLOCK(sb) + journal->j_start;
2425                 replay_count++;
2426                 if (cur_dblock == oldest_start)
2427                         break;
2428         }
2429
2430         if (oldest_trans_id == 0) {
2431                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2432                                "journal-1225: No valid " "transactions found");
2433         }
2434         /* j_start does not get set correctly if we don't replay any transactions.
2435          ** if we had a valid journal_header, set j_start to the first unflushed transaction value,
2436          ** copy the trans_id from the header
2437          */
2438         if (valid_journal_header && replay_count == 0) {
2439                 journal->j_start = le32_to_cpu(jh->j_first_unflushed_offset);
2440                 journal->j_trans_id =
2441                     le32_to_cpu(jh->j_last_flush_trans_id) + 1;
2442                 /* check for trans_id overflow */
2443                 if (journal->j_trans_id == 0)
2444                         journal->j_trans_id = 10;
2445                 journal->j_last_flush_trans_id =
2446                     le32_to_cpu(jh->j_last_flush_trans_id);
2447                 journal->j_mount_id = le32_to_cpu(jh->j_mount_id) + 1;
2448         } else {
2449                 journal->j_mount_id = newest_mount_id + 1;
2450         }
2451         reiserfs_debug(sb, REISERFS_DEBUG_CODE, "journal-1299: Setting "
2452                        "newest_mount_id to %lu", journal->j_mount_id);
2453         journal->j_first_unflushed_offset = journal->j_start;
2454         if (replay_count > 0) {
2455                 reiserfs_info(sb,
2456                               "replayed %d transactions in %lu seconds\n",
2457                               replay_count, get_seconds() - start);
2458         }
2459         /* needed to satisfy the locking in _update_journal_header_block */
2460         reiserfs_write_lock(sb);
2461         if (!bdev_read_only(sb->s_bdev) &&
2462             _update_journal_header_block(sb, journal->j_start,
2463                                          journal->j_last_flush_trans_id)) {
2464                 reiserfs_write_unlock(sb);
2465                 /* replay failed, caller must call free_journal_ram and abort
2466                  ** the mount
2467                  */
2468                 return -1;
2469         }
2470         reiserfs_write_unlock(sb);
2471         return 0;
2472 }
2473
2474 static struct reiserfs_journal_list *alloc_journal_list(struct super_block *s)
2475 {
2476         struct reiserfs_journal_list *jl;
2477         jl = kzalloc(sizeof(struct reiserfs_journal_list),
2478                      GFP_NOFS | __GFP_NOFAIL);
2479         INIT_LIST_HEAD(&jl->j_list);
2480         INIT_LIST_HEAD(&jl->j_working_list);
2481         INIT_LIST_HEAD(&jl->j_tail_bh_list);
2482         INIT_LIST_HEAD(&jl->j_bh_list);
2483         mutex_init(&jl->j_commit_mutex);
2484         SB_JOURNAL(s)->j_num_lists++;
2485         get_journal_list(jl);
2486         return jl;
2487 }
2488
2489 static void journal_list_init(struct super_block *sb)
2490 {
2491         SB_JOURNAL(sb)->j_current_jl = alloc_journal_list(sb);
2492 }
2493
2494 static void release_journal_dev(struct super_block *super,
2495                                struct reiserfs_journal *journal)
2496 {
2497         if (journal->j_dev_bd != NULL) {
2498                 blkdev_put(journal->j_dev_bd, journal->j_dev_mode);
2499                 journal->j_dev_bd = NULL;
2500         }
2501 }
2502
2503 static int journal_init_dev(struct super_block *super,
2504                             struct reiserfs_journal *journal,
2505                             const char *jdev_name)
2506 {
2507         int result;
2508         dev_t jdev;
2509         fmode_t blkdev_mode = FMODE_READ | FMODE_WRITE | FMODE_EXCL;
2510         char b[BDEVNAME_SIZE];
2511
2512         result = 0;
2513
2514         journal->j_dev_bd = NULL;
2515         jdev = SB_ONDISK_JOURNAL_DEVICE(super) ?
2516             new_decode_dev(SB_ONDISK_JOURNAL_DEVICE(super)) : super->s_dev;
2517
2518         if (bdev_read_only(super->s_bdev))
2519                 blkdev_mode = FMODE_READ;
2520
2521         /* there is no "jdev" option and journal is on separate device */
2522         if ((!jdev_name || !jdev_name[0])) {
2523                 if (jdev == super->s_dev)
2524                         blkdev_mode &= ~FMODE_EXCL;
2525                 journal->j_dev_bd = blkdev_get_by_dev(jdev, blkdev_mode,
2526                                                       journal);
2527                 journal->j_dev_mode = blkdev_mode;
2528                 if (IS_ERR(journal->j_dev_bd)) {
2529                         result = PTR_ERR(journal->j_dev_bd);
2530                         journal->j_dev_bd = NULL;
2531                         reiserfs_warning(super, "sh-458",
2532                                          "cannot init journal device '%s': %i",
2533                                          __bdevname(jdev, b), result);
2534                         return result;
2535                 } else if (jdev != super->s_dev)
2536                         set_blocksize(journal->j_dev_bd, super->s_blocksize);
2537
2538                 return 0;
2539         }
2540
2541         journal->j_dev_mode = blkdev_mode;
2542         journal->j_dev_bd = blkdev_get_by_path(jdev_name, blkdev_mode, journal);
2543         if (IS_ERR(journal->j_dev_bd)) {
2544                 result = PTR_ERR(journal->j_dev_bd);
2545                 journal->j_dev_bd = NULL;
2546                 reiserfs_warning(super,
2547                                  "journal_init_dev: Cannot open '%s': %i",
2548                                  jdev_name, result);
2549                 return result;
2550         }
2551
2552         set_blocksize(journal->j_dev_bd, super->s_blocksize);
2553         reiserfs_info(super,
2554                       "journal_init_dev: journal device: %s\n",
2555                       bdevname(journal->j_dev_bd, b));
2556         return 0;
2557 }
2558
2559 /**
2560  * When creating/tuning a file system user can assign some
2561  * journal params within boundaries which depend on the ratio
2562  * blocksize/standard_blocksize.
2563  *
2564  * For blocks >= standard_blocksize transaction size should
2565  * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more
2566  * then JOURNAL_TRANS_MAX_DEFAULT.
2567  *
2568  * For blocks < standard_blocksize these boundaries should be
2569  * decreased proportionally.
2570  */
2571 #define REISERFS_STANDARD_BLKSIZE (4096)
2572
2573 static int check_advise_trans_params(struct super_block *sb,
2574                                      struct reiserfs_journal *journal)
2575 {
2576         if (journal->j_trans_max) {
2577                 /* Non-default journal params.
2578                    Do sanity check for them. */
2579                 int ratio = 1;
2580                 if (sb->s_blocksize < REISERFS_STANDARD_BLKSIZE)
2581                         ratio = REISERFS_STANDARD_BLKSIZE / sb->s_blocksize;
2582
2583                 if (journal->j_trans_max > JOURNAL_TRANS_MAX_DEFAULT / ratio ||
2584                     journal->j_trans_max < JOURNAL_TRANS_MIN_DEFAULT / ratio ||
2585                     SB_ONDISK_JOURNAL_SIZE(sb) / journal->j_trans_max <
2586                     JOURNAL_MIN_RATIO) {
2587                         reiserfs_warning(sb, "sh-462",
2588                                          "bad transaction max size (%u). "
2589                                          "FSCK?", journal->j_trans_max);
2590                         return 1;
2591                 }
2592                 if (journal->j_max_batch != (journal->j_trans_max) *
2593                         JOURNAL_MAX_BATCH_DEFAULT/JOURNAL_TRANS_MAX_DEFAULT) {
2594                         reiserfs_warning(sb, "sh-463",
2595                                          "bad transaction max batch (%u). "
2596                                          "FSCK?", journal->j_max_batch);
2597                         return 1;
2598                 }
2599         } else {
2600                 /* Default journal params.
2601                    The file system was created by old version
2602                    of mkreiserfs, so some fields contain zeros,
2603                    and we need to advise proper values for them */
2604                 if (sb->s_blocksize != REISERFS_STANDARD_BLKSIZE) {
2605                         reiserfs_warning(sb, "sh-464", "bad blocksize (%u)",
2606                                          sb->s_blocksize);
2607                         return 1;
2608                 }
2609                 journal->j_trans_max = JOURNAL_TRANS_MAX_DEFAULT;
2610                 journal->j_max_batch = JOURNAL_MAX_BATCH_DEFAULT;
2611                 journal->j_max_commit_age = JOURNAL_MAX_COMMIT_AGE;
2612         }
2613         return 0;
2614 }
2615
2616 /*
2617 ** must be called once on fs mount.  calls journal_read for you
2618 */
2619 int journal_init(struct super_block *sb, const char *j_dev_name,
2620                  int old_format, unsigned int commit_max_age)
2621 {
2622         int num_cnodes = SB_ONDISK_JOURNAL_SIZE(sb) * 2;
2623         struct buffer_head *bhjh;
2624         struct reiserfs_super_block *rs;
2625         struct reiserfs_journal_header *jh;
2626         struct reiserfs_journal *journal;
2627         struct reiserfs_journal_list *jl;
2628         char b[BDEVNAME_SIZE];
2629         int ret;
2630
2631         journal = SB_JOURNAL(sb) = vzalloc(sizeof(struct reiserfs_journal));
2632         if (!journal) {
2633                 reiserfs_warning(sb, "journal-1256",
2634                                  "unable to get memory for journal structure");
2635                 return 1;
2636         }
2637         INIT_LIST_HEAD(&journal->j_bitmap_nodes);
2638         INIT_LIST_HEAD(&journal->j_prealloc_list);
2639         INIT_LIST_HEAD(&journal->j_working_list);
2640         INIT_LIST_HEAD(&journal->j_journal_list);
2641         journal->j_persistent_trans = 0;
2642         if (reiserfs_allocate_list_bitmaps(sb, journal->j_list_bitmap,
2643                                            reiserfs_bmap_count(sb)))
2644                 goto free_and_return;
2645
2646         allocate_bitmap_nodes(sb);
2647
2648         /* reserved for journal area support */
2649         SB_JOURNAL_1st_RESERVED_BLOCK(sb) = (old_format ?
2650                                                  REISERFS_OLD_DISK_OFFSET_IN_BYTES
2651                                                  / sb->s_blocksize +
2652                                                  reiserfs_bmap_count(sb) +
2653                                                  1 :
2654                                                  REISERFS_DISK_OFFSET_IN_BYTES /
2655                                                  sb->s_blocksize + 2);
2656
2657         /* Sanity check to see is the standard journal fitting within first bitmap
2658            (actual for small blocksizes) */
2659         if (!SB_ONDISK_JOURNAL_DEVICE(sb) &&
2660             (SB_JOURNAL_1st_RESERVED_BLOCK(sb) +
2661              SB_ONDISK_JOURNAL_SIZE(sb) > sb->s_blocksize * 8)) {
2662                 reiserfs_warning(sb, "journal-1393",
2663                                  "journal does not fit for area addressed "
2664                                  "by first of bitmap blocks. It starts at "
2665                                  "%u and its size is %u. Block size %ld",
2666                                  SB_JOURNAL_1st_RESERVED_BLOCK(sb),
2667                                  SB_ONDISK_JOURNAL_SIZE(sb),
2668                                  sb->s_blocksize);
2669                 goto free_and_return;
2670         }
2671
2672         if (journal_init_dev(sb, journal, j_dev_name) != 0) {
2673                 reiserfs_warning(sb, "sh-462",
2674                                  "unable to initialize jornal device");
2675                 goto free_and_return;
2676         }
2677
2678         rs = SB_DISK_SUPER_BLOCK(sb);
2679
2680         /* read journal header */
2681         bhjh = journal_bread(sb,
2682                              SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
2683                              SB_ONDISK_JOURNAL_SIZE(sb));
2684         if (!bhjh) {
2685                 reiserfs_warning(sb, "sh-459",
2686                                  "unable to read journal header");
2687                 goto free_and_return;
2688         }
2689         jh = (struct reiserfs_journal_header *)(bhjh->b_data);
2690
2691         /* make sure that journal matches to the super block */
2692         if (is_reiserfs_jr(rs)
2693             && (le32_to_cpu(jh->jh_journal.jp_journal_magic) !=
2694                 sb_jp_journal_magic(rs))) {
2695                 reiserfs_warning(sb, "sh-460",
2696                                  "journal header magic %x (device %s) does "
2697                                  "not match to magic found in super block %x",
2698                                  jh->jh_journal.jp_journal_magic,
2699                                  bdevname(journal->j_dev_bd, b),
2700                                  sb_jp_journal_magic(rs));
2701                 brelse(bhjh);
2702                 goto free_and_return;
2703         }
2704
2705         journal->j_trans_max = le32_to_cpu(jh->jh_journal.jp_journal_trans_max);
2706         journal->j_max_batch = le32_to_cpu(jh->jh_journal.jp_journal_max_batch);
2707         journal->j_max_commit_age =
2708             le32_to_cpu(jh->jh_journal.jp_journal_max_commit_age);
2709         journal->j_max_trans_age = JOURNAL_MAX_TRANS_AGE;
2710
2711         if (check_advise_trans_params(sb, journal) != 0)
2712                 goto free_and_return;
2713         journal->j_default_max_commit_age = journal->j_max_commit_age;
2714
2715         if (commit_max_age != 0) {
2716                 journal->j_max_commit_age = commit_max_age;
2717                 journal->j_max_trans_age = commit_max_age;
2718         }
2719
2720         reiserfs_info(sb, "journal params: device %s, size %u, "
2721                       "journal first block %u, max trans len %u, max batch %u, "
2722                       "max commit age %u, max trans age %u\n",
2723                       bdevname(journal->j_dev_bd, b),
2724                       SB_ONDISK_JOURNAL_SIZE(sb),
2725                       SB_ONDISK_JOURNAL_1st_BLOCK(sb),
2726                       journal->j_trans_max,
2727                       journal->j_max_batch,
2728                       journal->j_max_commit_age, journal->j_max_trans_age);
2729
2730         brelse(bhjh);
2731
2732         journal->j_list_bitmap_index = 0;
2733         journal_list_init(sb);
2734
2735         memset(journal->j_list_hash_table, 0,
2736                JOURNAL_HASH_SIZE * sizeof(struct reiserfs_journal_cnode *));
2737
2738         INIT_LIST_HEAD(&journal->j_dirty_buffers);
2739         spin_lock_init(&journal->j_dirty_buffers_lock);
2740
2741         journal->j_start = 0;
2742         journal->j_len = 0;
2743         journal->j_len_alloc = 0;
2744         atomic_set(&(journal->j_wcount), 0);
2745         atomic_set(&(journal->j_async_throttle), 0);
2746         journal->j_bcount = 0;
2747         journal->j_trans_start_time = 0;
2748         journal->j_last = NULL;
2749         journal->j_first = NULL;
2750         init_waitqueue_head(&(journal->j_join_wait));
2751         mutex_init(&journal->j_mutex);
2752         mutex_init(&journal->j_flush_mutex);
2753
2754         journal->j_trans_id = 10;
2755         journal->j_mount_id = 10;
2756         journal->j_state = 0;
2757         atomic_set(&(journal->j_jlock), 0);
2758         journal->j_cnode_free_list = allocate_cnodes(num_cnodes);
2759         journal->j_cnode_free_orig = journal->j_cnode_free_list;
2760         journal->j_cnode_free = journal->j_cnode_free_list ? num_cnodes : 0;
2761         journal->j_cnode_used = 0;
2762         journal->j_must_wait = 0;
2763
2764         if (journal->j_cnode_free == 0) {
2765                 reiserfs_warning(sb, "journal-2004", "Journal cnode memory "
2766                                  "allocation failed (%ld bytes). Journal is "
2767                                  "too large for available memory. Usually "
2768                                  "this is due to a journal that is too large.",
2769                                  sizeof (struct reiserfs_journal_cnode) * num_cnodes);
2770                 goto free_and_return;
2771         }
2772
2773         init_journal_hash(sb);
2774         jl = journal->j_current_jl;
2775
2776         /*
2777          * get_list_bitmap() may call flush_commit_list() which
2778          * requires the lock. Calling flush_commit_list() shouldn't happen
2779          * this early but I like to be paranoid.
2780          */
2781         reiserfs_write_lock(sb);
2782         jl->j_list_bitmap = get_list_bitmap(sb, jl);
2783         reiserfs_write_unlock(sb);
2784         if (!jl->j_list_bitmap) {
2785                 reiserfs_warning(sb, "journal-2005",
2786                                  "get_list_bitmap failed for journal list 0");
2787                 goto free_and_return;
2788         }
2789
2790         ret = journal_read(sb);
2791         if (ret < 0) {
2792                 reiserfs_warning(sb, "reiserfs-2006",
2793                                  "Replay Failure, unable to mount");
2794                 goto free_and_return;
2795         }
2796
2797         INIT_DELAYED_WORK(&journal->j_work, flush_async_commits);
2798         journal->j_work_sb = sb;
2799         return 0;
2800       free_and_return:
2801         free_journal_ram(sb);
2802         return 1;
2803 }
2804
2805 /*
2806 ** test for a polite end of the current transaction.  Used by file_write, and should
2807 ** be used by delete to make sure they don't write more than can fit inside a single
2808 ** transaction
2809 */
2810 int journal_transaction_should_end(struct reiserfs_transaction_handle *th,
2811                                    int new_alloc)
2812 {
2813         struct reiserfs_journal *journal = SB_JOURNAL(th->t_super);
2814         time_t now = get_seconds();
2815         /* cannot restart while nested */
2816         BUG_ON(!th->t_trans_id);
2817         if (th->t_refcount > 1)
2818                 return 0;
2819         if (journal->j_must_wait > 0 ||
2820             (journal->j_len_alloc + new_alloc) >= journal->j_max_batch ||
2821             atomic_read(&(journal->j_jlock)) ||
2822             (now - journal->j_trans_start_time) > journal->j_max_trans_age ||
2823             journal->j_cnode_free < (journal->j_trans_max * 3)) {
2824                 return 1;
2825         }
2826
2827         journal->j_len_alloc += new_alloc;
2828         th->t_blocks_allocated += new_alloc ;
2829         return 0;
2830 }
2831
2832 /* this must be called inside a transaction
2833 */
2834 void reiserfs_block_writes(struct reiserfs_transaction_handle *th)
2835 {
2836         struct reiserfs_journal *journal = SB_JOURNAL(th->t_super);
2837         BUG_ON(!th->t_trans_id);
2838         journal->j_must_wait = 1;
2839         set_bit(J_WRITERS_BLOCKED, &journal->j_state);
2840         return;
2841 }
2842
2843 /* this must be called without a transaction started
2844 */
2845 void reiserfs_allow_writes(struct super_block *s)
2846 {
2847         struct reiserfs_journal *journal = SB_JOURNAL(s);
2848         clear_bit(J_WRITERS_BLOCKED, &journal->j_state);
2849         wake_up(&journal->j_join_wait);
2850 }
2851
2852 /* this must be called without a transaction started
2853 */
2854 void reiserfs_wait_on_write_block(struct super_block *s)
2855 {
2856         struct reiserfs_journal *journal = SB_JOURNAL(s);
2857         wait_event(journal->j_join_wait,
2858                    !test_bit(J_WRITERS_BLOCKED, &journal->j_state));
2859 }
2860
2861 static void queue_log_writer(struct super_block *s)
2862 {
2863         wait_queue_t wait;
2864         struct reiserfs_journal *journal = SB_JOURNAL(s);
2865         set_bit(J_WRITERS_QUEUED, &journal->j_state);
2866
2867         /*
2868          * we don't want to use wait_event here because
2869          * we only want to wait once.
2870          */
2871         init_waitqueue_entry(&wait, current);
2872         add_wait_queue(&journal->j_join_wait, &wait);
2873         set_current_state(TASK_UNINTERRUPTIBLE);
2874         if (test_bit(J_WRITERS_QUEUED, &journal->j_state)) {
2875                 int depth = reiserfs_write_unlock_nested(s);
2876                 schedule();
2877                 reiserfs_write_lock_nested(s, depth);
2878         }
2879         __set_current_state(TASK_RUNNING);
2880         remove_wait_queue(&journal->j_join_wait, &wait);
2881 }
2882
2883 static void wake_queued_writers(struct super_block *s)
2884 {
2885         struct reiserfs_journal *journal = SB_JOURNAL(s);
2886         if (test_and_clear_bit(J_WRITERS_QUEUED, &journal->j_state))
2887                 wake_up(&journal->j_join_wait);
2888 }
2889
2890 static void let_transaction_grow(struct super_block *sb, unsigned int trans_id)
2891 {
2892         struct reiserfs_journal *journal = SB_JOURNAL(sb);
2893         unsigned long bcount = journal->j_bcount;
2894         while (1) {
2895                 int depth;
2896
2897                 depth = reiserfs_write_unlock_nested(sb);
2898                 schedule_timeout_uninterruptible(1);
2899                 reiserfs_write_lock_nested(sb, depth);
2900
2901                 journal->j_current_jl->j_state |= LIST_COMMIT_PENDING;
2902                 while ((atomic_read(&journal->j_wcount) > 0 ||
2903                         atomic_read(&journal->j_jlock)) &&
2904                        journal->j_trans_id == trans_id) {
2905                         queue_log_writer(sb);
2906                 }
2907                 if (journal->j_trans_id != trans_id)
2908                         break;
2909                 if (bcount == journal->j_bcount)
2910                         break;
2911                 bcount = journal->j_bcount;
2912         }
2913 }
2914
2915 /* join == true if you must join an existing transaction.
2916 ** join == false if you can deal with waiting for others to finish
2917 **
2918 ** this will block until the transaction is joinable.  send the number of blocks you
2919 ** expect to use in nblocks.
2920 */
2921 static int do_journal_begin_r(struct reiserfs_transaction_handle *th,
2922                               struct super_block *sb, unsigned long nblocks,
2923                               int join)
2924 {
2925         time_t now = get_seconds();
2926         unsigned int old_trans_id;
2927         struct reiserfs_journal *journal = SB_JOURNAL(sb);
2928         struct reiserfs_transaction_handle myth;
2929         int sched_count = 0;
2930         int retval;
2931         int depth;
2932
2933         reiserfs_check_lock_depth(sb, "journal_begin");
2934         BUG_ON(nblocks > journal->j_trans_max);
2935
2936         PROC_INFO_INC(sb, journal.journal_being);
2937         /* set here for journal_join */
2938         th->t_refcount = 1;
2939         th->t_super = sb;
2940
2941       relock:
2942         lock_journal(sb);
2943         if (join != JBEGIN_ABORT && reiserfs_is_journal_aborted(journal)) {
2944                 unlock_journal(sb);
2945                 retval = journal->j_errno;
2946                 goto out_fail;
2947         }
2948         journal->j_bcount++;
2949
2950         if (test_bit(J_WRITERS_BLOCKED, &journal->j_state)) {
2951                 unlock_journal(sb);
2952                 depth = reiserfs_write_unlock_nested(sb);
2953                 reiserfs_wait_on_write_block(sb);
2954                 reiserfs_write_lock_nested(sb, depth);
2955                 PROC_INFO_INC(sb, journal.journal_relock_writers);
2956                 goto relock;
2957         }
2958         now = get_seconds();
2959
2960         /* if there is no room in the journal OR
2961          ** if this transaction is too old, and we weren't called joinable, wait for it to finish before beginning
2962          ** we don't sleep if there aren't other writers
2963          */
2964
2965         if ((!join && journal->j_must_wait > 0) ||
2966             (!join
2967              && (journal->j_len_alloc + nblocks + 2) >= journal->j_max_batch)
2968             || (!join && atomic_read(&journal->j_wcount) > 0
2969                 && journal->j_trans_start_time > 0
2970                 && (now - journal->j_trans_start_time) >
2971                 journal->j_max_trans_age) || (!join
2972                                               && atomic_read(&journal->j_jlock))
2973             || (!join && journal->j_cnode_free < (journal->j_trans_max * 3))) {
2974
2975                 old_trans_id = journal->j_trans_id;
2976                 unlock_journal(sb);     /* allow others to finish this transaction */
2977
2978                 if (!join && (journal->j_len_alloc + nblocks + 2) >=
2979                     journal->j_max_batch &&
2980                     ((journal->j_len + nblocks + 2) * 100) <
2981                     (journal->j_len_alloc * 75)) {
2982                         if (atomic_read(&journal->j_wcount) > 10) {
2983                                 sched_count++;
2984                                 queue_log_writer(sb);
2985                                 goto relock;
2986                         }
2987                 }
2988                 /* don't mess with joining the transaction if all we have to do is
2989                  * wait for someone else to do a commit
2990                  */
2991                 if (atomic_read(&journal->j_jlock)) {
2992                         while (journal->j_trans_id == old_trans_id &&
2993                                atomic_read(&journal->j_jlock)) {
2994                                 queue_log_writer(sb);
2995                         }
2996                         goto relock;
2997                 }
2998                 retval = journal_join(&myth, sb, 1);
2999                 if (retval)
3000                         goto out_fail;
3001
3002                 /* someone might have ended the transaction while we joined */
3003                 if (old_trans_id != journal->j_trans_id) {
3004                         retval = do_journal_end(&myth, sb, 1, 0);
3005                 } else {
3006                         retval = do_journal_end(&myth, sb, 1, COMMIT_NOW);
3007                 }
3008
3009                 if (retval)
3010                         goto out_fail;
3011
3012                 PROC_INFO_INC(sb, journal.journal_relock_wcount);
3013                 goto relock;
3014         }
3015         /* we are the first writer, set trans_id */
3016         if (journal->j_trans_start_time == 0) {
3017                 journal->j_trans_start_time = get_seconds();
3018         }
3019         atomic_inc(&(journal->j_wcount));
3020         journal->j_len_alloc += nblocks;
3021         th->t_blocks_logged = 0;
3022         th->t_blocks_allocated = nblocks;
3023         th->t_trans_id = journal->j_trans_id;
3024         unlock_journal(sb);
3025         INIT_LIST_HEAD(&th->t_list);
3026         return 0;
3027
3028       out_fail:
3029         memset(th, 0, sizeof(*th));
3030         /* Re-set th->t_super, so we can properly keep track of how many
3031          * persistent transactions there are. We need to do this so if this
3032          * call is part of a failed restart_transaction, we can free it later */
3033         th->t_super = sb;
3034         return retval;
3035 }
3036
3037 struct reiserfs_transaction_handle *reiserfs_persistent_transaction(struct
3038                                                                     super_block
3039                                                                     *s,
3040                                                                     int nblocks)
3041 {
3042         int ret;
3043         struct reiserfs_transaction_handle *th;
3044
3045         /* if we're nesting into an existing transaction.  It will be
3046          ** persistent on its own
3047          */
3048         if (reiserfs_transaction_running(s)) {
3049                 th = current->journal_info;
3050                 th->t_refcount++;
3051                 BUG_ON(th->t_refcount < 2);
3052
3053                 return th;
3054         }
3055         th = kmalloc(sizeof(struct reiserfs_transaction_handle), GFP_NOFS);
3056         if (!th)
3057                 return NULL;
3058         ret = journal_begin(th, s, nblocks);
3059         if (ret) {
3060                 kfree(th);
3061                 return NULL;
3062         }
3063
3064         SB_JOURNAL(s)->j_persistent_trans++;
3065         return th;
3066 }
3067
3068 int reiserfs_end_persistent_transaction(struct reiserfs_transaction_handle *th)
3069 {
3070         struct super_block *s = th->t_super;
3071         int ret = 0;
3072         if (th->t_trans_id)
3073                 ret = journal_end(th, th->t_super, th->t_blocks_allocated);
3074         else
3075                 ret = -EIO;
3076         if (th->t_refcount == 0) {
3077                 SB_JOURNAL(s)->j_persistent_trans--;
3078                 kfree(th);
3079         }
3080         return ret;
3081 }
3082
3083 static int journal_join(struct reiserfs_transaction_handle *th,
3084                         struct super_block *sb, unsigned long nblocks)
3085 {
3086         struct reiserfs_transaction_handle *cur_th = current->journal_info;
3087
3088         /* this keeps do_journal_end from NULLing out the current->journal_info
3089          ** pointer
3090          */
3091         th->t_handle_save = cur_th;
3092         BUG_ON(cur_th && cur_th->t_refcount > 1);
3093         return do_journal_begin_r(th, sb, nblocks, JBEGIN_JOIN);
3094 }
3095
3096 int journal_join_abort(struct reiserfs_transaction_handle *th,
3097                        struct super_block *sb, unsigned long nblocks)
3098 {
3099         struct reiserfs_transaction_handle *cur_th = current->journal_info;
3100
3101         /* this keeps do_journal_end from NULLing out the current->journal_info
3102          ** pointer
3103          */
3104         th->t_handle_save = cur_th;
3105         BUG_ON(cur_th && cur_th->t_refcount > 1);
3106         return do_journal_begin_r(th, sb, nblocks, JBEGIN_ABORT);
3107 }
3108
3109 int journal_begin(struct reiserfs_transaction_handle *th,
3110                   struct super_block *sb, unsigned long nblocks)
3111 {
3112         struct reiserfs_transaction_handle *cur_th = current->journal_info;
3113         int ret;
3114
3115         th->t_handle_save = NULL;
3116         if (cur_th) {
3117                 /* we are nesting into the current transaction */
3118                 if (cur_th->t_super == sb) {
3119                         BUG_ON(!cur_th->t_refcount);
3120                         cur_th->t_refcount++;
3121                         memcpy(th, cur_th, sizeof(*th));
3122                         if (th->t_refcount <= 1)
3123                                 reiserfs_warning(sb, "reiserfs-2005",
3124                                                  "BAD: refcount <= 1, but "
3125                                                  "journal_info != 0");
3126                         return 0;
3127                 } else {
3128                         /* we've ended up with a handle from a different filesystem.
3129                          ** save it and restore on journal_end.  This should never
3130                          ** really happen...
3131                          */
3132                         reiserfs_warning(sb, "clm-2100",
3133                                          "nesting info a different FS");
3134                         th->t_handle_save = current->journal_info;
3135                         current->journal_info = th;
3136                 }
3137         } else {
3138                 current->journal_info = th;
3139         }
3140         ret = do_journal_begin_r(th, sb, nblocks, JBEGIN_REG);
3141         BUG_ON(current->journal_info != th);
3142
3143         /* I guess this boils down to being the reciprocal of clm-2100 above.
3144          * If do_journal_begin_r fails, we need to put it back, since journal_end
3145          * won't be called to do it. */
3146         if (ret)
3147                 current->journal_info = th->t_handle_save;
3148         else
3149                 BUG_ON(!th->t_refcount);
3150
3151         return ret;
3152 }
3153
3154 /*
3155 ** puts bh into the current transaction.  If it was already there, reorders removes the
3156 ** old pointers from the hash, and puts new ones in (to make sure replay happen in the right order).
3157 **
3158 ** if it was dirty, cleans and files onto the clean list.  I can't let it be dirty again until the
3159 ** transaction is committed.
3160 **
3161 ** if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3162 */
3163 int journal_mark_dirty(struct reiserfs_transaction_handle *th,
3164                        struct super_block *sb, struct buffer_head *bh)
3165 {
3166         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3167         struct reiserfs_journal_cnode *cn = NULL;
3168         int count_already_incd = 0;
3169         int prepared = 0;
3170         BUG_ON(!th->t_trans_id);
3171
3172         PROC_INFO_INC(sb, journal.mark_dirty);
3173         if (th->t_trans_id != journal->j_trans_id) {
3174                 reiserfs_panic(th->t_super, "journal-1577",
3175                                "handle trans id %ld != current trans id %ld",
3176                                th->t_trans_id, journal->j_trans_id);
3177         }
3178
3179         prepared = test_clear_buffer_journal_prepared(bh);
3180         clear_buffer_journal_restore_dirty(bh);
3181         /* already in this transaction, we are done */
3182         if (buffer_journaled(bh)) {
3183                 PROC_INFO_INC(sb, journal.mark_dirty_already);
3184                 return 0;
3185         }
3186
3187         /* this must be turned into a panic instead of a warning.  We can't allow
3188          ** a dirty or journal_dirty or locked buffer to be logged, as some changes
3189          ** could get to disk too early.  NOT GOOD.
3190          */
3191         if (!prepared || buffer_dirty(bh)) {
3192                 reiserfs_warning(sb, "journal-1777",
3193                                  "buffer %llu bad state "
3194                                  "%cPREPARED %cLOCKED %cDIRTY %cJDIRTY_WAIT",
3195                                  (unsigned long long)bh->b_blocknr,
3196                                  prepared ? ' ' : '!',
3197                                  buffer_locked(bh) ? ' ' : '!',
3198                                  buffer_dirty(bh) ? ' ' : '!',
3199                                  buffer_journal_dirty(bh) ? ' ' : '!');
3200         }
3201
3202         if (atomic_read(&(journal->j_wcount)) <= 0) {
3203                 reiserfs_warning(sb, "journal-1409",
3204                                  "returning because j_wcount was %d",
3205                                  atomic_read(&(journal->j_wcount)));
3206                 return 1;
3207         }
3208         /* this error means I've screwed up, and we've overflowed the transaction.
3209          ** Nothing can be done here, except make the FS readonly or panic.
3210          */
3211         if (journal->j_len >= journal->j_trans_max) {
3212                 reiserfs_panic(th->t_super, "journal-1413",
3213                                "j_len (%lu) is too big",
3214                                journal->j_len);
3215         }
3216
3217         if (buffer_journal_dirty(bh)) {
3218                 count_already_incd = 1;
3219                 PROC_INFO_INC(sb, journal.mark_dirty_notjournal);
3220                 clear_buffer_journal_dirty(bh);
3221         }
3222
3223         if (journal->j_len > journal->j_len_alloc) {
3224                 journal->j_len_alloc = journal->j_len + JOURNAL_PER_BALANCE_CNT;
3225         }
3226
3227         set_buffer_journaled(bh);
3228
3229         /* now put this guy on the end */
3230         if (!cn) {
3231                 cn = get_cnode(sb);
3232                 if (!cn) {
3233                         reiserfs_panic(sb, "journal-4", "get_cnode failed!");
3234                 }
3235
3236                 if (th->t_blocks_logged == th->t_blocks_allocated) {
3237                         th->t_blocks_allocated += JOURNAL_PER_BALANCE_CNT;
3238                         journal->j_len_alloc += JOURNAL_PER_BALANCE_CNT;
3239                 }
3240                 th->t_blocks_logged++;
3241                 journal->j_len++;
3242
3243                 cn->bh = bh;
3244                 cn->blocknr = bh->b_blocknr;
3245                 cn->sb = sb;
3246                 cn->jlist = NULL;
3247                 insert_journal_hash(journal->j_hash_table, cn);
3248                 if (!count_already_incd) {
3249                         get_bh(bh);
3250                 }
3251         }
3252         cn->next = NULL;
3253         cn->prev = journal->j_last;
3254         cn->bh = bh;
3255         if (journal->j_last) {
3256                 journal->j_last->next = cn;
3257                 journal->j_last = cn;
3258         } else {
3259                 journal->j_first = cn;
3260                 journal->j_last = cn;
3261         }
3262         reiserfs_schedule_old_flush(sb);
3263         return 0;
3264 }
3265
3266 int journal_end(struct reiserfs_transaction_handle *th,
3267                 struct super_block *sb, unsigned long nblocks)
3268 {
3269         if (!current->journal_info && th->t_refcount > 1)
3270                 reiserfs_warning(sb, "REISER-NESTING",
3271                                  "th NULL, refcount %d", th->t_refcount);
3272
3273         if (!th->t_trans_id) {
3274                 WARN_ON(1);
3275                 return -EIO;
3276         }
3277
3278         th->t_refcount--;
3279         if (th->t_refcount > 0) {
3280                 struct reiserfs_transaction_handle *cur_th =
3281                     current->journal_info;
3282
3283                 /* we aren't allowed to close a nested transaction on a different
3284                  ** filesystem from the one in the task struct
3285                  */
3286                 BUG_ON(cur_th->t_super != th->t_super);
3287
3288                 if (th != cur_th) {
3289                         memcpy(current->journal_info, th, sizeof(*th));
3290                         th->t_trans_id = 0;
3291                 }
3292                 return 0;
3293         } else {
3294                 return do_journal_end(th, sb, nblocks, 0);
3295         }
3296 }
3297
3298 /* removes from the current transaction, relsing and descrementing any counters.
3299 ** also files the removed buffer directly onto the clean list
3300 **
3301 ** called by journal_mark_freed when a block has been deleted
3302 **
3303 ** returns 1 if it cleaned and relsed the buffer. 0 otherwise
3304 */
3305 static int remove_from_transaction(struct super_block *sb,
3306                                    b_blocknr_t blocknr, int already_cleaned)
3307 {
3308         struct buffer_head *bh;
3309         struct reiserfs_journal_cnode *cn;
3310         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3311         int ret = 0;
3312
3313         cn = get_journal_hash_dev(sb, journal->j_hash_table, blocknr);
3314         if (!cn || !cn->bh) {
3315                 return ret;
3316         }
3317         bh = cn->bh;
3318         if (cn->prev) {
3319                 cn->prev->next = cn->next;
3320         }
3321         if (cn->next) {
3322                 cn->next->prev = cn->prev;
3323         }
3324         if (cn == journal->j_first) {
3325                 journal->j_first = cn->next;
3326         }
3327         if (cn == journal->j_last) {
3328                 journal->j_last = cn->prev;
3329         }
3330         if (bh)
3331                 remove_journal_hash(sb, journal->j_hash_table, NULL,
3332                                     bh->b_blocknr, 0);
3333         clear_buffer_journaled(bh);     /* don't log this one */
3334
3335         if (!already_cleaned) {
3336                 clear_buffer_journal_dirty(bh);
3337                 clear_buffer_dirty(bh);
3338                 clear_buffer_journal_test(bh);
3339                 put_bh(bh);
3340                 if (atomic_read(&(bh->b_count)) < 0) {
3341                         reiserfs_warning(sb, "journal-1752",
3342                                          "b_count < 0");
3343                 }
3344                 ret = 1;
3345         }
3346         journal->j_len--;
3347         journal->j_len_alloc--;
3348         free_cnode(sb, cn);
3349         return ret;
3350 }
3351
3352 /*
3353 ** for any cnode in a journal list, it can only be dirtied of all the
3354 ** transactions that include it are committed to disk.
3355 ** this checks through each transaction, and returns 1 if you are allowed to dirty,
3356 ** and 0 if you aren't
3357 **
3358 ** it is called by dirty_journal_list, which is called after flush_commit_list has gotten all the log
3359 ** blocks for a given transaction on disk
3360 **
3361 */
3362 static int can_dirty(struct reiserfs_journal_cnode *cn)
3363 {
3364         struct super_block *sb = cn->sb;
3365         b_blocknr_t blocknr = cn->blocknr;
3366         struct reiserfs_journal_cnode *cur = cn->hprev;
3367         int can_dirty = 1;
3368
3369         /* first test hprev.  These are all newer than cn, so any node here
3370          ** with the same block number and dev means this node can't be sent
3371          ** to disk right now.
3372          */
3373         while (cur && can_dirty) {
3374                 if (cur->jlist && cur->bh && cur->blocknr && cur->sb == sb &&
3375                     cur->blocknr == blocknr) {
3376                         can_dirty = 0;
3377                 }
3378                 cur = cur->hprev;
3379         }
3380         /* then test hnext.  These are all older than cn.  As long as they
3381          ** are committed to the log, it is safe to write cn to disk
3382          */
3383         cur = cn->hnext;
3384         while (cur && can_dirty) {
3385                 if (cur->jlist && cur->jlist->j_len > 0 &&
3386                     atomic_read(&(cur->jlist->j_commit_left)) > 0 && cur->bh &&
3387                     cur->blocknr && cur->sb == sb && cur->blocknr == blocknr) {
3388                         can_dirty = 0;
3389                 }
3390                 cur = cur->hnext;
3391         }
3392         return can_dirty;
3393 }
3394
3395 /* syncs the commit blocks, but does not force the real buffers to disk
3396 ** will wait until the current transaction is done/committed before returning
3397 */
3398 int journal_end_sync(struct reiserfs_transaction_handle *th,
3399                      struct super_block *sb, unsigned long nblocks)
3400 {
3401         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3402
3403         BUG_ON(!th->t_trans_id);
3404         /* you can sync while nested, very, very bad */
3405         BUG_ON(th->t_refcount > 1);
3406         if (journal->j_len == 0) {
3407                 reiserfs_prepare_for_journal(sb, SB_BUFFER_WITH_SB(sb),
3408                                              1);
3409                 journal_mark_dirty(th, sb, SB_BUFFER_WITH_SB(sb));
3410         }
3411         return do_journal_end(th, sb, nblocks, COMMIT_NOW | WAIT);
3412 }
3413
3414 /*
3415 ** writeback the pending async commits to disk
3416 */
3417 static void flush_async_commits(struct work_struct *work)
3418 {
3419         struct reiserfs_journal *journal =
3420                 container_of(work, struct reiserfs_journal, j_work.work);
3421         struct super_block *sb = journal->j_work_sb;
3422         struct reiserfs_journal_list *jl;
3423         struct list_head *entry;
3424
3425         reiserfs_write_lock(sb);
3426         if (!list_empty(&journal->j_journal_list)) {
3427                 /* last entry is the youngest, commit it and you get everything */
3428                 entry = journal->j_journal_list.prev;
3429                 jl = JOURNAL_LIST_ENTRY(entry);
3430                 flush_commit_list(sb, jl, 1);
3431         }
3432         reiserfs_write_unlock(sb);
3433 }
3434
3435 /*
3436 ** flushes any old transactions to disk
3437 ** ends the current transaction if it is too old
3438 */
3439 void reiserfs_flush_old_commits(struct super_block *sb)
3440 {
3441         time_t now;
3442         struct reiserfs_transaction_handle th;
3443         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3444
3445         now = get_seconds();
3446         /* safety check so we don't flush while we are replaying the log during
3447          * mount
3448          */
3449         if (list_empty(&journal->j_journal_list))
3450                 return;
3451
3452         /* check the current transaction.  If there are no writers, and it is
3453          * too old, finish it, and force the commit blocks to disk
3454          */
3455         if (atomic_read(&journal->j_wcount) <= 0 &&
3456             journal->j_trans_start_time > 0 &&
3457             journal->j_len > 0 &&
3458             (now - journal->j_trans_start_time) > journal->j_max_trans_age) {
3459                 if (!journal_join(&th, sb, 1)) {
3460                         reiserfs_prepare_for_journal(sb,
3461                                                      SB_BUFFER_WITH_SB(sb),
3462                                                      1);
3463                         journal_mark_dirty(&th, sb,
3464                                            SB_BUFFER_WITH_SB(sb));
3465
3466                         /* we're only being called from kreiserfsd, it makes no sense to do
3467                          ** an async commit so that kreiserfsd can do it later
3468                          */
3469                         do_journal_end(&th, sb, 1, COMMIT_NOW | WAIT);
3470                 }
3471         }
3472 }
3473
3474 /*
3475 ** returns 0 if do_journal_end should return right away, returns 1 if do_journal_end should finish the commit
3476 **
3477 ** if the current transaction is too old, but still has writers, this will wait on j_join_wait until all
3478 ** the writers are done.  By the time it wakes up, the transaction it was called has already ended, so it just
3479 ** flushes the commit list and returns 0.
3480 **
3481 ** Won't batch when flush or commit_now is set.  Also won't batch when others are waiting on j_join_wait.
3482 **
3483 ** Note, we can't allow the journal_end to proceed while there are still writers in the log.
3484 */
3485 static int check_journal_end(struct reiserfs_transaction_handle *th,
3486                              struct super_block *sb, unsigned long nblocks,
3487                              int flags)
3488 {
3489
3490         time_t now;
3491         int flush = flags & FLUSH_ALL;
3492         int commit_now = flags & COMMIT_NOW;
3493         int wait_on_commit = flags & WAIT;
3494         struct reiserfs_journal_list *jl;
3495         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3496
3497         BUG_ON(!th->t_trans_id);
3498
3499         if (th->t_trans_id != journal->j_trans_id) {
3500                 reiserfs_panic(th->t_super, "journal-1577",
3501                                "handle trans id %ld != current trans id %ld",
3502                                th->t_trans_id, journal->j_trans_id);
3503         }
3504
3505         journal->j_len_alloc -= (th->t_blocks_allocated - th->t_blocks_logged);
3506         if (atomic_read(&(journal->j_wcount)) > 0) {    /* <= 0 is allowed.  unmounting might not call begin */
3507                 atomic_dec(&(journal->j_wcount));
3508         }
3509
3510         /* BUG, deal with case where j_len is 0, but people previously freed blocks need to be released
3511          ** will be dealt with by next transaction that actually writes something, but should be taken
3512          ** care of in this trans
3513          */
3514         BUG_ON(journal->j_len == 0);
3515
3516         /* if wcount > 0, and we are called to with flush or commit_now,
3517          ** we wait on j_join_wait.  We will wake up when the last writer has
3518          ** finished the transaction, and started it on its way to the disk.
3519          ** Then, we flush the commit or journal list, and just return 0
3520          ** because the rest of journal end was already done for this transaction.
3521          */
3522         if (atomic_read(&(journal->j_wcount)) > 0) {
3523                 if (flush || commit_now) {
3524                         unsigned trans_id;
3525
3526                         jl = journal->j_current_jl;
3527                         trans_id = jl->j_trans_id;
3528                         if (wait_on_commit)
3529                                 jl->j_state |= LIST_COMMIT_PENDING;
3530                         atomic_set(&(journal->j_jlock), 1);
3531                         if (flush) {
3532                                 journal->j_next_full_flush = 1;
3533                         }
3534                         unlock_journal(sb);
3535
3536                         /* sleep while the current transaction is still j_jlocked */
3537                         while (journal->j_trans_id == trans_id) {
3538                                 if (atomic_read(&journal->j_jlock)) {
3539                                         queue_log_writer(sb);
3540                                 } else {
3541                                         lock_journal(sb);
3542                                         if (journal->j_trans_id == trans_id) {
3543                                                 atomic_set(&(journal->j_jlock),
3544                                                            1);
3545                                         }
3546                                         unlock_journal(sb);
3547                                 }
3548                         }
3549                         BUG_ON(journal->j_trans_id == trans_id);
3550
3551                         if (commit_now
3552                             && journal_list_still_alive(sb, trans_id)
3553                             && wait_on_commit) {
3554                                 flush_commit_list(sb, jl, 1);
3555                         }
3556                         return 0;
3557                 }
3558                 unlock_journal(sb);
3559                 return 0;
3560         }
3561
3562         /* deal with old transactions where we are the last writers */
3563         now = get_seconds();
3564         if ((now - journal->j_trans_start_time) > journal->j_max_trans_age) {
3565                 commit_now = 1;
3566                 journal->j_next_async_flush = 1;
3567         }
3568         /* don't batch when someone is waiting on j_join_wait */
3569         /* don't batch when syncing the commit or flushing the whole trans */
3570         if (!(journal->j_must_wait > 0) && !(atomic_read(&(journal->j_jlock)))
3571             && !flush && !commit_now && (journal->j_len < journal->j_max_batch)
3572             && journal->j_len_alloc < journal->j_max_batch
3573             && journal->j_cnode_free > (journal->j_trans_max * 3)) {
3574                 journal->j_bcount++;
3575                 unlock_journal(sb);
3576                 return 0;
3577         }
3578
3579         if (journal->j_start > SB_ONDISK_JOURNAL_SIZE(sb)) {
3580                 reiserfs_panic(sb, "journal-003",
3581                                "j_start (%ld) is too high",
3582                                journal->j_start);
3583         }
3584         return 1;
3585 }
3586
3587 /*
3588 ** Does all the work that makes deleting blocks safe.
3589 ** when deleting a block mark BH_JNew, just remove it from the current transaction, clean it's buffer_head and move on.
3590 **
3591 ** otherwise:
3592 ** set a bit for the block in the journal bitmap.  That will prevent it from being allocated for unformatted nodes
3593 ** before this transaction has finished.
3594 **
3595 ** mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers.  That will prevent any old transactions with
3596 ** this block from trying to flush to the real location.  Since we aren't removing the cnode from the journal_list_hash,
3597 ** the block can't be reallocated yet.
3598 **
3599 ** Then remove it from the current transaction, decrementing any counters and filing it on the clean list.
3600 */
3601 int journal_mark_freed(struct reiserfs_transaction_handle *th,
3602                        struct super_block *sb, b_blocknr_t blocknr)
3603 {
3604         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3605         struct reiserfs_journal_cnode *cn = NULL;
3606         struct buffer_head *bh = NULL;
3607         struct reiserfs_list_bitmap *jb = NULL;
3608         int cleaned = 0;
3609         BUG_ON(!th->t_trans_id);
3610
3611         cn = get_journal_hash_dev(sb, journal->j_hash_table, blocknr);
3612         if (cn && cn->bh) {
3613                 bh = cn->bh;
3614                 get_bh(bh);
3615         }
3616         /* if it is journal new, we just remove it from this transaction */
3617         if (bh && buffer_journal_new(bh)) {
3618                 clear_buffer_journal_new(bh);
3619                 clear_prepared_bits(bh);
3620                 reiserfs_clean_and_file_buffer(bh);
3621                 cleaned = remove_from_transaction(sb, blocknr, cleaned);
3622         } else {
3623                 /* set the bit for this block in the journal bitmap for this transaction */
3624                 jb = journal->j_current_jl->j_list_bitmap;
3625                 if (!jb) {
3626                         reiserfs_panic(sb, "journal-1702",
3627                                        "journal_list_bitmap is NULL");
3628                 }
3629                 set_bit_in_list_bitmap(sb, blocknr, jb);
3630
3631                 /* Note, the entire while loop is not allowed to schedule.  */
3632
3633                 if (bh) {
3634                         clear_prepared_bits(bh);
3635                         reiserfs_clean_and_file_buffer(bh);
3636                 }
3637                 cleaned = remove_from_transaction(sb, blocknr, cleaned);
3638
3639                 /* find all older transactions with this block, make sure they don't try to write it out */
3640                 cn = get_journal_hash_dev(sb, journal->j_list_hash_table,
3641                                           blocknr);
3642                 while (cn) {
3643                         if (sb == cn->sb && blocknr == cn->blocknr) {
3644                                 set_bit(BLOCK_FREED, &cn->state);
3645                                 if (cn->bh) {
3646                                         if (!cleaned) {
3647                                                 /* remove_from_transaction will brelse the buffer if it was
3648                                                  ** in the current trans
3649                                                  */
3650                                                 clear_buffer_journal_dirty(cn->
3651                                                                            bh);
3652                                                 clear_buffer_dirty(cn->bh);
3653                                                 clear_buffer_journal_test(cn->
3654                                                                           bh);
3655                                                 cleaned = 1;
3656                                                 put_bh(cn->bh);
3657                                                 if (atomic_read
3658                                                     (&(cn->bh->b_count)) < 0) {
3659                                                         reiserfs_warning(sb,
3660                                                                  "journal-2138",
3661                                                                  "cn->bh->b_count < 0");
3662                                                 }
3663                                         }
3664                                         if (cn->jlist) {        /* since we are clearing the bh, we MUST dec nonzerolen */
3665                                                 atomic_dec(&
3666                                                            (cn->jlist->
3667                                                             j_nonzerolen));
3668                                         }
3669                                         cn->bh = NULL;
3670                                 }
3671                         }
3672                         cn = cn->hnext;
3673                 }
3674         }
3675
3676         if (bh)
3677                 release_buffer_page(bh); /* get_hash grabs the buffer */
3678         return 0;
3679 }
3680
3681 void reiserfs_update_inode_transaction(struct inode *inode)
3682 {
3683         struct reiserfs_journal *journal = SB_JOURNAL(inode->i_sb);
3684         REISERFS_I(inode)->i_jl = journal->j_current_jl;
3685         REISERFS_I(inode)->i_trans_id = journal->j_trans_id;
3686 }
3687
3688 /*
3689  * returns -1 on error, 0 if no commits/barriers were done and 1
3690  * if a transaction was actually committed and the barrier was done
3691  */
3692 static int __commit_trans_jl(struct inode *inode, unsigned long id,
3693                              struct reiserfs_journal_list *jl)
3694 {
3695         struct reiserfs_transaction_handle th;
3696         struct super_block *sb = inode->i_sb;
3697         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3698         int ret = 0;
3699
3700         /* is it from the current transaction, or from an unknown transaction? */
3701         if (id == journal->j_trans_id) {
3702                 jl = journal->j_current_jl;
3703                 /* try to let other writers come in and grow this transaction */
3704                 let_transaction_grow(sb, id);
3705                 if (journal->j_trans_id != id) {
3706                         goto flush_commit_only;
3707                 }
3708
3709                 ret = journal_begin(&th, sb, 1);
3710                 if (ret)
3711                         return ret;
3712
3713                 /* someone might have ended this transaction while we joined */
3714                 if (journal->j_trans_id != id) {
3715                         reiserfs_prepare_for_journal(sb, SB_BUFFER_WITH_SB(sb),
3716                                                      1);
3717                         journal_mark_dirty(&th, sb, SB_BUFFER_WITH_SB(sb));
3718                         ret = journal_end(&th, sb, 1);
3719                         goto flush_commit_only;
3720                 }
3721
3722                 ret = journal_end_sync(&th, sb, 1);
3723                 if (!ret)
3724                         ret = 1;
3725
3726         } else {
3727                 /* this gets tricky, we have to make sure the journal list in
3728                  * the inode still exists.  We know the list is still around
3729                  * if we've got a larger transaction id than the oldest list
3730                  */
3731               flush_commit_only:
3732                 if (journal_list_still_alive(inode->i_sb, id)) {
3733                         /*
3734                          * we only set ret to 1 when we know for sure
3735                          * the barrier hasn't been started yet on the commit
3736                          * block.
3737                          */
3738                         if (atomic_read(&jl->j_commit_left) > 1)
3739                                 ret = 1;
3740                         flush_commit_list(sb, jl, 1);
3741                         if (journal->j_errno)
3742                                 ret = journal->j_errno;
3743                 }
3744         }
3745         /* otherwise the list is gone, and long since committed */
3746         return ret;
3747 }
3748
3749 int reiserfs_commit_for_inode(struct inode *inode)
3750 {
3751         unsigned int id = REISERFS_I(inode)->i_trans_id;
3752         struct reiserfs_journal_list *jl = REISERFS_I(inode)->i_jl;
3753
3754         /* for the whole inode, assume unset id means it was
3755          * changed in the current transaction.  More conservative
3756          */
3757         if (!id || !jl) {
3758                 reiserfs_update_inode_transaction(inode);
3759                 id = REISERFS_I(inode)->i_trans_id;
3760                 /* jl will be updated in __commit_trans_jl */
3761         }
3762
3763         return __commit_trans_jl(inode, id, jl);
3764 }
3765
3766 void reiserfs_restore_prepared_buffer(struct super_block *sb,
3767                                       struct buffer_head *bh)
3768 {
3769         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3770         PROC_INFO_INC(sb, journal.restore_prepared);
3771         if (!bh) {
3772                 return;
3773         }
3774         if (test_clear_buffer_journal_restore_dirty(bh) &&
3775             buffer_journal_dirty(bh)) {
3776                 struct reiserfs_journal_cnode *cn;
3777                 reiserfs_write_lock(sb);
3778                 cn = get_journal_hash_dev(sb,
3779                                           journal->j_list_hash_table,
3780                                           bh->b_blocknr);
3781                 if (cn && can_dirty(cn)) {
3782                         set_buffer_journal_test(bh);
3783                         mark_buffer_dirty(bh);
3784                 }
3785                 reiserfs_write_unlock(sb);
3786         }
3787         clear_buffer_journal_prepared(bh);
3788 }
3789
3790 extern struct tree_balance *cur_tb;
3791 /*
3792 ** before we can change a metadata block, we have to make sure it won't
3793 ** be written to disk while we are altering it.  So, we must:
3794 ** clean it
3795 ** wait on it.
3796 **
3797 */
3798 int reiserfs_prepare_for_journal(struct super_block *sb,
3799                                  struct buffer_head *bh, int wait)
3800 {
3801         PROC_INFO_INC(sb, journal.prepare);
3802
3803         if (!trylock_buffer(bh)) {
3804                 if (!wait)
3805                         return 0;
3806                 lock_buffer(bh);
3807         }
3808         set_buffer_journal_prepared(bh);
3809         if (test_clear_buffer_dirty(bh) && buffer_journal_dirty(bh)) {
3810                 clear_buffer_journal_test(bh);
3811                 set_buffer_journal_restore_dirty(bh);
3812         }
3813         unlock_buffer(bh);
3814         return 1;
3815 }
3816
3817 /*
3818 ** long and ugly.  If flush, will not return until all commit
3819 ** blocks and all real buffers in the trans are on disk.
3820 ** If no_async, won't return until all commit blocks are on disk.
3821 **
3822 ** keep reading, there are comments as you go along
3823 **
3824 ** If the journal is aborted, we just clean up. Things like flushing
3825 ** journal lists, etc just won't happen.
3826 */
3827 static int do_journal_end(struct reiserfs_transaction_handle *th,
3828                           struct super_block *sb, unsigned long nblocks,
3829                           int flags)
3830 {
3831         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3832         struct reiserfs_journal_cnode *cn, *next, *jl_cn;
3833         struct reiserfs_journal_cnode *last_cn = NULL;
3834         struct reiserfs_journal_desc *desc;
3835         struct reiserfs_journal_commit *commit;
3836         struct buffer_head *c_bh;       /* commit bh */
3837         struct buffer_head *d_bh;       /* desc bh */
3838         int cur_write_start = 0;        /* start index of current log write */
3839         int old_start;
3840         int i;
3841         int flush;
3842         int wait_on_commit;
3843         struct reiserfs_journal_list *jl, *temp_jl;
3844         struct list_head *entry, *safe;
3845         unsigned long jindex;
3846         unsigned int commit_trans_id;
3847         int trans_half;
3848         int depth;
3849
3850         BUG_ON(th->t_refcount > 1);
3851         BUG_ON(!th->t_trans_id);
3852
3853         /* protect flush_older_commits from doing mistakes if the
3854            transaction ID counter gets overflowed.  */
3855         if (th->t_trans_id == ~0U)
3856                 flags |= FLUSH_ALL | COMMIT_NOW | WAIT;
3857         flush = flags & FLUSH_ALL;
3858         wait_on_commit = flags & WAIT;
3859
3860         current->journal_info = th->t_handle_save;
3861         reiserfs_check_lock_depth(sb, "journal end");
3862         if (journal->j_len == 0) {
3863                 reiserfs_prepare_for_journal(sb, SB_BUFFER_WITH_SB(sb),
3864                                              1);
3865                 journal_mark_dirty(th, sb, SB_BUFFER_WITH_SB(sb));
3866         }
3867
3868         lock_journal(sb);
3869         if (journal->j_next_full_flush) {
3870                 flags |= FLUSH_ALL;
3871                 flush = 1;
3872         }
3873         if (journal->j_next_async_flush) {
3874                 flags |= COMMIT_NOW | WAIT;
3875                 wait_on_commit = 1;
3876         }
3877
3878         /* check_journal_end locks the journal, and unlocks if it does not return 1
3879          ** it tells us if we should continue with the journal_end, or just return
3880          */
3881         if (!check_journal_end(th, sb, nblocks, flags)) {
3882                 reiserfs_schedule_old_flush(sb);
3883                 wake_queued_writers(sb);
3884                 reiserfs_async_progress_wait(sb);
3885                 goto out;
3886         }
3887
3888         /* check_journal_end might set these, check again */
3889         if (journal->j_next_full_flush) {
3890                 flush = 1;
3891         }
3892
3893         /*
3894          ** j must wait means we have to flush the log blocks, and the real blocks for
3895          ** this transaction
3896          */
3897         if (journal->j_must_wait > 0) {
3898                 flush = 1;
3899         }
3900 #ifdef REISERFS_PREALLOCATE
3901         /* quota ops might need to nest, setup the journal_info pointer for them
3902          * and raise the refcount so that it is > 0. */
3903         current->journal_info = th;
3904         th->t_refcount++;
3905         reiserfs_discard_all_prealloc(th);      /* it should not involve new blocks into
3906                                                  * the transaction */
3907         th->t_refcount--;
3908         current->journal_info = th->t_handle_save;
3909 #endif
3910
3911         /* setup description block */
3912         d_bh =
3913             journal_getblk(sb,
3914                            SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
3915                            journal->j_start);
3916         set_buffer_uptodate(d_bh);
3917         desc = (struct reiserfs_journal_desc *)(d_bh)->b_data;
3918         memset(d_bh->b_data, 0, d_bh->b_size);
3919         memcpy(get_journal_desc_magic(d_bh), JOURNAL_DESC_MAGIC, 8);
3920         set_desc_trans_id(desc, journal->j_trans_id);
3921
3922         /* setup commit block.  Don't write (keep it clean too) this one until after everyone else is written */
3923         c_bh = journal_getblk(sb, SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
3924                               ((journal->j_start + journal->j_len +
3925                                 1) % SB_ONDISK_JOURNAL_SIZE(sb)));
3926         commit = (struct reiserfs_journal_commit *)c_bh->b_data;
3927         memset(c_bh->b_data, 0, c_bh->b_size);
3928         set_commit_trans_id(commit, journal->j_trans_id);
3929         set_buffer_uptodate(c_bh);
3930
3931         /* init this journal list */
3932         jl = journal->j_current_jl;
3933
3934         /* we lock the commit before doing anything because
3935          * we want to make sure nobody tries to run flush_commit_list until
3936          * the new transaction is fully setup, and we've already flushed the
3937          * ordered bh list
3938          */
3939         reiserfs_mutex_lock_safe(&jl->j_commit_mutex, sb);
3940
3941         /* save the transaction id in case we need to commit it later */
3942         commit_trans_id = jl->j_trans_id;
3943
3944         atomic_set(&jl->j_older_commits_done, 0);
3945         jl->j_trans_id = journal->j_trans_id;
3946         jl->j_timestamp = journal->j_trans_start_time;
3947         jl->j_commit_bh = c_bh;
3948         jl->j_start = journal->j_start;
3949         jl->j_len = journal->j_len;
3950         atomic_set(&jl->j_nonzerolen, journal->j_len);
3951         atomic_set(&jl->j_commit_left, journal->j_len + 2);
3952         jl->j_realblock = NULL;
3953
3954         /* The ENTIRE FOR LOOP MUST not cause schedule to occur.
3955          **  for each real block, add it to the journal list hash,
3956          ** copy into real block index array in the commit or desc block
3957          */
3958         trans_half = journal_trans_half(sb->s_blocksize);
3959         for (i = 0, cn = journal->j_first; cn; cn = cn->next, i++) {
3960                 if (buffer_journaled(cn->bh)) {
3961                         jl_cn = get_cnode(sb);
3962                         if (!jl_cn) {
3963                                 reiserfs_panic(sb, "journal-1676",
3964                                                "get_cnode returned NULL");
3965                         }
3966                         if (i == 0) {
3967                                 jl->j_realblock = jl_cn;
3968                         }
3969                         jl_cn->prev = last_cn;
3970                         jl_cn->next = NULL;
3971                         if (last_cn) {
3972                                 last_cn->next = jl_cn;
3973                         }
3974                         last_cn = jl_cn;
3975                         /* make sure the block we are trying to log is not a block
3976                            of journal or reserved area */
3977
3978                         if (is_block_in_log_or_reserved_area
3979                             (sb, cn->bh->b_blocknr)) {
3980                                 reiserfs_panic(sb, "journal-2332",
3981                                                "Trying to log block %lu, "
3982                                                "which is a log block",
3983                                                cn->bh->b_blocknr);
3984                         }
3985                         jl_cn->blocknr = cn->bh->b_blocknr;
3986                         jl_cn->state = 0;
3987                         jl_cn->sb = sb;
3988                         jl_cn->bh = cn->bh;
3989                         jl_cn->jlist = jl;
3990                         insert_journal_hash(journal->j_list_hash_table, jl_cn);
3991                         if (i < trans_half) {
3992                                 desc->j_realblock[i] =
3993                                     cpu_to_le32(cn->bh->b_blocknr);
3994                         } else {
3995                                 commit->j_realblock[i - trans_half] =
3996                                     cpu_to_le32(cn->bh->b_blocknr);
3997                         }
3998                 } else {
3999                         i--;
4000                 }
4001         }
4002         set_desc_trans_len(desc, journal->j_len);
4003         set_desc_mount_id(desc, journal->j_mount_id);
4004         set_desc_trans_id(desc, journal->j_trans_id);
4005         set_commit_trans_len(commit, journal->j_len);
4006
4007         /* special check in case all buffers in the journal were marked for not logging */
4008         BUG_ON(journal->j_len == 0);
4009
4010         /* we're about to dirty all the log blocks, mark the description block
4011          * dirty now too.  Don't mark the commit block dirty until all the
4012          * others are on disk
4013          */
4014         mark_buffer_dirty(d_bh);
4015
4016         /* first data block is j_start + 1, so add one to cur_write_start wherever you use it */
4017         cur_write_start = journal->j_start;
4018         cn = journal->j_first;
4019         jindex = 1;             /* start at one so we don't get the desc again */
4020         while (cn) {
4021                 clear_buffer_journal_new(cn->bh);
4022                 /* copy all the real blocks into log area.  dirty log blocks */
4023                 if (buffer_journaled(cn->bh)) {
4024                         struct buffer_head *tmp_bh;
4025                         char *addr;
4026                         struct page *page;
4027                         tmp_bh =
4028                             journal_getblk(sb,
4029                                            SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
4030                                            ((cur_write_start +
4031                                              jindex) %
4032                                             SB_ONDISK_JOURNAL_SIZE(sb)));
4033                         set_buffer_uptodate(tmp_bh);
4034                         page = cn->bh->b_page;
4035                         addr = kmap(page);
4036                         memcpy(tmp_bh->b_data,
4037                                addr + offset_in_page(cn->bh->b_data),
4038                                cn->bh->b_size);
4039                         kunmap(page);
4040                         mark_buffer_dirty(tmp_bh);
4041                         jindex++;
4042                         set_buffer_journal_dirty(cn->bh);
4043                         clear_buffer_journaled(cn->bh);
4044                 } else {
4045                         /* JDirty cleared sometime during transaction.  don't log this one */
4046                         reiserfs_warning(sb, "journal-2048",
4047                                          "BAD, buffer in journal hash, "
4048                                          "but not JDirty!");
4049                         brelse(cn->bh);
4050                 }
4051                 next = cn->next;
4052                 free_cnode(sb, cn);
4053                 cn = next;
4054                 reiserfs_cond_resched(sb);
4055         }
4056
4057         /* we are done  with both the c_bh and d_bh, but
4058          ** c_bh must be written after all other commit blocks,
4059          ** so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4060          */
4061
4062         journal->j_current_jl = alloc_journal_list(sb);
4063
4064         /* now it is safe to insert this transaction on the main list */
4065         list_add_tail(&jl->j_list, &journal->j_journal_list);
4066         list_add_tail(&jl->j_working_list, &journal->j_working_list);
4067         journal->j_num_work_lists++;
4068
4069         /* reset journal values for the next transaction */
4070         old_start = journal->j_start;
4071         journal->j_start =
4072             (journal->j_start + journal->j_len +
4073              2) % SB_ONDISK_JOURNAL_SIZE(sb);
4074         atomic_set(&(journal->j_wcount), 0);
4075         journal->j_bcount = 0;
4076         journal->j_last = NULL;
4077         journal->j_first = NULL;
4078         journal->j_len = 0;
4079         journal->j_trans_start_time = 0;
4080         /* check for trans_id overflow */
4081         if (++journal->j_trans_id == 0)
4082                 journal->j_trans_id = 10;
4083         journal->j_current_jl->j_trans_id = journal->j_trans_id;
4084         journal->j_must_wait = 0;
4085         journal->j_len_alloc = 0;
4086         journal->j_next_full_flush = 0;
4087         journal->j_next_async_flush = 0;
4088         init_journal_hash(sb);
4089
4090         // make sure reiserfs_add_jh sees the new current_jl before we
4091         // write out the tails
4092         smp_mb();
4093
4094         /* tail conversion targets have to hit the disk before we end the
4095          * transaction.  Otherwise a later transaction might repack the tail
4096          * before this transaction commits, leaving the data block unflushed and
4097          * clean, if we crash before the later transaction commits, the data block
4098          * is lost.
4099          */
4100         if (!list_empty(&jl->j_tail_bh_list)) {
4101                 depth = reiserfs_write_unlock_nested(sb);
4102                 write_ordered_buffers(&journal->j_dirty_buffers_lock,
4103                                       journal, jl, &jl->j_tail_bh_list);
4104                 reiserfs_write_lock_nested(sb, depth);
4105         }
4106         BUG_ON(!list_empty(&jl->j_tail_bh_list));
4107         mutex_unlock(&jl->j_commit_mutex);
4108
4109         /* honor the flush wishes from the caller, simple commits can
4110          ** be done outside the journal lock, they are done below
4111          **
4112          ** if we don't flush the commit list right now, we put it into
4113          ** the work queue so the people waiting on the async progress work
4114          ** queue don't wait for this proc to flush journal lists and such.
4115          */
4116         if (flush) {
4117                 flush_commit_list(sb, jl, 1);
4118                 flush_journal_list(sb, jl, 1);
4119         } else if (!(jl->j_state & LIST_COMMIT_PENDING))
4120                 queue_delayed_work(REISERFS_SB(sb)->commit_wq,
4121                                    &journal->j_work, HZ / 10);
4122
4123         /* if the next transaction has any chance of wrapping, flush
4124          ** transactions that might get overwritten.  If any journal lists are very
4125          ** old flush them as well.
4126          */
4127       first_jl:
4128         list_for_each_safe(entry, safe, &journal->j_journal_list) {
4129                 temp_jl = JOURNAL_LIST_ENTRY(entry);
4130                 if (journal->j_start <= temp_jl->j_start) {
4131                         if ((journal->j_start + journal->j_trans_max + 1) >=
4132                             temp_jl->j_start) {
4133                                 flush_used_journal_lists(sb, temp_jl);
4134                                 goto first_jl;
4135                         } else if ((journal->j_start +
4136                                     journal->j_trans_max + 1) <
4137                                    SB_ONDISK_JOURNAL_SIZE(sb)) {
4138                                 /* if we don't cross into the next transaction and we don't
4139                                  * wrap, there is no way we can overlap any later transactions
4140                                  * break now
4141                                  */
4142                                 break;
4143                         }
4144                 } else if ((journal->j_start +
4145                             journal->j_trans_max + 1) >
4146                            SB_ONDISK_JOURNAL_SIZE(sb)) {
4147                         if (((journal->j_start + journal->j_trans_max + 1) %
4148                              SB_ONDISK_JOURNAL_SIZE(sb)) >=
4149                             temp_jl->j_start) {
4150                                 flush_used_journal_lists(sb, temp_jl);
4151                                 goto first_jl;
4152                         } else {
4153                                 /* we don't overlap anything from out start to the end of the
4154                                  * log, and our wrapped portion doesn't overlap anything at
4155                                  * the start of the log.  We can break
4156                                  */
4157                                 break;
4158                         }
4159                 }
4160         }
4161
4162         journal->j_current_jl->j_list_bitmap =
4163             get_list_bitmap(sb, journal->j_current_jl);
4164
4165         if (!(journal->j_current_jl->j_list_bitmap)) {
4166                 reiserfs_panic(sb, "journal-1996",
4167                                "could not get a list bitmap");
4168         }
4169
4170         atomic_set(&(journal->j_jlock), 0);
4171         unlock_journal(sb);
4172         /* wake up any body waiting to join. */
4173         clear_bit(J_WRITERS_QUEUED, &journal->j_state);
4174         wake_up(&(journal->j_join_wait));
4175
4176         if (!flush && wait_on_commit &&
4177             journal_list_still_alive(sb, commit_trans_id)) {
4178                 flush_commit_list(sb, jl, 1);
4179         }
4180       out:
4181         reiserfs_check_lock_depth(sb, "journal end2");
4182
4183         memset(th, 0, sizeof(*th));
4184         /* Re-set th->t_super, so we can properly keep track of how many
4185          * persistent transactions there are. We need to do this so if this
4186          * call is part of a failed restart_transaction, we can free it later */
4187         th->t_super = sb;
4188
4189         return journal->j_errno;
4190 }
4191
4192 /* Send the file system read only and refuse new transactions */
4193 void reiserfs_abort_journal(struct super_block *sb, int errno)
4194 {
4195         struct reiserfs_journal *journal = SB_JOURNAL(sb);
4196         if (test_bit(J_ABORTED, &journal->j_state))
4197                 return;
4198
4199         if (!journal->j_errno)
4200                 journal->j_errno = errno;
4201
4202         sb->s_flags |= MS_RDONLY;
4203         set_bit(J_ABORTED, &journal->j_state);
4204
4205 #ifdef CONFIG_REISERFS_CHECK
4206         dump_stack();
4207 #endif
4208 }