fs/reiserfs/journal.c

   1 /*
   2 ** Write ahead logging implementation copyright Chris Mason 2000
   3 **
   4 ** The background commits make this code very interelated, 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 <linux/reiserfs_fs.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/smp_lock.h>
  47 #include <linux/buffer_head.h>
  48 #include <linux/workqueue.h>
  49 #include <linux/writeback.h>
  50 #include <linux/blkdev.h>
  51 #include <linux/backing-dev.h>
  52 #include <linux/uaccess.h>
  53
  54 #include <asm/system.h>
  55
  56 /* gets a struct reiserfs_journal_list * from a list head */
  57 #define JOURNAL_LIST_ENTRY(h) (list_entry((h), struct reiserfs_journal_list, \
  58                                j_list))
  59 #define JOURNAL_WORK_ENTRY(h) (list_entry((h), struct reiserfs_journal_list, \
  60                                j_working_list))
  61
  62 /* the number of mounted filesystems.  This is used to decide when to
  63 ** start and kill the commit workqueue
  64 */
  65 static int reiserfs_mounted_fs_count;
  66
  67 static struct workqueue_struct *commit_wq;
  68
  69 #define JOURNAL_TRANS_HALF 1018 /* must be correct to keep the desc and commit
  70                                    structs at 4k */
  71 #define BUFNR 64                /*read ahead */
  72
  73 /* cnode stat bits.  Move these into reiserfs_fs.h */
  74
  75 #define BLOCK_FREED 2           /* this block was freed, and can't be written.  */
  76 #define BLOCK_FREED_HOLDER 3    /* this block was freed during this transaction, and can't be written */
  77
  78 #define BLOCK_NEEDS_FLUSH 4     /* used in flush_journal_list */
  79 #define BLOCK_DIRTIED 5
  80
  81 /* journal list state bits */
  82 #define LIST_TOUCHED 1
  83 #define LIST_DIRTY   2
  84 #define LIST_COMMIT_PENDING  4  /* someone will commit this list */
  85
  86 /* flags for do_journal_end */
  87 #define FLUSH_ALL   1           /* flush commit and real blocks */
  88 #define COMMIT_NOW  2           /* end and commit this transaction */
  89 #define WAIT        4           /* wait for the log blocks to hit the disk */
  90
  91 static int do_journal_end(struct reiserfs_transaction_handle *,
  92                           struct super_block *, unsigned long nblocks,
  93                           int flags);
  94 static int flush_journal_list(struct super_block *s,
  95                               struct reiserfs_journal_list *jl, int flushall);
  96 static int flush_commit_list(struct super_block *s,
  97                              struct reiserfs_journal_list *jl, int flushall);
  98 static int can_dirty(struct reiserfs_journal_cnode *cn);
  99 static int journal_join(struct reiserfs_transaction_handle *th,
 100                         struct super_block *sb, unsigned long nblocks);
 101 static int release_journal_dev(struct super_block *super,
 102                                struct reiserfs_journal *journal);
 103 static int dirty_one_transaction(struct super_block *s,
 104                                  struct reiserfs_journal_list *jl);
 105 static void flush_async_commits(struct work_struct *work);
 106 static void queue_log_writer(struct super_block *s);
 107
 108 /* values for join in do_journal_begin_r */
 109 enum {
 110         JBEGIN_REG = 0,         /* regular journal begin */
 111         JBEGIN_JOIN = 1,        /* join the running transaction if at all possible */
 112         JBEGIN_ABORT = 2,       /* called from cleanup code, ignores aborted flag */
 113 };
 114
 115 static int do_journal_begin_r(struct reiserfs_transaction_handle *th,
 116                               struct super_block *sb,
 117                               unsigned long nblocks, int join);
 118
 119 static void init_journal_hash(struct super_block *sb)
 120 {
 121         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 122         memset(journal->j_hash_table, 0,
 123                JOURNAL_HASH_SIZE * sizeof(struct reiserfs_journal_cnode *));
 124 }
 125
 126 /*
 127 ** clears BH_Dirty and sticks the buffer on the clean list.  Called because I can't allow refile_buffer to
 128 ** make schedule happen after I've freed a block.  Look at remove_from_transaction and journal_mark_freed for
 129 ** more details.
 130 */
 131 static int reiserfs_clean_and_file_buffer(struct buffer_head *bh)
 132 {
 133         if (bh) {
 134                 clear_buffer_dirty(bh);
 135                 clear_buffer_journal_test(bh);
 136         }
 137         return 0;
 138 }
 139
 140 static void disable_barrier(struct super_block *s)
 141 {
 142         REISERFS_SB(s)->s_mount_opt &= ~(1 << REISERFS_BARRIER_FLUSH);
 143         printk("reiserfs: disabling flush barriers on %s\n",
 144                reiserfs_bdevname(s));
 145 }
 146
 147 static struct reiserfs_bitmap_node *allocate_bitmap_node(struct super_block
 148                                                          *sb)
 149 {
 150         struct reiserfs_bitmap_node *bn;
 151         static int id;
 152
 153         bn = kmalloc(sizeof(struct reiserfs_bitmap_node), GFP_NOFS);
 154         if (!bn) {
 155                 return NULL;
 156         }
 157         bn->data = kzalloc(sb->s_blocksize, GFP_NOFS);
 158         if (!bn->data) {
 159                 kfree(bn);
 160                 return NULL;
 161         }
 162         bn->id = id++;
 163         INIT_LIST_HEAD(&bn->list);
 164         return bn;
 165 }
 166
 167 static struct reiserfs_bitmap_node *get_bitmap_node(struct super_block *sb)
 168 {
 169         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 170         struct reiserfs_bitmap_node *bn = NULL;
 171         struct list_head *entry = journal->j_bitmap_nodes.next;
 172
 173         journal->j_used_bitmap_nodes++;
 174       repeat:
 175
 176         if (entry != &journal->j_bitmap_nodes) {
 177                 bn = list_entry(entry, struct reiserfs_bitmap_node, list);
 178                 list_del(entry);
 179                 memset(bn->data, 0, sb->s_blocksize);
 180                 journal->j_free_bitmap_nodes--;
 181                 return bn;
 182         }
 183         bn = allocate_bitmap_node(sb);
 184         if (!bn) {
 185                 yield();
 186                 goto repeat;
 187         }
 188         return bn;
 189 }
 190 static inline void free_bitmap_node(struct super_block *sb,
 191                                     struct reiserfs_bitmap_node *bn)
 192 {
 193         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 194         journal->j_used_bitmap_nodes--;
 195         if (journal->j_free_bitmap_nodes > REISERFS_MAX_BITMAP_NODES) {
 196                 kfree(bn->data);
 197                 kfree(bn);
 198         } else {
 199                 list_add(&bn->list, &journal->j_bitmap_nodes);
 200                 journal->j_free_bitmap_nodes++;
 201         }
 202 }
 203
 204 static void allocate_bitmap_nodes(struct super_block *sb)
 205 {
 206         int i;
 207         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 208         struct reiserfs_bitmap_node *bn = NULL;
 209         for (i = 0; i < REISERFS_MIN_BITMAP_NODES; i++) {
 210                 bn = allocate_bitmap_node(sb);
 211                 if (bn) {
 212                         list_add(&bn->list, &journal->j_bitmap_nodes);
 213                         journal->j_free_bitmap_nodes++;
 214                 } else {
 215                         break;  /* this is ok, we'll try again when more are needed */
 216                 }
 217         }
 218 }
 219
 220 static int set_bit_in_list_bitmap(struct super_block *sb,
 221                                   b_blocknr_t block,
 222                                   struct reiserfs_list_bitmap *jb)
 223 {
 224         unsigned int bmap_nr = block / (sb->s_blocksize << 3);
 225         unsigned int bit_nr = block % (sb->s_blocksize << 3);
 226
 227         if (!jb->bitmaps[bmap_nr]) {
 228                 jb->bitmaps[bmap_nr] = get_bitmap_node(sb);
 229         }
 230         set_bit(bit_nr, (unsigned long *)jb->bitmaps[bmap_nr]->data);
 231         return 0;
 232 }
 233
 234 static void cleanup_bitmap_list(struct super_block *sb,
 235                                 struct reiserfs_list_bitmap *jb)
 236 {
 237         int i;
 238         if (jb->bitmaps == NULL)
 239                 return;
 240
 241         for (i = 0; i < reiserfs_bmap_count(sb); i++) {
 242                 if (jb->bitmaps[i]) {
 243                         free_bitmap_node(sb, jb->bitmaps[i]);
 244                         jb->bitmaps[i] = NULL;
 245                 }
 246         }
 247 }
 248
 249 /*
 250 ** only call this on FS unmount.
 251 */
 252 static int free_list_bitmaps(struct super_block *sb,
 253                              struct reiserfs_list_bitmap *jb_array)
 254 {
 255         int i;
 256         struct reiserfs_list_bitmap *jb;
 257         for (i = 0; i < JOURNAL_NUM_BITMAPS; i++) {
 258                 jb = jb_array + i;
 259                 jb->journal_list = NULL;
 260                 cleanup_bitmap_list(sb, jb);
 261                 vfree(jb->bitmaps);
 262                 jb->bitmaps = NULL;
 263         }
 264         return 0;
 265 }
 266
 267 static int free_bitmap_nodes(struct super_block *sb)
 268 {
 269         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 270         struct list_head *next = journal->j_bitmap_nodes.next;
 271         struct reiserfs_bitmap_node *bn;
 272
 273         while (next != &journal->j_bitmap_nodes) {
 274                 bn = list_entry(next, struct reiserfs_bitmap_node, list);
 275                 list_del(next);
 276                 kfree(bn->data);
 277                 kfree(bn);
 278                 next = journal->j_bitmap_nodes.next;
 279                 journal->j_free_bitmap_nodes--;
 280         }
 281
 282         return 0;
 283 }
 284
 285 /*
 286 ** get memory for JOURNAL_NUM_BITMAPS worth of bitmaps.
 287 ** jb_array is the array to be filled in.
 288 */
 289 int reiserfs_allocate_list_bitmaps(struct super_block *sb,
 290                                    struct reiserfs_list_bitmap *jb_array,
 291                                    unsigned int bmap_nr)
 292 {
 293         int i;
 294         int failed = 0;
 295         struct reiserfs_list_bitmap *jb;
 296         int mem = bmap_nr * sizeof(struct reiserfs_bitmap_node *);
 297
 298         for (i = 0; i < JOURNAL_NUM_BITMAPS; i++) {
 299                 jb = jb_array + i;
 300                 jb->journal_list = NULL;
 301                 jb->bitmaps = vmalloc(mem);
 302                 if (!jb->bitmaps) {
 303                         reiserfs_warning(sb, "clm-2000", "unable to "
 304                                          "allocate bitmaps for journal lists");
 305                         failed = 1;
 306                         break;
 307                 }
 308                 memset(jb->bitmaps, 0, mem);
 309         }
 310         if (failed) {
 311                 free_list_bitmaps(sb, jb_array);
 312                 return -1;
 313         }
 314         return 0;
 315 }
 316
 317 /*
 318 ** find an available list bitmap.  If you can't find one, flush a commit list
 319 ** and try again
 320 */
 321 static struct reiserfs_list_bitmap *get_list_bitmap(struct super_block *sb,
 322                                                     struct reiserfs_journal_list
 323                                                     *jl)
 324 {
 325         int i, j;
 326         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 327         struct reiserfs_list_bitmap *jb = NULL;
 328
 329         for (j = 0; j < (JOURNAL_NUM_BITMAPS * 3); j++) {
 330                 i = journal->j_list_bitmap_index;
 331                 journal->j_list_bitmap_index = (i + 1) % JOURNAL_NUM_BITMAPS;
 332                 jb = journal->j_list_bitmap + i;
 333                 if (journal->j_list_bitmap[i].journal_list) {
 334                         flush_commit_list(sb,
 335                                           journal->j_list_bitmap[i].
 336                                           journal_list, 1);
 337                         if (!journal->j_list_bitmap[i].journal_list) {
 338                                 break;
 339                         }
 340                 } else {
 341                         break;
 342                 }
 343         }
 344         if (jb->journal_list) { /* double check to make sure if flushed correctly */
 345                 return NULL;
 346         }
 347         jb->journal_list = jl;
 348         return jb;
 349 }
 350
 351 /*
 352 ** allocates a new chunk of X nodes, and links them all together as a list.
 353 ** Uses the cnode->next and cnode->prev pointers
 354 ** returns NULL on failure
 355 */
 356 static struct reiserfs_journal_cnode *allocate_cnodes(int num_cnodes)
 357 {
 358         struct reiserfs_journal_cnode *head;
 359         int i;
 360         if (num_cnodes <= 0) {
 361                 return NULL;
 362         }
 363         head = vmalloc(num_cnodes * sizeof(struct reiserfs_journal_cnode));
 364         if (!head) {
 365                 return NULL;
 366         }
 367         memset(head, 0, num_cnodes * sizeof(struct reiserfs_journal_cnode));
 368         head[0].prev = NULL;
 369         head[0].next = head + 1;
 370         for (i = 1; i < num_cnodes; i++) {
 371                 head[i].prev = head + (i - 1);
 372                 head[i].next = head + (i + 1);  /* if last one, overwrite it after the if */
 373         }
 374         head[num_cnodes - 1].next = NULL;
 375         return head;
 376 }
 377
 378 /*
 379 ** pulls a cnode off the free list, or returns NULL on failure
 380 */
 381 static struct reiserfs_journal_cnode *get_cnode(struct super_block *sb)
 382 {
 383         struct reiserfs_journal_cnode *cn;
 384         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 385
 386         reiserfs_check_lock_depth(sb, "get_cnode");
 387
 388         if (journal->j_cnode_free <= 0) {
 389                 return NULL;
 390         }
 391         journal->j_cnode_used++;
 392         journal->j_cnode_free--;
 393         cn = journal->j_cnode_free_list;
 394         if (!cn) {
 395                 return cn;
 396         }
 397         if (cn->next) {
 398                 cn->next->prev = NULL;
 399         }
 400         journal->j_cnode_free_list = cn->next;
 401         memset(cn, 0, sizeof(struct reiserfs_journal_cnode));
 402         return cn;
 403 }
 404
 405 /*
 406 ** returns a cnode to the free list
 407 */
 408 static void free_cnode(struct super_block *sb,
 409                        struct reiserfs_journal_cnode *cn)
 410 {
 411         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 412
 413         reiserfs_check_lock_depth(sb, "free_cnode");
 414
 415         journal->j_cnode_used--;
 416         journal->j_cnode_free++;
 417         /* memset(cn, 0, sizeof(struct reiserfs_journal_cnode)) ; */
 418         cn->next = journal->j_cnode_free_list;
 419         if (journal->j_cnode_free_list) {
 420                 journal->j_cnode_free_list->prev = cn;
 421         }
 422         cn->prev = NULL;        /* not needed with the memset, but I might kill the memset, and forget to do this */
 423         journal->j_cnode_free_list = cn;
 424 }
 425
 426 static void clear_prepared_bits(struct buffer_head *bh)
 427 {
 428         clear_buffer_journal_prepared(bh);
 429         clear_buffer_journal_restore_dirty(bh);
 430 }
 431
 432 /* return a cnode with same dev, block number and size in table, or null if not found */
 433 static inline struct reiserfs_journal_cnode *get_journal_hash_dev(struct
 434                                                                   super_block
 435                                                                   *sb,
 436                                                                   struct
 437                                                                   reiserfs_journal_cnode
 438                                                                   **table,
 439                                                                   long bl)
 440 {
 441         struct reiserfs_journal_cnode *cn;
 442         cn = journal_hash(table, sb, bl);
 443         while (cn) {
 444                 if (cn->blocknr == bl && cn->sb == sb)
 445                         return cn;
 446                 cn = cn->hnext;
 447         }
 448         return (struct reiserfs_journal_cnode *)0;
 449 }
 450
 451 /*
 452 ** this actually means 'can this block be reallocated yet?'.  If you set search_all, a block can only be allocated
 453 ** if it is not in the current transaction, was not freed by the current transaction, and has no chance of ever
 454 ** being overwritten by a replay after crashing.
 455 **
 456 ** If you don't set search_all, a block can only be allocated if it is not in the current transaction.  Since deleting
 457 ** a block removes it from the current transaction, this case should never happen.  If you don't set search_all, make
 458 ** sure you never write the block without logging it.
 459 **
 460 ** next_zero_bit is a suggestion about the next block to try for find_forward.
 461 ** when bl is rejected because it is set in a journal list bitmap, we search
 462 ** for the next zero bit in the bitmap that rejected bl.  Then, we return that
 463 ** through next_zero_bit for find_forward to try.
 464 **
 465 ** Just because we return something in next_zero_bit does not mean we won't
 466 ** reject it on the next call to reiserfs_in_journal
 467 **
 468 */
 469 int reiserfs_in_journal(struct super_block *sb,
 470                         unsigned int bmap_nr, int bit_nr, int search_all,
 471                         b_blocknr_t * next_zero_bit)
 472 {
 473         struct reiserfs_journal *journal = SB_JOURNAL(sb);
 474         struct reiserfs_journal_cnode *cn;
 475         struct reiserfs_list_bitmap *jb;
 476         int i;
 477         unsigned long bl;
 478
 479         *next_zero_bit = 0;     /* always start this at zero. */
 480
 481         PROC_INFO_INC(sb, journal.in_journal);
 482         /* If we aren't doing a search_all, this is a metablock, and it will be logged before use.
 483          ** if we crash before the transaction that freed it commits,  this transaction won't
 484          ** have committed either, and the block will never be written
 485          */
 486         if (search_all) {
 487                 for (i = 0; i < JOURNAL_NUM_BITMAPS; i++) {
 488                         PROC_INFO_INC(sb, journal.in_journal_bitmap);
 489                         jb = journal->j_list_bitmap + i;
 490                         if (jb->journal_list && jb->bitmaps[bmap_nr] &&
 491                             test_bit(bit_nr,
 492                                      (unsigned long *)jb->bitmaps[bmap_nr]->
 493                                      data)) {
 494                                 *next_zero_bit =
 495                                     find_next_zero_bit((unsigned long *)
 496                                                        (jb->bitmaps[bmap_nr]->
 497                                                         data),
 498                                                        sb->s_blocksize << 3,
 499                                                        bit_nr + 1);
 500                                 return 1;
 501                         }
 502                 }
 503         }
 504
 505         bl = bmap_nr * (sb->s_blocksize << 3) + bit_nr;
 506         /* is it in any old transactions? */
 507         if (search_all
 508             && (cn =
 509                 get_journal_hash_dev(sb, journal->j_list_hash_table, bl))) {
 510                 return 1;
 511         }
 512
 513         /* is it in the current transaction.  This should never happen */
 514         if ((cn = get_journal_hash_dev(sb, journal->j_hash_table, bl))) {
 515                 BUG();
 516                 return 1;
 517         }
 518
 519         PROC_INFO_INC(sb, journal.in_journal_reusable);
 520         /* safe for reuse */
 521         return 0;
 522 }
 523
 524 /* insert cn into table
 525 */
 526 static inline void insert_journal_hash(struct reiserfs_journal_cnode **table,
 527                                        struct reiserfs_journal_cnode *cn)
 528 {
 529         struct reiserfs_journal_cnode *cn_orig;
 530
 531         cn_orig = journal_hash(table, cn->sb, cn->blocknr);
 532         cn->hnext = cn_orig;
 533         cn->hprev = NULL;
 534         if (cn_orig) {
 535                 cn_orig->hprev = cn;
 536         }
 537         journal_hash(table, cn->sb, cn->blocknr) = cn;
 538 }
 539
 540 /* lock the current transaction */
 541 static inline void lock_journal(struct super_block *sb)
 542 {
 543         PROC_INFO_INC(sb, journal.lock_journal);
 544
 545         reiserfs_mutex_lock_safe(&SB_JOURNAL(sb)->j_mutex, sb);
 546 }
 547
 548 /* unlock the current transaction */
 549 static inline void unlock_journal(struct super_block *sb)
 550 {
 551         mutex_unlock(&SB_JOURNAL(sb)->j_mutex);
 552 }
 553
 554 static inline void get_journal_list(struct reiserfs_journal_list *jl)
 555 {
 556         jl->j_refcount++;
 557 }
 558
 559 static inline void put_journal_list(struct super_block *s,
 560                                     struct reiserfs_journal_list *jl)
 561 {
 562         if (jl->j_refcount < 1) {
 563                 reiserfs_panic(s, "journal-2", "trans id %u, refcount at %d",
 564                                jl->j_trans_id, jl->j_refcount);
 565         }
 566         if (--jl->j_refcount == 0)
 567                 kfree(jl);
 568 }
 569
 570 /*
 571 ** this used to be much more involved, and I'm keeping it just in case things get ugly again.
 572 ** it gets called by flush_commit_list, and cleans up any data stored about blocks freed during a
 573 ** transaction.
 574 */
 575 static void cleanup_freed_for_journal_list(struct super_block *sb,
 576                                            struct reiserfs_journal_list *jl)
 577 {
 578
 579         struct reiserfs_list_bitmap *jb = jl->j_list_bitmap;
 580         if (jb) {
 581                 cleanup_bitmap_list(sb, jb);
 582         }
 583         jl->j_list_bitmap->journal_list = NULL;
 584         jl->j_list_bitmap = NULL;
 585 }
 586
 587 static int journal_list_still_alive(struct super_block *s,
 588                                     unsigned int trans_id)
 589 {
 590         struct reiserfs_journal *journal = SB_JOURNAL(s);
 591         struct list_head *entry = &journal->j_journal_list;
 592         struct reiserfs_journal_list *jl;
 593
 594         if (!list_empty(entry)) {
 595                 jl = JOURNAL_LIST_ENTRY(entry->next);
 596                 if (jl->j_trans_id <= trans_id) {
 597                         return 1;
 598                 }
 599         }
 600         return 0;
 601 }
 602
 603 /*
 604  * If page->mapping was null, we failed to truncate this page for
 605  * some reason.  Most likely because it was truncated after being
 606  * logged via data=journal.
 607  *
 608  * This does a check to see if the buffer belongs to one of these
 609  * lost pages before doing the final put_bh.  If page->mapping was
 610  * null, it tries to free buffers on the page, which should make the
 611  * final page_cache_release drop the page from the lru.
 612  */
 613 static void release_buffer_page(struct buffer_head *bh)
 614 {
 615         struct page *page = bh->b_page;
 616         if (!page->mapping && trylock_page(page)) {
 617                 page_cache_get(page);
 618                 put_bh(bh);
 619                 if (!page->mapping)
 620                         try_to_free_buffers(page);
 621                 unlock_page(page);
 622                 page_cache_release(page);
 623         } else {
 624                 put_bh(bh);
 625         }
 626 }
 627
 628 static void reiserfs_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
 629 {
 630         char b[BDEVNAME_SIZE];
 631
 632         if (buffer_journaled(bh)) {
 633                 reiserfs_warning(NULL, "clm-2084",
 634                                  "pinned buffer %lu:%s sent to disk",
 635                                  bh->b_blocknr, bdevname(bh->b_bdev, b));
 636         }
 637         if (uptodate)
 638                 set_buffer_uptodate(bh);
 639         else
 640                 clear_buffer_uptodate(bh);
 641
 642         unlock_buffer(bh);
 643         release_buffer_page(bh);
 644 }
 645
 646 static void reiserfs_end_ordered_io(struct buffer_head *bh, int uptodate)
 647 {
 648         if (uptodate)
 649                 set_buffer_uptodate(bh);
 650         else
 651                 clear_buffer_uptodate(bh);
 652         unlock_buffer(bh);
 653         put_bh(bh);
 654 }
 655
 656 static void submit_logged_buffer(struct buffer_head *bh)
 657 {
 658         get_bh(bh);
 659         bh->b_end_io = reiserfs_end_buffer_io_sync;
 660         clear_buffer_journal_new(bh);
 661         clear_buffer_dirty(bh);
 662         if (!test_clear_buffer_journal_test(bh))
 663                 BUG();
 664         if (!buffer_uptodate(bh))
 665                 BUG();
 666         submit_bh(WRITE, bh);
 667 }
 668
 669 static void submit_ordered_buffer(struct buffer_head *bh)
 670 {
 671         get_bh(bh);
 672         bh->b_end_io = reiserfs_end_ordered_io;
 673         clear_buffer_dirty(bh);
 674         if (!buffer_uptodate(bh))
 675                 BUG();
 676         submit_bh(WRITE, bh);
 677 }
 678
 679 static int submit_barrier_buffer(struct buffer_head *bh)
 680 {
 681         get_bh(bh);
 682         bh->b_end_io = reiserfs_end_ordered_io;
 683         clear_buffer_dirty(bh);
 684         if (!buffer_uptodate(bh))
 685                 BUG();
 686         return submit_bh(WRITE_BARRIER, bh);
 687 }
 688
 689 static void check_barrier_completion(struct super_block *s,
 690                                      struct buffer_head *bh)
 691 {
 692         if (buffer_eopnotsupp(bh)) {
 693                 clear_buffer_eopnotsupp(bh);
 694                 disable_barrier(s);
 695                 set_buffer_uptodate(bh);
 696                 set_buffer_dirty(bh);
 697                 reiserfs_write_unlock(s);
 698                 sync_dirty_buffer(bh);
 699                 reiserfs_write_lock(s);
 700         }
 701 }
 702
 703 #define CHUNK_SIZE 32
 704 struct buffer_chunk {
 705         struct buffer_head *bh[CHUNK_SIZE];
 706         int nr;
 707 };
 708
 709 static void write_chunk(struct buffer_chunk *chunk)
 710 {
 711         int i;
 712         get_fs_excl();
 713         for (i = 0; i < chunk->nr; i++) {
 714                 submit_logged_buffer(chunk->bh[i]);
 715         }
 716         chunk->nr = 0;
 717         put_fs_excl();
 718 }
 719
 720 static void write_ordered_chunk(struct buffer_chunk *chunk)
 721 {
 722         int i;
 723         get_fs_excl();
 724         for (i = 0; i < chunk->nr; i++) {
 725                 submit_ordered_buffer(chunk->bh[i]);
 726         }
 727         chunk->nr = 0;
 728         put_fs_excl();
 729 }
 730
 731 static int add_to_chunk(struct buffer_chunk *chunk, struct buffer_head *bh,
 732                         spinlock_t * lock, void (fn) (struct buffer_chunk *))
 733 {
 734         int ret = 0;
 735         BUG_ON(chunk->nr >= CHUNK_SIZE);
 736         chunk->bh[chunk->nr++] = bh;
 737         if (chunk->nr >= CHUNK_SIZE) {
 738                 ret = 1;
 739                 if (lock)
 740                         spin_unlock(lock);
 741                 fn(chunk);
 742                 if (lock)
 743                         spin_lock(lock);
 744         }
 745         return ret;
 746 }
 747
 748 static atomic_t nr_reiserfs_jh = ATOMIC_INIT(0);
 749 static struct reiserfs_jh *alloc_jh(void)
 750 {
 751         struct reiserfs_jh *jh;
 752         while (1) {
 753                 jh = kmalloc(sizeof(*jh), GFP_NOFS);
 754                 if (jh) {
 755                         atomic_inc(&nr_reiserfs_jh);
 756                         return jh;
 757                 }
 758                 yield();
 759         }
 760 }
 761
 762 /*
 763  * we want to free the jh when the buffer has been written
 764  * and waited on
 765  */
 766 void reiserfs_free_jh(struct buffer_head *bh)
 767 {
 768         struct reiserfs_jh *jh;
 769
 770         jh = bh->b_private;
 771         if (jh) {
 772                 bh->b_private = NULL;
 773                 jh->bh = NULL;
 774                 list_del_init(&jh->list);
 775                 kfree(jh);
 776                 if (atomic_read(&nr_reiserfs_jh) <= 0)
 777                         BUG();
 778                 atomic_dec(&nr_reiserfs_jh);
 779                 put_bh(bh);
 780         }
 781 }
 782
 783 static inline int __add_jh(struct reiserfs_journal *j, struct buffer_head *bh,
 784                            int tail)
 785 {
 786         struct reiserfs_jh *jh;
 787
 788         if (bh->b_private) {
 789                 spin_lock(&j->j_dirty_buffers_lock);
 790                 if (!bh->b_private) {
 791                         spin_unlock(&j->j_dirty_buffers_lock);
 792                         goto no_jh;
 793                 }
 794                 jh = bh->b_private;
 795                 list_del_init(&jh->list);
 796         } else {
 797               no_jh:
 798                 get_bh(bh);
 799                 jh = alloc_jh();
 800                 spin_lock(&j->j_dirty_buffers_lock);
 801                 /* buffer must be locked for __add_jh, should be able to have
 802                  * two adds at the same time
 803                  */
 804                 BUG_ON(bh->b_private);
 805                 jh->bh = bh;
 806                 bh->b_private = jh;
 807         }
 808         jh->jl = j->j_current_jl;
 809         if (tail)
 810                 list_add_tail(&jh->list, &jh->jl->j_tail_bh_list);
 811         else {
 812                 list_add_tail(&jh->list, &jh->jl->j_bh_list);
 813         }
 814         spin_unlock(&j->j_dirty_buffers_lock);
 815         return 0;
 816 }
 817
 818 int reiserfs_add_tail_list(struct inode *inode, struct buffer_head *bh)
 819 {
 820         return __add_jh(SB_JOURNAL(inode->i_sb), bh, 1);
 821 }
 822 int reiserfs_add_ordered_list(struct inode *inode, struct buffer_head *bh)
 823 {
 824         return __add_jh(SB_JOURNAL(inode->i_sb), bh, 0);
 825 }
 826
 827 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list)
 828 static int write_ordered_buffers(spinlock_t * lock,
 829                                  struct reiserfs_journal *j,
 830                                  struct reiserfs_journal_list *jl,
 831                                  struct list_head *list)
 832 {
 833         struct buffer_head *bh;
 834         struct reiserfs_jh *jh;
 835         int ret = j->j_errno;
 836         struct buffer_chunk chunk;
 837         struct list_head tmp;
 838         INIT_LIST_HEAD(&tmp);
 839
 840         chunk.nr = 0;
 841         spin_lock(lock);
 842         while (!list_empty(list)) {
 843                 jh = JH_ENTRY(list->next);
 844                 bh = jh->bh;
 845                 get_bh(bh);
 846                 if (!trylock_buffer(bh)) {
 847                         if (!buffer_dirty(bh)) {
 848                                 list_move(&jh->list, &tmp);
 849                                 goto loop_next;
 850                         }
 851                         spin_unlock(lock);
 852                         if (chunk.nr)
 853                                 write_ordered_chunk(&chunk);
 854                         wait_on_buffer(bh);
 855                         cond_resched();
 856                         spin_lock(lock);
 857                         goto loop_next;
 858                 }
 859                 /* in theory, dirty non-uptodate buffers should never get here,
 860                  * but the upper layer io error paths still have a few quirks.
 861                  * Handle them here as gracefully as we can
 862                  */
 863                 if (!buffer_uptodate(bh) && buffer_dirty(bh)) {
 864                         clear_buffer_dirty(bh);
 865                         ret = -EIO;
 866                 }
 867                 if (buffer_dirty(bh)) {
 868                         list_move(&jh->list, &tmp);
 869                         add_to_chunk(&chunk, bh, lock, write_ordered_chunk);
 870                 } else {
 871                         reiserfs_free_jh(bh);
 872                         unlock_buffer(bh);
 873                 }
 874               loop_next:
 875                 put_bh(bh);
 876                 cond_resched_lock(lock);
 877         }
 878         if (chunk.nr) {
 879                 spin_unlock(lock);
 880                 write_ordered_chunk(&chunk);
 881                 spin_lock(lock);
 882         }
 883         while (!list_empty(&tmp)) {
 884                 jh = JH_ENTRY(tmp.prev);
 885                 bh = jh->bh;
 886                 get_bh(bh);
 887                 reiserfs_free_jh(bh);
 888
 889                 if (buffer_locked(bh)) {
 890                         spin_unlock(lock);
 891                         wait_on_buffer(bh);
 892                         spin_lock(lock);
 893                 }
 894                 if (!buffer_uptodate(bh)) {
 895                         ret = -EIO;
 896                 }
 897                 /* ugly interaction with invalidatepage here.
 898                  * reiserfs_invalidate_page will pin any buffer that has a valid
 899                  * journal head from an older transaction.  If someone else sets
 900                  * our buffer dirty after we write it in the first loop, and
 901                  * then someone truncates the page away, nobody will ever write
 902                  * the buffer. We're safe if we write the page one last time
 903                  * after freeing the journal header.
 904                  */
 905                 if (buffer_dirty(bh) && unlikely(bh->b_page->mapping == NULL)) {
 906                         spin_unlock(lock);
 907                         ll_rw_block(WRITE, 1, &bh);
 908                         spin_lock(lock);
 909                 }
 910                 put_bh(bh);
 911                 cond_resched_lock(lock);
 912         }
 913         spin_unlock(lock);
 914         return ret;
 915 }
 916
 917 static int flush_older_commits(struct super_block *s,
 918                                struct reiserfs_journal_list *jl)
 919 {
 920         struct reiserfs_journal *journal = SB_JOURNAL(s);
 921         struct reiserfs_journal_list *other_jl;
 922         struct reiserfs_journal_list *first_jl;
 923         struct list_head *entry;
 924         unsigned int trans_id = jl->j_trans_id;
 925         unsigned int other_trans_id;
 926         unsigned int first_trans_id;
 927
 928       find_first:
 929         /*
 930          * first we walk backwards to find the oldest uncommitted transation
 931          */
 932         first_jl = jl;
 933         entry = jl->j_list.prev;
 934         while (1) {
 935                 other_jl = JOURNAL_LIST_ENTRY(entry);
 936                 if (entry == &journal->j_journal_list ||
 937                     atomic_read(&other_jl->j_older_commits_done))
 938                         break;
 939
 940                 first_jl = other_jl;
 941                 entry = other_jl->j_list.prev;
 942         }
 943
 944         /* if we didn't find any older uncommitted transactions, return now */
 945         if (first_jl == jl) {
 946                 return 0;
 947         }
 948
 949         first_trans_id = first_jl->j_trans_id;
 950
 951         entry = &first_jl->j_list;
 952         while (1) {
 953                 other_jl = JOURNAL_LIST_ENTRY(entry);
 954                 other_trans_id = other_jl->j_trans_id;
 955
 956                 if (other_trans_id < trans_id) {
 957                         if (atomic_read(&other_jl->j_commit_left) != 0) {
 958                                 flush_commit_list(s, other_jl, 0);
 959
 960                                 /* list we were called with is gone, return */
 961                                 if (!journal_list_still_alive(s, trans_id))
 962                                         return 1;
 963
 964                                 /* the one we just flushed is gone, this means all
 965                                  * older lists are also gone, so first_jl is no longer
 966                                  * valid either.  Go back to the beginning.
 967                                  */
 968                                 if (!journal_list_still_alive
 969                                     (s, other_trans_id)) {
 970                                         goto find_first;
 971                                 }
 972                         }
 973                         entry = entry->next;
 974                         if (entry == &journal->j_journal_list)
 975                                 return 0;
 976                 } else {
 977                         return 0;
 978                 }
 979         }
 980         return 0;
 981 }
 982
 983 static int reiserfs_async_progress_wait(struct super_block *s)
 984 {
 985         DEFINE_WAIT(wait);
 986         struct reiserfs_journal *j = SB_JOURNAL(s);
 987
 988         if (atomic_read(&j->j_async_throttle)) {
 989                 reiserfs_write_unlock(s);
 990                 congestion_wait(BLK_RW_ASYNC, HZ / 10);
 991                 reiserfs_write_lock(s);
 992         }
 993
 994         return 0;
 995 }
 996
 997 /*
 998 ** if this journal list still has commit blocks unflushed, send them to disk.
 999 **
1000 ** log areas must be flushed in order (transaction 2 can't commit before transaction 1)
1001 ** Before the commit block can by written, every other log block must be safely on disk
1002 **
1003 */
1004 static int flush_commit_list(struct super_block *s,
1005                              struct reiserfs_journal_list *jl, int flushall)
1006 {
1007         int i;
1008         b_blocknr_t bn;
1009         struct buffer_head *tbh = NULL;
1010         unsigned int trans_id = jl->j_trans_id;
1011         struct reiserfs_journal *journal = SB_JOURNAL(s);
1012         int barrier = 0;
1013         int retval = 0;
1014         int write_len;
1015
1016         reiserfs_check_lock_depth(s, "flush_commit_list");
1017
1018         if (atomic_read(&jl->j_older_commits_done)) {
1019                 return 0;
1020         }
1021
1022         get_fs_excl();
1023
1024         /* before we can put our commit blocks on disk, we have to make sure everyone older than
1025          ** us is on disk too
1026          */
1027         BUG_ON(jl->j_len <= 0);
1028         BUG_ON(trans_id == journal->j_trans_id);
1029
1030         get_journal_list(jl);
1031         if (flushall) {
1032                 if (flush_older_commits(s, jl) == 1) {
1033                         /* list disappeared during flush_older_commits.  return */
1034                         goto put_jl;
1035                 }
1036         }
1037
1038         /* make sure nobody is trying to flush this one at the same time */
1039         reiserfs_mutex_lock_safe(&jl->j_commit_mutex, s);
1040
1041         if (!journal_list_still_alive(s, trans_id)) {
1042                 mutex_unlock(&jl->j_commit_mutex);
1043                 goto put_jl;
1044         }
1045         BUG_ON(jl->j_trans_id == 0);
1046
1047         /* this commit is done, exit */
1048         if (atomic_read(&(jl->j_commit_left)) <= 0) {
1049                 if (flushall) {
1050                         atomic_set(&(jl->j_older_commits_done), 1);
1051                 }
1052                 mutex_unlock(&jl->j_commit_mutex);
1053                 goto put_jl;
1054         }
1055
1056         if (!list_empty(&jl->j_bh_list)) {
1057                 int ret;
1058
1059                 /*
1060                  * We might sleep in numerous places inside
1061                  * write_ordered_buffers. Relax the write lock.
1062                  */
1063                 reiserfs_write_unlock(s);
1064                 ret = write_ordered_buffers(&journal->j_dirty_buffers_lock,
1065                                             journal, jl, &jl->j_bh_list);
1066                 if (ret < 0 && retval == 0)
1067                         retval = ret;
1068                 reiserfs_write_lock(s);
1069         }
1070         BUG_ON(!list_empty(&jl->j_bh_list));
1071         /*
1072          * for the description block and all the log blocks, submit any buffers
1073          * that haven't already reached the disk.  Try to write at least 256
1074          * log blocks. later on, we will only wait on blocks that correspond
1075          * to this transaction, but while we're unplugging we might as well
1076          * get a chunk of data on there.
1077          */
1078         atomic_inc(&journal->j_async_throttle);
1079         write_len = jl->j_len + 1;
1080         if (write_len < 256)
1081                 write_len = 256;
1082         for (i = 0 ; i < write_len ; i++) {
1083                 bn = SB_ONDISK_JOURNAL_1st_BLOCK(s) + (jl->j_start + i) %
1084                     SB_ONDISK_JOURNAL_SIZE(s);
1085                 tbh = journal_find_get_block(s, bn);
1086                 if (tbh) {
1087                         if (buffer_dirty(tbh)) {
1088                             reiserfs_write_unlock(s);
1089                             ll_rw_block(WRITE, 1, &tbh);
1090                             reiserfs_write_lock(s);
1091                         }
1092                         put_bh(tbh) ;
1093                 }
1094         }
1095         atomic_dec(&journal->j_async_throttle);
1096
1097         /* We're skipping the commit if there's an error */
1098         if (retval || reiserfs_is_journal_aborted(journal))
1099                 barrier = 0;
1100
1101         /* wait on everything written so far before writing the commit
1102          * if we are in barrier mode, send the commit down now
1103          */
1104         barrier = reiserfs_barrier_flush(s);
1105         if (barrier) {
1106                 int ret;
1107                 lock_buffer(jl->j_commit_bh);
1108                 ret = submit_barrier_buffer(jl->j_commit_bh);
1109                 if (ret == -EOPNOTSUPP) {
1110                         set_buffer_uptodate(jl->j_commit_bh);
1111                         disable_barrier(s);
1112                         barrier = 0;
1113                 }
1114         }
1115         for (i = 0; i < (jl->j_len + 1); i++) {
1116                 bn = SB_ONDISK_JOURNAL_1st_BLOCK(s) +
1117                     (jl->j_start + i) % SB_ONDISK_JOURNAL_SIZE(s);
1118                 tbh = journal_find_get_block(s, bn);
1119
1120                 reiserfs_write_unlock(s);
1121                 wait_on_buffer(tbh);
1122                 reiserfs_write_lock(s);
1123                 // since we're using ll_rw_blk above, it might have skipped over
1124                 // a locked buffer.  Double check here
1125                 //
1126                 /* redundant, sync_dirty_buffer() checks */
1127                 if (buffer_dirty(tbh)) {
1128                         reiserfs_write_unlock(s);
1129                         sync_dirty_buffer(tbh);
1130                         reiserfs_write_lock(s);
1131                 }
1132                 if (unlikely(!buffer_uptodate(tbh))) {
1133 #ifdef CONFIG_REISERFS_CHECK
1134                         reiserfs_warning(s, "journal-601",
1135                                          "buffer write failed");
1136 #endif
1137                         retval = -EIO;
1138                 }
1139                 put_bh(tbh);    /* once for journal_find_get_block */
1140                 put_bh(tbh);    /* once due to original getblk in do_journal_end */
1141                 atomic_dec(&(jl->j_commit_left));
1142         }
1143
1144         BUG_ON(atomic_read(&(jl->j_commit_left)) != 1);
1145
1146         if (!barrier) {
1147                 /* If there was a write error in the journal - we can't commit
1148                  * this transaction - it will be invalid and, if successful,
1149                  * will just end up propagating the write error out to
1150                  * the file system. */
1151                 if (likely(!retval && !reiserfs_is_journal_aborted (journal))) {
1152                         if (buffer_dirty(jl->j_commit_bh))
1153                                 BUG();
1154                         mark_buffer_dirty(jl->j_commit_bh) ;
1155                         reiserfs_write_unlock(s);
1156                         sync_dirty_buffer(jl->j_commit_bh) ;
1157                         reiserfs_write_lock(s);
1158                 }
1159         } else {
1160                 reiserfs_write_unlock(s);
1161                 wait_on_buffer(jl->j_commit_bh);
1162                 reiserfs_write_lock(s);
1163         }
1164
1165         check_barrier_completion(s, jl->j_commit_bh);
1166
1167         /* If there was a write error in the journal - we can't commit this
1168          * transaction - it will be invalid and, if successful, will just end
1169          * up propagating the write error out to the filesystem. */
1170         if (unlikely(!buffer_uptodate(jl->j_commit_bh))) {
1171 #ifdef CONFIG_REISERFS_CHECK
1172                 reiserfs_warning(s, "journal-615", "buffer write failed");
1173 #endif
1174                 retval = -EIO;
1175         }
1176         bforget(jl->j_commit_bh);
1177         if (journal->j_last_commit_id != 0 &&
1178             (jl->j_trans_id - journal->j_last_commit_id) != 1) {
1179                 reiserfs_warning(s, "clm-2200", "last commit %lu, current %lu",
1180                                  journal->j_last_commit_id, jl->j_trans_id);
1181         }
1182         journal->j_last_commit_id = jl->j_trans_id;
1183
1184         /* now, every commit block is on the disk.  It is safe to allow blocks freed during this transaction to be reallocated */
1185         cleanup_freed_for_journal_list(s, jl);
1186
1187         retval = retval ? retval : journal->j_errno;
1188
1189         /* mark the metadata dirty */
1190         if (!retval)
1191                 dirty_one_transaction(s, jl);
1192         atomic_dec(&(jl->j_commit_left));
1193
1194         if (flushall) {
1195                 atomic_set(&(jl->j_older_commits_done), 1);
1196         }
1197         mutex_unlock(&jl->j_commit_mutex);
1198       put_jl:
1199         put_journal_list(s, jl);
1200
1201         if (retval)
1202                 reiserfs_abort(s, retval, "Journal write error in %s",
1203                                __func__);
1204         put_fs_excl();
1205         return retval;
1206 }
1207
1208 /*
1209 ** flush_journal_list frequently needs to find a newer transaction for a given block.  This does that, or
1210 ** returns NULL if it can't find anything
1211 */
1212 static struct reiserfs_journal_list *find_newer_jl_for_cn(struct
1213                                                           reiserfs_journal_cnode
1214                                                           *cn)
1215 {
1216         struct super_block *sb = cn->sb;
1217         b_blocknr_t blocknr = cn->blocknr;
1218
1219         cn = cn->hprev;
1220         while (cn) {
1221                 if (cn->sb == sb && cn->blocknr == blocknr && cn->jlist) {
1222                         return cn->jlist;
1223                 }
1224                 cn = cn->hprev;
1225         }
1226         return NULL;
1227 }
1228
1229 static int newer_jl_done(struct reiserfs_journal_cnode *cn)
1230 {
1231         struct super_block *sb = cn->sb;
1232         b_blocknr_t blocknr = cn->blocknr;
1233
1234         cn = cn->hprev;
1235         while (cn) {
1236                 if (cn->sb == sb && cn->blocknr == blocknr && cn->jlist &&
1237                     atomic_read(&cn->jlist->j_commit_left) != 0)
1238                                     return 0;
1239                 cn = cn->hprev;
1240         }
1241         return 1;
1242 }
1243
1244 static void remove_journal_hash(struct super_block *,
1245                                 struct reiserfs_journal_cnode **,
1246                                 struct reiserfs_journal_list *, unsigned long,
1247                                 int);
1248
1249 /*
1250 ** once all the real blocks have been flushed, it is safe to remove them from the
1251 ** journal list for this transaction.  Aside from freeing the cnode, this also allows the
1252 ** block to be reallocated for data blocks if it had been deleted.
1253 */
1254 static void remove_all_from_journal_list(struct super_block *sb,
1255                                          struct reiserfs_journal_list *jl,
1256                                          int debug)
1257 {
1258         struct reiserfs_journal *journal = SB_JOURNAL(sb);
1259         struct reiserfs_journal_cnode *cn, *last;
1260         cn = jl->j_realblock;
1261
1262         /* which is better, to lock once around the whole loop, or
1263          ** to lock for each call to remove_journal_hash?
1264          */
1265         while (cn) {
1266                 if (cn->blocknr != 0) {
1267                         if (debug) {
1268                                 reiserfs_warning(sb, "reiserfs-2201",
1269                                                  "block %u, bh is %d, state %ld",
1270                                                  cn->blocknr, cn->bh ? 1 : 0,
1271                                                  cn->state);
1272                         }
1273                         cn->state = 0;
1274                         remove_journal_hash(sb, journal->j_list_hash_table,
1275                                             jl, cn->blocknr, 1);
1276                 }
1277                 last = cn;
1278                 cn = cn->next;
1279                 free_cnode(sb, last);
1280         }
1281         jl->j_realblock = NULL;
1282 }
1283
1284 /*
1285 ** if this timestamp is greater than the timestamp we wrote last to the header block, write it to the header block.
1286 ** once this is done, I can safely say the log area for this transaction won't ever be replayed, and I can start
1287 ** releasing blocks in this transaction for reuse as data blocks.
1288 ** called by flush_journal_list, before it calls remove_all_from_journal_list
1289 **
1290 */
1291 static int _update_journal_header_block(struct super_block *sb,
1292                                         unsigned long offset,
1293                                         unsigned int trans_id)
1294 {
1295         struct reiserfs_journal_header *jh;
1296         struct reiserfs_journal *journal = SB_JOURNAL(sb);
1297
1298         if (reiserfs_is_journal_aborted(journal))
1299                 return -EIO;
1300
1301         if (trans_id >= journal->j_last_flush_trans_id) {
1302                 if (buffer_locked((journal->j_header_bh))) {
1303                         reiserfs_write_unlock(sb);
1304                         wait_on_buffer((journal->j_header_bh));
1305                         reiserfs_write_lock(sb);
1306                         if (unlikely(!buffer_uptodate(journal->j_header_bh))) {
1307 #ifdef CONFIG_REISERFS_CHECK
1308                                 reiserfs_warning(sb, "journal-699",
1309                                                  "buffer write failed");
1310 #endif
1311                                 return -EIO;
1312                         }
1313                 }
1314                 journal->j_last_flush_trans_id = trans_id;
1315                 journal->j_first_unflushed_offset = offset;
1316                 jh = (struct reiserfs_journal_header *)(journal->j_header_bh->
1317                                                         b_data);
1318                 jh->j_last_flush_trans_id = cpu_to_le32(trans_id);
1319                 jh->j_first_unflushed_offset = cpu_to_le32(offset);
1320                 jh->j_mount_id = cpu_to_le32(journal->j_mount_id);
1321
1322                 if (reiserfs_barrier_flush(sb)) {
1323                         int ret;
1324                         lock_buffer(journal->j_header_bh);
1325                         ret = submit_barrier_buffer(journal->j_header_bh);
1326                         if (ret == -EOPNOTSUPP) {
1327                                 set_buffer_uptodate(journal->j_header_bh);
1328                                 disable_barrier(sb);
1329                                 goto sync;
1330                         }
1331                         reiserfs_write_unlock(sb);
1332                         wait_on_buffer(journal->j_header_bh);
1333                         reiserfs_write_lock(sb);
1334                         check_barrier_completion(sb, journal->j_header_bh);
1335                 } else {
1336                       sync:
1337                         set_buffer_dirty(journal->j_header_bh);
1338                         reiserfs_write_unlock(sb);
1339                         sync_dirty_buffer(journal->j_header_bh);
1340                         reiserfs_write_lock(sb);
1341                 }
1342                 if (!buffer_uptodate(journal->j_header_bh)) {
1343                         reiserfs_warning(sb, "journal-837",
1344                                          "IO error during journal replay");
1345                         return -EIO;
1346                 }
1347         }
1348         return 0;
1349 }
1350
1351 static int update_journal_header_block(struct super_block *sb,
1352                                        unsigned long offset,
1353                                        unsigned int trans_id)
1354 {
1355         return _update_journal_header_block(sb, offset, trans_id);
1356 }
1357
1358 /*
1359 ** flush any and all journal lists older than you are
1360 ** can only be called from flush_journal_list
1361 */
1362 static int flush_older_journal_lists(struct super_block *sb,
1363                                      struct reiserfs_journal_list *jl)
1364 {
1365         struct list_head *entry;
1366         struct reiserfs_journal_list *other_jl;
1367         struct reiserfs_journal *journal = SB_JOURNAL(sb);
1368         unsigned int trans_id = jl->j_trans_id;
1369
1370         /* we know we are the only ones flushing things, no extra race
1371          * protection is required.
1372          */
1373       restart:
1374         entry = journal->j_journal_list.next;
1375         /* Did we wrap? */
1376         if (entry == &journal->j_journal_list)
1377                 return 0;
1378         other_jl = JOURNAL_LIST_ENTRY(entry);
1379         if (other_jl->j_trans_id < trans_id) {
1380                 BUG_ON(other_jl->j_refcount <= 0);
1381                 /* do not flush all */
1382                 flush_journal_list(sb, other_jl, 0);
1383
1384                 /* other_jl is now deleted from the list */
1385                 goto restart;
1386         }
1387         return 0;
1388 }
1389
1390 static void del_from_work_list(struct super_block *s,
1391                                struct reiserfs_journal_list *jl)
1392 {
1393         struct reiserfs_journal *journal = SB_JOURNAL(s);
1394         if (!list_empty(&jl->j_working_list)) {
1395                 list_del_init(&jl->j_working_list);
1396                 journal->j_num_work_lists--;
1397         }
1398 }
1399
1400 /* flush a journal list, both commit and real blocks
1401 **
1402 ** always set flushall to 1, unless you are calling from inside
1403 ** flush_journal_list
1404 **
1405 ** IMPORTANT.  This can only be called while there are no journal writers,
1406 ** and the journal is locked.  That means it can only be called from
1407 ** do_journal_end, or by journal_release
1408 */
1409 static int flush_journal_list(struct super_block *s,
1410                               struct reiserfs_journal_list *jl, int flushall)
1411 {
1412         struct reiserfs_journal_list *pjl;
1413         struct reiserfs_journal_cnode *cn, *last;
1414         int count;
1415         int was_jwait = 0;
1416         int was_dirty = 0;
1417         struct buffer_head *saved_bh;
1418         unsigned long j_len_saved = jl->j_len;
1419         struct reiserfs_journal *journal = SB_JOURNAL(s);
1420         int err = 0;
1421
1422         BUG_ON(j_len_saved <= 0);
1423
1424         if (atomic_read(&journal->j_wcount) != 0) {
1425                 reiserfs_warning(s, "clm-2048", "called with wcount %d",
1426                                  atomic_read(&journal->j_wcount));
1427         }
1428         BUG_ON(jl->j_trans_id == 0);
1429
1430         /* if flushall == 0, the lock is already held */
1431         if (flushall) {
1432                 reiserfs_mutex_lock_safe(&journal->j_flush_mutex, s);
1433         } else if (mutex_trylock(&journal->j_flush_mutex)) {
1434                 BUG();
1435         }
1436
1437         count = 0;
1438         if (j_len_saved > journal->j_trans_max) {
1439                 reiserfs_panic(s, "journal-715", "length is %lu, trans id %lu",
1440                                j_len_saved, jl->j_trans_id);
1441                 return 0;
1442         }
1443
1444         get_fs_excl();
1445
1446         /* if all the work is already done, get out of here */
1447         if (atomic_read(&(jl->j_nonzerolen)) <= 0 &&
1448             atomic_read(&(jl->j_commit_left)) <= 0) {
1449                 goto flush_older_and_return;
1450         }
1451
1452         /* start by putting the commit list on disk.  This will also flush
1453          ** the commit lists of any olders transactions
1454          */
1455         flush_commit_list(s, jl, 1);
1456
1457         if (!(jl->j_state & LIST_DIRTY)
1458             && !reiserfs_is_journal_aborted(journal))
1459                 BUG();
1460
1461         /* are we done now? */
1462         if (atomic_read(&(jl->j_nonzerolen)) <= 0 &&
1463             atomic_read(&(jl->j_commit_left)) <= 0) {
1464                 goto flush_older_and_return;
1465         }
1466
1467         /* loop through each cnode, see if we need to write it,
1468          ** or wait on a more recent transaction, or just ignore it
1469          */
1470         if (atomic_read(&(journal->j_wcount)) != 0) {
1471                 reiserfs_panic(s, "journal-844", "journal list is flushing, "
1472                                "wcount is not 0");
1473         }
1474         cn = jl->j_realblock;
1475         while (cn) {
1476                 was_jwait = 0;
1477                 was_dirty = 0;
1478                 saved_bh = NULL;
1479                 /* blocknr of 0 is no longer in the hash, ignore it */
1480                 if (cn->blocknr == 0) {
1481                         goto free_cnode;
1482                 }
1483
1484                 /* This transaction failed commit. Don't write out to the disk */
1485                 if (!(jl->j_state & LIST_DIRTY))
1486                         goto free_cnode;
1487
1488                 pjl = find_newer_jl_for_cn(cn);
1489                 /* the order is important here.  We check pjl to make sure we
1490                  ** don't clear BH_JDirty_wait if we aren't the one writing this
1491                  ** block to disk
1492                  */
1493                 if (!pjl && cn->bh) {
1494                         saved_bh = cn->bh;
1495
1496                         /* we do this to make sure nobody releases the buffer while
1497                          ** we are working with it
1498                          */
1499                         get_bh(saved_bh);
1500
1501                         if (buffer_journal_dirty(saved_bh)) {
1502                                 BUG_ON(!can_dirty(cn));
1503                                 was_jwait = 1;
1504                                 was_dirty = 1;
1505                         } else if (can_dirty(cn)) {
1506                                 /* everything with !pjl && jwait should be writable */
1507                                 BUG();
1508                         }
1509                 }
1510
1511                 /* if someone has this block in a newer transaction, just make
1512                  ** sure they are committed, and don't try writing it to disk
1513                  */
1514                 if (pjl) {
1515                         if (atomic_read(&pjl->j_commit_left))
1516                                 flush_commit_list(s, pjl, 1);
1517                         goto free_cnode;
1518                 }
1519
1520                 /* bh == NULL when the block got to disk on its own, OR,
1521                  ** the block got freed in a future transaction
1522                  */
1523                 if (saved_bh == NULL) {
1524                         goto free_cnode;
1525                 }
1526
1527                 /* this should never happen.  kupdate_one_transaction has this list
1528                  ** locked while it works, so we should never see a buffer here that
1529                  ** is not marked JDirty_wait
1530                  */
1531                 if ((!was_jwait) && !buffer_locked(saved_bh)) {
1532                         reiserfs_warning(s, "journal-813",
1533                                          "BAD! buffer %llu %cdirty %cjwait, "
1534                                          "not in a newer tranasction",
1535                                          (unsigned long long)saved_bh->
1536                                          b_blocknr, was_dirty ? ' ' : '!',
1537                                          was_jwait ? ' ' : '!');
1538                 }
1539                 if (was_dirty) {
1540                         /* we inc again because saved_bh gets decremented at free_cnode */
1541                         get_bh(saved_bh);
1542                         set_bit(BLOCK_NEEDS_FLUSH, &cn->state);
1543                         lock_buffer(saved_bh);
1544                         BUG_ON(cn->blocknr != saved_bh->b_blocknr);
1545                         if (buffer_dirty(saved_bh))
1546                                 submit_logged_buffer(saved_bh);
1547                         else
1548                                 unlock_buffer(saved_bh);
1549                         count++;
1550                 } else {
1551                         reiserfs_warning(s, "clm-2082",
1552                                          "Unable to flush buffer %llu in %s",
1553                                          (unsigned long long)saved_bh->
1554                                          b_blocknr, __func__);
1555                 }
1556               free_cnode:
1557                 last = cn;
1558                 cn = cn->next;
1559                 if (saved_bh) {
1560                         /* we incremented this to keep others from taking the buffer head away */
1561                         put_bh(saved_bh);
1562                         if (atomic_read(&(saved_bh->b_count)) < 0) {
1563                                 reiserfs_warning(s, "journal-945",
1564                                                  "saved_bh->b_count < 0");
1565                         }
1566                 }
1567         }
1568         if (count > 0) {
1569                 cn = jl->j_realblock;
1570                 while (cn) {
1571                         if (test_bit(BLOCK_NEEDS_FLUSH, &cn->state)) {
1572                                 if (!cn->bh) {
1573                                         reiserfs_panic(s, "journal-1011",
1574                                                        "cn->bh is NULL");
1575                                 }
1576
1577                                 reiserfs_write_unlock(s);
1578                                 wait_on_buffer(cn->bh);
1579                                 reiserfs_write_lock(s);
1580
1581                                 if (!cn->bh) {
1582                                         reiserfs_panic(s, "journal-1012",
1583                                                        "cn->bh is NULL");
1584                                 }
1585                                 if (unlikely(!buffer_uptodate(cn->bh))) {
1586 #ifdef CONFIG_REISERFS_CHECK
1587                                         reiserfs_warning(s, "journal-949",
1588                                                          "buffer write failed");
1589 #endif
1590                                         err = -EIO;
1591                                 }
1592                                 /* note, we must clear the JDirty_wait bit after the up to date
1593                                  ** check, otherwise we race against our flushpage routine
1594                                  */
1595                                 BUG_ON(!test_clear_buffer_journal_dirty
1596                                        (cn->bh));
1597
1598                                 /* drop one ref for us */
1599                                 put_bh(cn->bh);
1600                                 /* drop one ref for journal_mark_dirty */
1601                                 release_buffer_page(cn->bh);
1602                         }
1603                         cn = cn->next;
1604                 }
1605         }
1606
1607         if (err)
1608                 reiserfs_abort(s, -EIO,
1609                                "Write error while pushing transaction to disk in %s",
1610                                __func__);
1611       flush_older_and_return:
1612
1613         /* before we can update the journal header block, we _must_ flush all
1614          ** real blocks from all older transactions to disk.  This is because
1615          ** once the header block is updated, this transaction will not be
1616          ** replayed after a crash
1617          */
1618         if (flushall) {
1619                 flush_older_journal_lists(s, jl);
1620         }
1621
1622         err = journal->j_errno;
1623         /* before we can remove everything from the hash tables for this
1624          ** transaction, we must make sure it can never be replayed
1625          **
1626          ** since we are only called from do_journal_end, we know for sure there
1627          ** are no allocations going on while we are flushing journal lists.  So,
1628          ** we only need to update the journal header block for the last list
1629          ** being flushed
1630          */
1631         if (!err && flushall) {
1632                 err =
1633                     update_journal_header_block(s,
1634                                                 (jl->j_start + jl->j_len +
1635                                                  2) % SB_ONDISK_JOURNAL_SIZE(s),
1636                                                 jl->j_trans_id);
1637                 if (err)
1638                         reiserfs_abort(s, -EIO,
1639                                        "Write error while updating journal header in %s",
1640                                        __func__);
1641         }
1642         remove_all_from_journal_list(s, jl, 0);
1643         list_del_init(&jl->j_list);
1644         journal->j_num_lists--;
1645         del_from_work_list(s, jl);
1646
1647         if (journal->j_last_flush_id != 0 &&
1648             (jl->j_trans_id - journal->j_last_flush_id) != 1) {
1649                 reiserfs_warning(s, "clm-2201", "last flush %lu, current %lu",
1650                                  journal->j_last_flush_id, jl->j_trans_id);
1651         }
1652         journal->j_last_flush_id = jl->j_trans_id;
1653
1654         /* not strictly required since we are freeing the list, but it should
1655          * help find code using dead lists later on
1656          */
1657         jl->j_len = 0;
1658         atomic_set(&(jl->j_nonzerolen), 0);
1659         jl->j_start = 0;
1660         jl->j_realblock = NULL;
1661         jl->j_commit_bh = NULL;
1662         jl->j_trans_id = 0;
1663         jl->j_state = 0;
1664         put_journal_list(s, jl);
1665         if (flushall)
1666                 mutex_unlock(&journal->j_flush_mutex);
1667         put_fs_excl();
1668         return err;
1669 }
1670
1671 static int test_transaction(struct super_block *s,
1672                             struct reiserfs_journal_list *jl)
1673 {
1674         struct reiserfs_journal_cnode *cn;
1675
1676         if (jl->j_len == 0 || atomic_read(&jl->j_nonzerolen) == 0)
1677                 return 1;
1678
1679         cn = jl->j_realblock;
1680         while (cn) {
1681                 /* if the blocknr == 0, this has been cleared from the hash,
1682                  ** skip it
1683                  */
1684                 if (cn->blocknr == 0) {
1685                         goto next;
1686                 }
1687                 if (cn->bh && !newer_jl_done(cn))
1688                         return 0;
1689               next:
1690                 cn = cn->next;
1691                 cond_resched();
1692         }
1693         return 0;
1694 }
1695
1696 static int write_one_transaction(struct super_block *s,
1697                                  struct reiserfs_journal_list *jl,
1698                                  struct buffer_chunk *chunk)
1699 {
1700         struct reiserfs_journal_cnode *cn;
1701         int ret = 0;
1702
1703         jl->j_state |= LIST_TOUCHED;
1704         del_from_work_list(s, jl);
1705         if (jl->j_len == 0 || atomic_read(&jl->j_nonzerolen) == 0) {
1706                 return 0;
1707         }
1708
1709         cn = jl->j_realblock;
1710         while (cn) {
1711                 /* if the blocknr == 0, this has been cleared from the hash,
1712                  ** skip it
1713                  */
1714                 if (cn->blocknr == 0) {
1715                         goto next;
1716                 }
1717                 if (cn->bh && can_dirty(cn) && buffer_dirty(cn->bh)) {
1718                         struct buffer_head *tmp_bh;
1719                         /* we can race against journal_mark_freed when we try
1720                          * to lock_buffer(cn->bh), so we have to inc the buffer
1721                          * count, and recheck things after locking
1722                          */
1723                         tmp_bh = cn->bh;
1724                         get_bh(tmp_bh);
1725                         lock_buffer(tmp_bh);
1726                         if (cn->bh && can_dirty(cn) && buffer_dirty(tmp_bh)) {
1727                                 if (!buffer_journal_dirty(tmp_bh) ||
1728                                     buffer_journal_prepared(tmp_bh))
1729                                         BUG();
1730                                 add_to_chunk(chunk, tmp_bh, NULL, write_chunk);
1731                                 ret++;
1732                         } else {
1733                                 /* note, cn->bh might be null now */
1734                                 unlock_buffer(tmp_bh);
1735                         }
1736                         put_bh(tmp_bh);
1737                 }
1738               next:
1739                 cn = cn->next;
1740                 cond_resched();
1741         }
1742         return ret;
1743 }
1744
1745 /* used by flush_commit_list */
1746 static int dirty_one_transaction(struct super_block *s,
1747                                  struct reiserfs_journal_list *jl)
1748 {
1749         struct reiserfs_journal_cnode *cn;
1750         struct reiserfs_journal_list *pjl;
1751         int ret = 0;
1752
1753         jl->j_state |= LIST_DIRTY;
1754         cn = jl->j_realblock;
1755         while (cn) {
1756                 /* look for a more recent transaction that logged this
1757                  ** buffer.  Only the most recent transaction with a buffer in
1758                  ** it is allowed to send that buffer to disk
1759                  */
1760                 pjl = find_newer_jl_for_cn(cn);
1761                 if (!pjl && cn->blocknr && cn->bh
1762                     && buffer_journal_dirty(cn->bh)) {
1763                         BUG_ON(!can_dirty(cn));
1764                         /* if the buffer is prepared, it will either be logged
1765                          * or restored.  If restored, we need to make sure
1766                          * it actually gets marked dirty
1767                          */
1768                         clear_buffer_journal_new(cn->bh);
1769                         if (buffer_journal_prepared(cn->bh)) {
1770                                 set_buffer_journal_restore_dirty(cn->bh);
1771                         } else {
1772                                 set_buffer_journal_test(cn->bh);
1773                                 mark_buffer_dirty(cn->bh);
1774                         }
1775                 }
1776                 cn = cn->next;
1777         }
1778         return ret;
1779 }
1780
1781 static int kupdate_transactions(struct super_block *s,
1782                                 struct reiserfs_journal_list *jl,
1783                                 struct reiserfs_journal_list **next_jl,
1784                                 unsigned int *next_trans_id,
1785                                 int num_blocks, int num_trans)
1786 {
1787         int ret = 0;
1788         int written = 0;
1789         int transactions_flushed = 0;
1790         unsigned int orig_trans_id = jl->j_trans_id;
1791         struct buffer_chunk chunk;
1792         struct list_head *entry;
1793         struct reiserfs_journal *journal = SB_JOURNAL(s);
1794         chunk.nr = 0;
1795
1796         reiserfs_mutex_lock_safe(&journal->j_flush_mutex, s);
1797         if (!journal_list_still_alive(s, orig_trans_id)) {
1798                 goto done;
1799         }
1800
1801         /* we've got j_flush_mutex held, nobody is going to delete any
1802          * of these lists out from underneath us
1803          */
1804         while ((num_trans && transactions_flushed < num_trans) ||
1805                (!num_trans && written < num_blocks)) {
1806
1807                 if (jl->j_len == 0 || (jl->j_state & LIST_TOUCHED) ||
1808                     atomic_read(&jl->j_commit_left)
1809                     || !(jl->j_state & LIST_DIRTY)) {
1810                         del_from_work_list(s, jl);
1811                         break;
1812                 }
1813                 ret = write_one_transaction(s, jl, &chunk);
1814
1815                 if (ret < 0)
1816                         goto done;
1817                 transactions_flushed++;
1818                 written += ret;
1819                 entry = jl->j_list.next;
1820
1821                 /* did we wrap? */
1822                 if (entry == &journal->j_journal_list) {
1823                         break;
1824                 }
1825                 jl = JOURNAL_LIST_ENTRY(entry);
1826
1827                 /* don't bother with older transactions */
1828                 if (jl->j_trans_id <= orig_trans_id)
1829                         break;
1830         }
1831         if (chunk.nr) {
1832                 write_chunk(&chunk);
1833         }
1834
1835       done:
1836         mutex_unlock(&journal->j_flush_mutex);
1837         return ret;
1838 }
1839
1840 /* for o_sync and fsync heavy applications, they tend to use
1841 ** all the journa list slots with tiny transactions.  These
1842 ** trigger lots and lots of calls to update the header block, which
1843 ** adds seeks and slows things down.
1844 **
1845 ** This function tries to clear out a large chunk of the journal lists
1846 ** at once, which makes everything faster since only the newest journal
1847 ** list updates the header block
1848 */
1849 static int flush_used_journal_lists(struct super_block *s,
1850                                     struct reiserfs_journal_list *jl)
1851 {
1852         unsigned long len = 0;
1853         unsigned long cur_len;
1854         int ret;
1855         int i;
1856         int limit = 256;
1857         struct reiserfs_journal_list *tjl;
1858         struct reiserfs_journal_list *flush_jl;
1859         unsigned int trans_id;
1860         struct reiserfs_journal *journal = SB_JOURNAL(s);
1861
1862         flush_jl = tjl = jl;
1863
1864         /* in data logging mode, try harder to flush a lot of blocks */
1865         if (reiserfs_data_log(s))
1866                 limit = 1024;
1867         /* flush for 256 transactions or limit blocks, whichever comes first */
1868         for (i = 0; i < 256 && len < limit; i++) {
1869                 if (atomic_read(&tjl->j_commit_left) ||
1870                     tjl->j_trans_id < jl->j_trans_id) {
1871                         break;
1872                 }
1873                 cur_len = atomic_read(&tjl->j_nonzerolen);
1874                 if (cur_len > 0) {
1875                         tjl->j_state &= ~LIST_TOUCHED;
1876                 }
1877                 len += cur_len;
1878                 flush_jl = tjl;
1879                 if (tjl->j_list.next == &journal->j_journal_list)
1880                         break;
1881                 tjl = JOURNAL_LIST_ENTRY(tjl->j_list.next);
1882         }
1883         /* try to find a group of blocks we can flush across all the
1884          ** transactions, but only bother if we've actually spanned
1885          ** across multiple lists
1886          */
1887         if (flush_jl != jl) {
1888                 ret = kupdate_transactions(s, jl, &tjl, &trans_id, len, i);
1889         }
1890         flush_journal_list(s, flush_jl, 1);
1891         return 0;
1892 }
1893
1894 /*
1895 ** removes any nodes in table with name block and dev as bh.
1896 ** only touchs the hnext and hprev pointers.
1897 */
1898 void remove_journal_hash(struct super_block *sb,
1899                          struct reiserfs_journal_cnode **table,
1900                          struct reiserfs_journal_list *jl,
1901                          unsigned long block, int remove_freed)
1902 {
1903         struct reiserfs_journal_cnode *cur;
1904         struct reiserfs_journal_cnode **head;
1905
1906         head = &(journal_hash(table, sb, block));
1907         if (!head) {
1908                 return;
1909         }
1910         cur = *head;
1911         while (cur) {
1912                 if (cur->blocknr == block && cur->sb == sb
1913                     && (jl == NULL || jl == cur->jlist)
1914                     && (!test_bit(BLOCK_FREED, &cur->state) || remove_freed)) {
1915                         if (cur->hnext) {
1916                                 cur->hnext->hprev = cur->hprev;
1917                         }
1918                         if (cur->hprev) {
1919                                 cur->hprev->hnext = cur->hnext;
1920                         } else {
1921                                 *head = cur->hnext;
1922                         }
1923                         cur->blocknr = 0;
1924                         cur->sb = NULL;
1925                         cur->state = 0;
1926                         if (cur->bh && cur->jlist)      /* anybody who clears the cur->bh will also dec the nonzerolen */
1927                                 atomic_dec(&(cur->jlist->j_nonzerolen));
1928                         cur->bh = NULL;
1929                         cur->jlist = NULL;
1930                 }
1931                 cur = cur->hnext;
1932         }
1933 }
1934
1935 static void free_journal_ram(struct super_block *sb)
1936 {
1937         struct reiserfs_journal *journal = SB_JOURNAL(sb);
1938         kfree(journal->j_current_jl);
1939         journal->j_num_lists--;
1940
1941         vfree(journal->j_cnode_free_orig);
1942         free_list_bitmaps(sb, journal->j_list_bitmap);
1943         free_bitmap_nodes(sb);  /* must be after free_list_bitmaps */
1944         if (journal->j_header_bh) {
1945                 brelse(journal->j_header_bh);
1946         }
1947         /* j_header_bh is on the journal dev, make sure not to release the journal
1948          * dev until we brelse j_header_bh
1949          */
1950         release_journal_dev(sb, journal);
1951         vfree(journal);
1952 }
1953
1954 /*
1955 ** call on unmount.  Only set error to 1 if you haven't made your way out
1956 ** of read_super() yet.  Any other caller must keep error at 0.
1957 */
1958 static int do_journal_release(struct reiserfs_transaction_handle *th,
1959                               struct super_block *sb, int error)
1960 {
1961         struct reiserfs_transaction_handle myth;
1962         int flushed = 0;
1963         struct reiserfs_journal *journal = SB_JOURNAL(sb);
1964
1965         /* we only want to flush out transactions if we were called with error == 0
1966          */
1967         if (!error && !(sb->s_flags & MS_RDONLY)) {
1968                 /* end the current trans */
1969                 BUG_ON(!th->t_trans_id);
1970                 do_journal_end(th, sb, 10, FLUSH_ALL);
1971
1972                 /* make sure something gets logged to force our way into the flush code */
1973                 if (!journal_join(&myth, sb, 1)) {
1974                         reiserfs_prepare_for_journal(sb,
1975                                                      SB_BUFFER_WITH_SB(sb),
1976                                                      1);
1977                         journal_mark_dirty(&myth, sb,
1978                                            SB_BUFFER_WITH_SB(sb));
1979                         do_journal_end(&myth, sb, 1, FLUSH_ALL);
1980                         flushed = 1;
1981                 }
1982         }
1983
1984         /* this also catches errors during the do_journal_end above */
1985         if (!error && reiserfs_is_journal_aborted(journal)) {
1986                 memset(&myth, 0, sizeof(myth));
1987                 if (!journal_join_abort(&myth, sb, 1)) {
1988                         reiserfs_prepare_for_journal(sb,
1989                                                      SB_BUFFER_WITH_SB(sb),
1990                                                      1);
1991                         journal_mark_dirty(&myth, sb,
1992                                            SB_BUFFER_WITH_SB(sb));
1993                         do_journal_end(&myth, sb, 1, FLUSH_ALL);
1994                 }
1995         }
1996
1997         reiserfs_mounted_fs_count--;
1998         /* wait for all commits to finish */
1999         cancel_delayed_work(&SB_JOURNAL(sb)->j_work);
2000
2001         /*
2002          * We must release the write lock here because
2003          * the workqueue job (flush_async_commit) needs this lock
2004          */
2005         reiserfs_write_unlock(sb);
2006         flush_workqueue(commit_wq);
2007
2008         if (!reiserfs_mounted_fs_count) {
2009                 destroy_workqueue(commit_wq);
2010                 commit_wq = NULL;
2011         }
2012
2013         free_journal_ram(sb);
2014
2015         reiserfs_write_lock(sb);
2016
2017         return 0;
2018 }
2019
2020 /*
2021 ** call on unmount.  flush all journal trans, release all alloc'd ram
2022 */
2023 int journal_release(struct reiserfs_transaction_handle *th,
2024                     struct super_block *sb)
2025 {
2026         return do_journal_release(th, sb, 0);
2027 }
2028
2029 /*
2030 ** only call from an error condition inside reiserfs_read_super!
2031 */
2032 int journal_release_error(struct reiserfs_transaction_handle *th,
2033                           struct super_block *sb)
2034 {
2035         return do_journal_release(th, sb, 1);
2036 }
2037
2038 /* compares description block with commit block.  returns 1 if they differ, 0 if they are the same */
2039 static int journal_compare_desc_commit(struct super_block *sb,
2040                                        struct reiserfs_journal_desc *desc,
2041                                        struct reiserfs_journal_commit *commit)
2042 {
2043         if (get_commit_trans_id(commit) != get_desc_trans_id(desc) ||
2044             get_commit_trans_len(commit) != get_desc_trans_len(desc) ||
2045             get_commit_trans_len(commit) > SB_JOURNAL(sb)->j_trans_max ||
2046             get_commit_trans_len(commit) <= 0) {
2047                 return 1;
2048         }
2049         return 0;
2050 }
2051
2052 /* returns 0 if it did not find a description block
2053 ** returns -1 if it found a corrupt commit block
2054 ** returns 1 if both desc and commit were valid
2055 */
2056 static int journal_transaction_is_valid(struct super_block *sb,
2057                                         struct buffer_head *d_bh,
2058                                         unsigned int *oldest_invalid_trans_id,
2059                                         unsigned long *newest_mount_id)
2060 {
2061         struct reiserfs_journal_desc *desc;
2062         struct reiserfs_journal_commit *commit;
2063         struct buffer_head *c_bh;
2064         unsigned long offset;
2065
2066         if (!d_bh)
2067                 return 0;
2068
2069         desc = (struct reiserfs_journal_desc *)d_bh->b_data;
2070         if (get_desc_trans_len(desc) > 0
2071             && !memcmp(get_journal_desc_magic(d_bh), JOURNAL_DESC_MAGIC, 8)) {
2072                 if (oldest_invalid_trans_id && *oldest_invalid_trans_id
2073                     && get_desc_trans_id(desc) > *oldest_invalid_trans_id) {
2074                         reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2075                                        "journal-986: transaction "
2076                                        "is valid returning because trans_id %d is greater than "
2077                                        "oldest_invalid %lu",
2078                                        get_desc_trans_id(desc),
2079                                        *oldest_invalid_trans_id);
2080                         return 0;
2081                 }
2082                 if (newest_mount_id
2083                     && *newest_mount_id > get_desc_mount_id(desc)) {
2084                         reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2085                                        "journal-1087: transaction "
2086                                        "is valid returning because mount_id %d is less than "
2087                                        "newest_mount_id %lu",
2088                                        get_desc_mount_id(desc),
2089                                        *newest_mount_id);
2090                         return -1;
2091                 }
2092                 if (get_desc_trans_len(desc) > SB_JOURNAL(sb)->j_trans_max) {
2093                         reiserfs_warning(sb, "journal-2018",
2094                                          "Bad transaction length %d "
2095                                          "encountered, ignoring transaction",
2096                                          get_desc_trans_len(desc));
2097                         return -1;
2098                 }
2099                 offset = d_bh->b_blocknr - SB_ONDISK_JOURNAL_1st_BLOCK(sb);
2100
2101                 /* ok, we have a journal description block, lets see if the transaction was valid */
2102                 c_bh =
2103                     journal_bread(sb,
2104                                   SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
2105                                   ((offset + get_desc_trans_len(desc) +
2106                                     1) % SB_ONDISK_JOURNAL_SIZE(sb)));
2107                 if (!c_bh)
2108                         return 0;
2109                 commit = (struct reiserfs_journal_commit *)c_bh->b_data;
2110                 if (journal_compare_desc_commit(sb, desc, commit)) {
2111                         reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2112                                        "journal_transaction_is_valid, commit offset %ld had bad "
2113                                        "time %d or length %d",
2114                                        c_bh->b_blocknr -
2115                                        SB_ONDISK_JOURNAL_1st_BLOCK(sb),
2116                                        get_commit_trans_id(commit),
2117                                        get_commit_trans_len(commit));
2118                         brelse(c_bh);
2119                         if (oldest_invalid_trans_id) {
2120                                 *oldest_invalid_trans_id =
2121                                     get_desc_trans_id(desc);
2122                                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2123                                                "journal-1004: "
2124                                                "transaction_is_valid setting oldest invalid trans_id "
2125                                                "to %d",
2126                                                get_desc_trans_id(desc));
2127                         }
2128                         return -1;
2129                 }
2130                 brelse(c_bh);
2131                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2132                                "journal-1006: found valid "
2133                                "transaction start offset %llu, len %d id %d",
2134                                d_bh->b_blocknr -
2135                                SB_ONDISK_JOURNAL_1st_BLOCK(sb),
2136                                get_desc_trans_len(desc),
2137                                get_desc_trans_id(desc));
2138                 return 1;
2139         } else {
2140                 return 0;
2141         }
2142 }
2143
2144 static void brelse_array(struct buffer_head **heads, int num)
2145 {
2146         int i;
2147         for (i = 0; i < num; i++) {
2148                 brelse(heads[i]);
2149         }
2150 }
2151
2152 /*
2153 ** given the start, and values for the oldest acceptable transactions,
2154 ** this either reads in a replays a transaction, or returns because the transaction
2155 ** is invalid, or too old.
2156 */
2157 static int journal_read_transaction(struct super_block *sb,
2158                                     unsigned long cur_dblock,
2159                                     unsigned long oldest_start,
2160                                     unsigned int oldest_trans_id,
2161                                     unsigned long newest_mount_id)
2162 {
2163         struct reiserfs_journal *journal = SB_JOURNAL(sb);
2164         struct reiserfs_journal_desc *desc;
2165         struct reiserfs_journal_commit *commit;
2166         unsigned int trans_id = 0;
2167         struct buffer_head *c_bh;
2168         struct buffer_head *d_bh;
2169         struct buffer_head **log_blocks = NULL;
2170         struct buffer_head **real_blocks = NULL;
2171         unsigned int trans_offset;
2172         int i;
2173         int trans_half;
2174
2175         d_bh = journal_bread(sb, cur_dblock);
2176         if (!d_bh)
2177                 return 1;
2178         desc = (struct reiserfs_journal_desc *)d_bh->b_data;
2179         trans_offset = d_bh->b_blocknr - SB_ONDISK_JOURNAL_1st_BLOCK(sb);
2180         reiserfs_debug(sb, REISERFS_DEBUG_CODE, "journal-1037: "
2181                        "journal_read_transaction, offset %llu, len %d mount_id %d",
2182                        d_bh->b_blocknr - SB_ONDISK_JOURNAL_1st_BLOCK(sb),
2183                        get_desc_trans_len(desc), get_desc_mount_id(desc));
2184         if (get_desc_trans_id(desc) < oldest_trans_id) {
2185                 reiserfs_debug(sb, REISERFS_DEBUG_CODE, "journal-1039: "
2186                                "journal_read_trans skipping because %lu is too old",
2187                                cur_dblock -
2188                                SB_ONDISK_JOURNAL_1st_BLOCK(sb));
2189                 brelse(d_bh);
2190                 return 1;
2191         }
2192         if (get_desc_mount_id(desc) != newest_mount_id) {
2193                 reiserfs_debug(sb, REISERFS_DEBUG_CODE, "journal-1146: "
2194                                "journal_read_trans skipping because %d is != "
2195                                "newest_mount_id %lu", get_desc_mount_id(desc),
2196                                newest_mount_id);
2197                 brelse(d_bh);
2198                 return 1;
2199         }
2200         c_bh = journal_bread(sb, SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
2201                              ((trans_offset + get_desc_trans_len(desc) + 1) %
2202                               SB_ONDISK_JOURNAL_SIZE(sb)));
2203         if (!c_bh) {
2204                 brelse(d_bh);
2205                 return 1;
2206         }
2207         commit = (struct reiserfs_journal_commit *)c_bh->b_data;
2208         if (journal_compare_desc_commit(sb, desc, commit)) {
2209                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2210                                "journal_read_transaction, "
2211                                "commit offset %llu had bad time %d or length %d",
2212                                c_bh->b_blocknr -
2213                                SB_ONDISK_JOURNAL_1st_BLOCK(sb),
2214                                get_commit_trans_id(commit),
2215                                get_commit_trans_len(commit));
2216                 brelse(c_bh);
2217                 brelse(d_bh);
2218                 return 1;
2219         }
2220
2221         if (bdev_read_only(sb->s_bdev)) {
2222                 reiserfs_warning(sb, "clm-2076",
2223                                  "device is readonly, unable to replay log");
2224                 brelse(c_bh);
2225                 brelse(d_bh);
2226                 return -EROFS;
2227         }
2228
2229         trans_id = get_desc_trans_id(desc);
2230         /* now we know we've got a good transaction, and it was inside the valid time ranges */
2231         log_blocks = kmalloc(get_desc_trans_len(desc) *
2232                              sizeof(struct buffer_head *), GFP_NOFS);
2233         real_blocks = kmalloc(get_desc_trans_len(desc) *
2234                               sizeof(struct buffer_head *), GFP_NOFS);
2235         if (!log_blocks || !real_blocks) {
2236                 brelse(c_bh);
2237                 brelse(d_bh);
2238                 kfree(log_blocks);
2239                 kfree(real_blocks);
2240                 reiserfs_warning(sb, "journal-1169",
2241                                  "kmalloc failed, unable to mount FS");
2242                 return -1;
2243         }
2244         /* get all the buffer heads */
2245         trans_half = journal_trans_half(sb->s_blocksize);
2246         for (i = 0; i < get_desc_trans_len(desc); i++) {
2247                 log_blocks[i] =
2248                     journal_getblk(sb,
2249                                    SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
2250                                    (trans_offset + 1 +
2251                                     i) % SB_ONDISK_JOURNAL_SIZE(sb));
2252                 if (i < trans_half) {
2253                         real_blocks[i] =
2254                             sb_getblk(sb,
2255                                       le32_to_cpu(desc->j_realblock[i]));
2256                 } else {
2257                         real_blocks[i] =
2258                             sb_getblk(sb,
2259                                       le32_to_cpu(commit->
2260                                                   j_realblock[i - trans_half]));
2261                 }
2262                 if (real_blocks[i]->b_blocknr > SB_BLOCK_COUNT(sb)) {
2263                         reiserfs_warning(sb, "journal-1207",
2264                                          "REPLAY FAILURE fsck required! "
2265                                          "Block to replay is outside of "
2266                                          "filesystem");
2267                         goto abort_replay;
2268                 }
2269                 /* make sure we don't try to replay onto log or reserved area */
2270                 if (is_block_in_log_or_reserved_area
2271                     (sb, real_blocks[i]->b_blocknr)) {
2272                         reiserfs_warning(sb, "journal-1204",
2273                                          "REPLAY FAILURE fsck required! "
2274                                          "Trying to replay onto a log block");
2275                       abort_replay:
2276                         brelse_array(log_blocks, i);
2277                         brelse_array(real_blocks, i);
2278                         brelse(c_bh);
2279                         brelse(d_bh);
2280                         kfree(log_blocks);
2281                         kfree(real_blocks);
2282                         return -1;
2283                 }
2284         }
2285         /* read in the log blocks, memcpy to the corresponding real block */
2286         ll_rw_block(READ, get_desc_trans_len(desc), log_blocks);
2287         for (i = 0; i < get_desc_trans_len(desc); i++) {
2288
2289                 reiserfs_write_unlock(sb);
2290                 wait_on_buffer(log_blocks[i]);
2291                 reiserfs_write_lock(sb);
2292
2293                 if (!buffer_uptodate(log_blocks[i])) {
2294                         reiserfs_warning(sb, "journal-1212",
2295                                          "REPLAY FAILURE fsck required! "
2296                                          "buffer write failed");
2297                         brelse_array(log_blocks + i,
2298                                      get_desc_trans_len(desc) - i);
2299                         brelse_array(real_blocks, get_desc_trans_len(desc));
2300                         brelse(c_bh);
2301                         brelse(d_bh);
2302                         kfree(log_blocks);
2303                         kfree(real_blocks);
2304                         return -1;
2305                 }
2306                 memcpy(real_blocks[i]->b_data, log_blocks[i]->b_data,
2307                        real_blocks[i]->b_size);
2308                 set_buffer_uptodate(real_blocks[i]);
2309                 brelse(log_blocks[i]);
2310         }
2311         /* flush out the real blocks */
2312         for (i = 0; i < get_desc_trans_len(desc); i++) {
2313                 set_buffer_dirty(real_blocks[i]);
2314                 ll_rw_block(SWRITE, 1, real_blocks + i);
2315         }
2316         for (i = 0; i < get_desc_trans_len(desc); i++) {
2317                 wait_on_buffer(real_blocks[i]);
2318                 if (!buffer_uptodate(real_blocks[i])) {
2319                         reiserfs_warning(sb, "journal-1226",
2320                                          "REPLAY FAILURE, fsck required! "
2321                                          "buffer write failed");
2322                         brelse_array(real_blocks + i,
2323                                      get_desc_trans_len(desc) - i);
2324                         brelse(c_bh);
2325                         brelse(d_bh);
2326                         kfree(log_blocks);
2327                         kfree(real_blocks);
2328                         return -1;
2329                 }
2330                 brelse(real_blocks[i]);
2331         }
2332         cur_dblock =
2333             SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
2334             ((trans_offset + get_desc_trans_len(desc) +
2335               2) % SB_ONDISK_JOURNAL_SIZE(sb));
2336         reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2337                        "journal-1095: setting journal " "start to offset %ld",
2338                        cur_dblock - SB_ONDISK_JOURNAL_1st_BLOCK(sb));
2339
2340         /* init starting values for the first transaction, in case this is the last transaction to be replayed. */
2341         journal->j_start = cur_dblock - SB_ONDISK_JOURNAL_1st_BLOCK(sb);
2342         journal->j_last_flush_trans_id = trans_id;
2343         journal->j_trans_id = trans_id + 1;
2344         /* check for trans_id overflow */
2345         if (journal->j_trans_id == 0)
2346                 journal->j_trans_id = 10;
2347         brelse(c_bh);
2348         brelse(d_bh);
2349         kfree(log_blocks);
2350         kfree(real_blocks);
2351         return 0;
2352 }
2353
2354 /* This function reads blocks starting from block and to max_block of bufsize
2355    size (but no more than BUFNR blocks at a time). This proved to improve
2356    mounting speed on self-rebuilding raid5 arrays at least.
2357    Right now it is only used from journal code. But later we might use it
2358    from other places.
2359    Note: Do not use journal_getblk/sb_getblk functions here! */
2360 static struct buffer_head *reiserfs_breada(struct block_device *dev,
2361                                            b_blocknr_t block, int bufsize,
2362                                            b_blocknr_t max_block)
2363 {
2364         struct buffer_head *bhlist[BUFNR];
2365         unsigned int blocks = BUFNR;
2366         struct buffer_head *bh;
2367         int i, j;
2368
2369         bh = __getblk(dev, block, bufsize);
2370         if (buffer_uptodate(bh))
2371                 return (bh);
2372
2373         if (block + BUFNR > max_block) {
2374                 blocks = max_block - block;
2375         }
2376         bhlist[0] = bh;
2377         j = 1;
2378         for (i = 1; i < blocks; i++) {
2379                 bh = __getblk(dev, block + i, bufsize);
2380                 if (buffer_uptodate(bh)) {
2381                         brelse(bh);
2382                         break;
2383                 } else
2384                         bhlist[j++] = bh;
2385         }
2386         ll_rw_block(READ, j, bhlist);
2387         for (i = 1; i < j; i++)
2388                 brelse(bhlist[i]);
2389         bh = bhlist[0];
2390         wait_on_buffer(bh);
2391         if (buffer_uptodate(bh))
2392                 return bh;
2393         brelse(bh);
2394         return NULL;
2395 }
2396
2397 /*
2398 ** read and replay the log
2399 ** on a clean unmount, the journal header's next unflushed pointer will be to an invalid
2400 ** transaction.  This tests that before finding all the transactions in the log, which makes normal mount times fast.
2401 **
2402 ** After a crash, this starts with the next unflushed transaction, and replays until it finds one too old, or invalid.
2403 **
2404 ** On exit, it sets things up so the first transaction will work correctly.
2405 */
2406 static int journal_read(struct super_block *sb)
2407 {
2408         struct reiserfs_journal *journal = SB_JOURNAL(sb);
2409         struct reiserfs_journal_desc *desc;
2410         unsigned int oldest_trans_id = 0;
2411         unsigned int oldest_invalid_trans_id = 0;
2412         time_t start;
2413         unsigned long oldest_start = 0;
2414         unsigned long cur_dblock = 0;
2415         unsigned long newest_mount_id = 9;
2416         struct buffer_head *d_bh;
2417         struct reiserfs_journal_header *jh;
2418         int valid_journal_header = 0;
2419         int replay_count = 0;
2420         int continue_replay = 1;
2421         int ret;
2422         char b[BDEVNAME_SIZE];
2423
2424         cur_dblock = SB_ONDISK_JOURNAL_1st_BLOCK(sb);
2425         reiserfs_info(sb, "checking transaction log (%s)\n",
2426                       bdevname(journal->j_dev_bd, b));
2427         start = get_seconds();
2428
2429         /* step 1, read in the journal header block.  Check the transaction it says
2430          ** is the first unflushed, and if that transaction is not valid,
2431          ** replay is done
2432          */
2433         journal->j_header_bh = journal_bread(sb,
2434                                              SB_ONDISK_JOURNAL_1st_BLOCK(sb)
2435                                              + SB_ONDISK_JOURNAL_SIZE(sb));
2436         if (!journal->j_header_bh) {
2437                 return 1;
2438         }
2439         jh = (struct reiserfs_journal_header *)(journal->j_header_bh->b_data);
2440         if (le32_to_cpu(jh->j_first_unflushed_offset) <
2441             SB_ONDISK_JOURNAL_SIZE(sb)
2442             && le32_to_cpu(jh->j_last_flush_trans_id) > 0) {
2443                 oldest_start =
2444                     SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
2445                     le32_to_cpu(jh->j_first_unflushed_offset);
2446                 oldest_trans_id = le32_to_cpu(jh->j_last_flush_trans_id) + 1;
2447                 newest_mount_id = le32_to_cpu(jh->j_mount_id);
2448                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2449                                "journal-1153: found in "
2450                                "header: first_unflushed_offset %d, last_flushed_trans_id "
2451                                "%lu", le32_to_cpu(jh->j_first_unflushed_offset),
2452                                le32_to_cpu(jh->j_last_flush_trans_id));
2453                 valid_journal_header = 1;
2454
2455                 /* now, we try to read the first unflushed offset.  If it is not valid,
2456                  ** there is nothing more we can do, and it makes no sense to read
2457                  ** through the whole log.
2458                  */
2459                 d_bh =
2460                     journal_bread(sb,
2461                                   SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
2462                                   le32_to_cpu(jh->j_first_unflushed_offset));
2463                 ret = journal_transaction_is_valid(sb, d_bh, NULL, NULL);
2464                 if (!ret) {
2465                         continue_replay = 0;
2466                 }
2467                 brelse(d_bh);
2468                 goto start_log_replay;
2469         }
2470
2471         /* ok, there are transactions that need to be replayed.  start with the first log block, find
2472          ** all the valid transactions, and pick out the oldest.
2473          */
2474         while (continue_replay
2475                && cur_dblock <
2476                (SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
2477                 SB_ONDISK_JOURNAL_SIZE(sb))) {
2478                 /* Note that it is required for blocksize of primary fs device and journal
2479                    device to be the same */
2480                 d_bh =
2481                     reiserfs_breada(journal->j_dev_bd, cur_dblock,
2482                                     sb->s_blocksize,
2483                                     SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
2484                                     SB_ONDISK_JOURNAL_SIZE(sb));
2485                 ret =
2486                     journal_transaction_is_valid(sb, d_bh,
2487                                                  &oldest_invalid_trans_id,
2488                                                  &newest_mount_id);
2489                 if (ret == 1) {
2490                         desc = (struct reiserfs_journal_desc *)d_bh->b_data;
2491                         if (oldest_start == 0) {        /* init all oldest_ values */
2492                                 oldest_trans_id = get_desc_trans_id(desc);
2493                                 oldest_start = d_bh->b_blocknr;
2494                                 newest_mount_id = get_desc_mount_id(desc);
2495                                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2496                                                "journal-1179: Setting "
2497                                                "oldest_start to offset %llu, trans_id %lu",
2498                                                oldest_start -
2499                                                SB_ONDISK_JOURNAL_1st_BLOCK
2500                                                (sb), oldest_trans_id);
2501                         } else if (oldest_trans_id > get_desc_trans_id(desc)) {
2502                                 /* one we just read was older */
2503                                 oldest_trans_id = get_desc_trans_id(desc);
2504                                 oldest_start = d_bh->b_blocknr;
2505                                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2506                                                "journal-1180: Resetting "
2507                                                "oldest_start to offset %lu, trans_id %lu",
2508                                                oldest_start -
2509                                                SB_ONDISK_JOURNAL_1st_BLOCK
2510                                                (sb), oldest_trans_id);
2511                         }
2512                         if (newest_mount_id < get_desc_mount_id(desc)) {
2513                                 newest_mount_id = get_desc_mount_id(desc);
2514                                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2515                                                "journal-1299: Setting "
2516                                                "newest_mount_id to %d",
2517                                                get_desc_mount_id(desc));
2518                         }
2519                         cur_dblock += get_desc_trans_len(desc) + 2;
2520                 } else {
2521                         cur_dblock++;
2522                 }
2523                 brelse(d_bh);
2524         }
2525
2526       start_log_replay:
2527         cur_dblock = oldest_start;
2528         if (oldest_trans_id) {
2529                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2530                                "journal-1206: Starting replay "
2531                                "from offset %llu, trans_id %lu",
2532                                cur_dblock - SB_ONDISK_JOURNAL_1st_BLOCK(sb),
2533                                oldest_trans_id);
2534
2535         }
2536         replay_count = 0;
2537         while (continue_replay && oldest_trans_id > 0) {
2538                 ret =
2539                     journal_read_transaction(sb, cur_dblock, oldest_start,
2540                                              oldest_trans_id, newest_mount_id);
2541                 if (ret < 0) {
2542                         return ret;
2543                 } else if (ret != 0) {
2544                         break;
2545                 }
2546                 cur_dblock =
2547                     SB_ONDISK_JOURNAL_1st_BLOCK(sb) + journal->j_start;
2548                 replay_count++;
2549                 if (cur_dblock == oldest_start)
2550                         break;
2551         }
2552
2553         if (oldest_trans_id == 0) {
2554                 reiserfs_debug(sb, REISERFS_DEBUG_CODE,
2555                                "journal-1225: No valid " "transactions found");
2556         }
2557         /* j_start does not get set correctly if we don't replay any transactions.
2558          ** if we had a valid journal_header, set j_start to the first unflushed transaction value,
2559          ** copy the trans_id from the header
2560          */
2561         if (valid_journal_header && replay_count == 0) {
2562                 journal->j_start = le32_to_cpu(jh->j_first_unflushed_offset);
2563                 journal->j_trans_id =
2564                     le32_to_cpu(jh->j_last_flush_trans_id) + 1;
2565                 /* check for trans_id overflow */
2566                 if (journal->j_trans_id == 0)
2567                         journal->j_trans_id = 10;
2568                 journal->j_last_flush_trans_id =
2569                     le32_to_cpu(jh->j_last_flush_trans_id);
2570                 journal->j_mount_id = le32_to_cpu(jh->j_mount_id) + 1;
2571         } else {
2572                 journal->j_mount_id = newest_mount_id + 1;
2573         }
2574         reiserfs_debug(sb, REISERFS_DEBUG_CODE, "journal-1299: Setting "
2575                        "newest_mount_id to %lu", journal->j_mount_id);
2576         journal->j_first_unflushed_offset = journal->j_start;
2577         if (replay_count > 0) {
2578                 reiserfs_info(sb,
2579                               "replayed %d transactions in %lu seconds\n",
2580                               replay_count, get_seconds() - start);
2581         }
2582         if (!bdev_read_only(sb->s_bdev) &&
2583             _update_journal_header_block(sb, journal->j_start,
2584                                          journal->j_last_flush_trans_id)) {
2585                 /* replay failed, caller must call free_journal_ram and abort
2586                  ** the mount
2587                  */
2588                 return -1;
2589         }
2590         return 0;
2591 }
2592
2593 static struct reiserfs_journal_list *alloc_journal_list(struct super_block *s)
2594 {
2595         struct reiserfs_journal_list *jl;
2596         jl = kzalloc(sizeof(struct reiserfs_journal_list),
2597                      GFP_NOFS | __GFP_NOFAIL);
2598         INIT_LIST_HEAD(&jl->j_list);
2599         INIT_LIST_HEAD(&jl->j_working_list);
2600         INIT_LIST_HEAD(&jl->j_tail_bh_list);
2601         INIT_LIST_HEAD(&jl->j_bh_list);
2602         mutex_init(&jl->j_commit_mutex);
2603         SB_JOURNAL(s)->j_num_lists++;
2604         get_journal_list(jl);
2605         return jl;
2606 }
2607
2608 static void journal_list_init(struct super_block *sb)
2609 {
2610         SB_JOURNAL(sb)->j_current_jl = alloc_journal_list(sb);
2611 }
2612
2613 static int release_journal_dev(struct super_block *super,
2614                                struct reiserfs_journal *journal)
2615 {
2616         int result;
2617
2618         result = 0;
2619
2620         if (journal->j_dev_bd != NULL) {
2621                 if (journal->j_dev_bd->bd_dev != super->s_dev)
2622                         bd_release(journal->j_dev_bd);
2623                 result = blkdev_put(journal->j_dev_bd, journal->j_dev_mode);
2624                 journal->j_dev_bd = NULL;
2625         }
2626
2627         if (result != 0) {
2628                 reiserfs_warning(super, "sh-457",
2629                                  "Cannot release journal device: %i", result);
2630         }
2631         return result;
2632 }
2633
2634 static int journal_init_dev(struct super_block *super,
2635                             struct reiserfs_journal *journal,
2636                             const char *jdev_name)
2637 {
2638         int result;
2639         dev_t jdev;
2640         fmode_t blkdev_mode = FMODE_READ | FMODE_WRITE;
2641         char b[BDEVNAME_SIZE];
2642
2643         result = 0;
2644
2645         journal->j_dev_bd = NULL;
2646         jdev = SB_ONDISK_JOURNAL_DEVICE(super) ?
2647             new_decode_dev(SB_ONDISK_JOURNAL_DEVICE(super)) : super->s_dev;
2648
2649         if (bdev_read_only(super->s_bdev))
2650                 blkdev_mode = FMODE_READ;
2651
2652         /* there is no "jdev" option and journal is on separate device */
2653         if ((!jdev_name || !jdev_name[0])) {
2654                 journal->j_dev_bd = open_by_devnum(jdev, blkdev_mode);
2655                 journal->j_dev_mode = blkdev_mode;
2656                 if (IS_ERR(journal->j_dev_bd)) {
2657                         result = PTR_ERR(journal->j_dev_bd);
2658                         journal->j_dev_bd = NULL;
2659                         reiserfs_warning(super, "sh-458",
2660                                          "cannot init journal device '%s': %i",
2661                                          __bdevname(jdev, b), result);
2662                         return result;
2663                 } else if (jdev != super->s_dev) {
2664                         result = bd_claim(journal->j_dev_bd, journal);
2665                         if (result) {
2666                                 blkdev_put(journal->j_dev_bd, blkdev_mode);
2667                                 return result;
2668                         }
2669
2670                         set_blocksize(journal->j_dev_bd, super->s_blocksize);
2671                 }
2672
2673                 return 0;
2674         }
2675
2676         journal->j_dev_mode = blkdev_mode;
2677         journal->j_dev_bd = open_bdev_exclusive(jdev_name,
2678                                                 blkdev_mode, journal);
2679         if (IS_ERR(journal->j_dev_bd)) {
2680                 result = PTR_ERR(journal->j_dev_bd);
2681                 journal->j_dev_bd = NULL;
2682                 reiserfs_warning(super,
2683                                  "journal_init_dev: Cannot open '%s': %i",
2684                                  jdev_name, result);
2685                 return result;
2686         }
2687
2688         set_blocksize(journal->j_dev_bd, super->s_blocksize);
2689         reiserfs_info(super,
2690                       "journal_init_dev: journal device: %s\n",
2691                       bdevname(journal->j_dev_bd, b));
2692         return 0;
2693 }
2694
2695 /**
2696  * When creating/tuning a file system user can assign some
2697  * journal params within boundaries which depend on the ratio
2698  * blocksize/standard_blocksize.
2699  *
2700  * For blocks >= standard_blocksize transaction size should
2701  * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more
2702  * then JOURNAL_TRANS_MAX_DEFAULT.
2703  *
2704  * For blocks < standard_blocksize these boundaries should be
2705  * decreased proportionally.
2706  */
2707 #define REISERFS_STANDARD_BLKSIZE (4096)
2708
2709 static int check_advise_trans_params(struct super_block *sb,
2710                                      struct reiserfs_journal *journal)
2711 {
2712         if (journal->j_trans_max) {
2713                 /* Non-default journal params.
2714                    Do sanity check for them. */
2715                 int ratio = 1;
2716                 if (sb->s_blocksize < REISERFS_STANDARD_BLKSIZE)
2717                         ratio = REISERFS_STANDARD_BLKSIZE / sb->s_blocksize;
2718
2719                 if (journal->j_trans_max > JOURNAL_TRANS_MAX_DEFAULT / ratio ||
2720                     journal->j_trans_max < JOURNAL_TRANS_MIN_DEFAULT / ratio ||
2721                     SB_ONDISK_JOURNAL_SIZE(sb) / journal->j_trans_max <
2722                     JOURNAL_MIN_RATIO) {
2723                         reiserfs_warning(sb, "sh-462",
2724                                          "bad transaction max size (%u). "
2725                                          "FSCK?", journal->j_trans_max);
2726                         return 1;
2727                 }
2728                 if (journal->j_max_batch != (journal->j_trans_max) *
2729                         JOURNAL_MAX_BATCH_DEFAULT/JOURNAL_TRANS_MAX_DEFAULT) {
2730                         reiserfs_warning(sb, "sh-463",
2731                                          "bad transaction max batch (%u). "
2732                                          "FSCK?", journal->j_max_batch);
2733                         return 1;
2734                 }
2735         } else {
2736                 /* Default journal params.
2737                    The file system was created by old version
2738                    of mkreiserfs, so some fields contain zeros,
2739                    and we need to advise proper values for them */
2740                 if (sb->s_blocksize != REISERFS_STANDARD_BLKSIZE) {
2741                         reiserfs_warning(sb, "sh-464", "bad blocksize (%u)",
2742                                          sb->s_blocksize);
2743                         return 1;
2744                 }
2745                 journal->j_trans_max = JOURNAL_TRANS_MAX_DEFAULT;
2746                 journal->j_max_batch = JOURNAL_MAX_BATCH_DEFAULT;
2747                 journal->j_max_commit_age = JOURNAL_MAX_COMMIT_AGE;
2748         }
2749         return 0;
2750 }
2751
2752 /*
2753 ** must be called once on fs mount.  calls journal_read for you
2754 */
2755 int journal_init(struct super_block *sb, const char *j_dev_name,
2756                  int old_format, unsigned int commit_max_age)
2757 {
2758         int num_cnodes = SB_ONDISK_JOURNAL_SIZE(sb) * 2;
2759         struct buffer_head *bhjh;
2760         struct reiserfs_super_block *rs;
2761         struct reiserfs_journal_header *jh;
2762         struct reiserfs_journal *journal;
2763         struct reiserfs_journal_list *jl;
2764         char b[BDEVNAME_SIZE];
2765         int ret;
2766
2767         /*
2768          * Unlock here to avoid various RECLAIM-FS-ON <-> IN-RECLAIM-FS
2769          * dependency inversion warnings.
2770          */
2771         reiserfs_write_unlock(sb);
2772         journal = SB_JOURNAL(sb) = vmalloc(sizeof(struct reiserfs_journal));
2773         if (!journal) {
2774                 reiserfs_warning(sb, "journal-1256",
2775                                  "unable to get memory for journal structure");
2776                 reiserfs_write_lock(sb);
2777                 return 1;
2778         }
2779         memset(journal, 0, sizeof(struct reiserfs_journal));
2780         INIT_LIST_HEAD(&journal->j_bitmap_nodes);
2781         INIT_LIST_HEAD(&journal->j_prealloc_list);
2782         INIT_LIST_HEAD(&journal->j_working_list);
2783         INIT_LIST_HEAD(&journal->j_journal_list);
2784         journal->j_persistent_trans = 0;
2785         ret = reiserfs_allocate_list_bitmaps(sb, journal->j_list_bitmap,
2786                                            reiserfs_bmap_count(sb));
2787         reiserfs_write_lock(sb);
2788         if (ret)
2789                 goto free_and_return;
2790
2791         allocate_bitmap_nodes(sb);
2792
2793         /* reserved for journal area support */
2794         SB_JOURNAL_1st_RESERVED_BLOCK(sb) = (old_format ?
2795                                                  REISERFS_OLD_DISK_OFFSET_IN_BYTES
2796                                                  / sb->s_blocksize +
2797                                                  reiserfs_bmap_count(sb) +
2798                                                  1 :
2799                                                  REISERFS_DISK_OFFSET_IN_BYTES /
2800                                                  sb->s_blocksize + 2);
2801
2802         /* Sanity check to see is the standard journal fitting withing first bitmap
2803            (actual for small blocksizes) */
2804         if (!SB_ONDISK_JOURNAL_DEVICE(sb) &&
2805             (SB_JOURNAL_1st_RESERVED_BLOCK(sb) +
2806              SB_ONDISK_JOURNAL_SIZE(sb) > sb->s_blocksize * 8)) {
2807                 reiserfs_warning(sb, "journal-1393",
2808                                  "journal does not fit for area addressed "
2809                                  "by first of bitmap blocks. It starts at "
2810                                  "%u and its size is %u. Block size %ld",
2811                                  SB_JOURNAL_1st_RESERVED_BLOCK(sb),
2812                                  SB_ONDISK_JOURNAL_SIZE(sb),
2813                                  sb->s_blocksize);
2814                 goto free_and_return;
2815         }
2816
2817         /*
2818          * We need to unlock here to avoid creating the following
2819          * dependency:
2820          * reiserfs_lock -> sysfs_mutex
2821          * Because the reiserfs mmap path creates the following dependency:
2822          * mm->mmap -> reiserfs_lock, hence we have
2823          * mm->mmap -> reiserfs_lock ->sysfs_mutex
2824          * This would ends up in a circular dependency with sysfs readdir path
2825          * which does sysfs_mutex -> mm->mmap_sem
2826          * This is fine because the reiserfs lock is useless in mount path,
2827          * at least until we call journal_begin. We keep it for paranoid
2828          * reasons.
2829          */
2830         reiserfs_write_unlock(sb);
2831         if (journal_init_dev(sb, journal, j_dev_name) != 0) {
2832                 reiserfs_write_lock(sb);
2833                 reiserfs_warning(sb, "sh-462",
2834                                  "unable to initialize jornal device");
2835                 goto free_and_return;
2836         }
2837         reiserfs_write_lock(sb);
2838
2839         rs = SB_DISK_SUPER_BLOCK(sb);
2840
2841         /* read journal header */
2842         bhjh = journal_bread(sb,
2843                              SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
2844                              SB_ONDISK_JOURNAL_SIZE(sb));
2845         if (!bhjh) {
2846                 reiserfs_warning(sb, "sh-459",
2847                                  "unable to read journal header");
2848                 goto free_and_return;
2849         }
2850         jh = (struct reiserfs_journal_header *)(bhjh->b_data);
2851
2852         /* make sure that journal matches to the super block */
2853         if (is_reiserfs_jr(rs)
2854             && (le32_to_cpu(jh->jh_journal.jp_journal_magic) !=
2855                 sb_jp_journal_magic(rs))) {
2856                 reiserfs_warning(sb, "sh-460",
2857                                  "journal header magic %x (device %s) does "
2858                                  "not match to magic found in super block %x",
2859                                  jh->jh_journal.jp_journal_magic,
2860                                  bdevname(journal->j_dev_bd, b),
2861                                  sb_jp_journal_magic(rs));
2862                 brelse(bhjh);
2863                 goto free_and_return;
2864         }
2865
2866         journal->j_trans_max = le32_to_cpu(jh->jh_journal.jp_journal_trans_max);
2867         journal->j_max_batch = le32_to_cpu(jh->jh_journal.jp_journal_max_batch);
2868         journal->j_max_commit_age =
2869             le32_to_cpu(jh->jh_journal.jp_journal_max_commit_age);
2870         journal->j_max_trans_age = JOURNAL_MAX_TRANS_AGE;
2871
2872         if (check_advise_trans_params(sb, journal) != 0)
2873                 goto free_and_return;
2874         journal->j_default_max_commit_age = journal->j_max_commit_age;
2875
2876         if (commit_max_age != 0) {
2877                 journal->j_max_commit_age = commit_max_age;
2878                 journal->j_max_trans_age = commit_max_age;
2879         }
2880
2881         reiserfs_info(sb, "journal params: device %s, size %u, "
2882                       "journal first block %u, max trans len %u, max batch %u, "
2883                       "max commit age %u, max trans age %u\n",
2884                       bdevname(journal->j_dev_bd, b),
2885                       SB_ONDISK_JOURNAL_SIZE(sb),
2886                       SB_ONDISK_JOURNAL_1st_BLOCK(sb),
2887                       journal->j_trans_max,
2888                       journal->j_max_batch,
2889                       journal->j_max_commit_age, journal->j_max_trans_age);
2890
2891         brelse(bhjh);
2892
2893         journal->j_list_bitmap_index = 0;
2894         journal_list_init(sb);
2895
2896         memset(journal->j_list_hash_table, 0,
2897                JOURNAL_HASH_SIZE * sizeof(struct reiserfs_journal_cnode *));
2898
2899         INIT_LIST_HEAD(&journal->j_dirty_buffers);
2900         spin_lock_init(&journal->j_dirty_buffers_lock);
2901
2902         journal->j_start = 0;
2903         journal->j_len = 0;
2904         journal->j_len_alloc = 0;
2905         atomic_set(&(journal->j_wcount), 0);
2906         atomic_set(&(journal->j_async_throttle), 0);
2907         journal->j_bcount = 0;
2908         journal->j_trans_start_time = 0;
2909         journal->j_last = NULL;
2910         journal->j_first = NULL;
2911         init_waitqueue_head(&(journal->j_join_wait));
2912         mutex_init(&journal->j_mutex);
2913         mutex_init(&journal->j_flush_mutex);
2914
2915         journal->j_trans_id = 10;
2916         journal->j_mount_id = 10;
2917         journal->j_state = 0;
2918         atomic_set(&(journal->j_jlock), 0);
2919         reiserfs_write_unlock(sb);
2920         journal->j_cnode_free_list = allocate_cnodes(num_cnodes);
2921         reiserfs_write_lock(sb);
2922         journal->j_cnode_free_orig = journal->j_cnode_free_list;
2923         journal->j_cnode_free = journal->j_cnode_free_list ? num_cnodes : 0;
2924         journal->j_cnode_used = 0;
2925         journal->j_must_wait = 0;
2926
2927         if (journal->j_cnode_free == 0) {
2928                 reiserfs_warning(sb, "journal-2004", "Journal cnode memory "
2929                                  "allocation failed (%ld bytes). Journal is "
2930                                  "too large for available memory. Usually "
2931                                  "this is due to a journal that is too large.",
2932                                  sizeof (struct reiserfs_journal_cnode) * num_cnodes);
2933                 goto free_and_return;
2934         }
2935
2936         init_journal_hash(sb);
2937         jl = journal->j_current_jl;
2938         jl->j_list_bitmap = get_list_bitmap(sb, jl);
2939         if (!jl->j_list_bitmap) {
2940                 reiserfs_warning(sb, "journal-2005",
2941                                  "get_list_bitmap failed for journal list 0");
2942                 goto free_and_return;
2943         }
2944         if (journal_read(sb) < 0) {
2945                 reiserfs_warning(sb, "reiserfs-2006",
2946                                  "Replay Failure, unable to mount");
2947                 goto free_and_return;
2948         }
2949
2950         reiserfs_mounted_fs_count++;
2951         if (reiserfs_mounted_fs_count <= 1) {
2952                 reiserfs_write_unlock(sb);
2953                 commit_wq = create_workqueue("reiserfs");
2954                 reiserfs_write_lock(sb);
2955         }
2956
2957         INIT_DELAYED_WORK(&journal->j_work, flush_async_commits);
2958         journal->j_work_sb = sb;
2959         return 0;
2960       free_and_return:
2961         free_journal_ram(sb);
2962         return 1;
2963 }
2964
2965 /*
2966 ** test for a polite end of the current transaction.  Used by file_write, and should
2967 ** be used by delete to make sure they don't write more than can fit inside a single
2968 ** transaction
2969 */
2970 int journal_transaction_should_end(struct reiserfs_transaction_handle *th,
2971                                    int new_alloc)
2972 {
2973         struct reiserfs_journal *journal = SB_JOURNAL(th->t_super);
2974         time_t now = get_seconds();
2975         /* cannot restart while nested */
2976         BUG_ON(!th->t_trans_id);
2977         if (th->t_refcount > 1)
2978                 return 0;
2979         if (journal->j_must_wait > 0 ||
2980             (journal->j_len_alloc + new_alloc) >= journal->j_max_batch ||
2981             atomic_read(&(journal->j_jlock)) ||
2982             (now - journal->j_trans_start_time) > journal->j_max_trans_age ||
2983             journal->j_cnode_free < (journal->j_trans_max * 3)) {
2984                 return 1;
2985         }
2986         /* protected by the BKL here */
2987         journal->j_len_alloc += new_alloc;
2988         th->t_blocks_allocated += new_alloc ;
2989         return 0;
2990 }
2991
2992 /* this must be called inside a transaction, and requires the
2993 ** kernel_lock to be held
2994 */
2995 void reiserfs_block_writes(struct reiserfs_transaction_handle *th)
2996 {
2997         struct reiserfs_journal *journal = SB_JOURNAL(th->t_super);
2998         BUG_ON(!th->t_trans_id);
2999         journal->j_must_wait = 1;
3000         set_bit(J_WRITERS_BLOCKED, &journal->j_state);
3001         return;
3002 }
3003
3004 /* this must be called without a transaction started, and does not
3005 ** require BKL
3006 */
3007 void reiserfs_allow_writes(struct super_block *s)
3008 {
3009         struct reiserfs_journal *journal = SB_JOURNAL(s);
3010         clear_bit(J_WRITERS_BLOCKED, &journal->j_state);
3011         wake_up(&journal->j_join_wait);
3012 }
3013
3014 /* this must be called without a transaction started, and does not
3015 ** require BKL
3016 */
3017 void reiserfs_wait_on_write_block(struct super_block *s)
3018 {
3019         struct reiserfs_journal *journal = SB_JOURNAL(s);
3020         wait_event(journal->j_join_wait,
3021                    !test_bit(J_WRITERS_BLOCKED, &journal->j_state));
3022 }
3023
3024 static void queue_log_writer(struct super_block *s)
3025 {
3026         wait_queue_t wait;
3027         struct reiserfs_journal *journal = SB_JOURNAL(s);
3028         set_bit(J_WRITERS_QUEUED, &journal->j_state);
3029
3030         /*
3031          * we don't want to use wait_event here because
3032          * we only want to wait once.
3033          */
3034         init_waitqueue_entry(&wait, current);
3035         add_wait_queue(&journal->j_join_wait, &wait);
3036         set_current_state(TASK_UNINTERRUPTIBLE);
3037         if (test_bit(J_WRITERS_QUEUED, &journal->j_state)) {
3038                 reiserfs_write_unlock(s);
3039                 schedule();
3040                 reiserfs_write_lock(s);
3041         }
3042         __set_current_state(TASK_RUNNING);
3043         remove_wait_queue(&journal->j_join_wait, &wait);
3044 }
3045
3046 static void wake_queued_writers(struct super_block *s)
3047 {
3048         struct reiserfs_journal *journal = SB_JOURNAL(s);
3049         if (test_and_clear_bit(J_WRITERS_QUEUED, &journal->j_state))
3050                 wake_up(&journal->j_join_wait);
3051 }
3052
3053 static void let_transaction_grow(struct super_block *sb, unsigned int trans_id)
3054 {
3055         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3056         unsigned long bcount = journal->j_bcount;
3057         while (1) {
3058                 reiserfs_write_unlock(sb);
3059                 schedule_timeout_uninterruptible(1);
3060                 reiserfs_write_lock(sb);
3061                 journal->j_current_jl->j_state |= LIST_COMMIT_PENDING;
3062                 while ((atomic_read(&journal->j_wcount) > 0 ||
3063                         atomic_read(&journal->j_jlock)) &&
3064                        journal->j_trans_id == trans_id) {
3065                         queue_log_writer(sb);
3066                 }
3067                 if (journal->j_trans_id != trans_id)
3068                         break;
3069                 if (bcount == journal->j_bcount)
3070                         break;
3071                 bcount = journal->j_bcount;
3072         }
3073 }
3074
3075 /* join == true if you must join an existing transaction.
3076 ** join == false if you can deal with waiting for others to finish
3077 **
3078 ** this will block until the transaction is joinable.  send the number of blocks you
3079 ** expect to use in nblocks.
3080 */
3081 static int do_journal_begin_r(struct reiserfs_transaction_handle *th,
3082                               struct super_block *sb, unsigned long nblocks,
3083                               int join)
3084 {
3085         time_t now = get_seconds();
3086         unsigned int old_trans_id;
3087         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3088         struct reiserfs_transaction_handle myth;
3089         int sched_count = 0;
3090         int retval;
3091
3092         reiserfs_check_lock_depth(sb, "journal_begin");
3093         BUG_ON(nblocks > journal->j_trans_max);
3094
3095         PROC_INFO_INC(sb, journal.journal_being);
3096         /* set here for journal_join */
3097         th->t_refcount = 1;
3098         th->t_super = sb;
3099
3100       relock:
3101         lock_journal(sb);
3102         if (join != JBEGIN_ABORT && reiserfs_is_journal_aborted(journal)) {
3103                 unlock_journal(sb);
3104                 retval = journal->j_errno;
3105                 goto out_fail;
3106         }
3107         journal->j_bcount++;
3108
3109         if (test_bit(J_WRITERS_BLOCKED, &journal->j_state)) {
3110                 unlock_journal(sb);
3111                 reiserfs_write_unlock(sb);
3112                 reiserfs_wait_on_write_block(sb);
3113                 reiserfs_write_lock(sb);
3114                 PROC_INFO_INC(sb, journal.journal_relock_writers);
3115                 goto relock;
3116         }
3117         now = get_seconds();
3118
3119         /* if there is no room in the journal OR
3120          ** if this transaction is too old, and we weren't called joinable, wait for it to finish before beginning
3121          ** we don't sleep if there aren't other writers
3122          */
3123
3124         if ((!join && journal->j_must_wait > 0) ||
3125             (!join
3126              && (journal->j_len_alloc + nblocks + 2) >= journal->j_max_batch)
3127             || (!join && atomic_read(&journal->j_wcount) > 0
3128                 && journal->j_trans_start_time > 0
3129                 && (now - journal->j_trans_start_time) >
3130                 journal->j_max_trans_age) || (!join
3131                                               && atomic_read(&journal->j_jlock))
3132             || (!join && journal->j_cnode_free < (journal->j_trans_max * 3))) {
3133
3134                 old_trans_id = journal->j_trans_id;
3135                 unlock_journal(sb);     /* allow others to finish this transaction */
3136
3137                 if (!join && (journal->j_len_alloc + nblocks + 2) >=
3138                     journal->j_max_batch &&
3139                     ((journal->j_len + nblocks + 2) * 100) <
3140                     (journal->j_len_alloc * 75)) {
3141                         if (atomic_read(&journal->j_wcount) > 10) {
3142                                 sched_count++;
3143                                 queue_log_writer(sb);
3144                                 goto relock;
3145                         }
3146                 }
3147                 /* don't mess with joining the transaction if all we have to do is
3148                  * wait for someone else to do a commit
3149                  */
3150                 if (atomic_read(&journal->j_jlock)) {
3151                         while (journal->j_trans_id == old_trans_id &&
3152                                atomic_read(&journal->j_jlock)) {
3153                                 queue_log_writer(sb);
3154                         }
3155                         goto relock;
3156                 }
3157                 retval = journal_join(&myth, sb, 1);
3158                 if (retval)
3159                         goto out_fail;
3160
3161                 /* someone might have ended the transaction while we joined */
3162                 if (old_trans_id != journal->j_trans_id) {
3163                         retval = do_journal_end(&myth, sb, 1, 0);
3164                 } else {
3165                         retval = do_journal_end(&myth, sb, 1, COMMIT_NOW);
3166                 }
3167
3168                 if (retval)
3169                         goto out_fail;
3170
3171                 PROC_INFO_INC(sb, journal.journal_relock_wcount);
3172                 goto relock;
3173         }
3174         /* we are the first writer, set trans_id */
3175         if (journal->j_trans_start_time == 0) {
3176                 journal->j_trans_start_time = get_seconds();
3177         }
3178         atomic_inc(&(journal->j_wcount));
3179         journal->j_len_alloc += nblocks;
3180         th->t_blocks_logged = 0;
3181         th->t_blocks_allocated = nblocks;
3182         th->t_trans_id = journal->j_trans_id;
3183         unlock_journal(sb);
3184         INIT_LIST_HEAD(&th->t_list);
3185         get_fs_excl();
3186         return 0;
3187
3188       out_fail:
3189         memset(th, 0, sizeof(*th));
3190         /* Re-set th->t_super, so we can properly keep track of how many
3191          * persistent transactions there are. We need to do this so if this
3192          * call is part of a failed restart_transaction, we can free it later */
3193         th->t_super = sb;
3194         return retval;
3195 }
3196
3197 struct reiserfs_transaction_handle *reiserfs_persistent_transaction(struct
3198                                                                     super_block
3199                                                                     *s,
3200                                                                     int nblocks)
3201 {
3202         int ret;
3203         struct reiserfs_transaction_handle *th;
3204
3205         /* if we're nesting into an existing transaction.  It will be
3206          ** persistent on its own
3207          */
3208         if (reiserfs_transaction_running(s)) {
3209                 th = current->journal_info;
3210                 th->t_refcount++;
3211                 BUG_ON(th->t_refcount < 2);
3212
3213                 return th;
3214         }
3215         th = kmalloc(sizeof(struct reiserfs_transaction_handle), GFP_NOFS);
3216         if (!th)
3217                 return NULL;
3218         ret = journal_begin(th, s, nblocks);
3219         if (ret) {
3220                 kfree(th);
3221                 return NULL;
3222         }
3223
3224         SB_JOURNAL(s)->j_persistent_trans++;
3225         return th;
3226 }
3227
3228 int reiserfs_end_persistent_transaction(struct reiserfs_transaction_handle *th)
3229 {
3230         struct super_block *s = th->t_super;
3231         int ret = 0;
3232         if (th->t_trans_id)
3233                 ret = journal_end(th, th->t_super, th->t_blocks_allocated);
3234         else
3235                 ret = -EIO;
3236         if (th->t_refcount == 0) {
3237                 SB_JOURNAL(s)->j_persistent_trans--;
3238                 kfree(th);
3239         }
3240         return ret;
3241 }
3242
3243 static int journal_join(struct reiserfs_transaction_handle *th,
3244                         struct super_block *sb, unsigned long nblocks)
3245 {
3246         struct reiserfs_transaction_handle *cur_th = current->journal_info;
3247
3248         /* this keeps do_journal_end from NULLing out the current->journal_info
3249          ** pointer
3250          */
3251         th->t_handle_save = cur_th;
3252         BUG_ON(cur_th && cur_th->t_refcount > 1);
3253         return do_journal_begin_r(th, sb, nblocks, JBEGIN_JOIN);
3254 }
3255
3256 int journal_join_abort(struct reiserfs_transaction_handle *th,
3257                        struct super_block *sb, unsigned long nblocks)
3258 {
3259         struct reiserfs_transaction_handle *cur_th = current->journal_info;
3260
3261         /* this keeps do_journal_end from NULLing out the current->journal_info
3262          ** pointer
3263          */
3264         th->t_handle_save = cur_th;
3265         BUG_ON(cur_th && cur_th->t_refcount > 1);
3266         return do_journal_begin_r(th, sb, nblocks, JBEGIN_ABORT);
3267 }
3268
3269 int journal_begin(struct reiserfs_transaction_handle *th,
3270                   struct super_block *sb, unsigned long nblocks)
3271 {
3272         struct reiserfs_transaction_handle *cur_th = current->journal_info;
3273         int ret;
3274
3275         th->t_handle_save = NULL;
3276         if (cur_th) {
3277                 /* we are nesting into the current transaction */
3278                 if (cur_th->t_super == sb) {
3279                         BUG_ON(!cur_th->t_refcount);
3280                         cur_th->t_refcount++;
3281                         memcpy(th, cur_th, sizeof(*th));
3282                         if (th->t_refcount <= 1)
3283                                 reiserfs_warning(sb, "reiserfs-2005",
3284                                                  "BAD: refcount <= 1, but "
3285                                                  "journal_info != 0");
3286                         return 0;
3287                 } else {
3288                         /* we've ended up with a handle from a different filesystem.
3289                          ** save it and restore on journal_end.  This should never
3290                          ** really happen...
3291                          */
3292                         reiserfs_warning(sb, "clm-2100",
3293                                          "nesting info a different FS");
3294                         th->t_handle_save = current->journal_info;
3295                         current->journal_info = th;
3296                 }
3297         } else {
3298                 current->journal_info = th;
3299         }
3300         ret = do_journal_begin_r(th, sb, nblocks, JBEGIN_REG);
3301         BUG_ON(current->journal_info != th);
3302
3303         /* I guess this boils down to being the reciprocal of clm-2100 above.
3304          * If do_journal_begin_r fails, we need to put it back, since journal_end
3305          * won't be called to do it. */
3306         if (ret)
3307                 current->journal_info = th->t_handle_save;
3308         else
3309                 BUG_ON(!th->t_refcount);
3310
3311         return ret;
3312 }
3313
3314 /*
3315 ** puts bh into the current transaction.  If it was already there, reorders removes the
3316 ** old pointers from the hash, and puts new ones in (to make sure replay happen in the right order).
3317 **
3318 ** if it was dirty, cleans and files onto the clean list.  I can't let it be dirty again until the
3319 ** transaction is committed.
3320 **
3321 ** if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len.
3322 */
3323 int journal_mark_dirty(struct reiserfs_transaction_handle *th,
3324                        struct super_block *sb, struct buffer_head *bh)
3325 {
3326         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3327         struct reiserfs_journal_cnode *cn = NULL;
3328         int count_already_incd = 0;
3329         int prepared = 0;
3330         BUG_ON(!th->t_trans_id);
3331
3332         PROC_INFO_INC(sb, journal.mark_dirty);
3333         if (th->t_trans_id != journal->j_trans_id) {
3334                 reiserfs_panic(th->t_super, "journal-1577",
3335                                "handle trans id %ld != current trans id %ld",
3336                                th->t_trans_id, journal->j_trans_id);
3337         }
3338
3339         sb->s_dirt = 1;
3340
3341         prepared = test_clear_buffer_journal_prepared(bh);
3342         clear_buffer_journal_restore_dirty(bh);
3343         /* already in this transaction, we are done */
3344         if (buffer_journaled(bh)) {
3345                 PROC_INFO_INC(sb, journal.mark_dirty_already);
3346                 return 0;
3347         }
3348
3349         /* this must be turned into a panic instead of a warning.  We can't allow
3350          ** a dirty or journal_dirty or locked buffer to be logged, as some changes
3351          ** could get to disk too early.  NOT GOOD.
3352          */
3353         if (!prepared || buffer_dirty(bh)) {
3354                 reiserfs_warning(sb, "journal-1777",
3355                                  "buffer %llu bad state "
3356                                  "%cPREPARED %cLOCKED %cDIRTY %cJDIRTY_WAIT",
3357                                  (unsigned long long)bh->b_blocknr,
3358                                  prepared ? ' ' : '!',
3359                                  buffer_locked(bh) ? ' ' : '!',
3360                                  buffer_dirty(bh) ? ' ' : '!',
3361                                  buffer_journal_dirty(bh) ? ' ' : '!');
3362         }
3363
3364         if (atomic_read(&(journal->j_wcount)) <= 0) {
3365                 reiserfs_warning(sb, "journal-1409",
3366                                  "returning because j_wcount was %d",
3367                                  atomic_read(&(journal->j_wcount)));
3368                 return 1;
3369         }
3370         /* this error means I've screwed up, and we've overflowed the transaction.
3371          ** Nothing can be done here, except make the FS readonly or panic.
3372          */
3373         if (journal->j_len >= journal->j_trans_max) {
3374                 reiserfs_panic(th->t_super, "journal-1413",
3375                                "j_len (%lu) is too big",
3376                                journal->j_len);
3377         }
3378
3379         if (buffer_journal_dirty(bh)) {
3380                 count_already_incd = 1;
3381                 PROC_INFO_INC(sb, journal.mark_dirty_notjournal);
3382                 clear_buffer_journal_dirty(bh);
3383         }
3384
3385         if (journal->j_len > journal->j_len_alloc) {
3386                 journal->j_len_alloc = journal->j_len + JOURNAL_PER_BALANCE_CNT;
3387         }
3388
3389         set_buffer_journaled(bh);
3390
3391         /* now put this guy on the end */
3392         if (!cn) {
3393                 cn = get_cnode(sb);
3394                 if (!cn) {
3395                         reiserfs_panic(sb, "journal-4", "get_cnode failed!");
3396                 }
3397
3398                 if (th->t_blocks_logged == th->t_blocks_allocated) {
3399                         th->t_blocks_allocated += JOURNAL_PER_BALANCE_CNT;
3400                         journal->j_len_alloc += JOURNAL_PER_BALANCE_CNT;
3401                 }
3402                 th->t_blocks_logged++;
3403                 journal->j_len++;
3404
3405                 cn->bh = bh;
3406                 cn->blocknr = bh->b_blocknr;
3407                 cn->sb = sb;
3408                 cn->jlist = NULL;
3409                 insert_journal_hash(journal->j_hash_table, cn);
3410                 if (!count_already_incd) {
3411                         get_bh(bh);
3412                 }
3413         }
3414         cn->next = NULL;
3415         cn->prev = journal->j_last;
3416         cn->bh = bh;
3417         if (journal->j_last) {
3418                 journal->j_last->next = cn;
3419                 journal->j_last = cn;
3420         } else {
3421                 journal->j_first = cn;
3422                 journal->j_last = cn;
3423         }
3424         return 0;
3425 }
3426
3427 int journal_end(struct reiserfs_transaction_handle *th,
3428                 struct super_block *sb, unsigned long nblocks)
3429 {
3430         if (!current->journal_info && th->t_refcount > 1)
3431                 reiserfs_warning(sb, "REISER-NESTING",
3432                                  "th NULL, refcount %d", th->t_refcount);
3433
3434         if (!th->t_trans_id) {
3435                 WARN_ON(1);
3436                 return -EIO;
3437         }
3438
3439         th->t_refcount--;
3440         if (th->t_refcount > 0) {
3441                 struct reiserfs_transaction_handle *cur_th =
3442                     current->journal_info;
3443
3444                 /* we aren't allowed to close a nested transaction on a different
3445                  ** filesystem from the one in the task struct
3446                  */
3447                 BUG_ON(cur_th->t_super != th->t_super);
3448
3449                 if (th != cur_th) {
3450                         memcpy(current->journal_info, th, sizeof(*th));
3451                         th->t_trans_id = 0;
3452                 }
3453                 return 0;
3454         } else {
3455                 return do_journal_end(th, sb, nblocks, 0);
3456         }
3457 }
3458
3459 /* removes from the current transaction, relsing and descrementing any counters.
3460 ** also files the removed buffer directly onto the clean list
3461 **
3462 ** called by journal_mark_freed when a block has been deleted
3463 **
3464 ** returns 1 if it cleaned and relsed the buffer. 0 otherwise
3465 */
3466 static int remove_from_transaction(struct super_block *sb,
3467                                    b_blocknr_t blocknr, int already_cleaned)
3468 {
3469         struct buffer_head *bh;
3470         struct reiserfs_journal_cnode *cn;
3471         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3472         int ret = 0;
3473
3474         cn = get_journal_hash_dev(sb, journal->j_hash_table, blocknr);
3475         if (!cn || !cn->bh) {
3476                 return ret;
3477         }
3478         bh = cn->bh;
3479         if (cn->prev) {
3480                 cn->prev->next = cn->next;
3481         }
3482         if (cn->next) {
3483                 cn->next->prev = cn->prev;
3484         }
3485         if (cn == journal->j_first) {
3486                 journal->j_first = cn->next;
3487         }
3488         if (cn == journal->j_last) {
3489                 journal->j_last = cn->prev;
3490         }
3491         if (bh)
3492                 remove_journal_hash(sb, journal->j_hash_table, NULL,
3493                                     bh->b_blocknr, 0);
3494         clear_buffer_journaled(bh);     /* don't log this one */
3495
3496         if (!already_cleaned) {
3497                 clear_buffer_journal_dirty(bh);
3498                 clear_buffer_dirty(bh);
3499                 clear_buffer_journal_test(bh);
3500                 put_bh(bh);
3501                 if (atomic_read(&(bh->b_count)) < 0) {
3502                         reiserfs_warning(sb, "journal-1752",
3503                                          "b_count < 0");
3504                 }
3505                 ret = 1;
3506         }
3507         journal->j_len--;
3508         journal->j_len_alloc--;
3509         free_cnode(sb, cn);
3510         return ret;
3511 }
3512
3513 /*
3514 ** for any cnode in a journal list, it can only be dirtied of all the
3515 ** transactions that include it are committed to disk.
3516 ** this checks through each transaction, and returns 1 if you are allowed to dirty,
3517 ** and 0 if you aren't
3518 **
3519 ** it is called by dirty_journal_list, which is called after flush_commit_list has gotten all the log
3520 ** blocks for a given transaction on disk
3521 **
3522 */
3523 static int can_dirty(struct reiserfs_journal_cnode *cn)
3524 {
3525         struct super_block *sb = cn->sb;
3526         b_blocknr_t blocknr = cn->blocknr;
3527         struct reiserfs_journal_cnode *cur = cn->hprev;
3528         int can_dirty = 1;
3529
3530         /* first test hprev.  These are all newer than cn, so any node here
3531          ** with the same block number and dev means this node can't be sent
3532          ** to disk right now.
3533          */
3534         while (cur && can_dirty) {
3535                 if (cur->jlist && cur->bh && cur->blocknr && cur->sb == sb &&
3536                     cur->blocknr == blocknr) {
3537                         can_dirty = 0;
3538                 }
3539                 cur = cur->hprev;
3540         }
3541         /* then test hnext.  These are all older than cn.  As long as they
3542          ** are committed to the log, it is safe to write cn to disk
3543          */
3544         cur = cn->hnext;
3545         while (cur && can_dirty) {
3546                 if (cur->jlist && cur->jlist->j_len > 0 &&
3547                     atomic_read(&(cur->jlist->j_commit_left)) > 0 && cur->bh &&
3548                     cur->blocknr && cur->sb == sb && cur->blocknr == blocknr) {
3549                         can_dirty = 0;
3550                 }
3551                 cur = cur->hnext;
3552         }
3553         return can_dirty;
3554 }
3555
3556 /* syncs the commit blocks, but does not force the real buffers to disk
3557 ** will wait until the current transaction is done/committed before returning
3558 */
3559 int journal_end_sync(struct reiserfs_transaction_handle *th,
3560                      struct super_block *sb, unsigned long nblocks)
3561 {
3562         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3563
3564         BUG_ON(!th->t_trans_id);
3565         /* you can sync while nested, very, very bad */
3566         BUG_ON(th->t_refcount > 1);
3567         if (journal->j_len == 0) {
3568                 reiserfs_prepare_for_journal(sb, SB_BUFFER_WITH_SB(sb),
3569                                              1);
3570                 journal_mark_dirty(th, sb, SB_BUFFER_WITH_SB(sb));
3571         }
3572         return do_journal_end(th, sb, nblocks, COMMIT_NOW | WAIT);
3573 }
3574
3575 /*
3576 ** writeback the pending async commits to disk
3577 */
3578 static void flush_async_commits(struct work_struct *work)
3579 {
3580         struct reiserfs_journal *journal =
3581                 container_of(work, struct reiserfs_journal, j_work.work);
3582         struct super_block *sb = journal->j_work_sb;
3583         struct reiserfs_journal_list *jl;
3584         struct list_head *entry;
3585
3586         reiserfs_write_lock(sb);
3587         if (!list_empty(&journal->j_journal_list)) {
3588                 /* last entry is the youngest, commit it and you get everything */
3589                 entry = journal->j_journal_list.prev;
3590                 jl = JOURNAL_LIST_ENTRY(entry);
3591                 flush_commit_list(sb, jl, 1);
3592         }
3593         reiserfs_write_unlock(sb);
3594 }
3595
3596 /*
3597 ** flushes any old transactions to disk
3598 ** ends the current transaction if it is too old
3599 */
3600 int reiserfs_flush_old_commits(struct super_block *sb)
3601 {
3602         time_t now;
3603         struct reiserfs_transaction_handle th;
3604         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3605
3606         now = get_seconds();
3607         /* safety check so we don't flush while we are replaying the log during
3608          * mount
3609          */
3610         if (list_empty(&journal->j_journal_list)) {
3611                 return 0;
3612         }
3613
3614         /* check the current transaction.  If there are no writers, and it is
3615          * too old, finish it, and force the commit blocks to disk
3616          */
3617         if (atomic_read(&journal->j_wcount) <= 0 &&
3618             journal->j_trans_start_time > 0 &&
3619             journal->j_len > 0 &&
3620             (now - journal->j_trans_start_time) > journal->j_max_trans_age) {
3621                 if (!journal_join(&th, sb, 1)) {
3622                         reiserfs_prepare_for_journal(sb,
3623                                                      SB_BUFFER_WITH_SB(sb),
3624                                                      1);
3625                         journal_mark_dirty(&th, sb,
3626                                            SB_BUFFER_WITH_SB(sb));
3627
3628                         /* we're only being called from kreiserfsd, it makes no sense to do
3629                          ** an async commit so that kreiserfsd can do it later
3630                          */
3631                         do_journal_end(&th, sb, 1, COMMIT_NOW | WAIT);
3632                 }
3633         }
3634         return sb->s_dirt;
3635 }
3636
3637 /*
3638 ** returns 0 if do_journal_end should return right away, returns 1 if do_journal_end should finish the commit
3639 **
3640 ** if the current transaction is too old, but still has writers, this will wait on j_join_wait until all
3641 ** the writers are done.  By the time it wakes up, the transaction it was called has already ended, so it just
3642 ** flushes the commit list and returns 0.
3643 **
3644 ** Won't batch when flush or commit_now is set.  Also won't batch when others are waiting on j_join_wait.
3645 **
3646 ** Note, we can't allow the journal_end to proceed while there are still writers in the log.
3647 */
3648 static int check_journal_end(struct reiserfs_transaction_handle *th,
3649                              struct super_block *sb, unsigned long nblocks,
3650                              int flags)
3651 {
3652
3653         time_t now;
3654         int flush = flags & FLUSH_ALL;
3655         int commit_now = flags & COMMIT_NOW;
3656         int wait_on_commit = flags & WAIT;
3657         struct reiserfs_journal_list *jl;
3658         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3659
3660         BUG_ON(!th->t_trans_id);
3661
3662         if (th->t_trans_id != journal->j_trans_id) {
3663                 reiserfs_panic(th->t_super, "journal-1577",
3664                                "handle trans id %ld != current trans id %ld",
3665                                th->t_trans_id, journal->j_trans_id);
3666         }
3667
3668         journal->j_len_alloc -= (th->t_blocks_allocated - th->t_blocks_logged);
3669         if (atomic_read(&(journal->j_wcount)) > 0) {    /* <= 0 is allowed.  unmounting might not call begin */
3670                 atomic_dec(&(journal->j_wcount));
3671         }
3672
3673         /* BUG, deal with case where j_len is 0, but people previously freed blocks need to be released
3674          ** will be dealt with by next transaction that actually writes something, but should be taken
3675          ** care of in this trans
3676          */
3677         BUG_ON(journal->j_len == 0);
3678
3679         /* if wcount > 0, and we are called to with flush or commit_now,
3680          ** we wait on j_join_wait.  We will wake up when the last writer has
3681          ** finished the transaction, and started it on its way to the disk.
3682          ** Then, we flush the commit or journal list, and just return 0
3683          ** because the rest of journal end was already done for this transaction.
3684          */
3685         if (atomic_read(&(journal->j_wcount)) > 0) {
3686                 if (flush || commit_now) {
3687                         unsigned trans_id;
3688
3689                         jl = journal->j_current_jl;
3690                         trans_id = jl->j_trans_id;
3691                         if (wait_on_commit)
3692                                 jl->j_state |= LIST_COMMIT_PENDING;
3693                         atomic_set(&(journal->j_jlock), 1);
3694                         if (flush) {
3695                                 journal->j_next_full_flush = 1;
3696                         }
3697                         unlock_journal(sb);
3698
3699                         /* sleep while the current transaction is still j_jlocked */
3700                         while (journal->j_trans_id == trans_id) {
3701                                 if (atomic_read(&journal->j_jlock)) {
3702                                         queue_log_writer(sb);
3703                                 } else {
3704                                         lock_journal(sb);
3705                                         if (journal->j_trans_id == trans_id) {
3706                                                 atomic_set(&(journal->j_jlock),
3707                                                            1);
3708                                         }
3709                                         unlock_journal(sb);
3710                                 }
3711                         }
3712                         BUG_ON(journal->j_trans_id == trans_id);
3713
3714                         if (commit_now
3715                             && journal_list_still_alive(sb, trans_id)
3716                             && wait_on_commit) {
3717                                 flush_commit_list(sb, jl, 1);
3718                         }
3719                         return 0;
3720                 }
3721                 unlock_journal(sb);
3722                 return 0;
3723         }
3724
3725         /* deal with old transactions where we are the last writers */
3726         now = get_seconds();
3727         if ((now - journal->j_trans_start_time) > journal->j_max_trans_age) {
3728                 commit_now = 1;
3729                 journal->j_next_async_flush = 1;
3730         }
3731         /* don't batch when someone is waiting on j_join_wait */
3732         /* don't batch when syncing the commit or flushing the whole trans */
3733         if (!(journal->j_must_wait > 0) && !(atomic_read(&(journal->j_jlock)))
3734             && !flush && !commit_now && (journal->j_len < journal->j_max_batch)
3735             && journal->j_len_alloc < journal->j_max_batch
3736             && journal->j_cnode_free > (journal->j_trans_max * 3)) {
3737                 journal->j_bcount++;
3738                 unlock_journal(sb);
3739                 return 0;
3740         }
3741
3742         if (journal->j_start > SB_ONDISK_JOURNAL_SIZE(sb)) {
3743                 reiserfs_panic(sb, "journal-003",
3744                                "j_start (%ld) is too high",
3745                                journal->j_start);
3746         }
3747         return 1;
3748 }
3749
3750 /*
3751 ** Does all the work that makes deleting blocks safe.
3752 ** when deleting a block mark BH_JNew, just remove it from the current transaction, clean it's buffer_head and move on.
3753 **
3754 ** otherwise:
3755 ** set a bit for the block in the journal bitmap.  That will prevent it from being allocated for unformatted nodes
3756 ** before this transaction has finished.
3757 **
3758 ** mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers.  That will prevent any old transactions with
3759 ** this block from trying to flush to the real location.  Since we aren't removing the cnode from the journal_list_hash,
3760 ** the block can't be reallocated yet.
3761 **
3762 ** Then remove it from the current transaction, decrementing any counters and filing it on the clean list.
3763 */
3764 int journal_mark_freed(struct reiserfs_transaction_handle *th,
3765                        struct super_block *sb, b_blocknr_t blocknr)
3766 {
3767         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3768         struct reiserfs_journal_cnode *cn = NULL;
3769         struct buffer_head *bh = NULL;
3770         struct reiserfs_list_bitmap *jb = NULL;
3771         int cleaned = 0;
3772         BUG_ON(!th->t_trans_id);
3773
3774         cn = get_journal_hash_dev(sb, journal->j_hash_table, blocknr);
3775         if (cn && cn->bh) {
3776                 bh = cn->bh;
3777                 get_bh(bh);
3778         }
3779         /* if it is journal new, we just remove it from this transaction */
3780         if (bh && buffer_journal_new(bh)) {
3781                 clear_buffer_journal_new(bh);
3782                 clear_prepared_bits(bh);
3783                 reiserfs_clean_and_file_buffer(bh);
3784                 cleaned = remove_from_transaction(sb, blocknr, cleaned);
3785         } else {
3786                 /* set the bit for this block in the journal bitmap for this transaction */
3787                 jb = journal->j_current_jl->j_list_bitmap;
3788                 if (!jb) {
3789                         reiserfs_panic(sb, "journal-1702",
3790                                        "journal_list_bitmap is NULL");
3791                 }
3792                 set_bit_in_list_bitmap(sb, blocknr, jb);
3793
3794                 /* Note, the entire while loop is not allowed to schedule.  */
3795
3796                 if (bh) {
3797                         clear_prepared_bits(bh);
3798                         reiserfs_clean_and_file_buffer(bh);
3799                 }
3800                 cleaned = remove_from_transaction(sb, blocknr, cleaned);
3801
3802                 /* find all older transactions with this block, make sure they don't try to write it out */
3803                 cn = get_journal_hash_dev(sb, journal->j_list_hash_table,
3804                                           blocknr);
3805                 while (cn) {
3806                         if (sb == cn->sb && blocknr == cn->blocknr) {
3807                                 set_bit(BLOCK_FREED, &cn->state);
3808                                 if (cn->bh) {
3809                                         if (!cleaned) {
3810                                                 /* remove_from_transaction will brelse the buffer if it was
3811                                                  ** in the current trans
3812                                                  */
3813                                                 clear_buffer_journal_dirty(cn->
3814                                                                            bh);
3815                                                 clear_buffer_dirty(cn->bh);
3816                                                 clear_buffer_journal_test(cn->
3817                                                                           bh);
3818                                                 cleaned = 1;
3819                                                 put_bh(cn->bh);
3820                                                 if (atomic_read
3821                                                     (&(cn->bh->b_count)) < 0) {
3822                                                         reiserfs_warning(sb,
3823                                                                  "journal-2138",
3824                                                                  "cn->bh->b_count < 0");
3825                                                 }
3826                                         }
3827                                         if (cn->jlist) {        /* since we are clearing the bh, we MUST dec nonzerolen */
3828                                                 atomic_dec(&
3829                                                            (cn->jlist->
3830                                                             j_nonzerolen));
3831                                         }
3832                                         cn->bh = NULL;
3833                                 }
3834                         }
3835                         cn = cn->hnext;
3836                 }
3837         }
3838
3839         if (bh)
3840                 release_buffer_page(bh); /* get_hash grabs the buffer */
3841         return 0;
3842 }
3843
3844 void reiserfs_update_inode_transaction(struct inode *inode)
3845 {
3846         struct reiserfs_journal *journal = SB_JOURNAL(inode->i_sb);
3847         REISERFS_I(inode)->i_jl = journal->j_current_jl;
3848         REISERFS_I(inode)->i_trans_id = journal->j_trans_id;
3849 }
3850
3851 /*
3852  * returns -1 on error, 0 if no commits/barriers were done and 1
3853  * if a transaction was actually committed and the barrier was done
3854  */
3855 static int __commit_trans_jl(struct inode *inode, unsigned long id,
3856                              struct reiserfs_journal_list *jl)
3857 {
3858         struct reiserfs_transaction_handle th;
3859         struct super_block *sb = inode->i_sb;
3860         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3861         int ret = 0;
3862
3863         /* is it from the current transaction, or from an unknown transaction? */
3864         if (id == journal->j_trans_id) {
3865                 jl = journal->j_current_jl;
3866                 /* try to let other writers come in and grow this transaction */
3867                 let_transaction_grow(sb, id);
3868                 if (journal->j_trans_id != id) {
3869                         goto flush_commit_only;
3870                 }
3871
3872                 ret = journal_begin(&th, sb, 1);
3873                 if (ret)
3874                         return ret;
3875
3876                 /* someone might have ended this transaction while we joined */
3877                 if (journal->j_trans_id != id) {
3878                         reiserfs_prepare_for_journal(sb, SB_BUFFER_WITH_SB(sb),
3879                                                      1);
3880                         journal_mark_dirty(&th, sb, SB_BUFFER_WITH_SB(sb));
3881                         ret = journal_end(&th, sb, 1);
3882                         goto flush_commit_only;
3883                 }
3884
3885                 ret = journal_end_sync(&th, sb, 1);
3886                 if (!ret)
3887                         ret = 1;
3888
3889         } else {
3890                 /* this gets tricky, we have to make sure the journal list in
3891                  * the inode still exists.  We know the list is still around
3892                  * if we've got a larger transaction id than the oldest list
3893                  */
3894               flush_commit_only:
3895                 if (journal_list_still_alive(inode->i_sb, id)) {
3896                         /*
3897                          * we only set ret to 1 when we know for sure
3898                          * the barrier hasn't been started yet on the commit
3899                          * block.
3900                          */
3901                         if (atomic_read(&jl->j_commit_left) > 1)
3902                                 ret = 1;
3903                         flush_commit_list(sb, jl, 1);
3904                         if (journal->j_errno)
3905                                 ret = journal->j_errno;
3906                 }
3907         }
3908         /* otherwise the list is gone, and long since committed */
3909         return ret;
3910 }
3911
3912 int reiserfs_commit_for_inode(struct inode *inode)
3913 {
3914         unsigned int id = REISERFS_I(inode)->i_trans_id;
3915         struct reiserfs_journal_list *jl = REISERFS_I(inode)->i_jl;
3916
3917         /* for the whole inode, assume unset id means it was
3918          * changed in the current transaction.  More conservative
3919          */
3920         if (!id || !jl) {
3921                 reiserfs_update_inode_transaction(inode);
3922                 id = REISERFS_I(inode)->i_trans_id;
3923                 /* jl will be updated in __commit_trans_jl */
3924         }
3925
3926         return __commit_trans_jl(inode, id, jl);
3927 }
3928
3929 void reiserfs_restore_prepared_buffer(struct super_block *sb,
3930                                       struct buffer_head *bh)
3931 {
3932         struct reiserfs_journal *journal = SB_JOURNAL(sb);
3933         PROC_INFO_INC(sb, journal.restore_prepared);
3934         if (!bh) {
3935                 return;
3936         }
3937         if (test_clear_buffer_journal_restore_dirty(bh) &&
3938             buffer_journal_dirty(bh)) {
3939                 struct reiserfs_journal_cnode *cn;
3940                 cn = get_journal_hash_dev(sb,
3941                                           journal->j_list_hash_table,
3942                                           bh->b_blocknr);
3943                 if (cn && can_dirty(cn)) {
3944                         set_buffer_journal_test(bh);
3945                         mark_buffer_dirty(bh);
3946                 }
3947         }
3948         clear_buffer_journal_prepared(bh);
3949 }
3950
3951 extern struct tree_balance *cur_tb;
3952 /*
3953 ** before we can change a metadata block, we have to make sure it won't
3954 ** be written to disk while we are altering it.  So, we must:
3955 ** clean it
3956 ** wait on it.
3957 **
3958 */
3959 int reiserfs_prepare_for_journal(struct super_block *sb,
3960                                  struct buffer_head *bh, int wait)
3961 {
3962         PROC_INFO_INC(sb, journal.prepare);
3963
3964         if (!trylock_buffer(bh)) {
3965                 if (!wait)
3966                         return 0;
3967                 lock_buffer(bh);
3968         }
3969         set_buffer_journal_prepared(bh);
3970         if (test_clear_buffer_dirty(bh) && buffer_journal_dirty(bh)) {
3971                 clear_buffer_journal_test(bh);
3972                 set_buffer_journal_restore_dirty(bh);
3973         }
3974         unlock_buffer(bh);
3975         return 1;
3976 }
3977
3978 static void flush_old_journal_lists(struct super_block *s)
3979 {
3980         struct reiserfs_journal *journal = SB_JOURNAL(s);
3981         struct reiserfs_journal_list *jl;
3982         struct list_head *entry;
3983         time_t now = get_seconds();
3984
3985         while (!list_empty(&journal->j_journal_list)) {
3986                 entry = journal->j_journal_list.next;
3987                 jl = JOURNAL_LIST_ENTRY(entry);
3988                 /* this check should always be run, to send old lists to disk */
3989                 if (jl->j_timestamp < (now - (JOURNAL_MAX_TRANS_AGE * 4)) &&
3990                     atomic_read(&jl->j_commit_left) == 0 &&
3991                     test_transaction(s, jl)) {
3992                         flush_used_journal_lists(s, jl);
3993                 } else {
3994                         break;
3995                 }
3996         }
3997 }
3998
3999 /*
4000 ** long and ugly.  If flush, will not return until all commit
4001 ** blocks and all real buffers in the trans are on disk.
4002 ** If no_async, won't return until all commit blocks are on disk.
4003 **
4004 ** keep reading, there are comments as you go along
4005 **
4006 ** If the journal is aborted, we just clean up. Things like flushing
4007 ** journal lists, etc just won't happen.
4008 */
4009 static int do_journal_end(struct reiserfs_transaction_handle *th,
4010                           struct super_block *sb, unsigned long nblocks,
4011                           int flags)
4012 {
4013         struct reiserfs_journal *journal = SB_JOURNAL(sb);
4014         struct reiserfs_journal_cnode *cn, *next, *jl_cn;
4015         struct reiserfs_journal_cnode *last_cn = NULL;
4016         struct reiserfs_journal_desc *desc;
4017         struct reiserfs_journal_commit *commit;
4018         struct buffer_head *c_bh;       /* commit bh */
4019         struct buffer_head *d_bh;       /* desc bh */
4020         int cur_write_start = 0;        /* start index of current log write */
4021         int old_start;
4022         int i;
4023         int flush;
4024         int wait_on_commit;
4025         struct reiserfs_journal_list *jl, *temp_jl;
4026         struct list_head *entry, *safe;
4027         unsigned long jindex;
4028         unsigned int commit_trans_id;
4029         int trans_half;
4030
4031         BUG_ON(th->t_refcount > 1);
4032         BUG_ON(!th->t_trans_id);
4033
4034         /* protect flush_older_commits from doing mistakes if the
4035            transaction ID counter gets overflowed.  */
4036         if (th->t_trans_id == ~0U)
4037                 flags |= FLUSH_ALL | COMMIT_NOW | WAIT;
4038         flush = flags & FLUSH_ALL;
4039         wait_on_commit = flags & WAIT;
4040
4041         put_fs_excl();
4042         current->journal_info = th->t_handle_save;
4043         reiserfs_check_lock_depth(sb, "journal end");
4044         if (journal->j_len == 0) {
4045                 reiserfs_prepare_for_journal(sb, SB_BUFFER_WITH_SB(sb),
4046                                              1);
4047                 journal_mark_dirty(th, sb, SB_BUFFER_WITH_SB(sb));
4048         }
4049
4050         lock_journal(sb);
4051         if (journal->j_next_full_flush) {
4052                 flags |= FLUSH_ALL;
4053                 flush = 1;
4054         }
4055         if (journal->j_next_async_flush) {
4056                 flags |= COMMIT_NOW | WAIT;
4057                 wait_on_commit = 1;
4058         }
4059
4060         /* check_journal_end locks the journal, and unlocks if it does not return 1
4061          ** it tells us if we should continue with the journal_end, or just return
4062          */
4063         if (!check_journal_end(th, sb, nblocks, flags)) {
4064                 sb->s_dirt = 1;
4065                 wake_queued_writers(sb);
4066                 reiserfs_async_progress_wait(sb);
4067                 goto out;
4068         }
4069
4070         /* check_journal_end might set these, check again */
4071         if (journal->j_next_full_flush) {
4072                 flush = 1;
4073         }
4074
4075         /*
4076          ** j must wait means we have to flush the log blocks, and the real blocks for
4077          ** this transaction
4078          */
4079         if (journal->j_must_wait > 0) {
4080                 flush = 1;
4081         }
4082 #ifdef REISERFS_PREALLOCATE
4083         /* quota ops might need to nest, setup the journal_info pointer for them
4084          * and raise the refcount so that it is > 0. */
4085         current->journal_info = th;
4086         th->t_refcount++;
4087         reiserfs_discard_all_prealloc(th);      /* it should not involve new blocks into
4088                                                  * the transaction */
4089         th->t_refcount--;
4090         current->journal_info = th->t_handle_save;
4091 #endif
4092
4093         /* setup description block */
4094         d_bh =
4095             journal_getblk(sb,
4096                            SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
4097                            journal->j_start);
4098         set_buffer_uptodate(d_bh);
4099         desc = (struct reiserfs_journal_desc *)(d_bh)->b_data;
4100         memset(d_bh->b_data, 0, d_bh->b_size);
4101         memcpy(get_journal_desc_magic(d_bh), JOURNAL_DESC_MAGIC, 8);
4102         set_desc_trans_id(desc, journal->j_trans_id);
4103
4104         /* setup commit block.  Don't write (keep it clean too) this one until after everyone else is written */
4105         c_bh = journal_getblk(sb, SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
4106                               ((journal->j_start + journal->j_len +
4107                                 1) % SB_ONDISK_JOURNAL_SIZE(sb)));
4108         commit = (struct reiserfs_journal_commit *)c_bh->b_data;
4109         memset(c_bh->b_data, 0, c_bh->b_size);
4110         set_commit_trans_id(commit, journal->j_trans_id);
4111         set_buffer_uptodate(c_bh);
4112
4113         /* init this journal list */
4114         jl = journal->j_current_jl;
4115
4116         /* we lock the commit before doing anything because
4117          * we want to make sure nobody tries to run flush_commit_list until
4118          * the new transaction is fully setup, and we've already flushed the
4119          * ordered bh list
4120          */
4121         reiserfs_mutex_lock_safe(&jl->j_commit_mutex, sb);
4122
4123         /* save the transaction id in case we need to commit it later */
4124         commit_trans_id = jl->j_trans_id;
4125
4126         atomic_set(&jl->j_older_commits_done, 0);
4127         jl->j_trans_id = journal->j_trans_id;
4128         jl->j_timestamp = journal->j_trans_start_time;
4129         jl->j_commit_bh = c_bh;
4130         jl->j_start = journal->j_start;
4131         jl->j_len = journal->j_len;
4132         atomic_set(&jl->j_nonzerolen, journal->j_len);
4133         atomic_set(&jl->j_commit_left, journal->j_len + 2);
4134         jl->j_realblock = NULL;
4135
4136         /* The ENTIRE FOR LOOP MUST not cause schedule to occur.
4137          **  for each real block, add it to the journal list hash,
4138          ** copy into real block index array in the commit or desc block
4139          */
4140         trans_half = journal_trans_half(sb->s_blocksize);
4141         for (i = 0, cn = journal->j_first; cn; cn = cn->next, i++) {
4142                 if (buffer_journaled(cn->bh)) {
4143                         jl_cn = get_cnode(sb);
4144                         if (!jl_cn) {
4145                                 reiserfs_panic(sb, "journal-1676",
4146                                                "get_cnode returned NULL");
4147                         }
4148                         if (i == 0) {
4149                                 jl->j_realblock = jl_cn;
4150                         }
4151                         jl_cn->prev = last_cn;
4152                         jl_cn->next = NULL;
4153                         if (last_cn) {
4154                                 last_cn->next = jl_cn;
4155                         }
4156                         last_cn = jl_cn;
4157                         /* make sure the block we are trying to log is not a block
4158                            of journal or reserved area */
4159
4160                         if (is_block_in_log_or_reserved_area
4161                             (sb, cn->bh->b_blocknr)) {
4162                                 reiserfs_panic(sb, "journal-2332",
4163                                                "Trying to log block %lu, "
4164                                                "which is a log block",
4165                                                cn->bh->b_blocknr);
4166                         }
4167                         jl_cn->blocknr = cn->bh->b_blocknr;
4168                         jl_cn->state = 0;
4169                         jl_cn->sb = sb;
4170                         jl_cn->bh = cn->bh;
4171                         jl_cn->jlist = jl;
4172                         insert_journal_hash(journal->j_list_hash_table, jl_cn);
4173                         if (i < trans_half) {
4174                                 desc->j_realblock[i] =
4175                                     cpu_to_le32(cn->bh->b_blocknr);
4176                         } else {
4177                                 commit->j_realblock[i - trans_half] =
4178                                     cpu_to_le32(cn->bh->b_blocknr);
4179                         }
4180                 } else {
4181                         i--;
4182                 }
4183         }
4184         set_desc_trans_len(desc, journal->j_len);
4185         set_desc_mount_id(desc, journal->j_mount_id);
4186         set_desc_trans_id(desc, journal->j_trans_id);
4187         set_commit_trans_len(commit, journal->j_len);
4188
4189         /* special check in case all buffers in the journal were marked for not logging */
4190         BUG_ON(journal->j_len == 0);
4191
4192         /* we're about to dirty all the log blocks, mark the description block
4193          * dirty now too.  Don't mark the commit block dirty until all the
4194          * others are on disk
4195          */
4196         mark_buffer_dirty(d_bh);
4197
4198         /* first data block is j_start + 1, so add one to cur_write_start wherever you use it */
4199         cur_write_start = journal->j_start;
4200         cn = journal->j_first;
4201         jindex = 1;             /* start at one so we don't get the desc again */
4202         while (cn) {
4203                 clear_buffer_journal_new(cn->bh);
4204                 /* copy all the real blocks into log area.  dirty log blocks */
4205                 if (buffer_journaled(cn->bh)) {
4206                         struct buffer_head *tmp_bh;
4207                         char *addr;
4208                         struct page *page;
4209                         tmp_bh =
4210                             journal_getblk(sb,
4211                                            SB_ONDISK_JOURNAL_1st_BLOCK(sb) +
4212                                            ((cur_write_start +
4213                                              jindex) %
4214                                             SB_ONDISK_JOURNAL_SIZE(sb)));
4215                         set_buffer_uptodate(tmp_bh);
4216                         page = cn->bh->b_page;
4217                         addr = kmap(page);
4218                         memcpy(tmp_bh->b_data,
4219                                addr + offset_in_page(cn->bh->b_data),
4220                                cn->bh->b_size);
4221                         kunmap(page);
4222                         mark_buffer_dirty(tmp_bh);
4223                         jindex++;
4224                         set_buffer_journal_dirty(cn->bh);
4225                         clear_buffer_journaled(cn->bh);
4226                 } else {
4227                         /* JDirty cleared sometime during transaction.  don't log this one */
4228                         reiserfs_warning(sb, "journal-2048",
4229                                          "BAD, buffer in journal hash, "
4230                                          "but not JDirty!");
4231                         brelse(cn->bh);
4232                 }
4233                 next = cn->next;
4234                 free_cnode(sb, cn);
4235                 cn = next;
4236                 reiserfs_write_unlock(sb);
4237                 cond_resched();
4238                 reiserfs_write_lock(sb);
4239         }
4240
4241         /* we are done  with both the c_bh and d_bh, but
4242          ** c_bh must be written after all other commit blocks,
4243          ** so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1.
4244          */
4245
4246         journal->j_current_jl = alloc_journal_list(sb);
4247
4248         /* now it is safe to insert this transaction on the main list */
4249         list_add_tail(&jl->j_list, &journal->j_journal_list);
4250         list_add_tail(&jl->j_working_list, &journal->j_working_list);
4251         journal->j_num_work_lists++;
4252
4253         /* reset journal values for the next transaction */
4254         old_start = journal->j_start;
4255         journal->j_start =
4256             (journal->j_start + journal->j_len +
4257              2) % SB_ONDISK_JOURNAL_SIZE(sb);
4258         atomic_set(&(journal->j_wcount), 0);
4259         journal->j_bcount = 0;
4260         journal->j_last = NULL;
4261         journal->j_first = NULL;
4262         journal->j_len = 0;
4263         journal->j_trans_start_time = 0;
4264         /* check for trans_id overflow */
4265         if (++journal->j_trans_id == 0)
4266                 journal->j_trans_id = 10;
4267         journal->j_current_jl->j_trans_id = journal->j_trans_id;
4268         journal->j_must_wait = 0;
4269         journal->j_len_alloc = 0;
4270         journal->j_next_full_flush = 0;
4271         journal->j_next_async_flush = 0;
4272         init_journal_hash(sb);
4273
4274         // make sure reiserfs_add_jh sees the new current_jl before we
4275         // write out the tails
4276         smp_mb();
4277
4278         /* tail conversion targets have to hit the disk before we end the
4279          * transaction.  Otherwise a later transaction might repack the tail
4280          * before this transaction commits, leaving the data block unflushed and
4281          * clean, if we crash before the later transaction commits, the data block
4282          * is lost.
4283          */
4284         if (!list_empty(&jl->j_tail_bh_list)) {
4285                 reiserfs_write_unlock(sb);
4286                 write_ordered_buffers(&journal->j_dirty_buffers_lock,
4287                                       journal, jl, &jl->j_tail_bh_list);
4288                 reiserfs_write_lock(sb);
4289         }
4290         BUG_ON(!list_empty(&jl->j_tail_bh_list));
4291         mutex_unlock(&jl->j_commit_mutex);
4292
4293         /* honor the flush wishes from the caller, simple commits can
4294          ** be done outside the journal lock, they are done below
4295          **
4296          ** if we don't flush the commit list right now, we put it into
4297          ** the work queue so the people waiting on the async progress work
4298          ** queue don't wait for this proc to flush journal lists and such.
4299          */
4300         if (flush) {
4301                 flush_commit_list(sb, jl, 1);
4302                 flush_journal_list(sb, jl, 1);
4303         } else if (!(jl->j_state & LIST_COMMIT_PENDING))
4304                 queue_delayed_work(commit_wq, &journal->j_work, HZ / 10);
4305
4306         /* if the next transaction has any chance of wrapping, flush
4307          ** transactions that might get overwritten.  If any journal lists are very
4308          ** old flush them as well.
4309          */
4310       first_jl:
4311         list_for_each_safe(entry, safe, &journal->j_journal_list) {
4312                 temp_jl = JOURNAL_LIST_ENTRY(entry);
4313                 if (journal->j_start <= temp_jl->j_start) {
4314                         if ((journal->j_start + journal->j_trans_max + 1) >=
4315                             temp_jl->j_start) {
4316                                 flush_used_journal_lists(sb, temp_jl);
4317                                 goto first_jl;
4318                         } else if ((journal->j_start +
4319                                     journal->j_trans_max + 1) <
4320                                    SB_ONDISK_JOURNAL_SIZE(sb)) {
4321                                 /* if we don't cross into the next transaction and we don't
4322                                  * wrap, there is no way we can overlap any later transactions
4323                                  * break now
4324                                  */
4325                                 break;
4326                         }
4327                 } else if ((journal->j_start +
4328                             journal->j_trans_max + 1) >
4329                            SB_ONDISK_JOURNAL_SIZE(sb)) {
4330                         if (((journal->j_start + journal->j_trans_max + 1) %
4331                              SB_ONDISK_JOURNAL_SIZE(sb)) >=
4332                             temp_jl->j_start) {
4333                                 flush_used_journal_lists(sb, temp_jl);
4334                                 goto first_jl;
4335                         } else {
4336                                 /* we don't overlap anything from out start to the end of the
4337                                  * log, and our wrapped portion doesn't overlap anything at
4338                                  * the start of the log.  We can break
4339                                  */
4340                                 break;
4341                         }
4342                 }
4343         }
4344         flush_old_journal_lists(sb);
4345
4346         journal->j_current_jl->j_list_bitmap =
4347             get_list_bitmap(sb, journal->j_current_jl);
4348
4349         if (!(journal->j_current_jl->j_list_bitmap)) {
4350                 reiserfs_panic(sb, "journal-1996",
4351                                "could not get a list bitmap");
4352         }
4353
4354         atomic_set(&(journal->j_jlock), 0);
4355         unlock_journal(sb);
4356         /* wake up any body waiting to join. */
4357         clear_bit(J_WRITERS_QUEUED, &journal->j_state);
4358         wake_up(&(journal->j_join_wait));
4359
4360         if (!flush && wait_on_commit &&
4361             journal_list_still_alive(sb, commit_trans_id)) {
4362                 flush_commit_list(sb, jl, 1);
4363         }
4364       out:
4365         reiserfs_check_lock_depth(sb, "journal end2");
4366
4367         memset(th, 0, sizeof(*th));
4368         /* Re-set th->t_super, so we can properly keep track of how many
4369          * persistent transactions there are. We need to do this so if this
4370          * call is part of a failed restart_transaction, we can free it later */
4371         th->t_super = sb;
4372
4373         return journal->j_errno;
4374 }
4375
4376 /* Send the file system read only and refuse new transactions */
4377 void reiserfs_abort_journal(struct super_block *sb, int errno)
4378 {
4379         struct reiserfs_journal *journal = SB_JOURNAL(sb);
4380         if (test_bit(J_ABORTED, &journal->j_state))
4381                 return;
4382
4383         if (!journal->j_errno)
4384                 journal->j_errno = errno;
4385
4386         sb->s_flags |= MS_RDONLY;
4387         set_bit(J_ABORTED, &journal->j_state);
4388
4389 #ifdef CONFIG_REISERFS_CHECK
4390         dump_stack();
4391 #endif
4392 }
4393