1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Write ahead logging implementation copyright Chris Mason 2000 4 * 5 * The background commits make this code very interrelated, and 6 * overly complex. I need to rethink things a bit....The major players: 7 * 8 * journal_begin -- call with the number of blocks you expect to log. 9 * If the current transaction is too 10 * old, it will block until the current transaction is 11 * finished, and then start a new one. 12 * Usually, your transaction will get joined in with 13 * previous ones for speed. 14 * 15 * journal_join -- same as journal_begin, but won't block on the current 16 * transaction regardless of age. Don't ever call 17 * this. Ever. There are only two places it should be 18 * called from, and they are both inside this file. 19 * 20 * journal_mark_dirty -- adds blocks into this transaction. clears any flags 21 * that might make them get sent to disk 22 * and then marks them BH_JDirty. Puts the buffer head 23 * into the current transaction hash. 24 * 25 * journal_end -- if the current transaction is batchable, it does nothing 26 * otherwise, it could do an async/synchronous commit, or 27 * a full flush of all log and real blocks in the 28 * transaction. 29 * 30 * flush_old_commits -- if the current transaction is too old, it is ended and 31 * commit blocks are sent to disk. Forces commit blocks 32 * to disk for all backgrounded commits that have been 33 * around too long. 34 * -- Note, if you call this as an immediate flush from 35 * within kupdate, it will ignore the immediate flag 36 */ 37 38 #include <linux/time.h> 39 #include <linux/semaphore.h> 40 #include <linux/vmalloc.h> 41 #include "reiserfs.h" 42 #include <linux/kernel.h> 43 #include <linux/errno.h> 44 #include <linux/fcntl.h> 45 #include <linux/stat.h> 46 #include <linux/string.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 #include <linux/slab.h> 54 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 60 /* must be correct to keep the desc and commit structs at 4k */ 61 #define JOURNAL_TRANS_HALF 1018 62 #define BUFNR 64 /*read ahead */ 63 64 /* cnode stat bits. Move these into reiserfs_fs.h */ 65 66 /* this block was freed, and can't be written. */ 67 #define BLOCK_FREED 2 68 /* this block was freed during this transaction, and can't be written */ 69 #define BLOCK_FREED_HOLDER 3 70 71 /* used in flush_journal_list */ 72 #define BLOCK_NEEDS_FLUSH 4 73 #define BLOCK_DIRTIED 5 74 75 /* journal list state bits */ 76 #define LIST_TOUCHED 1 77 #define LIST_DIRTY 2 78 #define LIST_COMMIT_PENDING 4 /* someone will commit this list */ 79 80 /* flags for do_journal_end */ 81 #define FLUSH_ALL 1 /* flush commit and real blocks */ 82 #define COMMIT_NOW 2 /* end and commit this transaction */ 83 #define WAIT 4 /* wait for the log blocks to hit the disk */ 84 85 static int do_journal_end(struct reiserfs_transaction_handle *, int flags); 86 static int flush_journal_list(struct super_block *s, 87 struct reiserfs_journal_list *jl, int flushall); 88 static int flush_commit_list(struct super_block *s, 89 struct reiserfs_journal_list *jl, int flushall); 90 static int can_dirty(struct reiserfs_journal_cnode *cn); 91 static int journal_join(struct reiserfs_transaction_handle *th, 92 struct super_block *sb); 93 static void release_journal_dev(struct super_block *super, 94 struct reiserfs_journal *journal); 95 static void dirty_one_transaction(struct super_block *s, 96 struct reiserfs_journal_list *jl); 97 static void flush_async_commits(struct work_struct *work); 98 static void queue_log_writer(struct super_block *s); 99 100 /* values for join in do_journal_begin_r */ 101 enum { 102 JBEGIN_REG = 0, /* regular journal begin */ 103 /* join the running transaction if at all possible */ 104 JBEGIN_JOIN = 1, 105 /* called from cleanup code, ignores aborted flag */ 106 JBEGIN_ABORT = 2, 107 }; 108 109 static int do_journal_begin_r(struct reiserfs_transaction_handle *th, 110 struct super_block *sb, 111 unsigned long nblocks, int join); 112 113 static void init_journal_hash(struct super_block *sb) 114 { 115 struct reiserfs_journal *journal = SB_JOURNAL(sb); 116 memset(journal->j_hash_table, 0, 117 JOURNAL_HASH_SIZE * sizeof(struct reiserfs_journal_cnode *)); 118 } 119 120 /* 121 * clears BH_Dirty and sticks the buffer on the clean list. Called because 122 * I can't allow refile_buffer to make schedule happen after I've freed a 123 * block. Look at remove_from_transaction and journal_mark_freed for 124 * more details. 125 */ 126 static int reiserfs_clean_and_file_buffer(struct buffer_head *bh) 127 { 128 if (bh) { 129 clear_buffer_dirty(bh); 130 clear_buffer_journal_test(bh); 131 } 132 return 0; 133 } 134 135 static struct reiserfs_bitmap_node *allocate_bitmap_node(struct super_block 136 *sb) 137 { 138 struct reiserfs_bitmap_node *bn; 139 static int id; 140 141 bn = kmalloc(sizeof(struct reiserfs_bitmap_node), GFP_NOFS); 142 if (!bn) { 143 return NULL; 144 } 145 bn->data = kzalloc(sb->s_blocksize, GFP_NOFS); 146 if (!bn->data) { 147 kfree(bn); 148 return NULL; 149 } 150 bn->id = id++; 151 INIT_LIST_HEAD(&bn->list); 152 return bn; 153 } 154 155 static struct reiserfs_bitmap_node *get_bitmap_node(struct super_block *sb) 156 { 157 struct reiserfs_journal *journal = SB_JOURNAL(sb); 158 struct reiserfs_bitmap_node *bn = NULL; 159 struct list_head *entry = journal->j_bitmap_nodes.next; 160 161 journal->j_used_bitmap_nodes++; 162 repeat: 163 164 if (entry != &journal->j_bitmap_nodes) { 165 bn = list_entry(entry, struct reiserfs_bitmap_node, list); 166 list_del(entry); 167 memset(bn->data, 0, sb->s_blocksize); 168 journal->j_free_bitmap_nodes--; 169 return bn; 170 } 171 bn = allocate_bitmap_node(sb); 172 if (!bn) { 173 yield(); 174 goto repeat; 175 } 176 return bn; 177 } 178 static inline void free_bitmap_node(struct super_block *sb, 179 struct reiserfs_bitmap_node *bn) 180 { 181 struct reiserfs_journal *journal = SB_JOURNAL(sb); 182 journal->j_used_bitmap_nodes--; 183 if (journal->j_free_bitmap_nodes > REISERFS_MAX_BITMAP_NODES) { 184 kfree(bn->data); 185 kfree(bn); 186 } else { 187 list_add(&bn->list, &journal->j_bitmap_nodes); 188 journal->j_free_bitmap_nodes++; 189 } 190 } 191 192 static void allocate_bitmap_nodes(struct super_block *sb) 193 { 194 int i; 195 struct reiserfs_journal *journal = SB_JOURNAL(sb); 196 struct reiserfs_bitmap_node *bn = NULL; 197 for (i = 0; i < REISERFS_MIN_BITMAP_NODES; i++) { 198 bn = allocate_bitmap_node(sb); 199 if (bn) { 200 list_add(&bn->list, &journal->j_bitmap_nodes); 201 journal->j_free_bitmap_nodes++; 202 } else { 203 /* this is ok, we'll try again when more are needed */ 204 break; 205 } 206 } 207 } 208 209 static int set_bit_in_list_bitmap(struct super_block *sb, 210 b_blocknr_t block, 211 struct reiserfs_list_bitmap *jb) 212 { 213 unsigned int bmap_nr = block / (sb->s_blocksize << 3); 214 unsigned int bit_nr = block % (sb->s_blocksize << 3); 215 216 if (!jb->bitmaps[bmap_nr]) { 217 jb->bitmaps[bmap_nr] = get_bitmap_node(sb); 218 } 219 set_bit(bit_nr, (unsigned long *)jb->bitmaps[bmap_nr]->data); 220 return 0; 221 } 222 223 static void cleanup_bitmap_list(struct super_block *sb, 224 struct reiserfs_list_bitmap *jb) 225 { 226 int i; 227 if (jb->bitmaps == NULL) 228 return; 229 230 for (i = 0; i < reiserfs_bmap_count(sb); i++) { 231 if (jb->bitmaps[i]) { 232 free_bitmap_node(sb, jb->bitmaps[i]); 233 jb->bitmaps[i] = NULL; 234 } 235 } 236 } 237 238 /* 239 * only call this on FS unmount. 240 */ 241 static int free_list_bitmaps(struct super_block *sb, 242 struct reiserfs_list_bitmap *jb_array) 243 { 244 int i; 245 struct reiserfs_list_bitmap *jb; 246 for (i = 0; i < JOURNAL_NUM_BITMAPS; i++) { 247 jb = jb_array + i; 248 jb->journal_list = NULL; 249 cleanup_bitmap_list(sb, jb); 250 vfree(jb->bitmaps); 251 jb->bitmaps = NULL; 252 } 253 return 0; 254 } 255 256 static int free_bitmap_nodes(struct super_block *sb) 257 { 258 struct reiserfs_journal *journal = SB_JOURNAL(sb); 259 struct list_head *next = journal->j_bitmap_nodes.next; 260 struct reiserfs_bitmap_node *bn; 261 262 while (next != &journal->j_bitmap_nodes) { 263 bn = list_entry(next, struct reiserfs_bitmap_node, list); 264 list_del(next); 265 kfree(bn->data); 266 kfree(bn); 267 next = journal->j_bitmap_nodes.next; 268 journal->j_free_bitmap_nodes--; 269 } 270 271 return 0; 272 } 273 274 /* 275 * get memory for JOURNAL_NUM_BITMAPS worth of bitmaps. 276 * jb_array is the array to be filled in. 277 */ 278 int reiserfs_allocate_list_bitmaps(struct super_block *sb, 279 struct reiserfs_list_bitmap *jb_array, 280 unsigned int bmap_nr) 281 { 282 int i; 283 int failed = 0; 284 struct reiserfs_list_bitmap *jb; 285 int mem = bmap_nr * sizeof(struct reiserfs_bitmap_node *); 286 287 for (i = 0; i < JOURNAL_NUM_BITMAPS; i++) { 288 jb = jb_array + i; 289 jb->journal_list = NULL; 290 jb->bitmaps = vzalloc(mem); 291 if (!jb->bitmaps) { 292 reiserfs_warning(sb, "clm-2000", "unable to " 293 "allocate bitmaps for journal lists"); 294 failed = 1; 295 break; 296 } 297 } 298 if (failed) { 299 free_list_bitmaps(sb, jb_array); 300 return -1; 301 } 302 return 0; 303 } 304 305 /* 306 * find an available list bitmap. If you can't find one, flush a commit list 307 * and try again 308 */ 309 static struct reiserfs_list_bitmap *get_list_bitmap(struct super_block *sb, 310 struct reiserfs_journal_list 311 *jl) 312 { 313 int i, j; 314 struct reiserfs_journal *journal = SB_JOURNAL(sb); 315 struct reiserfs_list_bitmap *jb = NULL; 316 317 for (j = 0; j < (JOURNAL_NUM_BITMAPS * 3); j++) { 318 i = journal->j_list_bitmap_index; 319 journal->j_list_bitmap_index = (i + 1) % JOURNAL_NUM_BITMAPS; 320 jb = journal->j_list_bitmap + i; 321 if (journal->j_list_bitmap[i].journal_list) { 322 flush_commit_list(sb, 323 journal->j_list_bitmap[i]. 324 journal_list, 1); 325 if (!journal->j_list_bitmap[i].journal_list) { 326 break; 327 } 328 } else { 329 break; 330 } 331 } 332 /* double check to make sure if flushed correctly */ 333 if (jb->journal_list) 334 return NULL; 335 jb->journal_list = jl; 336 return jb; 337 } 338 339 /* 340 * allocates a new chunk of X nodes, and links them all together as a list. 341 * Uses the cnode->next and cnode->prev pointers 342 * returns NULL on failure 343 */ 344 static struct reiserfs_journal_cnode *allocate_cnodes(int num_cnodes) 345 { 346 struct reiserfs_journal_cnode *head; 347 int i; 348 if (num_cnodes <= 0) { 349 return NULL; 350 } 351 head = vzalloc(array_size(num_cnodes, 352 sizeof(struct reiserfs_journal_cnode))); 353 if (!head) { 354 return NULL; 355 } 356 head[0].prev = NULL; 357 head[0].next = head + 1; 358 for (i = 1; i < num_cnodes; i++) { 359 head[i].prev = head + (i - 1); 360 head[i].next = head + (i + 1); /* if last one, overwrite it after the if */ 361 } 362 head[num_cnodes - 1].next = NULL; 363 return head; 364 } 365 366 /* pulls a cnode off the free list, or returns NULL on failure */ 367 static struct reiserfs_journal_cnode *get_cnode(struct super_block *sb) 368 { 369 struct reiserfs_journal_cnode *cn; 370 struct reiserfs_journal *journal = SB_JOURNAL(sb); 371 372 reiserfs_check_lock_depth(sb, "get_cnode"); 373 374 if (journal->j_cnode_free <= 0) { 375 return NULL; 376 } 377 journal->j_cnode_used++; 378 journal->j_cnode_free--; 379 cn = journal->j_cnode_free_list; 380 if (!cn) { 381 return cn; 382 } 383 if (cn->next) { 384 cn->next->prev = NULL; 385 } 386 journal->j_cnode_free_list = cn->next; 387 memset(cn, 0, sizeof(struct reiserfs_journal_cnode)); 388 return cn; 389 } 390 391 /* 392 * returns a cnode to the free list 393 */ 394 static void free_cnode(struct super_block *sb, 395 struct reiserfs_journal_cnode *cn) 396 { 397 struct reiserfs_journal *journal = SB_JOURNAL(sb); 398 399 reiserfs_check_lock_depth(sb, "free_cnode"); 400 401 journal->j_cnode_used--; 402 journal->j_cnode_free++; 403 /* memset(cn, 0, sizeof(struct reiserfs_journal_cnode)) ; */ 404 cn->next = journal->j_cnode_free_list; 405 if (journal->j_cnode_free_list) { 406 journal->j_cnode_free_list->prev = cn; 407 } 408 cn->prev = NULL; /* not needed with the memset, but I might kill the memset, and forget to do this */ 409 journal->j_cnode_free_list = cn; 410 } 411 412 static void clear_prepared_bits(struct buffer_head *bh) 413 { 414 clear_buffer_journal_prepared(bh); 415 clear_buffer_journal_restore_dirty(bh); 416 } 417 418 /* 419 * return a cnode with same dev, block number and size in table, 420 * or null if not found 421 */ 422 static inline struct reiserfs_journal_cnode *get_journal_hash_dev(struct 423 super_block 424 *sb, 425 struct 426 reiserfs_journal_cnode 427 **table, 428 long bl) 429 { 430 struct reiserfs_journal_cnode *cn; 431 cn = journal_hash(table, sb, bl); 432 while (cn) { 433 if (cn->blocknr == bl && cn->sb == sb) 434 return cn; 435 cn = cn->hnext; 436 } 437 return (struct reiserfs_journal_cnode *)0; 438 } 439 440 /* 441 * this actually means 'can this block be reallocated yet?'. If you set 442 * search_all, a block can only be allocated if it is not in the current 443 * transaction, was not freed by the current transaction, and has no chance 444 * of ever being overwritten by a replay after crashing. 445 * 446 * If you don't set search_all, a block can only be allocated if it is not 447 * in the current transaction. Since deleting a block removes it from the 448 * current transaction, this case should never happen. If you don't set 449 * search_all, make sure you never write the block without logging it. 450 * 451 * next_zero_bit is a suggestion about the next block to try for find_forward. 452 * when bl is rejected because it is set in a journal list bitmap, we search 453 * for the next zero bit in the bitmap that rejected bl. Then, we return 454 * that through next_zero_bit for find_forward to try. 455 * 456 * Just because we return something in next_zero_bit does not mean we won't 457 * reject it on the next call to reiserfs_in_journal 458 */ 459 int reiserfs_in_journal(struct super_block *sb, 460 unsigned int bmap_nr, int bit_nr, int search_all, 461 b_blocknr_t * next_zero_bit) 462 { 463 struct reiserfs_journal *journal = SB_JOURNAL(sb); 464 struct reiserfs_list_bitmap *jb; 465 int i; 466 unsigned long bl; 467 468 *next_zero_bit = 0; /* always start this at zero. */ 469 470 PROC_INFO_INC(sb, journal.in_journal); 471 /* 472 * If we aren't doing a search_all, this is a metablock, and it 473 * will be logged before use. if we crash before the transaction 474 * that freed it commits, this transaction won't have committed 475 * either, and the block will never be written 476 */ 477 if (search_all) { 478 for (i = 0; i < JOURNAL_NUM_BITMAPS; i++) { 479 PROC_INFO_INC(sb, journal.in_journal_bitmap); 480 jb = journal->j_list_bitmap + i; 481 if (jb->journal_list && jb->bitmaps[bmap_nr] && 482 test_bit(bit_nr, 483 (unsigned long *)jb->bitmaps[bmap_nr]-> 484 data)) { 485 *next_zero_bit = 486 find_next_zero_bit((unsigned long *) 487 (jb->bitmaps[bmap_nr]-> 488 data), 489 sb->s_blocksize << 3, 490 bit_nr + 1); 491 return 1; 492 } 493 } 494 } 495 496 bl = bmap_nr * (sb->s_blocksize << 3) + bit_nr; 497 /* is it in any old transactions? */ 498 if (search_all 499 && (get_journal_hash_dev(sb, journal->j_list_hash_table, bl))) { 500 return 1; 501 } 502 503 /* is it in the current transaction. This should never happen */ 504 if ((get_journal_hash_dev(sb, journal->j_hash_table, bl))) { 505 BUG(); 506 return 1; 507 } 508 509 PROC_INFO_INC(sb, journal.in_journal_reusable); 510 /* safe for reuse */ 511 return 0; 512 } 513 514 /* insert cn into table */ 515 static inline void insert_journal_hash(struct reiserfs_journal_cnode **table, 516 struct reiserfs_journal_cnode *cn) 517 { 518 struct reiserfs_journal_cnode *cn_orig; 519 520 cn_orig = journal_hash(table, cn->sb, cn->blocknr); 521 cn->hnext = cn_orig; 522 cn->hprev = NULL; 523 if (cn_orig) { 524 cn_orig->hprev = cn; 525 } 526 journal_hash(table, cn->sb, cn->blocknr) = cn; 527 } 528 529 /* lock the current transaction */ 530 static inline void lock_journal(struct super_block *sb) 531 { 532 PROC_INFO_INC(sb, journal.lock_journal); 533 534 reiserfs_mutex_lock_safe(&SB_JOURNAL(sb)->j_mutex, sb); 535 } 536 537 /* unlock the current transaction */ 538 static inline void unlock_journal(struct super_block *sb) 539 { 540 mutex_unlock(&SB_JOURNAL(sb)->j_mutex); 541 } 542 543 static inline void get_journal_list(struct reiserfs_journal_list *jl) 544 { 545 jl->j_refcount++; 546 } 547 548 static inline void put_journal_list(struct super_block *s, 549 struct reiserfs_journal_list *jl) 550 { 551 if (jl->j_refcount < 1) { 552 reiserfs_panic(s, "journal-2", "trans id %u, refcount at %d", 553 jl->j_trans_id, jl->j_refcount); 554 } 555 if (--jl->j_refcount == 0) 556 kfree(jl); 557 } 558 559 /* 560 * this used to be much more involved, and I'm keeping it just in case 561 * things get ugly again. it gets called by flush_commit_list, and 562 * cleans up any data stored about blocks freed during a transaction. 563 */ 564 static void cleanup_freed_for_journal_list(struct super_block *sb, 565 struct reiserfs_journal_list *jl) 566 { 567 568 struct reiserfs_list_bitmap *jb = jl->j_list_bitmap; 569 if (jb) { 570 cleanup_bitmap_list(sb, jb); 571 } 572 jl->j_list_bitmap->journal_list = NULL; 573 jl->j_list_bitmap = NULL; 574 } 575 576 static int journal_list_still_alive(struct super_block *s, 577 unsigned int trans_id) 578 { 579 struct reiserfs_journal *journal = SB_JOURNAL(s); 580 struct list_head *entry = &journal->j_journal_list; 581 struct reiserfs_journal_list *jl; 582 583 if (!list_empty(entry)) { 584 jl = JOURNAL_LIST_ENTRY(entry->next); 585 if (jl->j_trans_id <= trans_id) { 586 return 1; 587 } 588 } 589 return 0; 590 } 591 592 /* 593 * If page->mapping was null, we failed to truncate this page for 594 * some reason. Most likely because it was truncated after being 595 * logged via data=journal. 596 * 597 * This does a check to see if the buffer belongs to one of these 598 * lost pages before doing the final put_bh. If page->mapping was 599 * null, it tries to free buffers on the page, which should make the 600 * final put_page drop the page from the lru. 601 */ 602 static void release_buffer_page(struct buffer_head *bh) 603 { 604 struct folio *folio = bh->b_folio; 605 if (!folio->mapping && folio_trylock(folio)) { 606 folio_get(folio); 607 put_bh(bh); 608 if (!folio->mapping) 609 try_to_free_buffers(folio); 610 folio_unlock(folio); 611 folio_put(folio); 612 } else { 613 put_bh(bh); 614 } 615 } 616 617 static void reiserfs_end_buffer_io_sync(struct buffer_head *bh, int uptodate) 618 { 619 if (buffer_journaled(bh)) { 620 reiserfs_warning(NULL, "clm-2084", 621 "pinned buffer %lu:%pg sent to disk", 622 bh->b_blocknr, bh->b_bdev); 623 } 624 if (uptodate) 625 set_buffer_uptodate(bh); 626 else 627 clear_buffer_uptodate(bh); 628 629 unlock_buffer(bh); 630 release_buffer_page(bh); 631 } 632 633 static void reiserfs_end_ordered_io(struct buffer_head *bh, int uptodate) 634 { 635 if (uptodate) 636 set_buffer_uptodate(bh); 637 else 638 clear_buffer_uptodate(bh); 639 unlock_buffer(bh); 640 put_bh(bh); 641 } 642 643 static void submit_logged_buffer(struct buffer_head *bh) 644 { 645 get_bh(bh); 646 bh->b_end_io = reiserfs_end_buffer_io_sync; 647 clear_buffer_journal_new(bh); 648 clear_buffer_dirty(bh); 649 if (!test_clear_buffer_journal_test(bh)) 650 BUG(); 651 if (!buffer_uptodate(bh)) 652 BUG(); 653 submit_bh(REQ_OP_WRITE, bh); 654 } 655 656 static void submit_ordered_buffer(struct buffer_head *bh) 657 { 658 get_bh(bh); 659 bh->b_end_io = reiserfs_end_ordered_io; 660 clear_buffer_dirty(bh); 661 if (!buffer_uptodate(bh)) 662 BUG(); 663 submit_bh(REQ_OP_WRITE, bh); 664 } 665 666 #define CHUNK_SIZE 32 667 struct buffer_chunk { 668 struct buffer_head *bh[CHUNK_SIZE]; 669 int nr; 670 }; 671 672 static void write_chunk(struct buffer_chunk *chunk) 673 { 674 int i; 675 for (i = 0; i < chunk->nr; i++) { 676 submit_logged_buffer(chunk->bh[i]); 677 } 678 chunk->nr = 0; 679 } 680 681 static void write_ordered_chunk(struct buffer_chunk *chunk) 682 { 683 int i; 684 for (i = 0; i < chunk->nr; i++) { 685 submit_ordered_buffer(chunk->bh[i]); 686 } 687 chunk->nr = 0; 688 } 689 690 static int add_to_chunk(struct buffer_chunk *chunk, struct buffer_head *bh, 691 spinlock_t * lock, void (fn) (struct buffer_chunk *)) 692 { 693 int ret = 0; 694 BUG_ON(chunk->nr >= CHUNK_SIZE); 695 chunk->bh[chunk->nr++] = bh; 696 if (chunk->nr >= CHUNK_SIZE) { 697 ret = 1; 698 if (lock) { 699 spin_unlock(lock); 700 fn(chunk); 701 spin_lock(lock); 702 } else { 703 fn(chunk); 704 } 705 } 706 return ret; 707 } 708 709 static atomic_t nr_reiserfs_jh = ATOMIC_INIT(0); 710 static struct reiserfs_jh *alloc_jh(void) 711 { 712 struct reiserfs_jh *jh; 713 while (1) { 714 jh = kmalloc(sizeof(*jh), GFP_NOFS); 715 if (jh) { 716 atomic_inc(&nr_reiserfs_jh); 717 return jh; 718 } 719 yield(); 720 } 721 } 722 723 /* 724 * we want to free the jh when the buffer has been written 725 * and waited on 726 */ 727 void reiserfs_free_jh(struct buffer_head *bh) 728 { 729 struct reiserfs_jh *jh; 730 731 jh = bh->b_private; 732 if (jh) { 733 bh->b_private = NULL; 734 jh->bh = NULL; 735 list_del_init(&jh->list); 736 kfree(jh); 737 if (atomic_read(&nr_reiserfs_jh) <= 0) 738 BUG(); 739 atomic_dec(&nr_reiserfs_jh); 740 put_bh(bh); 741 } 742 } 743 744 static inline int __add_jh(struct reiserfs_journal *j, struct buffer_head *bh, 745 int tail) 746 { 747 struct reiserfs_jh *jh; 748 749 if (bh->b_private) { 750 spin_lock(&j->j_dirty_buffers_lock); 751 if (!bh->b_private) { 752 spin_unlock(&j->j_dirty_buffers_lock); 753 goto no_jh; 754 } 755 jh = bh->b_private; 756 list_del_init(&jh->list); 757 } else { 758 no_jh: 759 get_bh(bh); 760 jh = alloc_jh(); 761 spin_lock(&j->j_dirty_buffers_lock); 762 /* 763 * buffer must be locked for __add_jh, should be able to have 764 * two adds at the same time 765 */ 766 BUG_ON(bh->b_private); 767 jh->bh = bh; 768 bh->b_private = jh; 769 } 770 jh->jl = j->j_current_jl; 771 if (tail) 772 list_add_tail(&jh->list, &jh->jl->j_tail_bh_list); 773 else { 774 list_add_tail(&jh->list, &jh->jl->j_bh_list); 775 } 776 spin_unlock(&j->j_dirty_buffers_lock); 777 return 0; 778 } 779 780 int reiserfs_add_tail_list(struct inode *inode, struct buffer_head *bh) 781 { 782 return __add_jh(SB_JOURNAL(inode->i_sb), bh, 1); 783 } 784 int reiserfs_add_ordered_list(struct inode *inode, struct buffer_head *bh) 785 { 786 return __add_jh(SB_JOURNAL(inode->i_sb), bh, 0); 787 } 788 789 #define JH_ENTRY(l) list_entry((l), struct reiserfs_jh, list) 790 static int write_ordered_buffers(spinlock_t * lock, 791 struct reiserfs_journal *j, 792 struct reiserfs_journal_list *jl, 793 struct list_head *list) 794 { 795 struct buffer_head *bh; 796 struct reiserfs_jh *jh; 797 int ret = j->j_errno; 798 struct buffer_chunk chunk; 799 struct list_head tmp; 800 INIT_LIST_HEAD(&tmp); 801 802 chunk.nr = 0; 803 spin_lock(lock); 804 while (!list_empty(list)) { 805 jh = JH_ENTRY(list->next); 806 bh = jh->bh; 807 get_bh(bh); 808 if (!trylock_buffer(bh)) { 809 if (!buffer_dirty(bh)) { 810 list_move(&jh->list, &tmp); 811 goto loop_next; 812 } 813 spin_unlock(lock); 814 if (chunk.nr) 815 write_ordered_chunk(&chunk); 816 wait_on_buffer(bh); 817 cond_resched(); 818 spin_lock(lock); 819 goto loop_next; 820 } 821 /* 822 * in theory, dirty non-uptodate buffers should never get here, 823 * but the upper layer io error paths still have a few quirks. 824 * Handle them here as gracefully as we can 825 */ 826 if (!buffer_uptodate(bh) && buffer_dirty(bh)) { 827 clear_buffer_dirty(bh); 828 ret = -EIO; 829 } 830 if (buffer_dirty(bh)) { 831 list_move(&jh->list, &tmp); 832 add_to_chunk(&chunk, bh, lock, write_ordered_chunk); 833 } else { 834 reiserfs_free_jh(bh); 835 unlock_buffer(bh); 836 } 837 loop_next: 838 put_bh(bh); 839 cond_resched_lock(lock); 840 } 841 if (chunk.nr) { 842 spin_unlock(lock); 843 write_ordered_chunk(&chunk); 844 spin_lock(lock); 845 } 846 while (!list_empty(&tmp)) { 847 jh = JH_ENTRY(tmp.prev); 848 bh = jh->bh; 849 get_bh(bh); 850 reiserfs_free_jh(bh); 851 852 if (buffer_locked(bh)) { 853 spin_unlock(lock); 854 wait_on_buffer(bh); 855 spin_lock(lock); 856 } 857 if (!buffer_uptodate(bh)) { 858 ret = -EIO; 859 } 860 /* 861 * ugly interaction with invalidate_folio here. 862 * reiserfs_invalidate_folio will pin any buffer that has a 863 * valid journal head from an older transaction. If someone 864 * else sets our buffer dirty after we write it in the first 865 * loop, and then someone truncates the page away, nobody 866 * will ever write the buffer. We're safe if we write the 867 * page one last time after freeing the journal header. 868 */ 869 if (buffer_dirty(bh) && unlikely(bh->b_folio->mapping == NULL)) { 870 spin_unlock(lock); 871 write_dirty_buffer(bh, 0); 872 spin_lock(lock); 873 } 874 put_bh(bh); 875 cond_resched_lock(lock); 876 } 877 spin_unlock(lock); 878 return ret; 879 } 880 881 static int flush_older_commits(struct super_block *s, 882 struct reiserfs_journal_list *jl) 883 { 884 struct reiserfs_journal *journal = SB_JOURNAL(s); 885 struct reiserfs_journal_list *other_jl; 886 struct reiserfs_journal_list *first_jl; 887 struct list_head *entry; 888 unsigned int trans_id = jl->j_trans_id; 889 unsigned int other_trans_id; 890 891 find_first: 892 /* 893 * first we walk backwards to find the oldest uncommitted transation 894 */ 895 first_jl = jl; 896 entry = jl->j_list.prev; 897 while (1) { 898 other_jl = JOURNAL_LIST_ENTRY(entry); 899 if (entry == &journal->j_journal_list || 900 atomic_read(&other_jl->j_older_commits_done)) 901 break; 902 903 first_jl = other_jl; 904 entry = other_jl->j_list.prev; 905 } 906 907 /* if we didn't find any older uncommitted transactions, return now */ 908 if (first_jl == jl) { 909 return 0; 910 } 911 912 entry = &first_jl->j_list; 913 while (1) { 914 other_jl = JOURNAL_LIST_ENTRY(entry); 915 other_trans_id = other_jl->j_trans_id; 916 917 if (other_trans_id < trans_id) { 918 if (atomic_read(&other_jl->j_commit_left) != 0) { 919 flush_commit_list(s, other_jl, 0); 920 921 /* list we were called with is gone, return */ 922 if (!journal_list_still_alive(s, trans_id)) 923 return 1; 924 925 /* 926 * the one we just flushed is gone, this means 927 * all older lists are also gone, so first_jl 928 * is no longer valid either. Go back to the 929 * beginning. 930 */ 931 if (!journal_list_still_alive 932 (s, other_trans_id)) { 933 goto find_first; 934 } 935 } 936 entry = entry->next; 937 if (entry == &journal->j_journal_list) 938 return 0; 939 } else { 940 return 0; 941 } 942 } 943 return 0; 944 } 945 946 static int reiserfs_async_progress_wait(struct super_block *s) 947 { 948 struct reiserfs_journal *j = SB_JOURNAL(s); 949 950 if (atomic_read(&j->j_async_throttle)) { 951 int depth; 952 953 depth = reiserfs_write_unlock_nested(s); 954 wait_var_event_timeout(&j->j_async_throttle, 955 atomic_read(&j->j_async_throttle) == 0, 956 HZ / 10); 957 reiserfs_write_lock_nested(s, depth); 958 } 959 960 return 0; 961 } 962 963 /* 964 * if this journal list still has commit blocks unflushed, send them to disk. 965 * 966 * log areas must be flushed in order (transaction 2 can't commit before 967 * transaction 1) Before the commit block can by written, every other log 968 * block must be safely on disk 969 */ 970 static int flush_commit_list(struct super_block *s, 971 struct reiserfs_journal_list *jl, int flushall) 972 { 973 int i; 974 b_blocknr_t bn; 975 struct buffer_head *tbh = NULL; 976 unsigned int trans_id = jl->j_trans_id; 977 struct reiserfs_journal *journal = SB_JOURNAL(s); 978 int retval = 0; 979 int write_len; 980 int depth; 981 982 reiserfs_check_lock_depth(s, "flush_commit_list"); 983 984 if (atomic_read(&jl->j_older_commits_done)) { 985 return 0; 986 } 987 988 /* 989 * before we can put our commit blocks on disk, we have to make 990 * sure everyone older than us is on disk too 991 */ 992 BUG_ON(jl->j_len <= 0); 993 BUG_ON(trans_id == journal->j_trans_id); 994 995 get_journal_list(jl); 996 if (flushall) { 997 if (flush_older_commits(s, jl) == 1) { 998 /* 999 * list disappeared during flush_older_commits. 1000 * return 1001 */ 1002 goto put_jl; 1003 } 1004 } 1005 1006 /* make sure nobody is trying to flush this one at the same time */ 1007 reiserfs_mutex_lock_safe(&jl->j_commit_mutex, s); 1008 1009 if (!journal_list_still_alive(s, trans_id)) { 1010 mutex_unlock(&jl->j_commit_mutex); 1011 goto put_jl; 1012 } 1013 BUG_ON(jl->j_trans_id == 0); 1014 1015 /* this commit is done, exit */ 1016 if (atomic_read(&jl->j_commit_left) <= 0) { 1017 if (flushall) { 1018 atomic_set(&jl->j_older_commits_done, 1); 1019 } 1020 mutex_unlock(&jl->j_commit_mutex); 1021 goto put_jl; 1022 } 1023 1024 if (!list_empty(&jl->j_bh_list)) { 1025 int ret; 1026 1027 /* 1028 * We might sleep in numerous places inside 1029 * write_ordered_buffers. Relax the write lock. 1030 */ 1031 depth = reiserfs_write_unlock_nested(s); 1032 ret = write_ordered_buffers(&journal->j_dirty_buffers_lock, 1033 journal, jl, &jl->j_bh_list); 1034 if (ret < 0 && retval == 0) 1035 retval = ret; 1036 reiserfs_write_lock_nested(s, depth); 1037 } 1038 BUG_ON(!list_empty(&jl->j_bh_list)); 1039 /* 1040 * for the description block and all the log blocks, submit any buffers 1041 * that haven't already reached the disk. Try to write at least 256 1042 * log blocks. later on, we will only wait on blocks that correspond 1043 * to this transaction, but while we're unplugging we might as well 1044 * get a chunk of data on there. 1045 */ 1046 atomic_inc(&journal->j_async_throttle); 1047 write_len = jl->j_len + 1; 1048 if (write_len < 256) 1049 write_len = 256; 1050 for (i = 0 ; i < write_len ; i++) { 1051 bn = SB_ONDISK_JOURNAL_1st_BLOCK(s) + (jl->j_start + i) % 1052 SB_ONDISK_JOURNAL_SIZE(s); 1053 tbh = journal_find_get_block(s, bn); 1054 if (tbh) { 1055 if (buffer_dirty(tbh)) { 1056 depth = reiserfs_write_unlock_nested(s); 1057 write_dirty_buffer(tbh, 0); 1058 reiserfs_write_lock_nested(s, depth); 1059 } 1060 put_bh(tbh) ; 1061 } 1062 } 1063 if (atomic_dec_and_test(&journal->j_async_throttle)) 1064 wake_up_var(&journal->j_async_throttle); 1065 1066 for (i = 0; i < (jl->j_len + 1); i++) { 1067 bn = SB_ONDISK_JOURNAL_1st_BLOCK(s) + 1068 (jl->j_start + i) % SB_ONDISK_JOURNAL_SIZE(s); 1069 tbh = journal_find_get_block(s, bn); 1070 1071 depth = reiserfs_write_unlock_nested(s); 1072 __wait_on_buffer(tbh); 1073 reiserfs_write_lock_nested(s, depth); 1074 /* 1075 * since we're using ll_rw_blk above, it might have skipped 1076 * over a locked buffer. Double check here 1077 */ 1078 /* redundant, sync_dirty_buffer() checks */ 1079 if (buffer_dirty(tbh)) { 1080 depth = reiserfs_write_unlock_nested(s); 1081 sync_dirty_buffer(tbh); 1082 reiserfs_write_lock_nested(s, depth); 1083 } 1084 if (unlikely(!buffer_uptodate(tbh))) { 1085 #ifdef CONFIG_REISERFS_CHECK 1086 reiserfs_warning(s, "journal-601", 1087 "buffer write failed"); 1088 #endif 1089 retval = -EIO; 1090 } 1091 /* once for journal_find_get_block */ 1092 put_bh(tbh); 1093 /* once due to original getblk in do_journal_end */ 1094 put_bh(tbh); 1095 atomic_dec(&jl->j_commit_left); 1096 } 1097 1098 BUG_ON(atomic_read(&jl->j_commit_left) != 1); 1099 1100 /* 1101 * If there was a write error in the journal - we can't commit 1102 * this transaction - it will be invalid and, if successful, 1103 * will just end up propagating the write error out to 1104 * the file system. 1105 */ 1106 if (likely(!retval && !reiserfs_is_journal_aborted (journal))) { 1107 if (buffer_dirty(jl->j_commit_bh)) 1108 BUG(); 1109 mark_buffer_dirty(jl->j_commit_bh) ; 1110 depth = reiserfs_write_unlock_nested(s); 1111 if (reiserfs_barrier_flush(s)) 1112 __sync_dirty_buffer(jl->j_commit_bh, 1113 REQ_SYNC | REQ_PREFLUSH | REQ_FUA); 1114 else 1115 sync_dirty_buffer(jl->j_commit_bh); 1116 reiserfs_write_lock_nested(s, depth); 1117 } 1118 1119 /* 1120 * If there was a write error in the journal - we can't commit this 1121 * transaction - it will be invalid and, if successful, will just end 1122 * up propagating the write error out to the filesystem. 1123 */ 1124 if (unlikely(!buffer_uptodate(jl->j_commit_bh))) { 1125 #ifdef CONFIG_REISERFS_CHECK 1126 reiserfs_warning(s, "journal-615", "buffer write failed"); 1127 #endif 1128 retval = -EIO; 1129 } 1130 bforget(jl->j_commit_bh); 1131 if (journal->j_last_commit_id != 0 && 1132 (jl->j_trans_id - journal->j_last_commit_id) != 1) { 1133 reiserfs_warning(s, "clm-2200", "last commit %lu, current %lu", 1134 journal->j_last_commit_id, jl->j_trans_id); 1135 } 1136 journal->j_last_commit_id = jl->j_trans_id; 1137 1138 /* 1139 * now, every commit block is on the disk. It is safe to allow 1140 * blocks freed during this transaction to be reallocated 1141 */ 1142 cleanup_freed_for_journal_list(s, jl); 1143 1144 retval = retval ? retval : journal->j_errno; 1145 1146 /* mark the metadata dirty */ 1147 if (!retval) 1148 dirty_one_transaction(s, jl); 1149 atomic_dec(&jl->j_commit_left); 1150 1151 if (flushall) { 1152 atomic_set(&jl->j_older_commits_done, 1); 1153 } 1154 mutex_unlock(&jl->j_commit_mutex); 1155 put_jl: 1156 put_journal_list(s, jl); 1157 1158 if (retval) 1159 reiserfs_abort(s, retval, "Journal write error in %s", 1160 __func__); 1161 return retval; 1162 } 1163 1164 /* 1165 * flush_journal_list frequently needs to find a newer transaction for a 1166 * given block. This does that, or returns NULL if it can't find anything 1167 */ 1168 static struct reiserfs_journal_list *find_newer_jl_for_cn(struct 1169 reiserfs_journal_cnode 1170 *cn) 1171 { 1172 struct super_block *sb = cn->sb; 1173 b_blocknr_t blocknr = cn->blocknr; 1174 1175 cn = cn->hprev; 1176 while (cn) { 1177 if (cn->sb == sb && cn->blocknr == blocknr && cn->jlist) { 1178 return cn->jlist; 1179 } 1180 cn = cn->hprev; 1181 } 1182 return NULL; 1183 } 1184 1185 static void remove_journal_hash(struct super_block *, 1186 struct reiserfs_journal_cnode **, 1187 struct reiserfs_journal_list *, unsigned long, 1188 int); 1189 1190 /* 1191 * once all the real blocks have been flushed, it is safe to remove them 1192 * from the journal list for this transaction. Aside from freeing the 1193 * cnode, this also allows the block to be reallocated for data blocks 1194 * if it had been deleted. 1195 */ 1196 static void remove_all_from_journal_list(struct super_block *sb, 1197 struct reiserfs_journal_list *jl, 1198 int debug) 1199 { 1200 struct reiserfs_journal *journal = SB_JOURNAL(sb); 1201 struct reiserfs_journal_cnode *cn, *last; 1202 cn = jl->j_realblock; 1203 1204 /* 1205 * which is better, to lock once around the whole loop, or 1206 * to lock for each call to remove_journal_hash? 1207 */ 1208 while (cn) { 1209 if (cn->blocknr != 0) { 1210 if (debug) { 1211 reiserfs_warning(sb, "reiserfs-2201", 1212 "block %u, bh is %d, state %ld", 1213 cn->blocknr, cn->bh ? 1 : 0, 1214 cn->state); 1215 } 1216 cn->state = 0; 1217 remove_journal_hash(sb, journal->j_list_hash_table, 1218 jl, cn->blocknr, 1); 1219 } 1220 last = cn; 1221 cn = cn->next; 1222 free_cnode(sb, last); 1223 } 1224 jl->j_realblock = NULL; 1225 } 1226 1227 /* 1228 * if this timestamp is greater than the timestamp we wrote last to the 1229 * header block, write it to the header block. once this is done, I can 1230 * safely say the log area for this transaction won't ever be replayed, 1231 * and I can start releasing blocks in this transaction for reuse as data 1232 * blocks. called by flush_journal_list, before it calls 1233 * remove_all_from_journal_list 1234 */ 1235 static int _update_journal_header_block(struct super_block *sb, 1236 unsigned long offset, 1237 unsigned int trans_id) 1238 { 1239 struct reiserfs_journal_header *jh; 1240 struct reiserfs_journal *journal = SB_JOURNAL(sb); 1241 int depth; 1242 1243 if (reiserfs_is_journal_aborted(journal)) 1244 return -EIO; 1245 1246 if (trans_id >= journal->j_last_flush_trans_id) { 1247 if (buffer_locked((journal->j_header_bh))) { 1248 depth = reiserfs_write_unlock_nested(sb); 1249 __wait_on_buffer(journal->j_header_bh); 1250 reiserfs_write_lock_nested(sb, depth); 1251 if (unlikely(!buffer_uptodate(journal->j_header_bh))) { 1252 #ifdef CONFIG_REISERFS_CHECK 1253 reiserfs_warning(sb, "journal-699", 1254 "buffer write failed"); 1255 #endif 1256 return -EIO; 1257 } 1258 } 1259 journal->j_last_flush_trans_id = trans_id; 1260 journal->j_first_unflushed_offset = offset; 1261 jh = (struct reiserfs_journal_header *)(journal->j_header_bh-> 1262 b_data); 1263 jh->j_last_flush_trans_id = cpu_to_le32(trans_id); 1264 jh->j_first_unflushed_offset = cpu_to_le32(offset); 1265 jh->j_mount_id = cpu_to_le32(journal->j_mount_id); 1266 1267 set_buffer_dirty(journal->j_header_bh); 1268 depth = reiserfs_write_unlock_nested(sb); 1269 1270 if (reiserfs_barrier_flush(sb)) 1271 __sync_dirty_buffer(journal->j_header_bh, 1272 REQ_SYNC | REQ_PREFLUSH | REQ_FUA); 1273 else 1274 sync_dirty_buffer(journal->j_header_bh); 1275 1276 reiserfs_write_lock_nested(sb, depth); 1277 if (!buffer_uptodate(journal->j_header_bh)) { 1278 reiserfs_warning(sb, "journal-837", 1279 "IO error during journal replay"); 1280 return -EIO; 1281 } 1282 } 1283 return 0; 1284 } 1285 1286 static int update_journal_header_block(struct super_block *sb, 1287 unsigned long offset, 1288 unsigned int trans_id) 1289 { 1290 return _update_journal_header_block(sb, offset, trans_id); 1291 } 1292 1293 /* 1294 ** flush any and all journal lists older than you are 1295 ** can only be called from flush_journal_list 1296 */ 1297 static int flush_older_journal_lists(struct super_block *sb, 1298 struct reiserfs_journal_list *jl) 1299 { 1300 struct list_head *entry; 1301 struct reiserfs_journal_list *other_jl; 1302 struct reiserfs_journal *journal = SB_JOURNAL(sb); 1303 unsigned int trans_id = jl->j_trans_id; 1304 1305 /* 1306 * we know we are the only ones flushing things, no extra race 1307 * protection is required. 1308 */ 1309 restart: 1310 entry = journal->j_journal_list.next; 1311 /* Did we wrap? */ 1312 if (entry == &journal->j_journal_list) 1313 return 0; 1314 other_jl = JOURNAL_LIST_ENTRY(entry); 1315 if (other_jl->j_trans_id < trans_id) { 1316 BUG_ON(other_jl->j_refcount <= 0); 1317 /* do not flush all */ 1318 flush_journal_list(sb, other_jl, 0); 1319 1320 /* other_jl is now deleted from the list */ 1321 goto restart; 1322 } 1323 return 0; 1324 } 1325 1326 static void del_from_work_list(struct super_block *s, 1327 struct reiserfs_journal_list *jl) 1328 { 1329 struct reiserfs_journal *journal = SB_JOURNAL(s); 1330 if (!list_empty(&jl->j_working_list)) { 1331 list_del_init(&jl->j_working_list); 1332 journal->j_num_work_lists--; 1333 } 1334 } 1335 1336 /* 1337 * flush a journal list, both commit and real blocks 1338 * 1339 * always set flushall to 1, unless you are calling from inside 1340 * flush_journal_list 1341 * 1342 * IMPORTANT. This can only be called while there are no journal writers, 1343 * and the journal is locked. That means it can only be called from 1344 * do_journal_end, or by journal_release 1345 */ 1346 static int flush_journal_list(struct super_block *s, 1347 struct reiserfs_journal_list *jl, int flushall) 1348 { 1349 struct reiserfs_journal_list *pjl; 1350 struct reiserfs_journal_cnode *cn; 1351 int count; 1352 int was_jwait = 0; 1353 int was_dirty = 0; 1354 struct buffer_head *saved_bh; 1355 unsigned long j_len_saved = jl->j_len; 1356 struct reiserfs_journal *journal = SB_JOURNAL(s); 1357 int err = 0; 1358 int depth; 1359 1360 BUG_ON(j_len_saved <= 0); 1361 1362 if (atomic_read(&journal->j_wcount) != 0) { 1363 reiserfs_warning(s, "clm-2048", "called with wcount %d", 1364 atomic_read(&journal->j_wcount)); 1365 } 1366 1367 /* if flushall == 0, the lock is already held */ 1368 if (flushall) { 1369 reiserfs_mutex_lock_safe(&journal->j_flush_mutex, s); 1370 } else if (mutex_trylock(&journal->j_flush_mutex)) { 1371 BUG(); 1372 } 1373 1374 count = 0; 1375 if (j_len_saved > journal->j_trans_max) { 1376 reiserfs_panic(s, "journal-715", "length is %lu, trans id %lu", 1377 j_len_saved, jl->j_trans_id); 1378 return 0; 1379 } 1380 1381 /* if all the work is already done, get out of here */ 1382 if (atomic_read(&jl->j_nonzerolen) <= 0 && 1383 atomic_read(&jl->j_commit_left) <= 0) { 1384 goto flush_older_and_return; 1385 } 1386 1387 /* 1388 * start by putting the commit list on disk. This will also flush 1389 * the commit lists of any olders transactions 1390 */ 1391 flush_commit_list(s, jl, 1); 1392 1393 if (!(jl->j_state & LIST_DIRTY) 1394 && !reiserfs_is_journal_aborted(journal)) 1395 BUG(); 1396 1397 /* are we done now? */ 1398 if (atomic_read(&jl->j_nonzerolen) <= 0 && 1399 atomic_read(&jl->j_commit_left) <= 0) { 1400 goto flush_older_and_return; 1401 } 1402 1403 /* 1404 * loop through each cnode, see if we need to write it, 1405 * or wait on a more recent transaction, or just ignore it 1406 */ 1407 if (atomic_read(&journal->j_wcount) != 0) { 1408 reiserfs_panic(s, "journal-844", "journal list is flushing, " 1409 "wcount is not 0"); 1410 } 1411 cn = jl->j_realblock; 1412 while (cn) { 1413 was_jwait = 0; 1414 was_dirty = 0; 1415 saved_bh = NULL; 1416 /* blocknr of 0 is no longer in the hash, ignore it */ 1417 if (cn->blocknr == 0) { 1418 goto free_cnode; 1419 } 1420 1421 /* 1422 * This transaction failed commit. 1423 * Don't write out to the disk 1424 */ 1425 if (!(jl->j_state & LIST_DIRTY)) 1426 goto free_cnode; 1427 1428 pjl = find_newer_jl_for_cn(cn); 1429 /* 1430 * the order is important here. We check pjl to make sure we 1431 * don't clear BH_JDirty_wait if we aren't the one writing this 1432 * block to disk 1433 */ 1434 if (!pjl && cn->bh) { 1435 saved_bh = cn->bh; 1436 1437 /* 1438 * we do this to make sure nobody releases the 1439 * buffer while we are working with it 1440 */ 1441 get_bh(saved_bh); 1442 1443 if (buffer_journal_dirty(saved_bh)) { 1444 BUG_ON(!can_dirty(cn)); 1445 was_jwait = 1; 1446 was_dirty = 1; 1447 } else if (can_dirty(cn)) { 1448 /* 1449 * everything with !pjl && jwait 1450 * should be writable 1451 */ 1452 BUG(); 1453 } 1454 } 1455 1456 /* 1457 * if someone has this block in a newer transaction, just make 1458 * sure they are committed, and don't try writing it to disk 1459 */ 1460 if (pjl) { 1461 if (atomic_read(&pjl->j_commit_left)) 1462 flush_commit_list(s, pjl, 1); 1463 goto free_cnode; 1464 } 1465 1466 /* 1467 * bh == NULL when the block got to disk on its own, OR, 1468 * the block got freed in a future transaction 1469 */ 1470 if (saved_bh == NULL) { 1471 goto free_cnode; 1472 } 1473 1474 /* 1475 * this should never happen. kupdate_one_transaction has 1476 * this list locked while it works, so we should never see a 1477 * buffer here that is not marked JDirty_wait 1478 */ 1479 if ((!was_jwait) && !buffer_locked(saved_bh)) { 1480 reiserfs_warning(s, "journal-813", 1481 "BAD! buffer %llu %cdirty %cjwait, " 1482 "not in a newer transaction", 1483 (unsigned long long)saved_bh-> 1484 b_blocknr, was_dirty ? ' ' : '!', 1485 was_jwait ? ' ' : '!'); 1486 } 1487 if (was_dirty) { 1488 /* 1489 * we inc again because saved_bh gets decremented 1490 * at free_cnode 1491 */ 1492 get_bh(saved_bh); 1493 set_bit(BLOCK_NEEDS_FLUSH, &cn->state); 1494 lock_buffer(saved_bh); 1495 BUG_ON(cn->blocknr != saved_bh->b_blocknr); 1496 if (buffer_dirty(saved_bh)) 1497 submit_logged_buffer(saved_bh); 1498 else 1499 unlock_buffer(saved_bh); 1500 count++; 1501 } else { 1502 reiserfs_warning(s, "clm-2082", 1503 "Unable to flush buffer %llu in %s", 1504 (unsigned long long)saved_bh-> 1505 b_blocknr, __func__); 1506 } 1507 free_cnode: 1508 cn = cn->next; 1509 if (saved_bh) { 1510 /* 1511 * we incremented this to keep others from 1512 * taking the buffer head away 1513 */ 1514 put_bh(saved_bh); 1515 if (atomic_read(&saved_bh->b_count) < 0) { 1516 reiserfs_warning(s, "journal-945", 1517 "saved_bh->b_count < 0"); 1518 } 1519 } 1520 } 1521 if (count > 0) { 1522 cn = jl->j_realblock; 1523 while (cn) { 1524 if (test_bit(BLOCK_NEEDS_FLUSH, &cn->state)) { 1525 if (!cn->bh) { 1526 reiserfs_panic(s, "journal-1011", 1527 "cn->bh is NULL"); 1528 } 1529 1530 depth = reiserfs_write_unlock_nested(s); 1531 __wait_on_buffer(cn->bh); 1532 reiserfs_write_lock_nested(s, depth); 1533 1534 if (!cn->bh) { 1535 reiserfs_panic(s, "journal-1012", 1536 "cn->bh is NULL"); 1537 } 1538 if (unlikely(!buffer_uptodate(cn->bh))) { 1539 #ifdef CONFIG_REISERFS_CHECK 1540 reiserfs_warning(s, "journal-949", 1541 "buffer write failed"); 1542 #endif 1543 err = -EIO; 1544 } 1545 /* 1546 * note, we must clear the JDirty_wait bit 1547 * after the up to date check, otherwise we 1548 * race against our flushpage routine 1549 */ 1550 BUG_ON(!test_clear_buffer_journal_dirty 1551 (cn->bh)); 1552 1553 /* drop one ref for us */ 1554 put_bh(cn->bh); 1555 /* drop one ref for journal_mark_dirty */ 1556 release_buffer_page(cn->bh); 1557 } 1558 cn = cn->next; 1559 } 1560 } 1561 1562 if (err) 1563 reiserfs_abort(s, -EIO, 1564 "Write error while pushing transaction to disk in %s", 1565 __func__); 1566 flush_older_and_return: 1567 1568 /* 1569 * before we can update the journal header block, we _must_ flush all 1570 * real blocks from all older transactions to disk. This is because 1571 * once the header block is updated, this transaction will not be 1572 * replayed after a crash 1573 */ 1574 if (flushall) { 1575 flush_older_journal_lists(s, jl); 1576 } 1577 1578 err = journal->j_errno; 1579 /* 1580 * before we can remove everything from the hash tables for this 1581 * transaction, we must make sure it can never be replayed 1582 * 1583 * since we are only called from do_journal_end, we know for sure there 1584 * are no allocations going on while we are flushing journal lists. So, 1585 * we only need to update the journal header block for the last list 1586 * being flushed 1587 */ 1588 if (!err && flushall) { 1589 err = 1590 update_journal_header_block(s, 1591 (jl->j_start + jl->j_len + 1592 2) % SB_ONDISK_JOURNAL_SIZE(s), 1593 jl->j_trans_id); 1594 if (err) 1595 reiserfs_abort(s, -EIO, 1596 "Write error while updating journal header in %s", 1597 __func__); 1598 } 1599 remove_all_from_journal_list(s, jl, 0); 1600 list_del_init(&jl->j_list); 1601 journal->j_num_lists--; 1602 del_from_work_list(s, jl); 1603 1604 if (journal->j_last_flush_id != 0 && 1605 (jl->j_trans_id - journal->j_last_flush_id) != 1) { 1606 reiserfs_warning(s, "clm-2201", "last flush %lu, current %lu", 1607 journal->j_last_flush_id, jl->j_trans_id); 1608 } 1609 journal->j_last_flush_id = jl->j_trans_id; 1610 1611 /* 1612 * not strictly required since we are freeing the list, but it should 1613 * help find code using dead lists later on 1614 */ 1615 jl->j_len = 0; 1616 atomic_set(&jl->j_nonzerolen, 0); 1617 jl->j_start = 0; 1618 jl->j_realblock = NULL; 1619 jl->j_commit_bh = NULL; 1620 jl->j_trans_id = 0; 1621 jl->j_state = 0; 1622 put_journal_list(s, jl); 1623 if (flushall) 1624 mutex_unlock(&journal->j_flush_mutex); 1625 return err; 1626 } 1627 1628 static int write_one_transaction(struct super_block *s, 1629 struct reiserfs_journal_list *jl, 1630 struct buffer_chunk *chunk) 1631 { 1632 struct reiserfs_journal_cnode *cn; 1633 int ret = 0; 1634 1635 jl->j_state |= LIST_TOUCHED; 1636 del_from_work_list(s, jl); 1637 if (jl->j_len == 0 || atomic_read(&jl->j_nonzerolen) == 0) { 1638 return 0; 1639 } 1640 1641 cn = jl->j_realblock; 1642 while (cn) { 1643 /* 1644 * if the blocknr == 0, this has been cleared from the hash, 1645 * skip it 1646 */ 1647 if (cn->blocknr == 0) { 1648 goto next; 1649 } 1650 if (cn->bh && can_dirty(cn) && buffer_dirty(cn->bh)) { 1651 struct buffer_head *tmp_bh; 1652 /* 1653 * we can race against journal_mark_freed when we try 1654 * to lock_buffer(cn->bh), so we have to inc the buffer 1655 * count, and recheck things after locking 1656 */ 1657 tmp_bh = cn->bh; 1658 get_bh(tmp_bh); 1659 lock_buffer(tmp_bh); 1660 if (cn->bh && can_dirty(cn) && buffer_dirty(tmp_bh)) { 1661 if (!buffer_journal_dirty(tmp_bh) || 1662 buffer_journal_prepared(tmp_bh)) 1663 BUG(); 1664 add_to_chunk(chunk, tmp_bh, NULL, write_chunk); 1665 ret++; 1666 } else { 1667 /* note, cn->bh might be null now */ 1668 unlock_buffer(tmp_bh); 1669 } 1670 put_bh(tmp_bh); 1671 } 1672 next: 1673 cn = cn->next; 1674 cond_resched(); 1675 } 1676 return ret; 1677 } 1678 1679 /* used by flush_commit_list */ 1680 static void dirty_one_transaction(struct super_block *s, 1681 struct reiserfs_journal_list *jl) 1682 { 1683 struct reiserfs_journal_cnode *cn; 1684 struct reiserfs_journal_list *pjl; 1685 1686 jl->j_state |= LIST_DIRTY; 1687 cn = jl->j_realblock; 1688 while (cn) { 1689 /* 1690 * look for a more recent transaction that logged this 1691 * buffer. Only the most recent transaction with a buffer in 1692 * it is allowed to send that buffer to disk 1693 */ 1694 pjl = find_newer_jl_for_cn(cn); 1695 if (!pjl && cn->blocknr && cn->bh 1696 && buffer_journal_dirty(cn->bh)) { 1697 BUG_ON(!can_dirty(cn)); 1698 /* 1699 * if the buffer is prepared, it will either be logged 1700 * or restored. If restored, we need to make sure 1701 * it actually gets marked dirty 1702 */ 1703 clear_buffer_journal_new(cn->bh); 1704 if (buffer_journal_prepared(cn->bh)) { 1705 set_buffer_journal_restore_dirty(cn->bh); 1706 } else { 1707 set_buffer_journal_test(cn->bh); 1708 mark_buffer_dirty(cn->bh); 1709 } 1710 } 1711 cn = cn->next; 1712 } 1713 } 1714 1715 static int kupdate_transactions(struct super_block *s, 1716 struct reiserfs_journal_list *jl, 1717 struct reiserfs_journal_list **next_jl, 1718 unsigned int *next_trans_id, 1719 int num_blocks, int num_trans) 1720 { 1721 int ret = 0; 1722 int written = 0; 1723 int transactions_flushed = 0; 1724 unsigned int orig_trans_id = jl->j_trans_id; 1725 struct buffer_chunk chunk; 1726 struct list_head *entry; 1727 struct reiserfs_journal *journal = SB_JOURNAL(s); 1728 chunk.nr = 0; 1729 1730 reiserfs_mutex_lock_safe(&journal->j_flush_mutex, s); 1731 if (!journal_list_still_alive(s, orig_trans_id)) { 1732 goto done; 1733 } 1734 1735 /* 1736 * we've got j_flush_mutex held, nobody is going to delete any 1737 * of these lists out from underneath us 1738 */ 1739 while ((num_trans && transactions_flushed < num_trans) || 1740 (!num_trans && written < num_blocks)) { 1741 1742 if (jl->j_len == 0 || (jl->j_state & LIST_TOUCHED) || 1743 atomic_read(&jl->j_commit_left) 1744 || !(jl->j_state & LIST_DIRTY)) { 1745 del_from_work_list(s, jl); 1746 break; 1747 } 1748 ret = write_one_transaction(s, jl, &chunk); 1749 1750 if (ret < 0) 1751 goto done; 1752 transactions_flushed++; 1753 written += ret; 1754 entry = jl->j_list.next; 1755 1756 /* did we wrap? */ 1757 if (entry == &journal->j_journal_list) { 1758 break; 1759 } 1760 jl = JOURNAL_LIST_ENTRY(entry); 1761 1762 /* don't bother with older transactions */ 1763 if (jl->j_trans_id <= orig_trans_id) 1764 break; 1765 } 1766 if (chunk.nr) { 1767 write_chunk(&chunk); 1768 } 1769 1770 done: 1771 mutex_unlock(&journal->j_flush_mutex); 1772 return ret; 1773 } 1774 1775 /* 1776 * for o_sync and fsync heavy applications, they tend to use 1777 * all the journa list slots with tiny transactions. These 1778 * trigger lots and lots of calls to update the header block, which 1779 * adds seeks and slows things down. 1780 * 1781 * This function tries to clear out a large chunk of the journal lists 1782 * at once, which makes everything faster since only the newest journal 1783 * list updates the header block 1784 */ 1785 static int flush_used_journal_lists(struct super_block *s, 1786 struct reiserfs_journal_list *jl) 1787 { 1788 unsigned long len = 0; 1789 unsigned long cur_len; 1790 int i; 1791 int limit = 256; 1792 struct reiserfs_journal_list *tjl; 1793 struct reiserfs_journal_list *flush_jl; 1794 unsigned int trans_id; 1795 struct reiserfs_journal *journal = SB_JOURNAL(s); 1796 1797 flush_jl = tjl = jl; 1798 1799 /* in data logging mode, try harder to flush a lot of blocks */ 1800 if (reiserfs_data_log(s)) 1801 limit = 1024; 1802 /* flush for 256 transactions or limit blocks, whichever comes first */ 1803 for (i = 0; i < 256 && len < limit; i++) { 1804 if (atomic_read(&tjl->j_commit_left) || 1805 tjl->j_trans_id < jl->j_trans_id) { 1806 break; 1807 } 1808 cur_len = atomic_read(&tjl->j_nonzerolen); 1809 if (cur_len > 0) { 1810 tjl->j_state &= ~LIST_TOUCHED; 1811 } 1812 len += cur_len; 1813 flush_jl = tjl; 1814 if (tjl->j_list.next == &journal->j_journal_list) 1815 break; 1816 tjl = JOURNAL_LIST_ENTRY(tjl->j_list.next); 1817 } 1818 get_journal_list(jl); 1819 get_journal_list(flush_jl); 1820 1821 /* 1822 * try to find a group of blocks we can flush across all the 1823 * transactions, but only bother if we've actually spanned 1824 * across multiple lists 1825 */ 1826 if (flush_jl != jl) 1827 kupdate_transactions(s, jl, &tjl, &trans_id, len, i); 1828 1829 flush_journal_list(s, flush_jl, 1); 1830 put_journal_list(s, flush_jl); 1831 put_journal_list(s, jl); 1832 return 0; 1833 } 1834 1835 /* 1836 * removes any nodes in table with name block and dev as bh. 1837 * only touchs the hnext and hprev pointers. 1838 */ 1839 static void remove_journal_hash(struct super_block *sb, 1840 struct reiserfs_journal_cnode **table, 1841 struct reiserfs_journal_list *jl, 1842 unsigned long block, int remove_freed) 1843 { 1844 struct reiserfs_journal_cnode *cur; 1845 struct reiserfs_journal_cnode **head; 1846 1847 head = &(journal_hash(table, sb, block)); 1848 if (!head) { 1849 return; 1850 } 1851 cur = *head; 1852 while (cur) { 1853 if (cur->blocknr == block && cur->sb == sb 1854 && (jl == NULL || jl == cur->jlist) 1855 && (!test_bit(BLOCK_FREED, &cur->state) || remove_freed)) { 1856 if (cur->hnext) { 1857 cur->hnext->hprev = cur->hprev; 1858 } 1859 if (cur->hprev) { 1860 cur->hprev->hnext = cur->hnext; 1861 } else { 1862 *head = cur->hnext; 1863 } 1864 cur->blocknr = 0; 1865 cur->sb = NULL; 1866 cur->state = 0; 1867 /* 1868 * anybody who clears the cur->bh will also 1869 * dec the nonzerolen 1870 */ 1871 if (cur->bh && cur->jlist) 1872 atomic_dec(&cur->jlist->j_nonzerolen); 1873 cur->bh = NULL; 1874 cur->jlist = NULL; 1875 } 1876 cur = cur->hnext; 1877 } 1878 } 1879 1880 static void free_journal_ram(struct super_block *sb) 1881 { 1882 struct reiserfs_journal *journal = SB_JOURNAL(sb); 1883 kfree(journal->j_current_jl); 1884 journal->j_num_lists--; 1885 1886 vfree(journal->j_cnode_free_orig); 1887 free_list_bitmaps(sb, journal->j_list_bitmap); 1888 free_bitmap_nodes(sb); /* must be after free_list_bitmaps */ 1889 if (journal->j_header_bh) { 1890 brelse(journal->j_header_bh); 1891 } 1892 /* 1893 * j_header_bh is on the journal dev, make sure 1894 * not to release the journal dev until we brelse j_header_bh 1895 */ 1896 release_journal_dev(sb, journal); 1897 vfree(journal); 1898 } 1899 1900 /* 1901 * call on unmount. Only set error to 1 if you haven't made your way out 1902 * of read_super() yet. Any other caller must keep error at 0. 1903 */ 1904 static int do_journal_release(struct reiserfs_transaction_handle *th, 1905 struct super_block *sb, int error) 1906 { 1907 struct reiserfs_transaction_handle myth; 1908 struct reiserfs_journal *journal = SB_JOURNAL(sb); 1909 1910 /* 1911 * we only want to flush out transactions if we were 1912 * called with error == 0 1913 */ 1914 if (!error && !sb_rdonly(sb)) { 1915 /* end the current trans */ 1916 BUG_ON(!th->t_trans_id); 1917 do_journal_end(th, FLUSH_ALL); 1918 1919 /* 1920 * make sure something gets logged to force 1921 * our way into the flush code 1922 */ 1923 if (!journal_join(&myth, sb)) { 1924 reiserfs_prepare_for_journal(sb, 1925 SB_BUFFER_WITH_SB(sb), 1926 1); 1927 journal_mark_dirty(&myth, SB_BUFFER_WITH_SB(sb)); 1928 do_journal_end(&myth, FLUSH_ALL); 1929 } 1930 } 1931 1932 /* this also catches errors during the do_journal_end above */ 1933 if (!error && reiserfs_is_journal_aborted(journal)) { 1934 memset(&myth, 0, sizeof(myth)); 1935 if (!journal_join_abort(&myth, sb)) { 1936 reiserfs_prepare_for_journal(sb, 1937 SB_BUFFER_WITH_SB(sb), 1938 1); 1939 journal_mark_dirty(&myth, SB_BUFFER_WITH_SB(sb)); 1940 do_journal_end(&myth, FLUSH_ALL); 1941 } 1942 } 1943 1944 1945 /* 1946 * We must release the write lock here because 1947 * the workqueue job (flush_async_commit) needs this lock 1948 */ 1949 reiserfs_write_unlock(sb); 1950 1951 /* 1952 * Cancel flushing of old commits. Note that neither of these works 1953 * will be requeued because superblock is being shutdown and doesn't 1954 * have SB_ACTIVE set. 1955 */ 1956 reiserfs_cancel_old_flush(sb); 1957 /* wait for all commits to finish */ 1958 cancel_delayed_work_sync(&SB_JOURNAL(sb)->j_work); 1959 1960 free_journal_ram(sb); 1961 1962 reiserfs_write_lock(sb); 1963 1964 return 0; 1965 } 1966 1967 /* * call on unmount. flush all journal trans, release all alloc'd ram */ 1968 int journal_release(struct reiserfs_transaction_handle *th, 1969 struct super_block *sb) 1970 { 1971 return do_journal_release(th, sb, 0); 1972 } 1973 1974 /* only call from an error condition inside reiserfs_read_super! */ 1975 int journal_release_error(struct reiserfs_transaction_handle *th, 1976 struct super_block *sb) 1977 { 1978 return do_journal_release(th, sb, 1); 1979 } 1980 1981 /* 1982 * compares description block with commit block. 1983 * returns 1 if they differ, 0 if they are the same 1984 */ 1985 static int journal_compare_desc_commit(struct super_block *sb, 1986 struct reiserfs_journal_desc *desc, 1987 struct reiserfs_journal_commit *commit) 1988 { 1989 if (get_commit_trans_id(commit) != get_desc_trans_id(desc) || 1990 get_commit_trans_len(commit) != get_desc_trans_len(desc) || 1991 get_commit_trans_len(commit) > SB_JOURNAL(sb)->j_trans_max || 1992 get_commit_trans_len(commit) <= 0) { 1993 return 1; 1994 } 1995 return 0; 1996 } 1997 1998 /* 1999 * returns 0 if it did not find a description block 2000 * returns -1 if it found a corrupt commit block 2001 * returns 1 if both desc and commit were valid 2002 * NOTE: only called during fs mount 2003 */ 2004 static int journal_transaction_is_valid(struct super_block *sb, 2005 struct buffer_head *d_bh, 2006 unsigned int *oldest_invalid_trans_id, 2007 unsigned long *newest_mount_id) 2008 { 2009 struct reiserfs_journal_desc *desc; 2010 struct reiserfs_journal_commit *commit; 2011 struct buffer_head *c_bh; 2012 unsigned long offset; 2013 2014 if (!d_bh) 2015 return 0; 2016 2017 desc = (struct reiserfs_journal_desc *)d_bh->b_data; 2018 if (get_desc_trans_len(desc) > 0 2019 && !memcmp(get_journal_desc_magic(d_bh), JOURNAL_DESC_MAGIC, 8)) { 2020 if (oldest_invalid_trans_id && *oldest_invalid_trans_id 2021 && get_desc_trans_id(desc) > *oldest_invalid_trans_id) { 2022 reiserfs_debug(sb, REISERFS_DEBUG_CODE, 2023 "journal-986: transaction " 2024 "is valid returning because trans_id %d is greater than " 2025 "oldest_invalid %lu", 2026 get_desc_trans_id(desc), 2027 *oldest_invalid_trans_id); 2028 return 0; 2029 } 2030 if (newest_mount_id 2031 && *newest_mount_id > get_desc_mount_id(desc)) { 2032 reiserfs_debug(sb, REISERFS_DEBUG_CODE, 2033 "journal-1087: transaction " 2034 "is valid returning because mount_id %d is less than " 2035 "newest_mount_id %lu", 2036 get_desc_mount_id(desc), 2037 *newest_mount_id); 2038 return -1; 2039 } 2040 if (get_desc_trans_len(desc) > SB_JOURNAL(sb)->j_trans_max) { 2041 reiserfs_warning(sb, "journal-2018", 2042 "Bad transaction length %d " 2043 "encountered, ignoring transaction", 2044 get_desc_trans_len(desc)); 2045 return -1; 2046 } 2047 offset = d_bh->b_blocknr - SB_ONDISK_JOURNAL_1st_BLOCK(sb); 2048 2049 /* 2050 * ok, we have a journal description block, 2051 * let's see if the transaction was valid 2052 */ 2053 c_bh = 2054 journal_bread(sb, 2055 SB_ONDISK_JOURNAL_1st_BLOCK(sb) + 2056 ((offset + get_desc_trans_len(desc) + 2057 1) % SB_ONDISK_JOURNAL_SIZE(sb))); 2058 if (!c_bh) 2059 return 0; 2060 commit = (struct reiserfs_journal_commit *)c_bh->b_data; 2061 if (journal_compare_desc_commit(sb, desc, commit)) { 2062 reiserfs_debug(sb, REISERFS_DEBUG_CODE, 2063 "journal_transaction_is_valid, commit offset %ld had bad " 2064 "time %d or length %d", 2065 c_bh->b_blocknr - 2066 SB_ONDISK_JOURNAL_1st_BLOCK(sb), 2067 get_commit_trans_id(commit), 2068 get_commit_trans_len(commit)); 2069 brelse(c_bh); 2070 if (oldest_invalid_trans_id) { 2071 *oldest_invalid_trans_id = 2072 get_desc_trans_id(desc); 2073 reiserfs_debug(sb, REISERFS_DEBUG_CODE, 2074 "journal-1004: " 2075 "transaction_is_valid setting oldest invalid trans_id " 2076 "to %d", 2077 get_desc_trans_id(desc)); 2078 } 2079 return -1; 2080 } 2081 brelse(c_bh); 2082 reiserfs_debug(sb, REISERFS_DEBUG_CODE, 2083 "journal-1006: found valid " 2084 "transaction start offset %llu, len %d id %d", 2085 d_bh->b_blocknr - 2086 SB_ONDISK_JOURNAL_1st_BLOCK(sb), 2087 get_desc_trans_len(desc), 2088 get_desc_trans_id(desc)); 2089 return 1; 2090 } else { 2091 return 0; 2092 } 2093 } 2094 2095 static void brelse_array(struct buffer_head **heads, int num) 2096 { 2097 int i; 2098 for (i = 0; i < num; i++) { 2099 brelse(heads[i]); 2100 } 2101 } 2102 2103 /* 2104 * given the start, and values for the oldest acceptable transactions, 2105 * this either reads in a replays a transaction, or returns because the 2106 * transaction is invalid, or too old. 2107 * NOTE: only called during fs mount 2108 */ 2109 static int journal_read_transaction(struct super_block *sb, 2110 unsigned long cur_dblock, 2111 unsigned long oldest_start, 2112 unsigned int oldest_trans_id, 2113 unsigned long newest_mount_id) 2114 { 2115 struct reiserfs_journal *journal = SB_JOURNAL(sb); 2116 struct reiserfs_journal_desc *desc; 2117 struct reiserfs_journal_commit *commit; 2118 unsigned int trans_id = 0; 2119 struct buffer_head *c_bh; 2120 struct buffer_head *d_bh; 2121 struct buffer_head **log_blocks = NULL; 2122 struct buffer_head **real_blocks = NULL; 2123 unsigned int trans_offset; 2124 int i; 2125 int trans_half; 2126 2127 d_bh = journal_bread(sb, cur_dblock); 2128 if (!d_bh) 2129 return 1; 2130 desc = (struct reiserfs_journal_desc *)d_bh->b_data; 2131 trans_offset = d_bh->b_blocknr - SB_ONDISK_JOURNAL_1st_BLOCK(sb); 2132 reiserfs_debug(sb, REISERFS_DEBUG_CODE, "journal-1037: " 2133 "journal_read_transaction, offset %llu, len %d mount_id %d", 2134 d_bh->b_blocknr - SB_ONDISK_JOURNAL_1st_BLOCK(sb), 2135 get_desc_trans_len(desc), get_desc_mount_id(desc)); 2136 if (get_desc_trans_id(desc) < oldest_trans_id) { 2137 reiserfs_debug(sb, REISERFS_DEBUG_CODE, "journal-1039: " 2138 "journal_read_trans skipping because %lu is too old", 2139 cur_dblock - 2140 SB_ONDISK_JOURNAL_1st_BLOCK(sb)); 2141 brelse(d_bh); 2142 return 1; 2143 } 2144 if (get_desc_mount_id(desc) != newest_mount_id) { 2145 reiserfs_debug(sb, REISERFS_DEBUG_CODE, "journal-1146: " 2146 "journal_read_trans skipping because %d is != " 2147 "newest_mount_id %lu", get_desc_mount_id(desc), 2148 newest_mount_id); 2149 brelse(d_bh); 2150 return 1; 2151 } 2152 c_bh = journal_bread(sb, SB_ONDISK_JOURNAL_1st_BLOCK(sb) + 2153 ((trans_offset + get_desc_trans_len(desc) + 1) % 2154 SB_ONDISK_JOURNAL_SIZE(sb))); 2155 if (!c_bh) { 2156 brelse(d_bh); 2157 return 1; 2158 } 2159 commit = (struct reiserfs_journal_commit *)c_bh->b_data; 2160 if (journal_compare_desc_commit(sb, desc, commit)) { 2161 reiserfs_debug(sb, REISERFS_DEBUG_CODE, 2162 "journal_read_transaction, " 2163 "commit offset %llu had bad time %d or length %d", 2164 c_bh->b_blocknr - 2165 SB_ONDISK_JOURNAL_1st_BLOCK(sb), 2166 get_commit_trans_id(commit), 2167 get_commit_trans_len(commit)); 2168 brelse(c_bh); 2169 brelse(d_bh); 2170 return 1; 2171 } 2172 2173 if (bdev_read_only(sb->s_bdev)) { 2174 reiserfs_warning(sb, "clm-2076", 2175 "device is readonly, unable to replay log"); 2176 brelse(c_bh); 2177 brelse(d_bh); 2178 return -EROFS; 2179 } 2180 2181 trans_id = get_desc_trans_id(desc); 2182 /* 2183 * now we know we've got a good transaction, and it was 2184 * inside the valid time ranges 2185 */ 2186 log_blocks = kmalloc_array(get_desc_trans_len(desc), 2187 sizeof(struct buffer_head *), 2188 GFP_NOFS); 2189 real_blocks = kmalloc_array(get_desc_trans_len(desc), 2190 sizeof(struct buffer_head *), 2191 GFP_NOFS); 2192 if (!log_blocks || !real_blocks) { 2193 brelse(c_bh); 2194 brelse(d_bh); 2195 kfree(log_blocks); 2196 kfree(real_blocks); 2197 reiserfs_warning(sb, "journal-1169", 2198 "kmalloc failed, unable to mount FS"); 2199 return -1; 2200 } 2201 /* get all the buffer heads */ 2202 trans_half = journal_trans_half(sb->s_blocksize); 2203 for (i = 0; i < get_desc_trans_len(desc); i++) { 2204 log_blocks[i] = 2205 journal_getblk(sb, 2206 SB_ONDISK_JOURNAL_1st_BLOCK(sb) + 2207 (trans_offset + 1 + 2208 i) % SB_ONDISK_JOURNAL_SIZE(sb)); 2209 if (i < trans_half) { 2210 real_blocks[i] = 2211 sb_getblk(sb, 2212 le32_to_cpu(desc->j_realblock[i])); 2213 } else { 2214 real_blocks[i] = 2215 sb_getblk(sb, 2216 le32_to_cpu(commit-> 2217 j_realblock[i - trans_half])); 2218 } 2219 if (real_blocks[i]->b_blocknr > SB_BLOCK_COUNT(sb)) { 2220 reiserfs_warning(sb, "journal-1207", 2221 "REPLAY FAILURE fsck required! " 2222 "Block to replay is outside of " 2223 "filesystem"); 2224 goto abort_replay; 2225 } 2226 /* make sure we don't try to replay onto log or reserved area */ 2227 if (is_block_in_log_or_reserved_area 2228 (sb, real_blocks[i]->b_blocknr)) { 2229 reiserfs_warning(sb, "journal-1204", 2230 "REPLAY FAILURE fsck required! " 2231 "Trying to replay onto a log block"); 2232 abort_replay: 2233 brelse_array(log_blocks, i); 2234 brelse_array(real_blocks, i); 2235 brelse(c_bh); 2236 brelse(d_bh); 2237 kfree(log_blocks); 2238 kfree(real_blocks); 2239 return -1; 2240 } 2241 } 2242 /* read in the log blocks, memcpy to the corresponding real block */ 2243 bh_read_batch(get_desc_trans_len(desc), log_blocks); 2244 for (i = 0; i < get_desc_trans_len(desc); i++) { 2245 2246 wait_on_buffer(log_blocks[i]); 2247 if (!buffer_uptodate(log_blocks[i])) { 2248 reiserfs_warning(sb, "journal-1212", 2249 "REPLAY FAILURE fsck required! " 2250 "buffer write failed"); 2251 brelse_array(log_blocks + i, 2252 get_desc_trans_len(desc) - i); 2253 brelse_array(real_blocks, get_desc_trans_len(desc)); 2254 brelse(c_bh); 2255 brelse(d_bh); 2256 kfree(log_blocks); 2257 kfree(real_blocks); 2258 return -1; 2259 } 2260 memcpy(real_blocks[i]->b_data, log_blocks[i]->b_data, 2261 real_blocks[i]->b_size); 2262 set_buffer_uptodate(real_blocks[i]); 2263 brelse(log_blocks[i]); 2264 } 2265 /* flush out the real blocks */ 2266 for (i = 0; i < get_desc_trans_len(desc); i++) { 2267 set_buffer_dirty(real_blocks[i]); 2268 write_dirty_buffer(real_blocks[i], 0); 2269 } 2270 for (i = 0; i < get_desc_trans_len(desc); i++) { 2271 wait_on_buffer(real_blocks[i]); 2272 if (!buffer_uptodate(real_blocks[i])) { 2273 reiserfs_warning(sb, "journal-1226", 2274 "REPLAY FAILURE, fsck required! " 2275 "buffer write failed"); 2276 brelse_array(real_blocks + i, 2277 get_desc_trans_len(desc) - i); 2278 brelse(c_bh); 2279 brelse(d_bh); 2280 kfree(log_blocks); 2281 kfree(real_blocks); 2282 return -1; 2283 } 2284 brelse(real_blocks[i]); 2285 } 2286 cur_dblock = 2287 SB_ONDISK_JOURNAL_1st_BLOCK(sb) + 2288 ((trans_offset + get_desc_trans_len(desc) + 2289 2) % SB_ONDISK_JOURNAL_SIZE(sb)); 2290 reiserfs_debug(sb, REISERFS_DEBUG_CODE, 2291 "journal-1095: setting journal " "start to offset %ld", 2292 cur_dblock - SB_ONDISK_JOURNAL_1st_BLOCK(sb)); 2293 2294 /* 2295 * init starting values for the first transaction, in case 2296 * this is the last transaction to be replayed. 2297 */ 2298 journal->j_start = cur_dblock - SB_ONDISK_JOURNAL_1st_BLOCK(sb); 2299 journal->j_last_flush_trans_id = trans_id; 2300 journal->j_trans_id = trans_id + 1; 2301 /* check for trans_id overflow */ 2302 if (journal->j_trans_id == 0) 2303 journal->j_trans_id = 10; 2304 brelse(c_bh); 2305 brelse(d_bh); 2306 kfree(log_blocks); 2307 kfree(real_blocks); 2308 return 0; 2309 } 2310 2311 /* 2312 * This function reads blocks starting from block and to max_block of bufsize 2313 * size (but no more than BUFNR blocks at a time). This proved to improve 2314 * mounting speed on self-rebuilding raid5 arrays at least. 2315 * Right now it is only used from journal code. But later we might use it 2316 * from other places. 2317 * Note: Do not use journal_getblk/sb_getblk functions here! 2318 */ 2319 static struct buffer_head *reiserfs_breada(struct block_device *dev, 2320 b_blocknr_t block, int bufsize, 2321 b_blocknr_t max_block) 2322 { 2323 struct buffer_head *bhlist[BUFNR]; 2324 unsigned int blocks = BUFNR; 2325 struct buffer_head *bh; 2326 int i, j; 2327 2328 bh = __getblk(dev, block, bufsize); 2329 if (buffer_uptodate(bh)) 2330 return (bh); 2331 2332 if (block + BUFNR > max_block) { 2333 blocks = max_block - block; 2334 } 2335 bhlist[0] = bh; 2336 j = 1; 2337 for (i = 1; i < blocks; i++) { 2338 bh = __getblk(dev, block + i, bufsize); 2339 if (buffer_uptodate(bh)) { 2340 brelse(bh); 2341 break; 2342 } else 2343 bhlist[j++] = bh; 2344 } 2345 bh = bhlist[0]; 2346 bh_read_nowait(bh, 0); 2347 bh_readahead_batch(j - 1, &bhlist[1], 0); 2348 for (i = 1; i < j; i++) 2349 brelse(bhlist[i]); 2350 wait_on_buffer(bh); 2351 if (buffer_uptodate(bh)) 2352 return bh; 2353 brelse(bh); 2354 return NULL; 2355 } 2356 2357 /* 2358 * read and replay the log 2359 * on a clean unmount, the journal header's next unflushed pointer will be 2360 * to an invalid transaction. This tests that before finding all the 2361 * transactions in the log, which makes normal mount times fast. 2362 * 2363 * After a crash, this starts with the next unflushed transaction, and 2364 * replays until it finds one too old, or invalid. 2365 * 2366 * On exit, it sets things up so the first transaction will work correctly. 2367 * NOTE: only called during fs mount 2368 */ 2369 static int journal_read(struct super_block *sb) 2370 { 2371 struct reiserfs_journal *journal = SB_JOURNAL(sb); 2372 struct reiserfs_journal_desc *desc; 2373 unsigned int oldest_trans_id = 0; 2374 unsigned int oldest_invalid_trans_id = 0; 2375 time64_t start; 2376 unsigned long oldest_start = 0; 2377 unsigned long cur_dblock = 0; 2378 unsigned long newest_mount_id = 9; 2379 struct buffer_head *d_bh; 2380 struct reiserfs_journal_header *jh; 2381 int valid_journal_header = 0; 2382 int replay_count = 0; 2383 int continue_replay = 1; 2384 int ret; 2385 2386 cur_dblock = SB_ONDISK_JOURNAL_1st_BLOCK(sb); 2387 reiserfs_info(sb, "checking transaction log (%pg)\n", 2388 journal->j_dev_bd); 2389 start = ktime_get_seconds(); 2390 2391 /* 2392 * step 1, read in the journal header block. Check the transaction 2393 * it says is the first unflushed, and if that transaction is not 2394 * valid, replay is done 2395 */ 2396 journal->j_header_bh = journal_bread(sb, 2397 SB_ONDISK_JOURNAL_1st_BLOCK(sb) 2398 + SB_ONDISK_JOURNAL_SIZE(sb)); 2399 if (!journal->j_header_bh) { 2400 return 1; 2401 } 2402 jh = (struct reiserfs_journal_header *)(journal->j_header_bh->b_data); 2403 if (le32_to_cpu(jh->j_first_unflushed_offset) < 2404 SB_ONDISK_JOURNAL_SIZE(sb) 2405 && le32_to_cpu(jh->j_last_flush_trans_id) > 0) { 2406 oldest_start = 2407 SB_ONDISK_JOURNAL_1st_BLOCK(sb) + 2408 le32_to_cpu(jh->j_first_unflushed_offset); 2409 oldest_trans_id = le32_to_cpu(jh->j_last_flush_trans_id) + 1; 2410 newest_mount_id = le32_to_cpu(jh->j_mount_id); 2411 reiserfs_debug(sb, REISERFS_DEBUG_CODE, 2412 "journal-1153: found in " 2413 "header: first_unflushed_offset %d, last_flushed_trans_id " 2414 "%lu", le32_to_cpu(jh->j_first_unflushed_offset), 2415 le32_to_cpu(jh->j_last_flush_trans_id)); 2416 valid_journal_header = 1; 2417 2418 /* 2419 * now, we try to read the first unflushed offset. If it 2420 * is not valid, there is nothing more we can do, and it 2421 * makes no sense to read through the whole log. 2422 */ 2423 d_bh = 2424 journal_bread(sb, 2425 SB_ONDISK_JOURNAL_1st_BLOCK(sb) + 2426 le32_to_cpu(jh->j_first_unflushed_offset)); 2427 ret = journal_transaction_is_valid(sb, d_bh, NULL, NULL); 2428 if (!ret) { 2429 continue_replay = 0; 2430 } 2431 brelse(d_bh); 2432 goto start_log_replay; 2433 } 2434 2435 /* 2436 * ok, there are transactions that need to be replayed. start 2437 * with the first log block, find all the valid transactions, and 2438 * pick out the oldest. 2439 */ 2440 while (continue_replay 2441 && cur_dblock < 2442 (SB_ONDISK_JOURNAL_1st_BLOCK(sb) + 2443 SB_ONDISK_JOURNAL_SIZE(sb))) { 2444 /* 2445 * Note that it is required for blocksize of primary fs 2446 * device and journal device to be the same 2447 */ 2448 d_bh = 2449 reiserfs_breada(journal->j_dev_bd, cur_dblock, 2450 sb->s_blocksize, 2451 SB_ONDISK_JOURNAL_1st_BLOCK(sb) + 2452 SB_ONDISK_JOURNAL_SIZE(sb)); 2453 ret = 2454 journal_transaction_is_valid(sb, d_bh, 2455 &oldest_invalid_trans_id, 2456 &newest_mount_id); 2457 if (ret == 1) { 2458 desc = (struct reiserfs_journal_desc *)d_bh->b_data; 2459 if (oldest_start == 0) { /* init all oldest_ values */ 2460 oldest_trans_id = get_desc_trans_id(desc); 2461 oldest_start = d_bh->b_blocknr; 2462 newest_mount_id = get_desc_mount_id(desc); 2463 reiserfs_debug(sb, REISERFS_DEBUG_CODE, 2464 "journal-1179: Setting " 2465 "oldest_start to offset %llu, trans_id %lu", 2466 oldest_start - 2467 SB_ONDISK_JOURNAL_1st_BLOCK 2468 (sb), oldest_trans_id); 2469 } else if (oldest_trans_id > get_desc_trans_id(desc)) { 2470 /* one we just read was older */ 2471 oldest_trans_id = get_desc_trans_id(desc); 2472 oldest_start = d_bh->b_blocknr; 2473 reiserfs_debug(sb, REISERFS_DEBUG_CODE, 2474 "journal-1180: Resetting " 2475 "oldest_start to offset %lu, trans_id %lu", 2476 oldest_start - 2477 SB_ONDISK_JOURNAL_1st_BLOCK 2478 (sb), oldest_trans_id); 2479 } 2480 if (newest_mount_id < get_desc_mount_id(desc)) { 2481 newest_mount_id = get_desc_mount_id(desc); 2482 reiserfs_debug(sb, REISERFS_DEBUG_CODE, 2483 "journal-1299: Setting " 2484 "newest_mount_id to %d", 2485 get_desc_mount_id(desc)); 2486 } 2487 cur_dblock += get_desc_trans_len(desc) + 2; 2488 } else { 2489 cur_dblock++; 2490 } 2491 brelse(d_bh); 2492 } 2493 2494 start_log_replay: 2495 cur_dblock = oldest_start; 2496 if (oldest_trans_id) { 2497 reiserfs_debug(sb, REISERFS_DEBUG_CODE, 2498 "journal-1206: Starting replay " 2499 "from offset %llu, trans_id %lu", 2500 cur_dblock - SB_ONDISK_JOURNAL_1st_BLOCK(sb), 2501 oldest_trans_id); 2502 2503 } 2504 replay_count = 0; 2505 while (continue_replay && oldest_trans_id > 0) { 2506 ret = 2507 journal_read_transaction(sb, cur_dblock, oldest_start, 2508 oldest_trans_id, newest_mount_id); 2509 if (ret < 0) { 2510 return ret; 2511 } else if (ret != 0) { 2512 break; 2513 } 2514 cur_dblock = 2515 SB_ONDISK_JOURNAL_1st_BLOCK(sb) + journal->j_start; 2516 replay_count++; 2517 if (cur_dblock == oldest_start) 2518 break; 2519 } 2520 2521 if (oldest_trans_id == 0) { 2522 reiserfs_debug(sb, REISERFS_DEBUG_CODE, 2523 "journal-1225: No valid " "transactions found"); 2524 } 2525 /* 2526 * j_start does not get set correctly if we don't replay any 2527 * transactions. if we had a valid journal_header, set j_start 2528 * to the first unflushed transaction value, copy the trans_id 2529 * from the header 2530 */ 2531 if (valid_journal_header && replay_count == 0) { 2532 journal->j_start = le32_to_cpu(jh->j_first_unflushed_offset); 2533 journal->j_trans_id = 2534 le32_to_cpu(jh->j_last_flush_trans_id) + 1; 2535 /* check for trans_id overflow */ 2536 if (journal->j_trans_id == 0) 2537 journal->j_trans_id = 10; 2538 journal->j_last_flush_trans_id = 2539 le32_to_cpu(jh->j_last_flush_trans_id); 2540 journal->j_mount_id = le32_to_cpu(jh->j_mount_id) + 1; 2541 } else { 2542 journal->j_mount_id = newest_mount_id + 1; 2543 } 2544 reiserfs_debug(sb, REISERFS_DEBUG_CODE, "journal-1299: Setting " 2545 "newest_mount_id to %lu", journal->j_mount_id); 2546 journal->j_first_unflushed_offset = journal->j_start; 2547 if (replay_count > 0) { 2548 reiserfs_info(sb, 2549 "replayed %d transactions in %lu seconds\n", 2550 replay_count, ktime_get_seconds() - start); 2551 } 2552 /* needed to satisfy the locking in _update_journal_header_block */ 2553 reiserfs_write_lock(sb); 2554 if (!bdev_read_only(sb->s_bdev) && 2555 _update_journal_header_block(sb, journal->j_start, 2556 journal->j_last_flush_trans_id)) { 2557 reiserfs_write_unlock(sb); 2558 /* 2559 * replay failed, caller must call free_journal_ram and abort 2560 * the mount 2561 */ 2562 return -1; 2563 } 2564 reiserfs_write_unlock(sb); 2565 return 0; 2566 } 2567 2568 static struct reiserfs_journal_list *alloc_journal_list(struct super_block *s) 2569 { 2570 struct reiserfs_journal_list *jl; 2571 jl = kzalloc(sizeof(struct reiserfs_journal_list), 2572 GFP_NOFS | __GFP_NOFAIL); 2573 INIT_LIST_HEAD(&jl->j_list); 2574 INIT_LIST_HEAD(&jl->j_working_list); 2575 INIT_LIST_HEAD(&jl->j_tail_bh_list); 2576 INIT_LIST_HEAD(&jl->j_bh_list); 2577 mutex_init(&jl->j_commit_mutex); 2578 SB_JOURNAL(s)->j_num_lists++; 2579 get_journal_list(jl); 2580 return jl; 2581 } 2582 2583 static void journal_list_init(struct super_block *sb) 2584 { 2585 SB_JOURNAL(sb)->j_current_jl = alloc_journal_list(sb); 2586 } 2587 2588 static void release_journal_dev(struct super_block *super, 2589 struct reiserfs_journal *journal) 2590 { 2591 if (journal->j_dev_bd != NULL) { 2592 blkdev_put(journal->j_dev_bd, journal->j_dev_mode); 2593 journal->j_dev_bd = NULL; 2594 } 2595 } 2596 2597 static int journal_init_dev(struct super_block *super, 2598 struct reiserfs_journal *journal, 2599 const char *jdev_name) 2600 { 2601 int result; 2602 dev_t jdev; 2603 fmode_t blkdev_mode = FMODE_READ | FMODE_WRITE | FMODE_EXCL; 2604 2605 result = 0; 2606 2607 journal->j_dev_bd = NULL; 2608 jdev = SB_ONDISK_JOURNAL_DEVICE(super) ? 2609 new_decode_dev(SB_ONDISK_JOURNAL_DEVICE(super)) : super->s_dev; 2610 2611 if (bdev_read_only(super->s_bdev)) 2612 blkdev_mode = FMODE_READ; 2613 2614 /* there is no "jdev" option and journal is on separate device */ 2615 if ((!jdev_name || !jdev_name[0])) { 2616 if (jdev == super->s_dev) 2617 blkdev_mode &= ~FMODE_EXCL; 2618 journal->j_dev_bd = blkdev_get_by_dev(jdev, blkdev_mode, 2619 journal); 2620 journal->j_dev_mode = blkdev_mode; 2621 if (IS_ERR(journal->j_dev_bd)) { 2622 result = PTR_ERR(journal->j_dev_bd); 2623 journal->j_dev_bd = NULL; 2624 reiserfs_warning(super, "sh-458", 2625 "cannot init journal device unknown-block(%u,%u): %i", 2626 MAJOR(jdev), MINOR(jdev), result); 2627 return result; 2628 } else if (jdev != super->s_dev) 2629 set_blocksize(journal->j_dev_bd, super->s_blocksize); 2630 2631 return 0; 2632 } 2633 2634 journal->j_dev_mode = blkdev_mode; 2635 journal->j_dev_bd = blkdev_get_by_path(jdev_name, blkdev_mode, journal); 2636 if (IS_ERR(journal->j_dev_bd)) { 2637 result = PTR_ERR(journal->j_dev_bd); 2638 journal->j_dev_bd = NULL; 2639 reiserfs_warning(super, "sh-457", 2640 "journal_init_dev: Cannot open '%s': %i", 2641 jdev_name, result); 2642 return result; 2643 } 2644 2645 set_blocksize(journal->j_dev_bd, super->s_blocksize); 2646 reiserfs_info(super, 2647 "journal_init_dev: journal device: %pg\n", 2648 journal->j_dev_bd); 2649 return 0; 2650 } 2651 2652 /* 2653 * When creating/tuning a file system user can assign some 2654 * journal params within boundaries which depend on the ratio 2655 * blocksize/standard_blocksize. 2656 * 2657 * For blocks >= standard_blocksize transaction size should 2658 * be not less then JOURNAL_TRANS_MIN_DEFAULT, and not more 2659 * then JOURNAL_TRANS_MAX_DEFAULT. 2660 * 2661 * For blocks < standard_blocksize these boundaries should be 2662 * decreased proportionally. 2663 */ 2664 #define REISERFS_STANDARD_BLKSIZE (4096) 2665 2666 static int check_advise_trans_params(struct super_block *sb, 2667 struct reiserfs_journal *journal) 2668 { 2669 if (journal->j_trans_max) { 2670 /* Non-default journal params. Do sanity check for them. */ 2671 int ratio = 1; 2672 if (sb->s_blocksize < REISERFS_STANDARD_BLKSIZE) 2673 ratio = REISERFS_STANDARD_BLKSIZE / sb->s_blocksize; 2674 2675 if (journal->j_trans_max > JOURNAL_TRANS_MAX_DEFAULT / ratio || 2676 journal->j_trans_max < JOURNAL_TRANS_MIN_DEFAULT / ratio || 2677 SB_ONDISK_JOURNAL_SIZE(sb) / journal->j_trans_max < 2678 JOURNAL_MIN_RATIO) { 2679 reiserfs_warning(sb, "sh-462", 2680 "bad transaction max size (%u). " 2681 "FSCK?", journal->j_trans_max); 2682 return 1; 2683 } 2684 if (journal->j_max_batch != (journal->j_trans_max) * 2685 JOURNAL_MAX_BATCH_DEFAULT/JOURNAL_TRANS_MAX_DEFAULT) { 2686 reiserfs_warning(sb, "sh-463", 2687 "bad transaction max batch (%u). " 2688 "FSCK?", journal->j_max_batch); 2689 return 1; 2690 } 2691 } else { 2692 /* 2693 * Default journal params. 2694 * The file system was created by old version 2695 * of mkreiserfs, so some fields contain zeros, 2696 * and we need to advise proper values for them 2697 */ 2698 if (sb->s_blocksize != REISERFS_STANDARD_BLKSIZE) { 2699 reiserfs_warning(sb, "sh-464", "bad blocksize (%u)", 2700 sb->s_blocksize); 2701 return 1; 2702 } 2703 journal->j_trans_max = JOURNAL_TRANS_MAX_DEFAULT; 2704 journal->j_max_batch = JOURNAL_MAX_BATCH_DEFAULT; 2705 journal->j_max_commit_age = JOURNAL_MAX_COMMIT_AGE; 2706 } 2707 return 0; 2708 } 2709 2710 /* must be called once on fs mount. calls journal_read for you */ 2711 int journal_init(struct super_block *sb, const char *j_dev_name, 2712 int old_format, unsigned int commit_max_age) 2713 { 2714 int num_cnodes = SB_ONDISK_JOURNAL_SIZE(sb) * 2; 2715 struct buffer_head *bhjh; 2716 struct reiserfs_super_block *rs; 2717 struct reiserfs_journal_header *jh; 2718 struct reiserfs_journal *journal; 2719 struct reiserfs_journal_list *jl; 2720 int ret; 2721 2722 journal = SB_JOURNAL(sb) = vzalloc(sizeof(struct reiserfs_journal)); 2723 if (!journal) { 2724 reiserfs_warning(sb, "journal-1256", 2725 "unable to get memory for journal structure"); 2726 return 1; 2727 } 2728 INIT_LIST_HEAD(&journal->j_bitmap_nodes); 2729 INIT_LIST_HEAD(&journal->j_prealloc_list); 2730 INIT_LIST_HEAD(&journal->j_working_list); 2731 INIT_LIST_HEAD(&journal->j_journal_list); 2732 journal->j_persistent_trans = 0; 2733 if (reiserfs_allocate_list_bitmaps(sb, journal->j_list_bitmap, 2734 reiserfs_bmap_count(sb))) 2735 goto free_and_return; 2736 2737 allocate_bitmap_nodes(sb); 2738 2739 /* reserved for journal area support */ 2740 SB_JOURNAL_1st_RESERVED_BLOCK(sb) = (old_format ? 2741 REISERFS_OLD_DISK_OFFSET_IN_BYTES 2742 / sb->s_blocksize + 2743 reiserfs_bmap_count(sb) + 2744 1 : 2745 REISERFS_DISK_OFFSET_IN_BYTES / 2746 sb->s_blocksize + 2); 2747 2748 /* 2749 * Sanity check to see is the standard journal fitting 2750 * within first bitmap (actual for small blocksizes) 2751 */ 2752 if (!SB_ONDISK_JOURNAL_DEVICE(sb) && 2753 (SB_JOURNAL_1st_RESERVED_BLOCK(sb) + 2754 SB_ONDISK_JOURNAL_SIZE(sb) > sb->s_blocksize * 8)) { 2755 reiserfs_warning(sb, "journal-1393", 2756 "journal does not fit for area addressed " 2757 "by first of bitmap blocks. It starts at " 2758 "%u and its size is %u. Block size %ld", 2759 SB_JOURNAL_1st_RESERVED_BLOCK(sb), 2760 SB_ONDISK_JOURNAL_SIZE(sb), 2761 sb->s_blocksize); 2762 goto free_and_return; 2763 } 2764 2765 /* 2766 * Sanity check to see if journal first block is correct. 2767 * If journal first block is invalid it can cause 2768 * zeroing important superblock members. 2769 */ 2770 if (!SB_ONDISK_JOURNAL_DEVICE(sb) && 2771 SB_ONDISK_JOURNAL_1st_BLOCK(sb) < SB_JOURNAL_1st_RESERVED_BLOCK(sb)) { 2772 reiserfs_warning(sb, "journal-1393", 2773 "journal 1st super block is invalid: 1st reserved block %d, but actual 1st block is %d", 2774 SB_JOURNAL_1st_RESERVED_BLOCK(sb), 2775 SB_ONDISK_JOURNAL_1st_BLOCK(sb)); 2776 goto free_and_return; 2777 } 2778 2779 if (journal_init_dev(sb, journal, j_dev_name) != 0) { 2780 reiserfs_warning(sb, "sh-462", 2781 "unable to initialize journal device"); 2782 goto free_and_return; 2783 } 2784 2785 rs = SB_DISK_SUPER_BLOCK(sb); 2786 2787 /* read journal header */ 2788 bhjh = journal_bread(sb, 2789 SB_ONDISK_JOURNAL_1st_BLOCK(sb) + 2790 SB_ONDISK_JOURNAL_SIZE(sb)); 2791 if (!bhjh) { 2792 reiserfs_warning(sb, "sh-459", 2793 "unable to read journal header"); 2794 goto free_and_return; 2795 } 2796 jh = (struct reiserfs_journal_header *)(bhjh->b_data); 2797 2798 /* make sure that journal matches to the super block */ 2799 if (is_reiserfs_jr(rs) 2800 && (le32_to_cpu(jh->jh_journal.jp_journal_magic) != 2801 sb_jp_journal_magic(rs))) { 2802 reiserfs_warning(sb, "sh-460", 2803 "journal header magic %x (device %pg) does " 2804 "not match to magic found in super block %x", 2805 jh->jh_journal.jp_journal_magic, 2806 journal->j_dev_bd, 2807 sb_jp_journal_magic(rs)); 2808 brelse(bhjh); 2809 goto free_and_return; 2810 } 2811 2812 journal->j_trans_max = le32_to_cpu(jh->jh_journal.jp_journal_trans_max); 2813 journal->j_max_batch = le32_to_cpu(jh->jh_journal.jp_journal_max_batch); 2814 journal->j_max_commit_age = 2815 le32_to_cpu(jh->jh_journal.jp_journal_max_commit_age); 2816 journal->j_max_trans_age = JOURNAL_MAX_TRANS_AGE; 2817 2818 if (check_advise_trans_params(sb, journal) != 0) 2819 goto free_and_return; 2820 journal->j_default_max_commit_age = journal->j_max_commit_age; 2821 2822 if (commit_max_age != 0) { 2823 journal->j_max_commit_age = commit_max_age; 2824 journal->j_max_trans_age = commit_max_age; 2825 } 2826 2827 reiserfs_info(sb, "journal params: device %pg, size %u, " 2828 "journal first block %u, max trans len %u, max batch %u, " 2829 "max commit age %u, max trans age %u\n", 2830 journal->j_dev_bd, 2831 SB_ONDISK_JOURNAL_SIZE(sb), 2832 SB_ONDISK_JOURNAL_1st_BLOCK(sb), 2833 journal->j_trans_max, 2834 journal->j_max_batch, 2835 journal->j_max_commit_age, journal->j_max_trans_age); 2836 2837 brelse(bhjh); 2838 2839 journal->j_list_bitmap_index = 0; 2840 journal_list_init(sb); 2841 2842 memset(journal->j_list_hash_table, 0, 2843 JOURNAL_HASH_SIZE * sizeof(struct reiserfs_journal_cnode *)); 2844 2845 INIT_LIST_HEAD(&journal->j_dirty_buffers); 2846 spin_lock_init(&journal->j_dirty_buffers_lock); 2847 2848 journal->j_start = 0; 2849 journal->j_len = 0; 2850 journal->j_len_alloc = 0; 2851 atomic_set(&journal->j_wcount, 0); 2852 atomic_set(&journal->j_async_throttle, 0); 2853 journal->j_bcount = 0; 2854 journal->j_trans_start_time = 0; 2855 journal->j_last = NULL; 2856 journal->j_first = NULL; 2857 init_waitqueue_head(&journal->j_join_wait); 2858 mutex_init(&journal->j_mutex); 2859 mutex_init(&journal->j_flush_mutex); 2860 2861 journal->j_trans_id = 10; 2862 journal->j_mount_id = 10; 2863 journal->j_state = 0; 2864 atomic_set(&journal->j_jlock, 0); 2865 journal->j_cnode_free_list = allocate_cnodes(num_cnodes); 2866 journal->j_cnode_free_orig = journal->j_cnode_free_list; 2867 journal->j_cnode_free = journal->j_cnode_free_list ? num_cnodes : 0; 2868 journal->j_cnode_used = 0; 2869 journal->j_must_wait = 0; 2870 2871 if (journal->j_cnode_free == 0) { 2872 reiserfs_warning(sb, "journal-2004", "Journal cnode memory " 2873 "allocation failed (%ld bytes). Journal is " 2874 "too large for available memory. Usually " 2875 "this is due to a journal that is too large.", 2876 sizeof (struct reiserfs_journal_cnode) * num_cnodes); 2877 goto free_and_return; 2878 } 2879 2880 init_journal_hash(sb); 2881 jl = journal->j_current_jl; 2882 2883 /* 2884 * get_list_bitmap() may call flush_commit_list() which 2885 * requires the lock. Calling flush_commit_list() shouldn't happen 2886 * this early but I like to be paranoid. 2887 */ 2888 reiserfs_write_lock(sb); 2889 jl->j_list_bitmap = get_list_bitmap(sb, jl); 2890 reiserfs_write_unlock(sb); 2891 if (!jl->j_list_bitmap) { 2892 reiserfs_warning(sb, "journal-2005", 2893 "get_list_bitmap failed for journal list 0"); 2894 goto free_and_return; 2895 } 2896 2897 ret = journal_read(sb); 2898 if (ret < 0) { 2899 reiserfs_warning(sb, "reiserfs-2006", 2900 "Replay Failure, unable to mount"); 2901 goto free_and_return; 2902 } 2903 2904 INIT_DELAYED_WORK(&journal->j_work, flush_async_commits); 2905 journal->j_work_sb = sb; 2906 return 0; 2907 free_and_return: 2908 free_journal_ram(sb); 2909 return 1; 2910 } 2911 2912 /* 2913 * test for a polite end of the current transaction. Used by file_write, 2914 * and should be used by delete to make sure they don't write more than 2915 * can fit inside a single transaction 2916 */ 2917 int journal_transaction_should_end(struct reiserfs_transaction_handle *th, 2918 int new_alloc) 2919 { 2920 struct reiserfs_journal *journal = SB_JOURNAL(th->t_super); 2921 time64_t now = ktime_get_seconds(); 2922 /* cannot restart while nested */ 2923 BUG_ON(!th->t_trans_id); 2924 if (th->t_refcount > 1) 2925 return 0; 2926 if (journal->j_must_wait > 0 || 2927 (journal->j_len_alloc + new_alloc) >= journal->j_max_batch || 2928 atomic_read(&journal->j_jlock) || 2929 (now - journal->j_trans_start_time) > journal->j_max_trans_age || 2930 journal->j_cnode_free < (journal->j_trans_max * 3)) { 2931 return 1; 2932 } 2933 2934 journal->j_len_alloc += new_alloc; 2935 th->t_blocks_allocated += new_alloc ; 2936 return 0; 2937 } 2938 2939 /* this must be called inside a transaction */ 2940 void reiserfs_block_writes(struct reiserfs_transaction_handle *th) 2941 { 2942 struct reiserfs_journal *journal = SB_JOURNAL(th->t_super); 2943 BUG_ON(!th->t_trans_id); 2944 journal->j_must_wait = 1; 2945 set_bit(J_WRITERS_BLOCKED, &journal->j_state); 2946 return; 2947 } 2948 2949 /* this must be called without a transaction started */ 2950 void reiserfs_allow_writes(struct super_block *s) 2951 { 2952 struct reiserfs_journal *journal = SB_JOURNAL(s); 2953 clear_bit(J_WRITERS_BLOCKED, &journal->j_state); 2954 wake_up(&journal->j_join_wait); 2955 } 2956 2957 /* this must be called without a transaction started */ 2958 void reiserfs_wait_on_write_block(struct super_block *s) 2959 { 2960 struct reiserfs_journal *journal = SB_JOURNAL(s); 2961 wait_event(journal->j_join_wait, 2962 !test_bit(J_WRITERS_BLOCKED, &journal->j_state)); 2963 } 2964 2965 static void queue_log_writer(struct super_block *s) 2966 { 2967 wait_queue_entry_t wait; 2968 struct reiserfs_journal *journal = SB_JOURNAL(s); 2969 set_bit(J_WRITERS_QUEUED, &journal->j_state); 2970 2971 /* 2972 * we don't want to use wait_event here because 2973 * we only want to wait once. 2974 */ 2975 init_waitqueue_entry(&wait, current); 2976 add_wait_queue(&journal->j_join_wait, &wait); 2977 set_current_state(TASK_UNINTERRUPTIBLE); 2978 if (test_bit(J_WRITERS_QUEUED, &journal->j_state)) { 2979 int depth = reiserfs_write_unlock_nested(s); 2980 schedule(); 2981 reiserfs_write_lock_nested(s, depth); 2982 } 2983 __set_current_state(TASK_RUNNING); 2984 remove_wait_queue(&journal->j_join_wait, &wait); 2985 } 2986 2987 static void wake_queued_writers(struct super_block *s) 2988 { 2989 struct reiserfs_journal *journal = SB_JOURNAL(s); 2990 if (test_and_clear_bit(J_WRITERS_QUEUED, &journal->j_state)) 2991 wake_up(&journal->j_join_wait); 2992 } 2993 2994 static void let_transaction_grow(struct super_block *sb, unsigned int trans_id) 2995 { 2996 struct reiserfs_journal *journal = SB_JOURNAL(sb); 2997 unsigned long bcount = journal->j_bcount; 2998 while (1) { 2999 int depth; 3000 3001 depth = reiserfs_write_unlock_nested(sb); 3002 schedule_timeout_uninterruptible(1); 3003 reiserfs_write_lock_nested(sb, depth); 3004 3005 journal->j_current_jl->j_state |= LIST_COMMIT_PENDING; 3006 while ((atomic_read(&journal->j_wcount) > 0 || 3007 atomic_read(&journal->j_jlock)) && 3008 journal->j_trans_id == trans_id) { 3009 queue_log_writer(sb); 3010 } 3011 if (journal->j_trans_id != trans_id) 3012 break; 3013 if (bcount == journal->j_bcount) 3014 break; 3015 bcount = journal->j_bcount; 3016 } 3017 } 3018 3019 /* 3020 * join == true if you must join an existing transaction. 3021 * join == false if you can deal with waiting for others to finish 3022 * 3023 * this will block until the transaction is joinable. send the number of 3024 * blocks you expect to use in nblocks. 3025 */ 3026 static int do_journal_begin_r(struct reiserfs_transaction_handle *th, 3027 struct super_block *sb, unsigned long nblocks, 3028 int join) 3029 { 3030 time64_t now = ktime_get_seconds(); 3031 unsigned int old_trans_id; 3032 struct reiserfs_journal *journal = SB_JOURNAL(sb); 3033 struct reiserfs_transaction_handle myth; 3034 int retval; 3035 int depth; 3036 3037 reiserfs_check_lock_depth(sb, "journal_begin"); 3038 BUG_ON(nblocks > journal->j_trans_max); 3039 3040 PROC_INFO_INC(sb, journal.journal_being); 3041 /* set here for journal_join */ 3042 th->t_refcount = 1; 3043 th->t_super = sb; 3044 3045 relock: 3046 lock_journal(sb); 3047 if (join != JBEGIN_ABORT && reiserfs_is_journal_aborted(journal)) { 3048 unlock_journal(sb); 3049 retval = journal->j_errno; 3050 goto out_fail; 3051 } 3052 journal->j_bcount++; 3053 3054 if (test_bit(J_WRITERS_BLOCKED, &journal->j_state)) { 3055 unlock_journal(sb); 3056 depth = reiserfs_write_unlock_nested(sb); 3057 reiserfs_wait_on_write_block(sb); 3058 reiserfs_write_lock_nested(sb, depth); 3059 PROC_INFO_INC(sb, journal.journal_relock_writers); 3060 goto relock; 3061 } 3062 now = ktime_get_seconds(); 3063 3064 /* 3065 * if there is no room in the journal OR 3066 * if this transaction is too old, and we weren't called joinable, 3067 * wait for it to finish before beginning we don't sleep if there 3068 * aren't other writers 3069 */ 3070 3071 if ((!join && journal->j_must_wait > 0) || 3072 (!join 3073 && (journal->j_len_alloc + nblocks + 2) >= journal->j_max_batch) 3074 || (!join && atomic_read(&journal->j_wcount) > 0 3075 && journal->j_trans_start_time > 0 3076 && (now - journal->j_trans_start_time) > 3077 journal->j_max_trans_age) || (!join 3078 && atomic_read(&journal->j_jlock)) 3079 || (!join && journal->j_cnode_free < (journal->j_trans_max * 3))) { 3080 3081 old_trans_id = journal->j_trans_id; 3082 /* allow others to finish this transaction */ 3083 unlock_journal(sb); 3084 3085 if (!join && (journal->j_len_alloc + nblocks + 2) >= 3086 journal->j_max_batch && 3087 ((journal->j_len + nblocks + 2) * 100) < 3088 (journal->j_len_alloc * 75)) { 3089 if (atomic_read(&journal->j_wcount) > 10) { 3090 queue_log_writer(sb); 3091 goto relock; 3092 } 3093 } 3094 /* 3095 * don't mess with joining the transaction if all we 3096 * have to do is wait for someone else to do a commit 3097 */ 3098 if (atomic_read(&journal->j_jlock)) { 3099 while (journal->j_trans_id == old_trans_id && 3100 atomic_read(&journal->j_jlock)) { 3101 queue_log_writer(sb); 3102 } 3103 goto relock; 3104 } 3105 retval = journal_join(&myth, sb); 3106 if (retval) 3107 goto out_fail; 3108 3109 /* someone might have ended the transaction while we joined */ 3110 if (old_trans_id != journal->j_trans_id) { 3111 retval = do_journal_end(&myth, 0); 3112 } else { 3113 retval = do_journal_end(&myth, COMMIT_NOW); 3114 } 3115 3116 if (retval) 3117 goto out_fail; 3118 3119 PROC_INFO_INC(sb, journal.journal_relock_wcount); 3120 goto relock; 3121 } 3122 /* we are the first writer, set trans_id */ 3123 if (journal->j_trans_start_time == 0) { 3124 journal->j_trans_start_time = ktime_get_seconds(); 3125 } 3126 atomic_inc(&journal->j_wcount); 3127 journal->j_len_alloc += nblocks; 3128 th->t_blocks_logged = 0; 3129 th->t_blocks_allocated = nblocks; 3130 th->t_trans_id = journal->j_trans_id; 3131 unlock_journal(sb); 3132 INIT_LIST_HEAD(&th->t_list); 3133 return 0; 3134 3135 out_fail: 3136 memset(th, 0, sizeof(*th)); 3137 /* 3138 * Re-set th->t_super, so we can properly keep track of how many 3139 * persistent transactions there are. We need to do this so if this 3140 * call is part of a failed restart_transaction, we can free it later 3141 */ 3142 th->t_super = sb; 3143 return retval; 3144 } 3145 3146 struct reiserfs_transaction_handle *reiserfs_persistent_transaction(struct 3147 super_block 3148 *s, 3149 int nblocks) 3150 { 3151 int ret; 3152 struct reiserfs_transaction_handle *th; 3153 3154 /* 3155 * if we're nesting into an existing transaction. It will be 3156 * persistent on its own 3157 */ 3158 if (reiserfs_transaction_running(s)) { 3159 th = current->journal_info; 3160 th->t_refcount++; 3161 BUG_ON(th->t_refcount < 2); 3162 3163 return th; 3164 } 3165 th = kmalloc(sizeof(struct reiserfs_transaction_handle), GFP_NOFS); 3166 if (!th) 3167 return NULL; 3168 ret = journal_begin(th, s, nblocks); 3169 if (ret) { 3170 kfree(th); 3171 return NULL; 3172 } 3173 3174 SB_JOURNAL(s)->j_persistent_trans++; 3175 return th; 3176 } 3177 3178 int reiserfs_end_persistent_transaction(struct reiserfs_transaction_handle *th) 3179 { 3180 struct super_block *s = th->t_super; 3181 int ret = 0; 3182 if (th->t_trans_id) 3183 ret = journal_end(th); 3184 else 3185 ret = -EIO; 3186 if (th->t_refcount == 0) { 3187 SB_JOURNAL(s)->j_persistent_trans--; 3188 kfree(th); 3189 } 3190 return ret; 3191 } 3192 3193 static int journal_join(struct reiserfs_transaction_handle *th, 3194 struct super_block *sb) 3195 { 3196 struct reiserfs_transaction_handle *cur_th = current->journal_info; 3197 3198 /* 3199 * this keeps do_journal_end from NULLing out the 3200 * current->journal_info pointer 3201 */ 3202 th->t_handle_save = cur_th; 3203 BUG_ON(cur_th && cur_th->t_refcount > 1); 3204 return do_journal_begin_r(th, sb, 1, JBEGIN_JOIN); 3205 } 3206 3207 int journal_join_abort(struct reiserfs_transaction_handle *th, 3208 struct super_block *sb) 3209 { 3210 struct reiserfs_transaction_handle *cur_th = current->journal_info; 3211 3212 /* 3213 * this keeps do_journal_end from NULLing out the 3214 * current->journal_info pointer 3215 */ 3216 th->t_handle_save = cur_th; 3217 BUG_ON(cur_th && cur_th->t_refcount > 1); 3218 return do_journal_begin_r(th, sb, 1, JBEGIN_ABORT); 3219 } 3220 3221 int journal_begin(struct reiserfs_transaction_handle *th, 3222 struct super_block *sb, unsigned long nblocks) 3223 { 3224 struct reiserfs_transaction_handle *cur_th = current->journal_info; 3225 int ret; 3226 3227 th->t_handle_save = NULL; 3228 if (cur_th) { 3229 /* we are nesting into the current transaction */ 3230 if (cur_th->t_super == sb) { 3231 BUG_ON(!cur_th->t_refcount); 3232 cur_th->t_refcount++; 3233 memcpy(th, cur_th, sizeof(*th)); 3234 if (th->t_refcount <= 1) 3235 reiserfs_warning(sb, "reiserfs-2005", 3236 "BAD: refcount <= 1, but " 3237 "journal_info != 0"); 3238 return 0; 3239 } else { 3240 /* 3241 * we've ended up with a handle from a different 3242 * filesystem. save it and restore on journal_end. 3243 * This should never really happen... 3244 */ 3245 reiserfs_warning(sb, "clm-2100", 3246 "nesting info a different FS"); 3247 th->t_handle_save = current->journal_info; 3248 current->journal_info = th; 3249 } 3250 } else { 3251 current->journal_info = th; 3252 } 3253 ret = do_journal_begin_r(th, sb, nblocks, JBEGIN_REG); 3254 BUG_ON(current->journal_info != th); 3255 3256 /* 3257 * I guess this boils down to being the reciprocal of clm-2100 above. 3258 * If do_journal_begin_r fails, we need to put it back, since 3259 * journal_end won't be called to do it. */ 3260 if (ret) 3261 current->journal_info = th->t_handle_save; 3262 else 3263 BUG_ON(!th->t_refcount); 3264 3265 return ret; 3266 } 3267 3268 /* 3269 * puts bh into the current transaction. If it was already there, reorders 3270 * removes the old pointers from the hash, and puts new ones in (to make 3271 * sure replay happen in the right order). 3272 * 3273 * if it was dirty, cleans and files onto the clean list. I can't let it 3274 * be dirty again until the transaction is committed. 3275 * 3276 * if j_len, is bigger than j_len_alloc, it pushes j_len_alloc to 10 + j_len. 3277 */ 3278 int journal_mark_dirty(struct reiserfs_transaction_handle *th, 3279 struct buffer_head *bh) 3280 { 3281 struct super_block *sb = th->t_super; 3282 struct reiserfs_journal *journal = SB_JOURNAL(sb); 3283 struct reiserfs_journal_cnode *cn = NULL; 3284 int count_already_incd = 0; 3285 int prepared = 0; 3286 BUG_ON(!th->t_trans_id); 3287 3288 PROC_INFO_INC(sb, journal.mark_dirty); 3289 if (th->t_trans_id != journal->j_trans_id) { 3290 reiserfs_panic(th->t_super, "journal-1577", 3291 "handle trans id %ld != current trans id %ld", 3292 th->t_trans_id, journal->j_trans_id); 3293 } 3294 3295 prepared = test_clear_buffer_journal_prepared(bh); 3296 clear_buffer_journal_restore_dirty(bh); 3297 /* already in this transaction, we are done */ 3298 if (buffer_journaled(bh)) { 3299 PROC_INFO_INC(sb, journal.mark_dirty_already); 3300 return 0; 3301 } 3302 3303 /* 3304 * this must be turned into a panic instead of a warning. We can't 3305 * allow a dirty or journal_dirty or locked buffer to be logged, as 3306 * some changes could get to disk too early. NOT GOOD. 3307 */ 3308 if (!prepared || buffer_dirty(bh)) { 3309 reiserfs_warning(sb, "journal-1777", 3310 "buffer %llu bad state " 3311 "%cPREPARED %cLOCKED %cDIRTY %cJDIRTY_WAIT", 3312 (unsigned long long)bh->b_blocknr, 3313 prepared ? ' ' : '!', 3314 buffer_locked(bh) ? ' ' : '!', 3315 buffer_dirty(bh) ? ' ' : '!', 3316 buffer_journal_dirty(bh) ? ' ' : '!'); 3317 } 3318 3319 if (atomic_read(&journal->j_wcount) <= 0) { 3320 reiserfs_warning(sb, "journal-1409", 3321 "returning because j_wcount was %d", 3322 atomic_read(&journal->j_wcount)); 3323 return 1; 3324 } 3325 /* 3326 * this error means I've screwed up, and we've overflowed 3327 * the transaction. Nothing can be done here, except make the 3328 * FS readonly or panic. 3329 */ 3330 if (journal->j_len >= journal->j_trans_max) { 3331 reiserfs_panic(th->t_super, "journal-1413", 3332 "j_len (%lu) is too big", 3333 journal->j_len); 3334 } 3335 3336 if (buffer_journal_dirty(bh)) { 3337 count_already_incd = 1; 3338 PROC_INFO_INC(sb, journal.mark_dirty_notjournal); 3339 clear_buffer_journal_dirty(bh); 3340 } 3341 3342 if (journal->j_len > journal->j_len_alloc) { 3343 journal->j_len_alloc = journal->j_len + JOURNAL_PER_BALANCE_CNT; 3344 } 3345 3346 set_buffer_journaled(bh); 3347 3348 /* now put this guy on the end */ 3349 if (!cn) { 3350 cn = get_cnode(sb); 3351 if (!cn) { 3352 reiserfs_panic(sb, "journal-4", "get_cnode failed!"); 3353 } 3354 3355 if (th->t_blocks_logged == th->t_blocks_allocated) { 3356 th->t_blocks_allocated += JOURNAL_PER_BALANCE_CNT; 3357 journal->j_len_alloc += JOURNAL_PER_BALANCE_CNT; 3358 } 3359 th->t_blocks_logged++; 3360 journal->j_len++; 3361 3362 cn->bh = bh; 3363 cn->blocknr = bh->b_blocknr; 3364 cn->sb = sb; 3365 cn->jlist = NULL; 3366 insert_journal_hash(journal->j_hash_table, cn); 3367 if (!count_already_incd) { 3368 get_bh(bh); 3369 } 3370 } 3371 cn->next = NULL; 3372 cn->prev = journal->j_last; 3373 cn->bh = bh; 3374 if (journal->j_last) { 3375 journal->j_last->next = cn; 3376 journal->j_last = cn; 3377 } else { 3378 journal->j_first = cn; 3379 journal->j_last = cn; 3380 } 3381 reiserfs_schedule_old_flush(sb); 3382 return 0; 3383 } 3384 3385 int journal_end(struct reiserfs_transaction_handle *th) 3386 { 3387 struct super_block *sb = th->t_super; 3388 if (!current->journal_info && th->t_refcount > 1) 3389 reiserfs_warning(sb, "REISER-NESTING", 3390 "th NULL, refcount %d", th->t_refcount); 3391 3392 if (!th->t_trans_id) { 3393 WARN_ON(1); 3394 return -EIO; 3395 } 3396 3397 th->t_refcount--; 3398 if (th->t_refcount > 0) { 3399 struct reiserfs_transaction_handle *cur_th = 3400 current->journal_info; 3401 3402 /* 3403 * we aren't allowed to close a nested transaction on a 3404 * different filesystem from the one in the task struct 3405 */ 3406 BUG_ON(cur_th->t_super != th->t_super); 3407 3408 if (th != cur_th) { 3409 memcpy(current->journal_info, th, sizeof(*th)); 3410 th->t_trans_id = 0; 3411 } 3412 return 0; 3413 } else { 3414 return do_journal_end(th, 0); 3415 } 3416 } 3417 3418 /* 3419 * removes from the current transaction, relsing and descrementing any counters. 3420 * also files the removed buffer directly onto the clean list 3421 * 3422 * called by journal_mark_freed when a block has been deleted 3423 * 3424 * returns 1 if it cleaned and relsed the buffer. 0 otherwise 3425 */ 3426 static int remove_from_transaction(struct super_block *sb, 3427 b_blocknr_t blocknr, int already_cleaned) 3428 { 3429 struct buffer_head *bh; 3430 struct reiserfs_journal_cnode *cn; 3431 struct reiserfs_journal *journal = SB_JOURNAL(sb); 3432 int ret = 0; 3433 3434 cn = get_journal_hash_dev(sb, journal->j_hash_table, blocknr); 3435 if (!cn || !cn->bh) { 3436 return ret; 3437 } 3438 bh = cn->bh; 3439 if (cn->prev) { 3440 cn->prev->next = cn->next; 3441 } 3442 if (cn->next) { 3443 cn->next->prev = cn->prev; 3444 } 3445 if (cn == journal->j_first) { 3446 journal->j_first = cn->next; 3447 } 3448 if (cn == journal->j_last) { 3449 journal->j_last = cn->prev; 3450 } 3451 remove_journal_hash(sb, journal->j_hash_table, NULL, 3452 bh->b_blocknr, 0); 3453 clear_buffer_journaled(bh); /* don't log this one */ 3454 3455 if (!already_cleaned) { 3456 clear_buffer_journal_dirty(bh); 3457 clear_buffer_dirty(bh); 3458 clear_buffer_journal_test(bh); 3459 put_bh(bh); 3460 if (atomic_read(&bh->b_count) < 0) { 3461 reiserfs_warning(sb, "journal-1752", 3462 "b_count < 0"); 3463 } 3464 ret = 1; 3465 } 3466 journal->j_len--; 3467 journal->j_len_alloc--; 3468 free_cnode(sb, cn); 3469 return ret; 3470 } 3471 3472 /* 3473 * for any cnode in a journal list, it can only be dirtied of all the 3474 * transactions that include it are committed to disk. 3475 * this checks through each transaction, and returns 1 if you are allowed 3476 * to dirty, and 0 if you aren't 3477 * 3478 * it is called by dirty_journal_list, which is called after 3479 * flush_commit_list has gotten all the log blocks for a given 3480 * transaction on disk 3481 * 3482 */ 3483 static int can_dirty(struct reiserfs_journal_cnode *cn) 3484 { 3485 struct super_block *sb = cn->sb; 3486 b_blocknr_t blocknr = cn->blocknr; 3487 struct reiserfs_journal_cnode *cur = cn->hprev; 3488 int can_dirty = 1; 3489 3490 /* 3491 * first test hprev. These are all newer than cn, so any node here 3492 * with the same block number and dev means this node can't be sent 3493 * to disk right now. 3494 */ 3495 while (cur && can_dirty) { 3496 if (cur->jlist && cur->bh && cur->blocknr && cur->sb == sb && 3497 cur->blocknr == blocknr) { 3498 can_dirty = 0; 3499 } 3500 cur = cur->hprev; 3501 } 3502 /* 3503 * then test hnext. These are all older than cn. As long as they 3504 * are committed to the log, it is safe to write cn to disk 3505 */ 3506 cur = cn->hnext; 3507 while (cur && can_dirty) { 3508 if (cur->jlist && cur->jlist->j_len > 0 && 3509 atomic_read(&cur->jlist->j_commit_left) > 0 && cur->bh && 3510 cur->blocknr && cur->sb == sb && cur->blocknr == blocknr) { 3511 can_dirty = 0; 3512 } 3513 cur = cur->hnext; 3514 } 3515 return can_dirty; 3516 } 3517 3518 /* 3519 * syncs the commit blocks, but does not force the real buffers to disk 3520 * will wait until the current transaction is done/committed before returning 3521 */ 3522 int journal_end_sync(struct reiserfs_transaction_handle *th) 3523 { 3524 struct super_block *sb = th->t_super; 3525 struct reiserfs_journal *journal = SB_JOURNAL(sb); 3526 3527 BUG_ON(!th->t_trans_id); 3528 /* you can sync while nested, very, very bad */ 3529 BUG_ON(th->t_refcount > 1); 3530 if (journal->j_len == 0) { 3531 reiserfs_prepare_for_journal(sb, SB_BUFFER_WITH_SB(sb), 3532 1); 3533 journal_mark_dirty(th, SB_BUFFER_WITH_SB(sb)); 3534 } 3535 return do_journal_end(th, COMMIT_NOW | WAIT); 3536 } 3537 3538 /* writeback the pending async commits to disk */ 3539 static void flush_async_commits(struct work_struct *work) 3540 { 3541 struct reiserfs_journal *journal = 3542 container_of(work, struct reiserfs_journal, j_work.work); 3543 struct super_block *sb = journal->j_work_sb; 3544 struct reiserfs_journal_list *jl; 3545 struct list_head *entry; 3546 3547 reiserfs_write_lock(sb); 3548 if (!list_empty(&journal->j_journal_list)) { 3549 /* last entry is the youngest, commit it and you get everything */ 3550 entry = journal->j_journal_list.prev; 3551 jl = JOURNAL_LIST_ENTRY(entry); 3552 flush_commit_list(sb, jl, 1); 3553 } 3554 reiserfs_write_unlock(sb); 3555 } 3556 3557 /* 3558 * flushes any old transactions to disk 3559 * ends the current transaction if it is too old 3560 */ 3561 void reiserfs_flush_old_commits(struct super_block *sb) 3562 { 3563 time64_t now; 3564 struct reiserfs_transaction_handle th; 3565 struct reiserfs_journal *journal = SB_JOURNAL(sb); 3566 3567 now = ktime_get_seconds(); 3568 /* 3569 * safety check so we don't flush while we are replaying the log during 3570 * mount 3571 */ 3572 if (list_empty(&journal->j_journal_list)) 3573 return; 3574 3575 /* 3576 * check the current transaction. If there are no writers, and it is 3577 * too old, finish it, and force the commit blocks to disk 3578 */ 3579 if (atomic_read(&journal->j_wcount) <= 0 && 3580 journal->j_trans_start_time > 0 && 3581 journal->j_len > 0 && 3582 (now - journal->j_trans_start_time) > journal->j_max_trans_age) { 3583 if (!journal_join(&th, sb)) { 3584 reiserfs_prepare_for_journal(sb, 3585 SB_BUFFER_WITH_SB(sb), 3586 1); 3587 journal_mark_dirty(&th, SB_BUFFER_WITH_SB(sb)); 3588 3589 /* 3590 * we're only being called from kreiserfsd, it makes 3591 * no sense to do an async commit so that kreiserfsd 3592 * can do it later 3593 */ 3594 do_journal_end(&th, COMMIT_NOW | WAIT); 3595 } 3596 } 3597 } 3598 3599 /* 3600 * returns 0 if do_journal_end should return right away, returns 1 if 3601 * do_journal_end should finish the commit 3602 * 3603 * if the current transaction is too old, but still has writers, this will 3604 * wait on j_join_wait until all the writers are done. By the time it 3605 * wakes up, the transaction it was called has already ended, so it just 3606 * flushes the commit list and returns 0. 3607 * 3608 * Won't batch when flush or commit_now is set. Also won't batch when 3609 * others are waiting on j_join_wait. 3610 * 3611 * Note, we can't allow the journal_end to proceed while there are still 3612 * writers in the log. 3613 */ 3614 static int check_journal_end(struct reiserfs_transaction_handle *th, int flags) 3615 { 3616 3617 time64_t now; 3618 int flush = flags & FLUSH_ALL; 3619 int commit_now = flags & COMMIT_NOW; 3620 int wait_on_commit = flags & WAIT; 3621 struct reiserfs_journal_list *jl; 3622 struct super_block *sb = th->t_super; 3623 struct reiserfs_journal *journal = SB_JOURNAL(sb); 3624 3625 BUG_ON(!th->t_trans_id); 3626 3627 if (th->t_trans_id != journal->j_trans_id) { 3628 reiserfs_panic(th->t_super, "journal-1577", 3629 "handle trans id %ld != current trans id %ld", 3630 th->t_trans_id, journal->j_trans_id); 3631 } 3632 3633 journal->j_len_alloc -= (th->t_blocks_allocated - th->t_blocks_logged); 3634 /* <= 0 is allowed. unmounting might not call begin */ 3635 if (atomic_read(&journal->j_wcount) > 0) 3636 atomic_dec(&journal->j_wcount); 3637 3638 /* 3639 * BUG, deal with case where j_len is 0, but people previously 3640 * freed blocks need to be released will be dealt with by next 3641 * transaction that actually writes something, but should be taken 3642 * care of in this trans 3643 */ 3644 BUG_ON(journal->j_len == 0); 3645 3646 /* 3647 * if wcount > 0, and we are called to with flush or commit_now, 3648 * we wait on j_join_wait. We will wake up when the last writer has 3649 * finished the transaction, and started it on its way to the disk. 3650 * Then, we flush the commit or journal list, and just return 0 3651 * because the rest of journal end was already done for this 3652 * transaction. 3653 */ 3654 if (atomic_read(&journal->j_wcount) > 0) { 3655 if (flush || commit_now) { 3656 unsigned trans_id; 3657 3658 jl = journal->j_current_jl; 3659 trans_id = jl->j_trans_id; 3660 if (wait_on_commit) 3661 jl->j_state |= LIST_COMMIT_PENDING; 3662 atomic_set(&journal->j_jlock, 1); 3663 if (flush) { 3664 journal->j_next_full_flush = 1; 3665 } 3666 unlock_journal(sb); 3667 3668 /* 3669 * sleep while the current transaction is 3670 * still j_jlocked 3671 */ 3672 while (journal->j_trans_id == trans_id) { 3673 if (atomic_read(&journal->j_jlock)) { 3674 queue_log_writer(sb); 3675 } else { 3676 lock_journal(sb); 3677 if (journal->j_trans_id == trans_id) { 3678 atomic_set(&journal->j_jlock, 3679 1); 3680 } 3681 unlock_journal(sb); 3682 } 3683 } 3684 BUG_ON(journal->j_trans_id == trans_id); 3685 3686 if (commit_now 3687 && journal_list_still_alive(sb, trans_id) 3688 && wait_on_commit) { 3689 flush_commit_list(sb, jl, 1); 3690 } 3691 return 0; 3692 } 3693 unlock_journal(sb); 3694 return 0; 3695 } 3696 3697 /* deal with old transactions where we are the last writers */ 3698 now = ktime_get_seconds(); 3699 if ((now - journal->j_trans_start_time) > journal->j_max_trans_age) { 3700 commit_now = 1; 3701 journal->j_next_async_flush = 1; 3702 } 3703 /* don't batch when someone is waiting on j_join_wait */ 3704 /* don't batch when syncing the commit or flushing the whole trans */ 3705 if (!(journal->j_must_wait > 0) && !(atomic_read(&journal->j_jlock)) 3706 && !flush && !commit_now && (journal->j_len < journal->j_max_batch) 3707 && journal->j_len_alloc < journal->j_max_batch 3708 && journal->j_cnode_free > (journal->j_trans_max * 3)) { 3709 journal->j_bcount++; 3710 unlock_journal(sb); 3711 return 0; 3712 } 3713 3714 if (journal->j_start > SB_ONDISK_JOURNAL_SIZE(sb)) { 3715 reiserfs_panic(sb, "journal-003", 3716 "j_start (%ld) is too high", 3717 journal->j_start); 3718 } 3719 return 1; 3720 } 3721 3722 /* 3723 * Does all the work that makes deleting blocks safe. 3724 * when deleting a block mark BH_JNew, just remove it from the current 3725 * transaction, clean it's buffer_head and move on. 3726 * 3727 * otherwise: 3728 * set a bit for the block in the journal bitmap. That will prevent it from 3729 * being allocated for unformatted nodes before this transaction has finished. 3730 * 3731 * mark any cnodes for this block as BLOCK_FREED, and clear their bh pointers. 3732 * That will prevent any old transactions with this block from trying to flush 3733 * to the real location. Since we aren't removing the cnode from the 3734 * journal_list_hash, *the block can't be reallocated yet. 3735 * 3736 * Then remove it from the current transaction, decrementing any counters and 3737 * filing it on the clean list. 3738 */ 3739 int journal_mark_freed(struct reiserfs_transaction_handle *th, 3740 struct super_block *sb, b_blocknr_t blocknr) 3741 { 3742 struct reiserfs_journal *journal = SB_JOURNAL(sb); 3743 struct reiserfs_journal_cnode *cn = NULL; 3744 struct buffer_head *bh = NULL; 3745 struct reiserfs_list_bitmap *jb = NULL; 3746 int cleaned = 0; 3747 BUG_ON(!th->t_trans_id); 3748 3749 cn = get_journal_hash_dev(sb, journal->j_hash_table, blocknr); 3750 if (cn && cn->bh) { 3751 bh = cn->bh; 3752 get_bh(bh); 3753 } 3754 /* if it is journal new, we just remove it from this transaction */ 3755 if (bh && buffer_journal_new(bh)) { 3756 clear_buffer_journal_new(bh); 3757 clear_prepared_bits(bh); 3758 reiserfs_clean_and_file_buffer(bh); 3759 cleaned = remove_from_transaction(sb, blocknr, cleaned); 3760 } else { 3761 /* 3762 * set the bit for this block in the journal bitmap 3763 * for this transaction 3764 */ 3765 jb = journal->j_current_jl->j_list_bitmap; 3766 if (!jb) { 3767 reiserfs_panic(sb, "journal-1702", 3768 "journal_list_bitmap is NULL"); 3769 } 3770 set_bit_in_list_bitmap(sb, blocknr, jb); 3771 3772 /* Note, the entire while loop is not allowed to schedule. */ 3773 3774 if (bh) { 3775 clear_prepared_bits(bh); 3776 reiserfs_clean_and_file_buffer(bh); 3777 } 3778 cleaned = remove_from_transaction(sb, blocknr, cleaned); 3779 3780 /* 3781 * find all older transactions with this block, 3782 * make sure they don't try to write it out 3783 */ 3784 cn = get_journal_hash_dev(sb, journal->j_list_hash_table, 3785 blocknr); 3786 while (cn) { 3787 if (sb == cn->sb && blocknr == cn->blocknr) { 3788 set_bit(BLOCK_FREED, &cn->state); 3789 if (cn->bh) { 3790 /* 3791 * remove_from_transaction will brelse 3792 * the buffer if it was in the current 3793 * trans 3794 */ 3795 if (!cleaned) { 3796 clear_buffer_journal_dirty(cn-> 3797 bh); 3798 clear_buffer_dirty(cn->bh); 3799 clear_buffer_journal_test(cn-> 3800 bh); 3801 cleaned = 1; 3802 put_bh(cn->bh); 3803 if (atomic_read 3804 (&cn->bh->b_count) < 0) { 3805 reiserfs_warning(sb, 3806 "journal-2138", 3807 "cn->bh->b_count < 0"); 3808 } 3809 } 3810 /* 3811 * since we are clearing the bh, 3812 * we MUST dec nonzerolen 3813 */ 3814 if (cn->jlist) { 3815 atomic_dec(&cn->jlist-> 3816 j_nonzerolen); 3817 } 3818 cn->bh = NULL; 3819 } 3820 } 3821 cn = cn->hnext; 3822 } 3823 } 3824 3825 if (bh) 3826 release_buffer_page(bh); /* get_hash grabs the buffer */ 3827 return 0; 3828 } 3829 3830 void reiserfs_update_inode_transaction(struct inode *inode) 3831 { 3832 struct reiserfs_journal *journal = SB_JOURNAL(inode->i_sb); 3833 REISERFS_I(inode)->i_jl = journal->j_current_jl; 3834 REISERFS_I(inode)->i_trans_id = journal->j_trans_id; 3835 } 3836 3837 /* 3838 * returns -1 on error, 0 if no commits/barriers were done and 1 3839 * if a transaction was actually committed and the barrier was done 3840 */ 3841 static int __commit_trans_jl(struct inode *inode, unsigned long id, 3842 struct reiserfs_journal_list *jl) 3843 { 3844 struct reiserfs_transaction_handle th; 3845 struct super_block *sb = inode->i_sb; 3846 struct reiserfs_journal *journal = SB_JOURNAL(sb); 3847 int ret = 0; 3848 3849 /* 3850 * is it from the current transaction, 3851 * or from an unknown transaction? 3852 */ 3853 if (id == journal->j_trans_id) { 3854 jl = journal->j_current_jl; 3855 /* 3856 * try to let other writers come in and 3857 * grow this transaction 3858 */ 3859 let_transaction_grow(sb, id); 3860 if (journal->j_trans_id != id) { 3861 goto flush_commit_only; 3862 } 3863 3864 ret = journal_begin(&th, sb, 1); 3865 if (ret) 3866 return ret; 3867 3868 /* someone might have ended this transaction while we joined */ 3869 if (journal->j_trans_id != id) { 3870 reiserfs_prepare_for_journal(sb, SB_BUFFER_WITH_SB(sb), 3871 1); 3872 journal_mark_dirty(&th, SB_BUFFER_WITH_SB(sb)); 3873 ret = journal_end(&th); 3874 goto flush_commit_only; 3875 } 3876 3877 ret = journal_end_sync(&th); 3878 if (!ret) 3879 ret = 1; 3880 3881 } else { 3882 /* 3883 * this gets tricky, we have to make sure the journal list in 3884 * the inode still exists. We know the list is still around 3885 * if we've got a larger transaction id than the oldest list 3886 */ 3887 flush_commit_only: 3888 if (journal_list_still_alive(inode->i_sb, id)) { 3889 /* 3890 * we only set ret to 1 when we know for sure 3891 * the barrier hasn't been started yet on the commit 3892 * block. 3893 */ 3894 if (atomic_read(&jl->j_commit_left) > 1) 3895 ret = 1; 3896 flush_commit_list(sb, jl, 1); 3897 if (journal->j_errno) 3898 ret = journal->j_errno; 3899 } 3900 } 3901 /* otherwise the list is gone, and long since committed */ 3902 return ret; 3903 } 3904 3905 int reiserfs_commit_for_inode(struct inode *inode) 3906 { 3907 unsigned int id = REISERFS_I(inode)->i_trans_id; 3908 struct reiserfs_journal_list *jl = REISERFS_I(inode)->i_jl; 3909 3910 /* 3911 * for the whole inode, assume unset id means it was 3912 * changed in the current transaction. More conservative 3913 */ 3914 if (!id || !jl) { 3915 reiserfs_update_inode_transaction(inode); 3916 id = REISERFS_I(inode)->i_trans_id; 3917 /* jl will be updated in __commit_trans_jl */ 3918 } 3919 3920 return __commit_trans_jl(inode, id, jl); 3921 } 3922 3923 void reiserfs_restore_prepared_buffer(struct super_block *sb, 3924 struct buffer_head *bh) 3925 { 3926 struct reiserfs_journal *journal = SB_JOURNAL(sb); 3927 PROC_INFO_INC(sb, journal.restore_prepared); 3928 if (!bh) { 3929 return; 3930 } 3931 if (test_clear_buffer_journal_restore_dirty(bh) && 3932 buffer_journal_dirty(bh)) { 3933 struct reiserfs_journal_cnode *cn; 3934 reiserfs_write_lock(sb); 3935 cn = get_journal_hash_dev(sb, 3936 journal->j_list_hash_table, 3937 bh->b_blocknr); 3938 if (cn && can_dirty(cn)) { 3939 set_buffer_journal_test(bh); 3940 mark_buffer_dirty(bh); 3941 } 3942 reiserfs_write_unlock(sb); 3943 } 3944 clear_buffer_journal_prepared(bh); 3945 } 3946 3947 extern struct tree_balance *cur_tb; 3948 /* 3949 * before we can change a metadata block, we have to make sure it won't 3950 * be written to disk while we are altering it. So, we must: 3951 * clean it 3952 * wait on it. 3953 */ 3954 int reiserfs_prepare_for_journal(struct super_block *sb, 3955 struct buffer_head *bh, int wait) 3956 { 3957 PROC_INFO_INC(sb, journal.prepare); 3958 3959 if (!trylock_buffer(bh)) { 3960 if (!wait) 3961 return 0; 3962 lock_buffer(bh); 3963 } 3964 set_buffer_journal_prepared(bh); 3965 if (test_clear_buffer_dirty(bh) && buffer_journal_dirty(bh)) { 3966 clear_buffer_journal_test(bh); 3967 set_buffer_journal_restore_dirty(bh); 3968 } 3969 unlock_buffer(bh); 3970 return 1; 3971 } 3972 3973 /* 3974 * long and ugly. If flush, will not return until all commit 3975 * blocks and all real buffers in the trans are on disk. 3976 * If no_async, won't return until all commit blocks are on disk. 3977 * 3978 * keep reading, there are comments as you go along 3979 * 3980 * If the journal is aborted, we just clean up. Things like flushing 3981 * journal lists, etc just won't happen. 3982 */ 3983 static int do_journal_end(struct reiserfs_transaction_handle *th, int flags) 3984 { 3985 struct super_block *sb = th->t_super; 3986 struct reiserfs_journal *journal = SB_JOURNAL(sb); 3987 struct reiserfs_journal_cnode *cn, *next, *jl_cn; 3988 struct reiserfs_journal_cnode *last_cn = NULL; 3989 struct reiserfs_journal_desc *desc; 3990 struct reiserfs_journal_commit *commit; 3991 struct buffer_head *c_bh; /* commit bh */ 3992 struct buffer_head *d_bh; /* desc bh */ 3993 int cur_write_start = 0; /* start index of current log write */ 3994 int i; 3995 int flush; 3996 int wait_on_commit; 3997 struct reiserfs_journal_list *jl, *temp_jl; 3998 struct list_head *entry, *safe; 3999 unsigned long jindex; 4000 unsigned int commit_trans_id; 4001 int trans_half; 4002 int depth; 4003 4004 BUG_ON(th->t_refcount > 1); 4005 BUG_ON(!th->t_trans_id); 4006 BUG_ON(!th->t_super); 4007 4008 /* 4009 * protect flush_older_commits from doing mistakes if the 4010 * transaction ID counter gets overflowed. 4011 */ 4012 if (th->t_trans_id == ~0U) 4013 flags |= FLUSH_ALL | COMMIT_NOW | WAIT; 4014 flush = flags & FLUSH_ALL; 4015 wait_on_commit = flags & WAIT; 4016 4017 current->journal_info = th->t_handle_save; 4018 reiserfs_check_lock_depth(sb, "journal end"); 4019 if (journal->j_len == 0) { 4020 reiserfs_prepare_for_journal(sb, SB_BUFFER_WITH_SB(sb), 4021 1); 4022 journal_mark_dirty(th, SB_BUFFER_WITH_SB(sb)); 4023 } 4024 4025 lock_journal(sb); 4026 if (journal->j_next_full_flush) { 4027 flags |= FLUSH_ALL; 4028 flush = 1; 4029 } 4030 if (journal->j_next_async_flush) { 4031 flags |= COMMIT_NOW | WAIT; 4032 wait_on_commit = 1; 4033 } 4034 4035 /* 4036 * check_journal_end locks the journal, and unlocks if it does 4037 * not return 1 it tells us if we should continue with the 4038 * journal_end, or just return 4039 */ 4040 if (!check_journal_end(th, flags)) { 4041 reiserfs_schedule_old_flush(sb); 4042 wake_queued_writers(sb); 4043 reiserfs_async_progress_wait(sb); 4044 goto out; 4045 } 4046 4047 /* check_journal_end might set these, check again */ 4048 if (journal->j_next_full_flush) { 4049 flush = 1; 4050 } 4051 4052 /* 4053 * j must wait means we have to flush the log blocks, and the 4054 * real blocks for this transaction 4055 */ 4056 if (journal->j_must_wait > 0) { 4057 flush = 1; 4058 } 4059 #ifdef REISERFS_PREALLOCATE 4060 /* 4061 * quota ops might need to nest, setup the journal_info pointer 4062 * for them and raise the refcount so that it is > 0. 4063 */ 4064 current->journal_info = th; 4065 th->t_refcount++; 4066 4067 /* it should not involve new blocks into the transaction */ 4068 reiserfs_discard_all_prealloc(th); 4069 4070 th->t_refcount--; 4071 current->journal_info = th->t_handle_save; 4072 #endif 4073 4074 /* setup description block */ 4075 d_bh = 4076 journal_getblk(sb, 4077 SB_ONDISK_JOURNAL_1st_BLOCK(sb) + 4078 journal->j_start); 4079 set_buffer_uptodate(d_bh); 4080 desc = (struct reiserfs_journal_desc *)(d_bh)->b_data; 4081 memset(d_bh->b_data, 0, d_bh->b_size); 4082 memcpy(get_journal_desc_magic(d_bh), JOURNAL_DESC_MAGIC, 8); 4083 set_desc_trans_id(desc, journal->j_trans_id); 4084 4085 /* 4086 * setup commit block. Don't write (keep it clean too) this one 4087 * until after everyone else is written 4088 */ 4089 c_bh = journal_getblk(sb, SB_ONDISK_JOURNAL_1st_BLOCK(sb) + 4090 ((journal->j_start + journal->j_len + 4091 1) % SB_ONDISK_JOURNAL_SIZE(sb))); 4092 commit = (struct reiserfs_journal_commit *)c_bh->b_data; 4093 memset(c_bh->b_data, 0, c_bh->b_size); 4094 set_commit_trans_id(commit, journal->j_trans_id); 4095 set_buffer_uptodate(c_bh); 4096 4097 /* init this journal list */ 4098 jl = journal->j_current_jl; 4099 4100 /* 4101 * we lock the commit before doing anything because 4102 * we want to make sure nobody tries to run flush_commit_list until 4103 * the new transaction is fully setup, and we've already flushed the 4104 * ordered bh list 4105 */ 4106 reiserfs_mutex_lock_safe(&jl->j_commit_mutex, sb); 4107 4108 /* save the transaction id in case we need to commit it later */ 4109 commit_trans_id = jl->j_trans_id; 4110 4111 atomic_set(&jl->j_older_commits_done, 0); 4112 jl->j_trans_id = journal->j_trans_id; 4113 jl->j_timestamp = journal->j_trans_start_time; 4114 jl->j_commit_bh = c_bh; 4115 jl->j_start = journal->j_start; 4116 jl->j_len = journal->j_len; 4117 atomic_set(&jl->j_nonzerolen, journal->j_len); 4118 atomic_set(&jl->j_commit_left, journal->j_len + 2); 4119 jl->j_realblock = NULL; 4120 4121 /* 4122 * The ENTIRE FOR LOOP MUST not cause schedule to occur. 4123 * for each real block, add it to the journal list hash, 4124 * copy into real block index array in the commit or desc block 4125 */ 4126 trans_half = journal_trans_half(sb->s_blocksize); 4127 for (i = 0, cn = journal->j_first; cn; cn = cn->next, i++) { 4128 if (buffer_journaled(cn->bh)) { 4129 jl_cn = get_cnode(sb); 4130 if (!jl_cn) { 4131 reiserfs_panic(sb, "journal-1676", 4132 "get_cnode returned NULL"); 4133 } 4134 if (i == 0) { 4135 jl->j_realblock = jl_cn; 4136 } 4137 jl_cn->prev = last_cn; 4138 jl_cn->next = NULL; 4139 if (last_cn) { 4140 last_cn->next = jl_cn; 4141 } 4142 last_cn = jl_cn; 4143 /* 4144 * make sure the block we are trying to log 4145 * is not a block of journal or reserved area 4146 */ 4147 if (is_block_in_log_or_reserved_area 4148 (sb, cn->bh->b_blocknr)) { 4149 reiserfs_panic(sb, "journal-2332", 4150 "Trying to log block %lu, " 4151 "which is a log block", 4152 cn->bh->b_blocknr); 4153 } 4154 jl_cn->blocknr = cn->bh->b_blocknr; 4155 jl_cn->state = 0; 4156 jl_cn->sb = sb; 4157 jl_cn->bh = cn->bh; 4158 jl_cn->jlist = jl; 4159 insert_journal_hash(journal->j_list_hash_table, jl_cn); 4160 if (i < trans_half) { 4161 desc->j_realblock[i] = 4162 cpu_to_le32(cn->bh->b_blocknr); 4163 } else { 4164 commit->j_realblock[i - trans_half] = 4165 cpu_to_le32(cn->bh->b_blocknr); 4166 } 4167 } else { 4168 i--; 4169 } 4170 } 4171 set_desc_trans_len(desc, journal->j_len); 4172 set_desc_mount_id(desc, journal->j_mount_id); 4173 set_desc_trans_id(desc, journal->j_trans_id); 4174 set_commit_trans_len(commit, journal->j_len); 4175 4176 /* 4177 * special check in case all buffers in the journal 4178 * were marked for not logging 4179 */ 4180 BUG_ON(journal->j_len == 0); 4181 4182 /* 4183 * we're about to dirty all the log blocks, mark the description block 4184 * dirty now too. Don't mark the commit block dirty until all the 4185 * others are on disk 4186 */ 4187 mark_buffer_dirty(d_bh); 4188 4189 /* 4190 * first data block is j_start + 1, so add one to 4191 * cur_write_start wherever you use it 4192 */ 4193 cur_write_start = journal->j_start; 4194 cn = journal->j_first; 4195 jindex = 1; /* start at one so we don't get the desc again */ 4196 while (cn) { 4197 clear_buffer_journal_new(cn->bh); 4198 /* copy all the real blocks into log area. dirty log blocks */ 4199 if (buffer_journaled(cn->bh)) { 4200 struct buffer_head *tmp_bh; 4201 char *addr; 4202 struct page *page; 4203 tmp_bh = 4204 journal_getblk(sb, 4205 SB_ONDISK_JOURNAL_1st_BLOCK(sb) + 4206 ((cur_write_start + 4207 jindex) % 4208 SB_ONDISK_JOURNAL_SIZE(sb))); 4209 set_buffer_uptodate(tmp_bh); 4210 page = cn->bh->b_page; 4211 addr = kmap(page); 4212 memcpy(tmp_bh->b_data, 4213 addr + offset_in_page(cn->bh->b_data), 4214 cn->bh->b_size); 4215 kunmap(page); 4216 mark_buffer_dirty(tmp_bh); 4217 jindex++; 4218 set_buffer_journal_dirty(cn->bh); 4219 clear_buffer_journaled(cn->bh); 4220 } else { 4221 /* 4222 * JDirty cleared sometime during transaction. 4223 * don't log this one 4224 */ 4225 reiserfs_warning(sb, "journal-2048", 4226 "BAD, buffer in journal hash, " 4227 "but not JDirty!"); 4228 brelse(cn->bh); 4229 } 4230 next = cn->next; 4231 free_cnode(sb, cn); 4232 cn = next; 4233 reiserfs_cond_resched(sb); 4234 } 4235 4236 /* 4237 * we are done with both the c_bh and d_bh, but 4238 * c_bh must be written after all other commit blocks, 4239 * so we dirty/relse c_bh in flush_commit_list, with commit_left <= 1. 4240 */ 4241 4242 journal->j_current_jl = alloc_journal_list(sb); 4243 4244 /* now it is safe to insert this transaction on the main list */ 4245 list_add_tail(&jl->j_list, &journal->j_journal_list); 4246 list_add_tail(&jl->j_working_list, &journal->j_working_list); 4247 journal->j_num_work_lists++; 4248 4249 /* reset journal values for the next transaction */ 4250 journal->j_start = 4251 (journal->j_start + journal->j_len + 4252 2) % SB_ONDISK_JOURNAL_SIZE(sb); 4253 atomic_set(&journal->j_wcount, 0); 4254 journal->j_bcount = 0; 4255 journal->j_last = NULL; 4256 journal->j_first = NULL; 4257 journal->j_len = 0; 4258 journal->j_trans_start_time = 0; 4259 /* check for trans_id overflow */ 4260 if (++journal->j_trans_id == 0) 4261 journal->j_trans_id = 10; 4262 journal->j_current_jl->j_trans_id = journal->j_trans_id; 4263 journal->j_must_wait = 0; 4264 journal->j_len_alloc = 0; 4265 journal->j_next_full_flush = 0; 4266 journal->j_next_async_flush = 0; 4267 init_journal_hash(sb); 4268 4269 /* 4270 * make sure reiserfs_add_jh sees the new current_jl before we 4271 * write out the tails 4272 */ 4273 smp_mb(); 4274 4275 /* 4276 * tail conversion targets have to hit the disk before we end the 4277 * transaction. Otherwise a later transaction might repack the tail 4278 * before this transaction commits, leaving the data block unflushed 4279 * and clean, if we crash before the later transaction commits, the 4280 * data block is lost. 4281 */ 4282 if (!list_empty(&jl->j_tail_bh_list)) { 4283 depth = reiserfs_write_unlock_nested(sb); 4284 write_ordered_buffers(&journal->j_dirty_buffers_lock, 4285 journal, jl, &jl->j_tail_bh_list); 4286 reiserfs_write_lock_nested(sb, depth); 4287 } 4288 BUG_ON(!list_empty(&jl->j_tail_bh_list)); 4289 mutex_unlock(&jl->j_commit_mutex); 4290 4291 /* 4292 * honor the flush wishes from the caller, simple commits can 4293 * be done outside the journal lock, they are done below 4294 * 4295 * if we don't flush the commit list right now, we put it into 4296 * the work queue so the people waiting on the async progress work 4297 * queue don't wait for this proc to flush journal lists and such. 4298 */ 4299 if (flush) { 4300 flush_commit_list(sb, jl, 1); 4301 flush_journal_list(sb, jl, 1); 4302 } else if (!(jl->j_state & LIST_COMMIT_PENDING)) { 4303 /* 4304 * Avoid queueing work when sb is being shut down. Transaction 4305 * will be flushed on journal shutdown. 4306 */ 4307 if (sb->s_flags & SB_ACTIVE) 4308 queue_delayed_work(REISERFS_SB(sb)->commit_wq, 4309 &journal->j_work, HZ / 10); 4310 } 4311 4312 /* 4313 * if the next transaction has any chance of wrapping, flush 4314 * transactions that might get overwritten. If any journal lists 4315 * are very old flush them as well. 4316 */ 4317 first_jl: 4318 list_for_each_safe(entry, safe, &journal->j_journal_list) { 4319 temp_jl = JOURNAL_LIST_ENTRY(entry); 4320 if (journal->j_start <= temp_jl->j_start) { 4321 if ((journal->j_start + journal->j_trans_max + 1) >= 4322 temp_jl->j_start) { 4323 flush_used_journal_lists(sb, temp_jl); 4324 goto first_jl; 4325 } else if ((journal->j_start + 4326 journal->j_trans_max + 1) < 4327 SB_ONDISK_JOURNAL_SIZE(sb)) { 4328 /* 4329 * if we don't cross into the next 4330 * transaction and we don't wrap, there is 4331 * no way we can overlap any later transactions 4332 * break now 4333 */ 4334 break; 4335 } 4336 } else if ((journal->j_start + 4337 journal->j_trans_max + 1) > 4338 SB_ONDISK_JOURNAL_SIZE(sb)) { 4339 if (((journal->j_start + journal->j_trans_max + 1) % 4340 SB_ONDISK_JOURNAL_SIZE(sb)) >= 4341 temp_jl->j_start) { 4342 flush_used_journal_lists(sb, temp_jl); 4343 goto first_jl; 4344 } else { 4345 /* 4346 * we don't overlap anything from out start 4347 * to the end of the log, and our wrapped 4348 * portion doesn't overlap anything at 4349 * the start of the log. We can break 4350 */ 4351 break; 4352 } 4353 } 4354 } 4355 4356 journal->j_current_jl->j_list_bitmap = 4357 get_list_bitmap(sb, journal->j_current_jl); 4358 4359 if (!(journal->j_current_jl->j_list_bitmap)) { 4360 reiserfs_panic(sb, "journal-1996", 4361 "could not get a list bitmap"); 4362 } 4363 4364 atomic_set(&journal->j_jlock, 0); 4365 unlock_journal(sb); 4366 /* wake up any body waiting to join. */ 4367 clear_bit(J_WRITERS_QUEUED, &journal->j_state); 4368 wake_up(&journal->j_join_wait); 4369 4370 if (!flush && wait_on_commit && 4371 journal_list_still_alive(sb, commit_trans_id)) { 4372 flush_commit_list(sb, jl, 1); 4373 } 4374 out: 4375 reiserfs_check_lock_depth(sb, "journal end2"); 4376 4377 memset(th, 0, sizeof(*th)); 4378 /* 4379 * Re-set th->t_super, so we can properly keep track of how many 4380 * persistent transactions there are. We need to do this so if this 4381 * call is part of a failed restart_transaction, we can free it later 4382 */ 4383 th->t_super = sb; 4384 4385 return journal->j_errno; 4386 } 4387 4388 /* Send the file system read only and refuse new transactions */ 4389 void reiserfs_abort_journal(struct super_block *sb, int errno) 4390 { 4391 struct reiserfs_journal *journal = SB_JOURNAL(sb); 4392 if (test_bit(J_ABORTED, &journal->j_state)) 4393 return; 4394 4395 if (!journal->j_errno) 4396 journal->j_errno = errno; 4397 4398 sb->s_flags |= SB_RDONLY; 4399 set_bit(J_ABORTED, &journal->j_state); 4400 4401 #ifdef CONFIG_REISERFS_CHECK 4402 dump_stack(); 4403 #endif 4404 } 4405