fs/gfs2/log.c

   1 /*
   2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   3  * Copyright (C) 2004-2007 Red Hat, Inc.  All rights reserved.
   4  *
   5  * This copyrighted material is made available to anyone wishing to use,
   6  * modify, copy, or redistribute it subject to the terms and conditions
   7  * of the GNU General Public License version 2.
   8  */
   9
  10 #include <linux/sched.h>
  11 #include <linux/slab.h>
  12 #include <linux/spinlock.h>
  13 #include <linux/completion.h>
  14 #include <linux/buffer_head.h>
  15 #include <linux/gfs2_ondisk.h>
  16 #include <linux/crc32.h>
  17 #include <linux/delay.h>
  18 #include <linux/kthread.h>
  19 #include <linux/freezer.h>
  20 #include <linux/bio.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/writeback.h>
  23 #include <linux/list_sort.h>
  24
  25 #include "gfs2.h"
  26 #include "incore.h"
  27 #include "bmap.h"
  28 #include "glock.h"
  29 #include "log.h"
  30 #include "lops.h"
  31 #include "meta_io.h"
  32 #include "util.h"
  33 #include "dir.h"
  34 #include "trace_gfs2.h"
  35
  36 /**
  37  * gfs2_struct2blk - compute stuff
  38  * @sdp: the filesystem
  39  * @nstruct: the number of structures
  40  * @ssize: the size of the structures
  41  *
  42  * Compute the number of log descriptor blocks needed to hold a certain number
  43  * of structures of a certain size.
  44  *
  45  * Returns: the number of blocks needed (minimum is always 1)
  46  */
  47
  48 unsigned int gfs2_struct2blk(struct gfs2_sbd *sdp, unsigned int nstruct,
  49                              unsigned int ssize)
  50 {
  51         unsigned int blks;
  52         unsigned int first, second;
  53
  54         blks = 1;
  55         first = (sdp->sd_sb.sb_bsize - sizeof(struct gfs2_log_descriptor)) / ssize;
  56
  57         if (nstruct > first) {
  58                 second = (sdp->sd_sb.sb_bsize -
  59                           sizeof(struct gfs2_meta_header)) / ssize;
  60                 blks += DIV_ROUND_UP(nstruct - first, second);
  61         }
  62
  63         return blks;
  64 }
  65
  66 /**
  67  * gfs2_remove_from_ail - Remove an entry from the ail lists, updating counters
  68  * @mapping: The associated mapping (maybe NULL)
  69  * @bd: The gfs2_bufdata to remove
  70  *
  71  * The ail lock _must_ be held when calling this function
  72  *
  73  */
  74
  75 void gfs2_remove_from_ail(struct gfs2_bufdata *bd)
  76 {
  77         bd->bd_tr = NULL;
  78         list_del_init(&bd->bd_ail_st_list);
  79         list_del_init(&bd->bd_ail_gl_list);
  80         atomic_dec(&bd->bd_gl->gl_ail_count);
  81         brelse(bd->bd_bh);
  82 }
  83
  84 /**
  85  * gfs2_ail1_start_one - Start I/O on a part of the AIL
  86  * @sdp: the filesystem
  87  * @wbc: The writeback control structure
  88  * @ai: The ail structure
  89  *
  90  */
  91
  92 static int gfs2_ail1_start_one(struct gfs2_sbd *sdp,
  93                                struct writeback_control *wbc,
  94                                struct gfs2_trans *tr)
  95 __releases(&sdp->sd_ail_lock)
  96 __acquires(&sdp->sd_ail_lock)
  97 {
  98         struct gfs2_glock *gl = NULL;
  99         struct address_space *mapping;
 100         struct gfs2_bufdata *bd, *s;
 101         struct buffer_head *bh;
 102
 103         list_for_each_entry_safe_reverse(bd, s, &tr->tr_ail1_list, bd_ail_st_list) {
 104                 bh = bd->bd_bh;
 105
 106                 gfs2_assert(sdp, bd->bd_tr == tr);
 107
 108                 if (!buffer_busy(bh)) {
 109                         if (!buffer_uptodate(bh))
 110                                 gfs2_io_error_bh(sdp, bh);
 111                         list_move(&bd->bd_ail_st_list, &tr->tr_ail2_list);
 112                         continue;
 113                 }
 114
 115                 if (!buffer_dirty(bh))
 116                         continue;
 117                 if (gl == bd->bd_gl)
 118                         continue;
 119                 gl = bd->bd_gl;
 120                 list_move(&bd->bd_ail_st_list, &tr->tr_ail1_list);
 121                 mapping = bh->b_page->mapping;
 122                 if (!mapping)
 123                         continue;
 124                 spin_unlock(&sdp->sd_ail_lock);
 125                 generic_writepages(mapping, wbc);
 126                 spin_lock(&sdp->sd_ail_lock);
 127                 if (wbc->nr_to_write <= 0)
 128                         break;
 129                 return 1;
 130         }
 131
 132         return 0;
 133 }
 134
 135
 136 /**
 137  * gfs2_ail1_flush - start writeback of some ail1 entries
 138  * @sdp: The super block
 139  * @wbc: The writeback control structure
 140  *
 141  * Writes back some ail1 entries, according to the limits in the
 142  * writeback control structure
 143  */
 144
 145 void gfs2_ail1_flush(struct gfs2_sbd *sdp, struct writeback_control *wbc)
 146 {
 147         struct list_head *head = &sdp->sd_ail1_list;
 148         struct gfs2_trans *tr;
 149         struct blk_plug plug;
 150
 151         trace_gfs2_ail_flush(sdp, wbc, 1);
 152         blk_start_plug(&plug);
 153         spin_lock(&sdp->sd_ail_lock);
 154 restart:
 155         list_for_each_entry_reverse(tr, head, tr_list) {
 156                 if (wbc->nr_to_write <= 0)
 157                         break;
 158                 if (gfs2_ail1_start_one(sdp, wbc, tr))
 159                         goto restart;
 160         }
 161         spin_unlock(&sdp->sd_ail_lock);
 162         blk_finish_plug(&plug);
 163         trace_gfs2_ail_flush(sdp, wbc, 0);
 164 }
 165
 166 /**
 167  * gfs2_ail1_start - start writeback of all ail1 entries
 168  * @sdp: The superblock
 169  */
 170
 171 static void gfs2_ail1_start(struct gfs2_sbd *sdp)
 172 {
 173         struct writeback_control wbc = {
 174                 .sync_mode = WB_SYNC_NONE,
 175                 .nr_to_write = LONG_MAX,
 176                 .range_start = 0,
 177                 .range_end = LLONG_MAX,
 178         };
 179
 180         return gfs2_ail1_flush(sdp, &wbc);
 181 }
 182
 183 /**
 184  * gfs2_ail1_empty_one - Check whether or not a trans in the AIL has been synced
 185  * @sdp: the filesystem
 186  * @ai: the AIL entry
 187  *
 188  */
 189
 190 static void gfs2_ail1_empty_one(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 191 {
 192         struct gfs2_bufdata *bd, *s;
 193         struct buffer_head *bh;
 194
 195         list_for_each_entry_safe_reverse(bd, s, &tr->tr_ail1_list,
 196                                          bd_ail_st_list) {
 197                 bh = bd->bd_bh;
 198                 gfs2_assert(sdp, bd->bd_tr == tr);
 199                 if (buffer_busy(bh))
 200                         continue;
 201                 if (!buffer_uptodate(bh))
 202                         gfs2_io_error_bh(sdp, bh);
 203                 list_move(&bd->bd_ail_st_list, &tr->tr_ail2_list);
 204         }
 205
 206 }
 207
 208 /**
 209  * gfs2_ail1_empty - Try to empty the ail1 lists
 210  * @sdp: The superblock
 211  *
 212  * Tries to empty the ail1 lists, starting with the oldest first
 213  */
 214
 215 static int gfs2_ail1_empty(struct gfs2_sbd *sdp)
 216 {
 217         struct gfs2_trans *tr, *s;
 218         int oldest_tr = 1;
 219         int ret;
 220
 221         spin_lock(&sdp->sd_ail_lock);
 222         list_for_each_entry_safe_reverse(tr, s, &sdp->sd_ail1_list, tr_list) {
 223                 gfs2_ail1_empty_one(sdp, tr);
 224                 if (list_empty(&tr->tr_ail1_list) && oldest_tr)
 225                         list_move(&tr->tr_list, &sdp->sd_ail2_list);
 226                 else
 227                         oldest_tr = 0;
 228         }
 229         ret = list_empty(&sdp->sd_ail1_list);
 230         spin_unlock(&sdp->sd_ail_lock);
 231
 232         return ret;
 233 }
 234
 235 static void gfs2_ail1_wait(struct gfs2_sbd *sdp)
 236 {
 237         struct gfs2_trans *tr;
 238         struct gfs2_bufdata *bd;
 239         struct buffer_head *bh;
 240
 241         spin_lock(&sdp->sd_ail_lock);
 242         list_for_each_entry_reverse(tr, &sdp->sd_ail1_list, tr_list) {
 243                 list_for_each_entry(bd, &tr->tr_ail1_list, bd_ail_st_list) {
 244                         bh = bd->bd_bh;
 245                         if (!buffer_locked(bh))
 246                                 continue;
 247                         get_bh(bh);
 248                         spin_unlock(&sdp->sd_ail_lock);
 249                         wait_on_buffer(bh);
 250                         brelse(bh);
 251                         return;
 252                 }
 253         }
 254         spin_unlock(&sdp->sd_ail_lock);
 255 }
 256
 257 /**
 258  * gfs2_ail2_empty_one - Check whether or not a trans in the AIL has been synced
 259  * @sdp: the filesystem
 260  * @ai: the AIL entry
 261  *
 262  */
 263
 264 static void gfs2_ail2_empty_one(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 265 {
 266         struct list_head *head = &tr->tr_ail2_list;
 267         struct gfs2_bufdata *bd;
 268
 269         while (!list_empty(head)) {
 270                 bd = list_entry(head->prev, struct gfs2_bufdata,
 271                                 bd_ail_st_list);
 272                 gfs2_assert(sdp, bd->bd_tr == tr);
 273                 gfs2_remove_from_ail(bd);
 274         }
 275 }
 276
 277 static void ail2_empty(struct gfs2_sbd *sdp, unsigned int new_tail)
 278 {
 279         struct gfs2_trans *tr, *safe;
 280         unsigned int old_tail = sdp->sd_log_tail;
 281         int wrap = (new_tail < old_tail);
 282         int a, b, rm;
 283
 284         spin_lock(&sdp->sd_ail_lock);
 285
 286         list_for_each_entry_safe(tr, safe, &sdp->sd_ail2_list, tr_list) {
 287                 a = (old_tail <= tr->tr_first);
 288                 b = (tr->tr_first < new_tail);
 289                 rm = (wrap) ? (a || b) : (a && b);
 290                 if (!rm)
 291                         continue;
 292
 293                 gfs2_ail2_empty_one(sdp, tr);
 294                 list_del(&tr->tr_list);
 295                 gfs2_assert_warn(sdp, list_empty(&tr->tr_ail1_list));
 296                 gfs2_assert_warn(sdp, list_empty(&tr->tr_ail2_list));
 297                 kfree(tr);
 298         }
 299
 300         spin_unlock(&sdp->sd_ail_lock);
 301 }
 302
 303 /**
 304  * gfs2_log_release - Release a given number of log blocks
 305  * @sdp: The GFS2 superblock
 306  * @blks: The number of blocks
 307  *
 308  */
 309
 310 void gfs2_log_release(struct gfs2_sbd *sdp, unsigned int blks)
 311 {
 312
 313         atomic_add(blks, &sdp->sd_log_blks_free);
 314         trace_gfs2_log_blocks(sdp, blks);
 315         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
 316                                   sdp->sd_jdesc->jd_blocks);
 317         up_read(&sdp->sd_log_flush_lock);
 318 }
 319
 320 /**
 321  * gfs2_log_reserve - Make a log reservation
 322  * @sdp: The GFS2 superblock
 323  * @blks: The number of blocks to reserve
 324  *
 325  * Note that we never give out the last few blocks of the journal. Thats
 326  * due to the fact that there is a small number of header blocks
 327  * associated with each log flush. The exact number can't be known until
 328  * flush time, so we ensure that we have just enough free blocks at all
 329  * times to avoid running out during a log flush.
 330  *
 331  * We no longer flush the log here, instead we wake up logd to do that
 332  * for us. To avoid the thundering herd and to ensure that we deal fairly
 333  * with queued waiters, we use an exclusive wait. This means that when we
 334  * get woken with enough journal space to get our reservation, we need to
 335  * wake the next waiter on the list.
 336  *
 337  * Returns: errno
 338  */
 339
 340 int gfs2_log_reserve(struct gfs2_sbd *sdp, unsigned int blks)
 341 {
 342         unsigned reserved_blks = 7 * (4096 / sdp->sd_vfs->s_blocksize);
 343         unsigned wanted = blks + reserved_blks;
 344         DEFINE_WAIT(wait);
 345         int did_wait = 0;
 346         unsigned int free_blocks;
 347
 348         if (gfs2_assert_warn(sdp, blks) ||
 349             gfs2_assert_warn(sdp, blks <= sdp->sd_jdesc->jd_blocks))
 350                 return -EINVAL;
 351 retry:
 352         free_blocks = atomic_read(&sdp->sd_log_blks_free);
 353         if (unlikely(free_blocks <= wanted)) {
 354                 do {
 355                         prepare_to_wait_exclusive(&sdp->sd_log_waitq, &wait,
 356                                         TASK_UNINTERRUPTIBLE);
 357                         wake_up(&sdp->sd_logd_waitq);
 358                         did_wait = 1;
 359                         if (atomic_read(&sdp->sd_log_blks_free) <= wanted)
 360                                 io_schedule();
 361                         free_blocks = atomic_read(&sdp->sd_log_blks_free);
 362                 } while(free_blocks <= wanted);
 363                 finish_wait(&sdp->sd_log_waitq, &wait);
 364         }
 365         if (atomic_cmpxchg(&sdp->sd_log_blks_free, free_blocks,
 366                                 free_blocks - blks) != free_blocks)
 367                 goto retry;
 368         trace_gfs2_log_blocks(sdp, -blks);
 369
 370         /*
 371          * If we waited, then so might others, wake them up _after_ we get
 372          * our share of the log.
 373          */
 374         if (unlikely(did_wait))
 375                 wake_up(&sdp->sd_log_waitq);
 376
 377         down_read(&sdp->sd_log_flush_lock);
 378         if (unlikely(!test_bit(SDF_JOURNAL_LIVE, &sdp->sd_flags))) {
 379                 gfs2_log_release(sdp, blks);
 380                 return -EROFS;
 381         }
 382         return 0;
 383 }
 384
 385 /**
 386  * log_distance - Compute distance between two journal blocks
 387  * @sdp: The GFS2 superblock
 388  * @newer: The most recent journal block of the pair
 389  * @older: The older journal block of the pair
 390  *
 391  *   Compute the distance (in the journal direction) between two
 392  *   blocks in the journal
 393  *
 394  * Returns: the distance in blocks
 395  */
 396
 397 static inline unsigned int log_distance(struct gfs2_sbd *sdp, unsigned int newer,
 398                                         unsigned int older)
 399 {
 400         int dist;
 401
 402         dist = newer - older;
 403         if (dist < 0)
 404                 dist += sdp->sd_jdesc->jd_blocks;
 405
 406         return dist;
 407 }
 408
 409 /**
 410  * calc_reserved - Calculate the number of blocks to reserve when
 411  *                 refunding a transaction's unused buffers.
 412  * @sdp: The GFS2 superblock
 413  *
 414  * This is complex.  We need to reserve room for all our currently used
 415  * metadata buffers (e.g. normal file I/O rewriting file time stamps) and
 416  * all our journaled data buffers for journaled files (e.g. files in the
 417  * meta_fs like rindex, or files for which chattr +j was done.)
 418  * If we don't reserve enough space, gfs2_log_refund and gfs2_log_flush
 419  * will count it as free space (sd_log_blks_free) and corruption will follow.
 420  *
 421  * We can have metadata bufs and jdata bufs in the same journal.  So each
 422  * type gets its own log header, for which we need to reserve a block.
 423  * In fact, each type has the potential for needing more than one header
 424  * in cases where we have more buffers than will fit on a journal page.
 425  * Metadata journal entries take up half the space of journaled buffer entries.
 426  * Thus, metadata entries have buf_limit (502) and journaled buffers have
 427  * databuf_limit (251) before they cause a wrap around.
 428  *
 429  * Also, we need to reserve blocks for revoke journal entries and one for an
 430  * overall header for the lot.
 431  *
 432  * Returns: the number of blocks reserved
 433  */
 434 static unsigned int calc_reserved(struct gfs2_sbd *sdp)
 435 {
 436         unsigned int reserved = 0;
 437         unsigned int mbuf;
 438         unsigned int dbuf;
 439         struct gfs2_trans *tr = sdp->sd_log_tr;
 440
 441         if (tr) {
 442                 mbuf = tr->tr_num_buf_new - tr->tr_num_buf_rm;
 443                 dbuf = tr->tr_num_databuf_new - tr->tr_num_databuf_rm;
 444                 reserved = mbuf + dbuf;
 445                 /* Account for header blocks */
 446                 reserved += DIV_ROUND_UP(mbuf, buf_limit(sdp));
 447                 reserved += DIV_ROUND_UP(dbuf, databuf_limit(sdp));
 448         }
 449
 450         if (sdp->sd_log_commited_revoke > 0)
 451                 reserved += gfs2_struct2blk(sdp, sdp->sd_log_commited_revoke,
 452                                           sizeof(u64));
 453         /* One for the overall header */
 454         if (reserved)
 455                 reserved++;
 456         return reserved;
 457 }
 458
 459 static unsigned int current_tail(struct gfs2_sbd *sdp)
 460 {
 461         struct gfs2_trans *tr;
 462         unsigned int tail;
 463
 464         spin_lock(&sdp->sd_ail_lock);
 465
 466         if (list_empty(&sdp->sd_ail1_list)) {
 467                 tail = sdp->sd_log_head;
 468         } else {
 469                 tr = list_entry(sdp->sd_ail1_list.prev, struct gfs2_trans,
 470                                 tr_list);
 471                 tail = tr->tr_first;
 472         }
 473
 474         spin_unlock(&sdp->sd_ail_lock);
 475
 476         return tail;
 477 }
 478
 479 static void log_pull_tail(struct gfs2_sbd *sdp, unsigned int new_tail)
 480 {
 481         unsigned int dist = log_distance(sdp, new_tail, sdp->sd_log_tail);
 482
 483         ail2_empty(sdp, new_tail);
 484
 485         atomic_add(dist, &sdp->sd_log_blks_free);
 486         trace_gfs2_log_blocks(sdp, dist);
 487         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
 488                              sdp->sd_jdesc->jd_blocks);
 489
 490         sdp->sd_log_tail = new_tail;
 491 }
 492
 493
 494 static void log_flush_wait(struct gfs2_sbd *sdp)
 495 {
 496         DEFINE_WAIT(wait);
 497
 498         if (atomic_read(&sdp->sd_log_in_flight)) {
 499                 do {
 500                         prepare_to_wait(&sdp->sd_log_flush_wait, &wait,
 501                                         TASK_UNINTERRUPTIBLE);
 502                         if (atomic_read(&sdp->sd_log_in_flight))
 503                                 io_schedule();
 504                 } while(atomic_read(&sdp->sd_log_in_flight));
 505                 finish_wait(&sdp->sd_log_flush_wait, &wait);
 506         }
 507 }
 508
 509 static int ip_cmp(void *priv, struct list_head *a, struct list_head *b)
 510 {
 511         struct gfs2_inode *ipa, *ipb;
 512
 513         ipa = list_entry(a, struct gfs2_inode, i_ordered);
 514         ipb = list_entry(b, struct gfs2_inode, i_ordered);
 515
 516         if (ipa->i_no_addr < ipb->i_no_addr)
 517                 return -1;
 518         if (ipa->i_no_addr > ipb->i_no_addr)
 519                 return 1;
 520         return 0;
 521 }
 522
 523 static void gfs2_ordered_write(struct gfs2_sbd *sdp)
 524 {
 525         struct gfs2_inode *ip;
 526         LIST_HEAD(written);
 527
 528         spin_lock(&sdp->sd_ordered_lock);
 529         list_sort(NULL, &sdp->sd_log_le_ordered, &ip_cmp);
 530         while (!list_empty(&sdp->sd_log_le_ordered)) {
 531                 ip = list_entry(sdp->sd_log_le_ordered.next, struct gfs2_inode, i_ordered);
 532                 list_move(&ip->i_ordered, &written);
 533                 if (ip->i_inode.i_mapping->nrpages == 0)
 534                         continue;
 535                 spin_unlock(&sdp->sd_ordered_lock);
 536                 filemap_fdatawrite(ip->i_inode.i_mapping);
 537                 spin_lock(&sdp->sd_ordered_lock);
 538         }
 539         list_splice(&written, &sdp->sd_log_le_ordered);
 540         spin_unlock(&sdp->sd_ordered_lock);
 541 }
 542
 543 static void gfs2_ordered_wait(struct gfs2_sbd *sdp)
 544 {
 545         struct gfs2_inode *ip;
 546
 547         spin_lock(&sdp->sd_ordered_lock);
 548         while (!list_empty(&sdp->sd_log_le_ordered)) {
 549                 ip = list_entry(sdp->sd_log_le_ordered.next, struct gfs2_inode, i_ordered);
 550                 list_del(&ip->i_ordered);
 551                 WARN_ON(!test_and_clear_bit(GIF_ORDERED, &ip->i_flags));
 552                 if (ip->i_inode.i_mapping->nrpages == 0)
 553                         continue;
 554                 spin_unlock(&sdp->sd_ordered_lock);
 555                 filemap_fdatawait(ip->i_inode.i_mapping);
 556                 spin_lock(&sdp->sd_ordered_lock);
 557         }
 558         spin_unlock(&sdp->sd_ordered_lock);
 559 }
 560
 561 void gfs2_ordered_del_inode(struct gfs2_inode *ip)
 562 {
 563         struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
 564
 565         spin_lock(&sdp->sd_ordered_lock);
 566         if (test_and_clear_bit(GIF_ORDERED, &ip->i_flags))
 567                 list_del(&ip->i_ordered);
 568         spin_unlock(&sdp->sd_ordered_lock);
 569 }
 570
 571 void gfs2_add_revoke(struct gfs2_sbd *sdp, struct gfs2_bufdata *bd)
 572 {
 573         struct buffer_head *bh = bd->bd_bh;
 574         struct gfs2_glock *gl = bd->bd_gl;
 575
 576         bh->b_private = NULL;
 577         bd->bd_blkno = bh->b_blocknr;
 578         gfs2_remove_from_ail(bd); /* drops ref on bh */
 579         bd->bd_bh = NULL;
 580         bd->bd_ops = &gfs2_revoke_lops;
 581         sdp->sd_log_num_revoke++;
 582         atomic_inc(&gl->gl_revokes);
 583         set_bit(GLF_LFLUSH, &gl->gl_flags);
 584         list_add(&bd->bd_list, &sdp->sd_log_le_revoke);
 585 }
 586
 587 void gfs2_write_revokes(struct gfs2_sbd *sdp)
 588 {
 589         struct gfs2_trans *tr;
 590         struct gfs2_bufdata *bd, *tmp;
 591         int have_revokes = 0;
 592         int max_revokes = (sdp->sd_sb.sb_bsize - sizeof(struct gfs2_log_descriptor)) / sizeof(u64);
 593
 594         gfs2_ail1_empty(sdp);
 595         spin_lock(&sdp->sd_ail_lock);
 596         list_for_each_entry(tr, &sdp->sd_ail1_list, tr_list) {
 597                 list_for_each_entry(bd, &tr->tr_ail2_list, bd_ail_st_list) {
 598                         if (list_empty(&bd->bd_list)) {
 599                                 have_revokes = 1;
 600                                 goto done;
 601                         }
 602                 }
 603         }
 604 done:
 605         spin_unlock(&sdp->sd_ail_lock);
 606         if (have_revokes == 0)
 607                 return;
 608         while (sdp->sd_log_num_revoke > max_revokes)
 609                 max_revokes += (sdp->sd_sb.sb_bsize - sizeof(struct gfs2_meta_header)) / sizeof(u64);
 610         max_revokes -= sdp->sd_log_num_revoke;
 611         if (!sdp->sd_log_num_revoke) {
 612                 atomic_dec(&sdp->sd_log_blks_free);
 613                 /* If no blocks have been reserved, we need to also
 614                  * reserve a block for the header */
 615                 if (!sdp->sd_log_blks_reserved)
 616                         atomic_dec(&sdp->sd_log_blks_free);
 617         }
 618         gfs2_log_lock(sdp);
 619         spin_lock(&sdp->sd_ail_lock);
 620         list_for_each_entry(tr, &sdp->sd_ail1_list, tr_list) {
 621                 list_for_each_entry_safe(bd, tmp, &tr->tr_ail2_list, bd_ail_st_list) {
 622                         if (max_revokes == 0)
 623                                 goto out_of_blocks;
 624                         if (!list_empty(&bd->bd_list))
 625                                 continue;
 626                         gfs2_add_revoke(sdp, bd);
 627                         max_revokes--;
 628                 }
 629         }
 630 out_of_blocks:
 631         spin_unlock(&sdp->sd_ail_lock);
 632         gfs2_log_unlock(sdp);
 633
 634         if (!sdp->sd_log_num_revoke) {
 635                 atomic_inc(&sdp->sd_log_blks_free);
 636                 if (!sdp->sd_log_blks_reserved)
 637                         atomic_inc(&sdp->sd_log_blks_free);
 638         }
 639 }
 640
 641 /**
 642  * log_write_header - Get and initialize a journal header buffer
 643  * @sdp: The GFS2 superblock
 644  *
 645  * Returns: the initialized log buffer descriptor
 646  */
 647
 648 static void log_write_header(struct gfs2_sbd *sdp, u32 flags)
 649 {
 650         struct gfs2_log_header *lh;
 651         unsigned int tail;
 652         u32 hash;
 653         int rw = WRITE_FLUSH_FUA | REQ_META;
 654         struct page *page = mempool_alloc(gfs2_page_pool, GFP_NOIO);
 655         lh = page_address(page);
 656         clear_page(lh);
 657
 658         tail = current_tail(sdp);
 659
 660         lh->lh_header.mh_magic = cpu_to_be32(GFS2_MAGIC);
 661         lh->lh_header.mh_type = cpu_to_be32(GFS2_METATYPE_LH);
 662         lh->lh_header.__pad0 = cpu_to_be64(0);
 663         lh->lh_header.mh_format = cpu_to_be32(GFS2_FORMAT_LH);
 664         lh->lh_header.mh_jid = cpu_to_be32(sdp->sd_jdesc->jd_jid);
 665         lh->lh_sequence = cpu_to_be64(sdp->sd_log_sequence++);
 666         lh->lh_flags = cpu_to_be32(flags);
 667         lh->lh_tail = cpu_to_be32(tail);
 668         lh->lh_blkno = cpu_to_be32(sdp->sd_log_flush_head);
 669         hash = gfs2_disk_hash(page_address(page), sizeof(struct gfs2_log_header));
 670         lh->lh_hash = cpu_to_be32(hash);
 671
 672         if (test_bit(SDF_NOBARRIERS, &sdp->sd_flags)) {
 673                 gfs2_ordered_wait(sdp);
 674                 log_flush_wait(sdp);
 675                 rw = WRITE_SYNC | REQ_META | REQ_PRIO;
 676         }
 677
 678         sdp->sd_log_idle = (tail == sdp->sd_log_flush_head);
 679         gfs2_log_write_page(sdp, page);
 680         gfs2_log_flush_bio(sdp, rw);
 681         log_flush_wait(sdp);
 682
 683         if (sdp->sd_log_tail != tail)
 684                 log_pull_tail(sdp, tail);
 685 }
 686
 687 /**
 688  * gfs2_log_flush - flush incore transaction(s)
 689  * @sdp: the filesystem
 690  * @gl: The glock structure to flush.  If NULL, flush the whole incore log
 691  *
 692  */
 693
 694 void gfs2_log_flush(struct gfs2_sbd *sdp, struct gfs2_glock *gl,
 695                     enum gfs2_flush_type type)
 696 {
 697         struct gfs2_trans *tr;
 698
 699         down_write(&sdp->sd_log_flush_lock);
 700
 701         /* Log might have been flushed while we waited for the flush lock */
 702         if (gl && !test_bit(GLF_LFLUSH, &gl->gl_flags)) {
 703                 up_write(&sdp->sd_log_flush_lock);
 704                 return;
 705         }
 706         trace_gfs2_log_flush(sdp, 1);
 707
 708         sdp->sd_log_flush_head = sdp->sd_log_head;
 709         sdp->sd_log_flush_wrapped = 0;
 710         tr = sdp->sd_log_tr;
 711         if (tr) {
 712                 sdp->sd_log_tr = NULL;
 713                 INIT_LIST_HEAD(&tr->tr_ail1_list);
 714                 INIT_LIST_HEAD(&tr->tr_ail2_list);
 715                 tr->tr_first = sdp->sd_log_flush_head;
 716         }
 717
 718         gfs2_assert_withdraw(sdp,
 719                         sdp->sd_log_num_revoke == sdp->sd_log_commited_revoke);
 720
 721         gfs2_ordered_write(sdp);
 722         lops_before_commit(sdp, tr);
 723         gfs2_log_flush_bio(sdp, WRITE);
 724
 725         if (sdp->sd_log_head != sdp->sd_log_flush_head) {
 726                 log_flush_wait(sdp);
 727                 log_write_header(sdp, 0);
 728         } else if (sdp->sd_log_tail != current_tail(sdp) && !sdp->sd_log_idle){
 729                 atomic_dec(&sdp->sd_log_blks_free); /* Adjust for unreserved buffer */
 730                 trace_gfs2_log_blocks(sdp, -1);
 731                 log_write_header(sdp, 0);
 732         }
 733         lops_after_commit(sdp, tr);
 734
 735         gfs2_log_lock(sdp);
 736         sdp->sd_log_head = sdp->sd_log_flush_head;
 737         sdp->sd_log_blks_reserved = 0;
 738         sdp->sd_log_commited_revoke = 0;
 739
 740         spin_lock(&sdp->sd_ail_lock);
 741         if (tr && !list_empty(&tr->tr_ail1_list)) {
 742                 list_add(&tr->tr_list, &sdp->sd_ail1_list);
 743                 tr = NULL;
 744         }
 745         spin_unlock(&sdp->sd_ail_lock);
 746         gfs2_log_unlock(sdp);
 747
 748         if (atomic_read(&sdp->sd_log_freeze))
 749                 type = FREEZE_FLUSH;
 750         if (type != NORMAL_FLUSH) {
 751                 if (!sdp->sd_log_idle) {
 752                         for (;;) {
 753                                 gfs2_ail1_start(sdp);
 754                                 gfs2_ail1_wait(sdp);
 755                                 if (gfs2_ail1_empty(sdp))
 756                                         break;
 757                         }
 758                         atomic_dec(&sdp->sd_log_blks_free); /* Adjust for unreserved buffer */
 759                         trace_gfs2_log_blocks(sdp, -1);
 760                         sdp->sd_log_flush_wrapped = 0;
 761                         log_write_header(sdp, 0);
 762                         sdp->sd_log_head = sdp->sd_log_flush_head;
 763                 }
 764                 if (type == SHUTDOWN_FLUSH || type == FREEZE_FLUSH)
 765                         gfs2_log_shutdown(sdp);
 766                 if (type == FREEZE_FLUSH) {
 767                         int error;
 768
 769                         atomic_set(&sdp->sd_log_freeze, 0);
 770                         wake_up(&sdp->sd_log_frozen_wait);
 771                         error = gfs2_glock_nq_init(sdp->sd_freeze_gl,
 772                                                    LM_ST_SHARED, 0,
 773                                                    &sdp->sd_thaw_gh);
 774                         if (error) {
 775                                 printk(KERN_INFO "GFS2: couln't get freeze lock : %d\n", error);
 776                                 gfs2_assert_withdraw(sdp, 0);
 777                         }
 778                         else
 779                                 gfs2_glock_dq_uninit(&sdp->sd_thaw_gh);
 780                 }
 781         }
 782
 783         trace_gfs2_log_flush(sdp, 0);
 784         up_write(&sdp->sd_log_flush_lock);
 785
 786         kfree(tr);
 787 }
 788
 789 /**
 790  * gfs2_merge_trans - Merge a new transaction into a cached transaction
 791  * @old: Original transaction to be expanded
 792  * @new: New transaction to be merged
 793  */
 794
 795 static void gfs2_merge_trans(struct gfs2_trans *old, struct gfs2_trans *new)
 796 {
 797         WARN_ON_ONCE(old->tr_attached != 1);
 798
 799         old->tr_num_buf_new     += new->tr_num_buf_new;
 800         old->tr_num_databuf_new += new->tr_num_databuf_new;
 801         old->tr_num_buf_rm      += new->tr_num_buf_rm;
 802         old->tr_num_databuf_rm  += new->tr_num_databuf_rm;
 803         old->tr_num_revoke      += new->tr_num_revoke;
 804         old->tr_num_revoke_rm   += new->tr_num_revoke_rm;
 805
 806         list_splice_tail_init(&new->tr_databuf, &old->tr_databuf);
 807         list_splice_tail_init(&new->tr_buf, &old->tr_buf);
 808 }
 809
 810 static void log_refund(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 811 {
 812         unsigned int reserved;
 813         unsigned int unused;
 814         unsigned int maxres;
 815
 816         gfs2_log_lock(sdp);
 817
 818         if (sdp->sd_log_tr) {
 819                 gfs2_merge_trans(sdp->sd_log_tr, tr);
 820         } else if (tr->tr_num_buf_new || tr->tr_num_databuf_new) {
 821                 gfs2_assert_withdraw(sdp, tr->tr_alloced);
 822                 sdp->sd_log_tr = tr;
 823                 tr->tr_attached = 1;
 824         }
 825
 826         sdp->sd_log_commited_revoke += tr->tr_num_revoke - tr->tr_num_revoke_rm;
 827         reserved = calc_reserved(sdp);
 828         maxres = sdp->sd_log_blks_reserved + tr->tr_reserved;
 829         gfs2_assert_withdraw(sdp, maxres >= reserved);
 830         unused = maxres - reserved;
 831         atomic_add(unused, &sdp->sd_log_blks_free);
 832         trace_gfs2_log_blocks(sdp, unused);
 833         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
 834                              sdp->sd_jdesc->jd_blocks);
 835         sdp->sd_log_blks_reserved = reserved;
 836
 837         gfs2_log_unlock(sdp);
 838 }
 839
 840 /**
 841  * gfs2_log_commit - Commit a transaction to the log
 842  * @sdp: the filesystem
 843  * @tr: the transaction
 844  *
 845  * We wake up gfs2_logd if the number of pinned blocks exceed thresh1
 846  * or the total number of used blocks (pinned blocks plus AIL blocks)
 847  * is greater than thresh2.
 848  *
 849  * At mount time thresh1 is 1/3rd of journal size, thresh2 is 2/3rd of
 850  * journal size.
 851  *
 852  * Returns: errno
 853  */
 854
 855 void gfs2_log_commit(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 856 {
 857         log_refund(sdp, tr);
 858
 859         if (atomic_read(&sdp->sd_log_pinned) > atomic_read(&sdp->sd_log_thresh1) ||
 860             ((sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free)) >
 861             atomic_read(&sdp->sd_log_thresh2)))
 862                 wake_up(&sdp->sd_logd_waitq);
 863 }
 864
 865 /**
 866  * gfs2_log_shutdown - write a shutdown header into a journal
 867  * @sdp: the filesystem
 868  *
 869  */
 870
 871 void gfs2_log_shutdown(struct gfs2_sbd *sdp)
 872 {
 873         gfs2_assert_withdraw(sdp, !sdp->sd_log_blks_reserved);
 874         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_revoke);
 875         gfs2_assert_withdraw(sdp, list_empty(&sdp->sd_ail1_list));
 876
 877         sdp->sd_log_flush_head = sdp->sd_log_head;
 878         sdp->sd_log_flush_wrapped = 0;
 879
 880         log_write_header(sdp, GFS2_LOG_HEAD_UNMOUNT);
 881
 882         gfs2_assert_warn(sdp, sdp->sd_log_head == sdp->sd_log_tail);
 883         gfs2_assert_warn(sdp, list_empty(&sdp->sd_ail2_list));
 884
 885         sdp->sd_log_head = sdp->sd_log_flush_head;
 886         sdp->sd_log_tail = sdp->sd_log_head;
 887 }
 888
 889 static inline int gfs2_jrnl_flush_reqd(struct gfs2_sbd *sdp)
 890 {
 891         return (atomic_read(&sdp->sd_log_pinned) >= atomic_read(&sdp->sd_log_thresh1) || atomic_read(&sdp->sd_log_freeze));
 892 }
 893
 894 static inline int gfs2_ail_flush_reqd(struct gfs2_sbd *sdp)
 895 {
 896         unsigned int used_blocks = sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free);
 897         return used_blocks >= atomic_read(&sdp->sd_log_thresh2);
 898 }
 899
 900 /**
 901  * gfs2_logd - Update log tail as Active Items get flushed to in-place blocks
 902  * @sdp: Pointer to GFS2 superblock
 903  *
 904  * Also, periodically check to make sure that we're using the most recent
 905  * journal index.
 906  */
 907
 908 int gfs2_logd(void *data)
 909 {
 910         struct gfs2_sbd *sdp = data;
 911         unsigned long t = 1;
 912         DEFINE_WAIT(wait);
 913
 914         while (!kthread_should_stop()) {
 915
 916                 if (gfs2_jrnl_flush_reqd(sdp) || t == 0) {
 917                         gfs2_ail1_empty(sdp);
 918                         gfs2_log_flush(sdp, NULL, NORMAL_FLUSH);
 919                 }
 920
 921                 if (gfs2_ail_flush_reqd(sdp)) {
 922                         gfs2_ail1_start(sdp);
 923                         gfs2_ail1_wait(sdp);
 924                         gfs2_ail1_empty(sdp);
 925                         gfs2_log_flush(sdp, NULL, NORMAL_FLUSH);
 926                 }
 927
 928                 if (!gfs2_ail_flush_reqd(sdp))
 929                         wake_up(&sdp->sd_log_waitq);
 930
 931                 t = gfs2_tune_get(sdp, gt_logd_secs) * HZ;
 932
 933                 try_to_freeze();
 934
 935                 do {
 936                         prepare_to_wait(&sdp->sd_logd_waitq, &wait,
 937                                         TASK_INTERRUPTIBLE);
 938                         if (!gfs2_ail_flush_reqd(sdp) &&
 939                             !gfs2_jrnl_flush_reqd(sdp) &&
 940                             !kthread_should_stop())
 941                                 t = schedule_timeout(t);
 942                 } while(t && !gfs2_ail_flush_reqd(sdp) &&
 943                         !gfs2_jrnl_flush_reqd(sdp) &&
 944                         !kthread_should_stop());
 945                 finish_wait(&sdp->sd_logd_waitq, &wait);
 946         }
 947
 948         return 0;
 949 }
 950