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
  22 #include "gfs2.h"
  23 #include "incore.h"
  24 #include "bmap.h"
  25 #include "glock.h"
  26 #include "log.h"
  27 #include "lops.h"
  28 #include "meta_io.h"
  29 #include "util.h"
  30 #include "dir.h"
  31 #include "trace_gfs2.h"
  32
  33 #define PULL 1
  34
  35 /**
  36  * gfs2_struct2blk - compute stuff
  37  * @sdp: the filesystem
  38  * @nstruct: the number of structures
  39  * @ssize: the size of the structures
  40  *
  41  * Compute the number of log descriptor blocks needed to hold a certain number
  42  * of structures of a certain size.
  43  *
  44  * Returns: the number of blocks needed (minimum is always 1)
  45  */
  46
  47 unsigned int gfs2_struct2blk(struct gfs2_sbd *sdp, unsigned int nstruct,
  48                              unsigned int ssize)
  49 {
  50         unsigned int blks;
  51         unsigned int first, second;
  52
  53         blks = 1;
  54         first = (sdp->sd_sb.sb_bsize - sizeof(struct gfs2_log_descriptor)) / ssize;
  55
  56         if (nstruct > first) {
  57                 second = (sdp->sd_sb.sb_bsize -
  58                           sizeof(struct gfs2_meta_header)) / ssize;
  59                 blks += DIV_ROUND_UP(nstruct - first, second);
  60         }
  61
  62         return blks;
  63 }
  64
  65 /**
  66  * gfs2_remove_from_ail - Remove an entry from the ail lists, updating counters
  67  * @mapping: The associated mapping (maybe NULL)
  68  * @bd: The gfs2_bufdata to remove
  69  *
  70  * The log lock _must_ be held when calling this function
  71  *
  72  */
  73
  74 void gfs2_remove_from_ail(struct gfs2_bufdata *bd)
  75 {
  76         bd->bd_ail = NULL;
  77         list_del_init(&bd->bd_ail_st_list);
  78         list_del_init(&bd->bd_ail_gl_list);
  79         atomic_dec(&bd->bd_gl->gl_ail_count);
  80         brelse(bd->bd_bh);
  81 }
  82
  83 /**
  84  * gfs2_ail1_start_one - Start I/O on a part of the AIL
  85  * @sdp: the filesystem
  86  * @tr: the part of the AIL
  87  *
  88  */
  89
  90 static void gfs2_ail1_start_one(struct gfs2_sbd *sdp, struct gfs2_ail *ai)
  91 __releases(&sdp->sd_log_lock)
  92 __acquires(&sdp->sd_log_lock)
  93 {
  94         struct gfs2_bufdata *bd, *s;
  95         struct buffer_head *bh;
  96         int retry;
  97
  98         do {
  99                 retry = 0;
 100
 101                 list_for_each_entry_safe_reverse(bd, s, &ai->ai_ail1_list,
 102                                                  bd_ail_st_list) {
 103                         bh = bd->bd_bh;
 104
 105                         gfs2_assert(sdp, bd->bd_ail == ai);
 106
 107                         if (!buffer_busy(bh)) {
 108                                 if (!buffer_uptodate(bh))
 109                                         gfs2_io_error_bh(sdp, bh);
 110                                 list_move(&bd->bd_ail_st_list, &ai->ai_ail2_list);
 111                                 continue;
 112                         }
 113
 114                         if (!buffer_dirty(bh))
 115                                 continue;
 116
 117                         list_move(&bd->bd_ail_st_list, &ai->ai_ail1_list);
 118
 119                         get_bh(bh);
 120                         gfs2_log_unlock(sdp);
 121                         lock_buffer(bh);
 122                         if (test_clear_buffer_dirty(bh)) {
 123                                 bh->b_end_io = end_buffer_write_sync;
 124                                 submit_bh(WRITE_SYNC_PLUG, bh);
 125                         } else {
 126                                 unlock_buffer(bh);
 127                                 brelse(bh);
 128                         }
 129                         gfs2_log_lock(sdp);
 130
 131                         retry = 1;
 132                         break;
 133                 }
 134         } while (retry);
 135 }
 136
 137 /**
 138  * gfs2_ail1_empty_one - Check whether or not a trans in the AIL has been synced
 139  * @sdp: the filesystem
 140  * @ai: the AIL entry
 141  *
 142  */
 143
 144 static int gfs2_ail1_empty_one(struct gfs2_sbd *sdp, struct gfs2_ail *ai, int flags)
 145 {
 146         struct gfs2_bufdata *bd, *s;
 147         struct buffer_head *bh;
 148
 149         list_for_each_entry_safe_reverse(bd, s, &ai->ai_ail1_list,
 150                                          bd_ail_st_list) {
 151                 bh = bd->bd_bh;
 152
 153                 gfs2_assert(sdp, bd->bd_ail == ai);
 154
 155                 if (buffer_busy(bh)) {
 156                         if (flags & DIO_ALL)
 157                                 continue;
 158                         else
 159                                 break;
 160                 }
 161
 162                 if (!buffer_uptodate(bh))
 163                         gfs2_io_error_bh(sdp, bh);
 164
 165                 list_move(&bd->bd_ail_st_list, &ai->ai_ail2_list);
 166         }
 167
 168         return list_empty(&ai->ai_ail1_list);
 169 }
 170
 171 static void gfs2_ail1_start(struct gfs2_sbd *sdp)
 172 {
 173         struct list_head *head;
 174         u64 sync_gen;
 175         struct gfs2_ail *ai;
 176         int done = 0;
 177
 178         gfs2_log_lock(sdp);
 179         head = &sdp->sd_ail1_list;
 180         if (list_empty(head)) {
 181                 gfs2_log_unlock(sdp);
 182                 return;
 183         }
 184         sync_gen = sdp->sd_ail_sync_gen++;
 185
 186         while(!done) {
 187                 done = 1;
 188                 list_for_each_entry_reverse(ai, head, ai_list) {
 189                         if (ai->ai_sync_gen >= sync_gen)
 190                                 continue;
 191                         ai->ai_sync_gen = sync_gen;
 192                         gfs2_ail1_start_one(sdp, ai); /* This may drop log lock */
 193                         done = 0;
 194                         break;
 195                 }
 196         }
 197
 198         gfs2_log_unlock(sdp);
 199 }
 200
 201 static int gfs2_ail1_empty(struct gfs2_sbd *sdp, int flags)
 202 {
 203         struct gfs2_ail *ai, *s;
 204         int ret;
 205
 206         gfs2_log_lock(sdp);
 207
 208         list_for_each_entry_safe_reverse(ai, s, &sdp->sd_ail1_list, ai_list) {
 209                 if (gfs2_ail1_empty_one(sdp, ai, flags))
 210                         list_move(&ai->ai_list, &sdp->sd_ail2_list);
 211                 else if (!(flags & DIO_ALL))
 212                         break;
 213         }
 214
 215         ret = list_empty(&sdp->sd_ail1_list);
 216
 217         gfs2_log_unlock(sdp);
 218
 219         return ret;
 220 }
 221
 222
 223 /**
 224  * gfs2_ail2_empty_one - Check whether or not a trans in the AIL has been synced
 225  * @sdp: the filesystem
 226  * @ai: the AIL entry
 227  *
 228  */
 229
 230 static void gfs2_ail2_empty_one(struct gfs2_sbd *sdp, struct gfs2_ail *ai)
 231 {
 232         struct list_head *head = &ai->ai_ail2_list;
 233         struct gfs2_bufdata *bd;
 234
 235         while (!list_empty(head)) {
 236                 bd = list_entry(head->prev, struct gfs2_bufdata,
 237                                 bd_ail_st_list);
 238                 gfs2_assert(sdp, bd->bd_ail == ai);
 239                 gfs2_remove_from_ail(bd);
 240         }
 241 }
 242
 243 static void ail2_empty(struct gfs2_sbd *sdp, unsigned int new_tail)
 244 {
 245         struct gfs2_ail *ai, *safe;
 246         unsigned int old_tail = sdp->sd_log_tail;
 247         int wrap = (new_tail < old_tail);
 248         int a, b, rm;
 249
 250         gfs2_log_lock(sdp);
 251
 252         list_for_each_entry_safe(ai, safe, &sdp->sd_ail2_list, ai_list) {
 253                 a = (old_tail <= ai->ai_first);
 254                 b = (ai->ai_first < new_tail);
 255                 rm = (wrap) ? (a || b) : (a && b);
 256                 if (!rm)
 257                         continue;
 258
 259                 gfs2_ail2_empty_one(sdp, ai);
 260                 list_del(&ai->ai_list);
 261                 gfs2_assert_warn(sdp, list_empty(&ai->ai_ail1_list));
 262                 gfs2_assert_warn(sdp, list_empty(&ai->ai_ail2_list));
 263                 kfree(ai);
 264         }
 265
 266         gfs2_log_unlock(sdp);
 267 }
 268
 269 /**
 270  * gfs2_log_reserve - Make a log reservation
 271  * @sdp: The GFS2 superblock
 272  * @blks: The number of blocks to reserve
 273  *
 274  * Note that we never give out the last few blocks of the journal. Thats
 275  * due to the fact that there is a small number of header blocks
 276  * associated with each log flush. The exact number can't be known until
 277  * flush time, so we ensure that we have just enough free blocks at all
 278  * times to avoid running out during a log flush.
 279  *
 280  * We no longer flush the log here, instead we wake up logd to do that
 281  * for us. To avoid the thundering herd and to ensure that we deal fairly
 282  * with queued waiters, we use an exclusive wait. This means that when we
 283  * get woken with enough journal space to get our reservation, we need to
 284  * wake the next waiter on the list.
 285  *
 286  * Returns: errno
 287  */
 288
 289 int gfs2_log_reserve(struct gfs2_sbd *sdp, unsigned int blks)
 290 {
 291         unsigned reserved_blks = 6 * (4096 / sdp->sd_vfs->s_blocksize);
 292         unsigned wanted = blks + reserved_blks;
 293         DEFINE_WAIT(wait);
 294         int did_wait = 0;
 295         unsigned int free_blocks;
 296
 297         if (gfs2_assert_warn(sdp, blks) ||
 298             gfs2_assert_warn(sdp, blks <= sdp->sd_jdesc->jd_blocks))
 299                 return -EINVAL;
 300 retry:
 301         free_blocks = atomic_read(&sdp->sd_log_blks_free);
 302         if (unlikely(free_blocks <= wanted)) {
 303                 do {
 304                         prepare_to_wait_exclusive(&sdp->sd_log_waitq, &wait,
 305                                         TASK_UNINTERRUPTIBLE);
 306                         wake_up(&sdp->sd_logd_waitq);
 307                         did_wait = 1;
 308                         if (atomic_read(&sdp->sd_log_blks_free) <= wanted)
 309                                 io_schedule();
 310                         free_blocks = atomic_read(&sdp->sd_log_blks_free);
 311                 } while(free_blocks <= wanted);
 312                 finish_wait(&sdp->sd_log_waitq, &wait);
 313         }
 314         if (atomic_cmpxchg(&sdp->sd_log_blks_free, free_blocks,
 315                                 free_blocks - blks) != free_blocks)
 316                 goto retry;
 317         trace_gfs2_log_blocks(sdp, -blks);
 318
 319         /*
 320          * If we waited, then so might others, wake them up _after_ we get
 321          * our share of the log.
 322          */
 323         if (unlikely(did_wait))
 324                 wake_up(&sdp->sd_log_waitq);
 325
 326         down_read(&sdp->sd_log_flush_lock);
 327
 328         return 0;
 329 }
 330
 331 static u64 log_bmap(struct gfs2_sbd *sdp, unsigned int lbn)
 332 {
 333         struct gfs2_journal_extent *je;
 334
 335         list_for_each_entry(je, &sdp->sd_jdesc->extent_list, extent_list) {
 336                 if (lbn >= je->lblock && lbn < je->lblock + je->blocks)
 337                         return je->dblock + lbn - je->lblock;
 338         }
 339
 340         return -1;
 341 }
 342
 343 /**
 344  * log_distance - Compute distance between two journal blocks
 345  * @sdp: The GFS2 superblock
 346  * @newer: The most recent journal block of the pair
 347  * @older: The older journal block of the pair
 348  *
 349  *   Compute the distance (in the journal direction) between two
 350  *   blocks in the journal
 351  *
 352  * Returns: the distance in blocks
 353  */
 354
 355 static inline unsigned int log_distance(struct gfs2_sbd *sdp, unsigned int newer,
 356                                         unsigned int older)
 357 {
 358         int dist;
 359
 360         dist = newer - older;
 361         if (dist < 0)
 362                 dist += sdp->sd_jdesc->jd_blocks;
 363
 364         return dist;
 365 }
 366
 367 /**
 368  * calc_reserved - Calculate the number of blocks to reserve when
 369  *                 refunding a transaction's unused buffers.
 370  * @sdp: The GFS2 superblock
 371  *
 372  * This is complex.  We need to reserve room for all our currently used
 373  * metadata buffers (e.g. normal file I/O rewriting file time stamps) and
 374  * all our journaled data buffers for journaled files (e.g. files in the
 375  * meta_fs like rindex, or files for which chattr +j was done.)
 376  * If we don't reserve enough space, gfs2_log_refund and gfs2_log_flush
 377  * will count it as free space (sd_log_blks_free) and corruption will follow.
 378  *
 379  * We can have metadata bufs and jdata bufs in the same journal.  So each
 380  * type gets its own log header, for which we need to reserve a block.
 381  * In fact, each type has the potential for needing more than one header
 382  * in cases where we have more buffers than will fit on a journal page.
 383  * Metadata journal entries take up half the space of journaled buffer entries.
 384  * Thus, metadata entries have buf_limit (502) and journaled buffers have
 385  * databuf_limit (251) before they cause a wrap around.
 386  *
 387  * Also, we need to reserve blocks for revoke journal entries and one for an
 388  * overall header for the lot.
 389  *
 390  * Returns: the number of blocks reserved
 391  */
 392 static unsigned int calc_reserved(struct gfs2_sbd *sdp)
 393 {
 394         unsigned int reserved = 0;
 395         unsigned int mbuf_limit, metabufhdrs_needed;
 396         unsigned int dbuf_limit, databufhdrs_needed;
 397         unsigned int revokes = 0;
 398
 399         mbuf_limit = buf_limit(sdp);
 400         metabufhdrs_needed = (sdp->sd_log_commited_buf +
 401                               (mbuf_limit - 1)) / mbuf_limit;
 402         dbuf_limit = databuf_limit(sdp);
 403         databufhdrs_needed = (sdp->sd_log_commited_databuf +
 404                               (dbuf_limit - 1)) / dbuf_limit;
 405
 406         if (sdp->sd_log_commited_revoke > 0)
 407                 revokes = gfs2_struct2blk(sdp, sdp->sd_log_commited_revoke,
 408                                           sizeof(u64));
 409
 410         reserved = sdp->sd_log_commited_buf + metabufhdrs_needed +
 411                 sdp->sd_log_commited_databuf + databufhdrs_needed +
 412                 revokes;
 413         /* One for the overall header */
 414         if (reserved)
 415                 reserved++;
 416         return reserved;
 417 }
 418
 419 static unsigned int current_tail(struct gfs2_sbd *sdp)
 420 {
 421         struct gfs2_ail *ai;
 422         unsigned int tail;
 423
 424         gfs2_log_lock(sdp);
 425
 426         if (list_empty(&sdp->sd_ail1_list)) {
 427                 tail = sdp->sd_log_head;
 428         } else {
 429                 ai = list_entry(sdp->sd_ail1_list.prev, struct gfs2_ail, ai_list);
 430                 tail = ai->ai_first;
 431         }
 432
 433         gfs2_log_unlock(sdp);
 434
 435         return tail;
 436 }
 437
 438 void gfs2_log_incr_head(struct gfs2_sbd *sdp)
 439 {
 440         if (sdp->sd_log_flush_head == sdp->sd_log_tail)
 441                 BUG_ON(sdp->sd_log_flush_head != sdp->sd_log_head);
 442
 443         if (++sdp->sd_log_flush_head == sdp->sd_jdesc->jd_blocks) {
 444                 sdp->sd_log_flush_head = 0;
 445                 sdp->sd_log_flush_wrapped = 1;
 446         }
 447 }
 448
 449 /**
 450  * gfs2_log_write_endio - End of I/O for a log buffer
 451  * @bh: The buffer head
 452  * @uptodate: I/O Status
 453  *
 454  */
 455
 456 static void gfs2_log_write_endio(struct buffer_head *bh, int uptodate)
 457 {
 458         struct gfs2_sbd *sdp = bh->b_private;
 459         bh->b_private = NULL;
 460
 461         end_buffer_write_sync(bh, uptodate);
 462         if (atomic_dec_and_test(&sdp->sd_log_in_flight))
 463                 wake_up(&sdp->sd_log_flush_wait);
 464 }
 465
 466 /**
 467  * gfs2_log_get_buf - Get and initialize a buffer to use for log control data
 468  * @sdp: The GFS2 superblock
 469  *
 470  * Returns: the buffer_head
 471  */
 472
 473 struct buffer_head *gfs2_log_get_buf(struct gfs2_sbd *sdp)
 474 {
 475         u64 blkno = log_bmap(sdp, sdp->sd_log_flush_head);
 476         struct buffer_head *bh;
 477
 478         bh = sb_getblk(sdp->sd_vfs, blkno);
 479         lock_buffer(bh);
 480         memset(bh->b_data, 0, bh->b_size);
 481         set_buffer_uptodate(bh);
 482         clear_buffer_dirty(bh);
 483         gfs2_log_incr_head(sdp);
 484         atomic_inc(&sdp->sd_log_in_flight);
 485         bh->b_private = sdp;
 486         bh->b_end_io = gfs2_log_write_endio;
 487
 488         return bh;
 489 }
 490
 491 /**
 492  * gfs2_fake_write_endio -
 493  * @bh: The buffer head
 494  * @uptodate: The I/O Status
 495  *
 496  */
 497
 498 static void gfs2_fake_write_endio(struct buffer_head *bh, int uptodate)
 499 {
 500         struct buffer_head *real_bh = bh->b_private;
 501         struct gfs2_bufdata *bd = real_bh->b_private;
 502         struct gfs2_sbd *sdp = bd->bd_gl->gl_sbd;
 503
 504         end_buffer_write_sync(bh, uptodate);
 505         free_buffer_head(bh);
 506         unlock_buffer(real_bh);
 507         brelse(real_bh);
 508         if (atomic_dec_and_test(&sdp->sd_log_in_flight))
 509                 wake_up(&sdp->sd_log_flush_wait);
 510 }
 511
 512 /**
 513  * gfs2_log_fake_buf - Build a fake buffer head to write metadata buffer to log
 514  * @sdp: the filesystem
 515  * @data: the data the buffer_head should point to
 516  *
 517  * Returns: the log buffer descriptor
 518  */
 519
 520 struct buffer_head *gfs2_log_fake_buf(struct gfs2_sbd *sdp,
 521                                       struct buffer_head *real)
 522 {
 523         u64 blkno = log_bmap(sdp, sdp->sd_log_flush_head);
 524         struct buffer_head *bh;
 525
 526         bh = alloc_buffer_head(GFP_NOFS | __GFP_NOFAIL);
 527         atomic_set(&bh->b_count, 1);
 528         bh->b_state = (1 << BH_Mapped) | (1 << BH_Uptodate) | (1 << BH_Lock);
 529         set_bh_page(bh, real->b_page, bh_offset(real));
 530         bh->b_blocknr = blkno;
 531         bh->b_size = sdp->sd_sb.sb_bsize;
 532         bh->b_bdev = sdp->sd_vfs->s_bdev;
 533         bh->b_private = real;
 534         bh->b_end_io = gfs2_fake_write_endio;
 535
 536         gfs2_log_incr_head(sdp);
 537         atomic_inc(&sdp->sd_log_in_flight);
 538
 539         return bh;
 540 }
 541
 542 static void log_pull_tail(struct gfs2_sbd *sdp, unsigned int new_tail)
 543 {
 544         unsigned int dist = log_distance(sdp, new_tail, sdp->sd_log_tail);
 545
 546         ail2_empty(sdp, new_tail);
 547
 548         atomic_add(dist, &sdp->sd_log_blks_free);
 549         trace_gfs2_log_blocks(sdp, dist);
 550         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
 551                              sdp->sd_jdesc->jd_blocks);
 552
 553         sdp->sd_log_tail = new_tail;
 554 }
 555
 556 /**
 557  * log_write_header - Get and initialize a journal header buffer
 558  * @sdp: The GFS2 superblock
 559  *
 560  * Returns: the initialized log buffer descriptor
 561  */
 562
 563 static void log_write_header(struct gfs2_sbd *sdp, u32 flags, int pull)
 564 {
 565         u64 blkno = log_bmap(sdp, sdp->sd_log_flush_head);
 566         struct buffer_head *bh;
 567         struct gfs2_log_header *lh;
 568         unsigned int tail;
 569         u32 hash;
 570
 571         bh = sb_getblk(sdp->sd_vfs, blkno);
 572         lock_buffer(bh);
 573         memset(bh->b_data, 0, bh->b_size);
 574         set_buffer_uptodate(bh);
 575         clear_buffer_dirty(bh);
 576
 577         gfs2_ail1_empty(sdp, 0);
 578         tail = current_tail(sdp);
 579
 580         lh = (struct gfs2_log_header *)bh->b_data;
 581         memset(lh, 0, sizeof(struct gfs2_log_header));
 582         lh->lh_header.mh_magic = cpu_to_be32(GFS2_MAGIC);
 583         lh->lh_header.mh_type = cpu_to_be32(GFS2_METATYPE_LH);
 584         lh->lh_header.__pad0 = cpu_to_be64(0);
 585         lh->lh_header.mh_format = cpu_to_be32(GFS2_FORMAT_LH);
 586         lh->lh_header.mh_jid = cpu_to_be32(sdp->sd_jdesc->jd_jid);
 587         lh->lh_sequence = cpu_to_be64(sdp->sd_log_sequence++);
 588         lh->lh_flags = cpu_to_be32(flags);
 589         lh->lh_tail = cpu_to_be32(tail);
 590         lh->lh_blkno = cpu_to_be32(sdp->sd_log_flush_head);
 591         hash = gfs2_disk_hash(bh->b_data, sizeof(struct gfs2_log_header));
 592         lh->lh_hash = cpu_to_be32(hash);
 593
 594         bh->b_end_io = end_buffer_write_sync;
 595         if (test_bit(SDF_NOBARRIERS, &sdp->sd_flags))
 596                 goto skip_barrier;
 597         get_bh(bh);
 598         submit_bh(WRITE_SYNC | (1 << BIO_RW_BARRIER) | (1 << BIO_RW_META), bh);
 599         wait_on_buffer(bh);
 600         if (buffer_eopnotsupp(bh)) {
 601                 clear_buffer_eopnotsupp(bh);
 602                 set_buffer_uptodate(bh);
 603                 fs_info(sdp, "barrier sync failed - disabling barriers\n");
 604                 set_bit(SDF_NOBARRIERS, &sdp->sd_flags);
 605                 lock_buffer(bh);
 606 skip_barrier:
 607                 get_bh(bh);
 608                 submit_bh(WRITE_SYNC | (1 << BIO_RW_META), bh);
 609                 wait_on_buffer(bh);
 610         }
 611         if (!buffer_uptodate(bh))
 612                 gfs2_io_error_bh(sdp, bh);
 613         brelse(bh);
 614
 615         if (sdp->sd_log_tail != tail)
 616                 log_pull_tail(sdp, tail);
 617         else
 618                 gfs2_assert_withdraw(sdp, !pull);
 619
 620         sdp->sd_log_idle = (tail == sdp->sd_log_flush_head);
 621         gfs2_log_incr_head(sdp);
 622 }
 623
 624 static void log_flush_commit(struct gfs2_sbd *sdp)
 625 {
 626         DEFINE_WAIT(wait);
 627
 628         if (atomic_read(&sdp->sd_log_in_flight)) {
 629                 do {
 630                         prepare_to_wait(&sdp->sd_log_flush_wait, &wait,
 631                                         TASK_UNINTERRUPTIBLE);
 632                         if (atomic_read(&sdp->sd_log_in_flight))
 633                                 io_schedule();
 634                 } while(atomic_read(&sdp->sd_log_in_flight));
 635                 finish_wait(&sdp->sd_log_flush_wait, &wait);
 636         }
 637
 638         log_write_header(sdp, 0, 0);
 639 }
 640
 641 static void gfs2_ordered_write(struct gfs2_sbd *sdp)
 642 {
 643         struct gfs2_bufdata *bd;
 644         struct buffer_head *bh;
 645         LIST_HEAD(written);
 646
 647         gfs2_log_lock(sdp);
 648         while (!list_empty(&sdp->sd_log_le_ordered)) {
 649                 bd = list_entry(sdp->sd_log_le_ordered.next, struct gfs2_bufdata, bd_le.le_list);
 650                 list_move(&bd->bd_le.le_list, &written);
 651                 bh = bd->bd_bh;
 652                 if (!buffer_dirty(bh))
 653                         continue;
 654                 get_bh(bh);
 655                 gfs2_log_unlock(sdp);
 656                 lock_buffer(bh);
 657                 if (buffer_mapped(bh) && test_clear_buffer_dirty(bh)) {
 658                         bh->b_end_io = end_buffer_write_sync;
 659                         submit_bh(WRITE_SYNC_PLUG, bh);
 660                 } else {
 661                         unlock_buffer(bh);
 662                         brelse(bh);
 663                 }
 664                 gfs2_log_lock(sdp);
 665         }
 666         list_splice(&written, &sdp->sd_log_le_ordered);
 667         gfs2_log_unlock(sdp);
 668 }
 669
 670 static void gfs2_ordered_wait(struct gfs2_sbd *sdp)
 671 {
 672         struct gfs2_bufdata *bd;
 673         struct buffer_head *bh;
 674
 675         gfs2_log_lock(sdp);
 676         while (!list_empty(&sdp->sd_log_le_ordered)) {
 677                 bd = list_entry(sdp->sd_log_le_ordered.prev, struct gfs2_bufdata, bd_le.le_list);
 678                 bh = bd->bd_bh;
 679                 if (buffer_locked(bh)) {
 680                         get_bh(bh);
 681                         gfs2_log_unlock(sdp);
 682                         wait_on_buffer(bh);
 683                         brelse(bh);
 684                         gfs2_log_lock(sdp);
 685                         continue;
 686                 }
 687                 list_del_init(&bd->bd_le.le_list);
 688         }
 689         gfs2_log_unlock(sdp);
 690 }
 691
 692 /**
 693  * gfs2_log_flush - flush incore transaction(s)
 694  * @sdp: the filesystem
 695  * @gl: The glock structure to flush.  If NULL, flush the whole incore log
 696  *
 697  */
 698
 699 void __gfs2_log_flush(struct gfs2_sbd *sdp, struct gfs2_glock *gl)
 700 {
 701         struct gfs2_ail *ai;
 702
 703         down_write(&sdp->sd_log_flush_lock);
 704
 705         /* Log might have been flushed while we waited for the flush lock */
 706         if (gl && !test_bit(GLF_LFLUSH, &gl->gl_flags)) {
 707                 up_write(&sdp->sd_log_flush_lock);
 708                 return;
 709         }
 710         trace_gfs2_log_flush(sdp, 1);
 711
 712         ai = kzalloc(sizeof(struct gfs2_ail), GFP_NOFS | __GFP_NOFAIL);
 713         INIT_LIST_HEAD(&ai->ai_ail1_list);
 714         INIT_LIST_HEAD(&ai->ai_ail2_list);
 715
 716         if (sdp->sd_log_num_buf != sdp->sd_log_commited_buf) {
 717                 printk(KERN_INFO "GFS2: log buf %u %u\n", sdp->sd_log_num_buf,
 718                        sdp->sd_log_commited_buf);
 719                 gfs2_assert_withdraw(sdp, 0);
 720         }
 721         if (sdp->sd_log_num_databuf != sdp->sd_log_commited_databuf) {
 722                 printk(KERN_INFO "GFS2: log databuf %u %u\n",
 723                        sdp->sd_log_num_databuf, sdp->sd_log_commited_databuf);
 724                 gfs2_assert_withdraw(sdp, 0);
 725         }
 726         gfs2_assert_withdraw(sdp,
 727                         sdp->sd_log_num_revoke == sdp->sd_log_commited_revoke);
 728
 729         sdp->sd_log_flush_head = sdp->sd_log_head;
 730         sdp->sd_log_flush_wrapped = 0;
 731         ai->ai_first = sdp->sd_log_flush_head;
 732
 733         gfs2_ordered_write(sdp);
 734         lops_before_commit(sdp);
 735         gfs2_ordered_wait(sdp);
 736
 737         if (sdp->sd_log_head != sdp->sd_log_flush_head)
 738                 log_flush_commit(sdp);
 739         else if (sdp->sd_log_tail != current_tail(sdp) && !sdp->sd_log_idle){
 740                 gfs2_log_lock(sdp);
 741                 atomic_dec(&sdp->sd_log_blks_free); /* Adjust for unreserved buffer */
 742                 trace_gfs2_log_blocks(sdp, -1);
 743                 gfs2_log_unlock(sdp);
 744                 log_write_header(sdp, 0, PULL);
 745         }
 746         lops_after_commit(sdp, ai);
 747
 748         gfs2_log_lock(sdp);
 749         sdp->sd_log_head = sdp->sd_log_flush_head;
 750         sdp->sd_log_blks_reserved = 0;
 751         sdp->sd_log_commited_buf = 0;
 752         sdp->sd_log_commited_databuf = 0;
 753         sdp->sd_log_commited_revoke = 0;
 754
 755         if (!list_empty(&ai->ai_ail1_list)) {
 756                 list_add(&ai->ai_list, &sdp->sd_ail1_list);
 757                 ai = NULL;
 758         }
 759         gfs2_log_unlock(sdp);
 760         trace_gfs2_log_flush(sdp, 0);
 761         up_write(&sdp->sd_log_flush_lock);
 762
 763         kfree(ai);
 764 }
 765
 766 static void log_refund(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 767 {
 768         unsigned int reserved;
 769         unsigned int unused;
 770
 771         gfs2_log_lock(sdp);
 772
 773         sdp->sd_log_commited_buf += tr->tr_num_buf_new - tr->tr_num_buf_rm;
 774         sdp->sd_log_commited_databuf += tr->tr_num_databuf_new -
 775                 tr->tr_num_databuf_rm;
 776         gfs2_assert_withdraw(sdp, (((int)sdp->sd_log_commited_buf) >= 0) ||
 777                              (((int)sdp->sd_log_commited_databuf) >= 0));
 778         sdp->sd_log_commited_revoke += tr->tr_num_revoke - tr->tr_num_revoke_rm;
 779         reserved = calc_reserved(sdp);
 780         gfs2_assert_withdraw(sdp, sdp->sd_log_blks_reserved + tr->tr_reserved >= reserved);
 781         unused = sdp->sd_log_blks_reserved - reserved + tr->tr_reserved;
 782         atomic_add(unused, &sdp->sd_log_blks_free);
 783         trace_gfs2_log_blocks(sdp, unused);
 784         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
 785                              sdp->sd_jdesc->jd_blocks);
 786         sdp->sd_log_blks_reserved = reserved;
 787
 788         gfs2_log_unlock(sdp);
 789 }
 790
 791 static void buf_lo_incore_commit(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 792 {
 793         struct list_head *head = &tr->tr_list_buf;
 794         struct gfs2_bufdata *bd;
 795
 796         gfs2_log_lock(sdp);
 797         while (!list_empty(head)) {
 798                 bd = list_entry(head->next, struct gfs2_bufdata, bd_list_tr);
 799                 list_del_init(&bd->bd_list_tr);
 800                 tr->tr_num_buf--;
 801         }
 802         gfs2_log_unlock(sdp);
 803         gfs2_assert_warn(sdp, !tr->tr_num_buf);
 804 }
 805
 806 /**
 807  * gfs2_log_commit - Commit a transaction to the log
 808  * @sdp: the filesystem
 809  * @tr: the transaction
 810  *
 811  * We wake up gfs2_logd if the number of pinned blocks exceed thresh1
 812  * or the total number of used blocks (pinned blocks plus AIL blocks)
 813  * is greater than thresh2.
 814  *
 815  * At mount time thresh1 is 1/3rd of journal size, thresh2 is 2/3rd of
 816  * journal size.
 817  *
 818  * Returns: errno
 819  */
 820
 821 void gfs2_log_commit(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 822 {
 823         log_refund(sdp, tr);
 824         buf_lo_incore_commit(sdp, tr);
 825
 826         up_read(&sdp->sd_log_flush_lock);
 827
 828         if (atomic_read(&sdp->sd_log_pinned) > atomic_read(&sdp->sd_log_thresh1) ||
 829             ((sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free)) >
 830             atomic_read(&sdp->sd_log_thresh2)))
 831                 wake_up(&sdp->sd_logd_waitq);
 832 }
 833
 834 /**
 835  * gfs2_log_shutdown - write a shutdown header into a journal
 836  * @sdp: the filesystem
 837  *
 838  */
 839
 840 void gfs2_log_shutdown(struct gfs2_sbd *sdp)
 841 {
 842         down_write(&sdp->sd_log_flush_lock);
 843
 844         gfs2_assert_withdraw(sdp, !sdp->sd_log_blks_reserved);
 845         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_buf);
 846         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_revoke);
 847         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_rg);
 848         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_databuf);
 849         gfs2_assert_withdraw(sdp, list_empty(&sdp->sd_ail1_list));
 850
 851         sdp->sd_log_flush_head = sdp->sd_log_head;
 852         sdp->sd_log_flush_wrapped = 0;
 853
 854         log_write_header(sdp, GFS2_LOG_HEAD_UNMOUNT,
 855                          (sdp->sd_log_tail == current_tail(sdp)) ? 0 : PULL);
 856
 857         gfs2_assert_warn(sdp, atomic_read(&sdp->sd_log_blks_free) == sdp->sd_jdesc->jd_blocks);
 858         gfs2_assert_warn(sdp, sdp->sd_log_head == sdp->sd_log_tail);
 859         gfs2_assert_warn(sdp, list_empty(&sdp->sd_ail2_list));
 860
 861         sdp->sd_log_head = sdp->sd_log_flush_head;
 862         sdp->sd_log_tail = sdp->sd_log_head;
 863
 864         up_write(&sdp->sd_log_flush_lock);
 865 }
 866
 867
 868 /**
 869  * gfs2_meta_syncfs - sync all the buffers in a filesystem
 870  * @sdp: the filesystem
 871  *
 872  */
 873
 874 void gfs2_meta_syncfs(struct gfs2_sbd *sdp)
 875 {
 876         gfs2_log_flush(sdp, NULL);
 877         for (;;) {
 878                 gfs2_ail1_start(sdp);
 879                 if (gfs2_ail1_empty(sdp, DIO_ALL))
 880                         break;
 881                 msleep(10);
 882         }
 883 }
 884
 885 static inline int gfs2_jrnl_flush_reqd(struct gfs2_sbd *sdp)
 886 {
 887         return (atomic_read(&sdp->sd_log_pinned) >= atomic_read(&sdp->sd_log_thresh1));
 888 }
 889
 890 static inline int gfs2_ail_flush_reqd(struct gfs2_sbd *sdp)
 891 {
 892         unsigned int used_blocks = sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free);
 893         return used_blocks >= atomic_read(&sdp->sd_log_thresh2);
 894 }
 895
 896 /**
 897  * gfs2_logd - Update log tail as Active Items get flushed to in-place blocks
 898  * @sdp: Pointer to GFS2 superblock
 899  *
 900  * Also, periodically check to make sure that we're using the most recent
 901  * journal index.
 902  */
 903
 904 int gfs2_logd(void *data)
 905 {
 906         struct gfs2_sbd *sdp = data;
 907         unsigned long t = 1;
 908         DEFINE_WAIT(wait);
 909         unsigned preflush;
 910
 911         while (!kthread_should_stop()) {
 912
 913                 preflush = atomic_read(&sdp->sd_log_pinned);
 914                 if (gfs2_jrnl_flush_reqd(sdp) || t == 0) {
 915                         gfs2_ail1_empty(sdp, DIO_ALL);
 916                         gfs2_log_flush(sdp, NULL);
 917                         gfs2_ail1_empty(sdp, DIO_ALL);
 918                 }
 919
 920                 if (gfs2_ail_flush_reqd(sdp)) {
 921                         gfs2_ail1_start(sdp);
 922                         io_schedule();
 923                         gfs2_ail1_empty(sdp, 0);
 924                         gfs2_log_flush(sdp, NULL);
 925                         gfs2_ail1_empty(sdp, DIO_ALL);
 926                 }
 927
 928                 wake_up(&sdp->sd_log_waitq);
 929                 t = gfs2_tune_get(sdp, gt_logd_secs) * HZ;
 930                 if (freezing(current))
 931                         refrigerator();
 932
 933                 do {
 934                         prepare_to_wait(&sdp->sd_logd_waitq, &wait,
 935                                         TASK_UNINTERRUPTIBLE);
 936                         if (!gfs2_ail_flush_reqd(sdp) &&
 937                             !gfs2_jrnl_flush_reqd(sdp) &&
 938                             !kthread_should_stop())
 939                                 t = schedule_timeout(t);
 940                 } while(t && !gfs2_ail_flush_reqd(sdp) &&
 941                         !gfs2_jrnl_flush_reqd(sdp) &&
 942                         !kthread_should_stop());
 943                 finish_wait(&sdp->sd_logd_waitq, &wait);
 944         }
 945
 946         return 0;
 947 }
 948