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/writeback.h>
  22 #include <linux/list_sort.h>
  23
  24 #include "gfs2.h"
  25 #include "incore.h"
  26 #include "bmap.h"
  27 #include "glock.h"
  28 #include "log.h"
  29 #include "lops.h"
  30 #include "meta_io.h"
  31 #include "util.h"
  32 #include "dir.h"
  33 #include "trace_gfs2.h"
  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 ail 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_tr = 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  * @wbc: The writeback control structure
  87  * @ai: The ail structure
  88  *
  89  */
  90
  91 static int gfs2_ail1_start_one(struct gfs2_sbd *sdp,
  92                                struct writeback_control *wbc,
  93                                struct gfs2_trans *tr)
  94 __releases(&sdp->sd_ail_lock)
  95 __acquires(&sdp->sd_ail_lock)
  96 {
  97         struct gfs2_glock *gl = NULL;
  98         struct address_space *mapping;
  99         struct gfs2_bufdata *bd, *s;
 100         struct buffer_head *bh;
 101
 102         list_for_each_entry_safe_reverse(bd, s, &tr->tr_ail1_list, bd_ail_st_list) {
 103                 bh = bd->bd_bh;
 104
 105                 gfs2_assert(sdp, bd->bd_tr == tr);
 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, &tr->tr_ail2_list);
 111                         continue;
 112                 }
 113
 114                 if (!buffer_dirty(bh))
 115                         continue;
 116                 if (gl == bd->bd_gl)
 117                         continue;
 118                 gl = bd->bd_gl;
 119                 list_move(&bd->bd_ail_st_list, &tr->tr_ail1_list);
 120                 mapping = bh->b_page->mapping;
 121                 if (!mapping)
 122                         continue;
 123                 spin_unlock(&sdp->sd_ail_lock);
 124                 generic_writepages(mapping, wbc);
 125                 spin_lock(&sdp->sd_ail_lock);
 126                 if (wbc->nr_to_write <= 0)
 127                         break;
 128                 return 1;
 129         }
 130
 131         return 0;
 132 }
 133
 134
 135 /**
 136  * gfs2_ail1_flush - start writeback of some ail1 entries
 137  * @sdp: The super block
 138  * @wbc: The writeback control structure
 139  *
 140  * Writes back some ail1 entries, according to the limits in the
 141  * writeback control structure
 142  */
 143
 144 void gfs2_ail1_flush(struct gfs2_sbd *sdp, struct writeback_control *wbc)
 145 {
 146         struct list_head *head = &sdp->sd_ail1_list;
 147         struct gfs2_trans *tr;
 148
 149         trace_gfs2_ail_flush(sdp, wbc, 1);
 150         spin_lock(&sdp->sd_ail_lock);
 151 restart:
 152         list_for_each_entry_reverse(tr, head, tr_list) {
 153                 if (wbc->nr_to_write <= 0)
 154                         break;
 155                 if (gfs2_ail1_start_one(sdp, wbc, tr))
 156                         goto restart;
 157         }
 158         spin_unlock(&sdp->sd_ail_lock);
 159         trace_gfs2_ail_flush(sdp, wbc, 0);
 160 }
 161
 162 /**
 163  * gfs2_ail1_start - start writeback of all ail1 entries
 164  * @sdp: The superblock
 165  */
 166
 167 static void gfs2_ail1_start(struct gfs2_sbd *sdp)
 168 {
 169         struct writeback_control wbc = {
 170                 .sync_mode = WB_SYNC_NONE,
 171                 .nr_to_write = LONG_MAX,
 172                 .range_start = 0,
 173                 .range_end = LLONG_MAX,
 174         };
 175
 176         return gfs2_ail1_flush(sdp, &wbc);
 177 }
 178
 179 /**
 180  * gfs2_ail1_empty_one - Check whether or not a trans in the AIL has been synced
 181  * @sdp: the filesystem
 182  * @ai: the AIL entry
 183  *
 184  */
 185
 186 static void gfs2_ail1_empty_one(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 187 {
 188         struct gfs2_bufdata *bd, *s;
 189         struct buffer_head *bh;
 190
 191         list_for_each_entry_safe_reverse(bd, s, &tr->tr_ail1_list,
 192                                          bd_ail_st_list) {
 193                 bh = bd->bd_bh;
 194                 gfs2_assert(sdp, bd->bd_tr == tr);
 195                 if (buffer_busy(bh))
 196                         continue;
 197                 if (!buffer_uptodate(bh))
 198                         gfs2_io_error_bh(sdp, bh);
 199                 list_move(&bd->bd_ail_st_list, &tr->tr_ail2_list);
 200         }
 201
 202 }
 203
 204 /**
 205  * gfs2_ail1_empty - Try to empty the ail1 lists
 206  * @sdp: The superblock
 207  *
 208  * Tries to empty the ail1 lists, starting with the oldest first
 209  */
 210
 211 static int gfs2_ail1_empty(struct gfs2_sbd *sdp)
 212 {
 213         struct gfs2_trans *tr, *s;
 214         int ret;
 215
 216         spin_lock(&sdp->sd_ail_lock);
 217         list_for_each_entry_safe_reverse(tr, s, &sdp->sd_ail1_list, tr_list) {
 218                 gfs2_ail1_empty_one(sdp, tr);
 219                 if (list_empty(&tr->tr_ail1_list))
 220                         list_move(&tr->tr_list, &sdp->sd_ail2_list);
 221                 else
 222                         break;
 223         }
 224         ret = list_empty(&sdp->sd_ail1_list);
 225         spin_unlock(&sdp->sd_ail_lock);
 226
 227         return ret;
 228 }
 229
 230 static void gfs2_ail1_wait(struct gfs2_sbd *sdp)
 231 {
 232         struct gfs2_trans *tr;
 233         struct gfs2_bufdata *bd;
 234         struct buffer_head *bh;
 235
 236         spin_lock(&sdp->sd_ail_lock);
 237         list_for_each_entry_reverse(tr, &sdp->sd_ail1_list, tr_list) {
 238                 list_for_each_entry(bd, &tr->tr_ail1_list, bd_ail_st_list) {
 239                         bh = bd->bd_bh;
 240                         if (!buffer_locked(bh))
 241                                 continue;
 242                         get_bh(bh);
 243                         spin_unlock(&sdp->sd_ail_lock);
 244                         wait_on_buffer(bh);
 245                         brelse(bh);
 246                         return;
 247                 }
 248         }
 249         spin_unlock(&sdp->sd_ail_lock);
 250 }
 251
 252 /**
 253  * gfs2_ail2_empty_one - Check whether or not a trans in the AIL has been synced
 254  * @sdp: the filesystem
 255  * @ai: the AIL entry
 256  *
 257  */
 258
 259 static void gfs2_ail2_empty_one(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 260 {
 261         struct list_head *head = &tr->tr_ail2_list;
 262         struct gfs2_bufdata *bd;
 263
 264         while (!list_empty(head)) {
 265                 bd = list_entry(head->prev, struct gfs2_bufdata,
 266                                 bd_ail_st_list);
 267                 gfs2_assert(sdp, bd->bd_tr == tr);
 268                 gfs2_remove_from_ail(bd);
 269         }
 270 }
 271
 272 static void ail2_empty(struct gfs2_sbd *sdp, unsigned int new_tail)
 273 {
 274         struct gfs2_trans *tr, *safe;
 275         unsigned int old_tail = sdp->sd_log_tail;
 276         int wrap = (new_tail < old_tail);
 277         int a, b, rm;
 278
 279         spin_lock(&sdp->sd_ail_lock);
 280
 281         list_for_each_entry_safe(tr, safe, &sdp->sd_ail2_list, tr_list) {
 282                 a = (old_tail <= tr->tr_first);
 283                 b = (tr->tr_first < new_tail);
 284                 rm = (wrap) ? (a || b) : (a && b);
 285                 if (!rm)
 286                         continue;
 287
 288                 gfs2_ail2_empty_one(sdp, tr);
 289                 list_del(&tr->tr_list);
 290                 gfs2_assert_warn(sdp, list_empty(&tr->tr_ail1_list));
 291                 gfs2_assert_warn(sdp, list_empty(&tr->tr_ail2_list));
 292                 kfree(tr);
 293         }
 294
 295         spin_unlock(&sdp->sd_ail_lock);
 296 }
 297
 298 /**
 299  * gfs2_log_reserve - Make a log reservation
 300  * @sdp: The GFS2 superblock
 301  * @blks: The number of blocks to reserve
 302  *
 303  * Note that we never give out the last few blocks of the journal. Thats
 304  * due to the fact that there is a small number of header blocks
 305  * associated with each log flush. The exact number can't be known until
 306  * flush time, so we ensure that we have just enough free blocks at all
 307  * times to avoid running out during a log flush.
 308  *
 309  * We no longer flush the log here, instead we wake up logd to do that
 310  * for us. To avoid the thundering herd and to ensure that we deal fairly
 311  * with queued waiters, we use an exclusive wait. This means that when we
 312  * get woken with enough journal space to get our reservation, we need to
 313  * wake the next waiter on the list.
 314  *
 315  * Returns: errno
 316  */
 317
 318 int gfs2_log_reserve(struct gfs2_sbd *sdp, unsigned int blks)
 319 {
 320         unsigned reserved_blks = 6 * (4096 / sdp->sd_vfs->s_blocksize);
 321         unsigned wanted = blks + reserved_blks;
 322         DEFINE_WAIT(wait);
 323         int did_wait = 0;
 324         unsigned int free_blocks;
 325
 326         if (gfs2_assert_warn(sdp, blks) ||
 327             gfs2_assert_warn(sdp, blks <= sdp->sd_jdesc->jd_blocks))
 328                 return -EINVAL;
 329 retry:
 330         free_blocks = atomic_read(&sdp->sd_log_blks_free);
 331         if (unlikely(free_blocks <= wanted)) {
 332                 do {
 333                         prepare_to_wait_exclusive(&sdp->sd_log_waitq, &wait,
 334                                         TASK_UNINTERRUPTIBLE);
 335                         wake_up(&sdp->sd_logd_waitq);
 336                         did_wait = 1;
 337                         if (atomic_read(&sdp->sd_log_blks_free) <= wanted)
 338                                 io_schedule();
 339                         free_blocks = atomic_read(&sdp->sd_log_blks_free);
 340                 } while(free_blocks <= wanted);
 341                 finish_wait(&sdp->sd_log_waitq, &wait);
 342         }
 343         if (atomic_cmpxchg(&sdp->sd_log_blks_free, free_blocks,
 344                                 free_blocks - blks) != free_blocks)
 345                 goto retry;
 346         trace_gfs2_log_blocks(sdp, -blks);
 347
 348         /*
 349          * If we waited, then so might others, wake them up _after_ we get
 350          * our share of the log.
 351          */
 352         if (unlikely(did_wait))
 353                 wake_up(&sdp->sd_log_waitq);
 354
 355         down_read(&sdp->sd_log_flush_lock);
 356
 357         return 0;
 358 }
 359
 360 /**
 361  * log_distance - Compute distance between two journal blocks
 362  * @sdp: The GFS2 superblock
 363  * @newer: The most recent journal block of the pair
 364  * @older: The older journal block of the pair
 365  *
 366  *   Compute the distance (in the journal direction) between two
 367  *   blocks in the journal
 368  *
 369  * Returns: the distance in blocks
 370  */
 371
 372 static inline unsigned int log_distance(struct gfs2_sbd *sdp, unsigned int newer,
 373                                         unsigned int older)
 374 {
 375         int dist;
 376
 377         dist = newer - older;
 378         if (dist < 0)
 379                 dist += sdp->sd_jdesc->jd_blocks;
 380
 381         return dist;
 382 }
 383
 384 /**
 385  * calc_reserved - Calculate the number of blocks to reserve when
 386  *                 refunding a transaction's unused buffers.
 387  * @sdp: The GFS2 superblock
 388  *
 389  * This is complex.  We need to reserve room for all our currently used
 390  * metadata buffers (e.g. normal file I/O rewriting file time stamps) and
 391  * all our journaled data buffers for journaled files (e.g. files in the
 392  * meta_fs like rindex, or files for which chattr +j was done.)
 393  * If we don't reserve enough space, gfs2_log_refund and gfs2_log_flush
 394  * will count it as free space (sd_log_blks_free) and corruption will follow.
 395  *
 396  * We can have metadata bufs and jdata bufs in the same journal.  So each
 397  * type gets its own log header, for which we need to reserve a block.
 398  * In fact, each type has the potential for needing more than one header
 399  * in cases where we have more buffers than will fit on a journal page.
 400  * Metadata journal entries take up half the space of journaled buffer entries.
 401  * Thus, metadata entries have buf_limit (502) and journaled buffers have
 402  * databuf_limit (251) before they cause a wrap around.
 403  *
 404  * Also, we need to reserve blocks for revoke journal entries and one for an
 405  * overall header for the lot.
 406  *
 407  * Returns: the number of blocks reserved
 408  */
 409 static unsigned int calc_reserved(struct gfs2_sbd *sdp)
 410 {
 411         unsigned int reserved = 0;
 412         unsigned int mbuf_limit, metabufhdrs_needed;
 413         unsigned int dbuf_limit, databufhdrs_needed;
 414         unsigned int revokes = 0;
 415
 416         mbuf_limit = buf_limit(sdp);
 417         metabufhdrs_needed = (sdp->sd_log_commited_buf +
 418                               (mbuf_limit - 1)) / mbuf_limit;
 419         dbuf_limit = databuf_limit(sdp);
 420         databufhdrs_needed = (sdp->sd_log_commited_databuf +
 421                               (dbuf_limit - 1)) / dbuf_limit;
 422
 423         if (sdp->sd_log_commited_revoke > 0)
 424                 revokes = gfs2_struct2blk(sdp, sdp->sd_log_commited_revoke,
 425                                           sizeof(u64));
 426
 427         reserved = sdp->sd_log_commited_buf + metabufhdrs_needed +
 428                 sdp->sd_log_commited_databuf + databufhdrs_needed +
 429                 revokes;
 430         /* One for the overall header */
 431         if (reserved)
 432                 reserved++;
 433         return reserved;
 434 }
 435
 436 static unsigned int current_tail(struct gfs2_sbd *sdp)
 437 {
 438         struct gfs2_trans *tr;
 439         unsigned int tail;
 440
 441         spin_lock(&sdp->sd_ail_lock);
 442
 443         if (list_empty(&sdp->sd_ail1_list)) {
 444                 tail = sdp->sd_log_head;
 445         } else {
 446                 tr = list_entry(sdp->sd_ail1_list.prev, struct gfs2_trans,
 447                                 tr_list);
 448                 tail = tr->tr_first;
 449         }
 450
 451         spin_unlock(&sdp->sd_ail_lock);
 452
 453         return tail;
 454 }
 455
 456 static void log_pull_tail(struct gfs2_sbd *sdp, unsigned int new_tail)
 457 {
 458         unsigned int dist = log_distance(sdp, new_tail, sdp->sd_log_tail);
 459
 460         ail2_empty(sdp, new_tail);
 461
 462         atomic_add(dist, &sdp->sd_log_blks_free);
 463         trace_gfs2_log_blocks(sdp, dist);
 464         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
 465                              sdp->sd_jdesc->jd_blocks);
 466
 467         sdp->sd_log_tail = new_tail;
 468 }
 469
 470
 471 static void log_flush_wait(struct gfs2_sbd *sdp)
 472 {
 473         DEFINE_WAIT(wait);
 474
 475         if (atomic_read(&sdp->sd_log_in_flight)) {
 476                 do {
 477                         prepare_to_wait(&sdp->sd_log_flush_wait, &wait,
 478                                         TASK_UNINTERRUPTIBLE);
 479                         if (atomic_read(&sdp->sd_log_in_flight))
 480                                 io_schedule();
 481                 } while(atomic_read(&sdp->sd_log_in_flight));
 482                 finish_wait(&sdp->sd_log_flush_wait, &wait);
 483         }
 484 }
 485
 486 static int ip_cmp(void *priv, struct list_head *a, struct list_head *b)
 487 {
 488         struct gfs2_inode *ipa, *ipb;
 489
 490         ipa = list_entry(a, struct gfs2_inode, i_ordered);
 491         ipb = list_entry(b, struct gfs2_inode, i_ordered);
 492
 493         if (ipa->i_no_addr < ipb->i_no_addr)
 494                 return -1;
 495         if (ipa->i_no_addr > ipb->i_no_addr)
 496                 return 1;
 497         return 0;
 498 }
 499
 500 static void gfs2_ordered_write(struct gfs2_sbd *sdp)
 501 {
 502         struct gfs2_inode *ip;
 503         LIST_HEAD(written);
 504
 505         spin_lock(&sdp->sd_ordered_lock);
 506         list_sort(NULL, &sdp->sd_log_le_ordered, &ip_cmp);
 507         while (!list_empty(&sdp->sd_log_le_ordered)) {
 508                 ip = list_entry(sdp->sd_log_le_ordered.next, struct gfs2_inode, i_ordered);
 509                 list_move(&ip->i_ordered, &written);
 510                 if (ip->i_inode.i_mapping->nrpages == 0)
 511                         continue;
 512                 spin_unlock(&sdp->sd_ordered_lock);
 513                 filemap_fdatawrite(ip->i_inode.i_mapping);
 514                 spin_lock(&sdp->sd_ordered_lock);
 515         }
 516         list_splice(&written, &sdp->sd_log_le_ordered);
 517         spin_unlock(&sdp->sd_ordered_lock);
 518 }
 519
 520 static void gfs2_ordered_wait(struct gfs2_sbd *sdp)
 521 {
 522         struct gfs2_inode *ip;
 523
 524         spin_lock(&sdp->sd_ordered_lock);
 525         while (!list_empty(&sdp->sd_log_le_ordered)) {
 526                 ip = list_entry(sdp->sd_log_le_ordered.next, struct gfs2_inode, i_ordered);
 527                 list_del(&ip->i_ordered);
 528                 WARN_ON(!test_and_clear_bit(GIF_ORDERED, &ip->i_flags));
 529                 if (ip->i_inode.i_mapping->nrpages == 0)
 530                         continue;
 531                 spin_unlock(&sdp->sd_ordered_lock);
 532                 filemap_fdatawait(ip->i_inode.i_mapping);
 533                 spin_lock(&sdp->sd_ordered_lock);
 534         }
 535         spin_unlock(&sdp->sd_ordered_lock);
 536 }
 537
 538 void gfs2_ordered_del_inode(struct gfs2_inode *ip)
 539 {
 540         struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
 541
 542         spin_lock(&sdp->sd_ordered_lock);
 543         if (test_and_clear_bit(GIF_ORDERED, &ip->i_flags))
 544                 list_del(&ip->i_ordered);
 545         spin_unlock(&sdp->sd_ordered_lock);
 546 }
 547
 548 /**
 549  * log_write_header - Get and initialize a journal header buffer
 550  * @sdp: The GFS2 superblock
 551  *
 552  * Returns: the initialized log buffer descriptor
 553  */
 554
 555 static void log_write_header(struct gfs2_sbd *sdp, u32 flags)
 556 {
 557         struct gfs2_log_header *lh;
 558         unsigned int tail;
 559         u32 hash;
 560         int rw = WRITE_FLUSH_FUA | REQ_META;
 561         struct page *page = mempool_alloc(gfs2_page_pool, GFP_NOIO);
 562         lh = page_address(page);
 563         clear_page(lh);
 564
 565         gfs2_ail1_empty(sdp);
 566         tail = current_tail(sdp);
 567
 568         lh->lh_header.mh_magic = cpu_to_be32(GFS2_MAGIC);
 569         lh->lh_header.mh_type = cpu_to_be32(GFS2_METATYPE_LH);
 570         lh->lh_header.__pad0 = cpu_to_be64(0);
 571         lh->lh_header.mh_format = cpu_to_be32(GFS2_FORMAT_LH);
 572         lh->lh_header.mh_jid = cpu_to_be32(sdp->sd_jdesc->jd_jid);
 573         lh->lh_sequence = cpu_to_be64(sdp->sd_log_sequence++);
 574         lh->lh_flags = cpu_to_be32(flags);
 575         lh->lh_tail = cpu_to_be32(tail);
 576         lh->lh_blkno = cpu_to_be32(sdp->sd_log_flush_head);
 577         hash = gfs2_disk_hash(page_address(page), sizeof(struct gfs2_log_header));
 578         lh->lh_hash = cpu_to_be32(hash);
 579
 580         if (test_bit(SDF_NOBARRIERS, &sdp->sd_flags)) {
 581                 gfs2_ordered_wait(sdp);
 582                 log_flush_wait(sdp);
 583                 rw = WRITE_SYNC | REQ_META | REQ_PRIO;
 584         }
 585
 586         sdp->sd_log_idle = (tail == sdp->sd_log_flush_head);
 587         gfs2_log_write_page(sdp, page);
 588         gfs2_log_flush_bio(sdp, rw);
 589         log_flush_wait(sdp);
 590
 591         if (sdp->sd_log_tail != tail)
 592                 log_pull_tail(sdp, tail);
 593 }
 594
 595 /**
 596  * gfs2_log_flush - flush incore transaction(s)
 597  * @sdp: the filesystem
 598  * @gl: The glock structure to flush.  If NULL, flush the whole incore log
 599  *
 600  */
 601
 602 void gfs2_log_flush(struct gfs2_sbd *sdp, struct gfs2_glock *gl)
 603 {
 604         struct gfs2_trans *tr;
 605
 606         down_write(&sdp->sd_log_flush_lock);
 607
 608         /* Log might have been flushed while we waited for the flush lock */
 609         if (gl && !test_bit(GLF_LFLUSH, &gl->gl_flags)) {
 610                 up_write(&sdp->sd_log_flush_lock);
 611                 return;
 612         }
 613         trace_gfs2_log_flush(sdp, 1);
 614
 615         tr = sdp->sd_log_tr;
 616         if (tr) {
 617                 sdp->sd_log_tr = NULL;
 618                 INIT_LIST_HEAD(&tr->tr_ail1_list);
 619                 INIT_LIST_HEAD(&tr->tr_ail2_list);
 620         }
 621
 622         if (sdp->sd_log_num_buf != sdp->sd_log_commited_buf) {
 623                 printk(KERN_INFO "GFS2: log buf %u %u\n", sdp->sd_log_num_buf,
 624                        sdp->sd_log_commited_buf);
 625                 gfs2_assert_withdraw(sdp, 0);
 626         }
 627         if (sdp->sd_log_num_databuf != sdp->sd_log_commited_databuf) {
 628                 printk(KERN_INFO "GFS2: log databuf %u %u\n",
 629                        sdp->sd_log_num_databuf, sdp->sd_log_commited_databuf);
 630                 gfs2_assert_withdraw(sdp, 0);
 631         }
 632         gfs2_assert_withdraw(sdp,
 633                         sdp->sd_log_num_revoke == sdp->sd_log_commited_revoke);
 634
 635         sdp->sd_log_flush_head = sdp->sd_log_head;
 636         sdp->sd_log_flush_wrapped = 0;
 637         if (tr)
 638                 tr->tr_first = sdp->sd_log_flush_head;
 639
 640         gfs2_ordered_write(sdp);
 641         lops_before_commit(sdp);
 642         gfs2_log_flush_bio(sdp, WRITE);
 643
 644         if (sdp->sd_log_head != sdp->sd_log_flush_head) {
 645                 log_write_header(sdp, 0);
 646         } else if (sdp->sd_log_tail != current_tail(sdp) && !sdp->sd_log_idle){
 647                 atomic_dec(&sdp->sd_log_blks_free); /* Adjust for unreserved buffer */
 648                 trace_gfs2_log_blocks(sdp, -1);
 649                 log_write_header(sdp, 0);
 650         }
 651         lops_after_commit(sdp, tr);
 652
 653         gfs2_log_lock(sdp);
 654         sdp->sd_log_head = sdp->sd_log_flush_head;
 655         sdp->sd_log_blks_reserved = 0;
 656         sdp->sd_log_commited_buf = 0;
 657         sdp->sd_log_commited_databuf = 0;
 658         sdp->sd_log_commited_revoke = 0;
 659
 660         spin_lock(&sdp->sd_ail_lock);
 661         if (tr && !list_empty(&tr->tr_ail1_list)) {
 662                 list_add(&tr->tr_list, &sdp->sd_ail1_list);
 663                 tr = NULL;
 664         }
 665         spin_unlock(&sdp->sd_ail_lock);
 666         gfs2_log_unlock(sdp);
 667         trace_gfs2_log_flush(sdp, 0);
 668         up_write(&sdp->sd_log_flush_lock);
 669
 670         kfree(tr);
 671 }
 672
 673 static void log_refund(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 674 {
 675         unsigned int reserved;
 676         unsigned int unused;
 677
 678         gfs2_log_lock(sdp);
 679
 680         sdp->sd_log_commited_buf += tr->tr_num_buf_new - tr->tr_num_buf_rm;
 681         sdp->sd_log_commited_databuf += tr->tr_num_databuf_new -
 682                 tr->tr_num_databuf_rm;
 683         gfs2_assert_withdraw(sdp, (((int)sdp->sd_log_commited_buf) >= 0) ||
 684                              (((int)sdp->sd_log_commited_databuf) >= 0));
 685         sdp->sd_log_commited_revoke += tr->tr_num_revoke - tr->tr_num_revoke_rm;
 686         reserved = calc_reserved(sdp);
 687         gfs2_assert_withdraw(sdp, sdp->sd_log_blks_reserved + tr->tr_reserved >= reserved);
 688         unused = sdp->sd_log_blks_reserved - reserved + tr->tr_reserved;
 689         atomic_add(unused, &sdp->sd_log_blks_free);
 690         trace_gfs2_log_blocks(sdp, unused);
 691         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
 692                              sdp->sd_jdesc->jd_blocks);
 693         sdp->sd_log_blks_reserved = reserved;
 694
 695         if (sdp->sd_log_tr == NULL &&
 696             (tr->tr_num_buf_new || tr->tr_num_databuf_new)) {
 697                 gfs2_assert_withdraw(sdp, tr->tr_t_gh.gh_gl);
 698                 sdp->sd_log_tr = tr;
 699                 tr->tr_attached = 1;
 700         }
 701         gfs2_log_unlock(sdp);
 702 }
 703
 704 /**
 705  * gfs2_log_commit - Commit a transaction to the log
 706  * @sdp: the filesystem
 707  * @tr: the transaction
 708  *
 709  * We wake up gfs2_logd if the number of pinned blocks exceed thresh1
 710  * or the total number of used blocks (pinned blocks plus AIL blocks)
 711  * is greater than thresh2.
 712  *
 713  * At mount time thresh1 is 1/3rd of journal size, thresh2 is 2/3rd of
 714  * journal size.
 715  *
 716  * Returns: errno
 717  */
 718
 719 void gfs2_log_commit(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 720 {
 721         log_refund(sdp, tr);
 722
 723         if (atomic_read(&sdp->sd_log_pinned) > atomic_read(&sdp->sd_log_thresh1) ||
 724             ((sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free)) >
 725             atomic_read(&sdp->sd_log_thresh2)))
 726                 wake_up(&sdp->sd_logd_waitq);
 727 }
 728
 729 /**
 730  * gfs2_log_shutdown - write a shutdown header into a journal
 731  * @sdp: the filesystem
 732  *
 733  */
 734
 735 void gfs2_log_shutdown(struct gfs2_sbd *sdp)
 736 {
 737         down_write(&sdp->sd_log_flush_lock);
 738
 739         gfs2_assert_withdraw(sdp, !sdp->sd_log_blks_reserved);
 740         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_buf);
 741         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_revoke);
 742         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_rg);
 743         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_databuf);
 744         gfs2_assert_withdraw(sdp, list_empty(&sdp->sd_ail1_list));
 745
 746         sdp->sd_log_flush_head = sdp->sd_log_head;
 747         sdp->sd_log_flush_wrapped = 0;
 748
 749         log_write_header(sdp, GFS2_LOG_HEAD_UNMOUNT);
 750
 751         gfs2_assert_warn(sdp, atomic_read(&sdp->sd_log_blks_free) == sdp->sd_jdesc->jd_blocks);
 752         gfs2_assert_warn(sdp, sdp->sd_log_head == sdp->sd_log_tail);
 753         gfs2_assert_warn(sdp, list_empty(&sdp->sd_ail2_list));
 754
 755         sdp->sd_log_head = sdp->sd_log_flush_head;
 756         sdp->sd_log_tail = sdp->sd_log_head;
 757
 758         up_write(&sdp->sd_log_flush_lock);
 759 }
 760
 761
 762 /**
 763  * gfs2_meta_syncfs - sync all the buffers in a filesystem
 764  * @sdp: the filesystem
 765  *
 766  */
 767
 768 void gfs2_meta_syncfs(struct gfs2_sbd *sdp)
 769 {
 770         gfs2_log_flush(sdp, NULL);
 771         for (;;) {
 772                 gfs2_ail1_start(sdp);
 773                 gfs2_ail1_wait(sdp);
 774                 if (gfs2_ail1_empty(sdp))
 775                         break;
 776         }
 777         gfs2_log_flush(sdp, NULL);
 778 }
 779
 780 static inline int gfs2_jrnl_flush_reqd(struct gfs2_sbd *sdp)
 781 {
 782         return (atomic_read(&sdp->sd_log_pinned) >= atomic_read(&sdp->sd_log_thresh1));
 783 }
 784
 785 static inline int gfs2_ail_flush_reqd(struct gfs2_sbd *sdp)
 786 {
 787         unsigned int used_blocks = sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free);
 788         return used_blocks >= atomic_read(&sdp->sd_log_thresh2);
 789 }
 790
 791 /**
 792  * gfs2_logd - Update log tail as Active Items get flushed to in-place blocks
 793  * @sdp: Pointer to GFS2 superblock
 794  *
 795  * Also, periodically check to make sure that we're using the most recent
 796  * journal index.
 797  */
 798
 799 int gfs2_logd(void *data)
 800 {
 801         struct gfs2_sbd *sdp = data;
 802         unsigned long t = 1;
 803         DEFINE_WAIT(wait);
 804
 805         while (!kthread_should_stop()) {
 806
 807                 if (gfs2_jrnl_flush_reqd(sdp) || t == 0) {
 808                         gfs2_ail1_empty(sdp);
 809                         gfs2_log_flush(sdp, NULL);
 810                 }
 811
 812                 if (gfs2_ail_flush_reqd(sdp)) {
 813                         gfs2_ail1_start(sdp);
 814                         gfs2_ail1_wait(sdp);
 815                         gfs2_ail1_empty(sdp);
 816                         gfs2_log_flush(sdp, NULL);
 817                 }
 818
 819                 if (!gfs2_ail_flush_reqd(sdp))
 820                         wake_up(&sdp->sd_log_waitq);
 821
 822                 t = gfs2_tune_get(sdp, gt_logd_secs) * HZ;
 823
 824                 try_to_freeze();
 825
 826                 do {
 827                         prepare_to_wait(&sdp->sd_logd_waitq, &wait,
 828                                         TASK_INTERRUPTIBLE);
 829                         if (!gfs2_ail_flush_reqd(sdp) &&
 830                             !gfs2_jrnl_flush_reqd(sdp) &&
 831                             !kthread_should_stop())
 832                                 t = schedule_timeout(t);
 833                 } while(t && !gfs2_ail_flush_reqd(sdp) &&
 834                         !gfs2_jrnl_flush_reqd(sdp) &&
 835                         !kthread_should_stop());
 836                 finish_wait(&sdp->sd_logd_waitq, &wait);
 837         }
 838
 839         return 0;
 840 }
 841