Pileus Git - ~andy/linux/blob - 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 oldest_tr = 1;
 215         int ret;
 216
 217         spin_lock(&sdp->sd_ail_lock);
 218         list_for_each_entry_safe_reverse(tr, s, &sdp->sd_ail1_list, tr_list) {
 219                 gfs2_ail1_empty_one(sdp, tr);
 220                 if (list_empty(&tr->tr_ail1_list) && oldest_tr)
 221                         list_move(&tr->tr_list, &sdp->sd_ail2_list);
 222                 else
 223                         oldest_tr = 0;
 224         }
 225         ret = list_empty(&sdp->sd_ail1_list);
 226         spin_unlock(&sdp->sd_ail_lock);
 227
 228         return ret;
 229 }
 230
 231 static void gfs2_ail1_wait(struct gfs2_sbd *sdp)
 232 {
 233         struct gfs2_trans *tr;
 234         struct gfs2_bufdata *bd;
 235         struct buffer_head *bh;
 236
 237         spin_lock(&sdp->sd_ail_lock);
 238         list_for_each_entry_reverse(tr, &sdp->sd_ail1_list, tr_list) {
 239                 list_for_each_entry(bd, &tr->tr_ail1_list, bd_ail_st_list) {
 240                         bh = bd->bd_bh;
 241                         if (!buffer_locked(bh))
 242                                 continue;
 243                         get_bh(bh);
 244                         spin_unlock(&sdp->sd_ail_lock);
 245                         wait_on_buffer(bh);
 246                         brelse(bh);
 247                         return;
 248                 }
 249         }
 250         spin_unlock(&sdp->sd_ail_lock);
 251 }
 252
 253 /**
 254  * gfs2_ail2_empty_one - Check whether or not a trans in the AIL has been synced
 255  * @sdp: the filesystem
 256  * @ai: the AIL entry
 257  *
 258  */
 259
 260 static void gfs2_ail2_empty_one(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 261 {
 262         struct list_head *head = &tr->tr_ail2_list;
 263         struct gfs2_bufdata *bd;
 264
 265         while (!list_empty(head)) {
 266                 bd = list_entry(head->prev, struct gfs2_bufdata,
 267                                 bd_ail_st_list);
 268                 gfs2_assert(sdp, bd->bd_tr == tr);
 269                 gfs2_remove_from_ail(bd);
 270         }
 271 }
 272
 273 static void ail2_empty(struct gfs2_sbd *sdp, unsigned int new_tail)
 274 {
 275         struct gfs2_trans *tr, *safe;
 276         unsigned int old_tail = sdp->sd_log_tail;
 277         int wrap = (new_tail < old_tail);
 278         int a, b, rm;
 279
 280         spin_lock(&sdp->sd_ail_lock);
 281
 282         list_for_each_entry_safe(tr, safe, &sdp->sd_ail2_list, tr_list) {
 283                 a = (old_tail <= tr->tr_first);
 284                 b = (tr->tr_first < new_tail);
 285                 rm = (wrap) ? (a || b) : (a && b);
 286                 if (!rm)
 287                         continue;
 288
 289                 gfs2_ail2_empty_one(sdp, tr);
 290                 list_del(&tr->tr_list);
 291                 gfs2_assert_warn(sdp, list_empty(&tr->tr_ail1_list));
 292                 gfs2_assert_warn(sdp, list_empty(&tr->tr_ail2_list));
 293                 kfree(tr);
 294         }
 295
 296         spin_unlock(&sdp->sd_ail_lock);
 297 }
 298
 299 /**
 300  * gfs2_log_reserve - Make a log reservation
 301  * @sdp: The GFS2 superblock
 302  * @blks: The number of blocks to reserve
 303  *
 304  * Note that we never give out the last few blocks of the journal. Thats
 305  * due to the fact that there is a small number of header blocks
 306  * associated with each log flush. The exact number can't be known until
 307  * flush time, so we ensure that we have just enough free blocks at all
 308  * times to avoid running out during a log flush.
 309  *
 310  * We no longer flush the log here, instead we wake up logd to do that
 311  * for us. To avoid the thundering herd and to ensure that we deal fairly
 312  * with queued waiters, we use an exclusive wait. This means that when we
 313  * get woken with enough journal space to get our reservation, we need to
 314  * wake the next waiter on the list.
 315  *
 316  * Returns: errno
 317  */
 318
 319 int gfs2_log_reserve(struct gfs2_sbd *sdp, unsigned int blks)
 320 {
 321         unsigned reserved_blks = 7 * (4096 / sdp->sd_vfs->s_blocksize);
 322         unsigned wanted = blks + reserved_blks;
 323         DEFINE_WAIT(wait);
 324         int did_wait = 0;
 325         unsigned int free_blocks;
 326
 327         if (gfs2_assert_warn(sdp, blks) ||
 328             gfs2_assert_warn(sdp, blks <= sdp->sd_jdesc->jd_blocks))
 329                 return -EINVAL;
 330 retry:
 331         free_blocks = atomic_read(&sdp->sd_log_blks_free);
 332         if (unlikely(free_blocks <= wanted)) {
 333                 do {
 334                         prepare_to_wait_exclusive(&sdp->sd_log_waitq, &wait,
 335                                         TASK_UNINTERRUPTIBLE);
 336                         wake_up(&sdp->sd_logd_waitq);
 337                         did_wait = 1;
 338                         if (atomic_read(&sdp->sd_log_blks_free) <= wanted)
 339                                 io_schedule();
 340                         free_blocks = atomic_read(&sdp->sd_log_blks_free);
 341                 } while(free_blocks <= wanted);
 342                 finish_wait(&sdp->sd_log_waitq, &wait);
 343         }
 344         if (atomic_cmpxchg(&sdp->sd_log_blks_free, free_blocks,
 345                                 free_blocks - blks) != free_blocks)
 346                 goto retry;
 347         trace_gfs2_log_blocks(sdp, -blks);
 348
 349         /*
 350          * If we waited, then so might others, wake them up _after_ we get
 351          * our share of the log.
 352          */
 353         if (unlikely(did_wait))
 354                 wake_up(&sdp->sd_log_waitq);
 355
 356         down_read(&sdp->sd_log_flush_lock);
 357
 358         return 0;
 359 }
 360
 361 /**
 362  * log_distance - Compute distance between two journal blocks
 363  * @sdp: The GFS2 superblock
 364  * @newer: The most recent journal block of the pair
 365  * @older: The older journal block of the pair
 366  *
 367  *   Compute the distance (in the journal direction) between two
 368  *   blocks in the journal
 369  *
 370  * Returns: the distance in blocks
 371  */
 372
 373 static inline unsigned int log_distance(struct gfs2_sbd *sdp, unsigned int newer,
 374                                         unsigned int older)
 375 {
 376         int dist;
 377
 378         dist = newer - older;
 379         if (dist < 0)
 380                 dist += sdp->sd_jdesc->jd_blocks;
 381
 382         return dist;
 383 }
 384
 385 /**
 386  * calc_reserved - Calculate the number of blocks to reserve when
 387  *                 refunding a transaction's unused buffers.
 388  * @sdp: The GFS2 superblock
 389  *
 390  * This is complex.  We need to reserve room for all our currently used
 391  * metadata buffers (e.g. normal file I/O rewriting file time stamps) and
 392  * all our journaled data buffers for journaled files (e.g. files in the
 393  * meta_fs like rindex, or files for which chattr +j was done.)
 394  * If we don't reserve enough space, gfs2_log_refund and gfs2_log_flush
 395  * will count it as free space (sd_log_blks_free) and corruption will follow.
 396  *
 397  * We can have metadata bufs and jdata bufs in the same journal.  So each
 398  * type gets its own log header, for which we need to reserve a block.
 399  * In fact, each type has the potential for needing more than one header
 400  * in cases where we have more buffers than will fit on a journal page.
 401  * Metadata journal entries take up half the space of journaled buffer entries.
 402  * Thus, metadata entries have buf_limit (502) and journaled buffers have
 403  * databuf_limit (251) before they cause a wrap around.
 404  *
 405  * Also, we need to reserve blocks for revoke journal entries and one for an
 406  * overall header for the lot.
 407  *
 408  * Returns: the number of blocks reserved
 409  */
 410 static unsigned int calc_reserved(struct gfs2_sbd *sdp)
 411 {
 412         unsigned int reserved = 0;
 413         unsigned int mbuf_limit, metabufhdrs_needed;
 414         unsigned int dbuf_limit, databufhdrs_needed;
 415         unsigned int revokes = 0;
 416
 417         mbuf_limit = buf_limit(sdp);
 418         metabufhdrs_needed = (sdp->sd_log_commited_buf +
 419                               (mbuf_limit - 1)) / mbuf_limit;
 420         dbuf_limit = databuf_limit(sdp);
 421         databufhdrs_needed = (sdp->sd_log_commited_databuf +
 422                               (dbuf_limit - 1)) / dbuf_limit;
 423
 424         if (sdp->sd_log_commited_revoke > 0)
 425                 revokes = gfs2_struct2blk(sdp, sdp->sd_log_commited_revoke,
 426                                           sizeof(u64));
 427
 428         reserved = sdp->sd_log_commited_buf + metabufhdrs_needed +
 429                 sdp->sd_log_commited_databuf + databufhdrs_needed +
 430                 revokes;
 431         /* One for the overall header */
 432         if (reserved)
 433                 reserved++;
 434         return reserved;
 435 }
 436
 437 static unsigned int current_tail(struct gfs2_sbd *sdp)
 438 {
 439         struct gfs2_trans *tr;
 440         unsigned int tail;
 441
 442         spin_lock(&sdp->sd_ail_lock);
 443
 444         if (list_empty(&sdp->sd_ail1_list)) {
 445                 tail = sdp->sd_log_head;
 446         } else {
 447                 tr = list_entry(sdp->sd_ail1_list.prev, struct gfs2_trans,
 448                                 tr_list);
 449                 tail = tr->tr_first;
 450         }
 451
 452         spin_unlock(&sdp->sd_ail_lock);
 453
 454         return tail;
 455 }
 456
 457 static void log_pull_tail(struct gfs2_sbd *sdp, unsigned int new_tail)
 458 {
 459         unsigned int dist = log_distance(sdp, new_tail, sdp->sd_log_tail);
 460
 461         ail2_empty(sdp, new_tail);
 462
 463         atomic_add(dist, &sdp->sd_log_blks_free);
 464         trace_gfs2_log_blocks(sdp, dist);
 465         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
 466                              sdp->sd_jdesc->jd_blocks);
 467
 468         sdp->sd_log_tail = new_tail;
 469 }
 470
 471
 472 static void log_flush_wait(struct gfs2_sbd *sdp)
 473 {
 474         DEFINE_WAIT(wait);
 475
 476         if (atomic_read(&sdp->sd_log_in_flight)) {
 477                 do {
 478                         prepare_to_wait(&sdp->sd_log_flush_wait, &wait,
 479                                         TASK_UNINTERRUPTIBLE);
 480                         if (atomic_read(&sdp->sd_log_in_flight))
 481                                 io_schedule();
 482                 } while(atomic_read(&sdp->sd_log_in_flight));
 483                 finish_wait(&sdp->sd_log_flush_wait, &wait);
 484         }
 485 }
 486
 487 static int ip_cmp(void *priv, struct list_head *a, struct list_head *b)
 488 {
 489         struct gfs2_inode *ipa, *ipb;
 490
 491         ipa = list_entry(a, struct gfs2_inode, i_ordered);
 492         ipb = list_entry(b, struct gfs2_inode, i_ordered);
 493
 494         if (ipa->i_no_addr < ipb->i_no_addr)
 495                 return -1;
 496         if (ipa->i_no_addr > ipb->i_no_addr)
 497                 return 1;
 498         return 0;
 499 }
 500
 501 static void gfs2_ordered_write(struct gfs2_sbd *sdp)
 502 {
 503         struct gfs2_inode *ip;
 504         LIST_HEAD(written);
 505
 506         spin_lock(&sdp->sd_ordered_lock);
 507         list_sort(NULL, &sdp->sd_log_le_ordered, &ip_cmp);
 508         while (!list_empty(&sdp->sd_log_le_ordered)) {
 509                 ip = list_entry(sdp->sd_log_le_ordered.next, struct gfs2_inode, i_ordered);
 510                 list_move(&ip->i_ordered, &written);
 511                 if (ip->i_inode.i_mapping->nrpages == 0)
 512                         continue;
 513                 spin_unlock(&sdp->sd_ordered_lock);
 514                 filemap_fdatawrite(ip->i_inode.i_mapping);
 515                 spin_lock(&sdp->sd_ordered_lock);
 516         }
 517         list_splice(&written, &sdp->sd_log_le_ordered);
 518         spin_unlock(&sdp->sd_ordered_lock);
 519 }
 520
 521 static void gfs2_ordered_wait(struct gfs2_sbd *sdp)
 522 {
 523         struct gfs2_inode *ip;
 524
 525         spin_lock(&sdp->sd_ordered_lock);
 526         while (!list_empty(&sdp->sd_log_le_ordered)) {
 527                 ip = list_entry(sdp->sd_log_le_ordered.next, struct gfs2_inode, i_ordered);
 528                 list_del(&ip->i_ordered);
 529                 WARN_ON(!test_and_clear_bit(GIF_ORDERED, &ip->i_flags));
 530                 if (ip->i_inode.i_mapping->nrpages == 0)
 531                         continue;
 532                 spin_unlock(&sdp->sd_ordered_lock);
 533                 filemap_fdatawait(ip->i_inode.i_mapping);
 534                 spin_lock(&sdp->sd_ordered_lock);
 535         }
 536         spin_unlock(&sdp->sd_ordered_lock);
 537 }
 538
 539 void gfs2_ordered_del_inode(struct gfs2_inode *ip)
 540 {
 541         struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
 542
 543         spin_lock(&sdp->sd_ordered_lock);
 544         if (test_and_clear_bit(GIF_ORDERED, &ip->i_flags))
 545                 list_del(&ip->i_ordered);
 546         spin_unlock(&sdp->sd_ordered_lock);
 547 }
 548
 549 void gfs2_add_revoke(struct gfs2_sbd *sdp, struct gfs2_bufdata *bd)
 550 {
 551         struct buffer_head *bh = bd->bd_bh;
 552         struct gfs2_glock *gl = bd->bd_gl;
 553
 554         bh->b_private = NULL;
 555         bd->bd_blkno = bh->b_blocknr;
 556         gfs2_remove_from_ail(bd); /* drops ref on bh */
 557         bd->bd_bh = NULL;
 558         bd->bd_ops = &gfs2_revoke_lops;
 559         sdp->sd_log_num_revoke++;
 560         atomic_inc(&gl->gl_revokes);
 561         set_bit(GLF_LFLUSH, &gl->gl_flags);
 562         list_add(&bd->bd_list, &sdp->sd_log_le_revoke);
 563 }
 564
 565 void gfs2_write_revokes(struct gfs2_sbd *sdp)
 566 {
 567         struct gfs2_trans *tr;
 568         struct gfs2_bufdata *bd, *tmp;
 569         int have_revokes = 0;
 570         int max_revokes = (sdp->sd_sb.sb_bsize - sizeof(struct gfs2_log_descriptor)) / sizeof(u64);
 571
 572         gfs2_ail1_empty(sdp);
 573         spin_lock(&sdp->sd_ail_lock);
 574         list_for_each_entry(tr, &sdp->sd_ail1_list, tr_list) {
 575                 list_for_each_entry(bd, &tr->tr_ail2_list, bd_ail_st_list) {
 576                         if (list_empty(&bd->bd_list)) {
 577                                 have_revokes = 1;
 578                                 goto done;
 579                         }
 580                 }
 581         }
 582 done:
 583         spin_unlock(&sdp->sd_ail_lock);
 584         if (have_revokes == 0)
 585                 return;
 586         while (sdp->sd_log_num_revoke > max_revokes)
 587                 max_revokes += (sdp->sd_sb.sb_bsize - sizeof(struct gfs2_meta_header)) / sizeof(u64);
 588         max_revokes -= sdp->sd_log_num_revoke;
 589         if (!sdp->sd_log_num_revoke) {
 590                 atomic_dec(&sdp->sd_log_blks_free);
 591                 /* If no blocks have been reserved, we need to also
 592                  * reserve a block for the header */
 593                 if (!sdp->sd_log_blks_reserved)
 594                         atomic_dec(&sdp->sd_log_blks_free);
 595         }
 596         gfs2_log_lock(sdp);
 597         spin_lock(&sdp->sd_ail_lock);
 598         list_for_each_entry(tr, &sdp->sd_ail1_list, tr_list) {
 599                 list_for_each_entry_safe(bd, tmp, &tr->tr_ail2_list, bd_ail_st_list) {
 600                         if (max_revokes == 0)
 601                                 goto out_of_blocks;
 602                         if (!list_empty(&bd->bd_list))
 603                                 continue;
 604                         gfs2_add_revoke(sdp, bd);
 605                         max_revokes--;
 606                 }
 607         }
 608 out_of_blocks:
 609         spin_unlock(&sdp->sd_ail_lock);
 610         gfs2_log_unlock(sdp);
 611
 612         if (!sdp->sd_log_num_revoke) {
 613                 atomic_inc(&sdp->sd_log_blks_free);
 614                 if (!sdp->sd_log_blks_reserved)
 615                         atomic_inc(&sdp->sd_log_blks_free);
 616         }
 617 }
 618
 619 /**
 620  * log_write_header - Get and initialize a journal header buffer
 621  * @sdp: The GFS2 superblock
 622  *
 623  * Returns: the initialized log buffer descriptor
 624  */
 625
 626 static void log_write_header(struct gfs2_sbd *sdp, u32 flags)
 627 {
 628         struct gfs2_log_header *lh;
 629         unsigned int tail;
 630         u32 hash;
 631         int rw = WRITE_FLUSH_FUA | REQ_META;
 632         struct page *page = mempool_alloc(gfs2_page_pool, GFP_NOIO);
 633         lh = page_address(page);
 634         clear_page(lh);
 635
 636         tail = current_tail(sdp);
 637
 638         lh->lh_header.mh_magic = cpu_to_be32(GFS2_MAGIC);
 639         lh->lh_header.mh_type = cpu_to_be32(GFS2_METATYPE_LH);
 640         lh->lh_header.__pad0 = cpu_to_be64(0);
 641         lh->lh_header.mh_format = cpu_to_be32(GFS2_FORMAT_LH);
 642         lh->lh_header.mh_jid = cpu_to_be32(sdp->sd_jdesc->jd_jid);
 643         lh->lh_sequence = cpu_to_be64(sdp->sd_log_sequence++);
 644         lh->lh_flags = cpu_to_be32(flags);
 645         lh->lh_tail = cpu_to_be32(tail);
 646         lh->lh_blkno = cpu_to_be32(sdp->sd_log_flush_head);
 647         hash = gfs2_disk_hash(page_address(page), sizeof(struct gfs2_log_header));
 648         lh->lh_hash = cpu_to_be32(hash);
 649
 650         if (test_bit(SDF_NOBARRIERS, &sdp->sd_flags)) {
 651                 gfs2_ordered_wait(sdp);
 652                 log_flush_wait(sdp);
 653                 rw = WRITE_SYNC | REQ_META | REQ_PRIO;
 654         }
 655
 656         sdp->sd_log_idle = (tail == sdp->sd_log_flush_head);
 657         gfs2_log_write_page(sdp, page);
 658         gfs2_log_flush_bio(sdp, rw);
 659         log_flush_wait(sdp);
 660
 661         if (sdp->sd_log_tail != tail)
 662                 log_pull_tail(sdp, tail);
 663 }
 664
 665 /**
 666  * gfs2_log_flush - flush incore transaction(s)
 667  * @sdp: the filesystem
 668  * @gl: The glock structure to flush.  If NULL, flush the whole incore log
 669  *
 670  */
 671
 672 void gfs2_log_flush(struct gfs2_sbd *sdp, struct gfs2_glock *gl)
 673 {
 674         struct gfs2_trans *tr;
 675
 676         down_write(&sdp->sd_log_flush_lock);
 677
 678         /* Log might have been flushed while we waited for the flush lock */
 679         if (gl && !test_bit(GLF_LFLUSH, &gl->gl_flags)) {
 680                 up_write(&sdp->sd_log_flush_lock);
 681                 return;
 682         }
 683         trace_gfs2_log_flush(sdp, 1);
 684
 685         tr = sdp->sd_log_tr;
 686         if (tr) {
 687                 sdp->sd_log_tr = NULL;
 688                 INIT_LIST_HEAD(&tr->tr_ail1_list);
 689                 INIT_LIST_HEAD(&tr->tr_ail2_list);
 690         }
 691
 692         if (sdp->sd_log_num_buf != sdp->sd_log_commited_buf) {
 693                 printk(KERN_INFO "GFS2: log buf %u %u\n", sdp->sd_log_num_buf,
 694                        sdp->sd_log_commited_buf);
 695                 gfs2_assert_withdraw(sdp, 0);
 696         }
 697         if (sdp->sd_log_num_databuf != sdp->sd_log_commited_databuf) {
 698                 printk(KERN_INFO "GFS2: log databuf %u %u\n",
 699                        sdp->sd_log_num_databuf, sdp->sd_log_commited_databuf);
 700                 gfs2_assert_withdraw(sdp, 0);
 701         }
 702         gfs2_assert_withdraw(sdp,
 703                         sdp->sd_log_num_revoke == sdp->sd_log_commited_revoke);
 704
 705         sdp->sd_log_flush_head = sdp->sd_log_head;
 706         sdp->sd_log_flush_wrapped = 0;
 707         if (tr)
 708                 tr->tr_first = sdp->sd_log_flush_head;
 709
 710         gfs2_ordered_write(sdp);
 711         lops_before_commit(sdp);
 712         gfs2_log_flush_bio(sdp, WRITE);
 713
 714         if (sdp->sd_log_head != sdp->sd_log_flush_head) {
 715                 log_write_header(sdp, 0);
 716         } else if (sdp->sd_log_tail != current_tail(sdp) && !sdp->sd_log_idle){
 717                 atomic_dec(&sdp->sd_log_blks_free); /* Adjust for unreserved buffer */
 718                 trace_gfs2_log_blocks(sdp, -1);
 719                 log_write_header(sdp, 0);
 720         }
 721         lops_after_commit(sdp, tr);
 722
 723         gfs2_log_lock(sdp);
 724         sdp->sd_log_head = sdp->sd_log_flush_head;
 725         sdp->sd_log_blks_reserved = 0;
 726         sdp->sd_log_commited_buf = 0;
 727         sdp->sd_log_commited_databuf = 0;
 728         sdp->sd_log_commited_revoke = 0;
 729
 730         spin_lock(&sdp->sd_ail_lock);
 731         if (tr && !list_empty(&tr->tr_ail1_list)) {
 732                 list_add(&tr->tr_list, &sdp->sd_ail1_list);
 733                 tr = NULL;
 734         }
 735         spin_unlock(&sdp->sd_ail_lock);
 736         gfs2_log_unlock(sdp);
 737         trace_gfs2_log_flush(sdp, 0);
 738         up_write(&sdp->sd_log_flush_lock);
 739
 740         kfree(tr);
 741 }
 742
 743 static void log_refund(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 744 {
 745         unsigned int reserved;
 746         unsigned int unused;
 747
 748         gfs2_log_lock(sdp);
 749
 750         sdp->sd_log_commited_buf += tr->tr_num_buf_new - tr->tr_num_buf_rm;
 751         sdp->sd_log_commited_databuf += tr->tr_num_databuf_new -
 752                 tr->tr_num_databuf_rm;
 753         gfs2_assert_withdraw(sdp, (((int)sdp->sd_log_commited_buf) >= 0) ||
 754                              (((int)sdp->sd_log_commited_databuf) >= 0));
 755         sdp->sd_log_commited_revoke += tr->tr_num_revoke - tr->tr_num_revoke_rm;
 756         reserved = calc_reserved(sdp);
 757         gfs2_assert_withdraw(sdp, sdp->sd_log_blks_reserved + tr->tr_reserved >= reserved);
 758         unused = sdp->sd_log_blks_reserved - reserved + tr->tr_reserved;
 759         atomic_add(unused, &sdp->sd_log_blks_free);
 760         trace_gfs2_log_blocks(sdp, unused);
 761         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
 762                              sdp->sd_jdesc->jd_blocks);
 763         sdp->sd_log_blks_reserved = reserved;
 764
 765         if (sdp->sd_log_tr == NULL &&
 766             (tr->tr_num_buf_new || tr->tr_num_databuf_new)) {
 767                 gfs2_assert_withdraw(sdp, tr->tr_t_gh.gh_gl);
 768                 sdp->sd_log_tr = tr;
 769                 tr->tr_attached = 1;
 770         }
 771         gfs2_log_unlock(sdp);
 772 }
 773
 774 /**
 775  * gfs2_log_commit - Commit a transaction to the log
 776  * @sdp: the filesystem
 777  * @tr: the transaction
 778  *
 779  * We wake up gfs2_logd if the number of pinned blocks exceed thresh1
 780  * or the total number of used blocks (pinned blocks plus AIL blocks)
 781  * is greater than thresh2.
 782  *
 783  * At mount time thresh1 is 1/3rd of journal size, thresh2 is 2/3rd of
 784  * journal size.
 785  *
 786  * Returns: errno
 787  */
 788
 789 void gfs2_log_commit(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 790 {
 791         log_refund(sdp, tr);
 792
 793         if (atomic_read(&sdp->sd_log_pinned) > atomic_read(&sdp->sd_log_thresh1) ||
 794             ((sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free)) >
 795             atomic_read(&sdp->sd_log_thresh2)))
 796                 wake_up(&sdp->sd_logd_waitq);
 797 }
 798
 799 /**
 800  * gfs2_log_shutdown - write a shutdown header into a journal
 801  * @sdp: the filesystem
 802  *
 803  */
 804
 805 void gfs2_log_shutdown(struct gfs2_sbd *sdp)
 806 {
 807         down_write(&sdp->sd_log_flush_lock);
 808
 809         gfs2_assert_withdraw(sdp, !sdp->sd_log_blks_reserved);
 810         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_buf);
 811         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_revoke);
 812         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_rg);
 813         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_databuf);
 814         gfs2_assert_withdraw(sdp, list_empty(&sdp->sd_ail1_list));
 815
 816         sdp->sd_log_flush_head = sdp->sd_log_head;
 817         sdp->sd_log_flush_wrapped = 0;
 818
 819         log_write_header(sdp, GFS2_LOG_HEAD_UNMOUNT);
 820
 821         gfs2_assert_warn(sdp, atomic_read(&sdp->sd_log_blks_free) == sdp->sd_jdesc->jd_blocks);
 822         gfs2_assert_warn(sdp, sdp->sd_log_head == sdp->sd_log_tail);
 823         gfs2_assert_warn(sdp, list_empty(&sdp->sd_ail2_list));
 824
 825         sdp->sd_log_head = sdp->sd_log_flush_head;
 826         sdp->sd_log_tail = sdp->sd_log_head;
 827
 828         up_write(&sdp->sd_log_flush_lock);
 829 }
 830
 831
 832 /**
 833  * gfs2_meta_syncfs - sync all the buffers in a filesystem
 834  * @sdp: the filesystem
 835  *
 836  */
 837
 838 void gfs2_meta_syncfs(struct gfs2_sbd *sdp)
 839 {
 840         gfs2_log_flush(sdp, NULL);
 841         for (;;) {
 842                 gfs2_ail1_start(sdp);
 843                 gfs2_ail1_wait(sdp);
 844                 if (gfs2_ail1_empty(sdp))
 845                         break;
 846         }
 847         gfs2_log_flush(sdp, NULL);
 848 }
 849
 850 static inline int gfs2_jrnl_flush_reqd(struct gfs2_sbd *sdp)
 851 {
 852         return (atomic_read(&sdp->sd_log_pinned) >= atomic_read(&sdp->sd_log_thresh1));
 853 }
 854
 855 static inline int gfs2_ail_flush_reqd(struct gfs2_sbd *sdp)
 856 {
 857         unsigned int used_blocks = sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free);
 858         return used_blocks >= atomic_read(&sdp->sd_log_thresh2);
 859 }
 860
 861 /**
 862  * gfs2_logd - Update log tail as Active Items get flushed to in-place blocks
 863  * @sdp: Pointer to GFS2 superblock
 864  *
 865  * Also, periodically check to make sure that we're using the most recent
 866  * journal index.
 867  */
 868
 869 int gfs2_logd(void *data)
 870 {
 871         struct gfs2_sbd *sdp = data;
 872         unsigned long t = 1;
 873         DEFINE_WAIT(wait);
 874
 875         while (!kthread_should_stop()) {
 876
 877                 if (gfs2_jrnl_flush_reqd(sdp) || t == 0) {
 878                         gfs2_ail1_empty(sdp);
 879                         gfs2_log_flush(sdp, NULL);
 880                 }
 881
 882                 if (gfs2_ail_flush_reqd(sdp)) {
 883                         gfs2_ail1_start(sdp);
 884                         gfs2_ail1_wait(sdp);
 885                         gfs2_ail1_empty(sdp);
 886                         gfs2_log_flush(sdp, NULL);
 887                 }
 888
 889                 if (!gfs2_ail_flush_reqd(sdp))
 890                         wake_up(&sdp->sd_log_waitq);
 891
 892                 t = gfs2_tune_get(sdp, gt_logd_secs) * HZ;
 893
 894                 try_to_freeze();
 895
 896                 do {
 897                         prepare_to_wait(&sdp->sd_logd_waitq, &wait,
 898                                         TASK_INTERRUPTIBLE);
 899                         if (!gfs2_ail_flush_reqd(sdp) &&
 900                             !gfs2_jrnl_flush_reqd(sdp) &&
 901                             !kthread_should_stop())
 902                                 t = schedule_timeout(t);
 903                 } while(t && !gfs2_ail_flush_reqd(sdp) &&
 904                         !gfs2_jrnl_flush_reqd(sdp) &&
 905                         !kthread_should_stop());
 906                 finish_wait(&sdp->sd_logd_waitq, &wait);
 907         }
 908
 909         return 0;
 910 }
 911