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
  23 #include "gfs2.h"
  24 #include "incore.h"
  25 #include "bmap.h"
  26 #include "glock.h"
  27 #include "log.h"
  28 #include "lops.h"
  29 #include "meta_io.h"
  30 #include "util.h"
  31 #include "dir.h"
  32 #include "trace_gfs2.h"
  33
  34 #define PULL 1
  35
  36 /**
  37  * gfs2_struct2blk - compute stuff
  38  * @sdp: the filesystem
  39  * @nstruct: the number of structures
  40  * @ssize: the size of the structures
  41  *
  42  * Compute the number of log descriptor blocks needed to hold a certain number
  43  * of structures of a certain size.
  44  *
  45  * Returns: the number of blocks needed (minimum is always 1)
  46  */
  47
  48 unsigned int gfs2_struct2blk(struct gfs2_sbd *sdp, unsigned int nstruct,
  49                              unsigned int ssize)
  50 {
  51         unsigned int blks;
  52         unsigned int first, second;
  53
  54         blks = 1;
  55         first = (sdp->sd_sb.sb_bsize - sizeof(struct gfs2_log_descriptor)) / ssize;
  56
  57         if (nstruct > first) {
  58                 second = (sdp->sd_sb.sb_bsize -
  59                           sizeof(struct gfs2_meta_header)) / ssize;
  60                 blks += DIV_ROUND_UP(nstruct - first, second);
  61         }
  62
  63         return blks;
  64 }
  65
  66 /**
  67  * gfs2_remove_from_ail - Remove an entry from the ail lists, updating counters
  68  * @mapping: The associated mapping (maybe NULL)
  69  * @bd: The gfs2_bufdata to remove
  70  *
  71  * The ail lock _must_ be held when calling this function
  72  *
  73  */
  74
  75 void gfs2_remove_from_ail(struct gfs2_bufdata *bd)
  76 {
  77         bd->bd_ail = NULL;
  78         list_del_init(&bd->bd_ail_st_list);
  79         list_del_init(&bd->bd_ail_gl_list);
  80         atomic_dec(&bd->bd_gl->gl_ail_count);
  81         brelse(bd->bd_bh);
  82 }
  83
  84 /**
  85  * gfs2_ail1_start_one - Start I/O on a part of the AIL
  86  * @sdp: the filesystem
  87  * @wbc: The writeback control structure
  88  * @ai: The ail structure
  89  *
  90  */
  91
  92 static int gfs2_ail1_start_one(struct gfs2_sbd *sdp,
  93                                struct writeback_control *wbc,
  94                                struct gfs2_ail *ai)
  95 __releases(&sdp->sd_ail_lock)
  96 __acquires(&sdp->sd_ail_lock)
  97 {
  98         struct gfs2_glock *gl = NULL;
  99         struct address_space *mapping;
 100         struct gfs2_bufdata *bd, *s;
 101         struct buffer_head *bh;
 102
 103         list_for_each_entry_safe_reverse(bd, s, &ai->ai_ail1_list, bd_ail_st_list) {
 104                 bh = bd->bd_bh;
 105
 106                 gfs2_assert(sdp, bd->bd_ail == ai);
 107
 108                 if (!buffer_busy(bh)) {
 109                         if (!buffer_uptodate(bh))
 110                                 gfs2_io_error_bh(sdp, bh);
 111                         list_move(&bd->bd_ail_st_list, &ai->ai_ail2_list);
 112                         continue;
 113                 }
 114
 115                 if (!buffer_dirty(bh))
 116                         continue;
 117                 if (gl == bd->bd_gl)
 118                         continue;
 119                 gl = bd->bd_gl;
 120                 list_move(&bd->bd_ail_st_list, &ai->ai_ail1_list);
 121                 mapping = bh->b_page->mapping;
 122                 if (!mapping)
 123                         continue;
 124                 spin_unlock(&sdp->sd_ail_lock);
 125                 generic_writepages(mapping, wbc);
 126                 spin_lock(&sdp->sd_ail_lock);
 127                 if (wbc->nr_to_write <= 0)
 128                         break;
 129                 return 1;
 130         }
 131
 132         return 0;
 133 }
 134
 135
 136 /**
 137  * gfs2_ail1_flush - start writeback of some ail1 entries
 138  * @sdp: The super block
 139  * @wbc: The writeback control structure
 140  *
 141  * Writes back some ail1 entries, according to the limits in the
 142  * writeback control structure
 143  */
 144
 145 void gfs2_ail1_flush(struct gfs2_sbd *sdp, struct writeback_control *wbc)
 146 {
 147         struct list_head *head = &sdp->sd_ail1_list;
 148         struct gfs2_ail *ai;
 149
 150         trace_gfs2_ail_flush(sdp, wbc, 1);
 151         spin_lock(&sdp->sd_ail_lock);
 152 restart:
 153         list_for_each_entry_reverse(ai, head, ai_list) {
 154                 if (wbc->nr_to_write <= 0)
 155                         break;
 156                 if (gfs2_ail1_start_one(sdp, wbc, ai))
 157                         goto restart;
 158         }
 159         spin_unlock(&sdp->sd_ail_lock);
 160         trace_gfs2_ail_flush(sdp, wbc, 0);
 161 }
 162
 163 /**
 164  * gfs2_ail1_start - start writeback of all ail1 entries
 165  * @sdp: The superblock
 166  */
 167
 168 static void gfs2_ail1_start(struct gfs2_sbd *sdp)
 169 {
 170         struct writeback_control wbc = {
 171                 .sync_mode = WB_SYNC_NONE,
 172                 .nr_to_write = LONG_MAX,
 173                 .range_start = 0,
 174                 .range_end = LLONG_MAX,
 175         };
 176
 177         return gfs2_ail1_flush(sdp, &wbc);
 178 }
 179
 180 /**
 181  * gfs2_ail1_empty_one - Check whether or not a trans in the AIL has been synced
 182  * @sdp: the filesystem
 183  * @ai: the AIL entry
 184  *
 185  */
 186
 187 static void gfs2_ail1_empty_one(struct gfs2_sbd *sdp, struct gfs2_ail *ai)
 188 {
 189         struct gfs2_bufdata *bd, *s;
 190         struct buffer_head *bh;
 191
 192         list_for_each_entry_safe_reverse(bd, s, &ai->ai_ail1_list,
 193                                          bd_ail_st_list) {
 194                 bh = bd->bd_bh;
 195                 gfs2_assert(sdp, bd->bd_ail == ai);
 196                 if (buffer_busy(bh))
 197                         continue;
 198                 if (!buffer_uptodate(bh))
 199                         gfs2_io_error_bh(sdp, bh);
 200                 list_move(&bd->bd_ail_st_list, &ai->ai_ail2_list);
 201         }
 202
 203 }
 204
 205 /**
 206  * gfs2_ail1_empty - Try to empty the ail1 lists
 207  * @sdp: The superblock
 208  *
 209  * Tries to empty the ail1 lists, starting with the oldest first
 210  */
 211
 212 static int gfs2_ail1_empty(struct gfs2_sbd *sdp)
 213 {
 214         struct gfs2_ail *ai, *s;
 215         int ret;
 216
 217         spin_lock(&sdp->sd_ail_lock);
 218         list_for_each_entry_safe_reverse(ai, s, &sdp->sd_ail1_list, ai_list) {
 219                 gfs2_ail1_empty_one(sdp, ai);
 220                 if (list_empty(&ai->ai_ail1_list))
 221                         list_move(&ai->ai_list, &sdp->sd_ail2_list);
 222                 else
 223                         break;
 224         }
 225         ret = list_empty(&sdp->sd_ail1_list);
 226         spin_unlock(&sdp->sd_ail_lock);
 227
 228         return ret;
 229 }
 230
 231
 232 /**
 233  * gfs2_ail2_empty_one - Check whether or not a trans in the AIL has been synced
 234  * @sdp: the filesystem
 235  * @ai: the AIL entry
 236  *
 237  */
 238
 239 static void gfs2_ail2_empty_one(struct gfs2_sbd *sdp, struct gfs2_ail *ai)
 240 {
 241         struct list_head *head = &ai->ai_ail2_list;
 242         struct gfs2_bufdata *bd;
 243
 244         while (!list_empty(head)) {
 245                 bd = list_entry(head->prev, struct gfs2_bufdata,
 246                                 bd_ail_st_list);
 247                 gfs2_assert(sdp, bd->bd_ail == ai);
 248                 gfs2_remove_from_ail(bd);
 249         }
 250 }
 251
 252 static void ail2_empty(struct gfs2_sbd *sdp, unsigned int new_tail)
 253 {
 254         struct gfs2_ail *ai, *safe;
 255         unsigned int old_tail = sdp->sd_log_tail;
 256         int wrap = (new_tail < old_tail);
 257         int a, b, rm;
 258
 259         spin_lock(&sdp->sd_ail_lock);
 260
 261         list_for_each_entry_safe(ai, safe, &sdp->sd_ail2_list, ai_list) {
 262                 a = (old_tail <= ai->ai_first);
 263                 b = (ai->ai_first < new_tail);
 264                 rm = (wrap) ? (a || b) : (a && b);
 265                 if (!rm)
 266                         continue;
 267
 268                 gfs2_ail2_empty_one(sdp, ai);
 269                 list_del(&ai->ai_list);
 270                 gfs2_assert_warn(sdp, list_empty(&ai->ai_ail1_list));
 271                 gfs2_assert_warn(sdp, list_empty(&ai->ai_ail2_list));
 272                 kfree(ai);
 273         }
 274
 275         spin_unlock(&sdp->sd_ail_lock);
 276 }
 277
 278 /**
 279  * gfs2_log_reserve - Make a log reservation
 280  * @sdp: The GFS2 superblock
 281  * @blks: The number of blocks to reserve
 282  *
 283  * Note that we never give out the last few blocks of the journal. Thats
 284  * due to the fact that there is a small number of header blocks
 285  * associated with each log flush. The exact number can't be known until
 286  * flush time, so we ensure that we have just enough free blocks at all
 287  * times to avoid running out during a log flush.
 288  *
 289  * We no longer flush the log here, instead we wake up logd to do that
 290  * for us. To avoid the thundering herd and to ensure that we deal fairly
 291  * with queued waiters, we use an exclusive wait. This means that when we
 292  * get woken with enough journal space to get our reservation, we need to
 293  * wake the next waiter on the list.
 294  *
 295  * Returns: errno
 296  */
 297
 298 int gfs2_log_reserve(struct gfs2_sbd *sdp, unsigned int blks)
 299 {
 300         unsigned reserved_blks = 6 * (4096 / sdp->sd_vfs->s_blocksize);
 301         unsigned wanted = blks + reserved_blks;
 302         DEFINE_WAIT(wait);
 303         int did_wait = 0;
 304         unsigned int free_blocks;
 305
 306         if (gfs2_assert_warn(sdp, blks) ||
 307             gfs2_assert_warn(sdp, blks <= sdp->sd_jdesc->jd_blocks))
 308                 return -EINVAL;
 309 retry:
 310         free_blocks = atomic_read(&sdp->sd_log_blks_free);
 311         if (unlikely(free_blocks <= wanted)) {
 312                 do {
 313                         prepare_to_wait_exclusive(&sdp->sd_log_waitq, &wait,
 314                                         TASK_UNINTERRUPTIBLE);
 315                         wake_up(&sdp->sd_logd_waitq);
 316                         did_wait = 1;
 317                         if (atomic_read(&sdp->sd_log_blks_free) <= wanted)
 318                                 io_schedule();
 319                         free_blocks = atomic_read(&sdp->sd_log_blks_free);
 320                 } while(free_blocks <= wanted);
 321                 finish_wait(&sdp->sd_log_waitq, &wait);
 322         }
 323         if (atomic_cmpxchg(&sdp->sd_log_blks_free, free_blocks,
 324                                 free_blocks - blks) != free_blocks)
 325                 goto retry;
 326         trace_gfs2_log_blocks(sdp, -blks);
 327
 328         /*
 329          * If we waited, then so might others, wake them up _after_ we get
 330          * our share of the log.
 331          */
 332         if (unlikely(did_wait))
 333                 wake_up(&sdp->sd_log_waitq);
 334
 335         down_read(&sdp->sd_log_flush_lock);
 336
 337         return 0;
 338 }
 339
 340 static u64 log_bmap(struct gfs2_sbd *sdp, unsigned int lbn)
 341 {
 342         struct gfs2_journal_extent *je;
 343
 344         list_for_each_entry(je, &sdp->sd_jdesc->extent_list, extent_list) {
 345                 if (lbn >= je->lblock && lbn < je->lblock + je->blocks)
 346                         return je->dblock + lbn - je->lblock;
 347         }
 348
 349         return -1;
 350 }
 351
 352 /**
 353  * log_distance - Compute distance between two journal blocks
 354  * @sdp: The GFS2 superblock
 355  * @newer: The most recent journal block of the pair
 356  * @older: The older journal block of the pair
 357  *
 358  *   Compute the distance (in the journal direction) between two
 359  *   blocks in the journal
 360  *
 361  * Returns: the distance in blocks
 362  */
 363
 364 static inline unsigned int log_distance(struct gfs2_sbd *sdp, unsigned int newer,
 365                                         unsigned int older)
 366 {
 367         int dist;
 368
 369         dist = newer - older;
 370         if (dist < 0)
 371                 dist += sdp->sd_jdesc->jd_blocks;
 372
 373         return dist;
 374 }
 375
 376 /**
 377  * calc_reserved - Calculate the number of blocks to reserve when
 378  *                 refunding a transaction's unused buffers.
 379  * @sdp: The GFS2 superblock
 380  *
 381  * This is complex.  We need to reserve room for all our currently used
 382  * metadata buffers (e.g. normal file I/O rewriting file time stamps) and
 383  * all our journaled data buffers for journaled files (e.g. files in the
 384  * meta_fs like rindex, or files for which chattr +j was done.)
 385  * If we don't reserve enough space, gfs2_log_refund and gfs2_log_flush
 386  * will count it as free space (sd_log_blks_free) and corruption will follow.
 387  *
 388  * We can have metadata bufs and jdata bufs in the same journal.  So each
 389  * type gets its own log header, for which we need to reserve a block.
 390  * In fact, each type has the potential for needing more than one header
 391  * in cases where we have more buffers than will fit on a journal page.
 392  * Metadata journal entries take up half the space of journaled buffer entries.
 393  * Thus, metadata entries have buf_limit (502) and journaled buffers have
 394  * databuf_limit (251) before they cause a wrap around.
 395  *
 396  * Also, we need to reserve blocks for revoke journal entries and one for an
 397  * overall header for the lot.
 398  *
 399  * Returns: the number of blocks reserved
 400  */
 401 static unsigned int calc_reserved(struct gfs2_sbd *sdp)
 402 {
 403         unsigned int reserved = 0;
 404         unsigned int mbuf_limit, metabufhdrs_needed;
 405         unsigned int dbuf_limit, databufhdrs_needed;
 406         unsigned int revokes = 0;
 407
 408         mbuf_limit = buf_limit(sdp);
 409         metabufhdrs_needed = (sdp->sd_log_commited_buf +
 410                               (mbuf_limit - 1)) / mbuf_limit;
 411         dbuf_limit = databuf_limit(sdp);
 412         databufhdrs_needed = (sdp->sd_log_commited_databuf +
 413                               (dbuf_limit - 1)) / dbuf_limit;
 414
 415         if (sdp->sd_log_commited_revoke > 0)
 416                 revokes = gfs2_struct2blk(sdp, sdp->sd_log_commited_revoke,
 417                                           sizeof(u64));
 418
 419         reserved = sdp->sd_log_commited_buf + metabufhdrs_needed +
 420                 sdp->sd_log_commited_databuf + databufhdrs_needed +
 421                 revokes;
 422         /* One for the overall header */
 423         if (reserved)
 424                 reserved++;
 425         return reserved;
 426 }
 427
 428 static unsigned int current_tail(struct gfs2_sbd *sdp)
 429 {
 430         struct gfs2_ail *ai;
 431         unsigned int tail;
 432
 433         spin_lock(&sdp->sd_ail_lock);
 434
 435         if (list_empty(&sdp->sd_ail1_list)) {
 436                 tail = sdp->sd_log_head;
 437         } else {
 438                 ai = list_entry(sdp->sd_ail1_list.prev, struct gfs2_ail, ai_list);
 439                 tail = ai->ai_first;
 440         }
 441
 442         spin_unlock(&sdp->sd_ail_lock);
 443
 444         return tail;
 445 }
 446
 447 void gfs2_log_incr_head(struct gfs2_sbd *sdp)
 448 {
 449         if (sdp->sd_log_flush_head == sdp->sd_log_tail)
 450                 BUG_ON(sdp->sd_log_flush_head != sdp->sd_log_head);
 451
 452         if (++sdp->sd_log_flush_head == sdp->sd_jdesc->jd_blocks) {
 453                 sdp->sd_log_flush_head = 0;
 454                 sdp->sd_log_flush_wrapped = 1;
 455         }
 456 }
 457
 458 /**
 459  * gfs2_log_write_endio - End of I/O for a log buffer
 460  * @bh: The buffer head
 461  * @uptodate: I/O Status
 462  *
 463  */
 464
 465 static void gfs2_log_write_endio(struct buffer_head *bh, int uptodate)
 466 {
 467         struct gfs2_sbd *sdp = bh->b_private;
 468         bh->b_private = NULL;
 469
 470         end_buffer_write_sync(bh, uptodate);
 471         if (atomic_dec_and_test(&sdp->sd_log_in_flight))
 472                 wake_up(&sdp->sd_log_flush_wait);
 473 }
 474
 475 /**
 476  * gfs2_log_get_buf - Get and initialize a buffer to use for log control data
 477  * @sdp: The GFS2 superblock
 478  *
 479  * Returns: the buffer_head
 480  */
 481
 482 struct buffer_head *gfs2_log_get_buf(struct gfs2_sbd *sdp)
 483 {
 484         u64 blkno = log_bmap(sdp, sdp->sd_log_flush_head);
 485         struct buffer_head *bh;
 486
 487         bh = sb_getblk(sdp->sd_vfs, blkno);
 488         lock_buffer(bh);
 489         memset(bh->b_data, 0, bh->b_size);
 490         set_buffer_uptodate(bh);
 491         clear_buffer_dirty(bh);
 492         gfs2_log_incr_head(sdp);
 493         atomic_inc(&sdp->sd_log_in_flight);
 494         bh->b_private = sdp;
 495         bh->b_end_io = gfs2_log_write_endio;
 496
 497         return bh;
 498 }
 499
 500 /**
 501  * gfs2_fake_write_endio -
 502  * @bh: The buffer head
 503  * @uptodate: The I/O Status
 504  *
 505  */
 506
 507 static void gfs2_fake_write_endio(struct buffer_head *bh, int uptodate)
 508 {
 509         struct buffer_head *real_bh = bh->b_private;
 510         struct gfs2_bufdata *bd = real_bh->b_private;
 511         struct gfs2_sbd *sdp = bd->bd_gl->gl_sbd;
 512
 513         end_buffer_write_sync(bh, uptodate);
 514         free_buffer_head(bh);
 515         unlock_buffer(real_bh);
 516         brelse(real_bh);
 517         if (atomic_dec_and_test(&sdp->sd_log_in_flight))
 518                 wake_up(&sdp->sd_log_flush_wait);
 519 }
 520
 521 /**
 522  * gfs2_log_fake_buf - Build a fake buffer head to write metadata buffer to log
 523  * @sdp: the filesystem
 524  * @data: the data the buffer_head should point to
 525  *
 526  * Returns: the log buffer descriptor
 527  */
 528
 529 struct buffer_head *gfs2_log_fake_buf(struct gfs2_sbd *sdp,
 530                                       struct buffer_head *real)
 531 {
 532         u64 blkno = log_bmap(sdp, sdp->sd_log_flush_head);
 533         struct buffer_head *bh;
 534
 535         bh = alloc_buffer_head(GFP_NOFS | __GFP_NOFAIL);
 536         atomic_set(&bh->b_count, 1);
 537         bh->b_state = (1 << BH_Mapped) | (1 << BH_Uptodate) | (1 << BH_Lock);
 538         set_bh_page(bh, real->b_page, bh_offset(real));
 539         bh->b_blocknr = blkno;
 540         bh->b_size = sdp->sd_sb.sb_bsize;
 541         bh->b_bdev = sdp->sd_vfs->s_bdev;
 542         bh->b_private = real;
 543         bh->b_end_io = gfs2_fake_write_endio;
 544
 545         gfs2_log_incr_head(sdp);
 546         atomic_inc(&sdp->sd_log_in_flight);
 547
 548         return bh;
 549 }
 550
 551 static void log_pull_tail(struct gfs2_sbd *sdp, unsigned int new_tail)
 552 {
 553         unsigned int dist = log_distance(sdp, new_tail, sdp->sd_log_tail);
 554
 555         ail2_empty(sdp, new_tail);
 556
 557         atomic_add(dist, &sdp->sd_log_blks_free);
 558         trace_gfs2_log_blocks(sdp, dist);
 559         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
 560                              sdp->sd_jdesc->jd_blocks);
 561
 562         sdp->sd_log_tail = new_tail;
 563 }
 564
 565 /**
 566  * log_write_header - Get and initialize a journal header buffer
 567  * @sdp: The GFS2 superblock
 568  *
 569  * Returns: the initialized log buffer descriptor
 570  */
 571
 572 static void log_write_header(struct gfs2_sbd *sdp, u32 flags, int pull)
 573 {
 574         u64 blkno = log_bmap(sdp, sdp->sd_log_flush_head);
 575         struct buffer_head *bh;
 576         struct gfs2_log_header *lh;
 577         unsigned int tail;
 578         u32 hash;
 579
 580         bh = sb_getblk(sdp->sd_vfs, blkno);
 581         lock_buffer(bh);
 582         memset(bh->b_data, 0, bh->b_size);
 583         set_buffer_uptodate(bh);
 584         clear_buffer_dirty(bh);
 585
 586         gfs2_ail1_empty(sdp);
 587         tail = current_tail(sdp);
 588
 589         lh = (struct gfs2_log_header *)bh->b_data;
 590         memset(lh, 0, sizeof(struct gfs2_log_header));
 591         lh->lh_header.mh_magic = cpu_to_be32(GFS2_MAGIC);
 592         lh->lh_header.mh_type = cpu_to_be32(GFS2_METATYPE_LH);
 593         lh->lh_header.__pad0 = cpu_to_be64(0);
 594         lh->lh_header.mh_format = cpu_to_be32(GFS2_FORMAT_LH);
 595         lh->lh_header.mh_jid = cpu_to_be32(sdp->sd_jdesc->jd_jid);
 596         lh->lh_sequence = cpu_to_be64(sdp->sd_log_sequence++);
 597         lh->lh_flags = cpu_to_be32(flags);
 598         lh->lh_tail = cpu_to_be32(tail);
 599         lh->lh_blkno = cpu_to_be32(sdp->sd_log_flush_head);
 600         hash = gfs2_disk_hash(bh->b_data, sizeof(struct gfs2_log_header));
 601         lh->lh_hash = cpu_to_be32(hash);
 602
 603         bh->b_end_io = end_buffer_write_sync;
 604         get_bh(bh);
 605         if (test_bit(SDF_NOBARRIERS, &sdp->sd_flags))
 606                 submit_bh(WRITE_SYNC | REQ_META, bh);
 607         else
 608                 submit_bh(WRITE_FLUSH_FUA | REQ_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, 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         spin_lock(&sdp->sd_ail_lock);
 756         if (!list_empty(&ai->ai_ail1_list)) {
 757                 list_add(&ai->ai_list, &sdp->sd_ail1_list);
 758                 ai = NULL;
 759         }
 760         spin_unlock(&sdp->sd_ail_lock);
 761         gfs2_log_unlock(sdp);
 762         trace_gfs2_log_flush(sdp, 0);
 763         up_write(&sdp->sd_log_flush_lock);
 764
 765         kfree(ai);
 766 }
 767
 768 static void log_refund(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 769 {
 770         unsigned int reserved;
 771         unsigned int unused;
 772
 773         gfs2_log_lock(sdp);
 774
 775         sdp->sd_log_commited_buf += tr->tr_num_buf_new - tr->tr_num_buf_rm;
 776         sdp->sd_log_commited_databuf += tr->tr_num_databuf_new -
 777                 tr->tr_num_databuf_rm;
 778         gfs2_assert_withdraw(sdp, (((int)sdp->sd_log_commited_buf) >= 0) ||
 779                              (((int)sdp->sd_log_commited_databuf) >= 0));
 780         sdp->sd_log_commited_revoke += tr->tr_num_revoke - tr->tr_num_revoke_rm;
 781         reserved = calc_reserved(sdp);
 782         gfs2_assert_withdraw(sdp, sdp->sd_log_blks_reserved + tr->tr_reserved >= reserved);
 783         unused = sdp->sd_log_blks_reserved - reserved + tr->tr_reserved;
 784         atomic_add(unused, &sdp->sd_log_blks_free);
 785         trace_gfs2_log_blocks(sdp, unused);
 786         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
 787                              sdp->sd_jdesc->jd_blocks);
 788         sdp->sd_log_blks_reserved = reserved;
 789
 790         gfs2_log_unlock(sdp);
 791 }
 792
 793 static void buf_lo_incore_commit(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 794 {
 795         struct list_head *head = &tr->tr_list_buf;
 796         struct gfs2_bufdata *bd;
 797
 798         gfs2_log_lock(sdp);
 799         while (!list_empty(head)) {
 800                 bd = list_entry(head->next, struct gfs2_bufdata, bd_list_tr);
 801                 list_del_init(&bd->bd_list_tr);
 802                 tr->tr_num_buf--;
 803         }
 804         gfs2_log_unlock(sdp);
 805         gfs2_assert_warn(sdp, !tr->tr_num_buf);
 806 }
 807
 808 /**
 809  * gfs2_log_commit - Commit a transaction to the log
 810  * @sdp: the filesystem
 811  * @tr: the transaction
 812  *
 813  * We wake up gfs2_logd if the number of pinned blocks exceed thresh1
 814  * or the total number of used blocks (pinned blocks plus AIL blocks)
 815  * is greater than thresh2.
 816  *
 817  * At mount time thresh1 is 1/3rd of journal size, thresh2 is 2/3rd of
 818  * journal size.
 819  *
 820  * Returns: errno
 821  */
 822
 823 void gfs2_log_commit(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 824 {
 825         log_refund(sdp, tr);
 826         buf_lo_incore_commit(sdp, tr);
 827
 828         up_read(&sdp->sd_log_flush_lock);
 829
 830         if (atomic_read(&sdp->sd_log_pinned) > atomic_read(&sdp->sd_log_thresh1) ||
 831             ((sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free)) >
 832             atomic_read(&sdp->sd_log_thresh2)))
 833                 wake_up(&sdp->sd_logd_waitq);
 834 }
 835
 836 /**
 837  * gfs2_log_shutdown - write a shutdown header into a journal
 838  * @sdp: the filesystem
 839  *
 840  */
 841
 842 void gfs2_log_shutdown(struct gfs2_sbd *sdp)
 843 {
 844         down_write(&sdp->sd_log_flush_lock);
 845
 846         gfs2_assert_withdraw(sdp, !sdp->sd_log_blks_reserved);
 847         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_buf);
 848         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_revoke);
 849         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_rg);
 850         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_databuf);
 851         gfs2_assert_withdraw(sdp, list_empty(&sdp->sd_ail1_list));
 852
 853         sdp->sd_log_flush_head = sdp->sd_log_head;
 854         sdp->sd_log_flush_wrapped = 0;
 855
 856         log_write_header(sdp, GFS2_LOG_HEAD_UNMOUNT,
 857                          (sdp->sd_log_tail == current_tail(sdp)) ? 0 : PULL);
 858
 859         gfs2_assert_warn(sdp, atomic_read(&sdp->sd_log_blks_free) == sdp->sd_jdesc->jd_blocks);
 860         gfs2_assert_warn(sdp, sdp->sd_log_head == sdp->sd_log_tail);
 861         gfs2_assert_warn(sdp, list_empty(&sdp->sd_ail2_list));
 862
 863         sdp->sd_log_head = sdp->sd_log_flush_head;
 864         sdp->sd_log_tail = sdp->sd_log_head;
 865
 866         up_write(&sdp->sd_log_flush_lock);
 867 }
 868
 869
 870 /**
 871  * gfs2_meta_syncfs - sync all the buffers in a filesystem
 872  * @sdp: the filesystem
 873  *
 874  */
 875
 876 void gfs2_meta_syncfs(struct gfs2_sbd *sdp)
 877 {
 878         gfs2_log_flush(sdp, NULL);
 879         for (;;) {
 880                 gfs2_ail1_start(sdp);
 881                 if (gfs2_ail1_empty(sdp))
 882                         break;
 883                 msleep(10);
 884         }
 885 }
 886
 887 static inline int gfs2_jrnl_flush_reqd(struct gfs2_sbd *sdp)
 888 {
 889         return (atomic_read(&sdp->sd_log_pinned) >= atomic_read(&sdp->sd_log_thresh1));
 890 }
 891
 892 static inline int gfs2_ail_flush_reqd(struct gfs2_sbd *sdp)
 893 {
 894         unsigned int used_blocks = sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free);
 895         return used_blocks >= atomic_read(&sdp->sd_log_thresh2);
 896 }
 897
 898 /**
 899  * gfs2_logd - Update log tail as Active Items get flushed to in-place blocks
 900  * @sdp: Pointer to GFS2 superblock
 901  *
 902  * Also, periodically check to make sure that we're using the most recent
 903  * journal index.
 904  */
 905
 906 int gfs2_logd(void *data)
 907 {
 908         struct gfs2_sbd *sdp = data;
 909         unsigned long t = 1;
 910         DEFINE_WAIT(wait);
 911         unsigned preflush;
 912
 913         while (!kthread_should_stop()) {
 914
 915                 preflush = atomic_read(&sdp->sd_log_pinned);
 916                 if (gfs2_jrnl_flush_reqd(sdp) || t == 0) {
 917                         gfs2_ail1_empty(sdp);
 918                         gfs2_log_flush(sdp, NULL);
 919                 }
 920
 921                 if (gfs2_ail_flush_reqd(sdp)) {
 922                         gfs2_ail1_start(sdp);
 923                         io_schedule();
 924                         gfs2_ail1_empty(sdp);
 925                         gfs2_log_flush(sdp, NULL);
 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_INTERRUPTIBLE);
 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