fs/ocfs2/buffer_head_io.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * io.c
   5  *
   6  * Buffer cache handling
   7  *
   8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  */
  25
  26 #include <linux/fs.h>
  27 #include <linux/types.h>
  28 #include <linux/highmem.h>
  29
  30 #include <cluster/masklog.h>
  31
  32 #include "ocfs2.h"
  33
  34 #include "alloc.h"
  35 #include "inode.h"
  36 #include "journal.h"
  37 #include "uptodate.h"
  38 #include "buffer_head_io.h"
  39 #include "ocfs2_trace.h"
  40
  41 /*
  42  * Bits on bh->b_state used by ocfs2.
  43  *
  44  * These MUST be after the JBD2 bits.  Hence, we use BH_JBDPrivateStart.
  45  */
  46 enum ocfs2_state_bits {
  47         BH_NeedsValidate = BH_JBDPrivateStart,
  48 };
  49
  50 /* Expand the magic b_state functions */
  51 BUFFER_FNS(NeedsValidate, needs_validate);
  52
  53 int ocfs2_write_block(struct ocfs2_super *osb, struct buffer_head *bh,
  54                       struct ocfs2_caching_info *ci)
  55 {
  56         int ret = 0;
  57
  58         trace_ocfs2_write_block((unsigned long long)bh->b_blocknr, ci);
  59
  60         BUG_ON(bh->b_blocknr < OCFS2_SUPER_BLOCK_BLKNO);
  61         BUG_ON(buffer_jbd(bh));
  62
  63         /* No need to check for a soft readonly file system here. non
  64          * journalled writes are only ever done on system files which
  65          * can get modified during recovery even if read-only. */
  66         if (ocfs2_is_hard_readonly(osb)) {
  67                 ret = -EROFS;
  68                 mlog_errno(ret);
  69                 goto out;
  70         }
  71
  72         ocfs2_metadata_cache_io_lock(ci);
  73
  74         lock_buffer(bh);
  75         set_buffer_uptodate(bh);
  76
  77         /* remove from dirty list before I/O. */
  78         clear_buffer_dirty(bh);
  79
  80         get_bh(bh); /* for end_buffer_write_sync() */
  81         bh->b_end_io = end_buffer_write_sync;
  82         submit_bh(WRITE, bh);
  83
  84         wait_on_buffer(bh);
  85
  86         if (buffer_uptodate(bh)) {
  87                 ocfs2_set_buffer_uptodate(ci, bh);
  88         } else {
  89                 /* We don't need to remove the clustered uptodate
  90                  * information for this bh as it's not marked locally
  91                  * uptodate. */
  92                 ret = -EIO;
  93                 put_bh(bh);
  94                 mlog_errno(ret);
  95         }
  96
  97         ocfs2_metadata_cache_io_unlock(ci);
  98 out:
  99         return ret;
 100 }
 101
 102 int ocfs2_read_blocks_sync(struct ocfs2_super *osb, u64 block,
 103                            unsigned int nr, struct buffer_head *bhs[])
 104 {
 105         int status = 0;
 106         unsigned int i;
 107         struct buffer_head *bh;
 108
 109         trace_ocfs2_read_blocks_sync((unsigned long long)block, nr);
 110
 111         if (!nr)
 112                 goto bail;
 113
 114         for (i = 0 ; i < nr ; i++) {
 115                 if (bhs[i] == NULL) {
 116                         bhs[i] = sb_getblk(osb->sb, block++);
 117                         if (bhs[i] == NULL) {
 118                                 status = -EIO;
 119                                 mlog_errno(status);
 120                                 goto bail;
 121                         }
 122                 }
 123                 bh = bhs[i];
 124
 125                 if (buffer_jbd(bh)) {
 126                         trace_ocfs2_read_blocks_sync_jbd(
 127                                         (unsigned long long)bh->b_blocknr);
 128                         continue;
 129                 }
 130
 131                 if (buffer_dirty(bh)) {
 132                         /* This should probably be a BUG, or
 133                          * at least return an error. */
 134                         mlog(ML_ERROR,
 135                              "trying to sync read a dirty "
 136                              "buffer! (blocknr = %llu), skipping\n",
 137                              (unsigned long long)bh->b_blocknr);
 138                         continue;
 139                 }
 140
 141                 lock_buffer(bh);
 142                 if (buffer_jbd(bh)) {
 143                         mlog(ML_ERROR,
 144                              "block %llu had the JBD bit set "
 145                              "while I was in lock_buffer!",
 146                              (unsigned long long)bh->b_blocknr);
 147                         BUG();
 148                 }
 149
 150                 clear_buffer_uptodate(bh);
 151                 get_bh(bh); /* for end_buffer_read_sync() */
 152                 bh->b_end_io = end_buffer_read_sync;
 153                 submit_bh(READ, bh);
 154         }
 155
 156         for (i = nr; i > 0; i--) {
 157                 bh = bhs[i - 1];
 158
 159                 /* No need to wait on the buffer if it's managed by JBD. */
 160                 if (!buffer_jbd(bh))
 161                         wait_on_buffer(bh);
 162
 163                 if (!buffer_uptodate(bh)) {
 164                         /* Status won't be cleared from here on out,
 165                          * so we can safely record this and loop back
 166                          * to cleanup the other buffers. */
 167                         status = -EIO;
 168                         put_bh(bh);
 169                         bhs[i - 1] = NULL;
 170                 }
 171         }
 172
 173 bail:
 174         return status;
 175 }
 176
 177 int ocfs2_read_blocks(struct ocfs2_caching_info *ci, u64 block, int nr,
 178                       struct buffer_head *bhs[], int flags,
 179                       int (*validate)(struct super_block *sb,
 180                                       struct buffer_head *bh))
 181 {
 182         int status = 0;
 183         int i, ignore_cache = 0;
 184         struct buffer_head *bh;
 185         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
 186
 187         trace_ocfs2_read_blocks_begin(ci, (unsigned long long)block, nr, flags);
 188
 189         BUG_ON(!ci);
 190         BUG_ON((flags & OCFS2_BH_READAHEAD) &&
 191                (flags & OCFS2_BH_IGNORE_CACHE));
 192
 193         if (bhs == NULL) {
 194                 status = -EINVAL;
 195                 mlog_errno(status);
 196                 goto bail;
 197         }
 198
 199         if (nr < 0) {
 200                 mlog(ML_ERROR, "asked to read %d blocks!\n", nr);
 201                 status = -EINVAL;
 202                 mlog_errno(status);
 203                 goto bail;
 204         }
 205
 206         if (nr == 0) {
 207                 status = 0;
 208                 goto bail;
 209         }
 210
 211         ocfs2_metadata_cache_io_lock(ci);
 212         for (i = 0 ; i < nr ; i++) {
 213                 if (bhs[i] == NULL) {
 214                         bhs[i] = sb_getblk(sb, block++);
 215                         if (bhs[i] == NULL) {
 216                                 ocfs2_metadata_cache_io_unlock(ci);
 217                                 status = -EIO;
 218                                 mlog_errno(status);
 219                                 goto bail;
 220                         }
 221                 }
 222                 bh = bhs[i];
 223                 ignore_cache = (flags & OCFS2_BH_IGNORE_CACHE);
 224
 225                 /* There are three read-ahead cases here which we need to
 226                  * be concerned with. All three assume a buffer has
 227                  * previously been submitted with OCFS2_BH_READAHEAD
 228                  * and it hasn't yet completed I/O.
 229                  *
 230                  * 1) The current request is sync to disk. This rarely
 231                  *    happens these days, and never when performance
 232                  *    matters - the code can just wait on the buffer
 233                  *    lock and re-submit.
 234                  *
 235                  * 2) The current request is cached, but not
 236                  *    readahead. ocfs2_buffer_uptodate() will return
 237                  *    false anyway, so we'll wind up waiting on the
 238                  *    buffer lock to do I/O. We re-check the request
 239                  *    with after getting the lock to avoid a re-submit.
 240                  *
 241                  * 3) The current request is readahead (and so must
 242                  *    also be a caching one). We short circuit if the
 243                  *    buffer is locked (under I/O) and if it's in the
 244                  *    uptodate cache. The re-check from #2 catches the
 245                  *    case that the previous read-ahead completes just
 246                  *    before our is-it-in-flight check.
 247                  */
 248
 249                 if (!ignore_cache && !ocfs2_buffer_uptodate(ci, bh)) {
 250                         mlog(ML_UPTODATE,
 251                              "bh (%llu), owner %llu not uptodate\n",
 252                              (unsigned long long)bh->b_blocknr,
 253                              (unsigned long long)ocfs2_metadata_cache_owner(ci));
 254                         /* We're using ignore_cache here to say
 255                          * "go to disk" */
 256                         ignore_cache = 1;
 257                 }
 258
 259                 trace_ocfs2_read_blocks_bh((unsigned long long)bh->b_blocknr,
 260                         ignore_cache, buffer_jbd(bh), buffer_dirty(bh));
 261
 262                 if (buffer_jbd(bh)) {
 263                         continue;
 264                 }
 265
 266                 if (ignore_cache) {
 267                         if (buffer_dirty(bh)) {
 268                                 /* This should probably be a BUG, or
 269                                  * at least return an error. */
 270                                 continue;
 271                         }
 272
 273                         /* A read-ahead request was made - if the
 274                          * buffer is already under read-ahead from a
 275                          * previously submitted request than we are
 276                          * done here. */
 277                         if ((flags & OCFS2_BH_READAHEAD)
 278                             && ocfs2_buffer_read_ahead(ci, bh))
 279                                 continue;
 280
 281                         lock_buffer(bh);
 282                         if (buffer_jbd(bh)) {
 283 #ifdef CATCH_BH_JBD_RACES
 284                                 mlog(ML_ERROR, "block %llu had the JBD bit set "
 285                                                "while I was in lock_buffer!",
 286                                      (unsigned long long)bh->b_blocknr);
 287                                 BUG();
 288 #else
 289                                 unlock_buffer(bh);
 290                                 continue;
 291 #endif
 292                         }
 293
 294                         /* Re-check ocfs2_buffer_uptodate() as a
 295                          * previously read-ahead buffer may have
 296                          * completed I/O while we were waiting for the
 297                          * buffer lock. */
 298                         if (!(flags & OCFS2_BH_IGNORE_CACHE)
 299                             && !(flags & OCFS2_BH_READAHEAD)
 300                             && ocfs2_buffer_uptodate(ci, bh)) {
 301                                 unlock_buffer(bh);
 302                                 continue;
 303                         }
 304
 305                         clear_buffer_uptodate(bh);
 306                         get_bh(bh); /* for end_buffer_read_sync() */
 307                         if (validate)
 308                                 set_buffer_needs_validate(bh);
 309                         bh->b_end_io = end_buffer_read_sync;
 310                         submit_bh(READ, bh);
 311                         continue;
 312                 }
 313         }
 314
 315         status = 0;
 316
 317         for (i = (nr - 1); i >= 0; i--) {
 318                 bh = bhs[i];
 319
 320                 if (!(flags & OCFS2_BH_READAHEAD)) {
 321                         /* We know this can't have changed as we hold the
 322                          * owner sem. Avoid doing any work on the bh if the
 323                          * journal has it. */
 324                         if (!buffer_jbd(bh))
 325                                 wait_on_buffer(bh);
 326
 327                         if (!buffer_uptodate(bh)) {
 328                                 /* Status won't be cleared from here on out,
 329                                  * so we can safely record this and loop back
 330                                  * to cleanup the other buffers. Don't need to
 331                                  * remove the clustered uptodate information
 332                                  * for this bh as it's not marked locally
 333                                  * uptodate. */
 334                                 status = -EIO;
 335                                 put_bh(bh);
 336                                 bhs[i] = NULL;
 337                                 continue;
 338                         }
 339
 340                         if (buffer_needs_validate(bh)) {
 341                                 /* We never set NeedsValidate if the
 342                                  * buffer was held by the journal, so
 343                                  * that better not have changed */
 344                                 BUG_ON(buffer_jbd(bh));
 345                                 clear_buffer_needs_validate(bh);
 346                                 status = validate(sb, bh);
 347                                 if (status) {
 348                                         put_bh(bh);
 349                                         bhs[i] = NULL;
 350                                         continue;
 351                                 }
 352                         }
 353                 }
 354
 355                 /* Always set the buffer in the cache, even if it was
 356                  * a forced read, or read-ahead which hasn't yet
 357                  * completed. */
 358                 ocfs2_set_buffer_uptodate(ci, bh);
 359         }
 360         ocfs2_metadata_cache_io_unlock(ci);
 361
 362         trace_ocfs2_read_blocks_end((unsigned long long)block, nr,
 363                                     flags, ignore_cache);
 364
 365 bail:
 366
 367         return status;
 368 }
 369
 370 /* Check whether the blkno is the super block or one of the backups. */
 371 static void ocfs2_check_super_or_backup(struct super_block *sb,
 372                                         sector_t blkno)
 373 {
 374         int i;
 375         u64 backup_blkno;
 376
 377         if (blkno == OCFS2_SUPER_BLOCK_BLKNO)
 378                 return;
 379
 380         for (i = 0; i < OCFS2_MAX_BACKUP_SUPERBLOCKS; i++) {
 381                 backup_blkno = ocfs2_backup_super_blkno(sb, i);
 382                 if (backup_blkno == blkno)
 383                         return;
 384         }
 385
 386         BUG();
 387 }
 388
 389 /*
 390  * Write super block and backups doesn't need to collaborate with journal,
 391  * so we don't need to lock ip_io_mutex and ci doesn't need to bea passed
 392  * into this function.
 393  */
 394 int ocfs2_write_super_or_backup(struct ocfs2_super *osb,
 395                                 struct buffer_head *bh)
 396 {
 397         int ret = 0;
 398         struct ocfs2_dinode *di = (struct ocfs2_dinode *)bh->b_data;
 399
 400         BUG_ON(buffer_jbd(bh));
 401         ocfs2_check_super_or_backup(osb->sb, bh->b_blocknr);
 402
 403         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb)) {
 404                 ret = -EROFS;
 405                 mlog_errno(ret);
 406                 goto out;
 407         }
 408
 409         lock_buffer(bh);
 410         set_buffer_uptodate(bh);
 411
 412         /* remove from dirty list before I/O. */
 413         clear_buffer_dirty(bh);
 414
 415         get_bh(bh); /* for end_buffer_write_sync() */
 416         bh->b_end_io = end_buffer_write_sync;
 417         ocfs2_compute_meta_ecc(osb->sb, bh->b_data, &di->i_check);
 418         submit_bh(WRITE, bh);
 419
 420         wait_on_buffer(bh);
 421
 422         if (!buffer_uptodate(bh)) {
 423                 ret = -EIO;
 424                 put_bh(bh);
 425                 mlog_errno(ret);
 426         }
 427
 428 out:
 429         return ret;
 430 }