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(REQ_OP_WRITE, 0, 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                 mlog_errno(ret);
  94         }
  95
  96         ocfs2_metadata_cache_io_unlock(ci);
  97 out:
  98         return ret;
  99 }
 100
 101 int ocfs2_read_blocks_sync(struct ocfs2_super *osb, u64 block,
 102                            unsigned int nr, struct buffer_head *bhs[])
 103 {
 104         int status = 0;
 105         unsigned int i;
 106         struct buffer_head *bh;
 107
 108         trace_ocfs2_read_blocks_sync((unsigned long long)block, nr);
 109
 110         if (!nr)
 111                 goto bail;
 112
 113         for (i = 0 ; i < nr ; i++) {
 114                 if (bhs[i] == NULL) {
 115                         bhs[i] = sb_getblk(osb->sb, block++);
 116                         if (bhs[i] == NULL) {
 117                                 status = -ENOMEM;
 118                                 mlog_errno(status);
 119                                 goto bail;
 120                         }
 121                 }
 122                 bh = bhs[i];
 123
 124                 if (buffer_jbd(bh)) {
 125                         trace_ocfs2_read_blocks_sync_jbd(
 126                                         (unsigned long long)bh->b_blocknr);
 127                         continue;
 128                 }
 129
 130                 if (buffer_dirty(bh)) {
 131                         /* This should probably be a BUG, or
 132                          * at least return an error. */
 133                         mlog(ML_ERROR,
 134                              "trying to sync read a dirty "
 135                              "buffer! (blocknr = %llu), skipping\n",
 136                              (unsigned long long)bh->b_blocknr);
 137                         continue;
 138                 }
 139
 140                 lock_buffer(bh);
 141                 if (buffer_jbd(bh)) {
 142 #ifdef CATCH_BH_JBD_RACES
 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 #else
 149                         unlock_buffer(bh);
 150                         continue;
 151 #endif
 152                 }
 153
 154                 get_bh(bh); /* for end_buffer_read_sync() */
 155                 bh->b_end_io = end_buffer_read_sync;
 156                 submit_bh(REQ_OP_READ, 0, bh);
 157         }
 158
 159         for (i = nr; i > 0; i--) {
 160                 bh = bhs[i - 1];
 161
 162                 /* No need to wait on the buffer if it's managed by JBD. */
 163                 if (!buffer_jbd(bh))
 164                         wait_on_buffer(bh);
 165
 166                 if (!buffer_uptodate(bh)) {
 167                         /* Status won't be cleared from here on out,
 168                          * so we can safely record this and loop back
 169                          * to cleanup the other buffers. */
 170                         status = -EIO;
 171                         put_bh(bh);
 172                         bhs[i - 1] = NULL;
 173                 }
 174         }
 175
 176 bail:
 177         return status;
 178 }
 179
 180 int ocfs2_read_blocks(struct ocfs2_caching_info *ci, u64 block, int nr,
 181                       struct buffer_head *bhs[], int flags,
 182                       int (*validate)(struct super_block *sb,
 183                                       struct buffer_head *bh))
 184 {
 185         int status = 0;
 186         int i, ignore_cache = 0;
 187         struct buffer_head *bh;
 188         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
 189
 190         trace_ocfs2_read_blocks_begin(ci, (unsigned long long)block, nr, flags);
 191
 192         BUG_ON(!ci);
 193         BUG_ON((flags & OCFS2_BH_READAHEAD) &&
 194                (flags & OCFS2_BH_IGNORE_CACHE));
 195
 196         if (bhs == NULL) {
 197                 status = -EINVAL;
 198                 mlog_errno(status);
 199                 goto bail;
 200         }
 201
 202         if (nr < 0) {
 203                 mlog(ML_ERROR, "asked to read %d blocks!\n", nr);
 204                 status = -EINVAL;
 205                 mlog_errno(status);
 206                 goto bail;
 207         }
 208
 209         if (nr == 0) {
 210                 status = 0;
 211                 goto bail;
 212         }
 213
 214         ocfs2_metadata_cache_io_lock(ci);
 215         for (i = 0 ; i < nr ; i++) {
 216                 if (bhs[i] == NULL) {
 217                         bhs[i] = sb_getblk(sb, block++);
 218                         if (bhs[i] == NULL) {
 219                                 ocfs2_metadata_cache_io_unlock(ci);
 220                                 status = -ENOMEM;
 221                                 mlog_errno(status);
 222                                 goto bail;
 223                         }
 224                 }
 225                 bh = bhs[i];
 226                 ignore_cache = (flags & OCFS2_BH_IGNORE_CACHE);
 227
 228                 /* There are three read-ahead cases here which we need to
 229                  * be concerned with. All three assume a buffer has
 230                  * previously been submitted with OCFS2_BH_READAHEAD
 231                  * and it hasn't yet completed I/O.
 232                  *
 233                  * 1) The current request is sync to disk. This rarely
 234                  *    happens these days, and never when performance
 235                  *    matters - the code can just wait on the buffer
 236                  *    lock and re-submit.
 237                  *
 238                  * 2) The current request is cached, but not
 239                  *    readahead. ocfs2_buffer_uptodate() will return
 240                  *    false anyway, so we'll wind up waiting on the
 241                  *    buffer lock to do I/O. We re-check the request
 242                  *    with after getting the lock to avoid a re-submit.
 243                  *
 244                  * 3) The current request is readahead (and so must
 245                  *    also be a caching one). We short circuit if the
 246                  *    buffer is locked (under I/O) and if it's in the
 247                  *    uptodate cache. The re-check from #2 catches the
 248                  *    case that the previous read-ahead completes just
 249                  *    before our is-it-in-flight check.
 250                  */
 251
 252                 if (!ignore_cache && !ocfs2_buffer_uptodate(ci, bh)) {
 253                         trace_ocfs2_read_blocks_from_disk(
 254                              (unsigned long long)bh->b_blocknr,
 255                              (unsigned long long)ocfs2_metadata_cache_owner(ci));
 256                         /* We're using ignore_cache here to say
 257                          * "go to disk" */
 258                         ignore_cache = 1;
 259                 }
 260
 261                 trace_ocfs2_read_blocks_bh((unsigned long long)bh->b_blocknr,
 262                         ignore_cache, buffer_jbd(bh), buffer_dirty(bh));
 263
 264                 if (buffer_jbd(bh)) {
 265                         continue;
 266                 }
 267
 268                 if (ignore_cache) {
 269                         if (buffer_dirty(bh)) {
 270                                 /* This should probably be a BUG, or
 271                                  * at least return an error. */
 272                                 continue;
 273                         }
 274
 275                         /* A read-ahead request was made - if the
 276                          * buffer is already under read-ahead from a
 277                          * previously submitted request than we are
 278                          * done here. */
 279                         if ((flags & OCFS2_BH_READAHEAD)
 280                             && ocfs2_buffer_read_ahead(ci, bh))
 281                                 continue;
 282
 283                         lock_buffer(bh);
 284                         if (buffer_jbd(bh)) {
 285 #ifdef CATCH_BH_JBD_RACES
 286                                 mlog(ML_ERROR, "block %llu had the JBD bit set "
 287                                                "while I was in lock_buffer!",
 288                                      (unsigned long long)bh->b_blocknr);
 289                                 BUG();
 290 #else
 291                                 unlock_buffer(bh);
 292                                 continue;
 293 #endif
 294                         }
 295
 296                         /* Re-check ocfs2_buffer_uptodate() as a
 297                          * previously read-ahead buffer may have
 298                          * completed I/O while we were waiting for the
 299                          * buffer lock. */
 300                         if (!(flags & OCFS2_BH_IGNORE_CACHE)
 301                             && !(flags & OCFS2_BH_READAHEAD)
 302                             && ocfs2_buffer_uptodate(ci, bh)) {
 303                                 unlock_buffer(bh);
 304                                 continue;
 305                         }
 306
 307                         get_bh(bh); /* for end_buffer_read_sync() */
 308                         if (validate)
 309                                 set_buffer_needs_validate(bh);
 310                         bh->b_end_io = end_buffer_read_sync;
 311                         submit_bh(REQ_OP_READ, 0, bh);
 312                         continue;
 313                 }
 314         }
 315
 316         status = 0;
 317
 318         for (i = (nr - 1); i >= 0; i--) {
 319                 bh = bhs[i];
 320
 321                 if (!(flags & OCFS2_BH_READAHEAD)) {
 322                         if (status) {
 323                                 /* Clear the rest of the buffers on error */
 324                                 put_bh(bh);
 325                                 bhs[i] = NULL;
 326                                 continue;
 327                         }
 328                         /* We know this can't have changed as we hold the
 329                          * owner sem. Avoid doing any work on the bh if the
 330                          * journal has it. */
 331                         if (!buffer_jbd(bh))
 332                                 wait_on_buffer(bh);
 333
 334                         if (!buffer_uptodate(bh)) {
 335                                 /* Status won't be cleared from here on out,
 336                                  * so we can safely record this and loop back
 337                                  * to cleanup the other buffers. Don't need to
 338                                  * remove the clustered uptodate information
 339                                  * for this bh as it's not marked locally
 340                                  * uptodate. */
 341                                 status = -EIO;
 342                                 clear_buffer_needs_validate(bh);
 343                                 put_bh(bh);
 344                                 bhs[i] = NULL;
 345                                 continue;
 346                         }
 347
 348                         if (buffer_needs_validate(bh)) {
 349                                 /* We never set NeedsValidate if the
 350                                  * buffer was held by the journal, so
 351                                  * that better not have changed */
 352                                 BUG_ON(buffer_jbd(bh));
 353                                 clear_buffer_needs_validate(bh);
 354                                 status = validate(sb, bh);
 355                                 if (status) {
 356                                         put_bh(bh);
 357                                         bhs[i] = NULL;
 358                                         continue;
 359                                 }
 360                         }
 361                 }
 362
 363                 /* Always set the buffer in the cache, even if it was
 364                  * a forced read, or read-ahead which hasn't yet
 365                  * completed. */
 366                 ocfs2_set_buffer_uptodate(ci, bh);
 367         }
 368         ocfs2_metadata_cache_io_unlock(ci);
 369
 370         trace_ocfs2_read_blocks_end((unsigned long long)block, nr,
 371                                     flags, ignore_cache);
 372
 373 bail:
 374
 375         return status;
 376 }
 377
 378 /* Check whether the blkno is the super block or one of the backups. */
 379 static void ocfs2_check_super_or_backup(struct super_block *sb,
 380                                         sector_t blkno)
 381 {
 382         int i;
 383         u64 backup_blkno;
 384
 385         if (blkno == OCFS2_SUPER_BLOCK_BLKNO)
 386                 return;
 387
 388         for (i = 0; i < OCFS2_MAX_BACKUP_SUPERBLOCKS; i++) {
 389                 backup_blkno = ocfs2_backup_super_blkno(sb, i);
 390                 if (backup_blkno == blkno)
 391                         return;
 392         }
 393
 394         BUG();
 395 }
 396
 397 /*
 398  * Write super block and backups doesn't need to collaborate with journal,
 399  * so we don't need to lock ip_io_mutex and ci doesn't need to bea passed
 400  * into this function.
 401  */
 402 int ocfs2_write_super_or_backup(struct ocfs2_super *osb,
 403                                 struct buffer_head *bh)
 404 {
 405         int ret = 0;
 406         struct ocfs2_dinode *di = (struct ocfs2_dinode *)bh->b_data;
 407
 408         BUG_ON(buffer_jbd(bh));
 409         ocfs2_check_super_or_backup(osb->sb, bh->b_blocknr);
 410
 411         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb)) {
 412                 ret = -EROFS;
 413                 mlog_errno(ret);
 414                 goto out;
 415         }
 416
 417         lock_buffer(bh);
 418         set_buffer_uptodate(bh);
 419
 420         /* remove from dirty list before I/O. */
 421         clear_buffer_dirty(bh);
 422
 423         get_bh(bh); /* for end_buffer_write_sync() */
 424         bh->b_end_io = end_buffer_write_sync;
 425         ocfs2_compute_meta_ecc(osb->sb, bh->b_data, &di->i_check);
 426         submit_bh(REQ_OP_WRITE, 0, bh);
 427
 428         wait_on_buffer(bh);
 429
 430         if (!buffer_uptodate(bh)) {
 431                 ret = -EIO;
 432                 mlog_errno(ret);
 433         }
 434
 435 out:
 436         return ret;
 437 }