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 #include <linux/bio.h>
  30
  31 #include <cluster/masklog.h>
  32
  33 #include "ocfs2.h"
  34
  35 #include "alloc.h"
  36 #include "inode.h"
  37 #include "journal.h"
  38 #include "uptodate.h"
  39 #include "buffer_head_io.h"
  40 #include "ocfs2_trace.h"
  41
  42 /*
  43  * Bits on bh->b_state used by ocfs2.
  44  *
  45  * These MUST be after the JBD2 bits.  Hence, we use BH_JBDPrivateStart.
  46  */
  47 enum ocfs2_state_bits {
  48         BH_NeedsValidate = BH_JBDPrivateStart,
  49 };
  50
  51 /* Expand the magic b_state functions */
  52 BUFFER_FNS(NeedsValidate, needs_validate);
  53
  54 int ocfs2_write_block(struct ocfs2_super *osb, struct buffer_head *bh,
  55                       struct ocfs2_caching_info *ci)
  56 {
  57         int ret = 0;
  58
  59         trace_ocfs2_write_block((unsigned long long)bh->b_blocknr, ci);
  60
  61         BUG_ON(bh->b_blocknr < OCFS2_SUPER_BLOCK_BLKNO);
  62         BUG_ON(buffer_jbd(bh));
  63
  64         /* No need to check for a soft readonly file system here. non
  65          * journalled writes are only ever done on system files which
  66          * can get modified during recovery even if read-only. */
  67         if (ocfs2_is_hard_readonly(osb)) {
  68                 ret = -EROFS;
  69                 mlog_errno(ret);
  70                 goto out;
  71         }
  72
  73         ocfs2_metadata_cache_io_lock(ci);
  74
  75         lock_buffer(bh);
  76         set_buffer_uptodate(bh);
  77
  78         /* remove from dirty list before I/O. */
  79         clear_buffer_dirty(bh);
  80
  81         get_bh(bh); /* for end_buffer_write_sync() */
  82         bh->b_end_io = end_buffer_write_sync;
  83         submit_bh(REQ_OP_WRITE, 0, bh);
  84
  85         wait_on_buffer(bh);
  86
  87         if (buffer_uptodate(bh)) {
  88                 ocfs2_set_buffer_uptodate(ci, bh);
  89         } else {
  90                 /* We don't need to remove the clustered uptodate
  91                  * information for this bh as it's not marked locally
  92                  * uptodate. */
  93                 ret = -EIO;
  94                 mlog_errno(ret);
  95         }
  96
  97         ocfs2_metadata_cache_io_unlock(ci);
  98 out:
  99         return ret;
 100 }
 101
 102 /* Caller must provide a bhs[] with all NULL or non-NULL entries, so it
 103  * will be easier to handle read failure.
 104  */
 105 int ocfs2_read_blocks_sync(struct ocfs2_super *osb, u64 block,
 106                            unsigned int nr, struct buffer_head *bhs[])
 107 {
 108         int status = 0;
 109         unsigned int i;
 110         struct buffer_head *bh;
 111         int new_bh = 0;
 112
 113         trace_ocfs2_read_blocks_sync((unsigned long long)block, nr);
 114
 115         if (!nr)
 116                 goto bail;
 117
 118         /* Don't put buffer head and re-assign it to NULL if it is allocated
 119          * outside since the caller can't be aware of this alternation!
 120          */
 121         new_bh = (bhs[0] == NULL);
 122
 123         for (i = 0 ; i < nr ; i++) {
 124                 if (bhs[i] == NULL) {
 125                         bhs[i] = sb_getblk(osb->sb, block++);
 126                         if (bhs[i] == NULL) {
 127                                 status = -ENOMEM;
 128                                 mlog_errno(status);
 129                                 break;
 130                         }
 131                 }
 132                 bh = bhs[i];
 133
 134                 if (buffer_jbd(bh)) {
 135                         trace_ocfs2_read_blocks_sync_jbd(
 136                                         (unsigned long long)bh->b_blocknr);
 137                         continue;
 138                 }
 139
 140                 if (buffer_dirty(bh)) {
 141                         /* This should probably be a BUG, or
 142                          * at least return an error. */
 143                         mlog(ML_ERROR,
 144                              "trying to sync read a dirty "
 145                              "buffer! (blocknr = %llu), skipping\n",
 146                              (unsigned long long)bh->b_blocknr);
 147                         continue;
 148                 }
 149
 150                 lock_buffer(bh);
 151                 if (buffer_jbd(bh)) {
 152 #ifdef CATCH_BH_JBD_RACES
 153                         mlog(ML_ERROR,
 154                              "block %llu had the JBD bit set "
 155                              "while I was in lock_buffer!",
 156                              (unsigned long long)bh->b_blocknr);
 157                         BUG();
 158 #else
 159                         unlock_buffer(bh);
 160                         continue;
 161 #endif
 162                 }
 163
 164                 clear_buffer_uptodate(bh);
 165                 get_bh(bh); /* for end_buffer_read_sync() */
 166                 bh->b_end_io = end_buffer_read_sync;
 167                 submit_bh(REQ_OP_READ, 0, bh);
 168         }
 169
 170 read_failure:
 171         for (i = nr; i > 0; i--) {
 172                 bh = bhs[i - 1];
 173
 174                 if (unlikely(status)) {
 175                         if (new_bh && bh) {
 176                                 /* If middle bh fails, let previous bh
 177                                  * finish its read and then put it to
 178                                  * aovoid bh leak
 179                                  */
 180                                 if (!buffer_jbd(bh))
 181                                         wait_on_buffer(bh);
 182                                 put_bh(bh);
 183                                 bhs[i - 1] = NULL;
 184                         } else if (bh && buffer_uptodate(bh)) {
 185                                 clear_buffer_uptodate(bh);
 186                         }
 187                         continue;
 188                 }
 189
 190                 /* No need to wait on the buffer if it's managed by JBD. */
 191                 if (!buffer_jbd(bh))
 192                         wait_on_buffer(bh);
 193
 194                 if (!buffer_uptodate(bh)) {
 195                         /* Status won't be cleared from here on out,
 196                          * so we can safely record this and loop back
 197                          * to cleanup the other buffers. */
 198                         status = -EIO;
 199                         goto read_failure;
 200                 }
 201         }
 202
 203 bail:
 204         return status;
 205 }
 206
 207 /* Caller must provide a bhs[] with all NULL or non-NULL entries, so it
 208  * will be easier to handle read failure.
 209  */
 210 int ocfs2_read_blocks(struct ocfs2_caching_info *ci, u64 block, int nr,
 211                       struct buffer_head *bhs[], int flags,
 212                       int (*validate)(struct super_block *sb,
 213                                       struct buffer_head *bh))
 214 {
 215         int status = 0;
 216         int i, ignore_cache = 0;
 217         struct buffer_head *bh;
 218         struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
 219         int new_bh = 0;
 220
 221         trace_ocfs2_read_blocks_begin(ci, (unsigned long long)block, nr, flags);
 222
 223         BUG_ON(!ci);
 224         BUG_ON((flags & OCFS2_BH_READAHEAD) &&
 225                (flags & OCFS2_BH_IGNORE_CACHE));
 226
 227         if (bhs == NULL) {
 228                 status = -EINVAL;
 229                 mlog_errno(status);
 230                 goto bail;
 231         }
 232
 233         if (nr < 0) {
 234                 mlog(ML_ERROR, "asked to read %d blocks!\n", nr);
 235                 status = -EINVAL;
 236                 mlog_errno(status);
 237                 goto bail;
 238         }
 239
 240         if (nr == 0) {
 241                 status = 0;
 242                 goto bail;
 243         }
 244
 245         /* Don't put buffer head and re-assign it to NULL if it is allocated
 246          * outside since the caller can't be aware of this alternation!
 247          */
 248         new_bh = (bhs[0] == NULL);
 249
 250         ocfs2_metadata_cache_io_lock(ci);
 251         for (i = 0 ; i < nr ; i++) {
 252                 if (bhs[i] == NULL) {
 253                         bhs[i] = sb_getblk(sb, block++);
 254                         if (bhs[i] == NULL) {
 255                                 ocfs2_metadata_cache_io_unlock(ci);
 256                                 status = -ENOMEM;
 257                                 mlog_errno(status);
 258                                 /* Don't forget to put previous bh! */
 259                                 break;
 260                         }
 261                 }
 262                 bh = bhs[i];
 263                 ignore_cache = (flags & OCFS2_BH_IGNORE_CACHE);
 264
 265                 /* There are three read-ahead cases here which we need to
 266                  * be concerned with. All three assume a buffer has
 267                  * previously been submitted with OCFS2_BH_READAHEAD
 268                  * and it hasn't yet completed I/O.
 269                  *
 270                  * 1) The current request is sync to disk. This rarely
 271                  *    happens these days, and never when performance
 272                  *    matters - the code can just wait on the buffer
 273                  *    lock and re-submit.
 274                  *
 275                  * 2) The current request is cached, but not
 276                  *    readahead. ocfs2_buffer_uptodate() will return
 277                  *    false anyway, so we'll wind up waiting on the
 278                  *    buffer lock to do I/O. We re-check the request
 279                  *    with after getting the lock to avoid a re-submit.
 280                  *
 281                  * 3) The current request is readahead (and so must
 282                  *    also be a caching one). We short circuit if the
 283                  *    buffer is locked (under I/O) and if it's in the
 284                  *    uptodate cache. The re-check from #2 catches the
 285                  *    case that the previous read-ahead completes just
 286                  *    before our is-it-in-flight check.
 287                  */
 288
 289                 if (!ignore_cache && !ocfs2_buffer_uptodate(ci, bh)) {
 290                         trace_ocfs2_read_blocks_from_disk(
 291                              (unsigned long long)bh->b_blocknr,
 292                              (unsigned long long)ocfs2_metadata_cache_owner(ci));
 293                         /* We're using ignore_cache here to say
 294                          * "go to disk" */
 295                         ignore_cache = 1;
 296                 }
 297
 298                 trace_ocfs2_read_blocks_bh((unsigned long long)bh->b_blocknr,
 299                         ignore_cache, buffer_jbd(bh), buffer_dirty(bh));
 300
 301                 if (buffer_jbd(bh)) {
 302                         continue;
 303                 }
 304
 305                 if (ignore_cache) {
 306                         if (buffer_dirty(bh)) {
 307                                 /* This should probably be a BUG, or
 308                                  * at least return an error. */
 309                                 continue;
 310                         }
 311
 312                         /* A read-ahead request was made - if the
 313                          * buffer is already under read-ahead from a
 314                          * previously submitted request than we are
 315                          * done here. */
 316                         if ((flags & OCFS2_BH_READAHEAD)
 317                             && ocfs2_buffer_read_ahead(ci, bh))
 318                                 continue;
 319
 320                         lock_buffer(bh);
 321                         if (buffer_jbd(bh)) {
 322 #ifdef CATCH_BH_JBD_RACES
 323                                 mlog(ML_ERROR, "block %llu had the JBD bit set "
 324                                                "while I was in lock_buffer!",
 325                                      (unsigned long long)bh->b_blocknr);
 326                                 BUG();
 327 #else
 328                                 unlock_buffer(bh);
 329                                 continue;
 330 #endif
 331                         }
 332
 333                         /* Re-check ocfs2_buffer_uptodate() as a
 334                          * previously read-ahead buffer may have
 335                          * completed I/O while we were waiting for the
 336                          * buffer lock. */
 337                         if (!(flags & OCFS2_BH_IGNORE_CACHE)
 338                             && !(flags & OCFS2_BH_READAHEAD)
 339                             && ocfs2_buffer_uptodate(ci, bh)) {
 340                                 unlock_buffer(bh);
 341                                 continue;
 342                         }
 343
 344                         clear_buffer_uptodate(bh);
 345                         get_bh(bh); /* for end_buffer_read_sync() */
 346                         if (validate)
 347                                 set_buffer_needs_validate(bh);
 348                         bh->b_end_io = end_buffer_read_sync;
 349                         submit_bh(REQ_OP_READ, 0, bh);
 350                         continue;
 351                 }
 352         }
 353
 354 read_failure:
 355         for (i = (nr - 1); i >= 0; i--) {
 356                 bh = bhs[i];
 357
 358                 if (!(flags & OCFS2_BH_READAHEAD)) {
 359                         if (unlikely(status)) {
 360                                 /* Clear the buffers on error including those
 361                                  * ever succeeded in reading
 362                                  */
 363                                 if (new_bh && bh) {
 364                                         /* If middle bh fails, let previous bh
 365                                          * finish its read and then put it to
 366                                          * aovoid bh leak
 367                                          */
 368                                         if (!buffer_jbd(bh))
 369                                                 wait_on_buffer(bh);
 370                                         put_bh(bh);
 371                                         bhs[i] = NULL;
 372                                 } else if (bh && buffer_uptodate(bh)) {
 373                                         clear_buffer_uptodate(bh);
 374                                 }
 375                                 continue;
 376                         }
 377                         /* We know this can't have changed as we hold the
 378                          * owner sem. Avoid doing any work on the bh if the
 379                          * journal has it. */
 380                         if (!buffer_jbd(bh))
 381                                 wait_on_buffer(bh);
 382
 383                         if (!buffer_uptodate(bh)) {
 384                                 /* Status won't be cleared from here on out,
 385                                  * so we can safely record this and loop back
 386                                  * to cleanup the other buffers. Don't need to
 387                                  * remove the clustered uptodate information
 388                                  * for this bh as it's not marked locally
 389                                  * uptodate. */
 390                                 status = -EIO;
 391                                 clear_buffer_needs_validate(bh);
 392                                 goto read_failure;
 393                         }
 394
 395                         if (buffer_needs_validate(bh)) {
 396                                 /* We never set NeedsValidate if the
 397                                  * buffer was held by the journal, so
 398                                  * that better not have changed */
 399                                 BUG_ON(buffer_jbd(bh));
 400                                 clear_buffer_needs_validate(bh);
 401                                 status = validate(sb, bh);
 402                                 if (status)
 403                                         goto read_failure;
 404                         }
 405                 }
 406
 407                 /* Always set the buffer in the cache, even if it was
 408                  * a forced read, or read-ahead which hasn't yet
 409                  * completed. */
 410                 ocfs2_set_buffer_uptodate(ci, bh);
 411         }
 412         ocfs2_metadata_cache_io_unlock(ci);
 413
 414         trace_ocfs2_read_blocks_end((unsigned long long)block, nr,
 415                                     flags, ignore_cache);
 416
 417 bail:
 418
 419         return status;
 420 }
 421
 422 /* Check whether the blkno is the super block or one of the backups. */
 423 static void ocfs2_check_super_or_backup(struct super_block *sb,
 424                                         sector_t blkno)
 425 {
 426         int i;
 427         u64 backup_blkno;
 428
 429         if (blkno == OCFS2_SUPER_BLOCK_BLKNO)
 430                 return;
 431
 432         for (i = 0; i < OCFS2_MAX_BACKUP_SUPERBLOCKS; i++) {
 433                 backup_blkno = ocfs2_backup_super_blkno(sb, i);
 434                 if (backup_blkno == blkno)
 435                         return;
 436         }
 437
 438         BUG();
 439 }
 440
 441 /*
 442  * Write super block and backups doesn't need to collaborate with journal,
 443  * so we don't need to lock ip_io_mutex and ci doesn't need to bea passed
 444  * into this function.
 445  */
 446 int ocfs2_write_super_or_backup(struct ocfs2_super *osb,
 447                                 struct buffer_head *bh)
 448 {
 449         int ret = 0;
 450         struct ocfs2_dinode *di = (struct ocfs2_dinode *)bh->b_data;
 451
 452         BUG_ON(buffer_jbd(bh));
 453         ocfs2_check_super_or_backup(osb->sb, bh->b_blocknr);
 454
 455         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb)) {
 456                 ret = -EROFS;
 457                 mlog_errno(ret);
 458                 goto out;
 459         }
 460
 461         lock_buffer(bh);
 462         set_buffer_uptodate(bh);
 463
 464         /* remove from dirty list before I/O. */
 465         clear_buffer_dirty(bh);
 466
 467         get_bh(bh); /* for end_buffer_write_sync() */
 468         bh->b_end_io = end_buffer_write_sync;
 469         ocfs2_compute_meta_ecc(osb->sb, bh->b_data, &di->i_check);
 470         submit_bh(REQ_OP_WRITE, 0, bh);
 471
 472         wait_on_buffer(bh);
 473
 474         if (!buffer_uptodate(bh)) {
 475                 ret = -EIO;
 476                 mlog_errno(ret);
 477         }
 478
 479 out:
 480         return ret;
 481 }