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