fs/ocfs2/aops.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public
  17  * License along with this program; if not, write to the
  18  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  19  * Boston, MA 021110-1307, USA.
  20  */
  21
  22 #include <linux/fs.h>
  23 #include <linux/slab.h>
  24 #include <linux/highmem.h>
  25 #include <linux/pagemap.h>
  26 #include <asm/byteorder.h>
  27 #include <linux/swap.h>
  28
  29 #define MLOG_MASK_PREFIX ML_FILE_IO
  30 #include <cluster/masklog.h>
  31
  32 #include "ocfs2.h"
  33
  34 #include "alloc.h"
  35 #include "aops.h"
  36 #include "dlmglue.h"
  37 #include "extent_map.h"
  38 #include "file.h"
  39 #include "inode.h"
  40 #include "journal.h"
  41 #include "suballoc.h"
  42 #include "super.h"
  43 #include "symlink.h"
  44
  45 #include "buffer_head_io.h"
  46
  47 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
  48                                    struct buffer_head *bh_result, int create)
  49 {
  50         int err = -EIO;
  51         int status;
  52         struct ocfs2_dinode *fe = NULL;
  53         struct buffer_head *bh = NULL;
  54         struct buffer_head *buffer_cache_bh = NULL;
  55         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  56         void *kaddr;
  57
  58         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
  59                    (unsigned long long)iblock, bh_result, create);
  60
  61         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
  62
  63         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
  64                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
  65                      (unsigned long long)iblock);
  66                 goto bail;
  67         }
  68
  69         status = ocfs2_read_block(OCFS2_SB(inode->i_sb),
  70                                   OCFS2_I(inode)->ip_blkno,
  71                                   &bh, OCFS2_BH_CACHED, inode);
  72         if (status < 0) {
  73                 mlog_errno(status);
  74                 goto bail;
  75         }
  76         fe = (struct ocfs2_dinode *) bh->b_data;
  77
  78         if (!OCFS2_IS_VALID_DINODE(fe)) {
  79                 mlog(ML_ERROR, "Invalid dinode #%llu: signature = %.*s\n",
  80                      (unsigned long long)fe->i_blkno, 7, fe->i_signature);
  81                 goto bail;
  82         }
  83
  84         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
  85                                                     le32_to_cpu(fe->i_clusters))) {
  86                 mlog(ML_ERROR, "block offset is outside the allocated size: "
  87                      "%llu\n", (unsigned long long)iblock);
  88                 goto bail;
  89         }
  90
  91         /* We don't use the page cache to create symlink data, so if
  92          * need be, copy it over from the buffer cache. */
  93         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  94                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  95                             iblock;
  96                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
  97                 if (!buffer_cache_bh) {
  98                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
  99                         goto bail;
 100                 }
 101
 102                 /* we haven't locked out transactions, so a commit
 103                  * could've happened. Since we've got a reference on
 104                  * the bh, even if it commits while we're doing the
 105                  * copy, the data is still good. */
 106                 if (buffer_jbd(buffer_cache_bh)
 107                     && ocfs2_inode_is_new(inode)) {
 108                         kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
 109                         if (!kaddr) {
 110                                 mlog(ML_ERROR, "couldn't kmap!\n");
 111                                 goto bail;
 112                         }
 113                         memcpy(kaddr + (bh_result->b_size * iblock),
 114                                buffer_cache_bh->b_data,
 115                                bh_result->b_size);
 116                         kunmap_atomic(kaddr, KM_USER0);
 117                         set_buffer_uptodate(bh_result);
 118                 }
 119                 brelse(buffer_cache_bh);
 120         }
 121
 122         map_bh(bh_result, inode->i_sb,
 123                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 124
 125         err = 0;
 126
 127 bail:
 128         if (bh)
 129                 brelse(bh);
 130
 131         mlog_exit(err);
 132         return err;
 133 }
 134
 135 static int ocfs2_get_block(struct inode *inode, sector_t iblock,
 136                            struct buffer_head *bh_result, int create)
 137 {
 138         int err = 0;
 139         u64 p_blkno, past_eof;
 140         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 141
 142         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
 143                    (unsigned long long)iblock, bh_result, create);
 144
 145         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
 146                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
 147                      inode, inode->i_ino);
 148
 149         if (S_ISLNK(inode->i_mode)) {
 150                 /* this always does I/O for some reason. */
 151                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
 152                 goto bail;
 153         }
 154
 155         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, NULL);
 156         if (err) {
 157                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
 158                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
 159                      (unsigned long long)p_blkno);
 160                 goto bail;
 161         }
 162
 163         /*
 164          * ocfs2 never allocates in this function - the only time we
 165          * need to use BH_New is when we're extending i_size on a file
 166          * system which doesn't support holes, in which case BH_New
 167          * allows block_prepare_write() to zero.
 168          */
 169         mlog_bug_on_msg(create && p_blkno == 0 && ocfs2_sparse_alloc(osb),
 170                         "ino %lu, iblock %llu\n", inode->i_ino,
 171                         (unsigned long long)iblock);
 172
 173         if (p_blkno)
 174                 map_bh(bh_result, inode->i_sb, p_blkno);
 175
 176         if (!ocfs2_sparse_alloc(osb)) {
 177                 if (p_blkno == 0) {
 178                         err = -EIO;
 179                         mlog(ML_ERROR,
 180                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
 181                              (unsigned long long)iblock,
 182                              (unsigned long long)p_blkno,
 183                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
 184                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
 185                         dump_stack();
 186                 }
 187
 188                 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 189                 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
 190                      (unsigned long long)past_eof);
 191
 192                 if (create && (iblock >= past_eof))
 193                         set_buffer_new(bh_result);
 194         }
 195
 196 bail:
 197         if (err < 0)
 198                 err = -EIO;
 199
 200         mlog_exit(err);
 201         return err;
 202 }
 203
 204 static int ocfs2_readpage(struct file *file, struct page *page)
 205 {
 206         struct inode *inode = page->mapping->host;
 207         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
 208         int ret, unlock = 1;
 209
 210         mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0));
 211
 212         ret = ocfs2_meta_lock_with_page(inode, NULL, 0, page);
 213         if (ret != 0) {
 214                 if (ret == AOP_TRUNCATED_PAGE)
 215                         unlock = 0;
 216                 mlog_errno(ret);
 217                 goto out;
 218         }
 219
 220         down_read(&OCFS2_I(inode)->ip_alloc_sem);
 221
 222         /*
 223          * i_size might have just been updated as we grabed the meta lock.  We
 224          * might now be discovering a truncate that hit on another node.
 225          * block_read_full_page->get_block freaks out if it is asked to read
 226          * beyond the end of a file, so we check here.  Callers
 227          * (generic_file_read, fault->nopage) are clever enough to check i_size
 228          * and notice that the page they just read isn't needed.
 229          *
 230          * XXX sys_readahead() seems to get that wrong?
 231          */
 232         if (start >= i_size_read(inode)) {
 233                 char *addr = kmap(page);
 234                 memset(addr, 0, PAGE_SIZE);
 235                 flush_dcache_page(page);
 236                 kunmap(page);
 237                 SetPageUptodate(page);
 238                 ret = 0;
 239                 goto out_alloc;
 240         }
 241
 242         ret = ocfs2_data_lock_with_page(inode, 0, page);
 243         if (ret != 0) {
 244                 if (ret == AOP_TRUNCATED_PAGE)
 245                         unlock = 0;
 246                 mlog_errno(ret);
 247                 goto out_alloc;
 248         }
 249
 250         ret = block_read_full_page(page, ocfs2_get_block);
 251         unlock = 0;
 252
 253         ocfs2_data_unlock(inode, 0);
 254 out_alloc:
 255         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 256         ocfs2_meta_unlock(inode, 0);
 257 out:
 258         if (unlock)
 259                 unlock_page(page);
 260         mlog_exit(ret);
 261         return ret;
 262 }
 263
 264 /* Note: Because we don't support holes, our allocation has
 265  * already happened (allocation writes zeros to the file data)
 266  * so we don't have to worry about ordered writes in
 267  * ocfs2_writepage.
 268  *
 269  * ->writepage is called during the process of invalidating the page cache
 270  * during blocked lock processing.  It can't block on any cluster locks
 271  * to during block mapping.  It's relying on the fact that the block
 272  * mapping can't have disappeared under the dirty pages that it is
 273  * being asked to write back.
 274  */
 275 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
 276 {
 277         int ret;
 278
 279         mlog_entry("(0x%p)\n", page);
 280
 281         ret = block_write_full_page(page, ocfs2_get_block, wbc);
 282
 283         mlog_exit(ret);
 284
 285         return ret;
 286 }
 287
 288 /*
 289  * This is called from ocfs2_write_zero_page() which has handled it's
 290  * own cluster locking and has ensured allocation exists for those
 291  * blocks to be written.
 292  */
 293 int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
 294                                unsigned from, unsigned to)
 295 {
 296         int ret;
 297
 298         down_read(&OCFS2_I(inode)->ip_alloc_sem);
 299
 300         ret = block_prepare_write(page, from, to, ocfs2_get_block);
 301
 302         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 303
 304         return ret;
 305 }
 306
 307 /* Taken from ext3. We don't necessarily need the full blown
 308  * functionality yet, but IMHO it's better to cut and paste the whole
 309  * thing so we can avoid introducing our own bugs (and easily pick up
 310  * their fixes when they happen) --Mark */
 311 static int walk_page_buffers(   handle_t *handle,
 312                                 struct buffer_head *head,
 313                                 unsigned from,
 314                                 unsigned to,
 315                                 int *partial,
 316                                 int (*fn)(      handle_t *handle,
 317                                                 struct buffer_head *bh))
 318 {
 319         struct buffer_head *bh;
 320         unsigned block_start, block_end;
 321         unsigned blocksize = head->b_size;
 322         int err, ret = 0;
 323         struct buffer_head *next;
 324
 325         for (   bh = head, block_start = 0;
 326                 ret == 0 && (bh != head || !block_start);
 327                 block_start = block_end, bh = next)
 328         {
 329                 next = bh->b_this_page;
 330                 block_end = block_start + blocksize;
 331                 if (block_end <= from || block_start >= to) {
 332                         if (partial && !buffer_uptodate(bh))
 333                                 *partial = 1;
 334                         continue;
 335                 }
 336                 err = (*fn)(handle, bh);
 337                 if (!ret)
 338                         ret = err;
 339         }
 340         return ret;
 341 }
 342
 343 handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
 344                                                          struct page *page,
 345                                                          unsigned from,
 346                                                          unsigned to)
 347 {
 348         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 349         handle_t *handle = NULL;
 350         int ret = 0;
 351
 352         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 353         if (!handle) {
 354                 ret = -ENOMEM;
 355                 mlog_errno(ret);
 356                 goto out;
 357         }
 358
 359         if (ocfs2_should_order_data(inode)) {
 360                 ret = walk_page_buffers(handle,
 361                                         page_buffers(page),
 362                                         from, to, NULL,
 363                                         ocfs2_journal_dirty_data);
 364                 if (ret < 0)
 365                         mlog_errno(ret);
 366         }
 367 out:
 368         if (ret) {
 369                 if (handle)
 370                         ocfs2_commit_trans(osb, handle);
 371                 handle = ERR_PTR(ret);
 372         }
 373         return handle;
 374 }
 375
 376 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 377 {
 378         sector_t status;
 379         u64 p_blkno = 0;
 380         int err = 0;
 381         struct inode *inode = mapping->host;
 382
 383         mlog_entry("(block = %llu)\n", (unsigned long long)block);
 384
 385         /* We don't need to lock journal system files, since they aren't
 386          * accessed concurrently from multiple nodes.
 387          */
 388         if (!INODE_JOURNAL(inode)) {
 389                 err = ocfs2_meta_lock(inode, NULL, 0);
 390                 if (err) {
 391                         if (err != -ENOENT)
 392                                 mlog_errno(err);
 393                         goto bail;
 394                 }
 395                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
 396         }
 397
 398         err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL);
 399
 400         if (!INODE_JOURNAL(inode)) {
 401                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
 402                 ocfs2_meta_unlock(inode, 0);
 403         }
 404
 405         if (err) {
 406                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 407                      (unsigned long long)block);
 408                 mlog_errno(err);
 409                 goto bail;
 410         }
 411
 412
 413 bail:
 414         status = err ? 0 : p_blkno;
 415
 416         mlog_exit((int)status);
 417
 418         return status;
 419 }
 420
 421 /*
 422  * TODO: Make this into a generic get_blocks function.
 423  *
 424  * From do_direct_io in direct-io.c:
 425  *  "So what we do is to permit the ->get_blocks function to populate
 426  *   bh.b_size with the size of IO which is permitted at this offset and
 427  *   this i_blkbits."
 428  *
 429  * This function is called directly from get_more_blocks in direct-io.c.
 430  *
 431  * called like this: dio->get_blocks(dio->inode, fs_startblk,
 432  *                                      fs_count, map_bh, dio->rw == WRITE);
 433  */
 434 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
 435                                      struct buffer_head *bh_result, int create)
 436 {
 437         int ret;
 438         u64 p_blkno, inode_blocks;
 439         int contig_blocks;
 440         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
 441         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
 442
 443         /* This function won't even be called if the request isn't all
 444          * nicely aligned and of the right size, so there's no need
 445          * for us to check any of that. */
 446
 447         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 448
 449         /*
 450          * Any write past EOF is not allowed because we'd be extending.
 451          */
 452         if (create && (iblock + max_blocks) > inode_blocks) {
 453                 ret = -EIO;
 454                 goto bail;
 455         }
 456
 457         /* This figures out the size of the next contiguous block, and
 458          * our logical offset */
 459         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
 460                                           &contig_blocks);
 461         if (ret) {
 462                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
 463                      (unsigned long long)iblock);
 464                 ret = -EIO;
 465                 goto bail;
 466         }
 467
 468         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)) && !p_blkno) {
 469                 ocfs2_error(inode->i_sb,
 470                             "Inode %llu has a hole at block %llu\n",
 471                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 472                             (unsigned long long)iblock);
 473                 ret = -EROFS;
 474                 goto bail;
 475         }
 476
 477         /*
 478          * get_more_blocks() expects us to describe a hole by clearing
 479          * the mapped bit on bh_result().
 480          */
 481         if (p_blkno)
 482                 map_bh(bh_result, inode->i_sb, p_blkno);
 483         else {
 484                 /*
 485                  * ocfs2_prepare_inode_for_write() should have caught
 486                  * the case where we'd be filling a hole and triggered
 487                  * a buffered write instead.
 488                  */
 489                 if (create) {
 490                         ret = -EIO;
 491                         mlog_errno(ret);
 492                         goto bail;
 493                 }
 494
 495                 clear_buffer_mapped(bh_result);
 496         }
 497
 498         /* make sure we don't map more than max_blocks blocks here as
 499            that's all the kernel will handle at this point. */
 500         if (max_blocks < contig_blocks)
 501                 contig_blocks = max_blocks;
 502         bh_result->b_size = contig_blocks << blocksize_bits;
 503 bail:
 504         return ret;
 505 }
 506
 507 /*
 508  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
 509  * particularly interested in the aio/dio case.  Like the core uses
 510  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
 511  * truncation on another.
 512  */
 513 static void ocfs2_dio_end_io(struct kiocb *iocb,
 514                              loff_t offset,
 515                              ssize_t bytes,
 516                              void *private)
 517 {
 518         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
 519
 520         /* this io's submitter should not have unlocked this before we could */
 521         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 522         ocfs2_iocb_clear_rw_locked(iocb);
 523         up_read(&inode->i_alloc_sem);
 524         ocfs2_rw_unlock(inode, 0);
 525 }
 526
 527 /*
 528  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
 529  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
 530  * do journalled data.
 531  */
 532 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
 533 {
 534         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 535
 536         journal_invalidatepage(journal, page, offset);
 537 }
 538
 539 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 540 {
 541         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 542
 543         if (!page_has_buffers(page))
 544                 return 0;
 545         return journal_try_to_free_buffers(journal, page, wait);
 546 }
 547
 548 static ssize_t ocfs2_direct_IO(int rw,
 549                                struct kiocb *iocb,
 550                                const struct iovec *iov,
 551                                loff_t offset,
 552                                unsigned long nr_segs)
 553 {
 554         struct file *file = iocb->ki_filp;
 555         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
 556         int ret;
 557
 558         mlog_entry_void();
 559
 560         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
 561                 /*
 562                  * We get PR data locks even for O_DIRECT.  This
 563                  * allows concurrent O_DIRECT I/O but doesn't let
 564                  * O_DIRECT with extending and buffered zeroing writes
 565                  * race.  If they did race then the buffered zeroing
 566                  * could be written back after the O_DIRECT I/O.  It's
 567                  * one thing to tell people not to mix buffered and
 568                  * O_DIRECT writes, but expecting them to understand
 569                  * that file extension is also an implicit buffered
 570                  * write is too much.  By getting the PR we force
 571                  * writeback of the buffered zeroing before
 572                  * proceeding.
 573                  */
 574                 ret = ocfs2_data_lock(inode, 0);
 575                 if (ret < 0) {
 576                         mlog_errno(ret);
 577                         goto out;
 578                 }
 579                 ocfs2_data_unlock(inode, 0);
 580         }
 581
 582         ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
 583                                             inode->i_sb->s_bdev, iov, offset,
 584                                             nr_segs,
 585                                             ocfs2_direct_IO_get_blocks,
 586                                             ocfs2_dio_end_io);
 587 out:
 588         mlog_exit(ret);
 589         return ret;
 590 }
 591
 592 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 593                                             u32 cpos,
 594                                             unsigned int *start,
 595                                             unsigned int *end)
 596 {
 597         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
 598
 599         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
 600                 unsigned int cpp;
 601
 602                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
 603
 604                 cluster_start = cpos % cpp;
 605                 cluster_start = cluster_start << osb->s_clustersize_bits;
 606
 607                 cluster_end = cluster_start + osb->s_clustersize;
 608         }
 609
 610         BUG_ON(cluster_start > PAGE_SIZE);
 611         BUG_ON(cluster_end > PAGE_SIZE);
 612
 613         if (start)
 614                 *start = cluster_start;
 615         if (end)
 616                 *end = cluster_end;
 617 }
 618
 619 /*
 620  * 'from' and 'to' are the region in the page to avoid zeroing.
 621  *
 622  * If pagesize > clustersize, this function will avoid zeroing outside
 623  * of the cluster boundary.
 624  *
 625  * from == to == 0 is code for "zero the entire cluster region"
 626  */
 627 static void ocfs2_clear_page_regions(struct page *page,
 628                                      struct ocfs2_super *osb, u32 cpos,
 629                                      unsigned from, unsigned to)
 630 {
 631         void *kaddr;
 632         unsigned int cluster_start, cluster_end;
 633
 634         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 635
 636         kaddr = kmap_atomic(page, KM_USER0);
 637
 638         if (from || to) {
 639                 if (from > cluster_start)
 640                         memset(kaddr + cluster_start, 0, from - cluster_start);
 641                 if (to < cluster_end)
 642                         memset(kaddr + to, 0, cluster_end - to);
 643         } else {
 644                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 645         }
 646
 647         kunmap_atomic(kaddr, KM_USER0);
 648 }
 649
 650 /*
 651  * Some of this taken from block_prepare_write(). We already have our
 652  * mapping by now though, and the entire write will be allocating or
 653  * it won't, so not much need to use BH_New.
 654  *
 655  * This will also skip zeroing, which is handled externally.
 656  */
 657 static int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 658                                  struct inode *inode, unsigned int from,
 659                                  unsigned int to, int new)
 660 {
 661         int ret = 0;
 662         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 663         unsigned int block_end, block_start;
 664         unsigned int bsize = 1 << inode->i_blkbits;
 665
 666         if (!page_has_buffers(page))
 667                 create_empty_buffers(page, bsize, 0);
 668
 669         head = page_buffers(page);
 670         for (bh = head, block_start = 0; bh != head || !block_start;
 671              bh = bh->b_this_page, block_start += bsize) {
 672                 block_end = block_start + bsize;
 673
 674                 /*
 675                  * Ignore blocks outside of our i/o range -
 676                  * they may belong to unallocated clusters.
 677                  */
 678                 if (block_start >= to ||
 679                     (block_start + bsize) <= from) {
 680                         if (PageUptodate(page))
 681                                 set_buffer_uptodate(bh);
 682                         continue;
 683                 }
 684
 685                 /*
 686                  * For an allocating write with cluster size >= page
 687                  * size, we always write the entire page.
 688                  */
 689
 690                 if (buffer_new(bh))
 691                         clear_buffer_new(bh);
 692
 693                 if (!buffer_mapped(bh)) {
 694                         map_bh(bh, inode->i_sb, *p_blkno);
 695                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 696                 }
 697
 698                 if (PageUptodate(page)) {
 699                         if (!buffer_uptodate(bh))
 700                                 set_buffer_uptodate(bh);
 701                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 702                      (block_start < from || block_end > to)) {
 703                         ll_rw_block(READ, 1, &bh);
 704                         *wait_bh++=bh;
 705                 }
 706
 707                 *p_blkno = *p_blkno + 1;
 708         }
 709
 710         /*
 711          * If we issued read requests - let them complete.
 712          */
 713         while(wait_bh > wait) {
 714                 wait_on_buffer(*--wait_bh);
 715                 if (!buffer_uptodate(*wait_bh))
 716                         ret = -EIO;
 717         }
 718
 719         if (ret == 0 || !new)
 720                 return ret;
 721
 722         /*
 723          * If we get -EIO above, zero out any newly allocated blocks
 724          * to avoid exposing stale data.
 725          */
 726         bh = head;
 727         block_start = 0;
 728         do {
 729                 void *kaddr;
 730
 731                 block_end = block_start + bsize;
 732                 if (block_end <= from)
 733                         goto next_bh;
 734                 if (block_start >= to)
 735                         break;
 736
 737                 kaddr = kmap_atomic(page, KM_USER0);
 738                 memset(kaddr+block_start, 0, bh->b_size);
 739                 flush_dcache_page(page);
 740                 kunmap_atomic(kaddr, KM_USER0);
 741                 set_buffer_uptodate(bh);
 742                 mark_buffer_dirty(bh);
 743
 744 next_bh:
 745                 block_start = block_end;
 746                 bh = bh->b_this_page;
 747         } while (bh != head);
 748
 749         return ret;
 750 }
 751
 752 /*
 753  * This will copy user data from the iovec in the buffered write
 754  * context.
 755  */
 756 int ocfs2_map_and_write_user_data(struct inode *inode,
 757                                   struct ocfs2_write_ctxt *wc, u64 *p_blkno,
 758                                   unsigned int *ret_from, unsigned int *ret_to)
 759 {
 760         int ret;
 761         unsigned int to, from, cluster_start, cluster_end;
 762         unsigned long bytes, src_from;
 763         char *dst;
 764         struct ocfs2_buffered_write_priv *bp = wc->w_private;
 765         const struct iovec *cur_iov = bp->b_cur_iov;
 766         char __user *buf;
 767         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 768
 769         ocfs2_figure_cluster_boundaries(osb, wc->w_cpos, &cluster_start,
 770                                         &cluster_end);
 771
 772         buf = cur_iov->iov_base + bp->b_cur_off;
 773         src_from = (unsigned long)buf & ~PAGE_CACHE_MASK;
 774
 775         from = wc->w_pos & (PAGE_CACHE_SIZE - 1);
 776
 777         /*
 778          * This is a lot of comparisons, but it reads quite
 779          * easily, which is important here.
 780          */
 781         /* Stay within the src page */
 782         bytes = PAGE_SIZE - src_from;
 783         /* Stay within the vector */
 784         bytes = min(bytes,
 785                     (unsigned long)(cur_iov->iov_len - bp->b_cur_off));
 786         /* Stay within count */
 787         bytes = min(bytes, (unsigned long)wc->w_count);
 788         /*
 789          * For clustersize > page size, just stay within
 790          * target page, otherwise we have to calculate pos
 791          * within the cluster and obey the rightmost
 792          * boundary.
 793          */
 794         if (wc->w_large_pages) {
 795                 /*
 796                  * For cluster size < page size, we have to
 797                  * calculate pos within the cluster and obey
 798                  * the rightmost boundary.
 799                  */
 800                 bytes = min(bytes, (unsigned long)(osb->s_clustersize
 801                                    - (wc->w_pos & (osb->s_clustersize - 1))));
 802         } else {
 803                 /*
 804                  * cluster size > page size is the most common
 805                  * case - we just stay within the target page
 806                  * boundary.
 807                  */
 808                 bytes = min(bytes, PAGE_CACHE_SIZE - from);
 809         }
 810
 811         to = from + bytes;
 812
 813         if (wc->w_this_page_new)
 814                 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode,
 815                                             cluster_start, cluster_end, 1);
 816         else
 817                 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode,
 818                                             from, to, 0);
 819         if (ret) {
 820                 mlog_errno(ret);
 821                 goto out;
 822         }
 823
 824         BUG_ON(from > PAGE_CACHE_SIZE);
 825         BUG_ON(to > PAGE_CACHE_SIZE);
 826         BUG_ON(from > osb->s_clustersize);
 827         BUG_ON(to > osb->s_clustersize);
 828
 829         dst = kmap(wc->w_this_page);
 830         memcpy(dst + from, bp->b_src_buf + src_from, bytes);
 831         kunmap(wc->w_this_page);
 832
 833         /*
 834          * XXX: This is slow, but simple. The caller of
 835          * ocfs2_buffered_write_cluster() is responsible for
 836          * passing through the iovecs, so it's difficult to
 837          * predict what our next step is in here after our
 838          * initial write. A future version should be pushing
 839          * that iovec manipulation further down.
 840          *
 841          * By setting this, we indicate that a copy from user
 842          * data was done, and subsequent calls for this
 843          * cluster will skip copying more data.
 844          */
 845         wc->w_finished_copy = 1;
 846
 847         *ret_from = from;
 848         *ret_to = to;
 849 out:
 850
 851         return bytes ? (unsigned int)bytes : ret;
 852 }
 853
 854 /*
 855  * Map, fill and write a page to disk.
 856  *
 857  * The work of copying data is done via callback.  Newly allocated
 858  * pages which don't take user data will be zero'd (set 'new' to
 859  * indicate an allocating write)
 860  *
 861  * Returns a negative error code or the number of bytes copied into
 862  * the page.
 863  */
 864 int ocfs2_write_data_page(struct inode *inode, handle_t *handle,
 865                           u64 *p_blkno, struct page *page,
 866                           struct ocfs2_write_ctxt *wc, int new)
 867 {
 868         int ret, copied = 0;
 869         unsigned int from = 0, to = 0;
 870         unsigned int cluster_start, cluster_end;
 871         unsigned int zero_from = 0, zero_to = 0;
 872
 873         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), wc->w_cpos,
 874                                         &cluster_start, &cluster_end);
 875
 876         if ((wc->w_pos >> PAGE_CACHE_SHIFT) == page->index
 877             && !wc->w_finished_copy) {
 878
 879                 wc->w_this_page = page;
 880                 wc->w_this_page_new = new;
 881                 ret = wc->w_write_data_page(inode, wc, p_blkno, &from, &to);
 882                 if (ret < 0) {
 883                         mlog_errno(ret);
 884                         goto out;
 885                 }
 886
 887                 copied = ret;
 888
 889                 zero_from = from;
 890                 zero_to = to;
 891                 if (new) {
 892                         from = cluster_start;
 893                         to = cluster_end;
 894                 }
 895         } else {
 896                 /*
 897                  * If we haven't allocated the new page yet, we
 898                  * shouldn't be writing it out without copying user
 899                  * data. This is likely a math error from the caller.
 900                  */
 901                 BUG_ON(!new);
 902
 903                 from = cluster_start;
 904                 to = cluster_end;
 905
 906                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 907                                             cluster_start, cluster_end, 1);
 908                 if (ret) {
 909                         mlog_errno(ret);
 910                         goto out;
 911                 }
 912         }
 913
 914         /*
 915          * Parts of newly allocated pages need to be zero'd.
 916          *
 917          * Above, we have also rewritten 'to' and 'from' - as far as
 918          * the rest of the function is concerned, the entire cluster
 919          * range inside of a page needs to be written.
 920          *
 921          * We can skip this if the page is up to date - it's already
 922          * been zero'd from being read in as a hole.
 923          */
 924         if (new && !PageUptodate(page))
 925                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
 926                                          wc->w_cpos, zero_from, zero_to);
 927
 928         flush_dcache_page(page);
 929
 930         if (ocfs2_should_order_data(inode)) {
 931                 ret = walk_page_buffers(handle,
 932                                         page_buffers(page),
 933                                         from, to, NULL,
 934                                         ocfs2_journal_dirty_data);
 935                 if (ret < 0)
 936                         mlog_errno(ret);
 937         }
 938
 939         /*
 940          * We don't use generic_commit_write() because we need to
 941          * handle our own i_size update.
 942          */
 943         ret = block_commit_write(page, from, to);
 944         if (ret)
 945                 mlog_errno(ret);
 946 out:
 947
 948         return copied ? copied : ret;
 949 }
 950
 951 /*
 952  * Do the actual write of some data into an inode. Optionally allocate
 953  * in order to fulfill the write.
 954  *
 955  * cpos is the logical cluster offset within the file to write at
 956  *
 957  * 'phys' is the physical mapping of that offset. a 'phys' value of
 958  * zero indicates that allocation is required. In this case, data_ac
 959  * and meta_ac should be valid (meta_ac can be null if metadata
 960  * allocation isn't required).
 961  */
 962 static ssize_t ocfs2_write(struct file *file, u32 phys, handle_t *handle,
 963                            struct buffer_head *di_bh,
 964                            struct ocfs2_alloc_context *data_ac,
 965                            struct ocfs2_alloc_context *meta_ac,
 966                            struct ocfs2_write_ctxt *wc)
 967 {
 968         int ret, i, numpages = 1, new;
 969         unsigned int copied = 0;
 970         u32 tmp_pos;
 971         u64 v_blkno, p_blkno;
 972         struct address_space *mapping = file->f_mapping;
 973         struct inode *inode = mapping->host;
 974         unsigned int cbits = OCFS2_SB(inode->i_sb)->s_clustersize_bits;
 975         unsigned long index, start;
 976         struct page **cpages;
 977
 978         new = phys == 0 ? 1 : 0;
 979
 980         /*
 981          * Figure out how many pages we'll be manipulating here. For
 982          * non-allocating write, or any writes where cluster size is
 983          * less than page size, we only need one page. Otherwise,
 984          * allocating writes of cluster size larger than page size
 985          * need cluster size pages.
 986          */
 987         if (new && !wc->w_large_pages)
 988                 numpages = (1 << cbits) / PAGE_SIZE;
 989
 990         cpages = kzalloc(sizeof(*cpages) * numpages, GFP_NOFS);
 991         if (!cpages) {
 992                 ret = -ENOMEM;
 993                 mlog_errno(ret);
 994                 return ret;
 995         }
 996
 997         /*
 998          * Fill our page array first. That way we've grabbed enough so
 999          * that we can zero and flush if we error after adding the
1000          * extent.
1001          */
1002         if (new) {
1003                 start = ocfs2_align_clusters_to_page_index(inode->i_sb,
1004                                                            wc->w_cpos);
1005                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, wc->w_cpos);
1006         } else {
1007                 start = wc->w_pos >> PAGE_CACHE_SHIFT;
1008                 v_blkno = wc->w_pos >> inode->i_sb->s_blocksize_bits;
1009         }
1010
1011         for(i = 0; i < numpages; i++) {
1012                 index = start + i;
1013
1014                 cpages[i] = grab_cache_page(mapping, index);
1015                 if (!cpages[i]) {
1016                         ret = -ENOMEM;
1017                         mlog_errno(ret);
1018                         goto out;
1019                 }
1020         }
1021
1022         if (new) {
1023                 /*
1024                  * This is safe to call with the page locks - it won't take
1025                  * any additional semaphores or cluster locks.
1026                  */
1027                 tmp_pos = wc->w_cpos;
1028                 ret = ocfs2_do_extend_allocation(OCFS2_SB(inode->i_sb), inode,
1029                                                  &tmp_pos, 1, di_bh, handle,
1030                                                  data_ac, meta_ac, NULL);
1031                 /*
1032                  * This shouldn't happen because we must have already
1033                  * calculated the correct meta data allocation required. The
1034                  * internal tree allocation code should know how to increase
1035                  * transaction credits itself.
1036                  *
1037                  * If need be, we could handle -EAGAIN for a
1038                  * RESTART_TRANS here.
1039                  */
1040                 mlog_bug_on_msg(ret == -EAGAIN,
1041                                 "Inode %llu: EAGAIN return during allocation.\n",
1042                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1043                 if (ret < 0) {
1044                         mlog_errno(ret);
1045                         goto out;
1046                 }
1047         }
1048
1049         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL);
1050         if (ret < 0) {
1051
1052                 /*
1053                  * XXX: Should we go readonly here?
1054                  */
1055
1056                 mlog_errno(ret);
1057                 goto out;
1058         }
1059
1060         BUG_ON(p_blkno == 0);
1061
1062         for(i = 0; i < numpages; i++) {
1063                 ret = ocfs2_write_data_page(inode, handle, &p_blkno, cpages[i],
1064                                             wc, new);
1065                 if (ret < 0) {
1066                         mlog_errno(ret);
1067                         goto out;
1068                 }
1069
1070                 copied += ret;
1071         }
1072
1073 out:
1074         for(i = 0; i < numpages; i++) {
1075                 unlock_page(cpages[i]);
1076                 mark_page_accessed(cpages[i]);
1077                 page_cache_release(cpages[i]);
1078         }
1079         kfree(cpages);
1080
1081         return copied ? copied : ret;
1082 }
1083
1084 static void ocfs2_write_ctxt_init(struct ocfs2_write_ctxt *wc,
1085                                   struct ocfs2_super *osb, loff_t pos,
1086                                   size_t count, ocfs2_page_writer *cb,
1087                                   void *cb_priv)
1088 {
1089         wc->w_count = count;
1090         wc->w_pos = pos;
1091         wc->w_cpos = wc->w_pos >> osb->s_clustersize_bits;
1092         wc->w_finished_copy = 0;
1093
1094         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
1095                 wc->w_large_pages = 1;
1096         else
1097                 wc->w_large_pages = 0;
1098
1099         wc->w_write_data_page = cb;
1100         wc->w_private = cb_priv;
1101 }
1102
1103 /*
1104  * Write a cluster to an inode. The cluster may not be allocated yet,
1105  * in which case it will be. This only exists for buffered writes -
1106  * O_DIRECT takes a more "traditional" path through the kernel.
1107  *
1108  * The caller is responsible for incrementing pos, written counts, etc
1109  *
1110  * For file systems that don't support sparse files, pre-allocation
1111  * and page zeroing up until cpos should be done prior to this
1112  * function call.
1113  *
1114  * Callers should be holding i_sem, and the rw cluster lock.
1115  *
1116  * Returns the number of user bytes written, or less than zero for
1117  * error.
1118  */
1119 ssize_t ocfs2_buffered_write_cluster(struct file *file, loff_t pos,
1120                                      size_t count, ocfs2_page_writer *actor,
1121                                      void *priv)
1122 {
1123         int ret, credits = OCFS2_INODE_UPDATE_CREDITS;
1124         ssize_t written = 0;
1125         u32 phys;
1126         struct inode *inode = file->f_mapping->host;
1127         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1128         struct buffer_head *di_bh = NULL;
1129         struct ocfs2_dinode *di;
1130         struct ocfs2_alloc_context *data_ac = NULL;
1131         struct ocfs2_alloc_context *meta_ac = NULL;
1132         handle_t *handle;
1133         struct ocfs2_write_ctxt wc;
1134
1135         ocfs2_write_ctxt_init(&wc, osb, pos, count, actor, priv);
1136
1137         ret = ocfs2_meta_lock(inode, &di_bh, 1);
1138         if (ret) {
1139                 mlog_errno(ret);
1140                 goto out;
1141         }
1142         di = (struct ocfs2_dinode *)di_bh->b_data;
1143
1144         /*
1145          * Take alloc sem here to prevent concurrent lookups. That way
1146          * the mapping, zeroing and tree manipulation within
1147          * ocfs2_write() will be safe against ->readpage(). This
1148          * should also serve to lock out allocation from a shared
1149          * writeable region.
1150          */
1151         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1152
1153         ret = ocfs2_get_clusters(inode, wc.w_cpos, &phys, NULL);
1154         if (ret) {
1155                 mlog_errno(ret);
1156                 goto out_meta;
1157         }
1158
1159         /* phys == 0 means that allocation is required. */
1160         if (phys == 0) {
1161                 ret = ocfs2_lock_allocators(inode, di, 1, &data_ac, &meta_ac);
1162                 if (ret) {
1163                         mlog_errno(ret);
1164                         goto out_meta;
1165                 }
1166
1167                 credits = ocfs2_calc_extend_credits(inode->i_sb, di, 1);
1168         }
1169
1170         ret = ocfs2_data_lock(inode, 1);
1171         if (ret) {
1172                 mlog_errno(ret);
1173                 goto out_meta;
1174         }
1175
1176         handle = ocfs2_start_trans(osb, credits);
1177         if (IS_ERR(handle)) {
1178                 ret = PTR_ERR(handle);
1179                 mlog_errno(ret);
1180                 goto out_data;
1181         }
1182
1183         written = ocfs2_write(file, phys, handle, di_bh, data_ac,
1184                               meta_ac, &wc);
1185         if (written < 0) {
1186                 ret = written;
1187                 mlog_errno(ret);
1188                 goto out_commit;
1189         }
1190
1191         ret = ocfs2_journal_access(handle, inode, di_bh,
1192                                    OCFS2_JOURNAL_ACCESS_WRITE);
1193         if (ret) {
1194                 mlog_errno(ret);
1195                 goto out_commit;
1196         }
1197
1198         pos += written;
1199         if (pos > inode->i_size) {
1200                 i_size_write(inode, pos);
1201                 mark_inode_dirty(inode);
1202         }
1203         inode->i_blocks = ocfs2_align_bytes_to_sectors((u64)(i_size_read(inode)));
1204         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1205         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1206         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1207         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1208
1209         ret = ocfs2_journal_dirty(handle, di_bh);
1210         if (ret)
1211                 mlog_errno(ret);
1212
1213 out_commit:
1214         ocfs2_commit_trans(osb, handle);
1215
1216 out_data:
1217         ocfs2_data_unlock(inode, 1);
1218
1219 out_meta:
1220         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1221         ocfs2_meta_unlock(inode, 1);
1222
1223 out:
1224         brelse(di_bh);
1225         if (data_ac)
1226                 ocfs2_free_alloc_context(data_ac);
1227         if (meta_ac)
1228                 ocfs2_free_alloc_context(meta_ac);
1229
1230         return written ? written : ret;
1231 }
1232
1233 const struct address_space_operations ocfs2_aops = {
1234         .readpage       = ocfs2_readpage,
1235         .writepage      = ocfs2_writepage,
1236         .bmap           = ocfs2_bmap,
1237         .sync_page      = block_sync_page,
1238         .direct_IO      = ocfs2_direct_IO,
1239         .invalidatepage = ocfs2_invalidatepage,
1240         .releasepage    = ocfs2_releasepage,
1241         .migratepage    = buffer_migrate_page,
1242 };