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 #include <linux/pipe_fs_i.h>
  29 #include <linux/mpage.h>
  30
  31 #define MLOG_MASK_PREFIX ML_FILE_IO
  32 #include <cluster/masklog.h>
  33
  34 #include "ocfs2.h"
  35
  36 #include "alloc.h"
  37 #include "aops.h"
  38 #include "dlmglue.h"
  39 #include "extent_map.h"
  40 #include "file.h"
  41 #include "inode.h"
  42 #include "journal.h"
  43 #include "suballoc.h"
  44 #include "super.h"
  45 #include "symlink.h"
  46
  47 #include "buffer_head_io.h"
  48
  49 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
  50                                    struct buffer_head *bh_result, int create)
  51 {
  52         int err = -EIO;
  53         int status;
  54         struct ocfs2_dinode *fe = NULL;
  55         struct buffer_head *bh = NULL;
  56         struct buffer_head *buffer_cache_bh = NULL;
  57         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  58         void *kaddr;
  59
  60         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
  61                    (unsigned long long)iblock, bh_result, create);
  62
  63         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
  64
  65         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
  66                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
  67                      (unsigned long long)iblock);
  68                 goto bail;
  69         }
  70
  71         status = ocfs2_read_block(OCFS2_SB(inode->i_sb),
  72                                   OCFS2_I(inode)->ip_blkno,
  73                                   &bh, OCFS2_BH_CACHED, inode);
  74         if (status < 0) {
  75                 mlog_errno(status);
  76                 goto bail;
  77         }
  78         fe = (struct ocfs2_dinode *) bh->b_data;
  79
  80         if (!OCFS2_IS_VALID_DINODE(fe)) {
  81                 mlog(ML_ERROR, "Invalid dinode #%llu: signature = %.*s\n",
  82                      (unsigned long long)le64_to_cpu(fe->i_blkno), 7,
  83                      fe->i_signature);
  84                 goto bail;
  85         }
  86
  87         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
  88                                                     le32_to_cpu(fe->i_clusters))) {
  89                 mlog(ML_ERROR, "block offset is outside the allocated size: "
  90                      "%llu\n", (unsigned long long)iblock);
  91                 goto bail;
  92         }
  93
  94         /* We don't use the page cache to create symlink data, so if
  95          * need be, copy it over from the buffer cache. */
  96         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  97                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  98                             iblock;
  99                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
 100                 if (!buffer_cache_bh) {
 101                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
 102                         goto bail;
 103                 }
 104
 105                 /* we haven't locked out transactions, so a commit
 106                  * could've happened. Since we've got a reference on
 107                  * the bh, even if it commits while we're doing the
 108                  * copy, the data is still good. */
 109                 if (buffer_jbd(buffer_cache_bh)
 110                     && ocfs2_inode_is_new(inode)) {
 111                         kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
 112                         if (!kaddr) {
 113                                 mlog(ML_ERROR, "couldn't kmap!\n");
 114                                 goto bail;
 115                         }
 116                         memcpy(kaddr + (bh_result->b_size * iblock),
 117                                buffer_cache_bh->b_data,
 118                                bh_result->b_size);
 119                         kunmap_atomic(kaddr, KM_USER0);
 120                         set_buffer_uptodate(bh_result);
 121                 }
 122                 brelse(buffer_cache_bh);
 123         }
 124
 125         map_bh(bh_result, inode->i_sb,
 126                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 127
 128         err = 0;
 129
 130 bail:
 131         if (bh)
 132                 brelse(bh);
 133
 134         mlog_exit(err);
 135         return err;
 136 }
 137
 138 static int ocfs2_get_block(struct inode *inode, sector_t iblock,
 139                            struct buffer_head *bh_result, int create)
 140 {
 141         int err = 0;
 142         unsigned int ext_flags;
 143         u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
 144         u64 p_blkno, count, past_eof;
 145         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 146
 147         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
 148                    (unsigned long long)iblock, bh_result, create);
 149
 150         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
 151                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
 152                      inode, inode->i_ino);
 153
 154         if (S_ISLNK(inode->i_mode)) {
 155                 /* this always does I/O for some reason. */
 156                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
 157                 goto bail;
 158         }
 159
 160         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
 161                                           &ext_flags);
 162         if (err) {
 163                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
 164                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
 165                      (unsigned long long)p_blkno);
 166                 goto bail;
 167         }
 168
 169         if (max_blocks < count)
 170                 count = max_blocks;
 171
 172         /*
 173          * ocfs2 never allocates in this function - the only time we
 174          * need to use BH_New is when we're extending i_size on a file
 175          * system which doesn't support holes, in which case BH_New
 176          * allows block_prepare_write() to zero.
 177          *
 178          * If we see this on a sparse file system, then a truncate has
 179          * raced us and removed the cluster. In this case, we clear
 180          * the buffers dirty and uptodate bits and let the buffer code
 181          * ignore it as a hole.
 182          */
 183         if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
 184                 clear_buffer_dirty(bh_result);
 185                 clear_buffer_uptodate(bh_result);
 186                 goto bail;
 187         }
 188
 189         /* Treat the unwritten extent as a hole for zeroing purposes. */
 190         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 191                 map_bh(bh_result, inode->i_sb, p_blkno);
 192
 193         bh_result->b_size = count << inode->i_blkbits;
 194
 195         if (!ocfs2_sparse_alloc(osb)) {
 196                 if (p_blkno == 0) {
 197                         err = -EIO;
 198                         mlog(ML_ERROR,
 199                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
 200                              (unsigned long long)iblock,
 201                              (unsigned long long)p_blkno,
 202                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
 203                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
 204                         dump_stack();
 205                 }
 206
 207                 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 208                 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
 209                      (unsigned long long)past_eof);
 210
 211                 if (create && (iblock >= past_eof))
 212                         set_buffer_new(bh_result);
 213         }
 214
 215 bail:
 216         if (err < 0)
 217                 err = -EIO;
 218
 219         mlog_exit(err);
 220         return err;
 221 }
 222
 223 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
 224                            struct buffer_head *di_bh)
 225 {
 226         void *kaddr;
 227         loff_t size;
 228         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 229
 230         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
 231                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
 232                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
 233                 return -EROFS;
 234         }
 235
 236         size = i_size_read(inode);
 237
 238         if (size > PAGE_CACHE_SIZE ||
 239             size > ocfs2_max_inline_data(inode->i_sb)) {
 240                 ocfs2_error(inode->i_sb,
 241                             "Inode %llu has with inline data has bad size: %Lu",
 242                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 243                             (unsigned long long)size);
 244                 return -EROFS;
 245         }
 246
 247         kaddr = kmap_atomic(page, KM_USER0);
 248         if (size)
 249                 memcpy(kaddr, di->id2.i_data.id_data, size);
 250         /* Clear the remaining part of the page */
 251         memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
 252         flush_dcache_page(page);
 253         kunmap_atomic(kaddr, KM_USER0);
 254
 255         SetPageUptodate(page);
 256
 257         return 0;
 258 }
 259
 260 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
 261 {
 262         int ret;
 263         struct buffer_head *di_bh = NULL;
 264         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 265
 266         BUG_ON(!PageLocked(page));
 267         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
 268
 269         ret = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &di_bh,
 270                                OCFS2_BH_CACHED, inode);
 271         if (ret) {
 272                 mlog_errno(ret);
 273                 goto out;
 274         }
 275
 276         ret = ocfs2_read_inline_data(inode, page, di_bh);
 277 out:
 278         unlock_page(page);
 279
 280         brelse(di_bh);
 281         return ret;
 282 }
 283
 284 static int ocfs2_readpage(struct file *file, struct page *page)
 285 {
 286         struct inode *inode = page->mapping->host;
 287         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 288         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
 289         int ret, unlock = 1;
 290
 291         mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0));
 292
 293         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
 294         if (ret != 0) {
 295                 if (ret == AOP_TRUNCATED_PAGE)
 296                         unlock = 0;
 297                 mlog_errno(ret);
 298                 goto out;
 299         }
 300
 301         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 302                 ret = AOP_TRUNCATED_PAGE;
 303                 goto out_inode_unlock;
 304         }
 305
 306         /*
 307          * i_size might have just been updated as we grabed the meta lock.  We
 308          * might now be discovering a truncate that hit on another node.
 309          * block_read_full_page->get_block freaks out if it is asked to read
 310          * beyond the end of a file, so we check here.  Callers
 311          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
 312          * and notice that the page they just read isn't needed.
 313          *
 314          * XXX sys_readahead() seems to get that wrong?
 315          */
 316         if (start >= i_size_read(inode)) {
 317                 zero_user(page, 0, PAGE_SIZE);
 318                 SetPageUptodate(page);
 319                 ret = 0;
 320                 goto out_alloc;
 321         }
 322
 323         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 324                 ret = ocfs2_readpage_inline(inode, page);
 325         else
 326                 ret = block_read_full_page(page, ocfs2_get_block);
 327         unlock = 0;
 328
 329 out_alloc:
 330         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 331 out_inode_unlock:
 332         ocfs2_inode_unlock(inode, 0);
 333 out:
 334         if (unlock)
 335                 unlock_page(page);
 336         mlog_exit(ret);
 337         return ret;
 338 }
 339
 340 /*
 341  * This is used only for read-ahead. Failures or difficult to handle
 342  * situations are safe to ignore.
 343  *
 344  * Right now, we don't bother with BH_Boundary - in-inode extent lists
 345  * are quite large (243 extents on 4k blocks), so most inodes don't
 346  * grow out to a tree. If need be, detecting boundary extents could
 347  * trivially be added in a future version of ocfs2_get_block().
 348  */
 349 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
 350                            struct list_head *pages, unsigned nr_pages)
 351 {
 352         int ret, err = -EIO;
 353         struct inode *inode = mapping->host;
 354         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 355         loff_t start;
 356         struct page *last;
 357
 358         /*
 359          * Use the nonblocking flag for the dlm code to avoid page
 360          * lock inversion, but don't bother with retrying.
 361          */
 362         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
 363         if (ret)
 364                 return err;
 365
 366         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 367                 ocfs2_inode_unlock(inode, 0);
 368                 return err;
 369         }
 370
 371         /*
 372          * Don't bother with inline-data. There isn't anything
 373          * to read-ahead in that case anyway...
 374          */
 375         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 376                 goto out_unlock;
 377
 378         /*
 379          * Check whether a remote node truncated this file - we just
 380          * drop out in that case as it's not worth handling here.
 381          */
 382         last = list_entry(pages->prev, struct page, lru);
 383         start = (loff_t)last->index << PAGE_CACHE_SHIFT;
 384         if (start >= i_size_read(inode))
 385                 goto out_unlock;
 386
 387         err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
 388
 389 out_unlock:
 390         up_read(&oi->ip_alloc_sem);
 391         ocfs2_inode_unlock(inode, 0);
 392
 393         return err;
 394 }
 395
 396 /* Note: Because we don't support holes, our allocation has
 397  * already happened (allocation writes zeros to the file data)
 398  * so we don't have to worry about ordered writes in
 399  * ocfs2_writepage.
 400  *
 401  * ->writepage is called during the process of invalidating the page cache
 402  * during blocked lock processing.  It can't block on any cluster locks
 403  * to during block mapping.  It's relying on the fact that the block
 404  * mapping can't have disappeared under the dirty pages that it is
 405  * being asked to write back.
 406  */
 407 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
 408 {
 409         int ret;
 410
 411         mlog_entry("(0x%p)\n", page);
 412
 413         ret = block_write_full_page(page, ocfs2_get_block, wbc);
 414
 415         mlog_exit(ret);
 416
 417         return ret;
 418 }
 419
 420 /*
 421  * This is called from ocfs2_write_zero_page() which has handled it's
 422  * own cluster locking and has ensured allocation exists for those
 423  * blocks to be written.
 424  */
 425 int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
 426                                unsigned from, unsigned to)
 427 {
 428         int ret;
 429
 430         ret = block_prepare_write(page, from, to, ocfs2_get_block);
 431
 432         return ret;
 433 }
 434
 435 /* Taken from ext3. We don't necessarily need the full blown
 436  * functionality yet, but IMHO it's better to cut and paste the whole
 437  * thing so we can avoid introducing our own bugs (and easily pick up
 438  * their fixes when they happen) --Mark */
 439 int walk_page_buffers(  handle_t *handle,
 440                         struct buffer_head *head,
 441                         unsigned from,
 442                         unsigned to,
 443                         int *partial,
 444                         int (*fn)(      handle_t *handle,
 445                                         struct buffer_head *bh))
 446 {
 447         struct buffer_head *bh;
 448         unsigned block_start, block_end;
 449         unsigned blocksize = head->b_size;
 450         int err, ret = 0;
 451         struct buffer_head *next;
 452
 453         for (   bh = head, block_start = 0;
 454                 ret == 0 && (bh != head || !block_start);
 455                 block_start = block_end, bh = next)
 456         {
 457                 next = bh->b_this_page;
 458                 block_end = block_start + blocksize;
 459                 if (block_end <= from || block_start >= to) {
 460                         if (partial && !buffer_uptodate(bh))
 461                                 *partial = 1;
 462                         continue;
 463                 }
 464                 err = (*fn)(handle, bh);
 465                 if (!ret)
 466                         ret = err;
 467         }
 468         return ret;
 469 }
 470
 471 handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
 472                                                          struct page *page,
 473                                                          unsigned from,
 474                                                          unsigned to)
 475 {
 476         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 477         handle_t *handle;
 478         int ret = 0;
 479
 480         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 481         if (IS_ERR(handle)) {
 482                 ret = -ENOMEM;
 483                 mlog_errno(ret);
 484                 goto out;
 485         }
 486
 487         if (ocfs2_should_order_data(inode)) {
 488                 ret = walk_page_buffers(handle,
 489                                         page_buffers(page),
 490                                         from, to, NULL,
 491                                         ocfs2_journal_dirty_data);
 492                 if (ret < 0)
 493                         mlog_errno(ret);
 494         }
 495 out:
 496         if (ret) {
 497                 if (!IS_ERR(handle))
 498                         ocfs2_commit_trans(osb, handle);
 499                 handle = ERR_PTR(ret);
 500         }
 501         return handle;
 502 }
 503
 504 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 505 {
 506         sector_t status;
 507         u64 p_blkno = 0;
 508         int err = 0;
 509         struct inode *inode = mapping->host;
 510
 511         mlog_entry("(block = %llu)\n", (unsigned long long)block);
 512
 513         /* We don't need to lock journal system files, since they aren't
 514          * accessed concurrently from multiple nodes.
 515          */
 516         if (!INODE_JOURNAL(inode)) {
 517                 err = ocfs2_inode_lock(inode, NULL, 0);
 518                 if (err) {
 519                         if (err != -ENOENT)
 520                                 mlog_errno(err);
 521                         goto bail;
 522                 }
 523                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
 524         }
 525
 526         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
 527                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
 528                                                   NULL);
 529
 530         if (!INODE_JOURNAL(inode)) {
 531                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
 532                 ocfs2_inode_unlock(inode, 0);
 533         }
 534
 535         if (err) {
 536                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 537                      (unsigned long long)block);
 538                 mlog_errno(err);
 539                 goto bail;
 540         }
 541
 542 bail:
 543         status = err ? 0 : p_blkno;
 544
 545         mlog_exit((int)status);
 546
 547         return status;
 548 }
 549
 550 /*
 551  * TODO: Make this into a generic get_blocks function.
 552  *
 553  * From do_direct_io in direct-io.c:
 554  *  "So what we do is to permit the ->get_blocks function to populate
 555  *   bh.b_size with the size of IO which is permitted at this offset and
 556  *   this i_blkbits."
 557  *
 558  * This function is called directly from get_more_blocks in direct-io.c.
 559  *
 560  * called like this: dio->get_blocks(dio->inode, fs_startblk,
 561  *                                      fs_count, map_bh, dio->rw == WRITE);
 562  */
 563 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
 564                                      struct buffer_head *bh_result, int create)
 565 {
 566         int ret;
 567         u64 p_blkno, inode_blocks, contig_blocks;
 568         unsigned int ext_flags;
 569         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
 570         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
 571
 572         /* This function won't even be called if the request isn't all
 573          * nicely aligned and of the right size, so there's no need
 574          * for us to check any of that. */
 575
 576         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 577
 578         /*
 579          * Any write past EOF is not allowed because we'd be extending.
 580          */
 581         if (create && (iblock + max_blocks) > inode_blocks) {
 582                 ret = -EIO;
 583                 goto bail;
 584         }
 585
 586         /* This figures out the size of the next contiguous block, and
 587          * our logical offset */
 588         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
 589                                           &contig_blocks, &ext_flags);
 590         if (ret) {
 591                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
 592                      (unsigned long long)iblock);
 593                 ret = -EIO;
 594                 goto bail;
 595         }
 596
 597         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)) && !p_blkno && create) {
 598                 ocfs2_error(inode->i_sb,
 599                             "Inode %llu has a hole at block %llu\n",
 600                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 601                             (unsigned long long)iblock);
 602                 ret = -EROFS;
 603                 goto bail;
 604         }
 605
 606         /*
 607          * get_more_blocks() expects us to describe a hole by clearing
 608          * the mapped bit on bh_result().
 609          *
 610          * Consider an unwritten extent as a hole.
 611          */
 612         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 613                 map_bh(bh_result, inode->i_sb, p_blkno);
 614         else {
 615                 /*
 616                  * ocfs2_prepare_inode_for_write() should have caught
 617                  * the case where we'd be filling a hole and triggered
 618                  * a buffered write instead.
 619                  */
 620                 if (create) {
 621                         ret = -EIO;
 622                         mlog_errno(ret);
 623                         goto bail;
 624                 }
 625
 626                 clear_buffer_mapped(bh_result);
 627         }
 628
 629         /* make sure we don't map more than max_blocks blocks here as
 630            that's all the kernel will handle at this point. */
 631         if (max_blocks < contig_blocks)
 632                 contig_blocks = max_blocks;
 633         bh_result->b_size = contig_blocks << blocksize_bits;
 634 bail:
 635         return ret;
 636 }
 637
 638 /*
 639  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
 640  * particularly interested in the aio/dio case.  Like the core uses
 641  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
 642  * truncation on another.
 643  */
 644 static void ocfs2_dio_end_io(struct kiocb *iocb,
 645                              loff_t offset,
 646                              ssize_t bytes,
 647                              void *private)
 648 {
 649         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
 650         int level;
 651
 652         /* this io's submitter should not have unlocked this before we could */
 653         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 654
 655         ocfs2_iocb_clear_rw_locked(iocb);
 656
 657         level = ocfs2_iocb_rw_locked_level(iocb);
 658         if (!level)
 659                 up_read(&inode->i_alloc_sem);
 660         ocfs2_rw_unlock(inode, level);
 661 }
 662
 663 /*
 664  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
 665  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
 666  * do journalled data.
 667  */
 668 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
 669 {
 670         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 671
 672         journal_invalidatepage(journal, page, offset);
 673 }
 674
 675 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 676 {
 677         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 678
 679         if (!page_has_buffers(page))
 680                 return 0;
 681         return journal_try_to_free_buffers(journal, page, wait);
 682 }
 683
 684 static ssize_t ocfs2_direct_IO(int rw,
 685                                struct kiocb *iocb,
 686                                const struct iovec *iov,
 687                                loff_t offset,
 688                                unsigned long nr_segs)
 689 {
 690         struct file *file = iocb->ki_filp;
 691         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
 692         int ret;
 693
 694         mlog_entry_void();
 695
 696         /*
 697          * Fallback to buffered I/O if we see an inode without
 698          * extents.
 699          */
 700         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 701                 return 0;
 702
 703         ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
 704                                             inode->i_sb->s_bdev, iov, offset,
 705                                             nr_segs,
 706                                             ocfs2_direct_IO_get_blocks,
 707                                             ocfs2_dio_end_io);
 708
 709         mlog_exit(ret);
 710         return ret;
 711 }
 712
 713 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 714                                             u32 cpos,
 715                                             unsigned int *start,
 716                                             unsigned int *end)
 717 {
 718         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
 719
 720         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
 721                 unsigned int cpp;
 722
 723                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
 724
 725                 cluster_start = cpos % cpp;
 726                 cluster_start = cluster_start << osb->s_clustersize_bits;
 727
 728                 cluster_end = cluster_start + osb->s_clustersize;
 729         }
 730
 731         BUG_ON(cluster_start > PAGE_SIZE);
 732         BUG_ON(cluster_end > PAGE_SIZE);
 733
 734         if (start)
 735                 *start = cluster_start;
 736         if (end)
 737                 *end = cluster_end;
 738 }
 739
 740 /*
 741  * 'from' and 'to' are the region in the page to avoid zeroing.
 742  *
 743  * If pagesize > clustersize, this function will avoid zeroing outside
 744  * of the cluster boundary.
 745  *
 746  * from == to == 0 is code for "zero the entire cluster region"
 747  */
 748 static void ocfs2_clear_page_regions(struct page *page,
 749                                      struct ocfs2_super *osb, u32 cpos,
 750                                      unsigned from, unsigned to)
 751 {
 752         void *kaddr;
 753         unsigned int cluster_start, cluster_end;
 754
 755         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 756
 757         kaddr = kmap_atomic(page, KM_USER0);
 758
 759         if (from || to) {
 760                 if (from > cluster_start)
 761                         memset(kaddr + cluster_start, 0, from - cluster_start);
 762                 if (to < cluster_end)
 763                         memset(kaddr + to, 0, cluster_end - to);
 764         } else {
 765                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 766         }
 767
 768         kunmap_atomic(kaddr, KM_USER0);
 769 }
 770
 771 /*
 772  * Nonsparse file systems fully allocate before we get to the write
 773  * code. This prevents ocfs2_write() from tagging the write as an
 774  * allocating one, which means ocfs2_map_page_blocks() might try to
 775  * read-in the blocks at the tail of our file. Avoid reading them by
 776  * testing i_size against each block offset.
 777  */
 778 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 779                                  unsigned int block_start)
 780 {
 781         u64 offset = page_offset(page) + block_start;
 782
 783         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 784                 return 1;
 785
 786         if (i_size_read(inode) > offset)
 787                 return 1;
 788
 789         return 0;
 790 }
 791
 792 /*
 793  * Some of this taken from block_prepare_write(). We already have our
 794  * mapping by now though, and the entire write will be allocating or
 795  * it won't, so not much need to use BH_New.
 796  *
 797  * This will also skip zeroing, which is handled externally.
 798  */
 799 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 800                           struct inode *inode, unsigned int from,
 801                           unsigned int to, int new)
 802 {
 803         int ret = 0;
 804         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 805         unsigned int block_end, block_start;
 806         unsigned int bsize = 1 << inode->i_blkbits;
 807
 808         if (!page_has_buffers(page))
 809                 create_empty_buffers(page, bsize, 0);
 810
 811         head = page_buffers(page);
 812         for (bh = head, block_start = 0; bh != head || !block_start;
 813              bh = bh->b_this_page, block_start += bsize) {
 814                 block_end = block_start + bsize;
 815
 816                 clear_buffer_new(bh);
 817
 818                 /*
 819                  * Ignore blocks outside of our i/o range -
 820                  * they may belong to unallocated clusters.
 821                  */
 822                 if (block_start >= to || block_end <= from) {
 823                         if (PageUptodate(page))
 824                                 set_buffer_uptodate(bh);
 825                         continue;
 826                 }
 827
 828                 /*
 829                  * For an allocating write with cluster size >= page
 830                  * size, we always write the entire page.
 831                  */
 832                 if (new)
 833                         set_buffer_new(bh);
 834
 835                 if (!buffer_mapped(bh)) {
 836                         map_bh(bh, inode->i_sb, *p_blkno);
 837                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 838                 }
 839
 840                 if (PageUptodate(page)) {
 841                         if (!buffer_uptodate(bh))
 842                                 set_buffer_uptodate(bh);
 843                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 844                            !buffer_new(bh) &&
 845                            ocfs2_should_read_blk(inode, page, block_start) &&
 846                            (block_start < from || block_end > to)) {
 847                         ll_rw_block(READ, 1, &bh);
 848                         *wait_bh++=bh;
 849                 }
 850
 851                 *p_blkno = *p_blkno + 1;
 852         }
 853
 854         /*
 855          * If we issued read requests - let them complete.
 856          */
 857         while(wait_bh > wait) {
 858                 wait_on_buffer(*--wait_bh);
 859                 if (!buffer_uptodate(*wait_bh))
 860                         ret = -EIO;
 861         }
 862
 863         if (ret == 0 || !new)
 864                 return ret;
 865
 866         /*
 867          * If we get -EIO above, zero out any newly allocated blocks
 868          * to avoid exposing stale data.
 869          */
 870         bh = head;
 871         block_start = 0;
 872         do {
 873                 block_end = block_start + bsize;
 874                 if (block_end <= from)
 875                         goto next_bh;
 876                 if (block_start >= to)
 877                         break;
 878
 879                 zero_user(page, block_start, bh->b_size);
 880                 set_buffer_uptodate(bh);
 881                 mark_buffer_dirty(bh);
 882
 883 next_bh:
 884                 block_start = block_end;
 885                 bh = bh->b_this_page;
 886         } while (bh != head);
 887
 888         return ret;
 889 }
 890
 891 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 892 #define OCFS2_MAX_CTXT_PAGES    1
 893 #else
 894 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
 895 #endif
 896
 897 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 898
 899 /*
 900  * Describe the state of a single cluster to be written to.
 901  */
 902 struct ocfs2_write_cluster_desc {
 903         u32             c_cpos;
 904         u32             c_phys;
 905         /*
 906          * Give this a unique field because c_phys eventually gets
 907          * filled.
 908          */
 909         unsigned        c_new;
 910         unsigned        c_unwritten;
 911 };
 912
 913 static inline int ocfs2_should_zero_cluster(struct ocfs2_write_cluster_desc *d)
 914 {
 915         return d->c_new || d->c_unwritten;
 916 }
 917
 918 struct ocfs2_write_ctxt {
 919         /* Logical cluster position / len of write */
 920         u32                             w_cpos;
 921         u32                             w_clen;
 922
 923         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 924
 925         /*
 926          * This is true if page_size > cluster_size.
 927          *
 928          * It triggers a set of special cases during write which might
 929          * have to deal with allocating writes to partial pages.
 930          */
 931         unsigned int                    w_large_pages;
 932
 933         /*
 934          * Pages involved in this write.
 935          *
 936          * w_target_page is the page being written to by the user.
 937          *
 938          * w_pages is an array of pages which always contains
 939          * w_target_page, and in the case of an allocating write with
 940          * page_size < cluster size, it will contain zero'd and mapped
 941          * pages adjacent to w_target_page which need to be written
 942          * out in so that future reads from that region will get
 943          * zero's.
 944          */
 945         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 946         unsigned int                    w_num_pages;
 947         struct page                     *w_target_page;
 948
 949         /*
 950          * ocfs2_write_end() uses this to know what the real range to
 951          * write in the target should be.
 952          */
 953         unsigned int                    w_target_from;
 954         unsigned int                    w_target_to;
 955
 956         /*
 957          * We could use journal_current_handle() but this is cleaner,
 958          * IMHO -Mark
 959          */
 960         handle_t                        *w_handle;
 961
 962         struct buffer_head              *w_di_bh;
 963
 964         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 965 };
 966
 967 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 968 {
 969         int i;
 970
 971         for(i = 0; i < num_pages; i++) {
 972                 if (pages[i]) {
 973                         unlock_page(pages[i]);
 974                         mark_page_accessed(pages[i]);
 975                         page_cache_release(pages[i]);
 976                 }
 977         }
 978 }
 979
 980 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 981 {
 982         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 983
 984         brelse(wc->w_di_bh);
 985         kfree(wc);
 986 }
 987
 988 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 989                                   struct ocfs2_super *osb, loff_t pos,
 990                                   unsigned len, struct buffer_head *di_bh)
 991 {
 992         u32 cend;
 993         struct ocfs2_write_ctxt *wc;
 994
 995         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 996         if (!wc)
 997                 return -ENOMEM;
 998
 999         wc->w_cpos = pos >> osb->s_clustersize_bits;
1000         cend = (pos + len - 1) >> osb->s_clustersize_bits;
1001         wc->w_clen = cend - wc->w_cpos + 1;
1002         get_bh(di_bh);
1003         wc->w_di_bh = di_bh;
1004
1005         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
1006                 wc->w_large_pages = 1;
1007         else
1008                 wc->w_large_pages = 0;
1009
1010         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
1011
1012         *wcp = wc;
1013
1014         return 0;
1015 }
1016
1017 /*
1018  * If a page has any new buffers, zero them out here, and mark them uptodate
1019  * and dirty so they'll be written out (in order to prevent uninitialised
1020  * block data from leaking). And clear the new bit.
1021  */
1022 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1023 {
1024         unsigned int block_start, block_end;
1025         struct buffer_head *head, *bh;
1026
1027         BUG_ON(!PageLocked(page));
1028         if (!page_has_buffers(page))
1029                 return;
1030
1031         bh = head = page_buffers(page);
1032         block_start = 0;
1033         do {
1034                 block_end = block_start + bh->b_size;
1035
1036                 if (buffer_new(bh)) {
1037                         if (block_end > from && block_start < to) {
1038                                 if (!PageUptodate(page)) {
1039                                         unsigned start, end;
1040
1041                                         start = max(from, block_start);
1042                                         end = min(to, block_end);
1043
1044                                         zero_user_segment(page, start, end);
1045                                         set_buffer_uptodate(bh);
1046                                 }
1047
1048                                 clear_buffer_new(bh);
1049                                 mark_buffer_dirty(bh);
1050                         }
1051                 }
1052
1053                 block_start = block_end;
1054                 bh = bh->b_this_page;
1055         } while (bh != head);
1056 }
1057
1058 /*
1059  * Only called when we have a failure during allocating write to write
1060  * zero's to the newly allocated region.
1061  */
1062 static void ocfs2_write_failure(struct inode *inode,
1063                                 struct ocfs2_write_ctxt *wc,
1064                                 loff_t user_pos, unsigned user_len)
1065 {
1066         int i;
1067         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1068                 to = user_pos + user_len;
1069         struct page *tmppage;
1070
1071         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1072
1073         for(i = 0; i < wc->w_num_pages; i++) {
1074                 tmppage = wc->w_pages[i];
1075
1076                 if (page_has_buffers(tmppage)) {
1077                         if (ocfs2_should_order_data(inode))
1078                                 walk_page_buffers(wc->w_handle,
1079                                                   page_buffers(tmppage),
1080                                                   from, to, NULL,
1081                                                   ocfs2_journal_dirty_data);
1082
1083                         block_commit_write(tmppage, from, to);
1084                 }
1085         }
1086 }
1087
1088 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1089                                         struct ocfs2_write_ctxt *wc,
1090                                         struct page *page, u32 cpos,
1091                                         loff_t user_pos, unsigned user_len,
1092                                         int new)
1093 {
1094         int ret;
1095         unsigned int map_from = 0, map_to = 0;
1096         unsigned int cluster_start, cluster_end;
1097         unsigned int user_data_from = 0, user_data_to = 0;
1098
1099         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1100                                         &cluster_start, &cluster_end);
1101
1102         if (page == wc->w_target_page) {
1103                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1104                 map_to = map_from + user_len;
1105
1106                 if (new)
1107                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1108                                                     cluster_start, cluster_end,
1109                                                     new);
1110                 else
1111                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1112                                                     map_from, map_to, new);
1113                 if (ret) {
1114                         mlog_errno(ret);
1115                         goto out;
1116                 }
1117
1118                 user_data_from = map_from;
1119                 user_data_to = map_to;
1120                 if (new) {
1121                         map_from = cluster_start;
1122                         map_to = cluster_end;
1123                 }
1124         } else {
1125                 /*
1126                  * If we haven't allocated the new page yet, we
1127                  * shouldn't be writing it out without copying user
1128                  * data. This is likely a math error from the caller.
1129                  */
1130                 BUG_ON(!new);
1131
1132                 map_from = cluster_start;
1133                 map_to = cluster_end;
1134
1135                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1136                                             cluster_start, cluster_end, new);
1137                 if (ret) {
1138                         mlog_errno(ret);
1139                         goto out;
1140                 }
1141         }
1142
1143         /*
1144          * Parts of newly allocated pages need to be zero'd.
1145          *
1146          * Above, we have also rewritten 'to' and 'from' - as far as
1147          * the rest of the function is concerned, the entire cluster
1148          * range inside of a page needs to be written.
1149          *
1150          * We can skip this if the page is up to date - it's already
1151          * been zero'd from being read in as a hole.
1152          */
1153         if (new && !PageUptodate(page))
1154                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1155                                          cpos, user_data_from, user_data_to);
1156
1157         flush_dcache_page(page);
1158
1159 out:
1160         return ret;
1161 }
1162
1163 /*
1164  * This function will only grab one clusters worth of pages.
1165  */
1166 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1167                                       struct ocfs2_write_ctxt *wc,
1168                                       u32 cpos, loff_t user_pos, int new,
1169                                       struct page *mmap_page)
1170 {
1171         int ret = 0, i;
1172         unsigned long start, target_index, index;
1173         struct inode *inode = mapping->host;
1174
1175         target_index = user_pos >> PAGE_CACHE_SHIFT;
1176
1177         /*
1178          * Figure out how many pages we'll be manipulating here. For
1179          * non allocating write, we just change the one
1180          * page. Otherwise, we'll need a whole clusters worth.
1181          */
1182         if (new) {
1183                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1184                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1185         } else {
1186                 wc->w_num_pages = 1;
1187                 start = target_index;
1188         }
1189
1190         for(i = 0; i < wc->w_num_pages; i++) {
1191                 index = start + i;
1192
1193                 if (index == target_index && mmap_page) {
1194                         /*
1195                          * ocfs2_pagemkwrite() is a little different
1196                          * and wants us to directly use the page
1197                          * passed in.
1198                          */
1199                         lock_page(mmap_page);
1200
1201                         if (mmap_page->mapping != mapping) {
1202                                 unlock_page(mmap_page);
1203                                 /*
1204                                  * Sanity check - the locking in
1205                                  * ocfs2_pagemkwrite() should ensure
1206                                  * that this code doesn't trigger.
1207                                  */
1208                                 ret = -EINVAL;
1209                                 mlog_errno(ret);
1210                                 goto out;
1211                         }
1212
1213                         page_cache_get(mmap_page);
1214                         wc->w_pages[i] = mmap_page;
1215                 } else {
1216                         wc->w_pages[i] = find_or_create_page(mapping, index,
1217                                                              GFP_NOFS);
1218                         if (!wc->w_pages[i]) {
1219                                 ret = -ENOMEM;
1220                                 mlog_errno(ret);
1221                                 goto out;
1222                         }
1223                 }
1224
1225                 if (index == target_index)
1226                         wc->w_target_page = wc->w_pages[i];
1227         }
1228 out:
1229         return ret;
1230 }
1231
1232 /*
1233  * Prepare a single cluster for write one cluster into the file.
1234  */
1235 static int ocfs2_write_cluster(struct address_space *mapping,
1236                                u32 phys, unsigned int unwritten,
1237                                struct ocfs2_alloc_context *data_ac,
1238                                struct ocfs2_alloc_context *meta_ac,
1239                                struct ocfs2_write_ctxt *wc, u32 cpos,
1240                                loff_t user_pos, unsigned user_len)
1241 {
1242         int ret, i, new, should_zero = 0;
1243         u64 v_blkno, p_blkno;
1244         struct inode *inode = mapping->host;
1245
1246         new = phys == 0 ? 1 : 0;
1247         if (new || unwritten)
1248                 should_zero = 1;
1249
1250         if (new) {
1251                 u32 tmp_pos;
1252
1253                 /*
1254                  * This is safe to call with the page locks - it won't take
1255                  * any additional semaphores or cluster locks.
1256                  */
1257                 tmp_pos = cpos;
1258                 ret = ocfs2_do_extend_allocation(OCFS2_SB(inode->i_sb), inode,
1259                                                  &tmp_pos, 1, 0, wc->w_di_bh,
1260                                                  wc->w_handle, data_ac,
1261                                                  meta_ac, NULL);
1262                 /*
1263                  * This shouldn't happen because we must have already
1264                  * calculated the correct meta data allocation required. The
1265                  * internal tree allocation code should know how to increase
1266                  * transaction credits itself.
1267                  *
1268                  * If need be, we could handle -EAGAIN for a
1269                  * RESTART_TRANS here.
1270                  */
1271                 mlog_bug_on_msg(ret == -EAGAIN,
1272                                 "Inode %llu: EAGAIN return during allocation.\n",
1273                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1274                 if (ret < 0) {
1275                         mlog_errno(ret);
1276                         goto out;
1277                 }
1278         } else if (unwritten) {
1279                 ret = ocfs2_mark_extent_written(inode, wc->w_di_bh,
1280                                                 wc->w_handle, cpos, 1, phys,
1281                                                 meta_ac, &wc->w_dealloc);
1282                 if (ret < 0) {
1283                         mlog_errno(ret);
1284                         goto out;
1285                 }
1286         }
1287
1288         if (should_zero)
1289                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1290         else
1291                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1292
1293         /*
1294          * The only reason this should fail is due to an inability to
1295          * find the extent added.
1296          */
1297         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1298                                           NULL);
1299         if (ret < 0) {
1300                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1301                             "at logical block %llu",
1302                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1303                             (unsigned long long)v_blkno);
1304                 goto out;
1305         }
1306
1307         BUG_ON(p_blkno == 0);
1308
1309         for(i = 0; i < wc->w_num_pages; i++) {
1310                 int tmpret;
1311
1312                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1313                                                       wc->w_pages[i], cpos,
1314                                                       user_pos, user_len,
1315                                                       should_zero);
1316                 if (tmpret) {
1317                         mlog_errno(tmpret);
1318                         if (ret == 0)
1319                                 tmpret = ret;
1320                 }
1321         }
1322
1323         /*
1324          * We only have cleanup to do in case of allocating write.
1325          */
1326         if (ret && new)
1327                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1328
1329 out:
1330
1331         return ret;
1332 }
1333
1334 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1335                                        struct ocfs2_alloc_context *data_ac,
1336                                        struct ocfs2_alloc_context *meta_ac,
1337                                        struct ocfs2_write_ctxt *wc,
1338                                        loff_t pos, unsigned len)
1339 {
1340         int ret, i;
1341         loff_t cluster_off;
1342         unsigned int local_len = len;
1343         struct ocfs2_write_cluster_desc *desc;
1344         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1345
1346         for (i = 0; i < wc->w_clen; i++) {
1347                 desc = &wc->w_desc[i];
1348
1349                 /*
1350                  * We have to make sure that the total write passed in
1351                  * doesn't extend past a single cluster.
1352                  */
1353                 local_len = len;
1354                 cluster_off = pos & (osb->s_clustersize - 1);
1355                 if ((cluster_off + local_len) > osb->s_clustersize)
1356                         local_len = osb->s_clustersize - cluster_off;
1357
1358                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1359                                           desc->c_unwritten, data_ac, meta_ac,
1360                                           wc, desc->c_cpos, pos, local_len);
1361                 if (ret) {
1362                         mlog_errno(ret);
1363                         goto out;
1364                 }
1365
1366                 len -= local_len;
1367                 pos += local_len;
1368         }
1369
1370         ret = 0;
1371 out:
1372         return ret;
1373 }
1374
1375 /*
1376  * ocfs2_write_end() wants to know which parts of the target page it
1377  * should complete the write on. It's easiest to compute them ahead of
1378  * time when a more complete view of the write is available.
1379  */
1380 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1381                                         struct ocfs2_write_ctxt *wc,
1382                                         loff_t pos, unsigned len, int alloc)
1383 {
1384         struct ocfs2_write_cluster_desc *desc;
1385
1386         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1387         wc->w_target_to = wc->w_target_from + len;
1388
1389         if (alloc == 0)
1390                 return;
1391
1392         /*
1393          * Allocating write - we may have different boundaries based
1394          * on page size and cluster size.
1395          *
1396          * NOTE: We can no longer compute one value from the other as
1397          * the actual write length and user provided length may be
1398          * different.
1399          */
1400
1401         if (wc->w_large_pages) {
1402                 /*
1403                  * We only care about the 1st and last cluster within
1404                  * our range and whether they should be zero'd or not. Either
1405                  * value may be extended out to the start/end of a
1406                  * newly allocated cluster.
1407                  */
1408                 desc = &wc->w_desc[0];
1409                 if (ocfs2_should_zero_cluster(desc))
1410                         ocfs2_figure_cluster_boundaries(osb,
1411                                                         desc->c_cpos,
1412                                                         &wc->w_target_from,
1413                                                         NULL);
1414
1415                 desc = &wc->w_desc[wc->w_clen - 1];
1416                 if (ocfs2_should_zero_cluster(desc))
1417                         ocfs2_figure_cluster_boundaries(osb,
1418                                                         desc->c_cpos,
1419                                                         NULL,
1420                                                         &wc->w_target_to);
1421         } else {
1422                 wc->w_target_from = 0;
1423                 wc->w_target_to = PAGE_CACHE_SIZE;
1424         }
1425 }
1426
1427 /*
1428  * Populate each single-cluster write descriptor in the write context
1429  * with information about the i/o to be done.
1430  *
1431  * Returns the number of clusters that will have to be allocated, as
1432  * well as a worst case estimate of the number of extent records that
1433  * would have to be created during a write to an unwritten region.
1434  */
1435 static int ocfs2_populate_write_desc(struct inode *inode,
1436                                      struct ocfs2_write_ctxt *wc,
1437                                      unsigned int *clusters_to_alloc,
1438                                      unsigned int *extents_to_split)
1439 {
1440         int ret;
1441         struct ocfs2_write_cluster_desc *desc;
1442         unsigned int num_clusters = 0;
1443         unsigned int ext_flags = 0;
1444         u32 phys = 0;
1445         int i;
1446
1447         *clusters_to_alloc = 0;
1448         *extents_to_split = 0;
1449
1450         for (i = 0; i < wc->w_clen; i++) {
1451                 desc = &wc->w_desc[i];
1452                 desc->c_cpos = wc->w_cpos + i;
1453
1454                 if (num_clusters == 0) {
1455                         /*
1456                          * Need to look up the next extent record.
1457                          */
1458                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1459                                                  &num_clusters, &ext_flags);
1460                         if (ret) {
1461                                 mlog_errno(ret);
1462                                 goto out;
1463                         }
1464
1465                         /*
1466                          * Assume worst case - that we're writing in
1467                          * the middle of the extent.
1468                          *
1469                          * We can assume that the write proceeds from
1470                          * left to right, in which case the extent
1471                          * insert code is smart enough to coalesce the
1472                          * next splits into the previous records created.
1473                          */
1474                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1475                                 *extents_to_split = *extents_to_split + 2;
1476                 } else if (phys) {
1477                         /*
1478                          * Only increment phys if it doesn't describe
1479                          * a hole.
1480                          */
1481                         phys++;
1482                 }
1483
1484                 desc->c_phys = phys;
1485                 if (phys == 0) {
1486                         desc->c_new = 1;
1487                         *clusters_to_alloc = *clusters_to_alloc + 1;
1488                 }
1489                 if (ext_flags & OCFS2_EXT_UNWRITTEN)
1490                         desc->c_unwritten = 1;
1491
1492                 num_clusters--;
1493         }
1494
1495         ret = 0;
1496 out:
1497         return ret;
1498 }
1499
1500 static int ocfs2_write_begin_inline(struct address_space *mapping,
1501                                     struct inode *inode,
1502                                     struct ocfs2_write_ctxt *wc)
1503 {
1504         int ret;
1505         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1506         struct page *page;
1507         handle_t *handle;
1508         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1509
1510         page = find_or_create_page(mapping, 0, GFP_NOFS);
1511         if (!page) {
1512                 ret = -ENOMEM;
1513                 mlog_errno(ret);
1514                 goto out;
1515         }
1516         /*
1517          * If we don't set w_num_pages then this page won't get unlocked
1518          * and freed on cleanup of the write context.
1519          */
1520         wc->w_pages[0] = wc->w_target_page = page;
1521         wc->w_num_pages = 1;
1522
1523         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1524         if (IS_ERR(handle)) {
1525                 ret = PTR_ERR(handle);
1526                 mlog_errno(ret);
1527                 goto out;
1528         }
1529
1530         ret = ocfs2_journal_access(handle, inode, wc->w_di_bh,
1531                                    OCFS2_JOURNAL_ACCESS_WRITE);
1532         if (ret) {
1533                 ocfs2_commit_trans(osb, handle);
1534
1535                 mlog_errno(ret);
1536                 goto out;
1537         }
1538
1539         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1540                 ocfs2_set_inode_data_inline(inode, di);
1541
1542         if (!PageUptodate(page)) {
1543                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1544                 if (ret) {
1545                         ocfs2_commit_trans(osb, handle);
1546
1547                         goto out;
1548                 }
1549         }
1550
1551         wc->w_handle = handle;
1552 out:
1553         return ret;
1554 }
1555
1556 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1557 {
1558         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1559
1560         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1561                 return 1;
1562         return 0;
1563 }
1564
1565 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1566                                           struct inode *inode, loff_t pos,
1567                                           unsigned len, struct page *mmap_page,
1568                                           struct ocfs2_write_ctxt *wc)
1569 {
1570         int ret, written = 0;
1571         loff_t end = pos + len;
1572         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1573
1574         mlog(0, "Inode %llu, write of %u bytes at off %llu. features: 0x%x\n",
1575              (unsigned long long)oi->ip_blkno, len, (unsigned long long)pos,
1576              oi->ip_dyn_features);
1577
1578         /*
1579          * Handle inodes which already have inline data 1st.
1580          */
1581         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1582                 if (mmap_page == NULL &&
1583                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1584                         goto do_inline_write;
1585
1586                 /*
1587                  * The write won't fit - we have to give this inode an
1588                  * inline extent list now.
1589                  */
1590                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1591                 if (ret)
1592                         mlog_errno(ret);
1593                 goto out;
1594         }
1595
1596         /*
1597          * Check whether the inode can accept inline data.
1598          */
1599         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1600                 return 0;
1601
1602         /*
1603          * Check whether the write can fit.
1604          */
1605         if (mmap_page || end > ocfs2_max_inline_data(inode->i_sb))
1606                 return 0;
1607
1608 do_inline_write:
1609         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1610         if (ret) {
1611                 mlog_errno(ret);
1612                 goto out;
1613         }
1614
1615         /*
1616          * This signals to the caller that the data can be written
1617          * inline.
1618          */
1619         written = 1;
1620 out:
1621         return written ? written : ret;
1622 }
1623
1624 /*
1625  * This function only does anything for file systems which can't
1626  * handle sparse files.
1627  *
1628  * What we want to do here is fill in any hole between the current end
1629  * of allocation and the end of our write. That way the rest of the
1630  * write path can treat it as an non-allocating write, which has no
1631  * special case code for sparse/nonsparse files.
1632  */
1633 static int ocfs2_expand_nonsparse_inode(struct inode *inode, loff_t pos,
1634                                         unsigned len,
1635                                         struct ocfs2_write_ctxt *wc)
1636 {
1637         int ret;
1638         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1639         loff_t newsize = pos + len;
1640
1641         if (ocfs2_sparse_alloc(osb))
1642                 return 0;
1643
1644         if (newsize <= i_size_read(inode))
1645                 return 0;
1646
1647         ret = ocfs2_extend_no_holes(inode, newsize, newsize - len);
1648         if (ret)
1649                 mlog_errno(ret);
1650
1651         return ret;
1652 }
1653
1654 int ocfs2_write_begin_nolock(struct address_space *mapping,
1655                              loff_t pos, unsigned len, unsigned flags,
1656                              struct page **pagep, void **fsdata,
1657                              struct buffer_head *di_bh, struct page *mmap_page)
1658 {
1659         int ret, credits = OCFS2_INODE_UPDATE_CREDITS;
1660         unsigned int clusters_to_alloc, extents_to_split;
1661         struct ocfs2_write_ctxt *wc;
1662         struct inode *inode = mapping->host;
1663         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1664         struct ocfs2_dinode *di;
1665         struct ocfs2_alloc_context *data_ac = NULL;
1666         struct ocfs2_alloc_context *meta_ac = NULL;
1667         handle_t *handle;
1668
1669         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1670         if (ret) {
1671                 mlog_errno(ret);
1672                 return ret;
1673         }
1674
1675         if (ocfs2_supports_inline_data(osb)) {
1676                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1677                                                      mmap_page, wc);
1678                 if (ret == 1) {
1679                         ret = 0;
1680                         goto success;
1681                 }
1682                 if (ret < 0) {
1683                         mlog_errno(ret);
1684                         goto out;
1685                 }
1686         }
1687
1688         ret = ocfs2_expand_nonsparse_inode(inode, pos, len, wc);
1689         if (ret) {
1690                 mlog_errno(ret);
1691                 goto out;
1692         }
1693
1694         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1695                                         &extents_to_split);
1696         if (ret) {
1697                 mlog_errno(ret);
1698                 goto out;
1699         }
1700
1701         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1702
1703         /*
1704          * We set w_target_from, w_target_to here so that
1705          * ocfs2_write_end() knows which range in the target page to
1706          * write out. An allocation requires that we write the entire
1707          * cluster range.
1708          */
1709         if (clusters_to_alloc || extents_to_split) {
1710                 /*
1711                  * XXX: We are stretching the limits of
1712                  * ocfs2_lock_allocators(). It greatly over-estimates
1713                  * the work to be done.
1714                  */
1715                 ret = ocfs2_lock_allocators(inode, di, clusters_to_alloc,
1716                                             extents_to_split, &data_ac, &meta_ac);
1717                 if (ret) {
1718                         mlog_errno(ret);
1719                         goto out;
1720                 }
1721
1722                 credits = ocfs2_calc_extend_credits(inode->i_sb, di,
1723                                                     clusters_to_alloc);
1724
1725         }
1726
1727         ocfs2_set_target_boundaries(osb, wc, pos, len,
1728                                     clusters_to_alloc + extents_to_split);
1729
1730         handle = ocfs2_start_trans(osb, credits);
1731         if (IS_ERR(handle)) {
1732                 ret = PTR_ERR(handle);
1733                 mlog_errno(ret);
1734                 goto out;
1735         }
1736
1737         wc->w_handle = handle;
1738
1739         /*
1740          * We don't want this to fail in ocfs2_write_end(), so do it
1741          * here.
1742          */
1743         ret = ocfs2_journal_access(handle, inode, wc->w_di_bh,
1744                                    OCFS2_JOURNAL_ACCESS_WRITE);
1745         if (ret) {
1746                 mlog_errno(ret);
1747                 goto out_commit;
1748         }
1749
1750         /*
1751          * Fill our page array first. That way we've grabbed enough so
1752          * that we can zero and flush if we error after adding the
1753          * extent.
1754          */
1755         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos,
1756                                          clusters_to_alloc + extents_to_split,
1757                                          mmap_page);
1758         if (ret) {
1759                 mlog_errno(ret);
1760                 goto out_commit;
1761         }
1762
1763         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1764                                           len);
1765         if (ret) {
1766                 mlog_errno(ret);
1767                 goto out_commit;
1768         }
1769
1770         if (data_ac)
1771                 ocfs2_free_alloc_context(data_ac);
1772         if (meta_ac)
1773                 ocfs2_free_alloc_context(meta_ac);
1774
1775 success:
1776         *pagep = wc->w_target_page;
1777         *fsdata = wc;
1778         return 0;
1779 out_commit:
1780         ocfs2_commit_trans(osb, handle);
1781
1782 out:
1783         ocfs2_free_write_ctxt(wc);
1784
1785         if (data_ac)
1786                 ocfs2_free_alloc_context(data_ac);
1787         if (meta_ac)
1788                 ocfs2_free_alloc_context(meta_ac);
1789         return ret;
1790 }
1791
1792 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1793                              loff_t pos, unsigned len, unsigned flags,
1794                              struct page **pagep, void **fsdata)
1795 {
1796         int ret;
1797         struct buffer_head *di_bh = NULL;
1798         struct inode *inode = mapping->host;
1799
1800         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1801         if (ret) {
1802                 mlog_errno(ret);
1803                 return ret;
1804         }
1805
1806         /*
1807          * Take alloc sem here to prevent concurrent lookups. That way
1808          * the mapping, zeroing and tree manipulation within
1809          * ocfs2_write() will be safe against ->readpage(). This
1810          * should also serve to lock out allocation from a shared
1811          * writeable region.
1812          */
1813         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1814
1815         ret = ocfs2_write_begin_nolock(mapping, pos, len, flags, pagep,
1816                                        fsdata, di_bh, NULL);
1817         if (ret) {
1818                 mlog_errno(ret);
1819                 goto out_fail;
1820         }
1821
1822         brelse(di_bh);
1823
1824         return 0;
1825
1826 out_fail:
1827         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1828
1829         brelse(di_bh);
1830         ocfs2_inode_unlock(inode, 1);
1831
1832         return ret;
1833 }
1834
1835 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1836                                    unsigned len, unsigned *copied,
1837                                    struct ocfs2_dinode *di,
1838                                    struct ocfs2_write_ctxt *wc)
1839 {
1840         void *kaddr;
1841
1842         if (unlikely(*copied < len)) {
1843                 if (!PageUptodate(wc->w_target_page)) {
1844                         *copied = 0;
1845                         return;
1846                 }
1847         }
1848
1849         kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1850         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1851         kunmap_atomic(kaddr, KM_USER0);
1852
1853         mlog(0, "Data written to inode at offset %llu. "
1854              "id_count = %u, copied = %u, i_dyn_features = 0x%x\n",
1855              (unsigned long long)pos, *copied,
1856              le16_to_cpu(di->id2.i_data.id_count),
1857              le16_to_cpu(di->i_dyn_features));
1858 }
1859
1860 int ocfs2_write_end_nolock(struct address_space *mapping,
1861                            loff_t pos, unsigned len, unsigned copied,
1862                            struct page *page, void *fsdata)
1863 {
1864         int i;
1865         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1866         struct inode *inode = mapping->host;
1867         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1868         struct ocfs2_write_ctxt *wc = fsdata;
1869         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1870         handle_t *handle = wc->w_handle;
1871         struct page *tmppage;
1872
1873         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1874                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1875                 goto out_write_size;
1876         }
1877
1878         if (unlikely(copied < len)) {
1879                 if (!PageUptodate(wc->w_target_page))
1880                         copied = 0;
1881
1882                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1883                                        start+len);
1884         }
1885         flush_dcache_page(wc->w_target_page);
1886
1887         for(i = 0; i < wc->w_num_pages; i++) {
1888                 tmppage = wc->w_pages[i];
1889
1890                 if (tmppage == wc->w_target_page) {
1891                         from = wc->w_target_from;
1892                         to = wc->w_target_to;
1893
1894                         BUG_ON(from > PAGE_CACHE_SIZE ||
1895                                to > PAGE_CACHE_SIZE ||
1896                                to < from);
1897                 } else {
1898                         /*
1899                          * Pages adjacent to the target (if any) imply
1900                          * a hole-filling write in which case we want
1901                          * to flush their entire range.
1902                          */
1903                         from = 0;
1904                         to = PAGE_CACHE_SIZE;
1905                 }
1906
1907                 if (page_has_buffers(tmppage)) {
1908                         if (ocfs2_should_order_data(inode))
1909                                 walk_page_buffers(wc->w_handle,
1910                                                   page_buffers(tmppage),
1911                                                   from, to, NULL,
1912                                                   ocfs2_journal_dirty_data);
1913                         block_commit_write(tmppage, from, to);
1914                 }
1915         }
1916
1917 out_write_size:
1918         pos += copied;
1919         if (pos > inode->i_size) {
1920                 i_size_write(inode, pos);
1921                 mark_inode_dirty(inode);
1922         }
1923         inode->i_blocks = ocfs2_inode_sector_count(inode);
1924         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1925         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1926         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1927         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1928         ocfs2_journal_dirty(handle, wc->w_di_bh);
1929
1930         ocfs2_commit_trans(osb, handle);
1931
1932         ocfs2_run_deallocs(osb, &wc->w_dealloc);
1933
1934         ocfs2_free_write_ctxt(wc);
1935
1936         return copied;
1937 }
1938
1939 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
1940                            loff_t pos, unsigned len, unsigned copied,
1941                            struct page *page, void *fsdata)
1942 {
1943         int ret;
1944         struct inode *inode = mapping->host;
1945
1946         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
1947
1948         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1949         ocfs2_inode_unlock(inode, 1);
1950
1951         return ret;
1952 }
1953
1954 const struct address_space_operations ocfs2_aops = {
1955         .readpage       = ocfs2_readpage,
1956         .readpages      = ocfs2_readpages,
1957         .writepage      = ocfs2_writepage,
1958         .write_begin    = ocfs2_write_begin,
1959         .write_end      = ocfs2_write_end,
1960         .bmap           = ocfs2_bmap,
1961         .sync_page      = block_sync_page,
1962         .direct_IO      = ocfs2_direct_IO,
1963         .invalidatepage = ocfs2_invalidatepage,
1964         .releasepage    = ocfs2_releasepage,
1965         .migratepage    = buffer_migrate_page,
1966 };