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
  30 #define MLOG_MASK_PREFIX ML_FILE_IO
  31 #include <cluster/masklog.h>
  32
  33 #include "ocfs2.h"
  34
  35 #include "alloc.h"
  36 #include "aops.h"
  37 #include "dlmglue.h"
  38 #include "extent_map.h"
  39 #include "file.h"
  40 #include "inode.h"
  41 #include "journal.h"
  42 #include "suballoc.h"
  43 #include "super.h"
  44 #include "symlink.h"
  45
  46 #include "buffer_head_io.h"
  47
  48 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
  49                                    struct buffer_head *bh_result, int create)
  50 {
  51         int err = -EIO;
  52         int status;
  53         struct ocfs2_dinode *fe = NULL;
  54         struct buffer_head *bh = NULL;
  55         struct buffer_head *buffer_cache_bh = NULL;
  56         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  57         void *kaddr;
  58
  59         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
  60                    (unsigned long long)iblock, bh_result, create);
  61
  62         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
  63
  64         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
  65                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
  66                      (unsigned long long)iblock);
  67                 goto bail;
  68         }
  69
  70         status = ocfs2_read_block(OCFS2_SB(inode->i_sb),
  71                                   OCFS2_I(inode)->ip_blkno,
  72                                   &bh, OCFS2_BH_CACHED, inode);
  73         if (status < 0) {
  74                 mlog_errno(status);
  75                 goto bail;
  76         }
  77         fe = (struct ocfs2_dinode *) bh->b_data;
  78
  79         if (!OCFS2_IS_VALID_DINODE(fe)) {
  80                 mlog(ML_ERROR, "Invalid dinode #%llu: signature = %.*s\n",
  81                      (unsigned long long)le64_to_cpu(fe->i_blkno), 7,
  82                      fe->i_signature);
  83                 goto bail;
  84         }
  85
  86         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
  87                                                     le32_to_cpu(fe->i_clusters))) {
  88                 mlog(ML_ERROR, "block offset is outside the allocated size: "
  89                      "%llu\n", (unsigned long long)iblock);
  90                 goto bail;
  91         }
  92
  93         /* We don't use the page cache to create symlink data, so if
  94          * need be, copy it over from the buffer cache. */
  95         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  96                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  97                             iblock;
  98                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
  99                 if (!buffer_cache_bh) {
 100                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
 101                         goto bail;
 102                 }
 103
 104                 /* we haven't locked out transactions, so a commit
 105                  * could've happened. Since we've got a reference on
 106                  * the bh, even if it commits while we're doing the
 107                  * copy, the data is still good. */
 108                 if (buffer_jbd(buffer_cache_bh)
 109                     && ocfs2_inode_is_new(inode)) {
 110                         kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
 111                         if (!kaddr) {
 112                                 mlog(ML_ERROR, "couldn't kmap!\n");
 113                                 goto bail;
 114                         }
 115                         memcpy(kaddr + (bh_result->b_size * iblock),
 116                                buffer_cache_bh->b_data,
 117                                bh_result->b_size);
 118                         kunmap_atomic(kaddr, KM_USER0);
 119                         set_buffer_uptodate(bh_result);
 120                 }
 121                 brelse(buffer_cache_bh);
 122         }
 123
 124         map_bh(bh_result, inode->i_sb,
 125                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 126
 127         err = 0;
 128
 129 bail:
 130         if (bh)
 131                 brelse(bh);
 132
 133         mlog_exit(err);
 134         return err;
 135 }
 136
 137 static int ocfs2_get_block(struct inode *inode, sector_t iblock,
 138                            struct buffer_head *bh_result, int create)
 139 {
 140         int err = 0;
 141         unsigned int ext_flags;
 142         u64 p_blkno, past_eof;
 143         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 144
 145         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
 146                    (unsigned long long)iblock, bh_result, create);
 147
 148         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
 149                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
 150                      inode, inode->i_ino);
 151
 152         if (S_ISLNK(inode->i_mode)) {
 153                 /* this always does I/O for some reason. */
 154                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
 155                 goto bail;
 156         }
 157
 158         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, NULL,
 159                                           &ext_flags);
 160         if (err) {
 161                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
 162                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
 163                      (unsigned long long)p_blkno);
 164                 goto bail;
 165         }
 166
 167         /*
 168          * ocfs2 never allocates in this function - the only time we
 169          * need to use BH_New is when we're extending i_size on a file
 170          * system which doesn't support holes, in which case BH_New
 171          * allows block_prepare_write() to zero.
 172          */
 173         mlog_bug_on_msg(create && p_blkno == 0 && ocfs2_sparse_alloc(osb),
 174                         "ino %lu, iblock %llu\n", inode->i_ino,
 175                         (unsigned long long)iblock);
 176
 177         /* Treat the unwritten extent as a hole for zeroing purposes. */
 178         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 179                 map_bh(bh_result, inode->i_sb, p_blkno);
 180
 181         if (!ocfs2_sparse_alloc(osb)) {
 182                 if (p_blkno == 0) {
 183                         err = -EIO;
 184                         mlog(ML_ERROR,
 185                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
 186                              (unsigned long long)iblock,
 187                              (unsigned long long)p_blkno,
 188                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
 189                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
 190                         dump_stack();
 191                 }
 192
 193                 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 194                 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
 195                      (unsigned long long)past_eof);
 196
 197                 if (create && (iblock >= past_eof))
 198                         set_buffer_new(bh_result);
 199         }
 200
 201 bail:
 202         if (err < 0)
 203                 err = -EIO;
 204
 205         mlog_exit(err);
 206         return err;
 207 }
 208
 209 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
 210                            struct buffer_head *di_bh)
 211 {
 212         void *kaddr;
 213         unsigned int size;
 214         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 215
 216         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
 217                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
 218                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
 219                 return -EROFS;
 220         }
 221
 222         size = i_size_read(inode);
 223
 224         if (size > PAGE_CACHE_SIZE ||
 225             size > ocfs2_max_inline_data(inode->i_sb)) {
 226                 ocfs2_error(inode->i_sb,
 227                             "Inode %llu has with inline data has bad size: %u",
 228                             (unsigned long long)OCFS2_I(inode)->ip_blkno, size);
 229                 return -EROFS;
 230         }
 231
 232         kaddr = kmap_atomic(page, KM_USER0);
 233         if (size)
 234                 memcpy(kaddr, di->id2.i_data.id_data, size);
 235         /* Clear the remaining part of the page */
 236         memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
 237         flush_dcache_page(page);
 238         kunmap_atomic(kaddr, KM_USER0);
 239
 240         SetPageUptodate(page);
 241
 242         return 0;
 243 }
 244
 245 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
 246 {
 247         int ret;
 248         struct buffer_head *di_bh = NULL;
 249         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 250
 251         BUG_ON(!PageLocked(page));
 252         BUG_ON(!OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL);
 253
 254         ret = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &di_bh,
 255                                OCFS2_BH_CACHED, inode);
 256         if (ret) {
 257                 mlog_errno(ret);
 258                 goto out;
 259         }
 260
 261         ret = ocfs2_read_inline_data(inode, page, di_bh);
 262 out:
 263         unlock_page(page);
 264
 265         brelse(di_bh);
 266         return ret;
 267 }
 268
 269 static int ocfs2_readpage(struct file *file, struct page *page)
 270 {
 271         struct inode *inode = page->mapping->host;
 272         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 273         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
 274         int ret, unlock = 1;
 275
 276         mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0));
 277
 278         ret = ocfs2_meta_lock_with_page(inode, NULL, 0, page);
 279         if (ret != 0) {
 280                 if (ret == AOP_TRUNCATED_PAGE)
 281                         unlock = 0;
 282                 mlog_errno(ret);
 283                 goto out;
 284         }
 285
 286         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 287                 ret = AOP_TRUNCATED_PAGE;
 288                 goto out_meta_unlock;
 289         }
 290
 291         /*
 292          * i_size might have just been updated as we grabed the meta lock.  We
 293          * might now be discovering a truncate that hit on another node.
 294          * block_read_full_page->get_block freaks out if it is asked to read
 295          * beyond the end of a file, so we check here.  Callers
 296          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
 297          * and notice that the page they just read isn't needed.
 298          *
 299          * XXX sys_readahead() seems to get that wrong?
 300          */
 301         if (start >= i_size_read(inode)) {
 302                 zero_user_page(page, 0, PAGE_SIZE, KM_USER0);
 303                 SetPageUptodate(page);
 304                 ret = 0;
 305                 goto out_alloc;
 306         }
 307
 308         ret = ocfs2_data_lock_with_page(inode, 0, page);
 309         if (ret != 0) {
 310                 if (ret == AOP_TRUNCATED_PAGE)
 311                         unlock = 0;
 312                 mlog_errno(ret);
 313                 goto out_alloc;
 314         }
 315
 316         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 317                 ret = ocfs2_readpage_inline(inode, page);
 318         else
 319                 ret = block_read_full_page(page, ocfs2_get_block);
 320         unlock = 0;
 321
 322         ocfs2_data_unlock(inode, 0);
 323 out_alloc:
 324         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 325 out_meta_unlock:
 326         ocfs2_meta_unlock(inode, 0);
 327 out:
 328         if (unlock)
 329                 unlock_page(page);
 330         mlog_exit(ret);
 331         return ret;
 332 }
 333
 334 /* Note: Because we don't support holes, our allocation has
 335  * already happened (allocation writes zeros to the file data)
 336  * so we don't have to worry about ordered writes in
 337  * ocfs2_writepage.
 338  *
 339  * ->writepage is called during the process of invalidating the page cache
 340  * during blocked lock processing.  It can't block on any cluster locks
 341  * to during block mapping.  It's relying on the fact that the block
 342  * mapping can't have disappeared under the dirty pages that it is
 343  * being asked to write back.
 344  */
 345 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
 346 {
 347         int ret;
 348
 349         mlog_entry("(0x%p)\n", page);
 350
 351         ret = block_write_full_page(page, ocfs2_get_block, wbc);
 352
 353         mlog_exit(ret);
 354
 355         return ret;
 356 }
 357
 358 /*
 359  * This is called from ocfs2_write_zero_page() which has handled it's
 360  * own cluster locking and has ensured allocation exists for those
 361  * blocks to be written.
 362  */
 363 int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
 364                                unsigned from, unsigned to)
 365 {
 366         int ret;
 367
 368         ret = block_prepare_write(page, from, to, ocfs2_get_block);
 369
 370         return ret;
 371 }
 372
 373 /* Taken from ext3. We don't necessarily need the full blown
 374  * functionality yet, but IMHO it's better to cut and paste the whole
 375  * thing so we can avoid introducing our own bugs (and easily pick up
 376  * their fixes when they happen) --Mark */
 377 int walk_page_buffers(  handle_t *handle,
 378                         struct buffer_head *head,
 379                         unsigned from,
 380                         unsigned to,
 381                         int *partial,
 382                         int (*fn)(      handle_t *handle,
 383                                         struct buffer_head *bh))
 384 {
 385         struct buffer_head *bh;
 386         unsigned block_start, block_end;
 387         unsigned blocksize = head->b_size;
 388         int err, ret = 0;
 389         struct buffer_head *next;
 390
 391         for (   bh = head, block_start = 0;
 392                 ret == 0 && (bh != head || !block_start);
 393                 block_start = block_end, bh = next)
 394         {
 395                 next = bh->b_this_page;
 396                 block_end = block_start + blocksize;
 397                 if (block_end <= from || block_start >= to) {
 398                         if (partial && !buffer_uptodate(bh))
 399                                 *partial = 1;
 400                         continue;
 401                 }
 402                 err = (*fn)(handle, bh);
 403                 if (!ret)
 404                         ret = err;
 405         }
 406         return ret;
 407 }
 408
 409 handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
 410                                                          struct page *page,
 411                                                          unsigned from,
 412                                                          unsigned to)
 413 {
 414         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 415         handle_t *handle = NULL;
 416         int ret = 0;
 417
 418         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 419         if (!handle) {
 420                 ret = -ENOMEM;
 421                 mlog_errno(ret);
 422                 goto out;
 423         }
 424
 425         if (ocfs2_should_order_data(inode)) {
 426                 ret = walk_page_buffers(handle,
 427                                         page_buffers(page),
 428                                         from, to, NULL,
 429                                         ocfs2_journal_dirty_data);
 430                 if (ret < 0)
 431                         mlog_errno(ret);
 432         }
 433 out:
 434         if (ret) {
 435                 if (handle)
 436                         ocfs2_commit_trans(osb, handle);
 437                 handle = ERR_PTR(ret);
 438         }
 439         return handle;
 440 }
 441
 442 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 443 {
 444         sector_t status;
 445         u64 p_blkno = 0;
 446         int err = 0;
 447         struct inode *inode = mapping->host;
 448
 449         mlog_entry("(block = %llu)\n", (unsigned long long)block);
 450
 451         /* We don't need to lock journal system files, since they aren't
 452          * accessed concurrently from multiple nodes.
 453          */
 454         if (!INODE_JOURNAL(inode)) {
 455                 err = ocfs2_meta_lock(inode, NULL, 0);
 456                 if (err) {
 457                         if (err != -ENOENT)
 458                                 mlog_errno(err);
 459                         goto bail;
 460                 }
 461                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
 462         }
 463
 464         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
 465                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
 466                                                   NULL);
 467
 468         if (!INODE_JOURNAL(inode)) {
 469                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
 470                 ocfs2_meta_unlock(inode, 0);
 471         }
 472
 473         if (err) {
 474                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 475                      (unsigned long long)block);
 476                 mlog_errno(err);
 477                 goto bail;
 478         }
 479
 480 bail:
 481         status = err ? 0 : p_blkno;
 482
 483         mlog_exit((int)status);
 484
 485         return status;
 486 }
 487
 488 /*
 489  * TODO: Make this into a generic get_blocks function.
 490  *
 491  * From do_direct_io in direct-io.c:
 492  *  "So what we do is to permit the ->get_blocks function to populate
 493  *   bh.b_size with the size of IO which is permitted at this offset and
 494  *   this i_blkbits."
 495  *
 496  * This function is called directly from get_more_blocks in direct-io.c.
 497  *
 498  * called like this: dio->get_blocks(dio->inode, fs_startblk,
 499  *                                      fs_count, map_bh, dio->rw == WRITE);
 500  */
 501 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
 502                                      struct buffer_head *bh_result, int create)
 503 {
 504         int ret;
 505         u64 p_blkno, inode_blocks, contig_blocks;
 506         unsigned int ext_flags;
 507         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
 508         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
 509
 510         /* This function won't even be called if the request isn't all
 511          * nicely aligned and of the right size, so there's no need
 512          * for us to check any of that. */
 513
 514         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 515
 516         /*
 517          * Any write past EOF is not allowed because we'd be extending.
 518          */
 519         if (create && (iblock + max_blocks) > inode_blocks) {
 520                 ret = -EIO;
 521                 goto bail;
 522         }
 523
 524         /* This figures out the size of the next contiguous block, and
 525          * our logical offset */
 526         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
 527                                           &contig_blocks, &ext_flags);
 528         if (ret) {
 529                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
 530                      (unsigned long long)iblock);
 531                 ret = -EIO;
 532                 goto bail;
 533         }
 534
 535         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)) && !p_blkno) {
 536                 ocfs2_error(inode->i_sb,
 537                             "Inode %llu has a hole at block %llu\n",
 538                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 539                             (unsigned long long)iblock);
 540                 ret = -EROFS;
 541                 goto bail;
 542         }
 543
 544         /*
 545          * get_more_blocks() expects us to describe a hole by clearing
 546          * the mapped bit on bh_result().
 547          *
 548          * Consider an unwritten extent as a hole.
 549          */
 550         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 551                 map_bh(bh_result, inode->i_sb, p_blkno);
 552         else {
 553                 /*
 554                  * ocfs2_prepare_inode_for_write() should have caught
 555                  * the case where we'd be filling a hole and triggered
 556                  * a buffered write instead.
 557                  */
 558                 if (create) {
 559                         ret = -EIO;
 560                         mlog_errno(ret);
 561                         goto bail;
 562                 }
 563
 564                 clear_buffer_mapped(bh_result);
 565         }
 566
 567         /* make sure we don't map more than max_blocks blocks here as
 568            that's all the kernel will handle at this point. */
 569         if (max_blocks < contig_blocks)
 570                 contig_blocks = max_blocks;
 571         bh_result->b_size = contig_blocks << blocksize_bits;
 572 bail:
 573         return ret;
 574 }
 575
 576 /*
 577  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
 578  * particularly interested in the aio/dio case.  Like the core uses
 579  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
 580  * truncation on another.
 581  */
 582 static void ocfs2_dio_end_io(struct kiocb *iocb,
 583                              loff_t offset,
 584                              ssize_t bytes,
 585                              void *private)
 586 {
 587         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
 588         int level;
 589
 590         /* this io's submitter should not have unlocked this before we could */
 591         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 592
 593         ocfs2_iocb_clear_rw_locked(iocb);
 594
 595         level = ocfs2_iocb_rw_locked_level(iocb);
 596         if (!level)
 597                 up_read(&inode->i_alloc_sem);
 598         ocfs2_rw_unlock(inode, level);
 599 }
 600
 601 /*
 602  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
 603  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
 604  * do journalled data.
 605  */
 606 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
 607 {
 608         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 609
 610         journal_invalidatepage(journal, page, offset);
 611 }
 612
 613 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 614 {
 615         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 616
 617         if (!page_has_buffers(page))
 618                 return 0;
 619         return journal_try_to_free_buffers(journal, page, wait);
 620 }
 621
 622 static ssize_t ocfs2_direct_IO(int rw,
 623                                struct kiocb *iocb,
 624                                const struct iovec *iov,
 625                                loff_t offset,
 626                                unsigned long nr_segs)
 627 {
 628         struct file *file = iocb->ki_filp;
 629         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
 630         int ret;
 631
 632         mlog_entry_void();
 633
 634         /*
 635          * Fallback to buffered I/O if we see an inode without
 636          * extents.
 637          */
 638         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 639                 return 0;
 640
 641         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
 642                 /*
 643                  * We get PR data locks even for O_DIRECT.  This
 644                  * allows concurrent O_DIRECT I/O but doesn't let
 645                  * O_DIRECT with extending and buffered zeroing writes
 646                  * race.  If they did race then the buffered zeroing
 647                  * could be written back after the O_DIRECT I/O.  It's
 648                  * one thing to tell people not to mix buffered and
 649                  * O_DIRECT writes, but expecting them to understand
 650                  * that file extension is also an implicit buffered
 651                  * write is too much.  By getting the PR we force
 652                  * writeback of the buffered zeroing before
 653                  * proceeding.
 654                  */
 655                 ret = ocfs2_data_lock(inode, 0);
 656                 if (ret < 0) {
 657                         mlog_errno(ret);
 658                         goto out;
 659                 }
 660                 ocfs2_data_unlock(inode, 0);
 661         }
 662
 663         ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
 664                                             inode->i_sb->s_bdev, iov, offset,
 665                                             nr_segs,
 666                                             ocfs2_direct_IO_get_blocks,
 667                                             ocfs2_dio_end_io);
 668 out:
 669         mlog_exit(ret);
 670         return ret;
 671 }
 672
 673 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 674                                             u32 cpos,
 675                                             unsigned int *start,
 676                                             unsigned int *end)
 677 {
 678         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
 679
 680         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
 681                 unsigned int cpp;
 682
 683                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
 684
 685                 cluster_start = cpos % cpp;
 686                 cluster_start = cluster_start << osb->s_clustersize_bits;
 687
 688                 cluster_end = cluster_start + osb->s_clustersize;
 689         }
 690
 691         BUG_ON(cluster_start > PAGE_SIZE);
 692         BUG_ON(cluster_end > PAGE_SIZE);
 693
 694         if (start)
 695                 *start = cluster_start;
 696         if (end)
 697                 *end = cluster_end;
 698 }
 699
 700 /*
 701  * 'from' and 'to' are the region in the page to avoid zeroing.
 702  *
 703  * If pagesize > clustersize, this function will avoid zeroing outside
 704  * of the cluster boundary.
 705  *
 706  * from == to == 0 is code for "zero the entire cluster region"
 707  */
 708 static void ocfs2_clear_page_regions(struct page *page,
 709                                      struct ocfs2_super *osb, u32 cpos,
 710                                      unsigned from, unsigned to)
 711 {
 712         void *kaddr;
 713         unsigned int cluster_start, cluster_end;
 714
 715         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 716
 717         kaddr = kmap_atomic(page, KM_USER0);
 718
 719         if (from || to) {
 720                 if (from > cluster_start)
 721                         memset(kaddr + cluster_start, 0, from - cluster_start);
 722                 if (to < cluster_end)
 723                         memset(kaddr + to, 0, cluster_end - to);
 724         } else {
 725                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 726         }
 727
 728         kunmap_atomic(kaddr, KM_USER0);
 729 }
 730
 731 /*
 732  * Nonsparse file systems fully allocate before we get to the write
 733  * code. This prevents ocfs2_write() from tagging the write as an
 734  * allocating one, which means ocfs2_map_page_blocks() might try to
 735  * read-in the blocks at the tail of our file. Avoid reading them by
 736  * testing i_size against each block offset.
 737  */
 738 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 739                                  unsigned int block_start)
 740 {
 741         u64 offset = page_offset(page) + block_start;
 742
 743         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 744                 return 1;
 745
 746         if (i_size_read(inode) > offset)
 747                 return 1;
 748
 749         return 0;
 750 }
 751
 752 /*
 753  * Some of this taken from block_prepare_write(). We already have our
 754  * mapping by now though, and the entire write will be allocating or
 755  * it won't, so not much need to use BH_New.
 756  *
 757  * This will also skip zeroing, which is handled externally.
 758  */
 759 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 760                           struct inode *inode, unsigned int from,
 761                           unsigned int to, int new)
 762 {
 763         int ret = 0;
 764         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 765         unsigned int block_end, block_start;
 766         unsigned int bsize = 1 << inode->i_blkbits;
 767
 768         if (!page_has_buffers(page))
 769                 create_empty_buffers(page, bsize, 0);
 770
 771         head = page_buffers(page);
 772         for (bh = head, block_start = 0; bh != head || !block_start;
 773              bh = bh->b_this_page, block_start += bsize) {
 774                 block_end = block_start + bsize;
 775
 776                 clear_buffer_new(bh);
 777
 778                 /*
 779                  * Ignore blocks outside of our i/o range -
 780                  * they may belong to unallocated clusters.
 781                  */
 782                 if (block_start >= to || block_end <= from) {
 783                         if (PageUptodate(page))
 784                                 set_buffer_uptodate(bh);
 785                         continue;
 786                 }
 787
 788                 /*
 789                  * For an allocating write with cluster size >= page
 790                  * size, we always write the entire page.
 791                  */
 792                 if (new)
 793                         set_buffer_new(bh);
 794
 795                 if (!buffer_mapped(bh)) {
 796                         map_bh(bh, inode->i_sb, *p_blkno);
 797                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 798                 }
 799
 800                 if (PageUptodate(page)) {
 801                         if (!buffer_uptodate(bh))
 802                                 set_buffer_uptodate(bh);
 803                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 804                            !buffer_new(bh) &&
 805                            ocfs2_should_read_blk(inode, page, block_start) &&
 806                            (block_start < from || block_end > to)) {
 807                         ll_rw_block(READ, 1, &bh);
 808                         *wait_bh++=bh;
 809                 }
 810
 811                 *p_blkno = *p_blkno + 1;
 812         }
 813
 814         /*
 815          * If we issued read requests - let them complete.
 816          */
 817         while(wait_bh > wait) {
 818                 wait_on_buffer(*--wait_bh);
 819                 if (!buffer_uptodate(*wait_bh))
 820                         ret = -EIO;
 821         }
 822
 823         if (ret == 0 || !new)
 824                 return ret;
 825
 826         /*
 827          * If we get -EIO above, zero out any newly allocated blocks
 828          * to avoid exposing stale data.
 829          */
 830         bh = head;
 831         block_start = 0;
 832         do {
 833                 block_end = block_start + bsize;
 834                 if (block_end <= from)
 835                         goto next_bh;
 836                 if (block_start >= to)
 837                         break;
 838
 839                 zero_user_page(page, block_start, bh->b_size, KM_USER0);
 840                 set_buffer_uptodate(bh);
 841                 mark_buffer_dirty(bh);
 842
 843 next_bh:
 844                 block_start = block_end;
 845                 bh = bh->b_this_page;
 846         } while (bh != head);
 847
 848         return ret;
 849 }
 850
 851 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 852 #define OCFS2_MAX_CTXT_PAGES    1
 853 #else
 854 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
 855 #endif
 856
 857 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 858
 859 /*
 860  * Describe the state of a single cluster to be written to.
 861  */
 862 struct ocfs2_write_cluster_desc {
 863         u32             c_cpos;
 864         u32             c_phys;
 865         /*
 866          * Give this a unique field because c_phys eventually gets
 867          * filled.
 868          */
 869         unsigned        c_new;
 870         unsigned        c_unwritten;
 871 };
 872
 873 static inline int ocfs2_should_zero_cluster(struct ocfs2_write_cluster_desc *d)
 874 {
 875         return d->c_new || d->c_unwritten;
 876 }
 877
 878 struct ocfs2_write_ctxt {
 879         /* Logical cluster position / len of write */
 880         u32                             w_cpos;
 881         u32                             w_clen;
 882
 883         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 884
 885         /*
 886          * This is true if page_size > cluster_size.
 887          *
 888          * It triggers a set of special cases during write which might
 889          * have to deal with allocating writes to partial pages.
 890          */
 891         unsigned int                    w_large_pages;
 892
 893         /*
 894          * Pages involved in this write.
 895          *
 896          * w_target_page is the page being written to by the user.
 897          *
 898          * w_pages is an array of pages which always contains
 899          * w_target_page, and in the case of an allocating write with
 900          * page_size < cluster size, it will contain zero'd and mapped
 901          * pages adjacent to w_target_page which need to be written
 902          * out in so that future reads from that region will get
 903          * zero's.
 904          */
 905         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 906         unsigned int                    w_num_pages;
 907         struct page                     *w_target_page;
 908
 909         /*
 910          * ocfs2_write_end() uses this to know what the real range to
 911          * write in the target should be.
 912          */
 913         unsigned int                    w_target_from;
 914         unsigned int                    w_target_to;
 915
 916         /*
 917          * We could use journal_current_handle() but this is cleaner,
 918          * IMHO -Mark
 919          */
 920         handle_t                        *w_handle;
 921
 922         struct buffer_head              *w_di_bh;
 923
 924         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 925 };
 926
 927 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 928 {
 929         int i;
 930
 931         for(i = 0; i < num_pages; i++) {
 932                 if (pages[i]) {
 933                         unlock_page(pages[i]);
 934                         mark_page_accessed(pages[i]);
 935                         page_cache_release(pages[i]);
 936                 }
 937         }
 938 }
 939
 940 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 941 {
 942         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 943
 944         brelse(wc->w_di_bh);
 945         kfree(wc);
 946 }
 947
 948 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 949                                   struct ocfs2_super *osb, loff_t pos,
 950                                   unsigned len, struct buffer_head *di_bh)
 951 {
 952         u32 cend;
 953         struct ocfs2_write_ctxt *wc;
 954
 955         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 956         if (!wc)
 957                 return -ENOMEM;
 958
 959         wc->w_cpos = pos >> osb->s_clustersize_bits;
 960         cend = (pos + len - 1) >> osb->s_clustersize_bits;
 961         wc->w_clen = cend - wc->w_cpos + 1;
 962         get_bh(di_bh);
 963         wc->w_di_bh = di_bh;
 964
 965         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
 966                 wc->w_large_pages = 1;
 967         else
 968                 wc->w_large_pages = 0;
 969
 970         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 971
 972         *wcp = wc;
 973
 974         return 0;
 975 }
 976
 977 /*
 978  * If a page has any new buffers, zero them out here, and mark them uptodate
 979  * and dirty so they'll be written out (in order to prevent uninitialised
 980  * block data from leaking). And clear the new bit.
 981  */
 982 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 983 {
 984         unsigned int block_start, block_end;
 985         struct buffer_head *head, *bh;
 986
 987         BUG_ON(!PageLocked(page));
 988         if (!page_has_buffers(page))
 989                 return;
 990
 991         bh = head = page_buffers(page);
 992         block_start = 0;
 993         do {
 994                 block_end = block_start + bh->b_size;
 995
 996                 if (buffer_new(bh)) {
 997                         if (block_end > from && block_start < to) {
 998                                 if (!PageUptodate(page)) {
 999                                         unsigned start, end;
1000
1001                                         start = max(from, block_start);
1002                                         end = min(to, block_end);
1003
1004                                         zero_user_page(page, start, end - start, KM_USER0);
1005                                         set_buffer_uptodate(bh);
1006                                 }
1007
1008                                 clear_buffer_new(bh);
1009                                 mark_buffer_dirty(bh);
1010                         }
1011                 }
1012
1013                 block_start = block_end;
1014                 bh = bh->b_this_page;
1015         } while (bh != head);
1016 }
1017
1018 /*
1019  * Only called when we have a failure during allocating write to write
1020  * zero's to the newly allocated region.
1021  */
1022 static void ocfs2_write_failure(struct inode *inode,
1023                                 struct ocfs2_write_ctxt *wc,
1024                                 loff_t user_pos, unsigned user_len)
1025 {
1026         int i;
1027         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1028                 to = user_pos + user_len;
1029         struct page *tmppage;
1030
1031         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1032
1033         for(i = 0; i < wc->w_num_pages; i++) {
1034                 tmppage = wc->w_pages[i];
1035
1036                 if (ocfs2_should_order_data(inode))
1037                         walk_page_buffers(wc->w_handle, page_buffers(tmppage),
1038                                           from, to, NULL,
1039                                           ocfs2_journal_dirty_data);
1040
1041                 block_commit_write(tmppage, from, to);
1042         }
1043 }
1044
1045 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1046                                         struct ocfs2_write_ctxt *wc,
1047                                         struct page *page, u32 cpos,
1048                                         loff_t user_pos, unsigned user_len,
1049                                         int new)
1050 {
1051         int ret;
1052         unsigned int map_from = 0, map_to = 0;
1053         unsigned int cluster_start, cluster_end;
1054         unsigned int user_data_from = 0, user_data_to = 0;
1055
1056         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1057                                         &cluster_start, &cluster_end);
1058
1059         if (page == wc->w_target_page) {
1060                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1061                 map_to = map_from + user_len;
1062
1063                 if (new)
1064                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1065                                                     cluster_start, cluster_end,
1066                                                     new);
1067                 else
1068                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1069                                                     map_from, map_to, new);
1070                 if (ret) {
1071                         mlog_errno(ret);
1072                         goto out;
1073                 }
1074
1075                 user_data_from = map_from;
1076                 user_data_to = map_to;
1077                 if (new) {
1078                         map_from = cluster_start;
1079                         map_to = cluster_end;
1080                 }
1081         } else {
1082                 /*
1083                  * If we haven't allocated the new page yet, we
1084                  * shouldn't be writing it out without copying user
1085                  * data. This is likely a math error from the caller.
1086                  */
1087                 BUG_ON(!new);
1088
1089                 map_from = cluster_start;
1090                 map_to = cluster_end;
1091
1092                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1093                                             cluster_start, cluster_end, new);
1094                 if (ret) {
1095                         mlog_errno(ret);
1096                         goto out;
1097                 }
1098         }
1099
1100         /*
1101          * Parts of newly allocated pages need to be zero'd.
1102          *
1103          * Above, we have also rewritten 'to' and 'from' - as far as
1104          * the rest of the function is concerned, the entire cluster
1105          * range inside of a page needs to be written.
1106          *
1107          * We can skip this if the page is up to date - it's already
1108          * been zero'd from being read in as a hole.
1109          */
1110         if (new && !PageUptodate(page))
1111                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1112                                          cpos, user_data_from, user_data_to);
1113
1114         flush_dcache_page(page);
1115
1116 out:
1117         return ret;
1118 }
1119
1120 /*
1121  * This function will only grab one clusters worth of pages.
1122  */
1123 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1124                                       struct ocfs2_write_ctxt *wc,
1125                                       u32 cpos, loff_t user_pos, int new,
1126                                       struct page *mmap_page)
1127 {
1128         int ret = 0, i;
1129         unsigned long start, target_index, index;
1130         struct inode *inode = mapping->host;
1131
1132         target_index = user_pos >> PAGE_CACHE_SHIFT;
1133
1134         /*
1135          * Figure out how many pages we'll be manipulating here. For
1136          * non allocating write, we just change the one
1137          * page. Otherwise, we'll need a whole clusters worth.
1138          */
1139         if (new) {
1140                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1141                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1142         } else {
1143                 wc->w_num_pages = 1;
1144                 start = target_index;
1145         }
1146
1147         for(i = 0; i < wc->w_num_pages; i++) {
1148                 index = start + i;
1149
1150                 if (index == target_index && mmap_page) {
1151                         /*
1152                          * ocfs2_pagemkwrite() is a little different
1153                          * and wants us to directly use the page
1154                          * passed in.
1155                          */
1156                         lock_page(mmap_page);
1157
1158                         if (mmap_page->mapping != mapping) {
1159                                 unlock_page(mmap_page);
1160                                 /*
1161                                  * Sanity check - the locking in
1162                                  * ocfs2_pagemkwrite() should ensure
1163                                  * that this code doesn't trigger.
1164                                  */
1165                                 ret = -EINVAL;
1166                                 mlog_errno(ret);
1167                                 goto out;
1168                         }
1169
1170                         page_cache_get(mmap_page);
1171                         wc->w_pages[i] = mmap_page;
1172                 } else {
1173                         wc->w_pages[i] = find_or_create_page(mapping, index,
1174                                                              GFP_NOFS);
1175                         if (!wc->w_pages[i]) {
1176                                 ret = -ENOMEM;
1177                                 mlog_errno(ret);
1178                                 goto out;
1179                         }
1180                 }
1181
1182                 if (index == target_index)
1183                         wc->w_target_page = wc->w_pages[i];
1184         }
1185 out:
1186         return ret;
1187 }
1188
1189 /*
1190  * Prepare a single cluster for write one cluster into the file.
1191  */
1192 static int ocfs2_write_cluster(struct address_space *mapping,
1193                                u32 phys, unsigned int unwritten,
1194                                struct ocfs2_alloc_context *data_ac,
1195                                struct ocfs2_alloc_context *meta_ac,
1196                                struct ocfs2_write_ctxt *wc, u32 cpos,
1197                                loff_t user_pos, unsigned user_len)
1198 {
1199         int ret, i, new, should_zero = 0;
1200         u64 v_blkno, p_blkno;
1201         struct inode *inode = mapping->host;
1202
1203         new = phys == 0 ? 1 : 0;
1204         if (new || unwritten)
1205                 should_zero = 1;
1206
1207         if (new) {
1208                 u32 tmp_pos;
1209
1210                 /*
1211                  * This is safe to call with the page locks - it won't take
1212                  * any additional semaphores or cluster locks.
1213                  */
1214                 tmp_pos = cpos;
1215                 ret = ocfs2_do_extend_allocation(OCFS2_SB(inode->i_sb), inode,
1216                                                  &tmp_pos, 1, 0, wc->w_di_bh,
1217                                                  wc->w_handle, data_ac,
1218                                                  meta_ac, NULL);
1219                 /*
1220                  * This shouldn't happen because we must have already
1221                  * calculated the correct meta data allocation required. The
1222                  * internal tree allocation code should know how to increase
1223                  * transaction credits itself.
1224                  *
1225                  * If need be, we could handle -EAGAIN for a
1226                  * RESTART_TRANS here.
1227                  */
1228                 mlog_bug_on_msg(ret == -EAGAIN,
1229                                 "Inode %llu: EAGAIN return during allocation.\n",
1230                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1231                 if (ret < 0) {
1232                         mlog_errno(ret);
1233                         goto out;
1234                 }
1235         } else if (unwritten) {
1236                 ret = ocfs2_mark_extent_written(inode, wc->w_di_bh,
1237                                                 wc->w_handle, cpos, 1, phys,
1238                                                 meta_ac, &wc->w_dealloc);
1239                 if (ret < 0) {
1240                         mlog_errno(ret);
1241                         goto out;
1242                 }
1243         }
1244
1245         if (should_zero)
1246                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1247         else
1248                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1249
1250         /*
1251          * The only reason this should fail is due to an inability to
1252          * find the extent added.
1253          */
1254         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1255                                           NULL);
1256         if (ret < 0) {
1257                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1258                             "at logical block %llu",
1259                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1260                             (unsigned long long)v_blkno);
1261                 goto out;
1262         }
1263
1264         BUG_ON(p_blkno == 0);
1265
1266         for(i = 0; i < wc->w_num_pages; i++) {
1267                 int tmpret;
1268
1269                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1270                                                       wc->w_pages[i], cpos,
1271                                                       user_pos, user_len,
1272                                                       should_zero);
1273                 if (tmpret) {
1274                         mlog_errno(tmpret);
1275                         if (ret == 0)
1276                                 tmpret = ret;
1277                 }
1278         }
1279
1280         /*
1281          * We only have cleanup to do in case of allocating write.
1282          */
1283         if (ret && new)
1284                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1285
1286 out:
1287
1288         return ret;
1289 }
1290
1291 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1292                                        struct ocfs2_alloc_context *data_ac,
1293                                        struct ocfs2_alloc_context *meta_ac,
1294                                        struct ocfs2_write_ctxt *wc,
1295                                        loff_t pos, unsigned len)
1296 {
1297         int ret, i;
1298         loff_t cluster_off;
1299         unsigned int local_len = len;
1300         struct ocfs2_write_cluster_desc *desc;
1301         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1302
1303         for (i = 0; i < wc->w_clen; i++) {
1304                 desc = &wc->w_desc[i];
1305
1306                 /*
1307                  * We have to make sure that the total write passed in
1308                  * doesn't extend past a single cluster.
1309                  */
1310                 local_len = len;
1311                 cluster_off = pos & (osb->s_clustersize - 1);
1312                 if ((cluster_off + local_len) > osb->s_clustersize)
1313                         local_len = osb->s_clustersize - cluster_off;
1314
1315                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1316                                           desc->c_unwritten, data_ac, meta_ac,
1317                                           wc, desc->c_cpos, pos, local_len);
1318                 if (ret) {
1319                         mlog_errno(ret);
1320                         goto out;
1321                 }
1322
1323                 len -= local_len;
1324                 pos += local_len;
1325         }
1326
1327         ret = 0;
1328 out:
1329         return ret;
1330 }
1331
1332 /*
1333  * ocfs2_write_end() wants to know which parts of the target page it
1334  * should complete the write on. It's easiest to compute them ahead of
1335  * time when a more complete view of the write is available.
1336  */
1337 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1338                                         struct ocfs2_write_ctxt *wc,
1339                                         loff_t pos, unsigned len, int alloc)
1340 {
1341         struct ocfs2_write_cluster_desc *desc;
1342
1343         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1344         wc->w_target_to = wc->w_target_from + len;
1345
1346         if (alloc == 0)
1347                 return;
1348
1349         /*
1350          * Allocating write - we may have different boundaries based
1351          * on page size and cluster size.
1352          *
1353          * NOTE: We can no longer compute one value from the other as
1354          * the actual write length and user provided length may be
1355          * different.
1356          */
1357
1358         if (wc->w_large_pages) {
1359                 /*
1360                  * We only care about the 1st and last cluster within
1361                  * our range and whether they should be zero'd or not. Either
1362                  * value may be extended out to the start/end of a
1363                  * newly allocated cluster.
1364                  */
1365                 desc = &wc->w_desc[0];
1366                 if (ocfs2_should_zero_cluster(desc))
1367                         ocfs2_figure_cluster_boundaries(osb,
1368                                                         desc->c_cpos,
1369                                                         &wc->w_target_from,
1370                                                         NULL);
1371
1372                 desc = &wc->w_desc[wc->w_clen - 1];
1373                 if (ocfs2_should_zero_cluster(desc))
1374                         ocfs2_figure_cluster_boundaries(osb,
1375                                                         desc->c_cpos,
1376                                                         NULL,
1377                                                         &wc->w_target_to);
1378         } else {
1379                 wc->w_target_from = 0;
1380                 wc->w_target_to = PAGE_CACHE_SIZE;
1381         }
1382 }
1383
1384 /*
1385  * Populate each single-cluster write descriptor in the write context
1386  * with information about the i/o to be done.
1387  *
1388  * Returns the number of clusters that will have to be allocated, as
1389  * well as a worst case estimate of the number of extent records that
1390  * would have to be created during a write to an unwritten region.
1391  */
1392 static int ocfs2_populate_write_desc(struct inode *inode,
1393                                      struct ocfs2_write_ctxt *wc,
1394                                      unsigned int *clusters_to_alloc,
1395                                      unsigned int *extents_to_split)
1396 {
1397         int ret;
1398         struct ocfs2_write_cluster_desc *desc;
1399         unsigned int num_clusters = 0;
1400         unsigned int ext_flags = 0;
1401         u32 phys = 0;
1402         int i;
1403
1404         *clusters_to_alloc = 0;
1405         *extents_to_split = 0;
1406
1407         for (i = 0; i < wc->w_clen; i++) {
1408                 desc = &wc->w_desc[i];
1409                 desc->c_cpos = wc->w_cpos + i;
1410
1411                 if (num_clusters == 0) {
1412                         /*
1413                          * Need to look up the next extent record.
1414                          */
1415                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1416                                                  &num_clusters, &ext_flags);
1417                         if (ret) {
1418                                 mlog_errno(ret);
1419                                 goto out;
1420                         }
1421
1422                         /*
1423                          * Assume worst case - that we're writing in
1424                          * the middle of the extent.
1425                          *
1426                          * We can assume that the write proceeds from
1427                          * left to right, in which case the extent
1428                          * insert code is smart enough to coalesce the
1429                          * next splits into the previous records created.
1430                          */
1431                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1432                                 *extents_to_split = *extents_to_split + 2;
1433                 } else if (phys) {
1434                         /*
1435                          * Only increment phys if it doesn't describe
1436                          * a hole.
1437                          */
1438                         phys++;
1439                 }
1440
1441                 desc->c_phys = phys;
1442                 if (phys == 0) {
1443                         desc->c_new = 1;
1444                         *clusters_to_alloc = *clusters_to_alloc + 1;
1445                 }
1446                 if (ext_flags & OCFS2_EXT_UNWRITTEN)
1447                         desc->c_unwritten = 1;
1448
1449                 num_clusters--;
1450         }
1451
1452         ret = 0;
1453 out:
1454         return ret;
1455 }
1456
1457 static int ocfs2_write_begin_inline(struct address_space *mapping,
1458                                     struct inode *inode,
1459                                     struct ocfs2_write_ctxt *wc)
1460 {
1461         int ret;
1462         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1463         struct page *page;
1464         handle_t *handle;
1465         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1466
1467         page = find_or_create_page(mapping, 0, GFP_NOFS);
1468         if (!page) {
1469                 ret = -ENOMEM;
1470                 mlog_errno(ret);
1471                 goto out;
1472         }
1473         /*
1474          * If we don't set w_num_pages then this page won't get unlocked
1475          * and freed on cleanup of the write context.
1476          */
1477         wc->w_pages[0] = wc->w_target_page = page;
1478         wc->w_num_pages = 1;
1479
1480         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1481         if (IS_ERR(handle)) {
1482                 ret = PTR_ERR(handle);
1483                 mlog_errno(ret);
1484                 goto out;
1485         }
1486
1487         ret = ocfs2_journal_access(handle, inode, wc->w_di_bh,
1488                                    OCFS2_JOURNAL_ACCESS_WRITE);
1489         if (ret) {
1490                 ocfs2_commit_trans(osb, handle);
1491
1492                 mlog_errno(ret);
1493                 goto out;
1494         }
1495
1496         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1497                 ocfs2_set_inode_data_inline(inode, di);
1498
1499         if (!PageUptodate(page)) {
1500                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1501                 if (ret) {
1502                         ocfs2_commit_trans(osb, handle);
1503
1504                         goto out;
1505                 }
1506         }
1507
1508         wc->w_handle = handle;
1509 out:
1510         return ret;
1511 }
1512
1513 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1514 {
1515         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1516
1517         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1518                 return 1;
1519         return 0;
1520 }
1521
1522 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1523                                           struct inode *inode, loff_t pos,
1524                                           unsigned len, struct page *mmap_page,
1525                                           struct ocfs2_write_ctxt *wc)
1526 {
1527         int ret, written = 0;
1528         loff_t end = pos + len;
1529         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1530
1531         mlog(0, "Inode %llu, write of %u bytes at off %llu. features: 0x%x\n",
1532              (unsigned long long)oi->ip_blkno, len, (unsigned long long)pos,
1533              oi->ip_dyn_features);
1534
1535         /*
1536          * Handle inodes which already have inline data 1st.
1537          */
1538         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1539                 if (mmap_page == NULL &&
1540                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1541                         goto do_inline_write;
1542
1543                 /*
1544                  * The write won't fit - we have to give this inode an
1545                  * inline extent list now.
1546                  */
1547                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1548                 if (ret)
1549                         mlog_errno(ret);
1550                 goto out;
1551         }
1552
1553         /*
1554          * Check whether the inode can accept inline data.
1555          */
1556         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1557                 return 0;
1558
1559         /*
1560          * Check whether the write can fit.
1561          */
1562         if (mmap_page || end > ocfs2_max_inline_data(inode->i_sb))
1563                 return 0;
1564
1565 do_inline_write:
1566         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1567         if (ret) {
1568                 mlog_errno(ret);
1569                 goto out;
1570         }
1571
1572         /*
1573          * This signals to the caller that the data can be written
1574          * inline.
1575          */
1576         written = 1;
1577 out:
1578         return written ? written : ret;
1579 }
1580
1581 /*
1582  * This function only does anything for file systems which can't
1583  * handle sparse files.
1584  *
1585  * What we want to do here is fill in any hole between the current end
1586  * of allocation and the end of our write. That way the rest of the
1587  * write path can treat it as an non-allocating write, which has no
1588  * special case code for sparse/nonsparse files.
1589  */
1590 static int ocfs2_expand_nonsparse_inode(struct inode *inode, loff_t pos,
1591                                         unsigned len,
1592                                         struct ocfs2_write_ctxt *wc)
1593 {
1594         int ret;
1595         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1596         loff_t newsize = pos + len;
1597
1598         if (ocfs2_sparse_alloc(osb))
1599                 return 0;
1600
1601         if (newsize <= i_size_read(inode))
1602                 return 0;
1603
1604         ret = ocfs2_extend_no_holes(inode, newsize, newsize - len);
1605         if (ret)
1606                 mlog_errno(ret);
1607
1608         return ret;
1609 }
1610
1611 int ocfs2_write_begin_nolock(struct address_space *mapping,
1612                              loff_t pos, unsigned len, unsigned flags,
1613                              struct page **pagep, void **fsdata,
1614                              struct buffer_head *di_bh, struct page *mmap_page)
1615 {
1616         int ret, credits = OCFS2_INODE_UPDATE_CREDITS;
1617         unsigned int clusters_to_alloc, extents_to_split;
1618         struct ocfs2_write_ctxt *wc;
1619         struct inode *inode = mapping->host;
1620         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1621         struct ocfs2_dinode *di;
1622         struct ocfs2_alloc_context *data_ac = NULL;
1623         struct ocfs2_alloc_context *meta_ac = NULL;
1624         handle_t *handle;
1625
1626         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1627         if (ret) {
1628                 mlog_errno(ret);
1629                 return ret;
1630         }
1631
1632         if (ocfs2_supports_inline_data(osb)) {
1633                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1634                                                      mmap_page, wc);
1635                 if (ret == 1) {
1636                         ret = 0;
1637                         goto success;
1638                 }
1639                 if (ret < 0) {
1640                         mlog_errno(ret);
1641                         goto out;
1642                 }
1643         }
1644
1645         ret = ocfs2_expand_nonsparse_inode(inode, pos, len, wc);
1646         if (ret) {
1647                 mlog_errno(ret);
1648                 goto out;
1649         }
1650
1651         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1652                                         &extents_to_split);
1653         if (ret) {
1654                 mlog_errno(ret);
1655                 goto out;
1656         }
1657
1658         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1659
1660         /*
1661          * We set w_target_from, w_target_to here so that
1662          * ocfs2_write_end() knows which range in the target page to
1663          * write out. An allocation requires that we write the entire
1664          * cluster range.
1665          */
1666         if (clusters_to_alloc || extents_to_split) {
1667                 /*
1668                  * XXX: We are stretching the limits of
1669                  * ocfs2_lock_allocators(). It greatly over-estimates
1670                  * the work to be done.
1671                  */
1672                 ret = ocfs2_lock_allocators(inode, di, clusters_to_alloc,
1673                                             extents_to_split, &data_ac, &meta_ac);
1674                 if (ret) {
1675                         mlog_errno(ret);
1676                         goto out;
1677                 }
1678
1679                 credits = ocfs2_calc_extend_credits(inode->i_sb, di,
1680                                                     clusters_to_alloc);
1681
1682         }
1683
1684         ocfs2_set_target_boundaries(osb, wc, pos, len,
1685                                     clusters_to_alloc + extents_to_split);
1686
1687         handle = ocfs2_start_trans(osb, credits);
1688         if (IS_ERR(handle)) {
1689                 ret = PTR_ERR(handle);
1690                 mlog_errno(ret);
1691                 goto out;
1692         }
1693
1694         wc->w_handle = handle;
1695
1696         /*
1697          * We don't want this to fail in ocfs2_write_end(), so do it
1698          * here.
1699          */
1700         ret = ocfs2_journal_access(handle, inode, wc->w_di_bh,
1701                                    OCFS2_JOURNAL_ACCESS_WRITE);
1702         if (ret) {
1703                 mlog_errno(ret);
1704                 goto out_commit;
1705         }
1706
1707         /*
1708          * Fill our page array first. That way we've grabbed enough so
1709          * that we can zero and flush if we error after adding the
1710          * extent.
1711          */
1712         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos,
1713                                          clusters_to_alloc + extents_to_split,
1714                                          mmap_page);
1715         if (ret) {
1716                 mlog_errno(ret);
1717                 goto out_commit;
1718         }
1719
1720         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1721                                           len);
1722         if (ret) {
1723                 mlog_errno(ret);
1724                 goto out_commit;
1725         }
1726
1727         if (data_ac)
1728                 ocfs2_free_alloc_context(data_ac);
1729         if (meta_ac)
1730                 ocfs2_free_alloc_context(meta_ac);
1731
1732 success:
1733         *pagep = wc->w_target_page;
1734         *fsdata = wc;
1735         return 0;
1736 out_commit:
1737         ocfs2_commit_trans(osb, handle);
1738
1739 out:
1740         ocfs2_free_write_ctxt(wc);
1741
1742         if (data_ac)
1743                 ocfs2_free_alloc_context(data_ac);
1744         if (meta_ac)
1745                 ocfs2_free_alloc_context(meta_ac);
1746         return ret;
1747 }
1748
1749 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1750                              loff_t pos, unsigned len, unsigned flags,
1751                              struct page **pagep, void **fsdata)
1752 {
1753         int ret;
1754         struct buffer_head *di_bh = NULL;
1755         struct inode *inode = mapping->host;
1756
1757         ret = ocfs2_meta_lock(inode, &di_bh, 1);
1758         if (ret) {
1759                 mlog_errno(ret);
1760                 return ret;
1761         }
1762
1763         /*
1764          * Take alloc sem here to prevent concurrent lookups. That way
1765          * the mapping, zeroing and tree manipulation within
1766          * ocfs2_write() will be safe against ->readpage(). This
1767          * should also serve to lock out allocation from a shared
1768          * writeable region.
1769          */
1770         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1771
1772         ret = ocfs2_data_lock(inode, 1);
1773         if (ret) {
1774                 mlog_errno(ret);
1775                 goto out_fail;
1776         }
1777
1778         ret = ocfs2_write_begin_nolock(mapping, pos, len, flags, pagep,
1779                                        fsdata, di_bh, NULL);
1780         if (ret) {
1781                 mlog_errno(ret);
1782                 goto out_fail_data;
1783         }
1784
1785         brelse(di_bh);
1786
1787         return 0;
1788
1789 out_fail_data:
1790         ocfs2_data_unlock(inode, 1);
1791 out_fail:
1792         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1793
1794         brelse(di_bh);
1795         ocfs2_meta_unlock(inode, 1);
1796
1797         return ret;
1798 }
1799
1800 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1801                                    unsigned len, unsigned *copied,
1802                                    struct ocfs2_dinode *di,
1803                                    struct ocfs2_write_ctxt *wc)
1804 {
1805         void *kaddr;
1806
1807         if (unlikely(*copied < len)) {
1808                 if (!PageUptodate(wc->w_target_page)) {
1809                         *copied = 0;
1810                         return;
1811                 }
1812         }
1813
1814         kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1815         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1816         kunmap_atomic(kaddr, KM_USER0);
1817
1818         mlog(0, "Data written to inode at offset %llu. "
1819              "id_count = %u, copied = %u, i_dyn_features = 0x%x\n",
1820              (unsigned long long)pos, *copied,
1821              le16_to_cpu(di->id2.i_data.id_count),
1822              le16_to_cpu(di->i_dyn_features));
1823 }
1824
1825 int ocfs2_write_end_nolock(struct address_space *mapping,
1826                            loff_t pos, unsigned len, unsigned copied,
1827                            struct page *page, void *fsdata)
1828 {
1829         int i;
1830         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1831         struct inode *inode = mapping->host;
1832         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1833         struct ocfs2_write_ctxt *wc = fsdata;
1834         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1835         handle_t *handle = wc->w_handle;
1836         struct page *tmppage;
1837
1838         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1839                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1840                 goto out_write_size;
1841         }
1842
1843         if (unlikely(copied < len)) {
1844                 if (!PageUptodate(wc->w_target_page))
1845                         copied = 0;
1846
1847                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1848                                        start+len);
1849         }
1850         flush_dcache_page(wc->w_target_page);
1851
1852         for(i = 0; i < wc->w_num_pages; i++) {
1853                 tmppage = wc->w_pages[i];
1854
1855                 if (tmppage == wc->w_target_page) {
1856                         from = wc->w_target_from;
1857                         to = wc->w_target_to;
1858
1859                         BUG_ON(from > PAGE_CACHE_SIZE ||
1860                                to > PAGE_CACHE_SIZE ||
1861                                to < from);
1862                 } else {
1863                         /*
1864                          * Pages adjacent to the target (if any) imply
1865                          * a hole-filling write in which case we want
1866                          * to flush their entire range.
1867                          */
1868                         from = 0;
1869                         to = PAGE_CACHE_SIZE;
1870                 }
1871
1872                 if (ocfs2_should_order_data(inode))
1873                         walk_page_buffers(wc->w_handle, page_buffers(tmppage),
1874                                           from, to, NULL,
1875                                           ocfs2_journal_dirty_data);
1876
1877                 block_commit_write(tmppage, from, to);
1878         }
1879
1880 out_write_size:
1881         pos += copied;
1882         if (pos > inode->i_size) {
1883                 i_size_write(inode, pos);
1884                 mark_inode_dirty(inode);
1885         }
1886         inode->i_blocks = ocfs2_inode_sector_count(inode);
1887         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1888         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1889         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1890         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1891         ocfs2_journal_dirty(handle, wc->w_di_bh);
1892
1893         ocfs2_commit_trans(osb, handle);
1894
1895         ocfs2_run_deallocs(osb, &wc->w_dealloc);
1896
1897         ocfs2_free_write_ctxt(wc);
1898
1899         return copied;
1900 }
1901
1902 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
1903                            loff_t pos, unsigned len, unsigned copied,
1904                            struct page *page, void *fsdata)
1905 {
1906         int ret;
1907         struct inode *inode = mapping->host;
1908
1909         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
1910
1911         ocfs2_data_unlock(inode, 1);
1912         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1913         ocfs2_meta_unlock(inode, 1);
1914
1915         return ret;
1916 }
1917
1918 const struct address_space_operations ocfs2_aops = {
1919         .readpage       = ocfs2_readpage,
1920         .writepage      = ocfs2_writepage,
1921         .write_begin    = ocfs2_write_begin,
1922         .write_end      = ocfs2_write_end,
1923         .bmap           = ocfs2_bmap,
1924         .sync_page      = block_sync_page,
1925         .direct_IO      = ocfs2_direct_IO,
1926         .invalidatepage = ocfs2_invalidatepage,
1927         .releasepage    = ocfs2_releasepage,
1928         .migratepage    = buffer_migrate_page,
1929 };