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