fs/ocfs2/aops.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/fs.h>
   7 #include <linux/slab.h>
   8 #include <linux/highmem.h>
   9 #include <linux/pagemap.h>
  10 #include <asm/byteorder.h>
  11 #include <linux/swap.h>
  12 #include <linux/mpage.h>
  13 #include <linux/quotaops.h>
  14 #include <linux/blkdev.h>
  15 #include <linux/uio.h>
  16 #include <linux/mm.h>
  17
  18 #include <cluster/masklog.h>
  19
  20 #include "ocfs2.h"
  21
  22 #include "alloc.h"
  23 #include "aops.h"
  24 #include "dlmglue.h"
  25 #include "extent_map.h"
  26 #include "file.h"
  27 #include "inode.h"
  28 #include "journal.h"
  29 #include "suballoc.h"
  30 #include "super.h"
  31 #include "symlink.h"
  32 #include "refcounttree.h"
  33 #include "ocfs2_trace.h"
  34
  35 #include "buffer_head_io.h"
  36 #include "dir.h"
  37 #include "namei.h"
  38 #include "sysfile.h"
  39
  40 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
  41                                    struct buffer_head *bh_result, int create)
  42 {
  43         int err = -EIO;
  44         int status;
  45         struct ocfs2_dinode *fe = NULL;
  46         struct buffer_head *bh = NULL;
  47         struct buffer_head *buffer_cache_bh = NULL;
  48         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  49         void *kaddr;
  50
  51         trace_ocfs2_symlink_get_block(
  52                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
  53                         (unsigned long long)iblock, bh_result, create);
  54
  55         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
  56
  57         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
  58                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
  59                      (unsigned long long)iblock);
  60                 goto bail;
  61         }
  62
  63         status = ocfs2_read_inode_block(inode, &bh);
  64         if (status < 0) {
  65                 mlog_errno(status);
  66                 goto bail;
  67         }
  68         fe = (struct ocfs2_dinode *) bh->b_data;
  69
  70         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
  71                                                     le32_to_cpu(fe->i_clusters))) {
  72                 err = -ENOMEM;
  73                 mlog(ML_ERROR, "block offset is outside the allocated size: "
  74                      "%llu\n", (unsigned long long)iblock);
  75                 goto bail;
  76         }
  77
  78         /* We don't use the page cache to create symlink data, so if
  79          * need be, copy it over from the buffer cache. */
  80         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  81                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  82                             iblock;
  83                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
  84                 if (!buffer_cache_bh) {
  85                         err = -ENOMEM;
  86                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
  87                         goto bail;
  88                 }
  89
  90                 /* we haven't locked out transactions, so a commit
  91                  * could've happened. Since we've got a reference on
  92                  * the bh, even if it commits while we're doing the
  93                  * copy, the data is still good. */
  94                 if (buffer_jbd(buffer_cache_bh)
  95                     && ocfs2_inode_is_new(inode)) {
  96                         kaddr = kmap_atomic(bh_result->b_page);
  97                         if (!kaddr) {
  98                                 mlog(ML_ERROR, "couldn't kmap!\n");
  99                                 goto bail;
 100                         }
 101                         memcpy(kaddr + (bh_result->b_size * iblock),
 102                                buffer_cache_bh->b_data,
 103                                bh_result->b_size);
 104                         kunmap_atomic(kaddr);
 105                         set_buffer_uptodate(bh_result);
 106                 }
 107                 brelse(buffer_cache_bh);
 108         }
 109
 110         map_bh(bh_result, inode->i_sb,
 111                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 112
 113         err = 0;
 114
 115 bail:
 116         brelse(bh);
 117
 118         return err;
 119 }
 120
 121 static int ocfs2_lock_get_block(struct inode *inode, sector_t iblock,
 122                     struct buffer_head *bh_result, int create)
 123 {
 124         int ret = 0;
 125         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 126
 127         down_read(&oi->ip_alloc_sem);
 128         ret = ocfs2_get_block(inode, iblock, bh_result, create);
 129         up_read(&oi->ip_alloc_sem);
 130
 131         return ret;
 132 }
 133
 134 int ocfs2_get_block(struct inode *inode, sector_t iblock,
 135                     struct buffer_head *bh_result, int create)
 136 {
 137         int err = 0;
 138         unsigned int ext_flags;
 139         u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
 140         u64 p_blkno, count, past_eof;
 141         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 142
 143         trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno,
 144                               (unsigned long long)iblock, bh_result, create);
 145
 146         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
 147                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
 148                      inode, inode->i_ino);
 149
 150         if (S_ISLNK(inode->i_mode)) {
 151                 /* this always does I/O for some reason. */
 152                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
 153                 goto bail;
 154         }
 155
 156         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
 157                                           &ext_flags);
 158         if (err) {
 159                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
 160                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
 161                      (unsigned long long)p_blkno);
 162                 goto bail;
 163         }
 164
 165         if (max_blocks < count)
 166                 count = max_blocks;
 167
 168         /*
 169          * ocfs2 never allocates in this function - the only time we
 170          * need to use BH_New is when we're extending i_size on a file
 171          * system which doesn't support holes, in which case BH_New
 172          * allows __block_write_begin() to zero.
 173          *
 174          * If we see this on a sparse file system, then a truncate has
 175          * raced us and removed the cluster. In this case, we clear
 176          * the buffers dirty and uptodate bits and let the buffer code
 177          * ignore it as a hole.
 178          */
 179         if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
 180                 clear_buffer_dirty(bh_result);
 181                 clear_buffer_uptodate(bh_result);
 182                 goto bail;
 183         }
 184
 185         /* Treat the unwritten extent as a hole for zeroing purposes. */
 186         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 187                 map_bh(bh_result, inode->i_sb, p_blkno);
 188
 189         bh_result->b_size = count << inode->i_blkbits;
 190
 191         if (!ocfs2_sparse_alloc(osb)) {
 192                 if (p_blkno == 0) {
 193                         err = -EIO;
 194                         mlog(ML_ERROR,
 195                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
 196                              (unsigned long long)iblock,
 197                              (unsigned long long)p_blkno,
 198                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
 199                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
 200                         dump_stack();
 201                         goto bail;
 202                 }
 203         }
 204
 205         past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 206
 207         trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno,
 208                                   (unsigned long long)past_eof);
 209         if (create && (iblock >= past_eof))
 210                 set_buffer_new(bh_result);
 211
 212 bail:
 213         if (err < 0)
 214                 err = -EIO;
 215
 216         return err;
 217 }
 218
 219 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
 220                            struct buffer_head *di_bh)
 221 {
 222         void *kaddr;
 223         loff_t size;
 224         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 225
 226         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
 227                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag\n",
 228                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
 229                 return -EROFS;
 230         }
 231
 232         size = i_size_read(inode);
 233
 234         if (size > PAGE_SIZE ||
 235             size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
 236                 ocfs2_error(inode->i_sb,
 237                             "Inode %llu has with inline data has bad size: %Lu\n",
 238                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 239                             (unsigned long long)size);
 240                 return -EROFS;
 241         }
 242
 243         kaddr = kmap_atomic(page);
 244         if (size)
 245                 memcpy(kaddr, di->id2.i_data.id_data, size);
 246         /* Clear the remaining part of the page */
 247         memset(kaddr + size, 0, PAGE_SIZE - size);
 248         flush_dcache_page(page);
 249         kunmap_atomic(kaddr);
 250
 251         SetPageUptodate(page);
 252
 253         return 0;
 254 }
 255
 256 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
 257 {
 258         int ret;
 259         struct buffer_head *di_bh = NULL;
 260
 261         BUG_ON(!PageLocked(page));
 262         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
 263
 264         ret = ocfs2_read_inode_block(inode, &di_bh);
 265         if (ret) {
 266                 mlog_errno(ret);
 267                 goto out;
 268         }
 269
 270         ret = ocfs2_read_inline_data(inode, page, di_bh);
 271 out:
 272         unlock_page(page);
 273
 274         brelse(di_bh);
 275         return ret;
 276 }
 277
 278 static int ocfs2_readpage(struct file *file, struct page *page)
 279 {
 280         struct inode *inode = page->mapping->host;
 281         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 282         loff_t start = (loff_t)page->index << PAGE_SHIFT;
 283         int ret, unlock = 1;
 284
 285         trace_ocfs2_readpage((unsigned long long)oi->ip_blkno,
 286                              (page ? page->index : 0));
 287
 288         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
 289         if (ret != 0) {
 290                 if (ret == AOP_TRUNCATED_PAGE)
 291                         unlock = 0;
 292                 mlog_errno(ret);
 293                 goto out;
 294         }
 295
 296         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 297                 /*
 298                  * Unlock the page and cycle ip_alloc_sem so that we don't
 299                  * busyloop waiting for ip_alloc_sem to unlock
 300                  */
 301                 ret = AOP_TRUNCATED_PAGE;
 302                 unlock_page(page);
 303                 unlock = 0;
 304                 down_read(&oi->ip_alloc_sem);
 305                 up_read(&oi->ip_alloc_sem);
 306                 goto out_inode_unlock;
 307         }
 308
 309         /*
 310          * i_size might have just been updated as we grabed the meta lock.  We
 311          * might now be discovering a truncate that hit on another node.
 312          * block_read_full_page->get_block freaks out if it is asked to read
 313          * beyond the end of a file, so we check here.  Callers
 314          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
 315          * and notice that the page they just read isn't needed.
 316          *
 317          * XXX sys_readahead() seems to get that wrong?
 318          */
 319         if (start >= i_size_read(inode)) {
 320                 zero_user(page, 0, PAGE_SIZE);
 321                 SetPageUptodate(page);
 322                 ret = 0;
 323                 goto out_alloc;
 324         }
 325
 326         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 327                 ret = ocfs2_readpage_inline(inode, page);
 328         else
 329                 ret = block_read_full_page(page, ocfs2_get_block);
 330         unlock = 0;
 331
 332 out_alloc:
 333         up_read(&oi->ip_alloc_sem);
 334 out_inode_unlock:
 335         ocfs2_inode_unlock(inode, 0);
 336 out:
 337         if (unlock)
 338                 unlock_page(page);
 339         return ret;
 340 }
 341
 342 /*
 343  * This is used only for read-ahead. Failures or difficult to handle
 344  * situations are safe to ignore.
 345  *
 346  * Right now, we don't bother with BH_Boundary - in-inode extent lists
 347  * are quite large (243 extents on 4k blocks), so most inodes don't
 348  * grow out to a tree. If need be, detecting boundary extents could
 349  * trivially be added in a future version of ocfs2_get_block().
 350  */
 351 static void ocfs2_readahead(struct readahead_control *rac)
 352 {
 353         int ret;
 354         struct inode *inode = rac->mapping->host;
 355         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 356
 357         /*
 358          * Use the nonblocking flag for the dlm code to avoid page
 359          * lock inversion, but don't bother with retrying.
 360          */
 361         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
 362         if (ret)
 363                 return;
 364
 365         if (down_read_trylock(&oi->ip_alloc_sem) == 0)
 366                 goto out_unlock;
 367
 368         /*
 369          * Don't bother with inline-data. There isn't anything
 370          * to read-ahead in that case anyway...
 371          */
 372         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 373                 goto out_up;
 374
 375         /*
 376          * Check whether a remote node truncated this file - we just
 377          * drop out in that case as it's not worth handling here.
 378          */
 379         if (readahead_pos(rac) >= i_size_read(inode))
 380                 goto out_up;
 381
 382         mpage_readahead(rac, ocfs2_get_block);
 383
 384 out_up:
 385         up_read(&oi->ip_alloc_sem);
 386 out_unlock:
 387         ocfs2_inode_unlock(inode, 0);
 388 }
 389
 390 /* Note: Because we don't support holes, our allocation has
 391  * already happened (allocation writes zeros to the file data)
 392  * so we don't have to worry about ordered writes in
 393  * ocfs2_writepage.
 394  *
 395  * ->writepage is called during the process of invalidating the page cache
 396  * during blocked lock processing.  It can't block on any cluster locks
 397  * to during block mapping.  It's relying on the fact that the block
 398  * mapping can't have disappeared under the dirty pages that it is
 399  * being asked to write back.
 400  */
 401 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
 402 {
 403         trace_ocfs2_writepage(
 404                 (unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno,
 405                 page->index);
 406
 407         return block_write_full_page(page, ocfs2_get_block, wbc);
 408 }
 409
 410 /* Taken from ext3. We don't necessarily need the full blown
 411  * functionality yet, but IMHO it's better to cut and paste the whole
 412  * thing so we can avoid introducing our own bugs (and easily pick up
 413  * their fixes when they happen) --Mark */
 414 int walk_page_buffers(  handle_t *handle,
 415                         struct buffer_head *head,
 416                         unsigned from,
 417                         unsigned to,
 418                         int *partial,
 419                         int (*fn)(      handle_t *handle,
 420                                         struct buffer_head *bh))
 421 {
 422         struct buffer_head *bh;
 423         unsigned block_start, block_end;
 424         unsigned blocksize = head->b_size;
 425         int err, ret = 0;
 426         struct buffer_head *next;
 427
 428         for (   bh = head, block_start = 0;
 429                 ret == 0 && (bh != head || !block_start);
 430                 block_start = block_end, bh = next)
 431         {
 432                 next = bh->b_this_page;
 433                 block_end = block_start + blocksize;
 434                 if (block_end <= from || block_start >= to) {
 435                         if (partial && !buffer_uptodate(bh))
 436                                 *partial = 1;
 437                         continue;
 438                 }
 439                 err = (*fn)(handle, bh);
 440                 if (!ret)
 441                         ret = err;
 442         }
 443         return ret;
 444 }
 445
 446 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 447 {
 448         sector_t status;
 449         u64 p_blkno = 0;
 450         int err = 0;
 451         struct inode *inode = mapping->host;
 452
 453         trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
 454                          (unsigned long long)block);
 455
 456         /*
 457          * The swap code (ab-)uses ->bmap to get a block mapping and then
 458          * bypasseѕ the file system for actual I/O.  We really can't allow
 459          * that on refcounted inodes, so we have to skip out here.  And yes,
 460          * 0 is the magic code for a bmap error..
 461          */
 462         if (ocfs2_is_refcount_inode(inode))
 463                 return 0;
 464
 465         /* We don't need to lock journal system files, since they aren't
 466          * accessed concurrently from multiple nodes.
 467          */
 468         if (!INODE_JOURNAL(inode)) {
 469                 err = ocfs2_inode_lock(inode, NULL, 0);
 470                 if (err) {
 471                         if (err != -ENOENT)
 472                                 mlog_errno(err);
 473                         goto bail;
 474                 }
 475                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
 476         }
 477
 478         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
 479                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
 480                                                   NULL);
 481
 482         if (!INODE_JOURNAL(inode)) {
 483                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
 484                 ocfs2_inode_unlock(inode, 0);
 485         }
 486
 487         if (err) {
 488                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 489                      (unsigned long long)block);
 490                 mlog_errno(err);
 491                 goto bail;
 492         }
 493
 494 bail:
 495         status = err ? 0 : p_blkno;
 496
 497         return status;
 498 }
 499
 500 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 501 {
 502         if (!page_has_buffers(page))
 503                 return 0;
 504         return try_to_free_buffers(page);
 505 }
 506
 507 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 508                                             u32 cpos,
 509                                             unsigned int *start,
 510                                             unsigned int *end)
 511 {
 512         unsigned int cluster_start = 0, cluster_end = PAGE_SIZE;
 513
 514         if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits)) {
 515                 unsigned int cpp;
 516
 517                 cpp = 1 << (PAGE_SHIFT - osb->s_clustersize_bits);
 518
 519                 cluster_start = cpos % cpp;
 520                 cluster_start = cluster_start << osb->s_clustersize_bits;
 521
 522                 cluster_end = cluster_start + osb->s_clustersize;
 523         }
 524
 525         BUG_ON(cluster_start > PAGE_SIZE);
 526         BUG_ON(cluster_end > PAGE_SIZE);
 527
 528         if (start)
 529                 *start = cluster_start;
 530         if (end)
 531                 *end = cluster_end;
 532 }
 533
 534 /*
 535  * 'from' and 'to' are the region in the page to avoid zeroing.
 536  *
 537  * If pagesize > clustersize, this function will avoid zeroing outside
 538  * of the cluster boundary.
 539  *
 540  * from == to == 0 is code for "zero the entire cluster region"
 541  */
 542 static void ocfs2_clear_page_regions(struct page *page,
 543                                      struct ocfs2_super *osb, u32 cpos,
 544                                      unsigned from, unsigned to)
 545 {
 546         void *kaddr;
 547         unsigned int cluster_start, cluster_end;
 548
 549         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 550
 551         kaddr = kmap_atomic(page);
 552
 553         if (from || to) {
 554                 if (from > cluster_start)
 555                         memset(kaddr + cluster_start, 0, from - cluster_start);
 556                 if (to < cluster_end)
 557                         memset(kaddr + to, 0, cluster_end - to);
 558         } else {
 559                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 560         }
 561
 562         kunmap_atomic(kaddr);
 563 }
 564
 565 /*
 566  * Nonsparse file systems fully allocate before we get to the write
 567  * code. This prevents ocfs2_write() from tagging the write as an
 568  * allocating one, which means ocfs2_map_page_blocks() might try to
 569  * read-in the blocks at the tail of our file. Avoid reading them by
 570  * testing i_size against each block offset.
 571  */
 572 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 573                                  unsigned int block_start)
 574 {
 575         u64 offset = page_offset(page) + block_start;
 576
 577         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 578                 return 1;
 579
 580         if (i_size_read(inode) > offset)
 581                 return 1;
 582
 583         return 0;
 584 }
 585
 586 /*
 587  * Some of this taken from __block_write_begin(). We already have our
 588  * mapping by now though, and the entire write will be allocating or
 589  * it won't, so not much need to use BH_New.
 590  *
 591  * This will also skip zeroing, which is handled externally.
 592  */
 593 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 594                           struct inode *inode, unsigned int from,
 595                           unsigned int to, int new)
 596 {
 597         int ret = 0;
 598         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 599         unsigned int block_end, block_start;
 600         unsigned int bsize = i_blocksize(inode);
 601
 602         if (!page_has_buffers(page))
 603                 create_empty_buffers(page, bsize, 0);
 604
 605         head = page_buffers(page);
 606         for (bh = head, block_start = 0; bh != head || !block_start;
 607              bh = bh->b_this_page, block_start += bsize) {
 608                 block_end = block_start + bsize;
 609
 610                 clear_buffer_new(bh);
 611
 612                 /*
 613                  * Ignore blocks outside of our i/o range -
 614                  * they may belong to unallocated clusters.
 615                  */
 616                 if (block_start >= to || block_end <= from) {
 617                         if (PageUptodate(page))
 618                                 set_buffer_uptodate(bh);
 619                         continue;
 620                 }
 621
 622                 /*
 623                  * For an allocating write with cluster size >= page
 624                  * size, we always write the entire page.
 625                  */
 626                 if (new)
 627                         set_buffer_new(bh);
 628
 629                 if (!buffer_mapped(bh)) {
 630                         map_bh(bh, inode->i_sb, *p_blkno);
 631                         clean_bdev_bh_alias(bh);
 632                 }
 633
 634                 if (PageUptodate(page)) {
 635                         set_buffer_uptodate(bh);
 636                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 637                            !buffer_new(bh) &&
 638                            ocfs2_should_read_blk(inode, page, block_start) &&
 639                            (block_start < from || block_end > to)) {
 640                         ll_rw_block(REQ_OP_READ, 0, 1, &bh);
 641                         *wait_bh++=bh;
 642                 }
 643
 644                 *p_blkno = *p_blkno + 1;
 645         }
 646
 647         /*
 648          * If we issued read requests - let them complete.
 649          */
 650         while(wait_bh > wait) {
 651                 wait_on_buffer(*--wait_bh);
 652                 if (!buffer_uptodate(*wait_bh))
 653                         ret = -EIO;
 654         }
 655
 656         if (ret == 0 || !new)
 657                 return ret;
 658
 659         /*
 660          * If we get -EIO above, zero out any newly allocated blocks
 661          * to avoid exposing stale data.
 662          */
 663         bh = head;
 664         block_start = 0;
 665         do {
 666                 block_end = block_start + bsize;
 667                 if (block_end <= from)
 668                         goto next_bh;
 669                 if (block_start >= to)
 670                         break;
 671
 672                 zero_user(page, block_start, bh->b_size);
 673                 set_buffer_uptodate(bh);
 674                 mark_buffer_dirty(bh);
 675
 676 next_bh:
 677                 block_start = block_end;
 678                 bh = bh->b_this_page;
 679         } while (bh != head);
 680
 681         return ret;
 682 }
 683
 684 #if (PAGE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 685 #define OCFS2_MAX_CTXT_PAGES    1
 686 #else
 687 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_SIZE)
 688 #endif
 689
 690 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 691
 692 struct ocfs2_unwritten_extent {
 693         struct list_head        ue_node;
 694         struct list_head        ue_ip_node;
 695         u32                     ue_cpos;
 696         u32                     ue_phys;
 697 };
 698
 699 /*
 700  * Describe the state of a single cluster to be written to.
 701  */
 702 struct ocfs2_write_cluster_desc {
 703         u32             c_cpos;
 704         u32             c_phys;
 705         /*
 706          * Give this a unique field because c_phys eventually gets
 707          * filled.
 708          */
 709         unsigned        c_new;
 710         unsigned        c_clear_unwritten;
 711         unsigned        c_needs_zero;
 712 };
 713
 714 struct ocfs2_write_ctxt {
 715         /* Logical cluster position / len of write */
 716         u32                             w_cpos;
 717         u32                             w_clen;
 718
 719         /* First cluster allocated in a nonsparse extend */
 720         u32                             w_first_new_cpos;
 721
 722         /* Type of caller. Must be one of buffer, mmap, direct.  */
 723         ocfs2_write_type_t              w_type;
 724
 725         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 726
 727         /*
 728          * This is true if page_size > cluster_size.
 729          *
 730          * It triggers a set of special cases during write which might
 731          * have to deal with allocating writes to partial pages.
 732          */
 733         unsigned int                    w_large_pages;
 734
 735         /*
 736          * Pages involved in this write.
 737          *
 738          * w_target_page is the page being written to by the user.
 739          *
 740          * w_pages is an array of pages which always contains
 741          * w_target_page, and in the case of an allocating write with
 742          * page_size < cluster size, it will contain zero'd and mapped
 743          * pages adjacent to w_target_page which need to be written
 744          * out in so that future reads from that region will get
 745          * zero's.
 746          */
 747         unsigned int                    w_num_pages;
 748         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 749         struct page                     *w_target_page;
 750
 751         /*
 752          * w_target_locked is used for page_mkwrite path indicating no unlocking
 753          * against w_target_page in ocfs2_write_end_nolock.
 754          */
 755         unsigned int                    w_target_locked:1;
 756
 757         /*
 758          * ocfs2_write_end() uses this to know what the real range to
 759          * write in the target should be.
 760          */
 761         unsigned int                    w_target_from;
 762         unsigned int                    w_target_to;
 763
 764         /*
 765          * We could use journal_current_handle() but this is cleaner,
 766          * IMHO -Mark
 767          */
 768         handle_t                        *w_handle;
 769
 770         struct buffer_head              *w_di_bh;
 771
 772         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 773
 774         struct list_head                w_unwritten_list;
 775         unsigned int                    w_unwritten_count;
 776 };
 777
 778 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 779 {
 780         int i;
 781
 782         for(i = 0; i < num_pages; i++) {
 783                 if (pages[i]) {
 784                         unlock_page(pages[i]);
 785                         mark_page_accessed(pages[i]);
 786                         put_page(pages[i]);
 787                 }
 788         }
 789 }
 790
 791 static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc)
 792 {
 793         int i;
 794
 795         /*
 796          * w_target_locked is only set to true in the page_mkwrite() case.
 797          * The intent is to allow us to lock the target page from write_begin()
 798          * to write_end(). The caller must hold a ref on w_target_page.
 799          */
 800         if (wc->w_target_locked) {
 801                 BUG_ON(!wc->w_target_page);
 802                 for (i = 0; i < wc->w_num_pages; i++) {
 803                         if (wc->w_target_page == wc->w_pages[i]) {
 804                                 wc->w_pages[i] = NULL;
 805                                 break;
 806                         }
 807                 }
 808                 mark_page_accessed(wc->w_target_page);
 809                 put_page(wc->w_target_page);
 810         }
 811         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 812 }
 813
 814 static void ocfs2_free_unwritten_list(struct inode *inode,
 815                                  struct list_head *head)
 816 {
 817         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 818         struct ocfs2_unwritten_extent *ue = NULL, *tmp = NULL;
 819
 820         list_for_each_entry_safe(ue, tmp, head, ue_node) {
 821                 list_del(&ue->ue_node);
 822                 spin_lock(&oi->ip_lock);
 823                 list_del(&ue->ue_ip_node);
 824                 spin_unlock(&oi->ip_lock);
 825                 kfree(ue);
 826         }
 827 }
 828
 829 static void ocfs2_free_write_ctxt(struct inode *inode,
 830                                   struct ocfs2_write_ctxt *wc)
 831 {
 832         ocfs2_free_unwritten_list(inode, &wc->w_unwritten_list);
 833         ocfs2_unlock_pages(wc);
 834         brelse(wc->w_di_bh);
 835         kfree(wc);
 836 }
 837
 838 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 839                                   struct ocfs2_super *osb, loff_t pos,
 840                                   unsigned len, ocfs2_write_type_t type,
 841                                   struct buffer_head *di_bh)
 842 {
 843         u32 cend;
 844         struct ocfs2_write_ctxt *wc;
 845
 846         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 847         if (!wc)
 848                 return -ENOMEM;
 849
 850         wc->w_cpos = pos >> osb->s_clustersize_bits;
 851         wc->w_first_new_cpos = UINT_MAX;
 852         cend = (pos + len - 1) >> osb->s_clustersize_bits;
 853         wc->w_clen = cend - wc->w_cpos + 1;
 854         get_bh(di_bh);
 855         wc->w_di_bh = di_bh;
 856         wc->w_type = type;
 857
 858         if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits))
 859                 wc->w_large_pages = 1;
 860         else
 861                 wc->w_large_pages = 0;
 862
 863         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 864         INIT_LIST_HEAD(&wc->w_unwritten_list);
 865
 866         *wcp = wc;
 867
 868         return 0;
 869 }
 870
 871 /*
 872  * If a page has any new buffers, zero them out here, and mark them uptodate
 873  * and dirty so they'll be written out (in order to prevent uninitialised
 874  * block data from leaking). And clear the new bit.
 875  */
 876 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 877 {
 878         unsigned int block_start, block_end;
 879         struct buffer_head *head, *bh;
 880
 881         BUG_ON(!PageLocked(page));
 882         if (!page_has_buffers(page))
 883                 return;
 884
 885         bh = head = page_buffers(page);
 886         block_start = 0;
 887         do {
 888                 block_end = block_start + bh->b_size;
 889
 890                 if (buffer_new(bh)) {
 891                         if (block_end > from && block_start < to) {
 892                                 if (!PageUptodate(page)) {
 893                                         unsigned start, end;
 894
 895                                         start = max(from, block_start);
 896                                         end = min(to, block_end);
 897
 898                                         zero_user_segment(page, start, end);
 899                                         set_buffer_uptodate(bh);
 900                                 }
 901
 902                                 clear_buffer_new(bh);
 903                                 mark_buffer_dirty(bh);
 904                         }
 905                 }
 906
 907                 block_start = block_end;
 908                 bh = bh->b_this_page;
 909         } while (bh != head);
 910 }
 911
 912 /*
 913  * Only called when we have a failure during allocating write to write
 914  * zero's to the newly allocated region.
 915  */
 916 static void ocfs2_write_failure(struct inode *inode,
 917                                 struct ocfs2_write_ctxt *wc,
 918                                 loff_t user_pos, unsigned user_len)
 919 {
 920         int i;
 921         unsigned from = user_pos & (PAGE_SIZE - 1),
 922                 to = user_pos + user_len;
 923         struct page *tmppage;
 924
 925         if (wc->w_target_page)
 926                 ocfs2_zero_new_buffers(wc->w_target_page, from, to);
 927
 928         for(i = 0; i < wc->w_num_pages; i++) {
 929                 tmppage = wc->w_pages[i];
 930
 931                 if (tmppage && page_has_buffers(tmppage)) {
 932                         if (ocfs2_should_order_data(inode))
 933                                 ocfs2_jbd2_inode_add_write(wc->w_handle, inode,
 934                                                            user_pos, user_len);
 935
 936                         block_commit_write(tmppage, from, to);
 937                 }
 938         }
 939 }
 940
 941 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
 942                                         struct ocfs2_write_ctxt *wc,
 943                                         struct page *page, u32 cpos,
 944                                         loff_t user_pos, unsigned user_len,
 945                                         int new)
 946 {
 947         int ret;
 948         unsigned int map_from = 0, map_to = 0;
 949         unsigned int cluster_start, cluster_end;
 950         unsigned int user_data_from = 0, user_data_to = 0;
 951
 952         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
 953                                         &cluster_start, &cluster_end);
 954
 955         /* treat the write as new if the a hole/lseek spanned across
 956          * the page boundary.
 957          */
 958         new = new | ((i_size_read(inode) <= page_offset(page)) &&
 959                         (page_offset(page) <= user_pos));
 960
 961         if (page == wc->w_target_page) {
 962                 map_from = user_pos & (PAGE_SIZE - 1);
 963                 map_to = map_from + user_len;
 964
 965                 if (new)
 966                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 967                                                     cluster_start, cluster_end,
 968                                                     new);
 969                 else
 970                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 971                                                     map_from, map_to, new);
 972                 if (ret) {
 973                         mlog_errno(ret);
 974                         goto out;
 975                 }
 976
 977                 user_data_from = map_from;
 978                 user_data_to = map_to;
 979                 if (new) {
 980                         map_from = cluster_start;
 981                         map_to = cluster_end;
 982                 }
 983         } else {
 984                 /*
 985                  * If we haven't allocated the new page yet, we
 986                  * shouldn't be writing it out without copying user
 987                  * data. This is likely a math error from the caller.
 988                  */
 989                 BUG_ON(!new);
 990
 991                 map_from = cluster_start;
 992                 map_to = cluster_end;
 993
 994                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 995                                             cluster_start, cluster_end, new);
 996                 if (ret) {
 997                         mlog_errno(ret);
 998                         goto out;
 999                 }
1000         }
1001
1002         /*
1003          * Parts of newly allocated pages need to be zero'd.
1004          *
1005          * Above, we have also rewritten 'to' and 'from' - as far as
1006          * the rest of the function is concerned, the entire cluster
1007          * range inside of a page needs to be written.
1008          *
1009          * We can skip this if the page is up to date - it's already
1010          * been zero'd from being read in as a hole.
1011          */
1012         if (new && !PageUptodate(page))
1013                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1014                                          cpos, user_data_from, user_data_to);
1015
1016         flush_dcache_page(page);
1017
1018 out:
1019         return ret;
1020 }
1021
1022 /*
1023  * This function will only grab one clusters worth of pages.
1024  */
1025 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1026                                       struct ocfs2_write_ctxt *wc,
1027                                       u32 cpos, loff_t user_pos,
1028                                       unsigned user_len, int new,
1029                                       struct page *mmap_page)
1030 {
1031         int ret = 0, i;
1032         unsigned long start, target_index, end_index, index;
1033         struct inode *inode = mapping->host;
1034         loff_t last_byte;
1035
1036         target_index = user_pos >> PAGE_SHIFT;
1037
1038         /*
1039          * Figure out how many pages we'll be manipulating here. For
1040          * non allocating write, we just change the one
1041          * page. Otherwise, we'll need a whole clusters worth.  If we're
1042          * writing past i_size, we only need enough pages to cover the
1043          * last page of the write.
1044          */
1045         if (new) {
1046                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1047                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1048                 /*
1049                  * We need the index *past* the last page we could possibly
1050                  * touch.  This is the page past the end of the write or
1051                  * i_size, whichever is greater.
1052                  */
1053                 last_byte = max(user_pos + user_len, i_size_read(inode));
1054                 BUG_ON(last_byte < 1);
1055                 end_index = ((last_byte - 1) >> PAGE_SHIFT) + 1;
1056                 if ((start + wc->w_num_pages) > end_index)
1057                         wc->w_num_pages = end_index - start;
1058         } else {
1059                 wc->w_num_pages = 1;
1060                 start = target_index;
1061         }
1062         end_index = (user_pos + user_len - 1) >> PAGE_SHIFT;
1063
1064         for(i = 0; i < wc->w_num_pages; i++) {
1065                 index = start + i;
1066
1067                 if (index >= target_index && index <= end_index &&
1068                     wc->w_type == OCFS2_WRITE_MMAP) {
1069                         /*
1070                          * ocfs2_pagemkwrite() is a little different
1071                          * and wants us to directly use the page
1072                          * passed in.
1073                          */
1074                         lock_page(mmap_page);
1075
1076                         /* Exit and let the caller retry */
1077                         if (mmap_page->mapping != mapping) {
1078                                 WARN_ON(mmap_page->mapping);
1079                                 unlock_page(mmap_page);
1080                                 ret = -EAGAIN;
1081                                 goto out;
1082                         }
1083
1084                         get_page(mmap_page);
1085                         wc->w_pages[i] = mmap_page;
1086                         wc->w_target_locked = true;
1087                 } else if (index >= target_index && index <= end_index &&
1088                            wc->w_type == OCFS2_WRITE_DIRECT) {
1089                         /* Direct write has no mapping page. */
1090                         wc->w_pages[i] = NULL;
1091                         continue;
1092                 } else {
1093                         wc->w_pages[i] = find_or_create_page(mapping, index,
1094                                                              GFP_NOFS);
1095                         if (!wc->w_pages[i]) {
1096                                 ret = -ENOMEM;
1097                                 mlog_errno(ret);
1098                                 goto out;
1099                         }
1100                 }
1101                 wait_for_stable_page(wc->w_pages[i]);
1102
1103                 if (index == target_index)
1104                         wc->w_target_page = wc->w_pages[i];
1105         }
1106 out:
1107         if (ret)
1108                 wc->w_target_locked = false;
1109         return ret;
1110 }
1111
1112 /*
1113  * Prepare a single cluster for write one cluster into the file.
1114  */
1115 static int ocfs2_write_cluster(struct address_space *mapping,
1116                                u32 *phys, unsigned int new,
1117                                unsigned int clear_unwritten,
1118                                unsigned int should_zero,
1119                                struct ocfs2_alloc_context *data_ac,
1120                                struct ocfs2_alloc_context *meta_ac,
1121                                struct ocfs2_write_ctxt *wc, u32 cpos,
1122                                loff_t user_pos, unsigned user_len)
1123 {
1124         int ret, i;
1125         u64 p_blkno;
1126         struct inode *inode = mapping->host;
1127         struct ocfs2_extent_tree et;
1128         int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1);
1129
1130         if (new) {
1131                 u32 tmp_pos;
1132
1133                 /*
1134                  * This is safe to call with the page locks - it won't take
1135                  * any additional semaphores or cluster locks.
1136                  */
1137                 tmp_pos = cpos;
1138                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1139                                            &tmp_pos, 1, !clear_unwritten,
1140                                            wc->w_di_bh, wc->w_handle,
1141                                            data_ac, meta_ac, NULL);
1142                 /*
1143                  * This shouldn't happen because we must have already
1144                  * calculated the correct meta data allocation required. The
1145                  * internal tree allocation code should know how to increase
1146                  * transaction credits itself.
1147                  *
1148                  * If need be, we could handle -EAGAIN for a
1149                  * RESTART_TRANS here.
1150                  */
1151                 mlog_bug_on_msg(ret == -EAGAIN,
1152                                 "Inode %llu: EAGAIN return during allocation.\n",
1153                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1154                 if (ret < 0) {
1155                         mlog_errno(ret);
1156                         goto out;
1157                 }
1158         } else if (clear_unwritten) {
1159                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1160                                               wc->w_di_bh);
1161                 ret = ocfs2_mark_extent_written(inode, &et,
1162                                                 wc->w_handle, cpos, 1, *phys,
1163                                                 meta_ac, &wc->w_dealloc);
1164                 if (ret < 0) {
1165                         mlog_errno(ret);
1166                         goto out;
1167                 }
1168         }
1169
1170         /*
1171          * The only reason this should fail is due to an inability to
1172          * find the extent added.
1173          */
1174         ret = ocfs2_get_clusters(inode, cpos, phys, NULL, NULL);
1175         if (ret < 0) {
1176                 mlog(ML_ERROR, "Get physical blkno failed for inode %llu, "
1177                             "at logical cluster %u",
1178                             (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
1179                 goto out;
1180         }
1181
1182         BUG_ON(*phys == 0);
1183
1184         p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *phys);
1185         if (!should_zero)
1186                 p_blkno += (user_pos >> inode->i_sb->s_blocksize_bits) & (u64)(bpc - 1);
1187
1188         for(i = 0; i < wc->w_num_pages; i++) {
1189                 int tmpret;
1190
1191                 /* This is the direct io target page. */
1192                 if (wc->w_pages[i] == NULL) {
1193                         p_blkno++;
1194                         continue;
1195                 }
1196
1197                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1198                                                       wc->w_pages[i], cpos,
1199                                                       user_pos, user_len,
1200                                                       should_zero);
1201                 if (tmpret) {
1202                         mlog_errno(tmpret);
1203                         if (ret == 0)
1204                                 ret = tmpret;
1205                 }
1206         }
1207
1208         /*
1209          * We only have cleanup to do in case of allocating write.
1210          */
1211         if (ret && new)
1212                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1213
1214 out:
1215
1216         return ret;
1217 }
1218
1219 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1220                                        struct ocfs2_alloc_context *data_ac,
1221                                        struct ocfs2_alloc_context *meta_ac,
1222                                        struct ocfs2_write_ctxt *wc,
1223                                        loff_t pos, unsigned len)
1224 {
1225         int ret, i;
1226         loff_t cluster_off;
1227         unsigned int local_len = len;
1228         struct ocfs2_write_cluster_desc *desc;
1229         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1230
1231         for (i = 0; i < wc->w_clen; i++) {
1232                 desc = &wc->w_desc[i];
1233
1234                 /*
1235                  * We have to make sure that the total write passed in
1236                  * doesn't extend past a single cluster.
1237                  */
1238                 local_len = len;
1239                 cluster_off = pos & (osb->s_clustersize - 1);
1240                 if ((cluster_off + local_len) > osb->s_clustersize)
1241                         local_len = osb->s_clustersize - cluster_off;
1242
1243                 ret = ocfs2_write_cluster(mapping, &desc->c_phys,
1244                                           desc->c_new,
1245                                           desc->c_clear_unwritten,
1246                                           desc->c_needs_zero,
1247                                           data_ac, meta_ac,
1248                                           wc, desc->c_cpos, pos, local_len);
1249                 if (ret) {
1250                         mlog_errno(ret);
1251                         goto out;
1252                 }
1253
1254                 len -= local_len;
1255                 pos += local_len;
1256         }
1257
1258         ret = 0;
1259 out:
1260         return ret;
1261 }
1262
1263 /*
1264  * ocfs2_write_end() wants to know which parts of the target page it
1265  * should complete the write on. It's easiest to compute them ahead of
1266  * time when a more complete view of the write is available.
1267  */
1268 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1269                                         struct ocfs2_write_ctxt *wc,
1270                                         loff_t pos, unsigned len, int alloc)
1271 {
1272         struct ocfs2_write_cluster_desc *desc;
1273
1274         wc->w_target_from = pos & (PAGE_SIZE - 1);
1275         wc->w_target_to = wc->w_target_from + len;
1276
1277         if (alloc == 0)
1278                 return;
1279
1280         /*
1281          * Allocating write - we may have different boundaries based
1282          * on page size and cluster size.
1283          *
1284          * NOTE: We can no longer compute one value from the other as
1285          * the actual write length and user provided length may be
1286          * different.
1287          */
1288
1289         if (wc->w_large_pages) {
1290                 /*
1291                  * We only care about the 1st and last cluster within
1292                  * our range and whether they should be zero'd or not. Either
1293                  * value may be extended out to the start/end of a
1294                  * newly allocated cluster.
1295                  */
1296                 desc = &wc->w_desc[0];
1297                 if (desc->c_needs_zero)
1298                         ocfs2_figure_cluster_boundaries(osb,
1299                                                         desc->c_cpos,
1300                                                         &wc->w_target_from,
1301                                                         NULL);
1302
1303                 desc = &wc->w_desc[wc->w_clen - 1];
1304                 if (desc->c_needs_zero)
1305                         ocfs2_figure_cluster_boundaries(osb,
1306                                                         desc->c_cpos,
1307                                                         NULL,
1308                                                         &wc->w_target_to);
1309         } else {
1310                 wc->w_target_from = 0;
1311                 wc->w_target_to = PAGE_SIZE;
1312         }
1313 }
1314
1315 /*
1316  * Check if this extent is marked UNWRITTEN by direct io. If so, we need not to
1317  * do the zero work. And should not to clear UNWRITTEN since it will be cleared
1318  * by the direct io procedure.
1319  * If this is a new extent that allocated by direct io, we should mark it in
1320  * the ip_unwritten_list.
1321  */
1322 static int ocfs2_unwritten_check(struct inode *inode,
1323                                  struct ocfs2_write_ctxt *wc,
1324                                  struct ocfs2_write_cluster_desc *desc)
1325 {
1326         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1327         struct ocfs2_unwritten_extent *ue = NULL, *new = NULL;
1328         int ret = 0;
1329
1330         if (!desc->c_needs_zero)
1331                 return 0;
1332
1333 retry:
1334         spin_lock(&oi->ip_lock);
1335         /* Needs not to zero no metter buffer or direct. The one who is zero
1336          * the cluster is doing zero. And he will clear unwritten after all
1337          * cluster io finished. */
1338         list_for_each_entry(ue, &oi->ip_unwritten_list, ue_ip_node) {
1339                 if (desc->c_cpos == ue->ue_cpos) {
1340                         BUG_ON(desc->c_new);
1341                         desc->c_needs_zero = 0;
1342                         desc->c_clear_unwritten = 0;
1343                         goto unlock;
1344                 }
1345         }
1346
1347         if (wc->w_type != OCFS2_WRITE_DIRECT)
1348                 goto unlock;
1349
1350         if (new == NULL) {
1351                 spin_unlock(&oi->ip_lock);
1352                 new = kmalloc(sizeof(struct ocfs2_unwritten_extent),
1353                              GFP_NOFS);
1354                 if (new == NULL) {
1355                         ret = -ENOMEM;
1356                         goto out;
1357                 }
1358                 goto retry;
1359         }
1360         /* This direct write will doing zero. */
1361         new->ue_cpos = desc->c_cpos;
1362         new->ue_phys = desc->c_phys;
1363         desc->c_clear_unwritten = 0;
1364         list_add_tail(&new->ue_ip_node, &oi->ip_unwritten_list);
1365         list_add_tail(&new->ue_node, &wc->w_unwritten_list);
1366         wc->w_unwritten_count++;
1367         new = NULL;
1368 unlock:
1369         spin_unlock(&oi->ip_lock);
1370 out:
1371         kfree(new);
1372         return ret;
1373 }
1374
1375 /*
1376  * Populate each single-cluster write descriptor in the write context
1377  * with information about the i/o to be done.
1378  *
1379  * Returns the number of clusters that will have to be allocated, as
1380  * well as a worst case estimate of the number of extent records that
1381  * would have to be created during a write to an unwritten region.
1382  */
1383 static int ocfs2_populate_write_desc(struct inode *inode,
1384                                      struct ocfs2_write_ctxt *wc,
1385                                      unsigned int *clusters_to_alloc,
1386                                      unsigned int *extents_to_split)
1387 {
1388         int ret;
1389         struct ocfs2_write_cluster_desc *desc;
1390         unsigned int num_clusters = 0;
1391         unsigned int ext_flags = 0;
1392         u32 phys = 0;
1393         int i;
1394
1395         *clusters_to_alloc = 0;
1396         *extents_to_split = 0;
1397
1398         for (i = 0; i < wc->w_clen; i++) {
1399                 desc = &wc->w_desc[i];
1400                 desc->c_cpos = wc->w_cpos + i;
1401
1402                 if (num_clusters == 0) {
1403                         /*
1404                          * Need to look up the next extent record.
1405                          */
1406                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1407                                                  &num_clusters, &ext_flags);
1408                         if (ret) {
1409                                 mlog_errno(ret);
1410                                 goto out;
1411                         }
1412
1413                         /* We should already CoW the refcountd extent. */
1414                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1415
1416                         /*
1417                          * Assume worst case - that we're writing in
1418                          * the middle of the extent.
1419                          *
1420                          * We can assume that the write proceeds from
1421                          * left to right, in which case the extent
1422                          * insert code is smart enough to coalesce the
1423                          * next splits into the previous records created.
1424                          */
1425                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1426                                 *extents_to_split = *extents_to_split + 2;
1427                 } else if (phys) {
1428                         /*
1429                          * Only increment phys if it doesn't describe
1430                          * a hole.
1431                          */
1432                         phys++;
1433                 }
1434
1435                 /*
1436                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1437                  * file that got extended.  w_first_new_cpos tells us
1438                  * where the newly allocated clusters are so we can
1439                  * zero them.
1440                  */
1441                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1442                         BUG_ON(phys == 0);
1443                         desc->c_needs_zero = 1;
1444                 }
1445
1446                 desc->c_phys = phys;
1447                 if (phys == 0) {
1448                         desc->c_new = 1;
1449                         desc->c_needs_zero = 1;
1450                         desc->c_clear_unwritten = 1;
1451                         *clusters_to_alloc = *clusters_to_alloc + 1;
1452                 }
1453
1454                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1455                         desc->c_clear_unwritten = 1;
1456                         desc->c_needs_zero = 1;
1457                 }
1458
1459                 ret = ocfs2_unwritten_check(inode, wc, desc);
1460                 if (ret) {
1461                         mlog_errno(ret);
1462                         goto out;
1463                 }
1464
1465                 num_clusters--;
1466         }
1467
1468         ret = 0;
1469 out:
1470         return ret;
1471 }
1472
1473 static int ocfs2_write_begin_inline(struct address_space *mapping,
1474                                     struct inode *inode,
1475                                     struct ocfs2_write_ctxt *wc)
1476 {
1477         int ret;
1478         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1479         struct page *page;
1480         handle_t *handle;
1481         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1482
1483         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1484         if (IS_ERR(handle)) {
1485                 ret = PTR_ERR(handle);
1486                 mlog_errno(ret);
1487                 goto out;
1488         }
1489
1490         page = find_or_create_page(mapping, 0, GFP_NOFS);
1491         if (!page) {
1492                 ocfs2_commit_trans(osb, handle);
1493                 ret = -ENOMEM;
1494                 mlog_errno(ret);
1495                 goto out;
1496         }
1497         /*
1498          * If we don't set w_num_pages then this page won't get unlocked
1499          * and freed on cleanup of the write context.
1500          */
1501         wc->w_pages[0] = wc->w_target_page = page;
1502         wc->w_num_pages = 1;
1503
1504         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1505                                       OCFS2_JOURNAL_ACCESS_WRITE);
1506         if (ret) {
1507                 ocfs2_commit_trans(osb, handle);
1508
1509                 mlog_errno(ret);
1510                 goto out;
1511         }
1512
1513         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1514                 ocfs2_set_inode_data_inline(inode, di);
1515
1516         if (!PageUptodate(page)) {
1517                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1518                 if (ret) {
1519                         ocfs2_commit_trans(osb, handle);
1520
1521                         goto out;
1522                 }
1523         }
1524
1525         wc->w_handle = handle;
1526 out:
1527         return ret;
1528 }
1529
1530 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1531 {
1532         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1533
1534         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1535                 return 1;
1536         return 0;
1537 }
1538
1539 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1540                                           struct inode *inode, loff_t pos,
1541                                           unsigned len, struct page *mmap_page,
1542                                           struct ocfs2_write_ctxt *wc)
1543 {
1544         int ret, written = 0;
1545         loff_t end = pos + len;
1546         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1547         struct ocfs2_dinode *di = NULL;
1548
1549         trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1550                                              len, (unsigned long long)pos,
1551                                              oi->ip_dyn_features);
1552
1553         /*
1554          * Handle inodes which already have inline data 1st.
1555          */
1556         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1557                 if (mmap_page == NULL &&
1558                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1559                         goto do_inline_write;
1560
1561                 /*
1562                  * The write won't fit - we have to give this inode an
1563                  * inline extent list now.
1564                  */
1565                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1566                 if (ret)
1567                         mlog_errno(ret);
1568                 goto out;
1569         }
1570
1571         /*
1572          * Check whether the inode can accept inline data.
1573          */
1574         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1575                 return 0;
1576
1577         /*
1578          * Check whether the write can fit.
1579          */
1580         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1581         if (mmap_page ||
1582             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1583                 return 0;
1584
1585 do_inline_write:
1586         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1587         if (ret) {
1588                 mlog_errno(ret);
1589                 goto out;
1590         }
1591
1592         /*
1593          * This signals to the caller that the data can be written
1594          * inline.
1595          */
1596         written = 1;
1597 out:
1598         return written ? written : ret;
1599 }
1600
1601 /*
1602  * This function only does anything for file systems which can't
1603  * handle sparse files.
1604  *
1605  * What we want to do here is fill in any hole between the current end
1606  * of allocation and the end of our write. That way the rest of the
1607  * write path can treat it as an non-allocating write, which has no
1608  * special case code for sparse/nonsparse files.
1609  */
1610 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1611                                         struct buffer_head *di_bh,
1612                                         loff_t pos, unsigned len,
1613                                         struct ocfs2_write_ctxt *wc)
1614 {
1615         int ret;
1616         loff_t newsize = pos + len;
1617
1618         BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1619
1620         if (newsize <= i_size_read(inode))
1621                 return 0;
1622
1623         ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1624         if (ret)
1625                 mlog_errno(ret);
1626
1627         /* There is no wc if this is call from direct. */
1628         if (wc)
1629                 wc->w_first_new_cpos =
1630                         ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1631
1632         return ret;
1633 }
1634
1635 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1636                            loff_t pos)
1637 {
1638         int ret = 0;
1639
1640         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1641         if (pos > i_size_read(inode))
1642                 ret = ocfs2_zero_extend(inode, di_bh, pos);
1643
1644         return ret;
1645 }
1646
1647 int ocfs2_write_begin_nolock(struct address_space *mapping,
1648                              loff_t pos, unsigned len, ocfs2_write_type_t type,
1649                              struct page **pagep, void **fsdata,
1650                              struct buffer_head *di_bh, struct page *mmap_page)
1651 {
1652         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1653         unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1654         struct ocfs2_write_ctxt *wc;
1655         struct inode *inode = mapping->host;
1656         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1657         struct ocfs2_dinode *di;
1658         struct ocfs2_alloc_context *data_ac = NULL;
1659         struct ocfs2_alloc_context *meta_ac = NULL;
1660         handle_t *handle;
1661         struct ocfs2_extent_tree et;
1662         int try_free = 1, ret1;
1663
1664 try_again:
1665         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, type, di_bh);
1666         if (ret) {
1667                 mlog_errno(ret);
1668                 return ret;
1669         }
1670
1671         if (ocfs2_supports_inline_data(osb)) {
1672                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1673                                                      mmap_page, wc);
1674                 if (ret == 1) {
1675                         ret = 0;
1676                         goto success;
1677                 }
1678                 if (ret < 0) {
1679                         mlog_errno(ret);
1680                         goto out;
1681                 }
1682         }
1683
1684         /* Direct io change i_size late, should not zero tail here. */
1685         if (type != OCFS2_WRITE_DIRECT) {
1686                 if (ocfs2_sparse_alloc(osb))
1687                         ret = ocfs2_zero_tail(inode, di_bh, pos);
1688                 else
1689                         ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
1690                                                            len, wc);
1691                 if (ret) {
1692                         mlog_errno(ret);
1693                         goto out;
1694                 }
1695         }
1696
1697         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1698         if (ret < 0) {
1699                 mlog_errno(ret);
1700                 goto out;
1701         } else if (ret == 1) {
1702                 clusters_need = wc->w_clen;
1703                 ret = ocfs2_refcount_cow(inode, di_bh,
1704                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1705                 if (ret) {
1706                         mlog_errno(ret);
1707                         goto out;
1708                 }
1709         }
1710
1711         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1712                                         &extents_to_split);
1713         if (ret) {
1714                 mlog_errno(ret);
1715                 goto out;
1716         }
1717         clusters_need += clusters_to_alloc;
1718
1719         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1720
1721         trace_ocfs2_write_begin_nolock(
1722                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1723                         (long long)i_size_read(inode),
1724                         le32_to_cpu(di->i_clusters),
1725                         pos, len, type, mmap_page,
1726                         clusters_to_alloc, extents_to_split);
1727
1728         /*
1729          * We set w_target_from, w_target_to here so that
1730          * ocfs2_write_end() knows which range in the target page to
1731          * write out. An allocation requires that we write the entire
1732          * cluster range.
1733          */
1734         if (clusters_to_alloc || extents_to_split) {
1735                 /*
1736                  * XXX: We are stretching the limits of
1737                  * ocfs2_lock_allocators(). It greatly over-estimates
1738                  * the work to be done.
1739                  */
1740                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1741                                               wc->w_di_bh);
1742                 ret = ocfs2_lock_allocators(inode, &et,
1743                                             clusters_to_alloc, extents_to_split,
1744                                             &data_ac, &meta_ac);
1745                 if (ret) {
1746                         mlog_errno(ret);
1747                         goto out;
1748                 }
1749
1750                 if (data_ac)
1751                         data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1752
1753                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1754                                                     &di->id2.i_list);
1755         } else if (type == OCFS2_WRITE_DIRECT)
1756                 /* direct write needs not to start trans if no extents alloc. */
1757                 goto success;
1758
1759         /*
1760          * We have to zero sparse allocated clusters, unwritten extent clusters,
1761          * and non-sparse clusters we just extended.  For non-sparse writes,
1762          * we know zeros will only be needed in the first and/or last cluster.
1763          */
1764         if (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1765                            wc->w_desc[wc->w_clen - 1].c_needs_zero))
1766                 cluster_of_pages = 1;
1767         else
1768                 cluster_of_pages = 0;
1769
1770         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1771
1772         handle = ocfs2_start_trans(osb, credits);
1773         if (IS_ERR(handle)) {
1774                 ret = PTR_ERR(handle);
1775                 mlog_errno(ret);
1776                 goto out;
1777         }
1778
1779         wc->w_handle = handle;
1780
1781         if (clusters_to_alloc) {
1782                 ret = dquot_alloc_space_nodirty(inode,
1783                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1784                 if (ret)
1785                         goto out_commit;
1786         }
1787
1788         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1789                                       OCFS2_JOURNAL_ACCESS_WRITE);
1790         if (ret) {
1791                 mlog_errno(ret);
1792                 goto out_quota;
1793         }
1794
1795         /*
1796          * Fill our page array first. That way we've grabbed enough so
1797          * that we can zero and flush if we error after adding the
1798          * extent.
1799          */
1800         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1801                                          cluster_of_pages, mmap_page);
1802         if (ret) {
1803                 /*
1804                  * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1805                  * the target page. In this case, we exit with no error and no target
1806                  * page. This will trigger the caller, page_mkwrite(), to re-try
1807                  * the operation.
1808                  */
1809                 if (type == OCFS2_WRITE_MMAP && ret == -EAGAIN) {
1810                         BUG_ON(wc->w_target_page);
1811                         ret = 0;
1812                         goto out_quota;
1813                 }
1814
1815                 mlog_errno(ret);
1816                 goto out_quota;
1817         }
1818
1819         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1820                                           len);
1821         if (ret) {
1822                 mlog_errno(ret);
1823                 goto out_quota;
1824         }
1825
1826         if (data_ac)
1827                 ocfs2_free_alloc_context(data_ac);
1828         if (meta_ac)
1829                 ocfs2_free_alloc_context(meta_ac);
1830
1831 success:
1832         if (pagep)
1833                 *pagep = wc->w_target_page;
1834         *fsdata = wc;
1835         return 0;
1836 out_quota:
1837         if (clusters_to_alloc)
1838                 dquot_free_space(inode,
1839                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1840 out_commit:
1841         ocfs2_commit_trans(osb, handle);
1842
1843 out:
1844         /*
1845          * The mmapped page won't be unlocked in ocfs2_free_write_ctxt(),
1846          * even in case of error here like ENOSPC and ENOMEM. So, we need
1847          * to unlock the target page manually to prevent deadlocks when
1848          * retrying again on ENOSPC, or when returning non-VM_FAULT_LOCKED
1849          * to VM code.
1850          */
1851         if (wc->w_target_locked)
1852                 unlock_page(mmap_page);
1853
1854         ocfs2_free_write_ctxt(inode, wc);
1855
1856         if (data_ac) {
1857                 ocfs2_free_alloc_context(data_ac);
1858                 data_ac = NULL;
1859         }
1860         if (meta_ac) {
1861                 ocfs2_free_alloc_context(meta_ac);
1862                 meta_ac = NULL;
1863         }
1864
1865         if (ret == -ENOSPC && try_free) {
1866                 /*
1867                  * Try to free some truncate log so that we can have enough
1868                  * clusters to allocate.
1869                  */
1870                 try_free = 0;
1871
1872                 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1873                 if (ret1 == 1)
1874                         goto try_again;
1875
1876                 if (ret1 < 0)
1877                         mlog_errno(ret1);
1878         }
1879
1880         return ret;
1881 }
1882
1883 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1884                              loff_t pos, unsigned len, unsigned flags,
1885                              struct page **pagep, void **fsdata)
1886 {
1887         int ret;
1888         struct buffer_head *di_bh = NULL;
1889         struct inode *inode = mapping->host;
1890
1891         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1892         if (ret) {
1893                 mlog_errno(ret);
1894                 return ret;
1895         }
1896
1897         /*
1898          * Take alloc sem here to prevent concurrent lookups. That way
1899          * the mapping, zeroing and tree manipulation within
1900          * ocfs2_write() will be safe against ->readpage(). This
1901          * should also serve to lock out allocation from a shared
1902          * writeable region.
1903          */
1904         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1905
1906         ret = ocfs2_write_begin_nolock(mapping, pos, len, OCFS2_WRITE_BUFFER,
1907                                        pagep, fsdata, di_bh, NULL);
1908         if (ret) {
1909                 mlog_errno(ret);
1910                 goto out_fail;
1911         }
1912
1913         brelse(di_bh);
1914
1915         return 0;
1916
1917 out_fail:
1918         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1919
1920         brelse(di_bh);
1921         ocfs2_inode_unlock(inode, 1);
1922
1923         return ret;
1924 }
1925
1926 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1927                                    unsigned len, unsigned *copied,
1928                                    struct ocfs2_dinode *di,
1929                                    struct ocfs2_write_ctxt *wc)
1930 {
1931         void *kaddr;
1932
1933         if (unlikely(*copied < len)) {
1934                 if (!PageUptodate(wc->w_target_page)) {
1935                         *copied = 0;
1936                         return;
1937                 }
1938         }
1939
1940         kaddr = kmap_atomic(wc->w_target_page);
1941         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1942         kunmap_atomic(kaddr);
1943
1944         trace_ocfs2_write_end_inline(
1945              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1946              (unsigned long long)pos, *copied,
1947              le16_to_cpu(di->id2.i_data.id_count),
1948              le16_to_cpu(di->i_dyn_features));
1949 }
1950
1951 int ocfs2_write_end_nolock(struct address_space *mapping,
1952                            loff_t pos, unsigned len, unsigned copied, void *fsdata)
1953 {
1954         int i, ret;
1955         unsigned from, to, start = pos & (PAGE_SIZE - 1);
1956         struct inode *inode = mapping->host;
1957         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1958         struct ocfs2_write_ctxt *wc = fsdata;
1959         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1960         handle_t *handle = wc->w_handle;
1961         struct page *tmppage;
1962
1963         BUG_ON(!list_empty(&wc->w_unwritten_list));
1964
1965         if (handle) {
1966                 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
1967                                 wc->w_di_bh, OCFS2_JOURNAL_ACCESS_WRITE);
1968                 if (ret) {
1969                         copied = ret;
1970                         mlog_errno(ret);
1971                         goto out;
1972                 }
1973         }
1974
1975         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1976                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1977                 goto out_write_size;
1978         }
1979
1980         if (unlikely(copied < len) && wc->w_target_page) {
1981                 if (!PageUptodate(wc->w_target_page))
1982                         copied = 0;
1983
1984                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1985                                        start+len);
1986         }
1987         if (wc->w_target_page)
1988                 flush_dcache_page(wc->w_target_page);
1989
1990         for(i = 0; i < wc->w_num_pages; i++) {
1991                 tmppage = wc->w_pages[i];
1992
1993                 /* This is the direct io target page. */
1994                 if (tmppage == NULL)
1995                         continue;
1996
1997                 if (tmppage == wc->w_target_page) {
1998                         from = wc->w_target_from;
1999                         to = wc->w_target_to;
2000
2001                         BUG_ON(from > PAGE_SIZE ||
2002                                to > PAGE_SIZE ||
2003                                to < from);
2004                 } else {
2005                         /*
2006                          * Pages adjacent to the target (if any) imply
2007                          * a hole-filling write in which case we want
2008                          * to flush their entire range.
2009                          */
2010                         from = 0;
2011                         to = PAGE_SIZE;
2012                 }
2013
2014                 if (page_has_buffers(tmppage)) {
2015                         if (handle && ocfs2_should_order_data(inode)) {
2016                                 loff_t start_byte =
2017                                         ((loff_t)tmppage->index << PAGE_SHIFT) +
2018                                         from;
2019                                 loff_t length = to - from;
2020                                 ocfs2_jbd2_inode_add_write(handle, inode,
2021                                                            start_byte, length);
2022                         }
2023                         block_commit_write(tmppage, from, to);
2024                 }
2025         }
2026
2027 out_write_size:
2028         /* Direct io do not update i_size here. */
2029         if (wc->w_type != OCFS2_WRITE_DIRECT) {
2030                 pos += copied;
2031                 if (pos > i_size_read(inode)) {
2032                         i_size_write(inode, pos);
2033                         mark_inode_dirty(inode);
2034                 }
2035                 inode->i_blocks = ocfs2_inode_sector_count(inode);
2036                 di->i_size = cpu_to_le64((u64)i_size_read(inode));
2037                 inode->i_mtime = inode->i_ctime = current_time(inode);
2038                 di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2039                 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2040                 if (handle)
2041                         ocfs2_update_inode_fsync_trans(handle, inode, 1);
2042         }
2043         if (handle)
2044                 ocfs2_journal_dirty(handle, wc->w_di_bh);
2045
2046 out:
2047         /* unlock pages before dealloc since it needs acquiring j_trans_barrier
2048          * lock, or it will cause a deadlock since journal commit threads holds
2049          * this lock and will ask for the page lock when flushing the data.
2050          * put it here to preserve the unlock order.
2051          */
2052         ocfs2_unlock_pages(wc);
2053
2054         if (handle)
2055                 ocfs2_commit_trans(osb, handle);
2056
2057         ocfs2_run_deallocs(osb, &wc->w_dealloc);
2058
2059         brelse(wc->w_di_bh);
2060         kfree(wc);
2061
2062         return copied;
2063 }
2064
2065 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2066                            loff_t pos, unsigned len, unsigned copied,
2067                            struct page *page, void *fsdata)
2068 {
2069         int ret;
2070         struct inode *inode = mapping->host;
2071
2072         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, fsdata);
2073
2074         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2075         ocfs2_inode_unlock(inode, 1);
2076
2077         return ret;
2078 }
2079
2080 struct ocfs2_dio_write_ctxt {
2081         struct list_head        dw_zero_list;
2082         unsigned                dw_zero_count;
2083         int                     dw_orphaned;
2084         pid_t                   dw_writer_pid;
2085 };
2086
2087 static struct ocfs2_dio_write_ctxt *
2088 ocfs2_dio_alloc_write_ctx(struct buffer_head *bh, int *alloc)
2089 {
2090         struct ocfs2_dio_write_ctxt *dwc = NULL;
2091
2092         if (bh->b_private)
2093                 return bh->b_private;
2094
2095         dwc = kmalloc(sizeof(struct ocfs2_dio_write_ctxt), GFP_NOFS);
2096         if (dwc == NULL)
2097                 return NULL;
2098         INIT_LIST_HEAD(&dwc->dw_zero_list);
2099         dwc->dw_zero_count = 0;
2100         dwc->dw_orphaned = 0;
2101         dwc->dw_writer_pid = task_pid_nr(current);
2102         bh->b_private = dwc;
2103         *alloc = 1;
2104
2105         return dwc;
2106 }
2107
2108 static void ocfs2_dio_free_write_ctx(struct inode *inode,
2109                                      struct ocfs2_dio_write_ctxt *dwc)
2110 {
2111         ocfs2_free_unwritten_list(inode, &dwc->dw_zero_list);
2112         kfree(dwc);
2113 }
2114
2115 /*
2116  * TODO: Make this into a generic get_blocks function.
2117  *
2118  * From do_direct_io in direct-io.c:
2119  *  "So what we do is to permit the ->get_blocks function to populate
2120  *   bh.b_size with the size of IO which is permitted at this offset and
2121  *   this i_blkbits."
2122  *
2123  * This function is called directly from get_more_blocks in direct-io.c.
2124  *
2125  * called like this: dio->get_blocks(dio->inode, fs_startblk,
2126  *                                      fs_count, map_bh, dio->rw == WRITE);
2127  */
2128 static int ocfs2_dio_wr_get_block(struct inode *inode, sector_t iblock,
2129                                struct buffer_head *bh_result, int create)
2130 {
2131         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2132         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2133         struct ocfs2_write_ctxt *wc;
2134         struct ocfs2_write_cluster_desc *desc = NULL;
2135         struct ocfs2_dio_write_ctxt *dwc = NULL;
2136         struct buffer_head *di_bh = NULL;
2137         u64 p_blkno;
2138         unsigned int i_blkbits = inode->i_sb->s_blocksize_bits;
2139         loff_t pos = iblock << i_blkbits;
2140         sector_t endblk = (i_size_read(inode) - 1) >> i_blkbits;
2141         unsigned len, total_len = bh_result->b_size;
2142         int ret = 0, first_get_block = 0;
2143
2144         len = osb->s_clustersize - (pos & (osb->s_clustersize - 1));
2145         len = min(total_len, len);
2146
2147         /*
2148          * bh_result->b_size is count in get_more_blocks according to write
2149          * "pos" and "end", we need map twice to return different buffer state:
2150          * 1. area in file size, not set NEW;
2151          * 2. area out file size, set  NEW.
2152          *
2153          *                 iblock    endblk
2154          * |--------|---------|---------|---------
2155          * |<-------area in file------->|
2156          */
2157
2158         if ((iblock <= endblk) &&
2159             ((iblock + ((len - 1) >> i_blkbits)) > endblk))
2160                 len = (endblk - iblock + 1) << i_blkbits;
2161
2162         mlog(0, "get block of %lu at %llu:%u req %u\n",
2163                         inode->i_ino, pos, len, total_len);
2164
2165         /*
2166          * Because we need to change file size in ocfs2_dio_end_io_write(), or
2167          * we may need to add it to orphan dir. So can not fall to fast path
2168          * while file size will be changed.
2169          */
2170         if (pos + total_len <= i_size_read(inode)) {
2171
2172                 /* This is the fast path for re-write. */
2173                 ret = ocfs2_lock_get_block(inode, iblock, bh_result, create);
2174                 if (buffer_mapped(bh_result) &&
2175                     !buffer_new(bh_result) &&
2176                     ret == 0)
2177                         goto out;
2178
2179                 /* Clear state set by ocfs2_get_block. */
2180                 bh_result->b_state = 0;
2181         }
2182
2183         dwc = ocfs2_dio_alloc_write_ctx(bh_result, &first_get_block);
2184         if (unlikely(dwc == NULL)) {
2185                 ret = -ENOMEM;
2186                 mlog_errno(ret);
2187                 goto out;
2188         }
2189
2190         if (ocfs2_clusters_for_bytes(inode->i_sb, pos + total_len) >
2191             ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode)) &&
2192             !dwc->dw_orphaned) {
2193                 /*
2194                  * when we are going to alloc extents beyond file size, add the
2195                  * inode to orphan dir, so we can recall those spaces when
2196                  * system crashed during write.
2197                  */
2198                 ret = ocfs2_add_inode_to_orphan(osb, inode);
2199                 if (ret < 0) {
2200                         mlog_errno(ret);
2201                         goto out;
2202                 }
2203                 dwc->dw_orphaned = 1;
2204         }
2205
2206         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2207         if (ret) {
2208                 mlog_errno(ret);
2209                 goto out;
2210         }
2211
2212         down_write(&oi->ip_alloc_sem);
2213
2214         if (first_get_block) {
2215                 if (ocfs2_sparse_alloc(osb))
2216                         ret = ocfs2_zero_tail(inode, di_bh, pos);
2217                 else
2218                         ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
2219                                                            total_len, NULL);
2220                 if (ret < 0) {
2221                         mlog_errno(ret);
2222                         goto unlock;
2223                 }
2224         }
2225
2226         ret = ocfs2_write_begin_nolock(inode->i_mapping, pos, len,
2227                                        OCFS2_WRITE_DIRECT, NULL,
2228                                        (void **)&wc, di_bh, NULL);
2229         if (ret) {
2230                 mlog_errno(ret);
2231                 goto unlock;
2232         }
2233
2234         desc = &wc->w_desc[0];
2235
2236         p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, desc->c_phys);
2237         BUG_ON(p_blkno == 0);
2238         p_blkno += iblock & (u64)(ocfs2_clusters_to_blocks(inode->i_sb, 1) - 1);
2239
2240         map_bh(bh_result, inode->i_sb, p_blkno);
2241         bh_result->b_size = len;
2242         if (desc->c_needs_zero)
2243                 set_buffer_new(bh_result);
2244
2245         if (iblock > endblk)
2246                 set_buffer_new(bh_result);
2247
2248         /* May sleep in end_io. It should not happen in a irq context. So defer
2249          * it to dio work queue. */
2250         set_buffer_defer_completion(bh_result);
2251
2252         if (!list_empty(&wc->w_unwritten_list)) {
2253                 struct ocfs2_unwritten_extent *ue = NULL;
2254
2255                 ue = list_first_entry(&wc->w_unwritten_list,
2256                                       struct ocfs2_unwritten_extent,
2257                                       ue_node);
2258                 BUG_ON(ue->ue_cpos != desc->c_cpos);
2259                 /* The physical address may be 0, fill it. */
2260                 ue->ue_phys = desc->c_phys;
2261
2262                 list_splice_tail_init(&wc->w_unwritten_list, &dwc->dw_zero_list);
2263                 dwc->dw_zero_count += wc->w_unwritten_count;
2264         }
2265
2266         ret = ocfs2_write_end_nolock(inode->i_mapping, pos, len, len, wc);
2267         BUG_ON(ret != len);
2268         ret = 0;
2269 unlock:
2270         up_write(&oi->ip_alloc_sem);
2271         ocfs2_inode_unlock(inode, 1);
2272         brelse(di_bh);
2273 out:
2274         if (ret < 0)
2275                 ret = -EIO;
2276         return ret;
2277 }
2278
2279 static int ocfs2_dio_end_io_write(struct inode *inode,
2280                                   struct ocfs2_dio_write_ctxt *dwc,
2281                                   loff_t offset,
2282                                   ssize_t bytes)
2283 {
2284         struct ocfs2_cached_dealloc_ctxt dealloc;
2285         struct ocfs2_extent_tree et;
2286         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2287         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2288         struct ocfs2_unwritten_extent *ue = NULL;
2289         struct buffer_head *di_bh = NULL;
2290         struct ocfs2_dinode *di;
2291         struct ocfs2_alloc_context *data_ac = NULL;
2292         struct ocfs2_alloc_context *meta_ac = NULL;
2293         handle_t *handle = NULL;
2294         loff_t end = offset + bytes;
2295         int ret = 0, credits = 0;
2296
2297         ocfs2_init_dealloc_ctxt(&dealloc);
2298
2299         /* We do clear unwritten, delete orphan, change i_size here. If neither
2300          * of these happen, we can skip all this. */
2301         if (list_empty(&dwc->dw_zero_list) &&
2302             end <= i_size_read(inode) &&
2303             !dwc->dw_orphaned)
2304                 goto out;
2305
2306         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2307         if (ret < 0) {
2308                 mlog_errno(ret);
2309                 goto out;
2310         }
2311
2312         down_write(&oi->ip_alloc_sem);
2313
2314         /* Delete orphan before acquire i_rwsem. */
2315         if (dwc->dw_orphaned) {
2316                 BUG_ON(dwc->dw_writer_pid != task_pid_nr(current));
2317
2318                 end = end > i_size_read(inode) ? end : 0;
2319
2320                 ret = ocfs2_del_inode_from_orphan(osb, inode, di_bh,
2321                                 !!end, end);
2322                 if (ret < 0)
2323                         mlog_errno(ret);
2324         }
2325
2326         di = (struct ocfs2_dinode *)di_bh->b_data;
2327
2328         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
2329
2330         /* Attach dealloc with extent tree in case that we may reuse extents
2331          * which are already unlinked from current extent tree due to extent
2332          * rotation and merging.
2333          */
2334         et.et_dealloc = &dealloc;
2335
2336         ret = ocfs2_lock_allocators(inode, &et, 0, dwc->dw_zero_count*2,
2337                                     &data_ac, &meta_ac);
2338         if (ret) {
2339                 mlog_errno(ret);
2340                 goto unlock;
2341         }
2342
2343         credits = ocfs2_calc_extend_credits(inode->i_sb, &di->id2.i_list);
2344
2345         handle = ocfs2_start_trans(osb, credits);
2346         if (IS_ERR(handle)) {
2347                 ret = PTR_ERR(handle);
2348                 mlog_errno(ret);
2349                 goto unlock;
2350         }
2351         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2352                                       OCFS2_JOURNAL_ACCESS_WRITE);
2353         if (ret) {
2354                 mlog_errno(ret);
2355                 goto commit;
2356         }
2357
2358         list_for_each_entry(ue, &dwc->dw_zero_list, ue_node) {
2359                 ret = ocfs2_mark_extent_written(inode, &et, handle,
2360                                                 ue->ue_cpos, 1,
2361                                                 ue->ue_phys,
2362                                                 meta_ac, &dealloc);
2363                 if (ret < 0) {
2364                         mlog_errno(ret);
2365                         break;
2366                 }
2367         }
2368
2369         if (end > i_size_read(inode)) {
2370                 ret = ocfs2_set_inode_size(handle, inode, di_bh, end);
2371                 if (ret < 0)
2372                         mlog_errno(ret);
2373         }
2374 commit:
2375         ocfs2_commit_trans(osb, handle);
2376 unlock:
2377         up_write(&oi->ip_alloc_sem);
2378         ocfs2_inode_unlock(inode, 1);
2379         brelse(di_bh);
2380 out:
2381         if (data_ac)
2382                 ocfs2_free_alloc_context(data_ac);
2383         if (meta_ac)
2384                 ocfs2_free_alloc_context(meta_ac);
2385         ocfs2_run_deallocs(osb, &dealloc);
2386         ocfs2_dio_free_write_ctx(inode, dwc);
2387
2388         return ret;
2389 }
2390
2391 /*
2392  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
2393  * particularly interested in the aio/dio case.  We use the rw_lock DLM lock
2394  * to protect io on one node from truncation on another.
2395  */
2396 static int ocfs2_dio_end_io(struct kiocb *iocb,
2397                             loff_t offset,
2398                             ssize_t bytes,
2399                             void *private)
2400 {
2401         struct inode *inode = file_inode(iocb->ki_filp);
2402         int level;
2403         int ret = 0;
2404
2405         /* this io's submitter should not have unlocked this before we could */
2406         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
2407
2408         if (bytes <= 0)
2409                 mlog_ratelimited(ML_ERROR, "Direct IO failed, bytes = %lld",
2410                                  (long long)bytes);
2411         if (private) {
2412                 if (bytes > 0)
2413                         ret = ocfs2_dio_end_io_write(inode, private, offset,
2414                                                      bytes);
2415                 else
2416                         ocfs2_dio_free_write_ctx(inode, private);
2417         }
2418
2419         ocfs2_iocb_clear_rw_locked(iocb);
2420
2421         level = ocfs2_iocb_rw_locked_level(iocb);
2422         ocfs2_rw_unlock(inode, level);
2423         return ret;
2424 }
2425
2426 static ssize_t ocfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2427 {
2428         struct file *file = iocb->ki_filp;
2429         struct inode *inode = file->f_mapping->host;
2430         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2431         get_block_t *get_block;
2432
2433         /*
2434          * Fallback to buffered I/O if we see an inode without
2435          * extents.
2436          */
2437         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2438                 return 0;
2439
2440         /* Fallback to buffered I/O if we do not support append dio. */
2441         if (iocb->ki_pos + iter->count > i_size_read(inode) &&
2442             !ocfs2_supports_append_dio(osb))
2443                 return 0;
2444
2445         if (iov_iter_rw(iter) == READ)
2446                 get_block = ocfs2_lock_get_block;
2447         else
2448                 get_block = ocfs2_dio_wr_get_block;
2449
2450         return __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
2451                                     iter, get_block,
2452                                     ocfs2_dio_end_io, NULL, 0);
2453 }
2454
2455 const struct address_space_operations ocfs2_aops = {
2456         .dirty_folio            = block_dirty_folio,
2457         .readpage               = ocfs2_readpage,
2458         .readahead              = ocfs2_readahead,
2459         .writepage              = ocfs2_writepage,
2460         .write_begin            = ocfs2_write_begin,
2461         .write_end              = ocfs2_write_end,
2462         .bmap                   = ocfs2_bmap,
2463         .direct_IO              = ocfs2_direct_IO,
2464         .invalidate_folio       = block_invalidate_folio,
2465         .releasepage            = ocfs2_releasepage,
2466         .migratepage            = buffer_migrate_page,
2467         .is_partially_uptodate  = block_is_partially_uptodate,
2468         .error_remove_page      = generic_error_remove_page,
2469 };