fs/ocfs2/aops.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* -*- mode: c; c-basic-offset: 8; -*-
   3  * vim: noexpandtab sw=8 ts=8 sts=0:
   4  *
   5  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   6  */
   7
   8 #include <linux/fs.h>
   9 #include <linux/slab.h>
  10 #include <linux/highmem.h>
  11 #include <linux/pagemap.h>
  12 #include <asm/byteorder.h>
  13 #include <linux/swap.h>
  14 #include <linux/pipe_fs_i.h>
  15 #include <linux/mpage.h>
  16 #include <linux/quotaops.h>
  17 #include <linux/blkdev.h>
  18 #include <linux/uio.h>
  19 #include <linux/mm.h>
  20
  21 #include <cluster/masklog.h>
  22
  23 #include "ocfs2.h"
  24
  25 #include "alloc.h"
  26 #include "aops.h"
  27 #include "dlmglue.h"
  28 #include "extent_map.h"
  29 #include "file.h"
  30 #include "inode.h"
  31 #include "journal.h"
  32 #include "suballoc.h"
  33 #include "super.h"
  34 #include "symlink.h"
  35 #include "refcounttree.h"
  36 #include "ocfs2_trace.h"
  37
  38 #include "buffer_head_io.h"
  39 #include "dir.h"
  40 #include "namei.h"
  41 #include "sysfile.h"
  42
  43 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
  44                                    struct buffer_head *bh_result, int create)
  45 {
  46         int err = -EIO;
  47         int status;
  48         struct ocfs2_dinode *fe = NULL;
  49         struct buffer_head *bh = NULL;
  50         struct buffer_head *buffer_cache_bh = NULL;
  51         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  52         void *kaddr;
  53
  54         trace_ocfs2_symlink_get_block(
  55                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
  56                         (unsigned long long)iblock, bh_result, create);
  57
  58         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
  59
  60         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
  61                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
  62                      (unsigned long long)iblock);
  63                 goto bail;
  64         }
  65
  66         status = ocfs2_read_inode_block(inode, &bh);
  67         if (status < 0) {
  68                 mlog_errno(status);
  69                 goto bail;
  70         }
  71         fe = (struct ocfs2_dinode *) bh->b_data;
  72
  73         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
  74                                                     le32_to_cpu(fe->i_clusters))) {
  75                 err = -ENOMEM;
  76                 mlog(ML_ERROR, "block offset is outside the allocated size: "
  77                      "%llu\n", (unsigned long long)iblock);
  78                 goto bail;
  79         }
  80
  81         /* We don't use the page cache to create symlink data, so if
  82          * need be, copy it over from the buffer cache. */
  83         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  84                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  85                             iblock;
  86                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
  87                 if (!buffer_cache_bh) {
  88                         err = -ENOMEM;
  89                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
  90                         goto bail;
  91                 }
  92
  93                 /* we haven't locked out transactions, so a commit
  94                  * could've happened. Since we've got a reference on
  95                  * the bh, even if it commits while we're doing the
  96                  * copy, the data is still good. */
  97                 if (buffer_jbd(buffer_cache_bh)
  98                     && ocfs2_inode_is_new(inode)) {
  99                         kaddr = kmap_atomic(bh_result->b_page);
 100                         if (!kaddr) {
 101                                 mlog(ML_ERROR, "couldn't kmap!\n");
 102                                 goto bail;
 103                         }
 104                         memcpy(kaddr + (bh_result->b_size * iblock),
 105                                buffer_cache_bh->b_data,
 106                                bh_result->b_size);
 107                         kunmap_atomic(kaddr);
 108                         set_buffer_uptodate(bh_result);
 109                 }
 110                 brelse(buffer_cache_bh);
 111         }
 112
 113         map_bh(bh_result, inode->i_sb,
 114                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 115
 116         err = 0;
 117
 118 bail:
 119         brelse(bh);
 120
 121         return err;
 122 }
 123
 124 static int ocfs2_lock_get_block(struct inode *inode, sector_t iblock,
 125                     struct buffer_head *bh_result, int create)
 126 {
 127         int ret = 0;
 128         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 129
 130         down_read(&oi->ip_alloc_sem);
 131         ret = ocfs2_get_block(inode, iblock, bh_result, create);
 132         up_read(&oi->ip_alloc_sem);
 133
 134         return ret;
 135 }
 136
 137 int ocfs2_get_block(struct inode *inode, sector_t iblock,
 138                     struct buffer_head *bh_result, int create)
 139 {
 140         int err = 0;
 141         unsigned int ext_flags;
 142         u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
 143         u64 p_blkno, count, past_eof;
 144         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 145
 146         trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno,
 147                               (unsigned long long)iblock, bh_result, create);
 148
 149         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
 150                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
 151                      inode, inode->i_ino);
 152
 153         if (S_ISLNK(inode->i_mode)) {
 154                 /* this always does I/O for some reason. */
 155                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
 156                 goto bail;
 157         }
 158
 159         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
 160                                           &ext_flags);
 161         if (err) {
 162                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
 163                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
 164                      (unsigned long long)p_blkno);
 165                 goto bail;
 166         }
 167
 168         if (max_blocks < count)
 169                 count = max_blocks;
 170
 171         /*
 172          * ocfs2 never allocates in this function - the only time we
 173          * need to use BH_New is when we're extending i_size on a file
 174          * system which doesn't support holes, in which case BH_New
 175          * allows __block_write_begin() to zero.
 176          *
 177          * If we see this on a sparse file system, then a truncate has
 178          * raced us and removed the cluster. In this case, we clear
 179          * the buffers dirty and uptodate bits and let the buffer code
 180          * ignore it as a hole.
 181          */
 182         if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
 183                 clear_buffer_dirty(bh_result);
 184                 clear_buffer_uptodate(bh_result);
 185                 goto bail;
 186         }
 187
 188         /* Treat the unwritten extent as a hole for zeroing purposes. */
 189         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 190                 map_bh(bh_result, inode->i_sb, p_blkno);
 191
 192         bh_result->b_size = count << inode->i_blkbits;
 193
 194         if (!ocfs2_sparse_alloc(osb)) {
 195                 if (p_blkno == 0) {
 196                         err = -EIO;
 197                         mlog(ML_ERROR,
 198                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
 199                              (unsigned long long)iblock,
 200                              (unsigned long long)p_blkno,
 201                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
 202                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
 203                         dump_stack();
 204                         goto bail;
 205                 }
 206         }
 207
 208         past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 209
 210         trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno,
 211                                   (unsigned long long)past_eof);
 212         if (create && (iblock >= past_eof))
 213                 set_buffer_new(bh_result);
 214
 215 bail:
 216         if (err < 0)
 217                 err = -EIO;
 218
 219         return err;
 220 }
 221
 222 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
 223                            struct buffer_head *di_bh)
 224 {
 225         void *kaddr;
 226         loff_t size;
 227         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 228
 229         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
 230                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag\n",
 231                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
 232                 return -EROFS;
 233         }
 234
 235         size = i_size_read(inode);
 236
 237         if (size > PAGE_SIZE ||
 238             size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
 239                 ocfs2_error(inode->i_sb,
 240                             "Inode %llu has with inline data has bad size: %Lu\n",
 241                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 242                             (unsigned long long)size);
 243                 return -EROFS;
 244         }
 245
 246         kaddr = kmap_atomic(page);
 247         if (size)
 248                 memcpy(kaddr, di->id2.i_data.id_data, size);
 249         /* Clear the remaining part of the page */
 250         memset(kaddr + size, 0, PAGE_SIZE - size);
 251         flush_dcache_page(page);
 252         kunmap_atomic(kaddr);
 253
 254         SetPageUptodate(page);
 255
 256         return 0;
 257 }
 258
 259 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
 260 {
 261         int ret;
 262         struct buffer_head *di_bh = NULL;
 263
 264         BUG_ON(!PageLocked(page));
 265         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
 266
 267         ret = ocfs2_read_inode_block(inode, &di_bh);
 268         if (ret) {
 269                 mlog_errno(ret);
 270                 goto out;
 271         }
 272
 273         ret = ocfs2_read_inline_data(inode, page, di_bh);
 274 out:
 275         unlock_page(page);
 276
 277         brelse(di_bh);
 278         return ret;
 279 }
 280
 281 static int ocfs2_readpage(struct file *file, struct page *page)
 282 {
 283         struct inode *inode = page->mapping->host;
 284         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 285         loff_t start = (loff_t)page->index << PAGE_SHIFT;
 286         int ret, unlock = 1;
 287
 288         trace_ocfs2_readpage((unsigned long long)oi->ip_blkno,
 289                              (page ? page->index : 0));
 290
 291         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
 292         if (ret != 0) {
 293                 if (ret == AOP_TRUNCATED_PAGE)
 294                         unlock = 0;
 295                 mlog_errno(ret);
 296                 goto out;
 297         }
 298
 299         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 300                 /*
 301                  * Unlock the page and cycle ip_alloc_sem so that we don't
 302                  * busyloop waiting for ip_alloc_sem to unlock
 303                  */
 304                 ret = AOP_TRUNCATED_PAGE;
 305                 unlock_page(page);
 306                 unlock = 0;
 307                 down_read(&oi->ip_alloc_sem);
 308                 up_read(&oi->ip_alloc_sem);
 309                 goto out_inode_unlock;
 310         }
 311
 312         /*
 313          * i_size might have just been updated as we grabed the meta lock.  We
 314          * might now be discovering a truncate that hit on another node.
 315          * block_read_full_page->get_block freaks out if it is asked to read
 316          * beyond the end of a file, so we check here.  Callers
 317          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
 318          * and notice that the page they just read isn't needed.
 319          *
 320          * XXX sys_readahead() seems to get that wrong?
 321          */
 322         if (start >= i_size_read(inode)) {
 323                 zero_user(page, 0, PAGE_SIZE);
 324                 SetPageUptodate(page);
 325                 ret = 0;
 326                 goto out_alloc;
 327         }
 328
 329         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 330                 ret = ocfs2_readpage_inline(inode, page);
 331         else
 332                 ret = block_read_full_page(page, ocfs2_get_block);
 333         unlock = 0;
 334
 335 out_alloc:
 336         up_read(&oi->ip_alloc_sem);
 337 out_inode_unlock:
 338         ocfs2_inode_unlock(inode, 0);
 339 out:
 340         if (unlock)
 341                 unlock_page(page);
 342         return ret;
 343 }
 344
 345 /*
 346  * This is used only for read-ahead. Failures or difficult to handle
 347  * situations are safe to ignore.
 348  *
 349  * Right now, we don't bother with BH_Boundary - in-inode extent lists
 350  * are quite large (243 extents on 4k blocks), so most inodes don't
 351  * grow out to a tree. If need be, detecting boundary extents could
 352  * trivially be added in a future version of ocfs2_get_block().
 353  */
 354 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
 355                            struct list_head *pages, unsigned nr_pages)
 356 {
 357         int ret, err = -EIO;
 358         struct inode *inode = mapping->host;
 359         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 360         loff_t start;
 361         struct page *last;
 362
 363         /*
 364          * Use the nonblocking flag for the dlm code to avoid page
 365          * lock inversion, but don't bother with retrying.
 366          */
 367         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
 368         if (ret)
 369                 return err;
 370
 371         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 372                 ocfs2_inode_unlock(inode, 0);
 373                 return err;
 374         }
 375
 376         /*
 377          * Don't bother with inline-data. There isn't anything
 378          * to read-ahead in that case anyway...
 379          */
 380         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 381                 goto out_unlock;
 382
 383         /*
 384          * Check whether a remote node truncated this file - we just
 385          * drop out in that case as it's not worth handling here.
 386          */
 387         last = lru_to_page(pages);
 388         start = (loff_t)last->index << PAGE_SHIFT;
 389         if (start >= i_size_read(inode))
 390                 goto out_unlock;
 391
 392         err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
 393
 394 out_unlock:
 395         up_read(&oi->ip_alloc_sem);
 396         ocfs2_inode_unlock(inode, 0);
 397
 398         return err;
 399 }
 400
 401 /* Note: Because we don't support holes, our allocation has
 402  * already happened (allocation writes zeros to the file data)
 403  * so we don't have to worry about ordered writes in
 404  * ocfs2_writepage.
 405  *
 406  * ->writepage is called during the process of invalidating the page cache
 407  * during blocked lock processing.  It can't block on any cluster locks
 408  * to during block mapping.  It's relying on the fact that the block
 409  * mapping can't have disappeared under the dirty pages that it is
 410  * being asked to write back.
 411  */
 412 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
 413 {
 414         trace_ocfs2_writepage(
 415                 (unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno,
 416                 page->index);
 417
 418         return block_write_full_page(page, ocfs2_get_block, wbc);
 419 }
 420
 421 /* Taken from ext3. We don't necessarily need the full blown
 422  * functionality yet, but IMHO it's better to cut and paste the whole
 423  * thing so we can avoid introducing our own bugs (and easily pick up
 424  * their fixes when they happen) --Mark */
 425 int walk_page_buffers(  handle_t *handle,
 426                         struct buffer_head *head,
 427                         unsigned from,
 428                         unsigned to,
 429                         int *partial,
 430                         int (*fn)(      handle_t *handle,
 431                                         struct buffer_head *bh))
 432 {
 433         struct buffer_head *bh;
 434         unsigned block_start, block_end;
 435         unsigned blocksize = head->b_size;
 436         int err, ret = 0;
 437         struct buffer_head *next;
 438
 439         for (   bh = head, block_start = 0;
 440                 ret == 0 && (bh != head || !block_start);
 441                 block_start = block_end, bh = next)
 442         {
 443                 next = bh->b_this_page;
 444                 block_end = block_start + blocksize;
 445                 if (block_end <= from || block_start >= to) {
 446                         if (partial && !buffer_uptodate(bh))
 447                                 *partial = 1;
 448                         continue;
 449                 }
 450                 err = (*fn)(handle, bh);
 451                 if (!ret)
 452                         ret = err;
 453         }
 454         return ret;
 455 }
 456
 457 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 458 {
 459         sector_t status;
 460         u64 p_blkno = 0;
 461         int err = 0;
 462         struct inode *inode = mapping->host;
 463
 464         trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
 465                          (unsigned long long)block);
 466
 467         /*
 468          * The swap code (ab-)uses ->bmap to get a block mapping and then
 469          * bypasseѕ the file system for actual I/O.  We really can't allow
 470          * that on refcounted inodes, so we have to skip out here.  And yes,
 471          * 0 is the magic code for a bmap error..
 472          */
 473         if (ocfs2_is_refcount_inode(inode))
 474                 return 0;
 475
 476         /* We don't need to lock journal system files, since they aren't
 477          * accessed concurrently from multiple nodes.
 478          */
 479         if (!INODE_JOURNAL(inode)) {
 480                 err = ocfs2_inode_lock(inode, NULL, 0);
 481                 if (err) {
 482                         if (err != -ENOENT)
 483                                 mlog_errno(err);
 484                         goto bail;
 485                 }
 486                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
 487         }
 488
 489         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
 490                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
 491                                                   NULL);
 492
 493         if (!INODE_JOURNAL(inode)) {
 494                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
 495                 ocfs2_inode_unlock(inode, 0);
 496         }
 497
 498         if (err) {
 499                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 500                      (unsigned long long)block);
 501                 mlog_errno(err);
 502                 goto bail;
 503         }
 504
 505 bail:
 506         status = err ? 0 : p_blkno;
 507
 508         return status;
 509 }
 510
 511 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 512 {
 513         if (!page_has_buffers(page))
 514                 return 0;
 515         return try_to_free_buffers(page);
 516 }
 517
 518 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 519                                             u32 cpos,
 520                                             unsigned int *start,
 521                                             unsigned int *end)
 522 {
 523         unsigned int cluster_start = 0, cluster_end = PAGE_SIZE;
 524
 525         if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits)) {
 526                 unsigned int cpp;
 527
 528                 cpp = 1 << (PAGE_SHIFT - osb->s_clustersize_bits);
 529
 530                 cluster_start = cpos % cpp;
 531                 cluster_start = cluster_start << osb->s_clustersize_bits;
 532
 533                 cluster_end = cluster_start + osb->s_clustersize;
 534         }
 535
 536         BUG_ON(cluster_start > PAGE_SIZE);
 537         BUG_ON(cluster_end > PAGE_SIZE);
 538
 539         if (start)
 540                 *start = cluster_start;
 541         if (end)
 542                 *end = cluster_end;
 543 }
 544
 545 /*
 546  * 'from' and 'to' are the region in the page to avoid zeroing.
 547  *
 548  * If pagesize > clustersize, this function will avoid zeroing outside
 549  * of the cluster boundary.
 550  *
 551  * from == to == 0 is code for "zero the entire cluster region"
 552  */
 553 static void ocfs2_clear_page_regions(struct page *page,
 554                                      struct ocfs2_super *osb, u32 cpos,
 555                                      unsigned from, unsigned to)
 556 {
 557         void *kaddr;
 558         unsigned int cluster_start, cluster_end;
 559
 560         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 561
 562         kaddr = kmap_atomic(page);
 563
 564         if (from || to) {
 565                 if (from > cluster_start)
 566                         memset(kaddr + cluster_start, 0, from - cluster_start);
 567                 if (to < cluster_end)
 568                         memset(kaddr + to, 0, cluster_end - to);
 569         } else {
 570                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 571         }
 572
 573         kunmap_atomic(kaddr);
 574 }
 575
 576 /*
 577  * Nonsparse file systems fully allocate before we get to the write
 578  * code. This prevents ocfs2_write() from tagging the write as an
 579  * allocating one, which means ocfs2_map_page_blocks() might try to
 580  * read-in the blocks at the tail of our file. Avoid reading them by
 581  * testing i_size against each block offset.
 582  */
 583 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 584                                  unsigned int block_start)
 585 {
 586         u64 offset = page_offset(page) + block_start;
 587
 588         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 589                 return 1;
 590
 591         if (i_size_read(inode) > offset)
 592                 return 1;
 593
 594         return 0;
 595 }
 596
 597 /*
 598  * Some of this taken from __block_write_begin(). We already have our
 599  * mapping by now though, and the entire write will be allocating or
 600  * it won't, so not much need to use BH_New.
 601  *
 602  * This will also skip zeroing, which is handled externally.
 603  */
 604 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 605                           struct inode *inode, unsigned int from,
 606                           unsigned int to, int new)
 607 {
 608         int ret = 0;
 609         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 610         unsigned int block_end, block_start;
 611         unsigned int bsize = i_blocksize(inode);
 612
 613         if (!page_has_buffers(page))
 614                 create_empty_buffers(page, bsize, 0);
 615
 616         head = page_buffers(page);
 617         for (bh = head, block_start = 0; bh != head || !block_start;
 618              bh = bh->b_this_page, block_start += bsize) {
 619                 block_end = block_start + bsize;
 620
 621                 clear_buffer_new(bh);
 622
 623                 /*
 624                  * Ignore blocks outside of our i/o range -
 625                  * they may belong to unallocated clusters.
 626                  */
 627                 if (block_start >= to || block_end <= from) {
 628                         if (PageUptodate(page))
 629                                 set_buffer_uptodate(bh);
 630                         continue;
 631                 }
 632
 633                 /*
 634                  * For an allocating write with cluster size >= page
 635                  * size, we always write the entire page.
 636                  */
 637                 if (new)
 638                         set_buffer_new(bh);
 639
 640                 if (!buffer_mapped(bh)) {
 641                         map_bh(bh, inode->i_sb, *p_blkno);
 642                         clean_bdev_bh_alias(bh);
 643                 }
 644
 645                 if (PageUptodate(page)) {
 646                         if (!buffer_uptodate(bh))
 647                                 set_buffer_uptodate(bh);
 648                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 649                            !buffer_new(bh) &&
 650                            ocfs2_should_read_blk(inode, page, block_start) &&
 651                            (block_start < from || block_end > to)) {
 652                         ll_rw_block(REQ_OP_READ, 0, 1, &bh);
 653                         *wait_bh++=bh;
 654                 }
 655
 656                 *p_blkno = *p_blkno + 1;
 657         }
 658
 659         /*
 660          * If we issued read requests - let them complete.
 661          */
 662         while(wait_bh > wait) {
 663                 wait_on_buffer(*--wait_bh);
 664                 if (!buffer_uptodate(*wait_bh))
 665                         ret = -EIO;
 666         }
 667
 668         if (ret == 0 || !new)
 669                 return ret;
 670
 671         /*
 672          * If we get -EIO above, zero out any newly allocated blocks
 673          * to avoid exposing stale data.
 674          */
 675         bh = head;
 676         block_start = 0;
 677         do {
 678                 block_end = block_start + bsize;
 679                 if (block_end <= from)
 680                         goto next_bh;
 681                 if (block_start >= to)
 682                         break;
 683
 684                 zero_user(page, block_start, bh->b_size);
 685                 set_buffer_uptodate(bh);
 686                 mark_buffer_dirty(bh);
 687
 688 next_bh:
 689                 block_start = block_end;
 690                 bh = bh->b_this_page;
 691         } while (bh != head);
 692
 693         return ret;
 694 }
 695
 696 #if (PAGE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 697 #define OCFS2_MAX_CTXT_PAGES    1
 698 #else
 699 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_SIZE)
 700 #endif
 701
 702 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 703
 704 struct ocfs2_unwritten_extent {
 705         struct list_head        ue_node;
 706         struct list_head        ue_ip_node;
 707         u32                     ue_cpos;
 708         u32                     ue_phys;
 709 };
 710
 711 /*
 712  * Describe the state of a single cluster to be written to.
 713  */
 714 struct ocfs2_write_cluster_desc {
 715         u32             c_cpos;
 716         u32             c_phys;
 717         /*
 718          * Give this a unique field because c_phys eventually gets
 719          * filled.
 720          */
 721         unsigned        c_new;
 722         unsigned        c_clear_unwritten;
 723         unsigned        c_needs_zero;
 724 };
 725
 726 struct ocfs2_write_ctxt {
 727         /* Logical cluster position / len of write */
 728         u32                             w_cpos;
 729         u32                             w_clen;
 730
 731         /* First cluster allocated in a nonsparse extend */
 732         u32                             w_first_new_cpos;
 733
 734         /* Type of caller. Must be one of buffer, mmap, direct.  */
 735         ocfs2_write_type_t              w_type;
 736
 737         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 738
 739         /*
 740          * This is true if page_size > cluster_size.
 741          *
 742          * It triggers a set of special cases during write which might
 743          * have to deal with allocating writes to partial pages.
 744          */
 745         unsigned int                    w_large_pages;
 746
 747         /*
 748          * Pages involved in this write.
 749          *
 750          * w_target_page is the page being written to by the user.
 751          *
 752          * w_pages is an array of pages which always contains
 753          * w_target_page, and in the case of an allocating write with
 754          * page_size < cluster size, it will contain zero'd and mapped
 755          * pages adjacent to w_target_page which need to be written
 756          * out in so that future reads from that region will get
 757          * zero's.
 758          */
 759         unsigned int                    w_num_pages;
 760         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 761         struct page                     *w_target_page;
 762
 763         /*
 764          * w_target_locked is used for page_mkwrite path indicating no unlocking
 765          * against w_target_page in ocfs2_write_end_nolock.
 766          */
 767         unsigned int                    w_target_locked:1;
 768
 769         /*
 770          * ocfs2_write_end() uses this to know what the real range to
 771          * write in the target should be.
 772          */
 773         unsigned int                    w_target_from;
 774         unsigned int                    w_target_to;
 775
 776         /*
 777          * We could use journal_current_handle() but this is cleaner,
 778          * IMHO -Mark
 779          */
 780         handle_t                        *w_handle;
 781
 782         struct buffer_head              *w_di_bh;
 783
 784         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 785
 786         struct list_head                w_unwritten_list;
 787         unsigned int                    w_unwritten_count;
 788 };
 789
 790 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 791 {
 792         int i;
 793
 794         for(i = 0; i < num_pages; i++) {
 795                 if (pages[i]) {
 796                         unlock_page(pages[i]);
 797                         mark_page_accessed(pages[i]);
 798                         put_page(pages[i]);
 799                 }
 800         }
 801 }
 802
 803 static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc)
 804 {
 805         int i;
 806
 807         /*
 808          * w_target_locked is only set to true in the page_mkwrite() case.
 809          * The intent is to allow us to lock the target page from write_begin()
 810          * to write_end(). The caller must hold a ref on w_target_page.
 811          */
 812         if (wc->w_target_locked) {
 813                 BUG_ON(!wc->w_target_page);
 814                 for (i = 0; i < wc->w_num_pages; i++) {
 815                         if (wc->w_target_page == wc->w_pages[i]) {
 816                                 wc->w_pages[i] = NULL;
 817                                 break;
 818                         }
 819                 }
 820                 mark_page_accessed(wc->w_target_page);
 821                 put_page(wc->w_target_page);
 822         }
 823         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 824 }
 825
 826 static void ocfs2_free_unwritten_list(struct inode *inode,
 827                                  struct list_head *head)
 828 {
 829         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 830         struct ocfs2_unwritten_extent *ue = NULL, *tmp = NULL;
 831
 832         list_for_each_entry_safe(ue, tmp, head, ue_node) {
 833                 list_del(&ue->ue_node);
 834                 spin_lock(&oi->ip_lock);
 835                 list_del(&ue->ue_ip_node);
 836                 spin_unlock(&oi->ip_lock);
 837                 kfree(ue);
 838         }
 839 }
 840
 841 static void ocfs2_free_write_ctxt(struct inode *inode,
 842                                   struct ocfs2_write_ctxt *wc)
 843 {
 844         ocfs2_free_unwritten_list(inode, &wc->w_unwritten_list);
 845         ocfs2_unlock_pages(wc);
 846         brelse(wc->w_di_bh);
 847         kfree(wc);
 848 }
 849
 850 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 851                                   struct ocfs2_super *osb, loff_t pos,
 852                                   unsigned len, ocfs2_write_type_t type,
 853                                   struct buffer_head *di_bh)
 854 {
 855         u32 cend;
 856         struct ocfs2_write_ctxt *wc;
 857
 858         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 859         if (!wc)
 860                 return -ENOMEM;
 861
 862         wc->w_cpos = pos >> osb->s_clustersize_bits;
 863         wc->w_first_new_cpos = UINT_MAX;
 864         cend = (pos + len - 1) >> osb->s_clustersize_bits;
 865         wc->w_clen = cend - wc->w_cpos + 1;
 866         get_bh(di_bh);
 867         wc->w_di_bh = di_bh;
 868         wc->w_type = type;
 869
 870         if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits))
 871                 wc->w_large_pages = 1;
 872         else
 873                 wc->w_large_pages = 0;
 874
 875         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 876         INIT_LIST_HEAD(&wc->w_unwritten_list);
 877
 878         *wcp = wc;
 879
 880         return 0;
 881 }
 882
 883 /*
 884  * If a page has any new buffers, zero them out here, and mark them uptodate
 885  * and dirty so they'll be written out (in order to prevent uninitialised
 886  * block data from leaking). And clear the new bit.
 887  */
 888 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 889 {
 890         unsigned int block_start, block_end;
 891         struct buffer_head *head, *bh;
 892
 893         BUG_ON(!PageLocked(page));
 894         if (!page_has_buffers(page))
 895                 return;
 896
 897         bh = head = page_buffers(page);
 898         block_start = 0;
 899         do {
 900                 block_end = block_start + bh->b_size;
 901
 902                 if (buffer_new(bh)) {
 903                         if (block_end > from && block_start < to) {
 904                                 if (!PageUptodate(page)) {
 905                                         unsigned start, end;
 906
 907                                         start = max(from, block_start);
 908                                         end = min(to, block_end);
 909
 910                                         zero_user_segment(page, start, end);
 911                                         set_buffer_uptodate(bh);
 912                                 }
 913
 914                                 clear_buffer_new(bh);
 915                                 mark_buffer_dirty(bh);
 916                         }
 917                 }
 918
 919                 block_start = block_end;
 920                 bh = bh->b_this_page;
 921         } while (bh != head);
 922 }
 923
 924 /*
 925  * Only called when we have a failure during allocating write to write
 926  * zero's to the newly allocated region.
 927  */
 928 static void ocfs2_write_failure(struct inode *inode,
 929                                 struct ocfs2_write_ctxt *wc,
 930                                 loff_t user_pos, unsigned user_len)
 931 {
 932         int i;
 933         unsigned from = user_pos & (PAGE_SIZE - 1),
 934                 to = user_pos + user_len;
 935         struct page *tmppage;
 936
 937         if (wc->w_target_page)
 938                 ocfs2_zero_new_buffers(wc->w_target_page, from, to);
 939
 940         for(i = 0; i < wc->w_num_pages; i++) {
 941                 tmppage = wc->w_pages[i];
 942
 943                 if (tmppage && page_has_buffers(tmppage)) {
 944                         if (ocfs2_should_order_data(inode))
 945                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
 946
 947                         block_commit_write(tmppage, from, to);
 948                 }
 949         }
 950 }
 951
 952 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
 953                                         struct ocfs2_write_ctxt *wc,
 954                                         struct page *page, u32 cpos,
 955                                         loff_t user_pos, unsigned user_len,
 956                                         int new)
 957 {
 958         int ret;
 959         unsigned int map_from = 0, map_to = 0;
 960         unsigned int cluster_start, cluster_end;
 961         unsigned int user_data_from = 0, user_data_to = 0;
 962
 963         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
 964                                         &cluster_start, &cluster_end);
 965
 966         /* treat the write as new if the a hole/lseek spanned across
 967          * the page boundary.
 968          */
 969         new = new | ((i_size_read(inode) <= page_offset(page)) &&
 970                         (page_offset(page) <= user_pos));
 971
 972         if (page == wc->w_target_page) {
 973                 map_from = user_pos & (PAGE_SIZE - 1);
 974                 map_to = map_from + user_len;
 975
 976                 if (new)
 977                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 978                                                     cluster_start, cluster_end,
 979                                                     new);
 980                 else
 981                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
 982                                                     map_from, map_to, new);
 983                 if (ret) {
 984                         mlog_errno(ret);
 985                         goto out;
 986                 }
 987
 988                 user_data_from = map_from;
 989                 user_data_to = map_to;
 990                 if (new) {
 991                         map_from = cluster_start;
 992                         map_to = cluster_end;
 993                 }
 994         } else {
 995                 /*
 996                  * If we haven't allocated the new page yet, we
 997                  * shouldn't be writing it out without copying user
 998                  * data. This is likely a math error from the caller.
 999                  */
1000                 BUG_ON(!new);
1001
1002                 map_from = cluster_start;
1003                 map_to = cluster_end;
1004
1005                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1006                                             cluster_start, cluster_end, new);
1007                 if (ret) {
1008                         mlog_errno(ret);
1009                         goto out;
1010                 }
1011         }
1012
1013         /*
1014          * Parts of newly allocated pages need to be zero'd.
1015          *
1016          * Above, we have also rewritten 'to' and 'from' - as far as
1017          * the rest of the function is concerned, the entire cluster
1018          * range inside of a page needs to be written.
1019          *
1020          * We can skip this if the page is up to date - it's already
1021          * been zero'd from being read in as a hole.
1022          */
1023         if (new && !PageUptodate(page))
1024                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1025                                          cpos, user_data_from, user_data_to);
1026
1027         flush_dcache_page(page);
1028
1029 out:
1030         return ret;
1031 }
1032
1033 /*
1034  * This function will only grab one clusters worth of pages.
1035  */
1036 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1037                                       struct ocfs2_write_ctxt *wc,
1038                                       u32 cpos, loff_t user_pos,
1039                                       unsigned user_len, int new,
1040                                       struct page *mmap_page)
1041 {
1042         int ret = 0, i;
1043         unsigned long start, target_index, end_index, index;
1044         struct inode *inode = mapping->host;
1045         loff_t last_byte;
1046
1047         target_index = user_pos >> PAGE_SHIFT;
1048
1049         /*
1050          * Figure out how many pages we'll be manipulating here. For
1051          * non allocating write, we just change the one
1052          * page. Otherwise, we'll need a whole clusters worth.  If we're
1053          * writing past i_size, we only need enough pages to cover the
1054          * last page of the write.
1055          */
1056         if (new) {
1057                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1058                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1059                 /*
1060                  * We need the index *past* the last page we could possibly
1061                  * touch.  This is the page past the end of the write or
1062                  * i_size, whichever is greater.
1063                  */
1064                 last_byte = max(user_pos + user_len, i_size_read(inode));
1065                 BUG_ON(last_byte < 1);
1066                 end_index = ((last_byte - 1) >> PAGE_SHIFT) + 1;
1067                 if ((start + wc->w_num_pages) > end_index)
1068                         wc->w_num_pages = end_index - start;
1069         } else {
1070                 wc->w_num_pages = 1;
1071                 start = target_index;
1072         }
1073         end_index = (user_pos + user_len - 1) >> PAGE_SHIFT;
1074
1075         for(i = 0; i < wc->w_num_pages; i++) {
1076                 index = start + i;
1077
1078                 if (index >= target_index && index <= end_index &&
1079                     wc->w_type == OCFS2_WRITE_MMAP) {
1080                         /*
1081                          * ocfs2_pagemkwrite() is a little different
1082                          * and wants us to directly use the page
1083                          * passed in.
1084                          */
1085                         lock_page(mmap_page);
1086
1087                         /* Exit and let the caller retry */
1088                         if (mmap_page->mapping != mapping) {
1089                                 WARN_ON(mmap_page->mapping);
1090                                 unlock_page(mmap_page);
1091                                 ret = -EAGAIN;
1092                                 goto out;
1093                         }
1094
1095                         get_page(mmap_page);
1096                         wc->w_pages[i] = mmap_page;
1097                         wc->w_target_locked = true;
1098                 } else if (index >= target_index && index <= end_index &&
1099                            wc->w_type == OCFS2_WRITE_DIRECT) {
1100                         /* Direct write has no mapping page. */
1101                         wc->w_pages[i] = NULL;
1102                         continue;
1103                 } else {
1104                         wc->w_pages[i] = find_or_create_page(mapping, index,
1105                                                              GFP_NOFS);
1106                         if (!wc->w_pages[i]) {
1107                                 ret = -ENOMEM;
1108                                 mlog_errno(ret);
1109                                 goto out;
1110                         }
1111                 }
1112                 wait_for_stable_page(wc->w_pages[i]);
1113
1114                 if (index == target_index)
1115                         wc->w_target_page = wc->w_pages[i];
1116         }
1117 out:
1118         if (ret)
1119                 wc->w_target_locked = false;
1120         return ret;
1121 }
1122
1123 /*
1124  * Prepare a single cluster for write one cluster into the file.
1125  */
1126 static int ocfs2_write_cluster(struct address_space *mapping,
1127                                u32 *phys, unsigned int new,
1128                                unsigned int clear_unwritten,
1129                                unsigned int should_zero,
1130                                struct ocfs2_alloc_context *data_ac,
1131                                struct ocfs2_alloc_context *meta_ac,
1132                                struct ocfs2_write_ctxt *wc, u32 cpos,
1133                                loff_t user_pos, unsigned user_len)
1134 {
1135         int ret, i;
1136         u64 p_blkno;
1137         struct inode *inode = mapping->host;
1138         struct ocfs2_extent_tree et;
1139         int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1);
1140
1141         if (new) {
1142                 u32 tmp_pos;
1143
1144                 /*
1145                  * This is safe to call with the page locks - it won't take
1146                  * any additional semaphores or cluster locks.
1147                  */
1148                 tmp_pos = cpos;
1149                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1150                                            &tmp_pos, 1, !clear_unwritten,
1151                                            wc->w_di_bh, wc->w_handle,
1152                                            data_ac, meta_ac, NULL);
1153                 /*
1154                  * This shouldn't happen because we must have already
1155                  * calculated the correct meta data allocation required. The
1156                  * internal tree allocation code should know how to increase
1157                  * transaction credits itself.
1158                  *
1159                  * If need be, we could handle -EAGAIN for a
1160                  * RESTART_TRANS here.
1161                  */
1162                 mlog_bug_on_msg(ret == -EAGAIN,
1163                                 "Inode %llu: EAGAIN return during allocation.\n",
1164                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1165                 if (ret < 0) {
1166                         mlog_errno(ret);
1167                         goto out;
1168                 }
1169         } else if (clear_unwritten) {
1170                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1171                                               wc->w_di_bh);
1172                 ret = ocfs2_mark_extent_written(inode, &et,
1173                                                 wc->w_handle, cpos, 1, *phys,
1174                                                 meta_ac, &wc->w_dealloc);
1175                 if (ret < 0) {
1176                         mlog_errno(ret);
1177                         goto out;
1178                 }
1179         }
1180
1181         /*
1182          * The only reason this should fail is due to an inability to
1183          * find the extent added.
1184          */
1185         ret = ocfs2_get_clusters(inode, cpos, phys, NULL, NULL);
1186         if (ret < 0) {
1187                 mlog(ML_ERROR, "Get physical blkno failed for inode %llu, "
1188                             "at logical cluster %u",
1189                             (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
1190                 goto out;
1191         }
1192
1193         BUG_ON(*phys == 0);
1194
1195         p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *phys);
1196         if (!should_zero)
1197                 p_blkno += (user_pos >> inode->i_sb->s_blocksize_bits) & (u64)(bpc - 1);
1198
1199         for(i = 0; i < wc->w_num_pages; i++) {
1200                 int tmpret;
1201
1202                 /* This is the direct io target page. */
1203                 if (wc->w_pages[i] == NULL) {
1204                         p_blkno++;
1205                         continue;
1206                 }
1207
1208                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1209                                                       wc->w_pages[i], cpos,
1210                                                       user_pos, user_len,
1211                                                       should_zero);
1212                 if (tmpret) {
1213                         mlog_errno(tmpret);
1214                         if (ret == 0)
1215                                 ret = tmpret;
1216                 }
1217         }
1218
1219         /*
1220          * We only have cleanup to do in case of allocating write.
1221          */
1222         if (ret && new)
1223                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1224
1225 out:
1226
1227         return ret;
1228 }
1229
1230 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1231                                        struct ocfs2_alloc_context *data_ac,
1232                                        struct ocfs2_alloc_context *meta_ac,
1233                                        struct ocfs2_write_ctxt *wc,
1234                                        loff_t pos, unsigned len)
1235 {
1236         int ret, i;
1237         loff_t cluster_off;
1238         unsigned int local_len = len;
1239         struct ocfs2_write_cluster_desc *desc;
1240         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1241
1242         for (i = 0; i < wc->w_clen; i++) {
1243                 desc = &wc->w_desc[i];
1244
1245                 /*
1246                  * We have to make sure that the total write passed in
1247                  * doesn't extend past a single cluster.
1248                  */
1249                 local_len = len;
1250                 cluster_off = pos & (osb->s_clustersize - 1);
1251                 if ((cluster_off + local_len) > osb->s_clustersize)
1252                         local_len = osb->s_clustersize - cluster_off;
1253
1254                 ret = ocfs2_write_cluster(mapping, &desc->c_phys,
1255                                           desc->c_new,
1256                                           desc->c_clear_unwritten,
1257                                           desc->c_needs_zero,
1258                                           data_ac, meta_ac,
1259                                           wc, desc->c_cpos, pos, local_len);
1260                 if (ret) {
1261                         mlog_errno(ret);
1262                         goto out;
1263                 }
1264
1265                 len -= local_len;
1266                 pos += local_len;
1267         }
1268
1269         ret = 0;
1270 out:
1271         return ret;
1272 }
1273
1274 /*
1275  * ocfs2_write_end() wants to know which parts of the target page it
1276  * should complete the write on. It's easiest to compute them ahead of
1277  * time when a more complete view of the write is available.
1278  */
1279 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1280                                         struct ocfs2_write_ctxt *wc,
1281                                         loff_t pos, unsigned len, int alloc)
1282 {
1283         struct ocfs2_write_cluster_desc *desc;
1284
1285         wc->w_target_from = pos & (PAGE_SIZE - 1);
1286         wc->w_target_to = wc->w_target_from + len;
1287
1288         if (alloc == 0)
1289                 return;
1290
1291         /*
1292          * Allocating write - we may have different boundaries based
1293          * on page size and cluster size.
1294          *
1295          * NOTE: We can no longer compute one value from the other as
1296          * the actual write length and user provided length may be
1297          * different.
1298          */
1299
1300         if (wc->w_large_pages) {
1301                 /*
1302                  * We only care about the 1st and last cluster within
1303                  * our range and whether they should be zero'd or not. Either
1304                  * value may be extended out to the start/end of a
1305                  * newly allocated cluster.
1306                  */
1307                 desc = &wc->w_desc[0];
1308                 if (desc->c_needs_zero)
1309                         ocfs2_figure_cluster_boundaries(osb,
1310                                                         desc->c_cpos,
1311                                                         &wc->w_target_from,
1312                                                         NULL);
1313
1314                 desc = &wc->w_desc[wc->w_clen - 1];
1315                 if (desc->c_needs_zero)
1316                         ocfs2_figure_cluster_boundaries(osb,
1317                                                         desc->c_cpos,
1318                                                         NULL,
1319                                                         &wc->w_target_to);
1320         } else {
1321                 wc->w_target_from = 0;
1322                 wc->w_target_to = PAGE_SIZE;
1323         }
1324 }
1325
1326 /*
1327  * Check if this extent is marked UNWRITTEN by direct io. If so, we need not to
1328  * do the zero work. And should not to clear UNWRITTEN since it will be cleared
1329  * by the direct io procedure.
1330  * If this is a new extent that allocated by direct io, we should mark it in
1331  * the ip_unwritten_list.
1332  */
1333 static int ocfs2_unwritten_check(struct inode *inode,
1334                                  struct ocfs2_write_ctxt *wc,
1335                                  struct ocfs2_write_cluster_desc *desc)
1336 {
1337         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1338         struct ocfs2_unwritten_extent *ue = NULL, *new = NULL;
1339         int ret = 0;
1340
1341         if (!desc->c_needs_zero)
1342                 return 0;
1343
1344 retry:
1345         spin_lock(&oi->ip_lock);
1346         /* Needs not to zero no metter buffer or direct. The one who is zero
1347          * the cluster is doing zero. And he will clear unwritten after all
1348          * cluster io finished. */
1349         list_for_each_entry(ue, &oi->ip_unwritten_list, ue_ip_node) {
1350                 if (desc->c_cpos == ue->ue_cpos) {
1351                         BUG_ON(desc->c_new);
1352                         desc->c_needs_zero = 0;
1353                         desc->c_clear_unwritten = 0;
1354                         goto unlock;
1355                 }
1356         }
1357
1358         if (wc->w_type != OCFS2_WRITE_DIRECT)
1359                 goto unlock;
1360
1361         if (new == NULL) {
1362                 spin_unlock(&oi->ip_lock);
1363                 new = kmalloc(sizeof(struct ocfs2_unwritten_extent),
1364                              GFP_NOFS);
1365                 if (new == NULL) {
1366                         ret = -ENOMEM;
1367                         goto out;
1368                 }
1369                 goto retry;
1370         }
1371         /* This direct write will doing zero. */
1372         new->ue_cpos = desc->c_cpos;
1373         new->ue_phys = desc->c_phys;
1374         desc->c_clear_unwritten = 0;
1375         list_add_tail(&new->ue_ip_node, &oi->ip_unwritten_list);
1376         list_add_tail(&new->ue_node, &wc->w_unwritten_list);
1377         wc->w_unwritten_count++;
1378         new = NULL;
1379 unlock:
1380         spin_unlock(&oi->ip_lock);
1381 out:
1382         kfree(new);
1383         return ret;
1384 }
1385
1386 /*
1387  * Populate each single-cluster write descriptor in the write context
1388  * with information about the i/o to be done.
1389  *
1390  * Returns the number of clusters that will have to be allocated, as
1391  * well as a worst case estimate of the number of extent records that
1392  * would have to be created during a write to an unwritten region.
1393  */
1394 static int ocfs2_populate_write_desc(struct inode *inode,
1395                                      struct ocfs2_write_ctxt *wc,
1396                                      unsigned int *clusters_to_alloc,
1397                                      unsigned int *extents_to_split)
1398 {
1399         int ret;
1400         struct ocfs2_write_cluster_desc *desc;
1401         unsigned int num_clusters = 0;
1402         unsigned int ext_flags = 0;
1403         u32 phys = 0;
1404         int i;
1405
1406         *clusters_to_alloc = 0;
1407         *extents_to_split = 0;
1408
1409         for (i = 0; i < wc->w_clen; i++) {
1410                 desc = &wc->w_desc[i];
1411                 desc->c_cpos = wc->w_cpos + i;
1412
1413                 if (num_clusters == 0) {
1414                         /*
1415                          * Need to look up the next extent record.
1416                          */
1417                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1418                                                  &num_clusters, &ext_flags);
1419                         if (ret) {
1420                                 mlog_errno(ret);
1421                                 goto out;
1422                         }
1423
1424                         /* We should already CoW the refcountd extent. */
1425                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1426
1427                         /*
1428                          * Assume worst case - that we're writing in
1429                          * the middle of the extent.
1430                          *
1431                          * We can assume that the write proceeds from
1432                          * left to right, in which case the extent
1433                          * insert code is smart enough to coalesce the
1434                          * next splits into the previous records created.
1435                          */
1436                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1437                                 *extents_to_split = *extents_to_split + 2;
1438                 } else if (phys) {
1439                         /*
1440                          * Only increment phys if it doesn't describe
1441                          * a hole.
1442                          */
1443                         phys++;
1444                 }
1445
1446                 /*
1447                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1448                  * file that got extended.  w_first_new_cpos tells us
1449                  * where the newly allocated clusters are so we can
1450                  * zero them.
1451                  */
1452                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1453                         BUG_ON(phys == 0);
1454                         desc->c_needs_zero = 1;
1455                 }
1456
1457                 desc->c_phys = phys;
1458                 if (phys == 0) {
1459                         desc->c_new = 1;
1460                         desc->c_needs_zero = 1;
1461                         desc->c_clear_unwritten = 1;
1462                         *clusters_to_alloc = *clusters_to_alloc + 1;
1463                 }
1464
1465                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1466                         desc->c_clear_unwritten = 1;
1467                         desc->c_needs_zero = 1;
1468                 }
1469
1470                 ret = ocfs2_unwritten_check(inode, wc, desc);
1471                 if (ret) {
1472                         mlog_errno(ret);
1473                         goto out;
1474                 }
1475
1476                 num_clusters--;
1477         }
1478
1479         ret = 0;
1480 out:
1481         return ret;
1482 }
1483
1484 static int ocfs2_write_begin_inline(struct address_space *mapping,
1485                                     struct inode *inode,
1486                                     struct ocfs2_write_ctxt *wc)
1487 {
1488         int ret;
1489         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1490         struct page *page;
1491         handle_t *handle;
1492         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1493
1494         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1495         if (IS_ERR(handle)) {
1496                 ret = PTR_ERR(handle);
1497                 mlog_errno(ret);
1498                 goto out;
1499         }
1500
1501         page = find_or_create_page(mapping, 0, GFP_NOFS);
1502         if (!page) {
1503                 ocfs2_commit_trans(osb, handle);
1504                 ret = -ENOMEM;
1505                 mlog_errno(ret);
1506                 goto out;
1507         }
1508         /*
1509          * If we don't set w_num_pages then this page won't get unlocked
1510          * and freed on cleanup of the write context.
1511          */
1512         wc->w_pages[0] = wc->w_target_page = page;
1513         wc->w_num_pages = 1;
1514
1515         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1516                                       OCFS2_JOURNAL_ACCESS_WRITE);
1517         if (ret) {
1518                 ocfs2_commit_trans(osb, handle);
1519
1520                 mlog_errno(ret);
1521                 goto out;
1522         }
1523
1524         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1525                 ocfs2_set_inode_data_inline(inode, di);
1526
1527         if (!PageUptodate(page)) {
1528                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1529                 if (ret) {
1530                         ocfs2_commit_trans(osb, handle);
1531
1532                         goto out;
1533                 }
1534         }
1535
1536         wc->w_handle = handle;
1537 out:
1538         return ret;
1539 }
1540
1541 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1542 {
1543         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1544
1545         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1546                 return 1;
1547         return 0;
1548 }
1549
1550 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1551                                           struct inode *inode, loff_t pos,
1552                                           unsigned len, struct page *mmap_page,
1553                                           struct ocfs2_write_ctxt *wc)
1554 {
1555         int ret, written = 0;
1556         loff_t end = pos + len;
1557         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1558         struct ocfs2_dinode *di = NULL;
1559
1560         trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1561                                              len, (unsigned long long)pos,
1562                                              oi->ip_dyn_features);
1563
1564         /*
1565          * Handle inodes which already have inline data 1st.
1566          */
1567         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1568                 if (mmap_page == NULL &&
1569                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1570                         goto do_inline_write;
1571
1572                 /*
1573                  * The write won't fit - we have to give this inode an
1574                  * inline extent list now.
1575                  */
1576                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1577                 if (ret)
1578                         mlog_errno(ret);
1579                 goto out;
1580         }
1581
1582         /*
1583          * Check whether the inode can accept inline data.
1584          */
1585         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1586                 return 0;
1587
1588         /*
1589          * Check whether the write can fit.
1590          */
1591         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1592         if (mmap_page ||
1593             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1594                 return 0;
1595
1596 do_inline_write:
1597         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1598         if (ret) {
1599                 mlog_errno(ret);
1600                 goto out;
1601         }
1602
1603         /*
1604          * This signals to the caller that the data can be written
1605          * inline.
1606          */
1607         written = 1;
1608 out:
1609         return written ? written : ret;
1610 }
1611
1612 /*
1613  * This function only does anything for file systems which can't
1614  * handle sparse files.
1615  *
1616  * What we want to do here is fill in any hole between the current end
1617  * of allocation and the end of our write. That way the rest of the
1618  * write path can treat it as an non-allocating write, which has no
1619  * special case code for sparse/nonsparse files.
1620  */
1621 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1622                                         struct buffer_head *di_bh,
1623                                         loff_t pos, unsigned len,
1624                                         struct ocfs2_write_ctxt *wc)
1625 {
1626         int ret;
1627         loff_t newsize = pos + len;
1628
1629         BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1630
1631         if (newsize <= i_size_read(inode))
1632                 return 0;
1633
1634         ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1635         if (ret)
1636                 mlog_errno(ret);
1637
1638         /* There is no wc if this is call from direct. */
1639         if (wc)
1640                 wc->w_first_new_cpos =
1641                         ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1642
1643         return ret;
1644 }
1645
1646 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1647                            loff_t pos)
1648 {
1649         int ret = 0;
1650
1651         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1652         if (pos > i_size_read(inode))
1653                 ret = ocfs2_zero_extend(inode, di_bh, pos);
1654
1655         return ret;
1656 }
1657
1658 int ocfs2_write_begin_nolock(struct address_space *mapping,
1659                              loff_t pos, unsigned len, ocfs2_write_type_t type,
1660                              struct page **pagep, void **fsdata,
1661                              struct buffer_head *di_bh, struct page *mmap_page)
1662 {
1663         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1664         unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1665         struct ocfs2_write_ctxt *wc;
1666         struct inode *inode = mapping->host;
1667         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1668         struct ocfs2_dinode *di;
1669         struct ocfs2_alloc_context *data_ac = NULL;
1670         struct ocfs2_alloc_context *meta_ac = NULL;
1671         handle_t *handle;
1672         struct ocfs2_extent_tree et;
1673         int try_free = 1, ret1;
1674
1675 try_again:
1676         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, type, di_bh);
1677         if (ret) {
1678                 mlog_errno(ret);
1679                 return ret;
1680         }
1681
1682         if (ocfs2_supports_inline_data(osb)) {
1683                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1684                                                      mmap_page, wc);
1685                 if (ret == 1) {
1686                         ret = 0;
1687                         goto success;
1688                 }
1689                 if (ret < 0) {
1690                         mlog_errno(ret);
1691                         goto out;
1692                 }
1693         }
1694
1695         /* Direct io change i_size late, should not zero tail here. */
1696         if (type != OCFS2_WRITE_DIRECT) {
1697                 if (ocfs2_sparse_alloc(osb))
1698                         ret = ocfs2_zero_tail(inode, di_bh, pos);
1699                 else
1700                         ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
1701                                                            len, wc);
1702                 if (ret) {
1703                         mlog_errno(ret);
1704                         goto out;
1705                 }
1706         }
1707
1708         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1709         if (ret < 0) {
1710                 mlog_errno(ret);
1711                 goto out;
1712         } else if (ret == 1) {
1713                 clusters_need = wc->w_clen;
1714                 ret = ocfs2_refcount_cow(inode, di_bh,
1715                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1716                 if (ret) {
1717                         mlog_errno(ret);
1718                         goto out;
1719                 }
1720         }
1721
1722         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1723                                         &extents_to_split);
1724         if (ret) {
1725                 mlog_errno(ret);
1726                 goto out;
1727         }
1728         clusters_need += clusters_to_alloc;
1729
1730         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1731
1732         trace_ocfs2_write_begin_nolock(
1733                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1734                         (long long)i_size_read(inode),
1735                         le32_to_cpu(di->i_clusters),
1736                         pos, len, type, mmap_page,
1737                         clusters_to_alloc, extents_to_split);
1738
1739         /*
1740          * We set w_target_from, w_target_to here so that
1741          * ocfs2_write_end() knows which range in the target page to
1742          * write out. An allocation requires that we write the entire
1743          * cluster range.
1744          */
1745         if (clusters_to_alloc || extents_to_split) {
1746                 /*
1747                  * XXX: We are stretching the limits of
1748                  * ocfs2_lock_allocators(). It greatly over-estimates
1749                  * the work to be done.
1750                  */
1751                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1752                                               wc->w_di_bh);
1753                 ret = ocfs2_lock_allocators(inode, &et,
1754                                             clusters_to_alloc, extents_to_split,
1755                                             &data_ac, &meta_ac);
1756                 if (ret) {
1757                         mlog_errno(ret);
1758                         goto out;
1759                 }
1760
1761                 if (data_ac)
1762                         data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1763
1764                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1765                                                     &di->id2.i_list);
1766         } else if (type == OCFS2_WRITE_DIRECT)
1767                 /* direct write needs not to start trans if no extents alloc. */
1768                 goto success;
1769
1770         /*
1771          * We have to zero sparse allocated clusters, unwritten extent clusters,
1772          * and non-sparse clusters we just extended.  For non-sparse writes,
1773          * we know zeros will only be needed in the first and/or last cluster.
1774          */
1775         if (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1776                            wc->w_desc[wc->w_clen - 1].c_needs_zero))
1777                 cluster_of_pages = 1;
1778         else
1779                 cluster_of_pages = 0;
1780
1781         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1782
1783         handle = ocfs2_start_trans(osb, credits);
1784         if (IS_ERR(handle)) {
1785                 ret = PTR_ERR(handle);
1786                 mlog_errno(ret);
1787                 goto out;
1788         }
1789
1790         wc->w_handle = handle;
1791
1792         if (clusters_to_alloc) {
1793                 ret = dquot_alloc_space_nodirty(inode,
1794                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1795                 if (ret)
1796                         goto out_commit;
1797         }
1798
1799         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1800                                       OCFS2_JOURNAL_ACCESS_WRITE);
1801         if (ret) {
1802                 mlog_errno(ret);
1803                 goto out_quota;
1804         }
1805
1806         /*
1807          * Fill our page array first. That way we've grabbed enough so
1808          * that we can zero and flush if we error after adding the
1809          * extent.
1810          */
1811         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1812                                          cluster_of_pages, mmap_page);
1813         if (ret && ret != -EAGAIN) {
1814                 mlog_errno(ret);
1815                 goto out_quota;
1816         }
1817
1818         /*
1819          * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1820          * the target page. In this case, we exit with no error and no target
1821          * page. This will trigger the caller, page_mkwrite(), to re-try
1822          * the operation.
1823          */
1824         if (ret == -EAGAIN) {
1825                 BUG_ON(wc->w_target_page);
1826                 ret = 0;
1827                 goto out_quota;
1828         }
1829
1830         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1831                                           len);
1832         if (ret) {
1833                 mlog_errno(ret);
1834                 goto out_quota;
1835         }
1836
1837         if (data_ac)
1838                 ocfs2_free_alloc_context(data_ac);
1839         if (meta_ac)
1840                 ocfs2_free_alloc_context(meta_ac);
1841
1842 success:
1843         if (pagep)
1844                 *pagep = wc->w_target_page;
1845         *fsdata = wc;
1846         return 0;
1847 out_quota:
1848         if (clusters_to_alloc)
1849                 dquot_free_space(inode,
1850                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1851 out_commit:
1852         ocfs2_commit_trans(osb, handle);
1853
1854 out:
1855         /*
1856          * The mmapped page won't be unlocked in ocfs2_free_write_ctxt(),
1857          * even in case of error here like ENOSPC and ENOMEM. So, we need
1858          * to unlock the target page manually to prevent deadlocks when
1859          * retrying again on ENOSPC, or when returning non-VM_FAULT_LOCKED
1860          * to VM code.
1861          */
1862         if (wc->w_target_locked)
1863                 unlock_page(mmap_page);
1864
1865         ocfs2_free_write_ctxt(inode, wc);
1866
1867         if (data_ac) {
1868                 ocfs2_free_alloc_context(data_ac);
1869                 data_ac = NULL;
1870         }
1871         if (meta_ac) {
1872                 ocfs2_free_alloc_context(meta_ac);
1873                 meta_ac = NULL;
1874         }
1875
1876         if (ret == -ENOSPC && try_free) {
1877                 /*
1878                  * Try to free some truncate log so that we can have enough
1879                  * clusters to allocate.
1880                  */
1881                 try_free = 0;
1882
1883                 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1884                 if (ret1 == 1)
1885                         goto try_again;
1886
1887                 if (ret1 < 0)
1888                         mlog_errno(ret1);
1889         }
1890
1891         return ret;
1892 }
1893
1894 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1895                              loff_t pos, unsigned len, unsigned flags,
1896                              struct page **pagep, void **fsdata)
1897 {
1898         int ret;
1899         struct buffer_head *di_bh = NULL;
1900         struct inode *inode = mapping->host;
1901
1902         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1903         if (ret) {
1904                 mlog_errno(ret);
1905                 return ret;
1906         }
1907
1908         /*
1909          * Take alloc sem here to prevent concurrent lookups. That way
1910          * the mapping, zeroing and tree manipulation within
1911          * ocfs2_write() will be safe against ->readpage(). This
1912          * should also serve to lock out allocation from a shared
1913          * writeable region.
1914          */
1915         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1916
1917         ret = ocfs2_write_begin_nolock(mapping, pos, len, OCFS2_WRITE_BUFFER,
1918                                        pagep, fsdata, di_bh, NULL);
1919         if (ret) {
1920                 mlog_errno(ret);
1921                 goto out_fail;
1922         }
1923
1924         brelse(di_bh);
1925
1926         return 0;
1927
1928 out_fail:
1929         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1930
1931         brelse(di_bh);
1932         ocfs2_inode_unlock(inode, 1);
1933
1934         return ret;
1935 }
1936
1937 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1938                                    unsigned len, unsigned *copied,
1939                                    struct ocfs2_dinode *di,
1940                                    struct ocfs2_write_ctxt *wc)
1941 {
1942         void *kaddr;
1943
1944         if (unlikely(*copied < len)) {
1945                 if (!PageUptodate(wc->w_target_page)) {
1946                         *copied = 0;
1947                         return;
1948                 }
1949         }
1950
1951         kaddr = kmap_atomic(wc->w_target_page);
1952         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1953         kunmap_atomic(kaddr);
1954
1955         trace_ocfs2_write_end_inline(
1956              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1957              (unsigned long long)pos, *copied,
1958              le16_to_cpu(di->id2.i_data.id_count),
1959              le16_to_cpu(di->i_dyn_features));
1960 }
1961
1962 int ocfs2_write_end_nolock(struct address_space *mapping,
1963                            loff_t pos, unsigned len, unsigned copied, void *fsdata)
1964 {
1965         int i, ret;
1966         unsigned from, to, start = pos & (PAGE_SIZE - 1);
1967         struct inode *inode = mapping->host;
1968         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1969         struct ocfs2_write_ctxt *wc = fsdata;
1970         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1971         handle_t *handle = wc->w_handle;
1972         struct page *tmppage;
1973
1974         BUG_ON(!list_empty(&wc->w_unwritten_list));
1975
1976         if (handle) {
1977                 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
1978                                 wc->w_di_bh, OCFS2_JOURNAL_ACCESS_WRITE);
1979                 if (ret) {
1980                         copied = ret;
1981                         mlog_errno(ret);
1982                         goto out;
1983                 }
1984         }
1985
1986         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1987                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1988                 goto out_write_size;
1989         }
1990
1991         if (unlikely(copied < len) && wc->w_target_page) {
1992                 if (!PageUptodate(wc->w_target_page))
1993                         copied = 0;
1994
1995                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1996                                        start+len);
1997         }
1998         if (wc->w_target_page)
1999                 flush_dcache_page(wc->w_target_page);
2000
2001         for(i = 0; i < wc->w_num_pages; i++) {
2002                 tmppage = wc->w_pages[i];
2003
2004                 /* This is the direct io target page. */
2005                 if (tmppage == NULL)
2006                         continue;
2007
2008                 if (tmppage == wc->w_target_page) {
2009                         from = wc->w_target_from;
2010                         to = wc->w_target_to;
2011
2012                         BUG_ON(from > PAGE_SIZE ||
2013                                to > PAGE_SIZE ||
2014                                to < from);
2015                 } else {
2016                         /*
2017                          * Pages adjacent to the target (if any) imply
2018                          * a hole-filling write in which case we want
2019                          * to flush their entire range.
2020                          */
2021                         from = 0;
2022                         to = PAGE_SIZE;
2023                 }
2024
2025                 if (page_has_buffers(tmppage)) {
2026                         if (handle && ocfs2_should_order_data(inode))
2027                                 ocfs2_jbd2_file_inode(handle, inode);
2028                         block_commit_write(tmppage, from, to);
2029                 }
2030         }
2031
2032 out_write_size:
2033         /* Direct io do not update i_size here. */
2034         if (wc->w_type != OCFS2_WRITE_DIRECT) {
2035                 pos += copied;
2036                 if (pos > i_size_read(inode)) {
2037                         i_size_write(inode, pos);
2038                         mark_inode_dirty(inode);
2039                 }
2040                 inode->i_blocks = ocfs2_inode_sector_count(inode);
2041                 di->i_size = cpu_to_le64((u64)i_size_read(inode));
2042                 inode->i_mtime = inode->i_ctime = current_time(inode);
2043                 di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2044                 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2045                 if (handle)
2046                         ocfs2_update_inode_fsync_trans(handle, inode, 1);
2047         }
2048         if (handle)
2049                 ocfs2_journal_dirty(handle, wc->w_di_bh);
2050
2051 out:
2052         /* unlock pages before dealloc since it needs acquiring j_trans_barrier
2053          * lock, or it will cause a deadlock since journal commit threads holds
2054          * this lock and will ask for the page lock when flushing the data.
2055          * put it here to preserve the unlock order.
2056          */
2057         ocfs2_unlock_pages(wc);
2058
2059         if (handle)
2060                 ocfs2_commit_trans(osb, handle);
2061
2062         ocfs2_run_deallocs(osb, &wc->w_dealloc);
2063
2064         brelse(wc->w_di_bh);
2065         kfree(wc);
2066
2067         return copied;
2068 }
2069
2070 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2071                            loff_t pos, unsigned len, unsigned copied,
2072                            struct page *page, void *fsdata)
2073 {
2074         int ret;
2075         struct inode *inode = mapping->host;
2076
2077         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, fsdata);
2078
2079         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2080         ocfs2_inode_unlock(inode, 1);
2081
2082         return ret;
2083 }
2084
2085 struct ocfs2_dio_write_ctxt {
2086         struct list_head        dw_zero_list;
2087         unsigned                dw_zero_count;
2088         int                     dw_orphaned;
2089         pid_t                   dw_writer_pid;
2090 };
2091
2092 static struct ocfs2_dio_write_ctxt *
2093 ocfs2_dio_alloc_write_ctx(struct buffer_head *bh, int *alloc)
2094 {
2095         struct ocfs2_dio_write_ctxt *dwc = NULL;
2096
2097         if (bh->b_private)
2098                 return bh->b_private;
2099
2100         dwc = kmalloc(sizeof(struct ocfs2_dio_write_ctxt), GFP_NOFS);
2101         if (dwc == NULL)
2102                 return NULL;
2103         INIT_LIST_HEAD(&dwc->dw_zero_list);
2104         dwc->dw_zero_count = 0;
2105         dwc->dw_orphaned = 0;
2106         dwc->dw_writer_pid = task_pid_nr(current);
2107         bh->b_private = dwc;
2108         *alloc = 1;
2109
2110         return dwc;
2111 }
2112
2113 static void ocfs2_dio_free_write_ctx(struct inode *inode,
2114                                      struct ocfs2_dio_write_ctxt *dwc)
2115 {
2116         ocfs2_free_unwritten_list(inode, &dwc->dw_zero_list);
2117         kfree(dwc);
2118 }
2119
2120 /*
2121  * TODO: Make this into a generic get_blocks function.
2122  *
2123  * From do_direct_io in direct-io.c:
2124  *  "So what we do is to permit the ->get_blocks function to populate
2125  *   bh.b_size with the size of IO which is permitted at this offset and
2126  *   this i_blkbits."
2127  *
2128  * This function is called directly from get_more_blocks in direct-io.c.
2129  *
2130  * called like this: dio->get_blocks(dio->inode, fs_startblk,
2131  *                                      fs_count, map_bh, dio->rw == WRITE);
2132  */
2133 static int ocfs2_dio_wr_get_block(struct inode *inode, sector_t iblock,
2134                                struct buffer_head *bh_result, int create)
2135 {
2136         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2137         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2138         struct ocfs2_write_ctxt *wc;
2139         struct ocfs2_write_cluster_desc *desc = NULL;
2140         struct ocfs2_dio_write_ctxt *dwc = NULL;
2141         struct buffer_head *di_bh = NULL;
2142         u64 p_blkno;
2143         unsigned int i_blkbits = inode->i_sb->s_blocksize_bits;
2144         loff_t pos = iblock << i_blkbits;
2145         sector_t endblk = (i_size_read(inode) - 1) >> i_blkbits;
2146         unsigned len, total_len = bh_result->b_size;
2147         int ret = 0, first_get_block = 0;
2148
2149         len = osb->s_clustersize - (pos & (osb->s_clustersize - 1));
2150         len = min(total_len, len);
2151
2152         /*
2153          * bh_result->b_size is count in get_more_blocks according to write
2154          * "pos" and "end", we need map twice to return different buffer state:
2155          * 1. area in file size, not set NEW;
2156          * 2. area out file size, set  NEW.
2157          *
2158          *                 iblock    endblk
2159          * |--------|---------|---------|---------
2160          * |<-------area in file------->|
2161          */
2162
2163         if ((iblock <= endblk) &&
2164             ((iblock + ((len - 1) >> i_blkbits)) > endblk))
2165                 len = (endblk - iblock + 1) << i_blkbits;
2166
2167         mlog(0, "get block of %lu at %llu:%u req %u\n",
2168                         inode->i_ino, pos, len, total_len);
2169
2170         /*
2171          * Because we need to change file size in ocfs2_dio_end_io_write(), or
2172          * we may need to add it to orphan dir. So can not fall to fast path
2173          * while file size will be changed.
2174          */
2175         if (pos + total_len <= i_size_read(inode)) {
2176
2177                 /* This is the fast path for re-write. */
2178                 ret = ocfs2_lock_get_block(inode, iblock, bh_result, create);
2179                 if (buffer_mapped(bh_result) &&
2180                     !buffer_new(bh_result) &&
2181                     ret == 0)
2182                         goto out;
2183
2184                 /* Clear state set by ocfs2_get_block. */
2185                 bh_result->b_state = 0;
2186         }
2187
2188         dwc = ocfs2_dio_alloc_write_ctx(bh_result, &first_get_block);
2189         if (unlikely(dwc == NULL)) {
2190                 ret = -ENOMEM;
2191                 mlog_errno(ret);
2192                 goto out;
2193         }
2194
2195         if (ocfs2_clusters_for_bytes(inode->i_sb, pos + total_len) >
2196             ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode)) &&
2197             !dwc->dw_orphaned) {
2198                 /*
2199                  * when we are going to alloc extents beyond file size, add the
2200                  * inode to orphan dir, so we can recall those spaces when
2201                  * system crashed during write.
2202                  */
2203                 ret = ocfs2_add_inode_to_orphan(osb, inode);
2204                 if (ret < 0) {
2205                         mlog_errno(ret);
2206                         goto out;
2207                 }
2208                 dwc->dw_orphaned = 1;
2209         }
2210
2211         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2212         if (ret) {
2213                 mlog_errno(ret);
2214                 goto out;
2215         }
2216
2217         down_write(&oi->ip_alloc_sem);
2218
2219         if (first_get_block) {
2220                 if (ocfs2_sparse_alloc(osb))
2221                         ret = ocfs2_zero_tail(inode, di_bh, pos);
2222                 else
2223                         ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
2224                                                            total_len, NULL);
2225                 if (ret < 0) {
2226                         mlog_errno(ret);
2227                         goto unlock;
2228                 }
2229         }
2230
2231         ret = ocfs2_write_begin_nolock(inode->i_mapping, pos, len,
2232                                        OCFS2_WRITE_DIRECT, NULL,
2233                                        (void **)&wc, di_bh, NULL);
2234         if (ret) {
2235                 mlog_errno(ret);
2236                 goto unlock;
2237         }
2238
2239         desc = &wc->w_desc[0];
2240
2241         p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, desc->c_phys);
2242         BUG_ON(p_blkno == 0);
2243         p_blkno += iblock & (u64)(ocfs2_clusters_to_blocks(inode->i_sb, 1) - 1);
2244
2245         map_bh(bh_result, inode->i_sb, p_blkno);
2246         bh_result->b_size = len;
2247         if (desc->c_needs_zero)
2248                 set_buffer_new(bh_result);
2249
2250         if (iblock > endblk)
2251                 set_buffer_new(bh_result);
2252
2253         /* May sleep in end_io. It should not happen in a irq context. So defer
2254          * it to dio work queue. */
2255         set_buffer_defer_completion(bh_result);
2256
2257         if (!list_empty(&wc->w_unwritten_list)) {
2258                 struct ocfs2_unwritten_extent *ue = NULL;
2259
2260                 ue = list_first_entry(&wc->w_unwritten_list,
2261                                       struct ocfs2_unwritten_extent,
2262                                       ue_node);
2263                 BUG_ON(ue->ue_cpos != desc->c_cpos);
2264                 /* The physical address may be 0, fill it. */
2265                 ue->ue_phys = desc->c_phys;
2266
2267                 list_splice_tail_init(&wc->w_unwritten_list, &dwc->dw_zero_list);
2268                 dwc->dw_zero_count += wc->w_unwritten_count;
2269         }
2270
2271         ret = ocfs2_write_end_nolock(inode->i_mapping, pos, len, len, wc);
2272         BUG_ON(ret != len);
2273         ret = 0;
2274 unlock:
2275         up_write(&oi->ip_alloc_sem);
2276         ocfs2_inode_unlock(inode, 1);
2277         brelse(di_bh);
2278 out:
2279         if (ret < 0)
2280                 ret = -EIO;
2281         return ret;
2282 }
2283
2284 static int ocfs2_dio_end_io_write(struct inode *inode,
2285                                   struct ocfs2_dio_write_ctxt *dwc,
2286                                   loff_t offset,
2287                                   ssize_t bytes)
2288 {
2289         struct ocfs2_cached_dealloc_ctxt dealloc;
2290         struct ocfs2_extent_tree et;
2291         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2292         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2293         struct ocfs2_unwritten_extent *ue = NULL;
2294         struct buffer_head *di_bh = NULL;
2295         struct ocfs2_dinode *di;
2296         struct ocfs2_alloc_context *data_ac = NULL;
2297         struct ocfs2_alloc_context *meta_ac = NULL;
2298         handle_t *handle = NULL;
2299         loff_t end = offset + bytes;
2300         int ret = 0, credits = 0, locked = 0;
2301
2302         ocfs2_init_dealloc_ctxt(&dealloc);
2303
2304         /* We do clear unwritten, delete orphan, change i_size here. If neither
2305          * of these happen, we can skip all this. */
2306         if (list_empty(&dwc->dw_zero_list) &&
2307             end <= i_size_read(inode) &&
2308             !dwc->dw_orphaned)
2309                 goto out;
2310
2311         /* ocfs2_file_write_iter will get i_mutex, so we need not lock if we
2312          * are in that context. */
2313         if (dwc->dw_writer_pid != task_pid_nr(current)) {
2314                 inode_lock(inode);
2315                 locked = 1;
2316         }
2317
2318         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2319         if (ret < 0) {
2320                 mlog_errno(ret);
2321                 goto out;
2322         }
2323
2324         down_write(&oi->ip_alloc_sem);
2325
2326         /* Delete orphan before acquire i_mutex. */
2327         if (dwc->dw_orphaned) {
2328                 BUG_ON(dwc->dw_writer_pid != task_pid_nr(current));
2329
2330                 end = end > i_size_read(inode) ? end : 0;
2331
2332                 ret = ocfs2_del_inode_from_orphan(osb, inode, di_bh,
2333                                 !!end, end);
2334                 if (ret < 0)
2335                         mlog_errno(ret);
2336         }
2337
2338         di = (struct ocfs2_dinode *)di_bh->b_data;
2339
2340         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
2341
2342         /* Attach dealloc with extent tree in case that we may reuse extents
2343          * which are already unlinked from current extent tree due to extent
2344          * rotation and merging.
2345          */
2346         et.et_dealloc = &dealloc;
2347
2348         ret = ocfs2_lock_allocators(inode, &et, 0, dwc->dw_zero_count*2,
2349                                     &data_ac, &meta_ac);
2350         if (ret) {
2351                 mlog_errno(ret);
2352                 goto unlock;
2353         }
2354
2355         credits = ocfs2_calc_extend_credits(inode->i_sb, &di->id2.i_list);
2356
2357         handle = ocfs2_start_trans(osb, credits);
2358         if (IS_ERR(handle)) {
2359                 ret = PTR_ERR(handle);
2360                 mlog_errno(ret);
2361                 goto unlock;
2362         }
2363         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2364                                       OCFS2_JOURNAL_ACCESS_WRITE);
2365         if (ret) {
2366                 mlog_errno(ret);
2367                 goto commit;
2368         }
2369
2370         list_for_each_entry(ue, &dwc->dw_zero_list, ue_node) {
2371                 ret = ocfs2_mark_extent_written(inode, &et, handle,
2372                                                 ue->ue_cpos, 1,
2373                                                 ue->ue_phys,
2374                                                 meta_ac, &dealloc);
2375                 if (ret < 0) {
2376                         mlog_errno(ret);
2377                         break;
2378                 }
2379         }
2380
2381         if (end > i_size_read(inode)) {
2382                 ret = ocfs2_set_inode_size(handle, inode, di_bh, end);
2383                 if (ret < 0)
2384                         mlog_errno(ret);
2385         }
2386 commit:
2387         ocfs2_commit_trans(osb, handle);
2388 unlock:
2389         up_write(&oi->ip_alloc_sem);
2390         ocfs2_inode_unlock(inode, 1);
2391         brelse(di_bh);
2392 out:
2393         if (data_ac)
2394                 ocfs2_free_alloc_context(data_ac);
2395         if (meta_ac)
2396                 ocfs2_free_alloc_context(meta_ac);
2397         ocfs2_run_deallocs(osb, &dealloc);
2398         if (locked)
2399                 inode_unlock(inode);
2400         ocfs2_dio_free_write_ctx(inode, dwc);
2401
2402         return ret;
2403 }
2404
2405 /*
2406  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
2407  * particularly interested in the aio/dio case.  We use the rw_lock DLM lock
2408  * to protect io on one node from truncation on another.
2409  */
2410 static int ocfs2_dio_end_io(struct kiocb *iocb,
2411                             loff_t offset,
2412                             ssize_t bytes,
2413                             void *private)
2414 {
2415         struct inode *inode = file_inode(iocb->ki_filp);
2416         int level;
2417         int ret = 0;
2418
2419         /* this io's submitter should not have unlocked this before we could */
2420         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
2421
2422         if (bytes <= 0)
2423                 mlog_ratelimited(ML_ERROR, "Direct IO failed, bytes = %lld",
2424                                  (long long)bytes);
2425         if (private) {
2426                 if (bytes > 0)
2427                         ret = ocfs2_dio_end_io_write(inode, private, offset,
2428                                                      bytes);
2429                 else
2430                         ocfs2_dio_free_write_ctx(inode, private);
2431         }
2432
2433         ocfs2_iocb_clear_rw_locked(iocb);
2434
2435         level = ocfs2_iocb_rw_locked_level(iocb);
2436         ocfs2_rw_unlock(inode, level);
2437         return ret;
2438 }
2439
2440 static ssize_t ocfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2441 {
2442         struct file *file = iocb->ki_filp;
2443         struct inode *inode = file->f_mapping->host;
2444         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2445         get_block_t *get_block;
2446
2447         /*
2448          * Fallback to buffered I/O if we see an inode without
2449          * extents.
2450          */
2451         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2452                 return 0;
2453
2454         /* Fallback to buffered I/O if we do not support append dio. */
2455         if (iocb->ki_pos + iter->count > i_size_read(inode) &&
2456             !ocfs2_supports_append_dio(osb))
2457                 return 0;
2458
2459         if (iov_iter_rw(iter) == READ)
2460                 get_block = ocfs2_lock_get_block;
2461         else
2462                 get_block = ocfs2_dio_wr_get_block;
2463
2464         return __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
2465                                     iter, get_block,
2466                                     ocfs2_dio_end_io, NULL, 0);
2467 }
2468
2469 const struct address_space_operations ocfs2_aops = {
2470         .readpage               = ocfs2_readpage,
2471         .readpages              = ocfs2_readpages,
2472         .writepage              = ocfs2_writepage,
2473         .write_begin            = ocfs2_write_begin,
2474         .write_end              = ocfs2_write_end,
2475         .bmap                   = ocfs2_bmap,
2476         .direct_IO              = ocfs2_direct_IO,
2477         .invalidatepage         = block_invalidatepage,
2478         .releasepage            = ocfs2_releasepage,
2479         .migratepage            = buffer_migrate_page,
2480         .is_partially_uptodate  = block_is_partially_uptodate,
2481         .error_remove_page      = generic_error_remove_page,
2482 };