fs/ocfs2/aops.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public
  17  * License along with this program; if not, write to the
  18  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  19  * Boston, MA 021110-1307, USA.
  20  */
  21
  22 #include <linux/fs.h>
  23 #include <linux/slab.h>
  24 #include <linux/highmem.h>
  25 #include <linux/pagemap.h>
  26 #include <asm/byteorder.h>
  27 #include <linux/swap.h>
  28 #include <linux/pipe_fs_i.h>
  29 #include <linux/mpage.h>
  30 #include <linux/quotaops.h>
  31 #include <linux/blkdev.h>
  32 #include <linux/uio.h>
  33
  34 #include <cluster/masklog.h>
  35
  36 #include "ocfs2.h"
  37
  38 #include "alloc.h"
  39 #include "aops.h"
  40 #include "dlmglue.h"
  41 #include "extent_map.h"
  42 #include "file.h"
  43 #include "inode.h"
  44 #include "journal.h"
  45 #include "suballoc.h"
  46 #include "super.h"
  47 #include "symlink.h"
  48 #include "refcounttree.h"
  49 #include "ocfs2_trace.h"
  50
  51 #include "buffer_head_io.h"
  52 #include "dir.h"
  53 #include "namei.h"
  54 #include "sysfile.h"
  55
  56 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
  57                                    struct buffer_head *bh_result, int create)
  58 {
  59         int err = -EIO;
  60         int status;
  61         struct ocfs2_dinode *fe = NULL;
  62         struct buffer_head *bh = NULL;
  63         struct buffer_head *buffer_cache_bh = NULL;
  64         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  65         void *kaddr;
  66
  67         trace_ocfs2_symlink_get_block(
  68                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
  69                         (unsigned long long)iblock, bh_result, create);
  70
  71         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
  72
  73         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
  74                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
  75                      (unsigned long long)iblock);
  76                 goto bail;
  77         }
  78
  79         status = ocfs2_read_inode_block(inode, &bh);
  80         if (status < 0) {
  81                 mlog_errno(status);
  82                 goto bail;
  83         }
  84         fe = (struct ocfs2_dinode *) bh->b_data;
  85
  86         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
  87                                                     le32_to_cpu(fe->i_clusters))) {
  88                 err = -ENOMEM;
  89                 mlog(ML_ERROR, "block offset is outside the allocated size: "
  90                      "%llu\n", (unsigned long long)iblock);
  91                 goto bail;
  92         }
  93
  94         /* We don't use the page cache to create symlink data, so if
  95          * need be, copy it over from the buffer cache. */
  96         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  97                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  98                             iblock;
  99                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
 100                 if (!buffer_cache_bh) {
 101                         err = -ENOMEM;
 102                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
 103                         goto bail;
 104                 }
 105
 106                 /* we haven't locked out transactions, so a commit
 107                  * could've happened. Since we've got a reference on
 108                  * the bh, even if it commits while we're doing the
 109                  * copy, the data is still good. */
 110                 if (buffer_jbd(buffer_cache_bh)
 111                     && ocfs2_inode_is_new(inode)) {
 112                         kaddr = kmap_atomic(bh_result->b_page);
 113                         if (!kaddr) {
 114                                 mlog(ML_ERROR, "couldn't kmap!\n");
 115                                 goto bail;
 116                         }
 117                         memcpy(kaddr + (bh_result->b_size * iblock),
 118                                buffer_cache_bh->b_data,
 119                                bh_result->b_size);
 120                         kunmap_atomic(kaddr);
 121                         set_buffer_uptodate(bh_result);
 122                 }
 123                 brelse(buffer_cache_bh);
 124         }
 125
 126         map_bh(bh_result, inode->i_sb,
 127                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 128
 129         err = 0;
 130
 131 bail:
 132         brelse(bh);
 133
 134         return err;
 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_CACHE_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_CACHE_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_CACHE_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(&OCFS2_I(inode)->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 = list_entry(pages->prev, struct page, lru);
 388         start = (loff_t)last->index << PAGE_CACHE_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         /* We don't need to lock journal system files, since they aren't
 468          * accessed concurrently from multiple nodes.
 469          */
 470         if (!INODE_JOURNAL(inode)) {
 471                 err = ocfs2_inode_lock(inode, NULL, 0);
 472                 if (err) {
 473                         if (err != -ENOENT)
 474                                 mlog_errno(err);
 475                         goto bail;
 476                 }
 477                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
 478         }
 479
 480         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
 481                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
 482                                                   NULL);
 483
 484         if (!INODE_JOURNAL(inode)) {
 485                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
 486                 ocfs2_inode_unlock(inode, 0);
 487         }
 488
 489         if (err) {
 490                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 491                      (unsigned long long)block);
 492                 mlog_errno(err);
 493                 goto bail;
 494         }
 495
 496 bail:
 497         status = err ? 0 : p_blkno;
 498
 499         return status;
 500 }
 501
 502 /*
 503  * TODO: Make this into a generic get_blocks function.
 504  *
 505  * From do_direct_io in direct-io.c:
 506  *  "So what we do is to permit the ->get_blocks function to populate
 507  *   bh.b_size with the size of IO which is permitted at this offset and
 508  *   this i_blkbits."
 509  *
 510  * This function is called directly from get_more_blocks in direct-io.c.
 511  *
 512  * called like this: dio->get_blocks(dio->inode, fs_startblk,
 513  *                                      fs_count, map_bh, dio->rw == WRITE);
 514  */
 515 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
 516                                      struct buffer_head *bh_result, int create)
 517 {
 518         int ret;
 519         u32 cpos = 0;
 520         int alloc_locked = 0;
 521         u64 p_blkno, inode_blocks, contig_blocks;
 522         unsigned int ext_flags;
 523         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
 524         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
 525         unsigned long len = bh_result->b_size;
 526         unsigned int clusters_to_alloc = 0, contig_clusters = 0;
 527
 528         cpos = ocfs2_blocks_to_clusters(inode->i_sb, iblock);
 529
 530         /* This function won't even be called if the request isn't all
 531          * nicely aligned and of the right size, so there's no need
 532          * for us to check any of that. */
 533
 534         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 535
 536         down_read(&OCFS2_I(inode)->ip_alloc_sem);
 537
 538         /* This figures out the size of the next contiguous block, and
 539          * our logical offset */
 540         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
 541                                           &contig_blocks, &ext_flags);
 542         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 543
 544         if (ret) {
 545                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
 546                      (unsigned long long)iblock);
 547                 ret = -EIO;
 548                 goto bail;
 549         }
 550
 551         /* We should already CoW the refcounted extent in case of create. */
 552         BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED));
 553
 554         /* allocate blocks if no p_blkno is found, and create == 1 */
 555         if (!p_blkno && create) {
 556                 ret = ocfs2_inode_lock(inode, NULL, 1);
 557                 if (ret < 0) {
 558                         mlog_errno(ret);
 559                         goto bail;
 560                 }
 561
 562                 alloc_locked = 1;
 563
 564                 down_write(&OCFS2_I(inode)->ip_alloc_sem);
 565
 566                 /* fill hole, allocate blocks can't be larger than the size
 567                  * of the hole */
 568                 clusters_to_alloc = ocfs2_clusters_for_bytes(inode->i_sb, len);
 569                 contig_clusters = ocfs2_clusters_for_blocks(inode->i_sb,
 570                                 contig_blocks);
 571                 if (clusters_to_alloc > contig_clusters)
 572                         clusters_to_alloc = contig_clusters;
 573
 574                 /* allocate extent and insert them into the extent tree */
 575                 ret = ocfs2_extend_allocation(inode, cpos,
 576                                 clusters_to_alloc, 0);
 577                 if (ret < 0) {
 578                         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 579                         mlog_errno(ret);
 580                         goto bail;
 581                 }
 582
 583                 ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
 584                                 &contig_blocks, &ext_flags);
 585                 if (ret < 0) {
 586                         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 587                         mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
 588                                         (unsigned long long)iblock);
 589                         ret = -EIO;
 590                         goto bail;
 591                 }
 592                 set_buffer_new(bh_result);
 593                 up_write(&OCFS2_I(inode)->ip_alloc_sem);
 594         }
 595
 596         /*
 597          * get_more_blocks() expects us to describe a hole by clearing
 598          * the mapped bit on bh_result().
 599          *
 600          * Consider an unwritten extent as a hole.
 601          */
 602         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 603                 map_bh(bh_result, inode->i_sb, p_blkno);
 604         else
 605                 clear_buffer_mapped(bh_result);
 606
 607         /* make sure we don't map more than max_blocks blocks here as
 608            that's all the kernel will handle at this point. */
 609         if (max_blocks < contig_blocks)
 610                 contig_blocks = max_blocks;
 611         bh_result->b_size = contig_blocks << blocksize_bits;
 612 bail:
 613         if (alloc_locked)
 614                 ocfs2_inode_unlock(inode, 1);
 615         return ret;
 616 }
 617
 618 /*
 619  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
 620  * particularly interested in the aio/dio case.  We use the rw_lock DLM lock
 621  * to protect io on one node from truncation on another.
 622  */
 623 static void ocfs2_dio_end_io(struct kiocb *iocb,
 624                              loff_t offset,
 625                              ssize_t bytes,
 626                              void *private)
 627 {
 628         struct inode *inode = file_inode(iocb->ki_filp);
 629         int level;
 630
 631         /* this io's submitter should not have unlocked this before we could */
 632         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 633
 634         if (ocfs2_iocb_is_unaligned_aio(iocb)) {
 635                 ocfs2_iocb_clear_unaligned_aio(iocb);
 636
 637                 mutex_unlock(&OCFS2_I(inode)->ip_unaligned_aio);
 638         }
 639
 640         /* Let rw unlock to be done later to protect append direct io write */
 641         if (offset + bytes <= i_size_read(inode)) {
 642                 ocfs2_iocb_clear_rw_locked(iocb);
 643
 644                 level = ocfs2_iocb_rw_locked_level(iocb);
 645                 ocfs2_rw_unlock(inode, level);
 646         }
 647 }
 648
 649 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 650 {
 651         if (!page_has_buffers(page))
 652                 return 0;
 653         return try_to_free_buffers(page);
 654 }
 655
 656 static int ocfs2_is_overwrite(struct ocfs2_super *osb,
 657                 struct inode *inode, loff_t offset)
 658 {
 659         int ret = 0;
 660         u32 v_cpos = 0;
 661         u32 p_cpos = 0;
 662         unsigned int num_clusters = 0;
 663         unsigned int ext_flags = 0;
 664
 665         v_cpos = ocfs2_bytes_to_clusters(osb->sb, offset);
 666         ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos,
 667                         &num_clusters, &ext_flags);
 668         if (ret < 0) {
 669                 mlog_errno(ret);
 670                 return ret;
 671         }
 672
 673         if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 674                 return 1;
 675
 676         return 0;
 677 }
 678
 679 static int ocfs2_direct_IO_zero_extend(struct ocfs2_super *osb,
 680                 struct inode *inode, loff_t offset,
 681                 u64 zero_len, int cluster_align)
 682 {
 683         u32 p_cpos = 0;
 684         u32 v_cpos = ocfs2_bytes_to_clusters(osb->sb, i_size_read(inode));
 685         unsigned int num_clusters = 0;
 686         unsigned int ext_flags = 0;
 687         int ret = 0;
 688
 689         if (offset <= i_size_read(inode) || cluster_align)
 690                 return 0;
 691
 692         ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos, &num_clusters,
 693                         &ext_flags);
 694         if (ret < 0) {
 695                 mlog_errno(ret);
 696                 return ret;
 697         }
 698
 699         if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
 700                 u64 s = i_size_read(inode);
 701                 sector_t sector = ((u64)p_cpos << (osb->s_clustersize_bits - 9)) +
 702                         (do_div(s, osb->s_clustersize) >> 9);
 703
 704                 ret = blkdev_issue_zeroout(osb->sb->s_bdev, sector,
 705                                 zero_len >> 9, GFP_NOFS, false);
 706                 if (ret < 0)
 707                         mlog_errno(ret);
 708         }
 709
 710         return ret;
 711 }
 712
 713 static int ocfs2_direct_IO_extend_no_holes(struct ocfs2_super *osb,
 714                 struct inode *inode, loff_t offset)
 715 {
 716         u64 zero_start, zero_len, total_zero_len;
 717         u32 p_cpos = 0, clusters_to_add;
 718         u32 v_cpos = ocfs2_bytes_to_clusters(osb->sb, i_size_read(inode));
 719         unsigned int num_clusters = 0;
 720         unsigned int ext_flags = 0;
 721         u32 size_div, offset_div;
 722         int ret = 0;
 723
 724         {
 725                 u64 o = offset;
 726                 u64 s = i_size_read(inode);
 727
 728                 offset_div = do_div(o, osb->s_clustersize);
 729                 size_div = do_div(s, osb->s_clustersize);
 730         }
 731
 732         if (offset <= i_size_read(inode))
 733                 return 0;
 734
 735         clusters_to_add = ocfs2_bytes_to_clusters(inode->i_sb, offset) -
 736                 ocfs2_bytes_to_clusters(inode->i_sb, i_size_read(inode));
 737         total_zero_len = offset - i_size_read(inode);
 738         if (clusters_to_add)
 739                 total_zero_len -= offset_div;
 740
 741         /* Allocate clusters to fill out holes, and this is only needed
 742          * when we add more than one clusters. Otherwise the cluster will
 743          * be allocated during direct IO */
 744         if (clusters_to_add > 1) {
 745                 ret = ocfs2_extend_allocation(inode,
 746                                 OCFS2_I(inode)->ip_clusters,
 747                                 clusters_to_add - 1, 0);
 748                 if (ret) {
 749                         mlog_errno(ret);
 750                         goto out;
 751                 }
 752         }
 753
 754         while (total_zero_len) {
 755                 ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos, &num_clusters,
 756                                 &ext_flags);
 757                 if (ret < 0) {
 758                         mlog_errno(ret);
 759                         goto out;
 760                 }
 761
 762                 zero_start = ocfs2_clusters_to_bytes(osb->sb, p_cpos) +
 763                         size_div;
 764                 zero_len = ocfs2_clusters_to_bytes(osb->sb, num_clusters) -
 765                         size_div;
 766                 zero_len = min(total_zero_len, zero_len);
 767
 768                 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
 769                         ret = blkdev_issue_zeroout(osb->sb->s_bdev,
 770                                         zero_start >> 9, zero_len >> 9,
 771                                         GFP_NOFS, false);
 772                         if (ret < 0) {
 773                                 mlog_errno(ret);
 774                                 goto out;
 775                         }
 776                 }
 777
 778                 total_zero_len -= zero_len;
 779                 v_cpos += ocfs2_bytes_to_clusters(osb->sb, zero_len + size_div);
 780
 781                 /* Only at first iteration can be cluster not aligned.
 782                  * So set size_div to 0 for the rest */
 783                 size_div = 0;
 784         }
 785
 786 out:
 787         return ret;
 788 }
 789
 790 static ssize_t ocfs2_direct_IO_write(struct kiocb *iocb,
 791                 struct iov_iter *iter,
 792                 loff_t offset)
 793 {
 794         ssize_t ret = 0;
 795         ssize_t written = 0;
 796         bool orphaned = false;
 797         int is_overwrite = 0;
 798         struct file *file = iocb->ki_filp;
 799         struct inode *inode = file_inode(file)->i_mapping->host;
 800         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 801         struct buffer_head *di_bh = NULL;
 802         size_t count = iter->count;
 803         journal_t *journal = osb->journal->j_journal;
 804         u64 zero_len_head, zero_len_tail;
 805         int cluster_align_head, cluster_align_tail;
 806         loff_t final_size = offset + count;
 807         int append_write = offset >= i_size_read(inode) ? 1 : 0;
 808         unsigned int num_clusters = 0;
 809         unsigned int ext_flags = 0;
 810
 811         {
 812                 u64 o = offset;
 813                 u64 s = i_size_read(inode);
 814
 815                 zero_len_head = do_div(o, 1 << osb->s_clustersize_bits);
 816                 cluster_align_head = !zero_len_head;
 817
 818                 zero_len_tail = osb->s_clustersize -
 819                         do_div(s, osb->s_clustersize);
 820                 if ((offset - i_size_read(inode)) < zero_len_tail)
 821                         zero_len_tail = offset - i_size_read(inode);
 822                 cluster_align_tail = !zero_len_tail;
 823         }
 824
 825         /*
 826          * when final_size > inode->i_size, inode->i_size will be
 827          * updated after direct write, so add the inode to orphan
 828          * dir first.
 829          */
 830         if (final_size > i_size_read(inode)) {
 831                 ret = ocfs2_add_inode_to_orphan(osb, inode);
 832                 if (ret < 0) {
 833                         mlog_errno(ret);
 834                         goto out;
 835                 }
 836                 orphaned = true;
 837         }
 838
 839         if (append_write) {
 840                 ret = ocfs2_inode_lock(inode, NULL, 1);
 841                 if (ret < 0) {
 842                         mlog_errno(ret);
 843                         goto clean_orphan;
 844                 }
 845
 846                 /* zeroing out the previously allocated cluster tail
 847                  * that but not zeroed */
 848                 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
 849                         down_read(&OCFS2_I(inode)->ip_alloc_sem);
 850                         ret = ocfs2_direct_IO_zero_extend(osb, inode, offset,
 851                                         zero_len_tail, cluster_align_tail);
 852                         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 853                 } else {
 854                         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 855                         ret = ocfs2_direct_IO_extend_no_holes(osb, inode,
 856                                         offset);
 857                         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 858                 }
 859                 if (ret < 0) {
 860                         mlog_errno(ret);
 861                         ocfs2_inode_unlock(inode, 1);
 862                         goto clean_orphan;
 863                 }
 864
 865                 is_overwrite = ocfs2_is_overwrite(osb, inode, offset);
 866                 if (is_overwrite < 0) {
 867                         mlog_errno(is_overwrite);
 868                         ret = is_overwrite;
 869                         ocfs2_inode_unlock(inode, 1);
 870                         goto clean_orphan;
 871                 }
 872
 873                 ocfs2_inode_unlock(inode, 1);
 874         }
 875
 876         written = __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev, iter,
 877                                        offset, ocfs2_direct_IO_get_blocks,
 878                                        ocfs2_dio_end_io, NULL, 0);
 879         /* overwrite aio may return -EIOCBQUEUED, and it is not an error */
 880         if ((written < 0) && (written != -EIOCBQUEUED)) {
 881                 loff_t i_size = i_size_read(inode);
 882
 883                 if (offset + count > i_size) {
 884                         ret = ocfs2_inode_lock(inode, &di_bh, 1);
 885                         if (ret < 0) {
 886                                 mlog_errno(ret);
 887                                 goto clean_orphan;
 888                         }
 889
 890                         if (i_size == i_size_read(inode)) {
 891                                 ret = ocfs2_truncate_file(inode, di_bh,
 892                                                 i_size);
 893                                 if (ret < 0) {
 894                                         if (ret != -ENOSPC)
 895                                                 mlog_errno(ret);
 896
 897                                         ocfs2_inode_unlock(inode, 1);
 898                                         brelse(di_bh);
 899                                         di_bh = NULL;
 900                                         goto clean_orphan;
 901                                 }
 902                         }
 903
 904                         ocfs2_inode_unlock(inode, 1);
 905                         brelse(di_bh);
 906                         di_bh = NULL;
 907
 908                         ret = jbd2_journal_force_commit(journal);
 909                         if (ret < 0)
 910                                 mlog_errno(ret);
 911                 }
 912         } else if (written > 0 && append_write && !is_overwrite &&
 913                         !cluster_align_head) {
 914                 /* zeroing out the allocated cluster head */
 915                 u32 p_cpos = 0;
 916                 u32 v_cpos = ocfs2_bytes_to_clusters(osb->sb, offset);
 917
 918                 ret = ocfs2_inode_lock(inode, NULL, 0);
 919                 if (ret < 0) {
 920                         mlog_errno(ret);
 921                         goto clean_orphan;
 922                 }
 923
 924                 ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos,
 925                                 &num_clusters, &ext_flags);
 926                 if (ret < 0) {
 927                         mlog_errno(ret);
 928                         ocfs2_inode_unlock(inode, 0);
 929                         goto clean_orphan;
 930                 }
 931
 932                 BUG_ON(!p_cpos || (ext_flags & OCFS2_EXT_UNWRITTEN));
 933
 934                 ret = blkdev_issue_zeroout(osb->sb->s_bdev,
 935                                 (u64)p_cpos << (osb->s_clustersize_bits - 9),
 936                                 zero_len_head >> 9, GFP_NOFS, false);
 937                 if (ret < 0)
 938                         mlog_errno(ret);
 939
 940                 ocfs2_inode_unlock(inode, 0);
 941         }
 942
 943 clean_orphan:
 944         if (orphaned) {
 945                 int tmp_ret;
 946                 int update_isize = written > 0 ? 1 : 0;
 947                 loff_t end = update_isize ? offset + written : 0;
 948
 949                 tmp_ret = ocfs2_inode_lock(inode, &di_bh, 1);
 950                 if (tmp_ret < 0) {
 951                         ret = tmp_ret;
 952                         mlog_errno(ret);
 953                         goto out;
 954                 }
 955
 956                 tmp_ret = ocfs2_del_inode_from_orphan(osb, inode, di_bh,
 957                                 update_isize, end);
 958                 if (tmp_ret < 0) {
 959                         ret = tmp_ret;
 960                         mlog_errno(ret);
 961                         brelse(di_bh);
 962                         goto out;
 963                 }
 964
 965                 ocfs2_inode_unlock(inode, 1);
 966                 brelse(di_bh);
 967
 968                 tmp_ret = jbd2_journal_force_commit(journal);
 969                 if (tmp_ret < 0) {
 970                         ret = tmp_ret;
 971                         mlog_errno(tmp_ret);
 972                 }
 973         }
 974
 975 out:
 976         if (ret >= 0)
 977                 ret = written;
 978         return ret;
 979 }
 980
 981 static ssize_t ocfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
 982                                loff_t offset)
 983 {
 984         struct file *file = iocb->ki_filp;
 985         struct inode *inode = file_inode(file)->i_mapping->host;
 986         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 987         int full_coherency = !(osb->s_mount_opt &
 988                         OCFS2_MOUNT_COHERENCY_BUFFERED);
 989
 990         /*
 991          * Fallback to buffered I/O if we see an inode without
 992          * extents.
 993          */
 994         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 995                 return 0;
 996
 997         /* Fallback to buffered I/O if we are appending and
 998          * concurrent O_DIRECT writes are allowed.
 999          */
1000         if (i_size_read(inode) <= offset && !full_coherency)
1001                 return 0;
1002
1003         if (iov_iter_rw(iter) == READ)
1004                 return __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
1005                                             iter, offset,
1006                                             ocfs2_direct_IO_get_blocks,
1007                                             ocfs2_dio_end_io, NULL, 0);
1008         else
1009                 return ocfs2_direct_IO_write(iocb, iter, offset);
1010 }
1011
1012 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
1013                                             u32 cpos,
1014                                             unsigned int *start,
1015                                             unsigned int *end)
1016 {
1017         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
1018
1019         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
1020                 unsigned int cpp;
1021
1022                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
1023
1024                 cluster_start = cpos % cpp;
1025                 cluster_start = cluster_start << osb->s_clustersize_bits;
1026
1027                 cluster_end = cluster_start + osb->s_clustersize;
1028         }
1029
1030         BUG_ON(cluster_start > PAGE_SIZE);
1031         BUG_ON(cluster_end > PAGE_SIZE);
1032
1033         if (start)
1034                 *start = cluster_start;
1035         if (end)
1036                 *end = cluster_end;
1037 }
1038
1039 /*
1040  * 'from' and 'to' are the region in the page to avoid zeroing.
1041  *
1042  * If pagesize > clustersize, this function will avoid zeroing outside
1043  * of the cluster boundary.
1044  *
1045  * from == to == 0 is code for "zero the entire cluster region"
1046  */
1047 static void ocfs2_clear_page_regions(struct page *page,
1048                                      struct ocfs2_super *osb, u32 cpos,
1049                                      unsigned from, unsigned to)
1050 {
1051         void *kaddr;
1052         unsigned int cluster_start, cluster_end;
1053
1054         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
1055
1056         kaddr = kmap_atomic(page);
1057
1058         if (from || to) {
1059                 if (from > cluster_start)
1060                         memset(kaddr + cluster_start, 0, from - cluster_start);
1061                 if (to < cluster_end)
1062                         memset(kaddr + to, 0, cluster_end - to);
1063         } else {
1064                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
1065         }
1066
1067         kunmap_atomic(kaddr);
1068 }
1069
1070 /*
1071  * Nonsparse file systems fully allocate before we get to the write
1072  * code. This prevents ocfs2_write() from tagging the write as an
1073  * allocating one, which means ocfs2_map_page_blocks() might try to
1074  * read-in the blocks at the tail of our file. Avoid reading them by
1075  * testing i_size against each block offset.
1076  */
1077 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
1078                                  unsigned int block_start)
1079 {
1080         u64 offset = page_offset(page) + block_start;
1081
1082         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
1083                 return 1;
1084
1085         if (i_size_read(inode) > offset)
1086                 return 1;
1087
1088         return 0;
1089 }
1090
1091 /*
1092  * Some of this taken from __block_write_begin(). We already have our
1093  * mapping by now though, and the entire write will be allocating or
1094  * it won't, so not much need to use BH_New.
1095  *
1096  * This will also skip zeroing, which is handled externally.
1097  */
1098 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
1099                           struct inode *inode, unsigned int from,
1100                           unsigned int to, int new)
1101 {
1102         int ret = 0;
1103         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
1104         unsigned int block_end, block_start;
1105         unsigned int bsize = 1 << inode->i_blkbits;
1106
1107         if (!page_has_buffers(page))
1108                 create_empty_buffers(page, bsize, 0);
1109
1110         head = page_buffers(page);
1111         for (bh = head, block_start = 0; bh != head || !block_start;
1112              bh = bh->b_this_page, block_start += bsize) {
1113                 block_end = block_start + bsize;
1114
1115                 clear_buffer_new(bh);
1116
1117                 /*
1118                  * Ignore blocks outside of our i/o range -
1119                  * they may belong to unallocated clusters.
1120                  */
1121                 if (block_start >= to || block_end <= from) {
1122                         if (PageUptodate(page))
1123                                 set_buffer_uptodate(bh);
1124                         continue;
1125                 }
1126
1127                 /*
1128                  * For an allocating write with cluster size >= page
1129                  * size, we always write the entire page.
1130                  */
1131                 if (new)
1132                         set_buffer_new(bh);
1133
1134                 if (!buffer_mapped(bh)) {
1135                         map_bh(bh, inode->i_sb, *p_blkno);
1136                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
1137                 }
1138
1139                 if (PageUptodate(page)) {
1140                         if (!buffer_uptodate(bh))
1141                                 set_buffer_uptodate(bh);
1142                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1143                            !buffer_new(bh) &&
1144                            ocfs2_should_read_blk(inode, page, block_start) &&
1145                            (block_start < from || block_end > to)) {
1146                         ll_rw_block(READ, 1, &bh);
1147                         *wait_bh++=bh;
1148                 }
1149
1150                 *p_blkno = *p_blkno + 1;
1151         }
1152
1153         /*
1154          * If we issued read requests - let them complete.
1155          */
1156         while(wait_bh > wait) {
1157                 wait_on_buffer(*--wait_bh);
1158                 if (!buffer_uptodate(*wait_bh))
1159                         ret = -EIO;
1160         }
1161
1162         if (ret == 0 || !new)
1163                 return ret;
1164
1165         /*
1166          * If we get -EIO above, zero out any newly allocated blocks
1167          * to avoid exposing stale data.
1168          */
1169         bh = head;
1170         block_start = 0;
1171         do {
1172                 block_end = block_start + bsize;
1173                 if (block_end <= from)
1174                         goto next_bh;
1175                 if (block_start >= to)
1176                         break;
1177
1178                 zero_user(page, block_start, bh->b_size);
1179                 set_buffer_uptodate(bh);
1180                 mark_buffer_dirty(bh);
1181
1182 next_bh:
1183                 block_start = block_end;
1184                 bh = bh->b_this_page;
1185         } while (bh != head);
1186
1187         return ret;
1188 }
1189
1190 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
1191 #define OCFS2_MAX_CTXT_PAGES    1
1192 #else
1193 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
1194 #endif
1195
1196 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
1197
1198 /*
1199  * Describe the state of a single cluster to be written to.
1200  */
1201 struct ocfs2_write_cluster_desc {
1202         u32             c_cpos;
1203         u32             c_phys;
1204         /*
1205          * Give this a unique field because c_phys eventually gets
1206          * filled.
1207          */
1208         unsigned        c_new;
1209         unsigned        c_unwritten;
1210         unsigned        c_needs_zero;
1211 };
1212
1213 struct ocfs2_write_ctxt {
1214         /* Logical cluster position / len of write */
1215         u32                             w_cpos;
1216         u32                             w_clen;
1217
1218         /* First cluster allocated in a nonsparse extend */
1219         u32                             w_first_new_cpos;
1220
1221         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
1222
1223         /*
1224          * This is true if page_size > cluster_size.
1225          *
1226          * It triggers a set of special cases during write which might
1227          * have to deal with allocating writes to partial pages.
1228          */
1229         unsigned int                    w_large_pages;
1230
1231         /*
1232          * Pages involved in this write.
1233          *
1234          * w_target_page is the page being written to by the user.
1235          *
1236          * w_pages is an array of pages which always contains
1237          * w_target_page, and in the case of an allocating write with
1238          * page_size < cluster size, it will contain zero'd and mapped
1239          * pages adjacent to w_target_page which need to be written
1240          * out in so that future reads from that region will get
1241          * zero's.
1242          */
1243         unsigned int                    w_num_pages;
1244         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
1245         struct page                     *w_target_page;
1246
1247         /*
1248          * w_target_locked is used for page_mkwrite path indicating no unlocking
1249          * against w_target_page in ocfs2_write_end_nolock.
1250          */
1251         unsigned int                    w_target_locked:1;
1252
1253         /*
1254          * ocfs2_write_end() uses this to know what the real range to
1255          * write in the target should be.
1256          */
1257         unsigned int                    w_target_from;
1258         unsigned int                    w_target_to;
1259
1260         /*
1261          * We could use journal_current_handle() but this is cleaner,
1262          * IMHO -Mark
1263          */
1264         handle_t                        *w_handle;
1265
1266         struct buffer_head              *w_di_bh;
1267
1268         struct ocfs2_cached_dealloc_ctxt w_dealloc;
1269 };
1270
1271 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
1272 {
1273         int i;
1274
1275         for(i = 0; i < num_pages; i++) {
1276                 if (pages[i]) {
1277                         unlock_page(pages[i]);
1278                         mark_page_accessed(pages[i]);
1279                         page_cache_release(pages[i]);
1280                 }
1281         }
1282 }
1283
1284 static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc)
1285 {
1286         int i;
1287
1288         /*
1289          * w_target_locked is only set to true in the page_mkwrite() case.
1290          * The intent is to allow us to lock the target page from write_begin()
1291          * to write_end(). The caller must hold a ref on w_target_page.
1292          */
1293         if (wc->w_target_locked) {
1294                 BUG_ON(!wc->w_target_page);
1295                 for (i = 0; i < wc->w_num_pages; i++) {
1296                         if (wc->w_target_page == wc->w_pages[i]) {
1297                                 wc->w_pages[i] = NULL;
1298                                 break;
1299                         }
1300                 }
1301                 mark_page_accessed(wc->w_target_page);
1302                 page_cache_release(wc->w_target_page);
1303         }
1304         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
1305 }
1306
1307 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
1308 {
1309         ocfs2_unlock_pages(wc);
1310         brelse(wc->w_di_bh);
1311         kfree(wc);
1312 }
1313
1314 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
1315                                   struct ocfs2_super *osb, loff_t pos,
1316                                   unsigned len, struct buffer_head *di_bh)
1317 {
1318         u32 cend;
1319         struct ocfs2_write_ctxt *wc;
1320
1321         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
1322         if (!wc)
1323                 return -ENOMEM;
1324
1325         wc->w_cpos = pos >> osb->s_clustersize_bits;
1326         wc->w_first_new_cpos = UINT_MAX;
1327         cend = (pos + len - 1) >> osb->s_clustersize_bits;
1328         wc->w_clen = cend - wc->w_cpos + 1;
1329         get_bh(di_bh);
1330         wc->w_di_bh = di_bh;
1331
1332         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
1333                 wc->w_large_pages = 1;
1334         else
1335                 wc->w_large_pages = 0;
1336
1337         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
1338
1339         *wcp = wc;
1340
1341         return 0;
1342 }
1343
1344 /*
1345  * If a page has any new buffers, zero them out here, and mark them uptodate
1346  * and dirty so they'll be written out (in order to prevent uninitialised
1347  * block data from leaking). And clear the new bit.
1348  */
1349 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1350 {
1351         unsigned int block_start, block_end;
1352         struct buffer_head *head, *bh;
1353
1354         BUG_ON(!PageLocked(page));
1355         if (!page_has_buffers(page))
1356                 return;
1357
1358         bh = head = page_buffers(page);
1359         block_start = 0;
1360         do {
1361                 block_end = block_start + bh->b_size;
1362
1363                 if (buffer_new(bh)) {
1364                         if (block_end > from && block_start < to) {
1365                                 if (!PageUptodate(page)) {
1366                                         unsigned start, end;
1367
1368                                         start = max(from, block_start);
1369                                         end = min(to, block_end);
1370
1371                                         zero_user_segment(page, start, end);
1372                                         set_buffer_uptodate(bh);
1373                                 }
1374
1375                                 clear_buffer_new(bh);
1376                                 mark_buffer_dirty(bh);
1377                         }
1378                 }
1379
1380                 block_start = block_end;
1381                 bh = bh->b_this_page;
1382         } while (bh != head);
1383 }
1384
1385 /*
1386  * Only called when we have a failure during allocating write to write
1387  * zero's to the newly allocated region.
1388  */
1389 static void ocfs2_write_failure(struct inode *inode,
1390                                 struct ocfs2_write_ctxt *wc,
1391                                 loff_t user_pos, unsigned user_len)
1392 {
1393         int i;
1394         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1395                 to = user_pos + user_len;
1396         struct page *tmppage;
1397
1398         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1399
1400         for(i = 0; i < wc->w_num_pages; i++) {
1401                 tmppage = wc->w_pages[i];
1402
1403                 if (page_has_buffers(tmppage)) {
1404                         if (ocfs2_should_order_data(inode))
1405                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1406
1407                         block_commit_write(tmppage, from, to);
1408                 }
1409         }
1410 }
1411
1412 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1413                                         struct ocfs2_write_ctxt *wc,
1414                                         struct page *page, u32 cpos,
1415                                         loff_t user_pos, unsigned user_len,
1416                                         int new)
1417 {
1418         int ret;
1419         unsigned int map_from = 0, map_to = 0;
1420         unsigned int cluster_start, cluster_end;
1421         unsigned int user_data_from = 0, user_data_to = 0;
1422
1423         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1424                                         &cluster_start, &cluster_end);
1425
1426         /* treat the write as new if the a hole/lseek spanned across
1427          * the page boundary.
1428          */
1429         new = new | ((i_size_read(inode) <= page_offset(page)) &&
1430                         (page_offset(page) <= user_pos));
1431
1432         if (page == wc->w_target_page) {
1433                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1434                 map_to = map_from + user_len;
1435
1436                 if (new)
1437                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1438                                                     cluster_start, cluster_end,
1439                                                     new);
1440                 else
1441                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1442                                                     map_from, map_to, new);
1443                 if (ret) {
1444                         mlog_errno(ret);
1445                         goto out;
1446                 }
1447
1448                 user_data_from = map_from;
1449                 user_data_to = map_to;
1450                 if (new) {
1451                         map_from = cluster_start;
1452                         map_to = cluster_end;
1453                 }
1454         } else {
1455                 /*
1456                  * If we haven't allocated the new page yet, we
1457                  * shouldn't be writing it out without copying user
1458                  * data. This is likely a math error from the caller.
1459                  */
1460                 BUG_ON(!new);
1461
1462                 map_from = cluster_start;
1463                 map_to = cluster_end;
1464
1465                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1466                                             cluster_start, cluster_end, new);
1467                 if (ret) {
1468                         mlog_errno(ret);
1469                         goto out;
1470                 }
1471         }
1472
1473         /*
1474          * Parts of newly allocated pages need to be zero'd.
1475          *
1476          * Above, we have also rewritten 'to' and 'from' - as far as
1477          * the rest of the function is concerned, the entire cluster
1478          * range inside of a page needs to be written.
1479          *
1480          * We can skip this if the page is up to date - it's already
1481          * been zero'd from being read in as a hole.
1482          */
1483         if (new && !PageUptodate(page))
1484                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1485                                          cpos, user_data_from, user_data_to);
1486
1487         flush_dcache_page(page);
1488
1489 out:
1490         return ret;
1491 }
1492
1493 /*
1494  * This function will only grab one clusters worth of pages.
1495  */
1496 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1497                                       struct ocfs2_write_ctxt *wc,
1498                                       u32 cpos, loff_t user_pos,
1499                                       unsigned user_len, int new,
1500                                       struct page *mmap_page)
1501 {
1502         int ret = 0, i;
1503         unsigned long start, target_index, end_index, index;
1504         struct inode *inode = mapping->host;
1505         loff_t last_byte;
1506
1507         target_index = user_pos >> PAGE_CACHE_SHIFT;
1508
1509         /*
1510          * Figure out how many pages we'll be manipulating here. For
1511          * non allocating write, we just change the one
1512          * page. Otherwise, we'll need a whole clusters worth.  If we're
1513          * writing past i_size, we only need enough pages to cover the
1514          * last page of the write.
1515          */
1516         if (new) {
1517                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1518                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1519                 /*
1520                  * We need the index *past* the last page we could possibly
1521                  * touch.  This is the page past the end of the write or
1522                  * i_size, whichever is greater.
1523                  */
1524                 last_byte = max(user_pos + user_len, i_size_read(inode));
1525                 BUG_ON(last_byte < 1);
1526                 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1;
1527                 if ((start + wc->w_num_pages) > end_index)
1528                         wc->w_num_pages = end_index - start;
1529         } else {
1530                 wc->w_num_pages = 1;
1531                 start = target_index;
1532         }
1533
1534         for(i = 0; i < wc->w_num_pages; i++) {
1535                 index = start + i;
1536
1537                 if (index == target_index && mmap_page) {
1538                         /*
1539                          * ocfs2_pagemkwrite() is a little different
1540                          * and wants us to directly use the page
1541                          * passed in.
1542                          */
1543                         lock_page(mmap_page);
1544
1545                         /* Exit and let the caller retry */
1546                         if (mmap_page->mapping != mapping) {
1547                                 WARN_ON(mmap_page->mapping);
1548                                 unlock_page(mmap_page);
1549                                 ret = -EAGAIN;
1550                                 goto out;
1551                         }
1552
1553                         page_cache_get(mmap_page);
1554                         wc->w_pages[i] = mmap_page;
1555                         wc->w_target_locked = true;
1556                 } else {
1557                         wc->w_pages[i] = find_or_create_page(mapping, index,
1558                                                              GFP_NOFS);
1559                         if (!wc->w_pages[i]) {
1560                                 ret = -ENOMEM;
1561                                 mlog_errno(ret);
1562                                 goto out;
1563                         }
1564                 }
1565                 wait_for_stable_page(wc->w_pages[i]);
1566
1567                 if (index == target_index)
1568                         wc->w_target_page = wc->w_pages[i];
1569         }
1570 out:
1571         if (ret)
1572                 wc->w_target_locked = false;
1573         return ret;
1574 }
1575
1576 /*
1577  * Prepare a single cluster for write one cluster into the file.
1578  */
1579 static int ocfs2_write_cluster(struct address_space *mapping,
1580                                u32 phys, unsigned int unwritten,
1581                                unsigned int should_zero,
1582                                struct ocfs2_alloc_context *data_ac,
1583                                struct ocfs2_alloc_context *meta_ac,
1584                                struct ocfs2_write_ctxt *wc, u32 cpos,
1585                                loff_t user_pos, unsigned user_len)
1586 {
1587         int ret, i, new;
1588         u64 v_blkno, p_blkno;
1589         struct inode *inode = mapping->host;
1590         struct ocfs2_extent_tree et;
1591
1592         new = phys == 0 ? 1 : 0;
1593         if (new) {
1594                 u32 tmp_pos;
1595
1596                 /*
1597                  * This is safe to call with the page locks - it won't take
1598                  * any additional semaphores or cluster locks.
1599                  */
1600                 tmp_pos = cpos;
1601                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1602                                            &tmp_pos, 1, 0, wc->w_di_bh,
1603                                            wc->w_handle, data_ac,
1604                                            meta_ac, NULL);
1605                 /*
1606                  * This shouldn't happen because we must have already
1607                  * calculated the correct meta data allocation required. The
1608                  * internal tree allocation code should know how to increase
1609                  * transaction credits itself.
1610                  *
1611                  * If need be, we could handle -EAGAIN for a
1612                  * RESTART_TRANS here.
1613                  */
1614                 mlog_bug_on_msg(ret == -EAGAIN,
1615                                 "Inode %llu: EAGAIN return during allocation.\n",
1616                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1617                 if (ret < 0) {
1618                         mlog_errno(ret);
1619                         goto out;
1620                 }
1621         } else if (unwritten) {
1622                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1623                                               wc->w_di_bh);
1624                 ret = ocfs2_mark_extent_written(inode, &et,
1625                                                 wc->w_handle, cpos, 1, phys,
1626                                                 meta_ac, &wc->w_dealloc);
1627                 if (ret < 0) {
1628                         mlog_errno(ret);
1629                         goto out;
1630                 }
1631         }
1632
1633         if (should_zero)
1634                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1635         else
1636                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1637
1638         /*
1639          * The only reason this should fail is due to an inability to
1640          * find the extent added.
1641          */
1642         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1643                                           NULL);
1644         if (ret < 0) {
1645                 mlog(ML_ERROR, "Get physical blkno failed for inode %llu, "
1646                             "at logical block %llu",
1647                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1648                             (unsigned long long)v_blkno);
1649                 goto out;
1650         }
1651
1652         BUG_ON(p_blkno == 0);
1653
1654         for(i = 0; i < wc->w_num_pages; i++) {
1655                 int tmpret;
1656
1657                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1658                                                       wc->w_pages[i], cpos,
1659                                                       user_pos, user_len,
1660                                                       should_zero);
1661                 if (tmpret) {
1662                         mlog_errno(tmpret);
1663                         if (ret == 0)
1664                                 ret = tmpret;
1665                 }
1666         }
1667
1668         /*
1669          * We only have cleanup to do in case of allocating write.
1670          */
1671         if (ret && new)
1672                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1673
1674 out:
1675
1676         return ret;
1677 }
1678
1679 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1680                                        struct ocfs2_alloc_context *data_ac,
1681                                        struct ocfs2_alloc_context *meta_ac,
1682                                        struct ocfs2_write_ctxt *wc,
1683                                        loff_t pos, unsigned len)
1684 {
1685         int ret, i;
1686         loff_t cluster_off;
1687         unsigned int local_len = len;
1688         struct ocfs2_write_cluster_desc *desc;
1689         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1690
1691         for (i = 0; i < wc->w_clen; i++) {
1692                 desc = &wc->w_desc[i];
1693
1694                 /*
1695                  * We have to make sure that the total write passed in
1696                  * doesn't extend past a single cluster.
1697                  */
1698                 local_len = len;
1699                 cluster_off = pos & (osb->s_clustersize - 1);
1700                 if ((cluster_off + local_len) > osb->s_clustersize)
1701                         local_len = osb->s_clustersize - cluster_off;
1702
1703                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1704                                           desc->c_unwritten,
1705                                           desc->c_needs_zero,
1706                                           data_ac, meta_ac,
1707                                           wc, desc->c_cpos, pos, local_len);
1708                 if (ret) {
1709                         mlog_errno(ret);
1710                         goto out;
1711                 }
1712
1713                 len -= local_len;
1714                 pos += local_len;
1715         }
1716
1717         ret = 0;
1718 out:
1719         return ret;
1720 }
1721
1722 /*
1723  * ocfs2_write_end() wants to know which parts of the target page it
1724  * should complete the write on. It's easiest to compute them ahead of
1725  * time when a more complete view of the write is available.
1726  */
1727 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1728                                         struct ocfs2_write_ctxt *wc,
1729                                         loff_t pos, unsigned len, int alloc)
1730 {
1731         struct ocfs2_write_cluster_desc *desc;
1732
1733         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1734         wc->w_target_to = wc->w_target_from + len;
1735
1736         if (alloc == 0)
1737                 return;
1738
1739         /*
1740          * Allocating write - we may have different boundaries based
1741          * on page size and cluster size.
1742          *
1743          * NOTE: We can no longer compute one value from the other as
1744          * the actual write length and user provided length may be
1745          * different.
1746          */
1747
1748         if (wc->w_large_pages) {
1749                 /*
1750                  * We only care about the 1st and last cluster within
1751                  * our range and whether they should be zero'd or not. Either
1752                  * value may be extended out to the start/end of a
1753                  * newly allocated cluster.
1754                  */
1755                 desc = &wc->w_desc[0];
1756                 if (desc->c_needs_zero)
1757                         ocfs2_figure_cluster_boundaries(osb,
1758                                                         desc->c_cpos,
1759                                                         &wc->w_target_from,
1760                                                         NULL);
1761
1762                 desc = &wc->w_desc[wc->w_clen - 1];
1763                 if (desc->c_needs_zero)
1764                         ocfs2_figure_cluster_boundaries(osb,
1765                                                         desc->c_cpos,
1766                                                         NULL,
1767                                                         &wc->w_target_to);
1768         } else {
1769                 wc->w_target_from = 0;
1770                 wc->w_target_to = PAGE_CACHE_SIZE;
1771         }
1772 }
1773
1774 /*
1775  * Populate each single-cluster write descriptor in the write context
1776  * with information about the i/o to be done.
1777  *
1778  * Returns the number of clusters that will have to be allocated, as
1779  * well as a worst case estimate of the number of extent records that
1780  * would have to be created during a write to an unwritten region.
1781  */
1782 static int ocfs2_populate_write_desc(struct inode *inode,
1783                                      struct ocfs2_write_ctxt *wc,
1784                                      unsigned int *clusters_to_alloc,
1785                                      unsigned int *extents_to_split)
1786 {
1787         int ret;
1788         struct ocfs2_write_cluster_desc *desc;
1789         unsigned int num_clusters = 0;
1790         unsigned int ext_flags = 0;
1791         u32 phys = 0;
1792         int i;
1793
1794         *clusters_to_alloc = 0;
1795         *extents_to_split = 0;
1796
1797         for (i = 0; i < wc->w_clen; i++) {
1798                 desc = &wc->w_desc[i];
1799                 desc->c_cpos = wc->w_cpos + i;
1800
1801                 if (num_clusters == 0) {
1802                         /*
1803                          * Need to look up the next extent record.
1804                          */
1805                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1806                                                  &num_clusters, &ext_flags);
1807                         if (ret) {
1808                                 mlog_errno(ret);
1809                                 goto out;
1810                         }
1811
1812                         /* We should already CoW the refcountd extent. */
1813                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1814
1815                         /*
1816                          * Assume worst case - that we're writing in
1817                          * the middle of the extent.
1818                          *
1819                          * We can assume that the write proceeds from
1820                          * left to right, in which case the extent
1821                          * insert code is smart enough to coalesce the
1822                          * next splits into the previous records created.
1823                          */
1824                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1825                                 *extents_to_split = *extents_to_split + 2;
1826                 } else if (phys) {
1827                         /*
1828                          * Only increment phys if it doesn't describe
1829                          * a hole.
1830                          */
1831                         phys++;
1832                 }
1833
1834                 /*
1835                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1836                  * file that got extended.  w_first_new_cpos tells us
1837                  * where the newly allocated clusters are so we can
1838                  * zero them.
1839                  */
1840                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1841                         BUG_ON(phys == 0);
1842                         desc->c_needs_zero = 1;
1843                 }
1844
1845                 desc->c_phys = phys;
1846                 if (phys == 0) {
1847                         desc->c_new = 1;
1848                         desc->c_needs_zero = 1;
1849                         *clusters_to_alloc = *clusters_to_alloc + 1;
1850                 }
1851
1852                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1853                         desc->c_unwritten = 1;
1854                         desc->c_needs_zero = 1;
1855                 }
1856
1857                 num_clusters--;
1858         }
1859
1860         ret = 0;
1861 out:
1862         return ret;
1863 }
1864
1865 static int ocfs2_write_begin_inline(struct address_space *mapping,
1866                                     struct inode *inode,
1867                                     struct ocfs2_write_ctxt *wc)
1868 {
1869         int ret;
1870         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1871         struct page *page;
1872         handle_t *handle;
1873         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1874
1875         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1876         if (IS_ERR(handle)) {
1877                 ret = PTR_ERR(handle);
1878                 mlog_errno(ret);
1879                 goto out;
1880         }
1881
1882         page = find_or_create_page(mapping, 0, GFP_NOFS);
1883         if (!page) {
1884                 ocfs2_commit_trans(osb, handle);
1885                 ret = -ENOMEM;
1886                 mlog_errno(ret);
1887                 goto out;
1888         }
1889         /*
1890          * If we don't set w_num_pages then this page won't get unlocked
1891          * and freed on cleanup of the write context.
1892          */
1893         wc->w_pages[0] = wc->w_target_page = page;
1894         wc->w_num_pages = 1;
1895
1896         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1897                                       OCFS2_JOURNAL_ACCESS_WRITE);
1898         if (ret) {
1899                 ocfs2_commit_trans(osb, handle);
1900
1901                 mlog_errno(ret);
1902                 goto out;
1903         }
1904
1905         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1906                 ocfs2_set_inode_data_inline(inode, di);
1907
1908         if (!PageUptodate(page)) {
1909                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1910                 if (ret) {
1911                         ocfs2_commit_trans(osb, handle);
1912
1913                         goto out;
1914                 }
1915         }
1916
1917         wc->w_handle = handle;
1918 out:
1919         return ret;
1920 }
1921
1922 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1923 {
1924         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1925
1926         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1927                 return 1;
1928         return 0;
1929 }
1930
1931 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1932                                           struct inode *inode, loff_t pos,
1933                                           unsigned len, struct page *mmap_page,
1934                                           struct ocfs2_write_ctxt *wc)
1935 {
1936         int ret, written = 0;
1937         loff_t end = pos + len;
1938         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1939         struct ocfs2_dinode *di = NULL;
1940
1941         trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1942                                              len, (unsigned long long)pos,
1943                                              oi->ip_dyn_features);
1944
1945         /*
1946          * Handle inodes which already have inline data 1st.
1947          */
1948         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1949                 if (mmap_page == NULL &&
1950                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1951                         goto do_inline_write;
1952
1953                 /*
1954                  * The write won't fit - we have to give this inode an
1955                  * inline extent list now.
1956                  */
1957                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1958                 if (ret)
1959                         mlog_errno(ret);
1960                 goto out;
1961         }
1962
1963         /*
1964          * Check whether the inode can accept inline data.
1965          */
1966         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1967                 return 0;
1968
1969         /*
1970          * Check whether the write can fit.
1971          */
1972         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1973         if (mmap_page ||
1974             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1975                 return 0;
1976
1977 do_inline_write:
1978         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1979         if (ret) {
1980                 mlog_errno(ret);
1981                 goto out;
1982         }
1983
1984         /*
1985          * This signals to the caller that the data can be written
1986          * inline.
1987          */
1988         written = 1;
1989 out:
1990         return written ? written : ret;
1991 }
1992
1993 /*
1994  * This function only does anything for file systems which can't
1995  * handle sparse files.
1996  *
1997  * What we want to do here is fill in any hole between the current end
1998  * of allocation and the end of our write. That way the rest of the
1999  * write path can treat it as an non-allocating write, which has no
2000  * special case code for sparse/nonsparse files.
2001  */
2002 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
2003                                         struct buffer_head *di_bh,
2004                                         loff_t pos, unsigned len,
2005                                         struct ocfs2_write_ctxt *wc)
2006 {
2007         int ret;
2008         loff_t newsize = pos + len;
2009
2010         BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
2011
2012         if (newsize <= i_size_read(inode))
2013                 return 0;
2014
2015         ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
2016         if (ret)
2017                 mlog_errno(ret);
2018
2019         wc->w_first_new_cpos =
2020                 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
2021
2022         return ret;
2023 }
2024
2025 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
2026                            loff_t pos)
2027 {
2028         int ret = 0;
2029
2030         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
2031         if (pos > i_size_read(inode))
2032                 ret = ocfs2_zero_extend(inode, di_bh, pos);
2033
2034         return ret;
2035 }
2036
2037 /*
2038  * Try to flush truncate logs if we can free enough clusters from it.
2039  * As for return value, "< 0" means error, "0" no space and "1" means
2040  * we have freed enough spaces and let the caller try to allocate again.
2041  */
2042 static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
2043                                           unsigned int needed)
2044 {
2045         tid_t target;
2046         int ret = 0;
2047         unsigned int truncated_clusters;
2048
2049         inode_lock(osb->osb_tl_inode);
2050         truncated_clusters = osb->truncated_clusters;
2051         inode_unlock(osb->osb_tl_inode);
2052
2053         /*
2054          * Check whether we can succeed in allocating if we free
2055          * the truncate log.
2056          */
2057         if (truncated_clusters < needed)
2058                 goto out;
2059
2060         ret = ocfs2_flush_truncate_log(osb);
2061         if (ret) {
2062                 mlog_errno(ret);
2063                 goto out;
2064         }
2065
2066         if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
2067                 jbd2_log_wait_commit(osb->journal->j_journal, target);
2068                 ret = 1;
2069         }
2070 out:
2071         return ret;
2072 }
2073
2074 int ocfs2_write_begin_nolock(struct file *filp,
2075                              struct address_space *mapping,
2076                              loff_t pos, unsigned len, unsigned flags,
2077                              struct page **pagep, void **fsdata,
2078                              struct buffer_head *di_bh, struct page *mmap_page)
2079 {
2080         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
2081         unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
2082         struct ocfs2_write_ctxt *wc;
2083         struct inode *inode = mapping->host;
2084         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2085         struct ocfs2_dinode *di;
2086         struct ocfs2_alloc_context *data_ac = NULL;
2087         struct ocfs2_alloc_context *meta_ac = NULL;
2088         handle_t *handle;
2089         struct ocfs2_extent_tree et;
2090         int try_free = 1, ret1;
2091
2092 try_again:
2093         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
2094         if (ret) {
2095                 mlog_errno(ret);
2096                 return ret;
2097         }
2098
2099         if (ocfs2_supports_inline_data(osb)) {
2100                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
2101                                                      mmap_page, wc);
2102                 if (ret == 1) {
2103                         ret = 0;
2104                         goto success;
2105                 }
2106                 if (ret < 0) {
2107                         mlog_errno(ret);
2108                         goto out;
2109                 }
2110         }
2111
2112         if (ocfs2_sparse_alloc(osb))
2113                 ret = ocfs2_zero_tail(inode, di_bh, pos);
2114         else
2115                 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len,
2116                                                    wc);
2117         if (ret) {
2118                 mlog_errno(ret);
2119                 goto out;
2120         }
2121
2122         ret = ocfs2_check_range_for_refcount(inode, pos, len);
2123         if (ret < 0) {
2124                 mlog_errno(ret);
2125                 goto out;
2126         } else if (ret == 1) {
2127                 clusters_need = wc->w_clen;
2128                 ret = ocfs2_refcount_cow(inode, di_bh,
2129                                          wc->w_cpos, wc->w_clen, UINT_MAX);
2130                 if (ret) {
2131                         mlog_errno(ret);
2132                         goto out;
2133                 }
2134         }
2135
2136         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
2137                                         &extents_to_split);
2138         if (ret) {
2139                 mlog_errno(ret);
2140                 goto out;
2141         }
2142         clusters_need += clusters_to_alloc;
2143
2144         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
2145
2146         trace_ocfs2_write_begin_nolock(
2147                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
2148                         (long long)i_size_read(inode),
2149                         le32_to_cpu(di->i_clusters),
2150                         pos, len, flags, mmap_page,
2151                         clusters_to_alloc, extents_to_split);
2152
2153         /*
2154          * We set w_target_from, w_target_to here so that
2155          * ocfs2_write_end() knows which range in the target page to
2156          * write out. An allocation requires that we write the entire
2157          * cluster range.
2158          */
2159         if (clusters_to_alloc || extents_to_split) {
2160                 /*
2161                  * XXX: We are stretching the limits of
2162                  * ocfs2_lock_allocators(). It greatly over-estimates
2163                  * the work to be done.
2164                  */
2165                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
2166                                               wc->w_di_bh);
2167                 ret = ocfs2_lock_allocators(inode, &et,
2168                                             clusters_to_alloc, extents_to_split,
2169                                             &data_ac, &meta_ac);
2170                 if (ret) {
2171                         mlog_errno(ret);
2172                         goto out;
2173                 }
2174
2175                 if (data_ac)
2176                         data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
2177
2178                 credits = ocfs2_calc_extend_credits(inode->i_sb,
2179                                                     &di->id2.i_list);
2180
2181         }
2182
2183         /*
2184          * We have to zero sparse allocated clusters, unwritten extent clusters,
2185          * and non-sparse clusters we just extended.  For non-sparse writes,
2186          * we know zeros will only be needed in the first and/or last cluster.
2187          */
2188         if (clusters_to_alloc || extents_to_split ||
2189             (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
2190                             wc->w_desc[wc->w_clen - 1].c_needs_zero)))
2191                 cluster_of_pages = 1;
2192         else
2193                 cluster_of_pages = 0;
2194
2195         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
2196
2197         handle = ocfs2_start_trans(osb, credits);
2198         if (IS_ERR(handle)) {
2199                 ret = PTR_ERR(handle);
2200                 mlog_errno(ret);
2201                 goto out;
2202         }
2203
2204         wc->w_handle = handle;
2205
2206         if (clusters_to_alloc) {
2207                 ret = dquot_alloc_space_nodirty(inode,
2208                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
2209                 if (ret)
2210                         goto out_commit;
2211         }
2212
2213         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
2214                                       OCFS2_JOURNAL_ACCESS_WRITE);
2215         if (ret) {
2216                 mlog_errno(ret);
2217                 goto out_quota;
2218         }
2219
2220         /*
2221          * Fill our page array first. That way we've grabbed enough so
2222          * that we can zero and flush if we error after adding the
2223          * extent.
2224          */
2225         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
2226                                          cluster_of_pages, mmap_page);
2227         if (ret && ret != -EAGAIN) {
2228                 mlog_errno(ret);
2229                 goto out_quota;
2230         }
2231
2232         /*
2233          * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
2234          * the target page. In this case, we exit with no error and no target
2235          * page. This will trigger the caller, page_mkwrite(), to re-try
2236          * the operation.
2237          */
2238         if (ret == -EAGAIN) {
2239                 BUG_ON(wc->w_target_page);
2240                 ret = 0;
2241                 goto out_quota;
2242         }
2243
2244         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
2245                                           len);
2246         if (ret) {
2247                 mlog_errno(ret);
2248                 goto out_quota;
2249         }
2250
2251         if (data_ac)
2252                 ocfs2_free_alloc_context(data_ac);
2253         if (meta_ac)
2254                 ocfs2_free_alloc_context(meta_ac);
2255
2256 success:
2257         *pagep = wc->w_target_page;
2258         *fsdata = wc;
2259         return 0;
2260 out_quota:
2261         if (clusters_to_alloc)
2262                 dquot_free_space(inode,
2263                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
2264 out_commit:
2265         ocfs2_commit_trans(osb, handle);
2266
2267 out:
2268         ocfs2_free_write_ctxt(wc);
2269
2270         if (data_ac) {
2271                 ocfs2_free_alloc_context(data_ac);
2272                 data_ac = NULL;
2273         }
2274         if (meta_ac) {
2275                 ocfs2_free_alloc_context(meta_ac);
2276                 meta_ac = NULL;
2277         }
2278
2279         if (ret == -ENOSPC && try_free) {
2280                 /*
2281                  * Try to free some truncate log so that we can have enough
2282                  * clusters to allocate.
2283                  */
2284                 try_free = 0;
2285
2286                 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
2287                 if (ret1 == 1)
2288                         goto try_again;
2289
2290                 if (ret1 < 0)
2291                         mlog_errno(ret1);
2292         }
2293
2294         return ret;
2295 }
2296
2297 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
2298                              loff_t pos, unsigned len, unsigned flags,
2299                              struct page **pagep, void **fsdata)
2300 {
2301         int ret;
2302         struct buffer_head *di_bh = NULL;
2303         struct inode *inode = mapping->host;
2304
2305         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2306         if (ret) {
2307                 mlog_errno(ret);
2308                 return ret;
2309         }
2310
2311         /*
2312          * Take alloc sem here to prevent concurrent lookups. That way
2313          * the mapping, zeroing and tree manipulation within
2314          * ocfs2_write() will be safe against ->readpage(). This
2315          * should also serve to lock out allocation from a shared
2316          * writeable region.
2317          */
2318         down_write(&OCFS2_I(inode)->ip_alloc_sem);
2319
2320         ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep,
2321                                        fsdata, di_bh, NULL);
2322         if (ret) {
2323                 mlog_errno(ret);
2324                 goto out_fail;
2325         }
2326
2327         brelse(di_bh);
2328
2329         return 0;
2330
2331 out_fail:
2332         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2333
2334         brelse(di_bh);
2335         ocfs2_inode_unlock(inode, 1);
2336
2337         return ret;
2338 }
2339
2340 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
2341                                    unsigned len, unsigned *copied,
2342                                    struct ocfs2_dinode *di,
2343                                    struct ocfs2_write_ctxt *wc)
2344 {
2345         void *kaddr;
2346
2347         if (unlikely(*copied < len)) {
2348                 if (!PageUptodate(wc->w_target_page)) {
2349                         *copied = 0;
2350                         return;
2351                 }
2352         }
2353
2354         kaddr = kmap_atomic(wc->w_target_page);
2355         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
2356         kunmap_atomic(kaddr);
2357
2358         trace_ocfs2_write_end_inline(
2359              (unsigned long long)OCFS2_I(inode)->ip_blkno,
2360              (unsigned long long)pos, *copied,
2361              le16_to_cpu(di->id2.i_data.id_count),
2362              le16_to_cpu(di->i_dyn_features));
2363 }
2364
2365 int ocfs2_write_end_nolock(struct address_space *mapping,
2366                            loff_t pos, unsigned len, unsigned copied,
2367                            struct page *page, void *fsdata)
2368 {
2369         int i, ret;
2370         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
2371         struct inode *inode = mapping->host;
2372         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2373         struct ocfs2_write_ctxt *wc = fsdata;
2374         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
2375         handle_t *handle = wc->w_handle;
2376         struct page *tmppage;
2377
2378         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
2379                         OCFS2_JOURNAL_ACCESS_WRITE);
2380         if (ret) {
2381                 copied = ret;
2382                 mlog_errno(ret);
2383                 goto out;
2384         }
2385
2386         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2387                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
2388                 goto out_write_size;
2389         }
2390
2391         if (unlikely(copied < len)) {
2392                 if (!PageUptodate(wc->w_target_page))
2393                         copied = 0;
2394
2395                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
2396                                        start+len);
2397         }
2398         flush_dcache_page(wc->w_target_page);
2399
2400         for(i = 0; i < wc->w_num_pages; i++) {
2401                 tmppage = wc->w_pages[i];
2402
2403                 if (tmppage == wc->w_target_page) {
2404                         from = wc->w_target_from;
2405                         to = wc->w_target_to;
2406
2407                         BUG_ON(from > PAGE_CACHE_SIZE ||
2408                                to > PAGE_CACHE_SIZE ||
2409                                to < from);
2410                 } else {
2411                         /*
2412                          * Pages adjacent to the target (if any) imply
2413                          * a hole-filling write in which case we want
2414                          * to flush their entire range.
2415                          */
2416                         from = 0;
2417                         to = PAGE_CACHE_SIZE;
2418                 }
2419
2420                 if (page_has_buffers(tmppage)) {
2421                         if (ocfs2_should_order_data(inode))
2422                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
2423                         block_commit_write(tmppage, from, to);
2424                 }
2425         }
2426
2427 out_write_size:
2428         pos += copied;
2429         if (pos > i_size_read(inode)) {
2430                 i_size_write(inode, pos);
2431                 mark_inode_dirty(inode);
2432         }
2433         inode->i_blocks = ocfs2_inode_sector_count(inode);
2434         di->i_size = cpu_to_le64((u64)i_size_read(inode));
2435         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2436         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2437         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2438         ocfs2_update_inode_fsync_trans(handle, inode, 1);
2439         ocfs2_journal_dirty(handle, wc->w_di_bh);
2440
2441 out:
2442         /* unlock pages before dealloc since it needs acquiring j_trans_barrier
2443          * lock, or it will cause a deadlock since journal commit threads holds
2444          * this lock and will ask for the page lock when flushing the data.
2445          * put it here to preserve the unlock order.
2446          */
2447         ocfs2_unlock_pages(wc);
2448
2449         ocfs2_commit_trans(osb, handle);
2450
2451         ocfs2_run_deallocs(osb, &wc->w_dealloc);
2452
2453         brelse(wc->w_di_bh);
2454         kfree(wc);
2455
2456         return copied;
2457 }
2458
2459 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2460                            loff_t pos, unsigned len, unsigned copied,
2461                            struct page *page, void *fsdata)
2462 {
2463         int ret;
2464         struct inode *inode = mapping->host;
2465
2466         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
2467
2468         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2469         ocfs2_inode_unlock(inode, 1);
2470
2471         return ret;
2472 }
2473
2474 const struct address_space_operations ocfs2_aops = {
2475         .readpage               = ocfs2_readpage,
2476         .readpages              = ocfs2_readpages,
2477         .writepage              = ocfs2_writepage,
2478         .write_begin            = ocfs2_write_begin,
2479         .write_end              = ocfs2_write_end,
2480         .bmap                   = ocfs2_bmap,
2481         .direct_IO              = ocfs2_direct_IO,
2482         .invalidatepage         = block_invalidatepage,
2483         .releasepage            = ocfs2_releasepage,
2484         .migratepage            = buffer_migrate_page,
2485         .is_partially_uptodate  = block_is_partially_uptodate,
2486         .error_remove_page      = generic_error_remove_page,
2487 };