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